Merge "Don't leak oat file when we fail to open a dex file."
diff --git a/Android.mk b/Android.mk
index 4351be9..92339af 100644
--- a/Android.mk
+++ b/Android.mk
@@ -46,9 +46,6 @@
 	rm -f $(ART_TEST_OUT)/*.odex
 	rm -f $(ART_TEST_OUT)/*.oat
 	rm -f $(ART_TEST_OUT)/*.art
-	rm -f $(DALVIK_CACHE_OUT)/*@classes.dex
-	rm -f $(DALVIK_CACHE_OUT)/*.oat
-	rm -f $(DALVIK_CACHE_OUT)/*.art
 	rm -f $(HOST_OUT_JAVA_LIBRARIES)/*.odex
 	rm -f $(HOST_OUT_JAVA_LIBRARIES)/*.oat
 	rm -f $(HOST_OUT_JAVA_LIBRARIES)/*.art
@@ -66,25 +63,31 @@
 	rm -f $(2ND_TARGET_OUT_INTERMEDIATES)/JAVA_LIBRARIES/*_intermediates/javalib.odex
 	rm -f $(2ND_TARGET_OUT_INTERMEDIATES)/APPS/*_intermediates/*.odex
 endif
-	rm -rf /tmp/test-*/dalvik-cache/*@classes.dex
-	rm -rf /tmp/android-data/dalvik-cache/*@classes.dex
+	rm -rf /tmp/test-*/dalvik-cache/*
+	rm -rf /tmp/android-data/dalvik-cache/*
 
 .PHONY: clean-oat-target
 clean-oat-target:
 	adb remount
-	adb shell rm $(ART_NATIVETEST_DIR)/*.odex
-	adb shell rm $(ART_NATIVETEST_DIR)/*.oat
-	adb shell rm $(ART_NATIVETEST_DIR)/*.art
-	adb shell rm $(ART_TEST_DIR)/*.odex
-	adb shell rm $(ART_TEST_DIR)/*.oat
-	adb shell rm $(ART_TEST_DIR)/*.art
-	adb shell rm $(ART_DALVIK_CACHE_DIR)/*.dex
-	adb shell rm $(ART_DALVIK_CACHE_DIR)/*.oat
-	adb shell rm $(ART_DALVIK_CACHE_DIR)/*.art
-	adb shell rm $(DEXPREOPT_BOOT_JAR_DIR)/*.oat
-	adb shell rm $(DEXPREOPT_BOOT_JAR_DIR)/*.art
-	adb shell rm system/app/*.odex
-	adb shell rm data/run-test/test-*/dalvik-cache/*@classes.dex
+	adb shell rm -f $(ART_NATIVETEST_DIR)/*.odex
+	adb shell rm -f $(ART_NATIVETEST_DIR)/*.oat
+	adb shell rm -f $(ART_NATIVETEST_DIR)/*.art
+	adb shell rm -f $(ART_TEST_DIR)/*.odex
+	adb shell rm -f $(ART_TEST_DIR)/*.oat
+	adb shell rm -f $(ART_TEST_DIR)/*.art
+ifdef TARGET_2ND_ARCH
+	adb shell rm -f $(2ND_ART_NATIVETEST_DIR)/*.odex
+	adb shell rm -f $(2ND_ART_NATIVETEST_DIR)/*.oat
+	adb shell rm -f $(2ND_ART_NATIVETEST_DIR)/*.art
+	adb shell rm -f $(2ND_ART_TEST_DIR)/*.odex
+	adb shell rm -f $(2ND_ART_TEST_DIR)/*.oat
+	adb shell rm -f $(2ND_ART_TEST_DIR)/*.art
+endif
+	adb shell rm -rf $(ART_DALVIK_CACHE_DIR)/*
+	adb shell rm -f $(DEXPREOPT_BOOT_JAR_DIR)/*.oat
+	adb shell rm -f $(DEXPREOPT_BOOT_JAR_DIR)/*.art
+	adb shell rm -f system/app/*.odex
+	adb shell rm -rf data/run-test/test-*/dalvik-cache/*
 
 ifneq ($(art_dont_bother),true)
 
diff --git a/compiler/dex/quick/codegen_util.cc b/compiler/dex/quick/codegen_util.cc
index c3f9b67..b030bb4 100644
--- a/compiler/dex/quick/codegen_util.cc
+++ b/compiler/dex/quick/codegen_util.cc
@@ -555,8 +555,8 @@
   return offset;
 }
 
-static int AssignLiteralPointerOffsetCommon(LIR* lir, CodeOffset offset) {
-  unsigned int element_size = sizeof(void*);
+static int AssignLiteralPointerOffsetCommon(LIR* lir, CodeOffset offset,
+                                            unsigned int element_size) {
   // Align to natural pointer size.
   offset = (offset + (element_size - 1)) & ~(element_size - 1);
   for (; lir != NULL; lir = lir->next) {
@@ -726,9 +726,10 @@
 /* Determine the offset of each literal field */
 int Mir2Lir::AssignLiteralOffset(CodeOffset offset) {
   offset = AssignLiteralOffsetCommon(literal_list_, offset);
-  offset = AssignLiteralPointerOffsetCommon(code_literal_list_, offset);
-  offset = AssignLiteralPointerOffsetCommon(method_literal_list_, offset);
-  offset = AssignLiteralPointerOffsetCommon(class_literal_list_, offset);
+  unsigned int ptr_size = GetInstructionSetPointerSize(cu_->instruction_set);
+  offset = AssignLiteralPointerOffsetCommon(code_literal_list_, offset, ptr_size);
+  offset = AssignLiteralPointerOffsetCommon(method_literal_list_, offset, ptr_size);
+  offset = AssignLiteralPointerOffsetCommon(class_literal_list_, offset, ptr_size);
   return offset;
 }
 
diff --git a/compiler/dex/quick/gen_invoke.cc b/compiler/dex/quick/gen_invoke.cc
index 05313a9..93a23a6 100644
--- a/compiler/dex/quick/gen_invoke.cc
+++ b/compiler/dex/quick/gen_invoke.cc
@@ -455,14 +455,14 @@
   if (direct_code != 0 && direct_method != 0) {
     switch (state) {
     case 0:  // Get the current Method* [sets kArg0]
-      if (direct_code != static_cast<unsigned int>(-1)) {
+      if (direct_code != static_cast<uintptr_t>(-1)) {
         if (cu->instruction_set != kX86 && cu->instruction_set != kX86_64) {
           cg->LoadConstant(cg->TargetReg(kInvokeTgt), direct_code);
         }
       } else if (cu->instruction_set != kX86 && cu->instruction_set != kX86_64) {
         cg->LoadCodeAddress(target_method, type, kInvokeTgt);
       }
-      if (direct_method != static_cast<unsigned int>(-1)) {
+      if (direct_method != static_cast<uintptr_t>(-1)) {
         cg->LoadConstant(cg->TargetReg(kArg0), direct_method);
       } else {
         cg->LoadMethodAddress(target_method, type, kArg0);
@@ -483,7 +483,7 @@
                       cg->TargetReg(kArg0));
       // Set up direct code if known.
       if (direct_code != 0) {
-        if (direct_code != static_cast<unsigned int>(-1)) {
+        if (direct_code != static_cast<uintptr_t>(-1)) {
           cg->LoadConstant(cg->TargetReg(kInvokeTgt), direct_code);
         } else if (cu->instruction_set != kX86 && cu->instruction_set != kX86_64) {
           CHECK_LT(target_method.dex_method_index, target_method.dex_file->NumMethodIds());
diff --git a/compiler/dex/quick/mir_to_lir.cc b/compiler/dex/quick/mir_to_lir.cc
index df7a7c1..6c5279e 100644
--- a/compiler/dex/quick/mir_to_lir.cc
+++ b/compiler/dex/quick/mir_to_lir.cc
@@ -42,7 +42,7 @@
   RegStorage reg_arg_high = wide ? GetArgMappingToPhysicalReg(in_position + 1) :
       RegStorage::InvalidReg();
 
-  int offset = StackVisitor::GetOutVROffset(in_position);
+  int offset = StackVisitor::GetOutVROffset(in_position, cu_->instruction_set);
   if (cu_->instruction_set == kX86 || cu_->instruction_set == kX86_64) {
     /*
      * When doing a call for x86, it moves the stack pointer in order to push return.
@@ -81,7 +81,7 @@
 }
 
 void Mir2Lir::LoadArgDirect(int in_position, RegLocation rl_dest) {
-  int offset = StackVisitor::GetOutVROffset(in_position);
+  int offset = StackVisitor::GetOutVROffset(in_position, cu_->instruction_set);
   if (cu_->instruction_set == kX86 || cu_->instruction_set == kX86_64) {
     /*
      * When doing a call for x86, it moves the stack pointer in order to push return.
diff --git a/compiler/dex/quick/ralloc_util.cc b/compiler/dex/quick/ralloc_util.cc
index 39783a2..6455572 100644
--- a/compiler/dex/quick/ralloc_util.cc
+++ b/compiler/dex/quick/ralloc_util.cc
@@ -1141,7 +1141,8 @@
 /* Returns sp-relative offset in bytes for a VReg */
 int Mir2Lir::VRegOffset(int v_reg) {
   return StackVisitor::GetVRegOffset(cu_->code_item, core_spill_mask_,
-                                     fp_spill_mask_, frame_size_, v_reg);
+                                     fp_spill_mask_, frame_size_, v_reg,
+                                     cu_->instruction_set);
 }
 
 /* Returns sp-relative offset in bytes for a SReg */
diff --git a/compiler/jni/quick/arm/calling_convention_arm.h b/compiler/jni/quick/arm/calling_convention_arm.h
index 00a239b..604ce1c 100644
--- a/compiler/jni/quick/arm/calling_convention_arm.h
+++ b/compiler/jni/quick/arm/calling_convention_arm.h
@@ -71,6 +71,11 @@
   ManagedRegister CurrentParamRegister() OVERRIDE;
   FrameOffset CurrentParamStackOffset() OVERRIDE;
 
+  // AAPCS mandates return values are extended.
+  bool RequiresSmallResultTypeExtension() const OVERRIDE {
+    return false;
+  }
+
  protected:
   size_t NumberOfOutgoingStackArgs() OVERRIDE;
 
diff --git a/compiler/jni/quick/arm64/calling_convention_arm64.h b/compiler/jni/quick/arm64/calling_convention_arm64.h
index 92f547c..9fd3265 100644
--- a/compiler/jni/quick/arm64/calling_convention_arm64.h
+++ b/compiler/jni/quick/arm64/calling_convention_arm64.h
@@ -68,6 +68,11 @@
   ManagedRegister CurrentParamRegister() OVERRIDE;
   FrameOffset CurrentParamStackOffset() OVERRIDE;
 
+  // aarch64 calling convention leaves upper bits undefined.
+  bool RequiresSmallResultTypeExtension() const OVERRIDE {
+    return true;
+  }
+
  protected:
   size_t NumberOfOutgoingStackArgs() OVERRIDE;
 
diff --git a/compiler/jni/quick/calling_convention.h b/compiler/jni/quick/calling_convention.h
index 4d25d1c..18afd58 100644
--- a/compiler/jni/quick/calling_convention.h
+++ b/compiler/jni/quick/calling_convention.h
@@ -287,6 +287,8 @@
   FrameOffset ReturnValueSaveLocation() const;
   // Register that holds result if it is integer.
   virtual ManagedRegister IntReturnRegister() = 0;
+  // Whether the compiler needs to ensure zero-/sign-extension of a small result type
+  virtual bool RequiresSmallResultTypeExtension() const = 0;
 
   // Callee save registers to spill prior to native code (which may clobber)
   virtual const std::vector<ManagedRegister>& CalleeSaveRegisters() const = 0;
diff --git a/compiler/jni/quick/jni_compiler.cc b/compiler/jni/quick/jni_compiler.cc
index 93b1b5a..9f439eb 100644
--- a/compiler/jni/quick/jni_compiler.cc
+++ b/compiler/jni/quick/jni_compiler.cc
@@ -314,7 +314,7 @@
           mr_conv->InterproceduralScratchRegister());
 
   // 10. Fix differences in result widths.
-  if (instruction_set == kX86 || instruction_set == kX86_64) {
+  if (main_jni_conv->RequiresSmallResultTypeExtension()) {
     if (main_jni_conv->GetReturnType() == Primitive::kPrimByte ||
         main_jni_conv->GetReturnType() == Primitive::kPrimShort) {
       __ SignExtend(main_jni_conv->ReturnRegister(),
diff --git a/compiler/jni/quick/mips/calling_convention_mips.h b/compiler/jni/quick/mips/calling_convention_mips.h
index e33fbad..8d82dce 100644
--- a/compiler/jni/quick/mips/calling_convention_mips.h
+++ b/compiler/jni/quick/mips/calling_convention_mips.h
@@ -71,6 +71,11 @@
   ManagedRegister CurrentParamRegister() OVERRIDE;
   FrameOffset CurrentParamStackOffset() OVERRIDE;
 
+  // Mips does not need to extend small return types.
+  bool RequiresSmallResultTypeExtension() const OVERRIDE {
+    return false;
+  }
+
  protected:
   size_t NumberOfOutgoingStackArgs() OVERRIDE;
 
diff --git a/compiler/jni/quick/x86/calling_convention_x86.h b/compiler/jni/quick/x86/calling_convention_x86.h
index 5b9069c..025eb6d 100644
--- a/compiler/jni/quick/x86/calling_convention_x86.h
+++ b/compiler/jni/quick/x86/calling_convention_x86.h
@@ -69,6 +69,11 @@
   ManagedRegister CurrentParamRegister() OVERRIDE;
   FrameOffset CurrentParamStackOffset() OVERRIDE;
 
+  // x86 needs to extend small return types.
+  bool RequiresSmallResultTypeExtension() const OVERRIDE {
+    return true;
+  }
+
  protected:
   size_t NumberOfOutgoingStackArgs() OVERRIDE;
 
diff --git a/compiler/jni/quick/x86_64/calling_convention_x86_64.h b/compiler/jni/quick/x86_64/calling_convention_x86_64.h
index d545774..1ba5353 100644
--- a/compiler/jni/quick/x86_64/calling_convention_x86_64.h
+++ b/compiler/jni/quick/x86_64/calling_convention_x86_64.h
@@ -69,6 +69,11 @@
   ManagedRegister CurrentParamRegister() OVERRIDE;
   FrameOffset CurrentParamStackOffset() OVERRIDE;
 
+  // x86-64 needs to extend small return types.
+  bool RequiresSmallResultTypeExtension() const OVERRIDE {
+    return true;
+  }
+
  protected:
   size_t NumberOfOutgoingStackArgs() OVERRIDE;
 
diff --git a/compiler/utils/arm64/assembler_arm64.cc b/compiler/utils/arm64/assembler_arm64.cc
index 1d87eaa..b4bb979 100644
--- a/compiler/utils/arm64/assembler_arm64.cc
+++ b/compiler/utils/arm64/assembler_arm64.cc
@@ -467,12 +467,26 @@
 #endif
 }
 
-void Arm64Assembler::SignExtend(ManagedRegister /*mreg*/, size_t /*size*/) {
-  UNIMPLEMENTED(FATAL) << "no sign extension necessary for Arm64";
+void Arm64Assembler::SignExtend(ManagedRegister mreg, size_t size) {
+  Arm64ManagedRegister reg = mreg.AsArm64();
+  CHECK(size == 1 || size == 2) << size;
+  CHECK(reg.IsWRegister()) << reg;
+  if (size == 1) {
+    ___ sxtb(reg_w(reg.AsWRegister()), reg_w(reg.AsWRegister()));
+  } else {
+    ___ sxth(reg_w(reg.AsWRegister()), reg_w(reg.AsWRegister()));
+  }
 }
 
-void Arm64Assembler::ZeroExtend(ManagedRegister /*mreg*/, size_t /*size*/) {
-  UNIMPLEMENTED(FATAL) << "no zero extension necessary for Arm64";
+void Arm64Assembler::ZeroExtend(ManagedRegister mreg, size_t size) {
+  Arm64ManagedRegister reg = mreg.AsArm64();
+  CHECK(size == 1 || size == 2) << size;
+  CHECK(reg.IsWRegister()) << reg;
+  if (size == 1) {
+    ___ uxtb(reg_w(reg.AsWRegister()), reg_w(reg.AsWRegister()));
+  } else {
+    ___ uxth(reg_w(reg.AsWRegister()), reg_w(reg.AsWRegister()));
+  }
 }
 
 void Arm64Assembler::VerifyObject(ManagedRegister /*src*/, bool /*could_be_null*/) {
diff --git a/dex2oat/dex2oat.cc b/dex2oat/dex2oat.cc
index 823b818..cdf26f1 100644
--- a/dex2oat/dex2oat.cc
+++ b/dex2oat/dex2oat.cc
@@ -1010,8 +1010,8 @@
   }
 
   if (compiler_filter_string == NULL) {
-    if (instruction_set == kX86_64 || instruction_set == kArm64) {
-      // TODO: currently x86-64 and arm64 are only interpreted.
+    if (instruction_set == kX86_64 || instruction_set == kArm64 || instruction_set == kMips) {
+      // TODO: implement/fix compilers for these architectures.
       compiler_filter_string = "interpret-only";
     } else if (image) {
       compiler_filter_string = "speed";
diff --git a/disassembler/disassembler_arm.cc b/disassembler/disassembler_arm.cc
index d6d2058..4e4a512 100644
--- a/disassembler/disassembler_arm.cc
+++ b/disassembler/disassembler_arm.cc
@@ -1360,7 +1360,7 @@
               } else if (Rn.r == 15) {
                 intptr_t lit_adr = reinterpret_cast<intptr_t>(instr_ptr);
                 lit_adr = RoundDown(lit_adr, 4) + 4 + imm12;
-                args << "  ; " << reinterpret_cast<void*>(*reinterpret_cast<int32_t*>(lit_adr));
+                args << StringPrintf("  ; 0x%08x", *reinterpret_cast<int32_t*>(lit_adr));
               }
             } else if (op3 == 3) {
               // LDRSH.W Rt, [Rn, #imm12]      - 111 11 00 11 011 nnnn tttt iiiiiiiiiiii
@@ -1373,7 +1373,7 @@
               } else if (Rn.r == 15) {
                 intptr_t lit_adr = reinterpret_cast<intptr_t>(instr_ptr);
                 lit_adr = RoundDown(lit_adr, 4) + 4 + imm12;
-                args << "  ; " << reinterpret_cast<void*>(*reinterpret_cast<int32_t*>(lit_adr));
+                args << StringPrintf("  ; 0x%08x", *reinterpret_cast<int32_t*>(lit_adr));
               }
             }
           }
@@ -1430,7 +1430,7 @@
             } else if (Rn.r == 15) {
               intptr_t lit_adr = reinterpret_cast<intptr_t>(instr_ptr);
               lit_adr = RoundDown(lit_adr, 4) + 4 + imm12;
-              args << "  ; " << reinterpret_cast<void*>(*reinterpret_cast<int32_t*>(lit_adr));
+              args << StringPrintf("  ; 0x%08x", *reinterpret_cast<int32_t*>(lit_adr));
             }
           } else if (op4 == 0) {
             // LDR.W Rt, [Rn, Rm{, LSL #imm2}] - 111 11 00 00 101 nnnn tttt 000000iimmmm
diff --git a/oatdump/oatdump.cc b/oatdump/oatdump.cc
index 1a67952..412a052 100644
--- a/oatdump/oatdump.cc
+++ b/oatdump/oatdump.cc
@@ -495,7 +495,8 @@
       } else {
         uint32_t offset = StackVisitor::GetVRegOffset(code_item, oat_method.GetCoreSpillMask(),
                                                       oat_method.GetFpSpillMask(),
-                                                      oat_method.GetFrameSizeInBytes(), reg);
+                                                      oat_method.GetFrameSizeInBytes(), reg,
+                                                      GetInstructionSet());
         os << "[sp + #" << offset << "]";
       }
     }
diff --git a/runtime/arch/stub_test.cc b/runtime/arch/stub_test.cc
index 7027b32..437beb5 100644
--- a/runtime/arch/stub_test.cc
+++ b/runtime/arch/stub_test.cc
@@ -633,4 +633,79 @@
 #endif
 }
 
+
+#if defined(__i386__) || defined(__arm__) || defined(__x86_64__)
+extern "C" void art_quick_string_compareto(void);
+#endif
+
+TEST_F(StubTest, StringCompareTo) {
+  TEST_DISABLED_FOR_HEAP_REFERENCE_POISONING();
+
+#if defined(__i386__) || defined(__arm__) || defined(__x86_64__)
+  // TODO: Check the "Unresolved" allocation stubs
+
+  Thread* self = Thread::Current();
+  ScopedObjectAccess soa(self);
+  // garbage is created during ClassLinker::Init
+
+  // Create some strings
+  // Use array so we can index into it and use a matrix for expected results
+  constexpr size_t string_count = 7;
+  const char* c[string_count] = { "", "", "a", "aa", "ab", "aac", "aac" };
+
+  SirtRef<mirror::String>* s[string_count];
+
+  for (size_t i = 0; i < string_count; ++i) {
+    s[i] = new SirtRef<mirror::String>(soa.Self(), mirror::String::AllocFromModifiedUtf8(soa.Self(),
+                                                                                         c[i]));
+  }
+
+  // TODO: wide characters
+
+  // Matrix of expectations. First component is first parameter. Note we only check against the
+  // sign, not the value.
+  int32_t expected[string_count][string_count] = {
+      {  0,  0, -1, -1, -1, -1, -1 },  // ""
+      {  0,  0, -1, -1, -1, -1, -1 },  // ""
+      {  1,  1,  0, -1, -1, -1, -1 },  // "a"
+      {  1,  1,  1,  0, -1, -1, -1 },  // "aa"
+      {  1,  1,  1,  1,  0,  1,  1 },  // "ab"
+      {  1,  1,  1,  1, -1,  0,  0 },  // "aac"
+      {  1,  1,  1,  1, -1,  0,  0 }   // "aac"
+  //    ""  ""   a  aa  ab  aac aac
+  };
+
+  // Play with it...
+
+  for (size_t x = 0; x < string_count; ++x) {
+    for (size_t y = 0; y < string_count; ++y) {
+      // Test string_compareto x y
+      size_t result = Invoke3(reinterpret_cast<size_t>(s[x]->get()),
+                              reinterpret_cast<size_t>(s[y]->get()), 0U,
+                              reinterpret_cast<uintptr_t>(&art_quick_string_compareto), self);
+
+      EXPECT_FALSE(self->IsExceptionPending());
+
+      // The result is a 32b signed integer
+      union {
+        size_t r;
+        int32_t i;
+      } conv;
+      conv.r = result;
+      int32_t e = expected[x][y];
+      EXPECT_TRUE(e == 0 ? conv.i == 0 : true) << "x=" << c[x] << " y=" << c[y];
+      EXPECT_TRUE(e < 0 ? conv.i < 0 : true)   << "x=" << c[x] << " y="  << c[y];
+      EXPECT_TRUE(e > 0 ? conv.i > 0 : true)   << "x=" << c[x] << " y=" << c[y];
+    }
+  }
+
+  // Tests done.
+#else
+  LOG(INFO) << "Skipping string_compareto as I don't know how to do that on " << kRuntimeISA;
+  // Force-print to std::cout so it's also outside the logcat.
+  std::cout << "Skipping string_compareto as I don't know how to do that on " << kRuntimeISA <<
+      std::endl;
+#endif
+}
+
 }  // namespace art
diff --git a/runtime/arch/x86_64/quick_entrypoints_x86_64.S b/runtime/arch/x86_64/quick_entrypoints_x86_64.S
index 4fefd20..cac6cfd 100644
--- a/runtime/arch/x86_64/quick_entrypoints_x86_64.S
+++ b/runtime/arch/x86_64/quick_entrypoints_x86_64.S
@@ -1169,5 +1169,47 @@
 UNIMPLEMENTED art_quick_deoptimize
 
 UNIMPLEMENTED art_quick_indexof
-UNIMPLEMENTED art_quick_string_compareto
+
+    /*
+     * String's compareTo.
+     *
+     * On entry:
+     *    rdi:   this string object (known non-null)
+     *    rsi:   comp string object (known non-null)
+     */
+DEFINE_FUNCTION art_quick_string_compareto
+    movl STRING_COUNT_OFFSET(%edi), %r8d
+    movl STRING_COUNT_OFFSET(%esi), %r9d
+    movl STRING_VALUE_OFFSET(%edi), %r10d
+    movl STRING_VALUE_OFFSET(%esi), %r11d
+    movl STRING_OFFSET_OFFSET(%edi), %eax
+    movl STRING_OFFSET_OFFSET(%esi), %ecx
+    /* Build pointers to the start of string data */
+    leal STRING_DATA_OFFSET(%r10d, %eax, 2), %esi
+    leal STRING_DATA_OFFSET(%r11d, %ecx, 2), %edi
+    /* Calculate min length and count diff */
+    movl  %r8d, %ecx
+    movl  %r8d, %eax
+    subl  %r9d, %eax
+    cmovg %r9d, %ecx
+    /*
+     * At this point we have:
+     *   eax: value to return if first part of strings are equal
+     *   ecx: minimum among the lengths of the two strings
+     *   esi: pointer to this string data
+     *   edi: pointer to comp string data
+     */
+    jecxz .Lkeep_length
+    repe cmpsw                    // find nonmatching chars in [%esi] and [%edi], up to length %ecx
+    jne .Lnot_equal
+.Lkeep_length:
+    ret
+    .balign 16
+.Lnot_equal:
+    movzwl  -2(%esi), %eax        // get last compared char from this string
+    movzwl  -2(%edi), %ecx        // get last compared char from comp string
+    subl  %ecx, %eax              // return the difference
+    ret
+END_FUNCTION art_quick_string_compareto
+
 UNIMPLEMENTED art_quick_memcmp16
diff --git a/runtime/entrypoints/quick/quick_trampoline_entrypoints.cc b/runtime/entrypoints/quick/quick_trampoline_entrypoints.cc
index 2b29591..680ffbe 100644
--- a/runtime/entrypoints/quick/quick_trampoline_entrypoints.cc
+++ b/runtime/entrypoints/quick/quick_trampoline_entrypoints.cc
@@ -34,12 +34,6 @@
 
 // Visits the arguments as saved to the stack by a Runtime::kRefAndArgs callee save frame.
 class QuickArgumentVisitor {
-  // Size of each spilled GPR.
-#ifdef __LP64__
-  static constexpr size_t kBytesPerGprSpillLocation = 8;
-#else
-  static constexpr size_t kBytesPerGprSpillLocation = 4;
-#endif
   // Number of bytes for each out register in the caller method's frame.
   static constexpr size_t kBytesStackArgLocation = 4;
 #if defined(__arm__)
@@ -61,13 +55,12 @@
   static constexpr bool kQuickSoftFloatAbi = true;  // This is a soft float ABI.
   static constexpr size_t kNumQuickGprArgs = 3;  // 3 arguments passed in GPRs.
   static constexpr size_t kNumQuickFprArgs = 0;  // 0 arguments passed in FPRs.
-  static constexpr size_t kBytesPerFprSpillLocation = 4;  // FPR spill size is 4 bytes.
   static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_Fpr1Offset = 0;  // Offset of first FPR arg.
   static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_Gpr1Offset = 8;  // Offset of first GPR arg.
   static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_LrOffset = 44;  // Offset of return address.
   static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_FrameSize = 48;  // Frame size.
   static size_t GprIndexToGprOffset(uint32_t gpr_index) {
-    return gpr_index * kBytesPerGprSpillLocation;
+    return gpr_index * GetBytesPerGprSpillLocation(kRuntimeISA);
   }
 #elif defined(__aarch64__)
   // The callee save frame is pointed to by SP.
@@ -93,13 +86,12 @@
   static constexpr bool kQuickSoftFloatAbi = false;  // This is a hard float ABI.
   static constexpr size_t kNumQuickGprArgs = 7;  // 7 arguments passed in GPRs.
   static constexpr size_t kNumQuickFprArgs = 8;  // 8 arguments passed in FPRs.
-  static constexpr size_t kBytesPerFprSpillLocation = 8;  // FPR spill size is 8 bytes.
   static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_Fpr1Offset =16;  // Offset of first FPR arg.
   static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_Gpr1Offset = 144;  // Offset of first GPR arg.
   static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_LrOffset = 296;  // Offset of return address.
   static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_FrameSize = 304;  // Frame size.
   static size_t GprIndexToGprOffset(uint32_t gpr_index) {
-    return gpr_index * kBytesPerGprSpillLocation;
+    return gpr_index * GetBytesPerGprSpillLocation(kRuntimeISA);
   }
 #elif defined(__mips__)
   // The callee save frame is pointed to by SP.
@@ -119,13 +111,12 @@
   static constexpr bool kQuickSoftFloatAbi = true;  // This is a soft float ABI.
   static constexpr size_t kNumQuickGprArgs = 3;  // 3 arguments passed in GPRs.
   static constexpr size_t kNumQuickFprArgs = 0;  // 0 arguments passed in FPRs.
-  static constexpr size_t kBytesPerFprSpillLocation = 4;  // FPR spill size is 4 bytes.
   static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_Fpr1Offset = 0;  // Offset of first FPR arg.
   static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_Gpr1Offset = 4;  // Offset of first GPR arg.
   static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_LrOffset = 60;  // Offset of return address.
   static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_FrameSize = 64;  // Frame size.
   static size_t GprIndexToGprOffset(uint32_t gpr_index) {
-    return gpr_index * kBytesPerGprSpillLocation;
+    return gpr_index * GetBytesPerGprSpillLocation(kRuntimeISA);
   }
 #elif defined(__i386__)
   // The callee save frame is pointed to by SP.
@@ -145,13 +136,12 @@
   static constexpr bool kQuickSoftFloatAbi = true;  // This is a soft float ABI.
   static constexpr size_t kNumQuickGprArgs = 3;  // 3 arguments passed in GPRs.
   static constexpr size_t kNumQuickFprArgs = 0;  // 0 arguments passed in FPRs.
-  static constexpr size_t kBytesPerFprSpillLocation = 8;  // FPR spill size is 8 bytes.
   static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_Fpr1Offset = 0;  // Offset of first FPR arg.
   static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_Gpr1Offset = 4;  // Offset of first GPR arg.
   static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_LrOffset = 28;  // Offset of return address.
   static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_FrameSize = 32;  // Frame size.
   static size_t GprIndexToGprOffset(uint32_t gpr_index) {
-    return gpr_index * kBytesPerGprSpillLocation;
+    return gpr_index * GetBytesPerGprSpillLocation(kRuntimeISA);
   }
 #elif defined(__x86_64__)
   // The callee save frame is pointed to by SP.
@@ -184,18 +174,17 @@
   static constexpr bool kQuickSoftFloatAbi = false;  // This is a hard float ABI.
   static constexpr size_t kNumQuickGprArgs = 5;  // 3 arguments passed in GPRs.
   static constexpr size_t kNumQuickFprArgs = 8;  // 0 arguments passed in FPRs.
-  static constexpr size_t kBytesPerFprSpillLocation = 8;  // FPR spill size is 8 bytes.
   static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_Fpr1Offset = 16;  // Offset of first FPR arg.
   static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_Gpr1Offset = 80;  // Offset of first GPR arg.
   static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_LrOffset = 168;  // Offset of return address.
   static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_FrameSize = 176;  // Frame size.
   static size_t GprIndexToGprOffset(uint32_t gpr_index) {
     switch (gpr_index) {
-      case 0: return (4 * kBytesPerGprSpillLocation);
-      case 1: return (1 * kBytesPerGprSpillLocation);
-      case 2: return (0 * kBytesPerGprSpillLocation);
-      case 3: return (5 * kBytesPerGprSpillLocation);
-      case 4: return (6 * kBytesPerGprSpillLocation);
+      case 0: return (4 * GetBytesPerGprSpillLocation(kRuntimeISA));
+      case 1: return (1 * GetBytesPerGprSpillLocation(kRuntimeISA));
+      case 2: return (0 * GetBytesPerGprSpillLocation(kRuntimeISA));
+      case 3: return (5 * GetBytesPerGprSpillLocation(kRuntimeISA));
+      case 4: return (6 * GetBytesPerGprSpillLocation(kRuntimeISA));
       default:
         LOG(FATAL) << "Unexpected GPR index: " << gpr_index;
         return 0;
@@ -248,7 +237,7 @@
       Primitive::Type type = GetParamPrimitiveType();
       if (UNLIKELY((type == Primitive::kPrimDouble) || (type == Primitive::kPrimFloat))) {
         if ((kNumQuickFprArgs != 0) && (fpr_index_ + 1 < kNumQuickFprArgs + 1)) {
-          return fpr_args_ + (fpr_index_ * kBytesPerFprSpillLocation);
+          return fpr_args_ + (fpr_index_ * GetBytesPerFprSpillLocation(kRuntimeISA));
         }
         return stack_args_ + (stack_index_ * kBytesStackArgLocation);
       }
@@ -260,7 +249,7 @@
   }
 
   bool IsSplitLongOrDouble() const {
-    if ((kBytesPerGprSpillLocation == 4) || (kBytesPerFprSpillLocation == 4)) {
+    if ((GetBytesPerGprSpillLocation(kRuntimeISA) == 4) || (GetBytesPerFprSpillLocation(kRuntimeISA) == 4)) {
       return is_split_long_or_double_;
     } else {
       return false;  // An optimization for when GPR and FPRs are 64bit.
@@ -341,7 +330,7 @@
         case Primitive::kPrimDouble:
         case Primitive::kPrimLong:
           if (kQuickSoftFloatAbi || (cur_type_ == Primitive::kPrimLong)) {
-            is_split_long_or_double_ = (kBytesPerGprSpillLocation == 4) &&
+            is_split_long_or_double_ = (GetBytesPerGprSpillLocation(kRuntimeISA) == 4) &&
                 ((gpr_index_ + 1) == kNumQuickGprArgs);
             Visit();
             if (!kQuickSoftFloatAbi || kNumQuickGprArgs == gpr_index_) {
@@ -354,7 +343,7 @@
             }
             if (gpr_index_ < kNumQuickGprArgs) {
               gpr_index_++;
-              if (kBytesPerGprSpillLocation == 4) {
+              if (GetBytesPerGprSpillLocation(kRuntimeISA) == 4) {
                 if (gpr_index_ < kNumQuickGprArgs) {
                   gpr_index_++;
                 } else if (kQuickSoftFloatAbi) {
@@ -363,12 +352,12 @@
               }
             }
           } else {
-            is_split_long_or_double_ = (kBytesPerFprSpillLocation == 4) &&
+            is_split_long_or_double_ = (GetBytesPerFprSpillLocation(kRuntimeISA) == 4) &&
                 ((fpr_index_ + 1) == kNumQuickFprArgs);
             Visit();
             if ((kNumQuickFprArgs != 0) && (fpr_index_ + 1 < kNumQuickFprArgs + 1)) {
               fpr_index_++;
-              if (kBytesPerFprSpillLocation == 4) {
+              if (GetBytesPerFprSpillLocation(kRuntimeISA) == 4) {
                 if ((kNumQuickFprArgs != 0) && (fpr_index_ + 1 < kNumQuickFprArgs + 1)) {
                   fpr_index_++;
                 }
@@ -393,12 +382,13 @@
                                              uint32_t shorty_len) {
     if (kQuickSoftFloatAbi) {
       CHECK_EQ(kNumQuickFprArgs, 0U);
-      return (kNumQuickGprArgs * kBytesPerGprSpillLocation) + kBytesPerGprSpillLocation /* ArtMethod* */;
+      return (kNumQuickGprArgs * GetBytesPerGprSpillLocation(kRuntimeISA))
+          + GetBytesPerGprSpillLocation(kRuntimeISA) /* ArtMethod* */;
     } else {
       // For now, there is no reg-spill area for the targets with
       // hard float ABI. So, the offset pointing to the first method's
       // parameter ('this' for non-static methods) should be returned.
-      return kBytesPerGprSpillLocation;  // Skip Method*.
+      return GetBytesPerGprSpillLocation(kRuntimeISA);  // Skip Method*.
     }
   }
 
diff --git a/runtime/gc/allocator/rosalloc-inl.h b/runtime/gc/allocator/rosalloc-inl.h
index ac0f67b..c69ca48 100644
--- a/runtime/gc/allocator/rosalloc-inl.h
+++ b/runtime/gc/allocator/rosalloc-inl.h
@@ -23,11 +23,17 @@
 namespace gc {
 namespace allocator {
 
+template<bool kThreadSafe>
 inline ALWAYS_INLINE void* RosAlloc::Alloc(Thread* self, size_t size, size_t* bytes_allocated) {
   if (UNLIKELY(size > kLargeSizeThreshold)) {
     return AllocLargeObject(self, size, bytes_allocated);
   }
-  void* m = AllocFromRun(self, size, bytes_allocated);
+  void* m;
+  if (kThreadSafe) {
+    m = AllocFromRun(self, size, bytes_allocated);
+  } else {
+    m = AllocFromRunThreadUnsafe(self, size, bytes_allocated);
+  }
   // Check if the returned memory is really all zero.
   if (kCheckZeroMemory && m != nullptr) {
     byte* bytes = reinterpret_cast<byte*>(m);
diff --git a/runtime/gc/allocator/rosalloc.cc b/runtime/gc/allocator/rosalloc.cc
index ff59016..f113030 100644
--- a/runtime/gc/allocator/rosalloc.cc
+++ b/runtime/gc/allocator/rosalloc.cc
@@ -67,11 +67,11 @@
              << std::hex << (intptr_t)(base_ + capacity_)
              << ", capacity=" << std::dec << capacity_
              << ", max_capacity=" << std::dec << max_capacity_;
-  memset(current_runs_, 0, sizeof(current_runs_));
   for (size_t i = 0; i < kNumOfSizeBrackets; i++) {
     size_bracket_lock_names[i] =
         StringPrintf("an rosalloc size bracket %d lock", static_cast<int>(i));
     size_bracket_locks_[i] = new Mutex(size_bracket_lock_names[i].c_str(), kRosAllocBracketLock);
+    current_runs_[i] = dedicated_full_run_;
   }
   DCHECK_EQ(footprint_, capacity_);
   size_t num_of_pages = footprint_ / kPageSize;
@@ -548,7 +548,7 @@
     DCHECK(!new_run->IsThreadLocal());
     DCHECK_EQ(new_run->first_search_vec_idx_, 0U);
     DCHECK(!new_run->to_be_bulk_freed_);
-    if (kUsePrefetchDuringAllocRun && idx <= kMaxThreadLocalSizeBracketIdx) {
+    if (kUsePrefetchDuringAllocRun && idx < kNumThreadLocalSizeBrackets) {
       // Take ownership of the cache lines if we are likely to be thread local run.
       if (kPrefetchNewRunDataByZeroing) {
         // Zeroing the data is sometimes faster than prefetching but it increases memory usage
@@ -584,6 +584,60 @@
   return AllocRun(self, idx);
 }
 
+void* RosAlloc::AllocFromCurrentRunUnlocked(Thread* self, size_t idx) {
+  Run* current_run = current_runs_[idx];
+  DCHECK(current_run != nullptr);
+  void* slot_addr = current_run->AllocSlot();
+  if (UNLIKELY(slot_addr == nullptr)) {
+    // The current run got full. Try to refill it.
+    DCHECK(current_run->IsFull());
+    if (kIsDebugBuild && current_run != dedicated_full_run_) {
+      full_runs_[idx].insert(current_run);
+      if (kTraceRosAlloc) {
+        LOG(INFO) << __FUNCTION__ << " : Inserted run 0x" << std::hex << reinterpret_cast<intptr_t>(current_run)
+                  << " into full_runs_[" << std::dec << idx << "]";
+      }
+      DCHECK(non_full_runs_[idx].find(current_run) == non_full_runs_[idx].end());
+      DCHECK(full_runs_[idx].find(current_run) != full_runs_[idx].end());
+    }
+    current_run = RefillRun(self, idx);
+    if (UNLIKELY(current_run == nullptr)) {
+      // Failed to allocate a new run, make sure that it is the dedicated full run.
+      current_runs_[idx] = dedicated_full_run_;
+      return nullptr;
+    }
+    DCHECK(current_run != nullptr);
+    DCHECK(non_full_runs_[idx].find(current_run) == non_full_runs_[idx].end());
+    DCHECK(full_runs_[idx].find(current_run) == full_runs_[idx].end());
+    current_run->SetIsThreadLocal(false);
+    current_runs_[idx] = current_run;
+    DCHECK(!current_run->IsFull());
+    slot_addr = current_run->AllocSlot();
+    // Must succeed now with a new run.
+    DCHECK(slot_addr != nullptr);
+  }
+  return slot_addr;
+}
+
+void* RosAlloc::AllocFromRunThreadUnsafe(Thread* self, size_t size, size_t* bytes_allocated) {
+  DCHECK_LE(size, kLargeSizeThreshold);
+  size_t bracket_size;
+  size_t idx = SizeToIndexAndBracketSize(size, &bracket_size);
+  DCHECK_EQ(idx, SizeToIndex(size));
+  DCHECK_EQ(bracket_size, IndexToBracketSize(idx));
+  DCHECK_EQ(bracket_size, bracketSizes[idx]);
+  DCHECK_LE(size, bracket_size);
+  DCHECK(size > 512 || bracket_size - size < 16);
+  Locks::mutator_lock_->AssertExclusiveHeld(self);
+  void* slot_addr = AllocFromCurrentRunUnlocked(self, idx);
+  if (LIKELY(slot_addr != nullptr)) {
+    DCHECK(bytes_allocated != nullptr);
+    *bytes_allocated = bracket_size;
+    // Caller verifies that it is all 0.
+  }
+  return slot_addr;
+}
+
 void* RosAlloc::AllocFromRun(Thread* self, size_t size, size_t* bytes_allocated) {
   DCHECK_LE(size, kLargeSizeThreshold);
   size_t bracket_size;
@@ -596,7 +650,7 @@
 
   void* slot_addr;
 
-  if (LIKELY(idx <= kMaxThreadLocalSizeBracketIdx)) {
+  if (LIKELY(idx < kNumThreadLocalSizeBrackets)) {
     // Use a thread-local run.
     Run* thread_local_run = reinterpret_cast<Run*>(self->GetRosAllocRun(idx));
     // Allow invalid since this will always fail the allocation.
@@ -631,7 +685,6 @@
         // No slots got freed. Try to refill the thread-local run.
         DCHECK(thread_local_run->IsFull());
         if (thread_local_run != dedicated_full_run_) {
-          self->SetRosAllocRun(idx, dedicated_full_run_);
           thread_local_run->SetIsThreadLocal(false);
           if (kIsDebugBuild) {
             full_runs_[idx].insert(thread_local_run);
@@ -646,8 +699,9 @@
         }
 
         thread_local_run = RefillRun(self, idx);
-        if (UNLIKELY(thread_local_run == NULL)) {
-          return NULL;
+        if (UNLIKELY(thread_local_run == nullptr)) {
+          self->SetRosAllocRun(idx, dedicated_full_run_);
+          return nullptr;
         }
         DCHECK(non_full_runs_[idx].find(thread_local_run) == non_full_runs_[idx].end());
         DCHECK(full_runs_[idx].find(thread_local_run) == full_runs_[idx].end());
@@ -656,12 +710,12 @@
         DCHECK(!thread_local_run->IsFull());
       }
 
-      DCHECK(thread_local_run != NULL);
+      DCHECK(thread_local_run != nullptr);
       DCHECK(!thread_local_run->IsFull());
       DCHECK(thread_local_run->IsThreadLocal());
       slot_addr = thread_local_run->AllocSlot();
       // Must succeed now with a new run.
-      DCHECK(slot_addr != NULL);
+      DCHECK(slot_addr != nullptr);
     }
     if (kTraceRosAlloc) {
       LOG(INFO) << "RosAlloc::AllocFromRun() thread-local : 0x" << std::hex << reinterpret_cast<intptr_t>(slot_addr)
@@ -671,48 +725,7 @@
   } else {
     // Use the (shared) current run.
     MutexLock mu(self, *size_bracket_locks_[idx]);
-    Run* current_run = current_runs_[idx];
-    if (UNLIKELY(current_run == NULL)) {
-      current_run = RefillRun(self, idx);
-      if (UNLIKELY(current_run == NULL)) {
-        return NULL;
-      }
-      DCHECK(non_full_runs_[idx].find(current_run) == non_full_runs_[idx].end());
-      DCHECK(full_runs_[idx].find(current_run) == full_runs_[idx].end());
-      current_run->SetIsThreadLocal(false);
-      current_runs_[idx] = current_run;
-      DCHECK(!current_run->IsFull());
-    }
-    DCHECK(current_run != NULL);
-    slot_addr = current_run->AllocSlot();
-    if (UNLIKELY(slot_addr == NULL)) {
-      // The current run got full. Try to refill it.
-      DCHECK(current_run->IsFull());
-      current_runs_[idx] = NULL;
-      if (kIsDebugBuild) {
-        // Insert it into full_runs and set the current run to NULL.
-        full_runs_[idx].insert(current_run);
-        if (kTraceRosAlloc) {
-          LOG(INFO) << "RosAlloc::AllocFromRun() : Inserted run 0x" << std::hex << reinterpret_cast<intptr_t>(current_run)
-                    << " into full_runs_[" << std::dec << idx << "]";
-        }
-      }
-      DCHECK(non_full_runs_[idx].find(current_run) == non_full_runs_[idx].end());
-      DCHECK(full_runs_[idx].find(current_run) != full_runs_[idx].end());
-      current_run = RefillRun(self, idx);
-      if (UNLIKELY(current_run == NULL)) {
-        return NULL;
-      }
-      DCHECK(current_run != NULL);
-      DCHECK(non_full_runs_[idx].find(current_run) == non_full_runs_[idx].end());
-      DCHECK(full_runs_[idx].find(current_run) == full_runs_[idx].end());
-      current_run->SetIsThreadLocal(false);
-      current_runs_[idx] = current_run;
-      DCHECK(!current_run->IsFull());
-      slot_addr = current_run->AllocSlot();
-      // Must succeed now with a new run.
-      DCHECK(slot_addr != NULL);
-    }
+    slot_addr = AllocFromCurrentRunUnlocked(self, idx);
     if (kTraceRosAlloc) {
       LOG(INFO) << "RosAlloc::AllocFromRun() : 0x" << std::hex << reinterpret_cast<intptr_t>(slot_addr)
                 << "-0x" << (reinterpret_cast<intptr_t>(slot_addr) + bracket_size)
@@ -741,7 +754,7 @@
   }
   if (LIKELY(run->IsThreadLocal())) {
     // It's a thread-local run. Just mark the thread-local free bit map and return.
-    DCHECK_LE(run->size_bracket_idx_, kMaxThreadLocalSizeBracketIdx);
+    DCHECK_LT(run->size_bracket_idx_, kNumThreadLocalSizeBrackets);
     DCHECK(non_full_runs_[idx].find(run) == non_full_runs_[idx].end());
     DCHECK(full_runs_[idx].find(run) == full_runs_[idx].end());
     run->MarkThreadLocalFreeBitMap(ptr);
@@ -766,7 +779,7 @@
       }
     }
     if (run == current_runs_[idx]) {
-      current_runs_[idx] = NULL;
+      current_runs_[idx] = dedicated_full_run_;
     }
     DCHECK(non_full_runs_[idx].find(run) == non_full_runs_[idx].end());
     DCHECK(full_runs_[idx].find(run) == full_runs_[idx].end());
@@ -1233,7 +1246,7 @@
     size_t idx = run->size_bracket_idx_;
     MutexLock mu(self, *size_bracket_locks_[idx]);
     if (run->IsThreadLocal()) {
-      DCHECK_LE(run->size_bracket_idx_, kMaxThreadLocalSizeBracketIdx);
+      DCHECK_LT(run->size_bracket_idx_, kNumThreadLocalSizeBrackets);
       DCHECK(non_full_runs_[idx].find(run) == non_full_runs_[idx].end());
       DCHECK(full_runs_[idx].find(run) == full_runs_[idx].end());
       run->UnionBulkFreeBitMapToThreadLocalFreeBitMap();
@@ -1627,7 +1640,7 @@
   Thread* self = Thread::Current();
   // Avoid race conditions on the bulk free bit maps with BulkFree() (GC).
   WriterMutexLock wmu(self, bulk_free_lock_);
-  for (size_t idx = 0; idx < kNumOfSizeBrackets; idx++) {
+  for (size_t idx = 0; idx < kNumThreadLocalSizeBrackets; idx++) {
     MutexLock mu(self, *size_bracket_locks_[idx]);
     Run* thread_local_run = reinterpret_cast<Run*>(thread->GetRosAllocRun(idx));
     CHECK(thread_local_run != nullptr);
@@ -1643,30 +1656,48 @@
       thread_local_run->MergeBulkFreeBitMapIntoAllocBitMap();
       DCHECK(non_full_runs_[idx].find(thread_local_run) == non_full_runs_[idx].end());
       DCHECK(full_runs_[idx].find(thread_local_run) == full_runs_[idx].end());
-      if (thread_local_run->IsFull()) {
-        if (kIsDebugBuild) {
-          full_runs_[idx].insert(thread_local_run);
-          DCHECK(full_runs_[idx].find(thread_local_run) != full_runs_[idx].end());
-          if (kTraceRosAlloc) {
-            LOG(INFO) << "RosAlloc::RevokeThreadLocalRuns() : Inserted run 0x" << std::hex
-                      << reinterpret_cast<intptr_t>(thread_local_run)
-                      << " into full_runs_[" << std::dec << idx << "]";
-          }
-        }
-      } else if (thread_local_run->IsAllFree()) {
-        MutexLock mu(self, lock_);
-        thread_local_run->ZeroHeader();
-        FreePages(self, thread_local_run, true);
-      } else {
-        non_full_runs_[idx].insert(thread_local_run);
-        DCHECK(non_full_runs_[idx].find(thread_local_run) != non_full_runs_[idx].end());
-        if (kTraceRosAlloc) {
-          LOG(INFO) << "RosAlloc::RevokeThreadLocalRuns() : Inserted run 0x" << std::hex
-                    << reinterpret_cast<intptr_t>(thread_local_run)
-                    << " into non_full_runs_[" << std::dec << idx << "]";
-        }
+      RevokeRun(self, idx, thread_local_run);
+    }
+  }
+}
+
+void RosAlloc::RevokeRun(Thread* self, size_t idx, Run* run) {
+  size_bracket_locks_[idx]->AssertHeld(self);
+  DCHECK(run != dedicated_full_run_);
+  if (run->IsFull()) {
+    if (kIsDebugBuild) {
+      full_runs_[idx].insert(run);
+      DCHECK(full_runs_[idx].find(run) != full_runs_[idx].end());
+      if (kTraceRosAlloc) {
+        LOG(INFO) << __FUNCTION__  << " : Inserted run 0x" << std::hex
+                  << reinterpret_cast<intptr_t>(run)
+                  << " into full_runs_[" << std::dec << idx << "]";
       }
     }
+  } else if (run->IsAllFree()) {
+    run->ZeroHeader();
+    MutexLock mu(self, lock_);
+    FreePages(self, run, true);
+  } else {
+    non_full_runs_[idx].insert(run);
+    DCHECK(non_full_runs_[idx].find(run) != non_full_runs_[idx].end());
+    if (kTraceRosAlloc) {
+      LOG(INFO) << __FUNCTION__ << " : Inserted run 0x" << std::hex
+                << reinterpret_cast<intptr_t>(run)
+                << " into non_full_runs_[" << std::dec << idx << "]";
+    }
+  }
+}
+
+void RosAlloc::RevokeThreadUnsafeCurrentRuns() {
+  // Revoke the current runs which share the same idx as thread local runs.
+  Thread* self = Thread::Current();
+  for (size_t idx = 0; idx < kNumThreadLocalSizeBrackets; ++idx) {
+    MutexLock mu(self, *size_bracket_locks_[idx]);
+    if (current_runs_[idx] != dedicated_full_run_) {
+      RevokeRun(self, idx, current_runs_[idx]);
+      current_runs_[idx] = dedicated_full_run_;
+    }
   }
 }
 
@@ -1679,6 +1710,7 @@
   for (Thread* thread : thread_list) {
     RevokeThreadLocalRuns(thread);
   }
+  RevokeThreadUnsafeCurrentRuns();
 }
 
 void RosAlloc::AssertThreadLocalRunsAreRevoked(Thread* thread) {
@@ -1686,7 +1718,7 @@
     Thread* self = Thread::Current();
     // Avoid race conditions on the bulk free bit maps with BulkFree() (GC).
     WriterMutexLock wmu(self, bulk_free_lock_);
-    for (size_t idx = 0; idx < kNumOfSizeBrackets; idx++) {
+    for (size_t idx = 0; idx < kNumThreadLocalSizeBrackets; idx++) {
       MutexLock mu(self, *size_bracket_locks_[idx]);
       Run* thread_local_run = reinterpret_cast<Run*>(thread->GetRosAllocRun(idx));
       DCHECK(thread_local_run == nullptr || thread_local_run == dedicated_full_run_);
@@ -1696,18 +1728,21 @@
 
 void RosAlloc::AssertAllThreadLocalRunsAreRevoked() {
   if (kIsDebugBuild) {
-    MutexLock mu(Thread::Current(), *Locks::runtime_shutdown_lock_);
-    MutexLock mu2(Thread::Current(), *Locks::thread_list_lock_);
+    Thread* self = Thread::Current();
+    MutexLock mu(self, *Locks::runtime_shutdown_lock_);
+    MutexLock mu2(self, *Locks::thread_list_lock_);
     std::list<Thread*> thread_list = Runtime::Current()->GetThreadList()->GetList();
     for (Thread* t : thread_list) {
       AssertThreadLocalRunsAreRevoked(t);
     }
+    for (size_t idx = 0; idx < kNumThreadLocalSizeBrackets; ++idx) {
+      MutexLock mu(self, *size_bracket_locks_[idx]);
+      CHECK_EQ(current_runs_[idx], dedicated_full_run_);
+    }
   }
 }
 
 void RosAlloc::Initialize() {
-  // Check the consistency of the number of size brackets.
-  DCHECK_EQ(Thread::kRosAllocNumOfSizeBrackets, kNumOfSizeBrackets);
   // bracketSizes.
   for (size_t i = 0; i < kNumOfSizeBrackets; i++) {
     if (i < kNumOfSizeBrackets - 2) {
@@ -1911,15 +1946,34 @@
           break;
         }
         case kPageMapRunPart:
-          LOG(FATAL) << "Unreachable - page map type: " << pm << std::endl << DumpPageMap();
-          break;
+          // Fall-through.
         default:
           LOG(FATAL) << "Unreachable - page map type: " << pm << std::endl << DumpPageMap();
           break;
       }
     }
   }
-
+  std::list<Thread*> threads = Runtime::Current()->GetThreadList()->GetList();
+  for (Thread* thread : threads) {
+    for (size_t i = 0; i < kNumThreadLocalSizeBrackets; ++i) {
+      MutexLock mu(self, *size_bracket_locks_[i]);
+      Run* thread_local_run = reinterpret_cast<Run*>(thread->GetRosAllocRun(i));
+      CHECK(thread_local_run != nullptr);
+      CHECK(thread_local_run->IsThreadLocal());
+      CHECK(thread_local_run == dedicated_full_run_ ||
+            thread_local_run->size_bracket_idx_ == i);
+    }
+  }
+  for (size_t i = 0; i < kNumOfSizeBrackets; i++) {
+    MutexLock mu(self, *size_bracket_locks_[i]);
+    Run* current_run = current_runs_[i];
+    CHECK(current_run != nullptr);
+    if (current_run != dedicated_full_run_) {
+      // The dedicated full run is currently marked as thread local.
+      CHECK(!current_run->IsThreadLocal());
+      CHECK_EQ(current_run->size_bracket_idx_, i);
+    }
+  }
   // Call Verify() here for the lock order.
   for (auto& run : runs) {
     run->Verify(self, this);
@@ -1952,7 +2006,7 @@
     std::list<Thread*> thread_list = Runtime::Current()->GetThreadList()->GetList();
     for (auto it = thread_list.begin(); it != thread_list.end(); ++it) {
       Thread* thread = *it;
-      for (size_t i = 0; i < kNumOfSizeBrackets; i++) {
+      for (size_t i = 0; i < kNumThreadLocalSizeBrackets; i++) {
         MutexLock mu(self, *rosalloc->size_bracket_locks_[i]);
         Run* thread_local_run = reinterpret_cast<Run*>(thread->GetRosAllocRun(i));
         if (thread_local_run == this) {
diff --git a/runtime/gc/allocator/rosalloc.h b/runtime/gc/allocator/rosalloc.h
index f7fa2da..21044f3 100644
--- a/runtime/gc/allocator/rosalloc.h
+++ b/runtime/gc/allocator/rosalloc.h
@@ -405,11 +405,6 @@
   // at a page-granularity.
   static const size_t kLargeSizeThreshold = 2048;
 
-  // We use use thread-local runs for the size Brackets whose indexes
-  // are less than or equal to this index. We use shared (current)
-  // runs for the rest.
-  static const size_t kMaxThreadLocalSizeBracketIdx = 10;
-
   // If true, check that the returned memory is actually zero.
   static constexpr bool kCheckZeroMemory = kIsDebugBuild;
 
@@ -442,6 +437,10 @@
   // The default value for page_release_size_threshold_.
   static constexpr size_t kDefaultPageReleaseSizeThreshold = 4 * MB;
 
+  // We use thread-local runs for the size Brackets whose indexes
+  // are less than this index. We use shared (current) runs for the rest.
+  static const size_t kNumThreadLocalSizeBrackets = 11;
+
  private:
   // The base address of the memory region that's managed by this allocator.
   byte* base_;
@@ -526,6 +525,12 @@
   // Allocate/free a run slot.
   void* AllocFromRun(Thread* self, size_t size, size_t* bytes_allocated)
       LOCKS_EXCLUDED(lock_);
+  // Allocate/free a run slot without acquiring locks.
+  // TODO: EXCLUSIVE_LOCKS_REQUIRED(Locks::mutator_lock_)
+  void* AllocFromRunThreadUnsafe(Thread* self, size_t size, size_t* bytes_allocated)
+      LOCKS_EXCLUDED(lock_);
+  void* AllocFromCurrentRunUnlocked(Thread* self, size_t idx);
+
   // Returns the bracket size.
   size_t FreeFromRun(Thread* self, void* ptr, Run* run)
       LOCKS_EXCLUDED(lock_);
@@ -543,11 +548,20 @@
   // Allocates large objects.
   void* AllocLargeObject(Thread* self, size_t size, size_t* bytes_allocated) LOCKS_EXCLUDED(lock_);
 
+  // Revoke a run by adding it to non_full_runs_ or freeing the pages.
+  void RevokeRun(Thread* self, size_t idx, Run* run);
+
+  // Revoke the current runs which share an index with the thread local runs.
+  void RevokeThreadUnsafeCurrentRuns();
+
  public:
   RosAlloc(void* base, size_t capacity, size_t max_capacity,
            PageReleaseMode page_release_mode,
            size_t page_release_size_threshold = kDefaultPageReleaseSizeThreshold);
   ~RosAlloc();
+  // If kThreadUnsafe is true then the allocator may avoid acquiring some locks as an optimization.
+  // If used, this may cause race conditions if multiple threads are allocating at the same time.
+  template<bool kThreadSafe = true>
   void* Alloc(Thread* self, size_t size, size_t* bytes_allocated)
       LOCKS_EXCLUDED(lock_);
   size_t Free(Thread* self, void* ptr)
diff --git a/runtime/gc/collector/semi_space.cc b/runtime/gc/collector/semi_space.cc
index 0b26019..b53ee10 100644
--- a/runtime/gc/collector/semi_space.cc
+++ b/runtime/gc/collector/semi_space.cc
@@ -180,7 +180,9 @@
     runtime->SetFaultMessage(oss.str());
     CHECK_EQ(self_->SetStateUnsafe(old_state), kRunnable);
   }
-
+  // Revoke the thread local buffers since the GC may allocate into a RosAllocSpace and this helps
+  // to prevent fragmentation.
+  RevokeAllThreadLocalBuffers();
   if (generational_) {
     if (gc_cause_ == kGcCauseExplicit || gc_cause_ == kGcCauseForNativeAlloc ||
         clear_soft_references_) {
@@ -332,11 +334,8 @@
 class SemiSpaceScanObjectVisitor {
  public:
   explicit SemiSpaceScanObjectVisitor(SemiSpace* ss) : semi_space_(ss) {}
-  void operator()(Object* obj) const SHARED_LOCKS_REQUIRED(Locks::mutator_lock_)
-      EXCLUSIVE_LOCKS_REQUIRED(Locks::heap_bitmap_lock_) {
-    // TODO: fix NO_THREAD_SAFETY_ANALYSIS. ScanObject() requires an
-    // exclusive lock on the mutator lock, but
-    // SpaceBitmap::VisitMarkedRange() only requires the shared lock.
+  void operator()(Object* obj) const EXCLUSIVE_LOCKS_REQUIRED(Locks::mutator_lock_,
+                                                              Locks::heap_bitmap_lock_) {
     DCHECK(obj != nullptr);
     semi_space_->ScanObject(obj);
   }
@@ -552,10 +551,11 @@
     // (pseudo-promote) it to the main free list space (as sort
     // of an old generation.)
     space::MallocSpace* promo_dest_space = GetHeap()->GetPrimaryFreeListSpace();
-    forward_address = promo_dest_space->Alloc(self_, object_size, &bytes_allocated, nullptr);
+    forward_address = promo_dest_space->AllocThreadUnsafe(self_, object_size, &bytes_allocated,
+                                                          nullptr);
     if (UNLIKELY(forward_address == nullptr)) {
       // If out of space, fall back to the to-space.
-      forward_address = to_space_->Alloc(self_, object_size, &bytes_allocated, nullptr);
+      forward_address = to_space_->AllocThreadUnsafe(self_, object_size, &bytes_allocated, nullptr);
     } else {
       bytes_promoted_ += bytes_allocated;
       // Dirty the card at the destionation as it may contain
@@ -599,7 +599,7 @@
     DCHECK(forward_address != nullptr);
   } else {
     // If it's allocated after the last GC (younger), copy it to the to-space.
-    forward_address = to_space_->Alloc(self_, object_size, &bytes_allocated, nullptr);
+    forward_address = to_space_->AllocThreadUnsafe(self_, object_size, &bytes_allocated, nullptr);
   }
   ++objects_moved_;
   bytes_moved_ += bytes_allocated;
diff --git a/runtime/gc/collector/semi_space.h b/runtime/gc/collector/semi_space.h
index 9b6df16..3b3e1b1 100644
--- a/runtime/gc/collector/semi_space.h
+++ b/runtime/gc/collector/semi_space.h
@@ -98,12 +98,10 @@
   // Returns the new address of the object.
   template<bool kPoisonReferences>
   void MarkObject(mirror::ObjectReference<kPoisonReferences, mirror::Object>* obj_ptr)
-      EXCLUSIVE_LOCKS_REQUIRED(Locks::heap_bitmap_lock_)
-      SHARED_LOCKS_REQUIRED(Locks::mutator_lock_);
+      EXCLUSIVE_LOCKS_REQUIRED(Locks::heap_bitmap_lock_, Locks::mutator_lock_);
 
   void ScanObject(mirror::Object* obj)
-      EXCLUSIVE_LOCKS_REQUIRED(Locks::heap_bitmap_lock_)
-      SHARED_LOCKS_REQUIRED(Locks::mutator_lock_);
+      EXCLUSIVE_LOCKS_REQUIRED(Locks::heap_bitmap_lock_, Locks::mutator_lock_);
 
   void VerifyNoFromSpaceReferences(mirror::Object* obj)
       SHARED_LOCKS_REQUIRED(Locks::heap_bitmap_lock_, Locks::mutator_lock_);
@@ -150,8 +148,7 @@
       SHARED_LOCKS_REQUIRED(Locks::heap_bitmap_lock_, Locks::mutator_lock_);
 
   virtual mirror::Object* MarkNonForwardedObject(mirror::Object* obj)
-      EXCLUSIVE_LOCKS_REQUIRED(Locks::heap_bitmap_lock_)
-      SHARED_LOCKS_REQUIRED(Locks::mutator_lock_);
+      EXCLUSIVE_LOCKS_REQUIRED(Locks::heap_bitmap_lock_, Locks::mutator_lock_);
 
   // Schedules an unmarked object for reference processing.
   void DelayReferenceReferent(mirror::Class* klass, mirror::Reference* reference)
diff --git a/runtime/gc/heap.cc b/runtime/gc/heap.cc
index 50ef7a6..5d517bb 100644
--- a/runtime/gc/heap.cc
+++ b/runtime/gc/heap.cc
@@ -358,16 +358,16 @@
     can_move_objects = !have_zygote_space_;
   }
   if (kUseRosAlloc) {
-    main_space_ = space::RosAllocSpace::CreateFromMemMap(mem_map, "main rosalloc space",
-                                                         kDefaultStartingSize, initial_size,
-                                                         growth_limit, capacity, low_memory_mode_,
-                                                         can_move_objects);
+    rosalloc_space_ = space::RosAllocSpace::CreateFromMemMap(
+        mem_map, "main rosalloc space", kDefaultStartingSize, initial_size, growth_limit, capacity,
+        low_memory_mode_, can_move_objects);
+    main_space_ = rosalloc_space_;
     CHECK(main_space_ != nullptr) << "Failed to create rosalloc space";
   } else {
-    main_space_ = space::DlMallocSpace::CreateFromMemMap(mem_map, "main dlmalloc space",
-                                                         kDefaultStartingSize, initial_size,
-                                                         growth_limit, capacity,
-                                                         can_move_objects);
+    dlmalloc_space_ = space::DlMallocSpace::CreateFromMemMap(
+        mem_map, "main dlmalloc space", kDefaultStartingSize, initial_size, growth_limit, capacity,
+        can_move_objects);
+    main_space_ = rosalloc_space_;
     CHECK(main_space_ != nullptr) << "Failed to create dlmalloc space";
   }
   main_space_->SetFootprintLimit(main_space_->Capacity());
@@ -580,7 +580,7 @@
   thread_pool_.reset(nullptr);
 }
 
-void Heap::AddSpace(space::Space* space, bool set_as_default) {
+void Heap::AddSpace(space::Space* space) {
   DCHECK(space != nullptr);
   WriterMutexLock mu(Thread::Current(), *Locks::heap_bitmap_lock_);
   if (space->IsContinuousSpace()) {
@@ -595,18 +595,6 @@
       mark_bitmap_->AddContinuousSpaceBitmap(mark_bitmap);
     }
     continuous_spaces_.push_back(continuous_space);
-    if (set_as_default) {
-      if (continuous_space->IsDlMallocSpace()) {
-        dlmalloc_space_ = continuous_space->AsDlMallocSpace();
-      } else if (continuous_space->IsRosAllocSpace()) {
-        // Revoke before if we already have a rosalloc_space_ so that we don't end up with non full
-        // runs from the previous one during the revoke after.
-        if (rosalloc_space_ != nullptr) {
-          rosalloc_space_->RevokeAllThreadLocalBuffers();
-        }
-        rosalloc_space_ = continuous_space->AsRosAllocSpace();
-      }
-    }
     // Ensure that spaces remain sorted in increasing order of start address.
     std::sort(continuous_spaces_.begin(), continuous_spaces_.end(),
               [](const space::ContinuousSpace* a, const space::ContinuousSpace* b) {
@@ -624,7 +612,16 @@
   }
 }
 
-void Heap::RemoveSpace(space::Space* space, bool unset_as_default) {
+void Heap::SetSpaceAsDefault(space::ContinuousSpace* continuous_space) {
+  WriterMutexLock mu(Thread::Current(), *Locks::heap_bitmap_lock_);
+  if (continuous_space->IsDlMallocSpace()) {
+    dlmalloc_space_ = continuous_space->AsDlMallocSpace();
+  } else if (continuous_space->IsRosAllocSpace()) {
+    rosalloc_space_ = continuous_space->AsRosAllocSpace();
+  }
+}
+
+void Heap::RemoveSpace(space::Space* space) {
   DCHECK(space != nullptr);
   WriterMutexLock mu(Thread::Current(), *Locks::heap_bitmap_lock_);
   if (space->IsContinuousSpace()) {
@@ -641,20 +638,6 @@
     auto it = std::find(continuous_spaces_.begin(), continuous_spaces_.end(), continuous_space);
     DCHECK(it != continuous_spaces_.end());
     continuous_spaces_.erase(it);
-    if (unset_as_default) {
-      if (continuous_space == dlmalloc_space_) {
-        dlmalloc_space_ = nullptr;
-      } else if (continuous_space == rosalloc_space_) {
-        rosalloc_space_ = nullptr;
-      }
-      if (continuous_space == main_space_) {
-        main_space_ = nullptr;
-      } else if (continuous_space == bump_pointer_space_) {
-        bump_pointer_space_ = nullptr;
-      } else if (continuous_space == temp_space_) {
-        temp_space_ = nullptr;
-      }
-    }
   } else {
     DCHECK(space->IsDiscontinuousSpace());
     space::DiscontinuousSpace* discontinuous_space = space->AsDiscontinuousSpace();
@@ -1470,7 +1453,7 @@
         // Remove the main space so that we don't try to trim it, this doens't work for debug
         // builds since RosAlloc attempts to read the magic number from a protected page.
         // TODO: Clean this up by getting rid of the remove_as_default parameter.
-        RemoveSpace(main_space_, false);
+        RemoveSpace(main_space_);
       }
       break;
     }
@@ -1479,7 +1462,7 @@
     case kCollectorTypeCMS: {
       if (IsMovingGc(collector_type_)) {
         // Compact to the main space from the bump pointer space, don't need to swap semispaces.
-        AddSpace(main_space_, false);
+        AddSpace(main_space_);
         main_space_->GetMemMap()->Protect(PROT_READ | PROT_WRITE);
         Compact(main_space_, bump_pointer_space_);
       }
@@ -1694,14 +1677,8 @@
       reset_main_space = true;
     }
     zygote_collector.SetToSpace(&target_space);
-
-    Runtime::Current()->GetThreadList()->SuspendAll();
+    zygote_collector.SetSwapSemiSpaces(false);
     zygote_collector.Run(kGcCauseCollectorTransition, false);
-    if (IsMovingGc(collector_type_)) {
-      SwapSemiSpaces();
-    }
-    Runtime::Current()->GetThreadList()->ResumeAll();
-
     if (reset_main_space) {
       main_space_->GetMemMap()->Protect(PROT_READ | PROT_WRITE);
       madvise(main_space_->Begin(), main_space_->Capacity(), MADV_DONTNEED);
@@ -1747,7 +1724,7 @@
                                                                         &non_moving_space_);
   delete old_alloc_space;
   CHECK(zygote_space != nullptr) << "Failed creating zygote space";
-  AddSpace(zygote_space, false);
+  AddSpace(zygote_space);
   non_moving_space_->SetFootprintLimit(non_moving_space_->Capacity());
   AddSpace(non_moving_space_);
   have_zygote_space_ = true;
@@ -2397,8 +2374,7 @@
 }
 
 void Heap::PreGcVerification(collector::GarbageCollector* gc) {
-  if (verify_pre_gc_heap_ || verify_missing_card_marks_ || verify_mod_union_table_ ||
-      verify_pre_gc_rosalloc_) {
+  if (verify_pre_gc_heap_ || verify_missing_card_marks_ || verify_mod_union_table_) {
     collector::GarbageCollector::ScopedPause pause(gc);
     PreGcVerificationPaused(gc);
   }
diff --git a/runtime/gc/heap.h b/runtime/gc/heap.h
index 5533f3d..d3b5cdc 100644
--- a/runtime/gc/heap.h
+++ b/runtime/gc/heap.h
@@ -283,11 +283,12 @@
   void RegisterGCAllocation(size_t bytes);
   void RegisterGCDeAllocation(size_t bytes);
 
-  // Public due to usage by tests.
-  void AddSpace(space::Space* space, bool set_as_default = true)
+  // Set the heap's private space pointers to be the same as the space based on it's type. Public
+  // due to usage by tests.
+  void SetSpaceAsDefault(space::ContinuousSpace* continuous_space)
       LOCKS_EXCLUDED(Locks::heap_bitmap_lock_);
-  void RemoveSpace(space::Space* space, bool unset_as_default = true)
-      LOCKS_EXCLUDED(Locks::heap_bitmap_lock_);
+  void AddSpace(space::Space* space) LOCKS_EXCLUDED(Locks::heap_bitmap_lock_);
+  void RemoveSpace(space::Space* space) LOCKS_EXCLUDED(Locks::heap_bitmap_lock_);
 
   // Set target ideal heap utilization ratio, implements
   // dalvik.system.VMRuntime.setTargetHeapUtilization.
diff --git a/runtime/gc/space/bump_pointer_space-inl.h b/runtime/gc/space/bump_pointer_space-inl.h
index 70ab64b..497a61f 100644
--- a/runtime/gc/space/bump_pointer_space-inl.h
+++ b/runtime/gc/space/bump_pointer_space-inl.h
@@ -36,6 +36,26 @@
   return ret;
 }
 
+inline mirror::Object* BumpPointerSpace::AllocThreadUnsafe(Thread* self, size_t num_bytes,
+                                                           size_t* bytes_allocated,
+                                                           size_t* usable_size) {
+  Locks::mutator_lock_->AssertExclusiveHeld(self);
+  num_bytes = RoundUp(num_bytes, kAlignment);
+  if (end_ + num_bytes > growth_end_) {
+    return nullptr;
+  }
+  mirror::Object* obj = reinterpret_cast<mirror::Object*>(end_);
+  end_ += num_bytes;
+  *bytes_allocated = num_bytes;
+  // Use the CAS free versions as an optimization.
+  objects_allocated_ = objects_allocated_ + 1;
+  bytes_allocated_ = bytes_allocated_ + num_bytes;
+  if (UNLIKELY(usable_size != nullptr)) {
+    *usable_size = num_bytes;
+  }
+  return obj;
+}
+
 inline mirror::Object* BumpPointerSpace::AllocNonvirtualWithoutAccounting(size_t num_bytes) {
   DCHECK(IsAligned<kAlignment>(num_bytes));
   byte* old_end;
diff --git a/runtime/gc/space/bump_pointer_space.h b/runtime/gc/space/bump_pointer_space.h
index e52a9a3..9e61f30 100644
--- a/runtime/gc/space/bump_pointer_space.h
+++ b/runtime/gc/space/bump_pointer_space.h
@@ -48,6 +48,11 @@
   // Allocate num_bytes, returns nullptr if the space is full.
   mirror::Object* Alloc(Thread* self, size_t num_bytes, size_t* bytes_allocated,
                         size_t* usable_size) OVERRIDE;
+  // Thread-unsafe allocation for when mutators are suspended, used by the semispace collector.
+  mirror::Object* AllocThreadUnsafe(Thread* self, size_t num_bytes, size_t* bytes_allocated,
+                                    size_t* usable_size)
+      OVERRIDE EXCLUSIVE_LOCKS_REQUIRED(Locks::mutator_lock_);
+
   mirror::Object* AllocNonvirtual(size_t num_bytes);
   mirror::Object* AllocNonvirtualWithoutAccounting(size_t num_bytes);
 
diff --git a/runtime/gc/space/rosalloc_space-inl.h b/runtime/gc/space/rosalloc_space-inl.h
index d270885..fbfef45 100644
--- a/runtime/gc/space/rosalloc_space-inl.h
+++ b/runtime/gc/space/rosalloc_space-inl.h
@@ -46,11 +46,15 @@
   return size_by_size;
 }
 
+template<bool kThreadSafe>
 inline mirror::Object* RosAllocSpace::AllocCommon(Thread* self, size_t num_bytes,
                                                   size_t* bytes_allocated, size_t* usable_size) {
   size_t rosalloc_size = 0;
+  if (!kThreadSafe) {
+    Locks::mutator_lock_->AssertExclusiveHeld(self);
+  }
   mirror::Object* result = reinterpret_cast<mirror::Object*>(
-      rosalloc_->Alloc(self, num_bytes, &rosalloc_size));
+      rosalloc_->Alloc<kThreadSafe>(self, num_bytes, &rosalloc_size));
   if (LIKELY(result != NULL)) {
     if (kDebugSpaces) {
       CHECK(Contains(result)) << "Allocation (" << reinterpret_cast<void*>(result)
diff --git a/runtime/gc/space/rosalloc_space.cc b/runtime/gc/space/rosalloc_space.cc
index f5c0e94..a1511e7 100644
--- a/runtime/gc/space/rosalloc_space.cc
+++ b/runtime/gc/space/rosalloc_space.cc
@@ -159,7 +159,7 @@
   }
   // Note RosAlloc zeroes memory internally.
   // Return the new allocation or NULL.
-  CHECK(!kDebugSpaces || result == NULL || Contains(result));
+  CHECK(!kDebugSpaces || result == nullptr || Contains(result));
   return result;
 }
 
diff --git a/runtime/gc/space/rosalloc_space.h b/runtime/gc/space/rosalloc_space.h
index a156738..2934af8 100644
--- a/runtime/gc/space/rosalloc_space.h
+++ b/runtime/gc/space/rosalloc_space.h
@@ -52,6 +52,11 @@
                         size_t* usable_size) OVERRIDE {
     return AllocNonvirtual(self, num_bytes, bytes_allocated, usable_size);
   }
+  mirror::Object* AllocThreadUnsafe(Thread* self, size_t num_bytes, size_t* bytes_allocated,
+                                    size_t* usable_size)
+      OVERRIDE EXCLUSIVE_LOCKS_REQUIRED(Locks::mutator_lock_) {
+    return AllocNonvirtualThreadUnsafe(self, num_bytes, bytes_allocated, usable_size);
+  }
   size_t AllocationSize(mirror::Object* obj, size_t* usable_size) OVERRIDE {
     return AllocationSizeNonvirtual(obj, usable_size);
   }
@@ -65,6 +70,11 @@
     // RosAlloc zeroes memory internally.
     return AllocCommon(self, num_bytes, bytes_allocated, usable_size);
   }
+  mirror::Object* AllocNonvirtualThreadUnsafe(Thread* self, size_t num_bytes,
+                                              size_t* bytes_allocated, size_t* usable_size) {
+    // RosAlloc zeroes memory internally. Pass in false for thread unsafe.
+    return AllocCommon<false>(self, num_bytes, bytes_allocated, usable_size);
+  }
 
   // TODO: NO_THREAD_SAFETY_ANALYSIS because SizeOf() requires that mutator_lock is held.
   size_t AllocationSizeNonvirtual(mirror::Object* obj, size_t* usable_size)
@@ -116,6 +126,7 @@
                 size_t starting_size, size_t initial_size, bool low_memory_mode);
 
  private:
+  template<bool kThreadSafe = true>
   mirror::Object* AllocCommon(Thread* self, size_t num_bytes, size_t* bytes_allocated,
                               size_t* usable_size);
 
diff --git a/runtime/gc/space/space.h b/runtime/gc/space/space.h
index 0a87a16..dcf5357 100644
--- a/runtime/gc/space/space.h
+++ b/runtime/gc/space/space.h
@@ -203,9 +203,17 @@
   // Allocate num_bytes without allowing growth. If the allocation
   // succeeds, the output parameter bytes_allocated will be set to the
   // actually allocated bytes which is >= num_bytes.
+  // Alloc can be called from multiple threads at the same time and must be thread-safe.
   virtual mirror::Object* Alloc(Thread* self, size_t num_bytes, size_t* bytes_allocated,
                                 size_t* usable_size) = 0;
 
+  // Thread-unsafe allocation for when mutators are suspended, used by the semispace collector.
+  virtual mirror::Object* AllocThreadUnsafe(Thread* self, size_t num_bytes, size_t* bytes_allocated,
+                                            size_t* usable_size)
+      EXCLUSIVE_LOCKS_REQUIRED(Locks::mutator_lock_) {
+    return Alloc(self, num_bytes, bytes_allocated, usable_size);
+  }
+
   // Return the storage space required by obj.
   virtual size_t AllocationSize(mirror::Object* obj, size_t* usable_size) = 0;
 
diff --git a/runtime/gc/space/space_test.h b/runtime/gc/space/space_test.h
index 28200df..3335e72 100644
--- a/runtime/gc/space/space_test.h
+++ b/runtime/gc/space/space_test.h
@@ -38,9 +38,13 @@
   SpaceTest() : byte_array_class_(nullptr) {
   }
 
-  void AddSpace(ContinuousSpace* space) {
-    // By passing true, AddSpace() does the revoke.
-    Runtime::Current()->GetHeap()->AddSpace(space, true);
+  void AddSpace(ContinuousSpace* space, bool revoke = true) {
+    Heap* heap = Runtime::Current()->GetHeap();
+    if (revoke) {
+      heap->RevokeAllThreadLocalBuffers();
+    }
+    heap->AddSpace(space);
+    heap->SetSpaceAsDefault(space);
   }
 
   mirror::Class* GetByteArrayClass(Thread* self) SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
@@ -227,15 +231,16 @@
   gc::Heap* heap = Runtime::Current()->GetHeap();
   space::Space* old_space = space;
   heap->RemoveSpace(old_space);
+  heap->RevokeAllThreadLocalBuffers();
   space::ZygoteSpace* zygote_space = space->CreateZygoteSpace("alloc space",
                                                               heap->IsLowMemoryMode(),
                                                               &space);
   delete old_space;
   // Add the zygote space.
-  AddSpace(zygote_space);
+  AddSpace(zygote_space, false);
 
   // Make space findable to the heap, will also delete space when runtime is cleaned up
-  AddSpace(space);
+  AddSpace(space, false);
 
   // Succeeds, fits without adjusting the footprint limit.
   ptr1.reset(Alloc(space, self, 1 * MB, &ptr1_bytes_allocated, &ptr1_usable_size));
diff --git a/runtime/instruction_set.cc b/runtime/instruction_set.cc
index cbcd2e0..c1931a9 100644
--- a/runtime/instruction_set.cc
+++ b/runtime/instruction_set.cc
@@ -86,6 +86,52 @@
   }
 }
 
+size_t GetBytesPerGprSpillLocation(InstructionSet isa) {
+  switch (isa) {
+    case kArm:
+      // Fall-through.
+    case kThumb2:
+      return 4;
+    case kArm64:
+      return 8;
+    case kX86:
+      return 4;
+    case kX86_64:
+      return 8;
+    case kMips:
+      return 4;
+    case kNone:
+      LOG(FATAL) << "ISA kNone does not have spills.";
+      return 0;
+    default:
+      LOG(FATAL) << "Unknown ISA " << isa;
+      return 0;
+  }
+}
+
+size_t GetBytesPerFprSpillLocation(InstructionSet isa) {
+  switch (isa) {
+    case kArm:
+      // Fall-through.
+    case kThumb2:
+      return 4;
+    case kArm64:
+      return 8;
+    case kX86:
+      return 8;
+    case kX86_64:
+      return 8;
+    case kMips:
+      return 4;
+    case kNone:
+      LOG(FATAL) << "ISA kNone does not have spills.";
+      return 0;
+    default:
+      LOG(FATAL) << "Unknown ISA " << isa;
+      return 0;
+  }
+}
+
 size_t GetInstructionSetAlignment(InstructionSet isa) {
   switch (isa) {
     case kArm:
diff --git a/runtime/instruction_set.h b/runtime/instruction_set.h
index 4bc35a7..bfbbbd6 100644
--- a/runtime/instruction_set.h
+++ b/runtime/instruction_set.h
@@ -41,6 +41,8 @@
 size_t GetInstructionSetPointerSize(InstructionSet isa);
 size_t GetInstructionSetAlignment(InstructionSet isa);
 bool Is64BitInstructionSet(InstructionSet isa);
+size_t GetBytesPerGprSpillLocation(InstructionSet isa);
+size_t GetBytesPerFprSpillLocation(InstructionSet isa);
 
 #if defined(__arm__)
 static constexpr InstructionSet kRuntimeISA = kArm;
diff --git a/runtime/quick_exception_handler.cc b/runtime/quick_exception_handler.cc
index a91fdf1..aee0d64 100644
--- a/runtime/quick_exception_handler.cc
+++ b/runtime/quick_exception_handler.cc
@@ -19,6 +19,7 @@
 #include "catch_block_stack_visitor.h"
 #include "deoptimize_stack_visitor.h"
 #include "entrypoints/entrypoint_utils.h"
+#include "mirror/art_method-inl.h"
 #include "sirt_ref-inl.h"
 
 namespace art {
diff --git a/runtime/stack.cc b/runtime/stack.cc
index ab3bd85..9c709ae 100644
--- a/runtime/stack.cc
+++ b/runtime/stack.cc
@@ -187,7 +187,7 @@
       uint32_t core_spills = m->GetCoreSpillMask();
       uint32_t fp_spills = m->GetFpSpillMask();
       size_t frame_size = m->GetFrameSizeInBytes();
-      int offset = GetVRegOffset(code_item, core_spills, fp_spills, frame_size, vreg);
+      int offset = GetVRegOffset(code_item, core_spills, fp_spills, frame_size, vreg, kRuntimeISA);
       byte* vreg_addr = reinterpret_cast<byte*>(GetCurrentQuickFrame()) + offset;
       *reinterpret_cast<uint32_t*>(vreg_addr) = new_value;
     }
diff --git a/runtime/stack.h b/runtime/stack.h
index ab903d6..afc4f25 100644
--- a/runtime/stack.h
+++ b/runtime/stack.h
@@ -19,9 +19,10 @@
 
 #include "dex_file.h"
 #include "instrumentation.h"
+#include "arch/context.h"
 #include "base/casts.h"
 #include "base/macros.h"
-#include "arch/context.h"
+#include "instruction_set.h"
 #include "mirror/object.h"
 #include "mirror/object_reference.h"
 #include "verify_object.h"
@@ -577,7 +578,7 @@
   uint32_t* GetVRegAddr(mirror::ArtMethod** cur_quick_frame, const DexFile::CodeItem* code_item,
                         uint32_t core_spills, uint32_t fp_spills, size_t frame_size,
                         uint16_t vreg) const {
-    int offset = GetVRegOffset(code_item, core_spills, fp_spills, frame_size, vreg);
+    int offset = GetVRegOffset(code_item, core_spills, fp_spills, frame_size, vreg, kRuntimeISA);
     DCHECK_EQ(cur_quick_frame, GetCurrentQuickFrame());
     byte* vreg_addr = reinterpret_cast<byte*>(cur_quick_frame) + offset;
     return reinterpret_cast<uint32_t*>(vreg_addr);
@@ -634,14 +635,15 @@
    */
   static int GetVRegOffset(const DexFile::CodeItem* code_item,
                            uint32_t core_spills, uint32_t fp_spills,
-                           size_t frame_size, int reg) {
+                           size_t frame_size, int reg, InstructionSet isa) {
     DCHECK_EQ(frame_size & (kStackAlignment - 1), 0U);
     DCHECK_NE(reg, static_cast<int>(kVRegInvalid));
-
-    int num_spills = __builtin_popcount(core_spills) + __builtin_popcount(fp_spills) + 1;  // Filler.
+    int spill_size = __builtin_popcount(core_spills) * GetBytesPerGprSpillLocation(isa)
+        + __builtin_popcount(fp_spills) * GetBytesPerFprSpillLocation(isa)
+        + sizeof(uint32_t);  // Filler.
     int num_ins = code_item->ins_size_;
     int num_regs = code_item->registers_size_ - num_ins;
-    int locals_start = frame_size - ((num_spills + num_regs) * sizeof(uint32_t));
+    int locals_start = frame_size - spill_size - num_regs * sizeof(uint32_t);
     if (reg == static_cast<int>(kVRegMethodPtrBaseReg)) {
       // The current method pointer corresponds to special location on stack.
       return 0;
@@ -654,19 +656,20 @@
        * temp is at offset -4 bytes from locals, the second is at -8 bytes from locals,
        * and so on.
        */
-      int relative_offset = (reg + std::abs(static_cast<int>(kVRegNonSpecialTempBaseReg)) - 1) * sizeof(uint32_t);
+      int relative_offset =
+          (reg + std::abs(static_cast<int>(kVRegNonSpecialTempBaseReg)) - 1) * sizeof(uint32_t);
       return locals_start + relative_offset;
     }  else if (reg < num_regs) {
       return locals_start + (reg * sizeof(uint32_t));
     } else {
       // Handle ins.
-      return frame_size + ((reg - num_regs) * sizeof(uint32_t)) + sizeof(StackReference<mirror::ArtMethod>);
+      return frame_size + ((reg - num_regs) * sizeof(uint32_t)) + GetBytesPerGprSpillLocation(isa);
     }
   }
 
-  static int GetOutVROffset(uint16_t out_num) {
+  static int GetOutVROffset(uint16_t out_num, InstructionSet isa) {
     // According to stack model, the first out is above the Method ptr.
-    return sizeof(StackReference<mirror::ArtMethod>) + (out_num * sizeof(uint32_t));
+    return GetBytesPerGprSpillLocation(isa) + (out_num * sizeof(uint32_t));
   }
 
   uintptr_t GetCurrentQuickFramePc() const {
diff --git a/runtime/thread.cc b/runtime/thread.cc
index 7470670..e67a64f 100644
--- a/runtime/thread.cc
+++ b/runtime/thread.cc
@@ -1018,7 +1018,8 @@
   tls32_.state_and_flags.as_struct.flags = 0;
   tls32_.state_and_flags.as_struct.state = kNative;
   memset(&tlsPtr_.held_mutexes[0], 0, sizeof(tlsPtr_.held_mutexes));
-  std::fill(tlsPtr_.rosalloc_runs, tlsPtr_.rosalloc_runs + kRosAllocNumOfSizeBrackets,
+  std::fill(tlsPtr_.rosalloc_runs,
+            tlsPtr_.rosalloc_runs + gc::allocator::RosAlloc::kNumThreadLocalSizeBrackets,
             gc::allocator::RosAlloc::GetDedicatedFullRun());
   for (uint32_t i = 0; i < kMaxCheckpoints; ++i) {
     tlsPtr_.checkpoint_functions[i] = nullptr;
diff --git a/runtime/thread.h b/runtime/thread.h
index f869285..8c17082 100644
--- a/runtime/thread.h
+++ b/runtime/thread.h
@@ -29,6 +29,7 @@
 #include "entrypoints/jni/jni_entrypoints.h"
 #include "entrypoints/portable/portable_entrypoints.h"
 #include "entrypoints/quick/quick_entrypoints.h"
+#include "gc/allocator/rosalloc.h"
 #include "globals.h"
 #include "jvalue.h"
 #include "object_callbacks.h"
@@ -783,9 +784,6 @@
     return tlsPtr_.thread_local_objects;
   }
 
-  // ROS alloc TLS.
-  static constexpr size_t kRosAllocNumOfSizeBrackets = 34;
-
   void* GetRosAllocRun(size_t index) const {
     return tlsPtr_.rosalloc_runs[index];
   }
@@ -1060,12 +1058,8 @@
     byte* thread_local_end;
     size_t thread_local_objects;
 
-    // Thread-local rosalloc runs. There are 34 size brackets in rosalloc
-    // runs (RosAlloc::kNumOfSizeBrackets). We can't refer to the
-    // RosAlloc class due to a header file circular dependency issue.
-    // To compensate, we check that the two values match at RosAlloc
-    // initialization time.
-    void* rosalloc_runs[kRosAllocNumOfSizeBrackets];
+    // There are RosAlloc::kNumThreadLocalSizeBrackets thread-local size brackets per thread.
+    void* rosalloc_runs[gc::allocator::RosAlloc::kNumThreadLocalSizeBrackets];
 
     // Thread-local allocation stack data/routines.
     mirror::Object** thread_local_alloc_stack_top;
diff --git a/test/JniTest/JniTest.java b/test/JniTest/JniTest.java
index d53cf5e..3c4ed35 100644
--- a/test/JniTest/JniTest.java
+++ b/test/JniTest/JniTest.java
@@ -24,6 +24,10 @@
         testCallStaticVoidMethodOnSubClass();
         testGetMirandaMethod();
         testZeroLengthByteBuffers();
+        testByteMethod();
+        testShortMethod();
+        testBooleanMethod();
+        testCharMethod();
     }
 
     private static native void testFindClassOnAttachedNativeThread();
@@ -79,4 +83,67 @@
     private static interface testGetMirandaMethod_MirandaInterface {
         public boolean inInterface();
     }
+
+    // Test sign-extension for values < 32b
+
+    native static byte byteMethod(byte b1, byte b2, byte b3, byte b4, byte b5, byte b6, byte b7,
+        byte b8, byte b9, byte b10);
+
+    private static void testByteMethod() {
+      byte returns[] = { 0, 1, 2, 127, -1, -2, -128 };
+      for (int i = 0; i < returns.length; i++) {
+        byte result = byteMethod((byte)i, (byte)2, (byte)(-3), (byte)4, (byte)(-5), (byte)6,
+            (byte)(-7), (byte)8, (byte)(-9), (byte)10);
+        if (returns[i] != result) {
+          System.out.println("Run " + i + " with " + returns[i] + " vs " + result);
+          throw new AssertionError();
+        }
+      }
+    }
+
+    native static short shortMethod(short s1, short s2, short s3, short s4, short s5, short s6, short s7,
+        short s8, short s9, short s10);
+
+    private static void testShortMethod() {
+      short returns[] = { 0, 1, 2, 127, 32767, -1, -2, -128, -32768 };
+      for (int i = 0; i < returns.length; i++) {
+        short result = shortMethod((short)i, (short)2, (short)(-3), (short)4, (short)(-5), (short)6,
+            (short)(-7), (short)8, (short)(-9), (short)10);
+        if (returns[i] != result) {
+          System.out.println("Run " + i + " with " + returns[i] + " vs " + result);
+          throw new AssertionError();
+        }
+      }
+    }
+
+    // Test zero-extension for values < 32b
+
+    native static boolean booleanMethod(boolean b1, boolean b2, boolean b3, boolean b4, boolean b5, boolean b6, boolean b7,
+        boolean b8, boolean b9, boolean b10);
+
+    private static void testBooleanMethod() {
+      if (booleanMethod(false, true, false, true, false, true, false, true, false, true)) {
+        throw new AssertionError();
+      }
+
+      if (!booleanMethod(true, true, false, true, false, true, false, true, false, true)) {
+        throw new AssertionError();
+      }
+    }
+
+    native static char charMethod(char c1, char c2, char c3, char c4, char c5, char c6, char c7,
+        char c8, char c9, char c10);
+
+    private static void testCharMethod() {
+      char returns[] = { (char)0, (char)1, (char)2, (char)127, (char)255, (char)256, (char)15000,
+          (char)34000 };
+      for (int i = 0; i < returns.length; i++) {
+        char result = charMethod((char)i, 'a', 'b', 'c', '0', '1', '2', (char)1234, (char)2345,
+            (char)3456);
+        if (returns[i] != result) {
+          System.out.println("Run " + i + " with " + (int)returns[i] + " vs " + (int)result);
+          throw new AssertionError();
+        }
+      }
+    }
 }
diff --git a/test/JniTest/jni_test.cc b/test/JniTest/jni_test.cc
index 33af94b..024ba53 100644
--- a/test/JniTest/jni_test.cc
+++ b/test/JniTest/jni_test.cc
@@ -137,3 +137,92 @@
   assert(env->GetDirectBufferAddress(byte_buffer) == &buffer[0]);
   assert(env->GetDirectBufferCapacity(byte_buffer) == 0);
 }
+
+constexpr size_t kByteReturnSize = 7;
+jbyte byte_returns[kByteReturnSize] = { 0, 1, 2, 127, -1, -2, -128 };
+
+extern "C" jbyte JNICALL Java_JniTest_byteMethod(JNIEnv* env, jclass klass, jbyte b1, jbyte b2,
+                                                    jbyte b3, jbyte b4, jbyte b5, jbyte b6,
+                                                    jbyte b7, jbyte b8, jbyte b9, jbyte b10) {
+  // We use b1 to drive the output.
+  assert(b2 == 2);
+  assert(b3 == -3);
+  assert(b4 == 4);
+  assert(b5 == -5);
+  assert(b6 == 6);
+  assert(b7 == -7);
+  assert(b8 == 8);
+  assert(b9 == -9);
+  assert(b10 == 10);
+
+  assert(0 <= b1);
+  assert(b1 < static_cast<jbyte>(kByteReturnSize));
+
+  return byte_returns[b1];
+}
+
+constexpr size_t kShortReturnSize = 9;
+jshort short_returns[kShortReturnSize] = { 0, 1, 2, 127, 32767, -1, -2, -128,
+    static_cast<jshort>(0x8000) };
+// The weird static_cast is because short int is only guaranteed down to -32767, not Java's -32768.
+
+extern "C" jshort JNICALL Java_JniTest_shortMethod(JNIEnv* env, jclass klass, jshort s1, jshort s2,
+                                                    jshort s3, jshort s4, jshort s5, jshort s6,
+                                                    jshort s7, jshort s8, jshort s9, jshort s10) {
+  // We use s1 to drive the output.
+  assert(s2 == 2);
+  assert(s3 == -3);
+  assert(s4 == 4);
+  assert(s5 == -5);
+  assert(s6 == 6);
+  assert(s7 == -7);
+  assert(s8 == 8);
+  assert(s9 == -9);
+  assert(s10 == 10);
+
+  assert(0 <= s1);
+  assert(s1 < static_cast<jshort>(kShortReturnSize));
+
+  return short_returns[s1];
+}
+
+extern "C" jboolean JNICALL Java_JniTest_booleanMethod(JNIEnv* env, jclass klass, jboolean b1,
+                                                       jboolean b2, jboolean b3, jboolean b4,
+                                                       jboolean b5, jboolean b6, jboolean b7,
+                                                       jboolean b8, jboolean b9, jboolean b10) {
+  // We use b1 to drive the output.
+  assert(b2 == JNI_TRUE);
+  assert(b3 == JNI_FALSE);
+  assert(b4 == JNI_TRUE);
+  assert(b5 == JNI_FALSE);
+  assert(b6 == JNI_TRUE);
+  assert(b7 == JNI_FALSE);
+  assert(b8 == JNI_TRUE);
+  assert(b9 == JNI_FALSE);
+  assert(b10 == JNI_TRUE);
+
+  assert(b1 == JNI_TRUE || b1 == JNI_FALSE);
+  return b1;
+}
+
+constexpr size_t kCharReturnSize = 8;
+jchar char_returns[kCharReturnSize] = { 0, 1, 2, 127, 255, 256, 15000, 34000 };
+
+extern "C" jchar JNICALL Java_JniTest_charMethod(JNIEnv* env, jclass klacc, jchar c1, jchar c2,
+                                                    jchar c3, jchar c4, jchar c5, jchar c6,
+                                                    jchar c7, jchar c8, jchar c9, jchar c10) {
+  // We use c1 to drive the output.
+  assert(c2 == 'a');
+  assert(c3 == 'b');
+  assert(c4 == 'c');
+  assert(c5 == '0');
+  assert(c6 == '1');
+  assert(c7 == '2');
+  assert(c8 == 1234);
+  assert(c9 == 2345);
+  assert(c10 == 3456);
+
+  assert(c1 < static_cast<jchar>(kCharReturnSize));
+
+  return char_returns[c1];
+}
diff --git a/test/etc/push-and-run-test-jar b/test/etc/push-and-run-test-jar
index 93d7e79..e0d2f1d 100755
--- a/test/etc/push-and-run-test-jar
+++ b/test/etc/push-and-run-test-jar
@@ -140,7 +140,7 @@
 fi
 
 if [ "$GDB" = "y" ]; then
-    gdb="/data/gdbserver$TARGET_SUFFIX :5039"
+    gdb="gdbserver$TARGET_SUFFIX :5039"
     gdbargs="$exe"
 fi
 
@@ -150,7 +150,7 @@
 
 JNI_OPTS="-Xjnigreflimit:512 -Xcheck:jni"
 
-cmdline="cd $DEX_LOCATION && mkdir dalvik-cache && export ANDROID_DATA=$DEX_LOCATION && export DEX_LOCATION=$DEX_LOCATION && \
+cmdline="cd $DEX_LOCATION && mkdir -p dalvik-cache/{arm,arm64,mips,x86,x86_64} && export ANDROID_DATA=$DEX_LOCATION && export DEX_LOCATION=$DEX_LOCATION && \
     $INVOKE_WITH $gdb /system/bin/dalvikvm$TARGET_SUFFIX $FLAGS $gdbargs -XXlib:$LIB $ZYGOTE $JNI_OPTS $INT_OPTS $DEBUGGER_OPTS $BOOT_OPT -cp $DEX_LOCATION/$TEST_NAME.jar Main"
 if [ "$DEV_MODE" = "y" ]; then
   echo $cmdline "$@"
diff --git a/tools/art b/tools/art
index c9c0d4f..e3f409c 100755
--- a/tools/art
+++ b/tools/art
@@ -48,7 +48,7 @@
 ANDROID_HOST_OUT=$PROG_DIR/..
 ANDROID_DATA=$PWD/android-data$$
 
-mkdir -p $ANDROID_DATA/dalvik-cache
+mkdir -p $ANDROID_DATA/dalvik-cache/{x86,x86_64}
 ANDROID_DATA=$ANDROID_DATA \
   ANDROID_ROOT=$ANDROID_HOST_OUT \
   LD_LIBRARY_PATH=$ANDROID_HOST_OUT/lib \