diff --git a/compiler/dex/compiler_enums.h b/compiler/dex/compiler_enums.h
index de9ac4b..caecb7a 100644
--- a/compiler/dex/compiler_enums.h
+++ b/compiler/dex/compiler_enums.h
@@ -527,6 +527,13 @@
 
 std::ostream& operator<<(std::ostream& os, const FixupKind& kind);
 
+enum VolatileKind {
+  kNotVolatile,      // Load/Store is not volatile
+  kVolatile          // Load/Store is volatile
+};
+
+std::ostream& operator<<(std::ostream& os, const VolatileKind& kind);
+
 }  // namespace art
 
 #endif  // ART_COMPILER_DEX_COMPILER_ENUMS_H_
diff --git a/compiler/dex/quick/arm/call_arm.cc b/compiler/dex/quick/arm/call_arm.cc
index 590c767..04d6898 100644
--- a/compiler/dex/quick/arm/call_arm.cc
+++ b/compiler/dex/quick/arm/call_arm.cc
@@ -316,9 +316,9 @@
   int ex_offset = Thread::ExceptionOffset<4>().Int32Value();
   RegLocation rl_result = EvalLoc(rl_dest, kRefReg, true);
   RegStorage reset_reg = AllocTempRef();
-  LoadRefDisp(rs_rARM_SELF, ex_offset, rl_result.reg);
+  LoadRefDisp(rs_rARM_SELF, ex_offset, rl_result.reg, kNotVolatile);
   LoadConstant(reset_reg, 0);
-  StoreRefDisp(rs_rARM_SELF, ex_offset, reset_reg);
+  StoreRefDisp(rs_rARM_SELF, ex_offset, reset_reg, kNotVolatile);
   FreeTemp(reset_reg);
   StoreValue(rl_dest, rl_result);
 }
diff --git a/compiler/dex/quick/arm/codegen_arm.h b/compiler/dex/quick/arm/codegen_arm.h
index 4499862..70dce7f 100644
--- a/compiler/dex/quick/arm/codegen_arm.h
+++ b/compiler/dex/quick/arm/codegen_arm.h
@@ -33,20 +33,16 @@
     LIR* CheckSuspendUsingLoad() OVERRIDE;
     RegStorage LoadHelper(ThreadOffset<4> offset) OVERRIDE;
     RegStorage LoadHelper(ThreadOffset<8> offset) OVERRIDE;
-    LIR* LoadBaseDispVolatile(RegStorage r_base, int displacement, RegStorage r_dest,
-                              OpSize size) OVERRIDE;
     LIR* LoadBaseDisp(RegStorage r_base, int displacement, RegStorage r_dest,
-                      OpSize size) OVERRIDE;
+                      OpSize size, VolatileKind is_volatile) OVERRIDE;
     LIR* LoadBaseIndexed(RegStorage r_base, RegStorage r_index, RegStorage r_dest, int scale,
                          OpSize size) OVERRIDE;
     LIR* LoadBaseIndexedDisp(RegStorage r_base, RegStorage r_index, int scale, int displacement,
                              RegStorage r_dest, OpSize size) OVERRIDE;
     LIR* LoadConstantNoClobber(RegStorage r_dest, int value);
     LIR* LoadConstantWide(RegStorage r_dest, int64_t value);
-    LIR* StoreBaseDispVolatile(RegStorage r_base, int displacement, RegStorage r_src,
-                               OpSize size) OVERRIDE;
     LIR* StoreBaseDisp(RegStorage r_base, int displacement, RegStorage r_src,
-                       OpSize size) OVERRIDE;
+                       OpSize size, VolatileKind is_volatile) OVERRIDE;
     LIR* StoreBaseIndexed(RegStorage r_base, RegStorage r_index, RegStorage r_src, int scale,
                           OpSize size) OVERRIDE;
     LIR* StoreBaseIndexedDisp(RegStorage r_base, RegStorage r_index, int scale, int displacement,
diff --git a/compiler/dex/quick/arm/int_arm.cc b/compiler/dex/quick/arm/int_arm.cc
index 916c528..e34d944 100644
--- a/compiler/dex/quick/arm/int_arm.cc
+++ b/compiler/dex/quick/arm/int_arm.cc
@@ -723,7 +723,7 @@
   } else {
     DCHECK(size == kSignedByte || size == kSignedHalf || size == k32);
     // Unaligned load with LDR and LDRSH is allowed on ARMv7 with SCTLR.A set to 0.
-    LoadBaseDisp(rl_address.reg, 0, rl_result.reg, size);
+    LoadBaseDisp(rl_address.reg, 0, rl_result.reg, size, kNotVolatile);
     StoreValue(rl_dest, rl_result);
   }
   return true;
@@ -737,13 +737,13 @@
   if (size == k64) {
     // Fake unaligned STRD by two unaligned STR instructions on ARMv7 with SCTLR.A set to 0.
     RegLocation rl_value = LoadValueWide(rl_src_value, kCoreReg);
-    StoreBaseDisp(rl_address.reg, 0, rl_value.reg.GetLow(), k32);
-    StoreBaseDisp(rl_address.reg, 4, rl_value.reg.GetHigh(), k32);
+    StoreBaseDisp(rl_address.reg, 0, rl_value.reg.GetLow(), k32, kNotVolatile);
+    StoreBaseDisp(rl_address.reg, 4, rl_value.reg.GetHigh(), k32, kNotVolatile);
   } else {
     DCHECK(size == kSignedByte || size == kSignedHalf || size == k32);
     // Unaligned store with STR and STRSH is allowed on ARMv7 with SCTLR.A set to 0.
     RegLocation rl_value = LoadValue(rl_src_value, kCoreReg);
-    StoreBaseDisp(rl_address.reg, 0, rl_value.reg, size);
+    StoreBaseDisp(rl_address.reg, 0, rl_value.reg, size, kNotVolatile);
   }
   return true;
 }
@@ -1230,7 +1230,7 @@
       }
       FreeTemp(reg_len);
     }
-    LoadBaseDisp(reg_ptr, data_offset, rl_result.reg, size);
+    LoadBaseDisp(reg_ptr, data_offset, rl_result.reg, size, kNotVolatile);
     MarkPossibleNullPointerException(opt_flags);
     if (!constant_index) {
       FreeTemp(reg_ptr);
@@ -1330,7 +1330,7 @@
       FreeTemp(reg_len);
     }
 
-    StoreBaseDisp(reg_ptr, data_offset, rl_src.reg, size);
+    StoreBaseDisp(reg_ptr, data_offset, rl_src.reg, size, kNotVolatile);
     MarkPossibleNullPointerException(opt_flags);
   } else {
     /* reg_ptr -> array data */
diff --git a/compiler/dex/quick/arm/utility_arm.cc b/compiler/dex/quick/arm/utility_arm.cc
index b236f99..bc8f95b 100644
--- a/compiler/dex/quick/arm/utility_arm.cc
+++ b/compiler/dex/quick/arm/utility_arm.cc
@@ -961,31 +961,37 @@
   return load;
 }
 
-LIR* ArmMir2Lir::LoadBaseDispVolatile(RegStorage r_base, int displacement, RegStorage r_dest,
-                                      OpSize size) {
-  // Only 64-bit load needs special handling.
-  if (UNLIKELY(size == k64 || size == kDouble)) {
-    DCHECK(!r_dest.IsFloat());  // See RegClassForFieldLoadSave().
-    // If the cpu supports LPAE, aligned LDRD is atomic - fall through to LoadBaseDisp().
-    if (!cu_->compiler_driver->GetInstructionSetFeatures().HasLpae()) {
-      // Use LDREXD for the atomic load. (Expect displacement > 0, don't optimize for == 0.)
-      RegStorage r_ptr = AllocTemp();
-      OpRegRegImm(kOpAdd, r_ptr, r_base, displacement);
-      LIR* lir = NewLIR3(kThumb2Ldrexd, r_dest.GetLowReg(), r_dest.GetHighReg(), r_ptr.GetReg());
-      FreeTemp(r_ptr);
-      return lir;
-    }
-  }
-  return LoadBaseDisp(r_base, displacement, r_dest, size);
-}
-
 LIR* ArmMir2Lir::LoadBaseDisp(RegStorage r_base, int displacement, RegStorage r_dest,
-                              OpSize size) {
+                              OpSize size, VolatileKind is_volatile) {
   // TODO: base this on target.
   if (size == kWord) {
     size = k32;
   }
-  return LoadBaseDispBody(r_base, displacement, r_dest, size);
+  LIR* load;
+  if (UNLIKELY(is_volatile == kVolatile &&
+               (size == k64 || size == kDouble) &&
+               !cu_->compiler_driver->GetInstructionSetFeatures().HasLpae())) {
+    // Only 64-bit load needs special handling.
+    // If the cpu supports LPAE, aligned LDRD is atomic - fall through to LoadBaseDisp().
+    DCHECK(!r_dest.IsFloat());  // See RegClassForFieldLoadSave().
+    // Use LDREXD for the atomic load. (Expect displacement > 0, don't optimize for == 0.)
+    RegStorage r_ptr = AllocTemp();
+    OpRegRegImm(kOpAdd, r_ptr, r_base, displacement);
+    LIR* lir = NewLIR3(kThumb2Ldrexd, r_dest.GetLowReg(), r_dest.GetHighReg(), r_ptr.GetReg());
+    FreeTemp(r_ptr);
+    return lir;
+  } else {
+    load = LoadBaseDispBody(r_base, displacement, r_dest, size);
+  }
+
+  if (UNLIKELY(is_volatile == kVolatile)) {
+    // Without context sensitive analysis, we must issue the most conservative barriers.
+    // In this case, either a load or store may follow so we issue both barriers.
+    GenMemBarrier(kLoadLoad);
+    GenMemBarrier(kLoadStore);
+  }
+
+  return load;
 }
 
 
@@ -1081,49 +1087,58 @@
   return store;
 }
 
-LIR* ArmMir2Lir::StoreBaseDispVolatile(RegStorage r_base, int displacement, RegStorage r_src,
-                                       OpSize size) {
-  // Only 64-bit store needs special handling.
-  if (UNLIKELY(size == k64 || size == kDouble)) {
-    DCHECK(!r_src.IsFloat());  // See RegClassForFieldLoadSave().
-    // If the cpu supports LPAE, aligned STRD is atomic - fall through to StoreBaseDisp().
-    if (!cu_->compiler_driver->GetInstructionSetFeatures().HasLpae()) {
-      // Use STREXD for the atomic store. (Expect displacement > 0, don't optimize for == 0.)
-      RegStorage r_ptr = AllocTemp();
-      OpRegRegImm(kOpAdd, r_ptr, r_base, displacement);
-      LIR* fail_target = NewLIR0(kPseudoTargetLabel);
-      // We have only 5 temporary registers available and if r_base, r_src and r_ptr already
-      // take 4, we can't directly allocate 2 more for LDREXD temps. In that case clobber r_ptr
-      // in LDREXD and recalculate it from r_base.
-      RegStorage r_temp = AllocTemp();
-      RegStorage r_temp_high = AllocFreeTemp();  // We may not have another temp.
-      if (r_temp_high.Valid()) {
-        NewLIR3(kThumb2Ldrexd, r_temp.GetReg(), r_temp_high.GetReg(), r_ptr.GetReg());
-        FreeTemp(r_temp_high);
-        FreeTemp(r_temp);
-      } else {
-        // If we don't have another temp, clobber r_ptr in LDREXD and reload it.
-        NewLIR3(kThumb2Ldrexd, r_temp.GetReg(), r_ptr.GetReg(), r_ptr.GetReg());
-        FreeTemp(r_temp);  // May need the temp for kOpAdd.
-        OpRegRegImm(kOpAdd, r_ptr, r_base, displacement);
-      }
-      LIR* lir = NewLIR4(kThumb2Strexd, r_temp.GetReg(), r_src.GetLowReg(), r_src.GetHighReg(),
-                         r_ptr.GetReg());
-      OpCmpImmBranch(kCondNe, r_temp, 0, fail_target);
-      FreeTemp(r_ptr);
-      return lir;
-    }
-  }
-  return StoreBaseDisp(r_base, displacement, r_src, size);
-}
-
 LIR* ArmMir2Lir::StoreBaseDisp(RegStorage r_base, int displacement, RegStorage r_src,
-                               OpSize size) {
-  // TODO: base this on target.
-  if (size == kWord) {
-    size = k32;
+                               OpSize size, VolatileKind is_volatile) {
+  if (UNLIKELY(is_volatile == kVolatile)) {
+    // There might have been a store before this volatile one so insert StoreStore barrier.
+    GenMemBarrier(kStoreStore);
   }
-  return StoreBaseDispBody(r_base, displacement, r_src, size);
+
+  LIR* store;
+  if (UNLIKELY(is_volatile == kVolatile &&
+               (size == k64 || size == kDouble) &&
+               !cu_->compiler_driver->GetInstructionSetFeatures().HasLpae())) {
+    // Only 64-bit store needs special handling.
+    // If the cpu supports LPAE, aligned STRD is atomic - fall through to StoreBaseDisp().
+    // Use STREXD for the atomic store. (Expect displacement > 0, don't optimize for == 0.)
+    DCHECK(!r_src.IsFloat());  // See RegClassForFieldLoadSave().
+    RegStorage r_ptr = AllocTemp();
+    OpRegRegImm(kOpAdd, r_ptr, r_base, displacement);
+    LIR* fail_target = NewLIR0(kPseudoTargetLabel);
+    // We have only 5 temporary registers available and if r_base, r_src and r_ptr already
+    // take 4, we can't directly allocate 2 more for LDREXD temps. In that case clobber r_ptr
+    // in LDREXD and recalculate it from r_base.
+    RegStorage r_temp = AllocTemp();
+    RegStorage r_temp_high = AllocFreeTemp();  // We may not have another temp.
+    if (r_temp_high.Valid()) {
+      NewLIR3(kThumb2Ldrexd, r_temp.GetReg(), r_temp_high.GetReg(), r_ptr.GetReg());
+      FreeTemp(r_temp_high);
+      FreeTemp(r_temp);
+    } else {
+      // If we don't have another temp, clobber r_ptr in LDREXD and reload it.
+      NewLIR3(kThumb2Ldrexd, r_temp.GetReg(), r_ptr.GetReg(), r_ptr.GetReg());
+      FreeTemp(r_temp);  // May need the temp for kOpAdd.
+      OpRegRegImm(kOpAdd, r_ptr, r_base, displacement);
+    }
+    store = NewLIR4(kThumb2Strexd, r_temp.GetReg(), r_src.GetLowReg(), r_src.GetHighReg(),
+                    r_ptr.GetReg());
+    OpCmpImmBranch(kCondNe, r_temp, 0, fail_target);
+    FreeTemp(r_ptr);
+  } else {
+    // TODO: base this on target.
+    if (size == kWord) {
+      size = k32;
+    }
+
+    store = StoreBaseDispBody(r_base, displacement, r_src, size);
+  }
+
+  if (UNLIKELY(is_volatile == kVolatile)) {
+    // A load might follow the volatile store so insert a StoreLoad barrier.
+    GenMemBarrier(kStoreLoad);
+  }
+
+  return store;
 }
 
 LIR* ArmMir2Lir::OpFpRegCopy(RegStorage r_dest, RegStorage r_src) {
diff --git a/compiler/dex/quick/arm64/arm64_lir.h b/compiler/dex/quick/arm64/arm64_lir.h
index 3f32c51..1f1a252 100644
--- a/compiler/dex/quick/arm64/arm64_lir.h
+++ b/compiler/dex/quick/arm64/arm64_lir.h
@@ -101,6 +101,7 @@
 // Temporary macros, used to mark code which wants to distinguish betweek zr/sp.
 #define A64_REG_IS_SP(reg_num) ((reg_num) == rwsp || (reg_num) == rsp)
 #define A64_REG_IS_ZR(reg_num) ((reg_num) == rwzr || (reg_num) == rxzr)
+#define A64_REGSTORAGE_IS_SP_OR_ZR(rs) (((rs).GetRegNum() & 0x1f) == 0x1f)
 
 enum Arm64ResourceEncodingPos {
   kArm64GPReg0   = 0,
diff --git a/compiler/dex/quick/arm64/assemble_arm64.cc b/compiler/dex/quick/arm64/assemble_arm64.cc
index 2a8da24..bee64f1 100644
--- a/compiler/dex/quick/arm64/assemble_arm64.cc
+++ b/compiler/dex/quick/arm64/assemble_arm64.cc
@@ -632,19 +632,19 @@
           if (static_cast<unsigned>(kind) < kFmtBitBlt) {
             bool is_zero = A64_REG_IS_ZR(operand);
 
-            if (kIsDebugBuild) {
+            if (kIsDebugBuild && (kFailOnSizeError || kReportSizeError)) {
               // Register usage checks: First establish register usage requirements based on the
               // format in `kind'.
               bool want_float = false;
               bool want_64_bit = false;
-              bool want_size_match = false;
+              bool want_var_size = true;
               bool want_zero = false;
               switch (kind) {
                 case kFmtRegX:
                   want_64_bit = true;
                   // Intentional fall-through.
                 case kFmtRegW:
-                  want_size_match = true;
+                  want_var_size = false;
                   // Intentional fall-through.
                 case kFmtRegR:
                   want_zero = true;
@@ -653,7 +653,7 @@
                   want_64_bit = true;
                   // Intentional fall-through.
                 case kFmtRegWOrSp:
-                  want_size_match = true;
+                  want_var_size = false;
                   break;
                 case kFmtRegROrSp:
                   break;
@@ -661,7 +661,7 @@
                   want_64_bit = true;
                   // Intentional fall-through.
                 case kFmtRegS:
-                  want_size_match = true;
+                  want_var_size = false;
                   // Intentional fall-through.
                 case kFmtRegF:
                   want_float = true;
@@ -672,21 +672,27 @@
                   break;
               }
 
+              // want_var_size == true means kind == kFmtReg{R,F}. In these two cases, we want
+              // the register size to be coherent with the instruction width.
+              if (want_var_size) {
+                want_64_bit = opcode_is_wide;
+              }
+
               // Now check that the requirements are satisfied.
               RegStorage reg(operand | RegStorage::kValid);
               const char *expected = nullptr;
               if (want_float) {
                 if (!reg.IsFloat()) {
                   expected = "float register";
-                } else if (want_size_match && (reg.IsDouble() != want_64_bit)) {
+                } else if (reg.IsDouble() != want_64_bit) {
                   expected = (want_64_bit) ? "double register" : "single register";
                 }
               } else {
                 if (reg.IsFloat()) {
                   expected = "core register";
-                } else if (want_size_match && (reg.Is64Bit() != want_64_bit)) {
+                } else if (reg.Is64Bit() != want_64_bit) {
                   expected = (want_64_bit) ? "x-register" : "w-register";
-                } else if (reg.GetRegNum() == 31 && is_zero != want_zero) {
+                } else if (A64_REGSTORAGE_IS_SP_OR_ZR(reg) && is_zero != want_zero) {
                   expected = (want_zero) ? "zero-register" : "sp-register";
                 }
               }
@@ -698,8 +704,13 @@
               if (expected != nullptr) {
                 LOG(WARNING) << "Method: " << PrettyMethod(cu_->method_idx, *cu_->dex_file)
                              << " @ 0x" << std::hex << lir->dalvik_offset;
-                LOG(FATAL) << "Bad argument n. " << i << " of " << encoder->name
-                           << ". Expected " << expected << ", got 0x" << std::hex << operand;
+                if (kFailOnSizeError) {
+                  LOG(FATAL) << "Bad argument n. " << i << " of " << encoder->name
+                             << ". Expected " << expected << ", got 0x" << std::hex << operand;
+                } else {
+                  LOG(WARNING) << "Bad argument n. " << i << " of " << encoder->name
+                               << ". Expected " << expected << ", got 0x" << std::hex << operand;
+                }
               }
             }
 
diff --git a/compiler/dex/quick/arm64/call_arm64.cc b/compiler/dex/quick/arm64/call_arm64.cc
index 1df576b..c3f4711 100644
--- a/compiler/dex/quick/arm64/call_arm64.cc
+++ b/compiler/dex/quick/arm64/call_arm64.cc
@@ -267,7 +267,7 @@
   MarkPossibleNullPointerException(opt_flags);
   LIR* slow_unlock_branch = OpCmpBranch(kCondNe, rs_w1, rs_w2, NULL);
   GenMemBarrier(kStoreLoad);
-  Store32Disp(rs_x0, mirror::Object::MonitorOffset().Int32Value(), rs_xzr);
+  Store32Disp(rs_x0, mirror::Object::MonitorOffset().Int32Value(), rs_wzr);
   LIR* unlock_success_branch = OpUnconditionalBranch(NULL);
 
   LIR* slow_path_target = NewLIR0(kPseudoTargetLabel);
@@ -289,8 +289,8 @@
 void Arm64Mir2Lir::GenMoveException(RegLocation rl_dest) {
   int ex_offset = Thread::ExceptionOffset<8>().Int32Value();
   RegLocation rl_result = EvalLoc(rl_dest, kRefReg, true);
-  LoadRefDisp(rs_rA64_SELF, ex_offset, rl_result.reg);
-  StoreRefDisp(rs_rA64_SELF, ex_offset, rs_xzr);
+  LoadRefDisp(rs_rA64_SELF, ex_offset, rl_result.reg, kNotVolatile);
+  StoreRefDisp(rs_rA64_SELF, ex_offset, rs_xzr, kNotVolatile);
   StoreValue(rl_dest, rl_result);
 }
 
diff --git a/compiler/dex/quick/arm64/codegen_arm64.h b/compiler/dex/quick/arm64/codegen_arm64.h
index f1270ec..68fa6f4 100644
--- a/compiler/dex/quick/arm64/codegen_arm64.h
+++ b/compiler/dex/quick/arm64/codegen_arm64.h
@@ -26,6 +26,11 @@
 
 class Arm64Mir2Lir : public Mir2Lir {
  protected:
+  // If we detect a size error, FATAL out.
+  static constexpr bool kFailOnSizeError = false && kIsDebugBuild;
+  // If we detect a size error, report to LOG.
+  static constexpr bool kReportSizeError = false && kIsDebugBuild;
+
   // TODO: consolidate 64-bit target support.
   class InToRegStorageMapper {
    public:
@@ -69,22 +74,25 @@
     LIR* CheckSuspendUsingLoad() OVERRIDE;
     RegStorage LoadHelper(ThreadOffset<4> offset) OVERRIDE;
     RegStorage LoadHelper(ThreadOffset<8> offset) OVERRIDE;
-    LIR* LoadBaseDispVolatile(RegStorage r_base, int displacement, RegStorage r_dest,
-                              OpSize size) OVERRIDE;
     LIR* LoadBaseDisp(RegStorage r_base, int displacement, RegStorage r_dest,
-                      OpSize size) OVERRIDE;
+                      OpSize size, VolatileKind is_volatile) OVERRIDE;
+    LIR* LoadRefDisp(RegStorage r_base, int displacement, RegStorage r_dest,
+                     VolatileKind is_volatile)
+        OVERRIDE;
     LIR* LoadBaseIndexed(RegStorage r_base, RegStorage r_index, RegStorage r_dest, int scale,
                          OpSize size) OVERRIDE;
+    LIR* LoadRefIndexed(RegStorage r_base, RegStorage r_index, RegStorage r_dest) OVERRIDE;
     LIR* LoadBaseIndexedDisp(RegStorage r_base, RegStorage r_index, int scale, int displacement,
                              RegStorage r_dest, OpSize size) OVERRIDE;
     LIR* LoadConstantNoClobber(RegStorage r_dest, int value);
     LIR* LoadConstantWide(RegStorage r_dest, int64_t value);
-    LIR* StoreBaseDispVolatile(RegStorage r_base, int displacement, RegStorage r_dest,
-                               OpSize size) OVERRIDE;
     LIR* StoreBaseDisp(RegStorage r_base, int displacement, RegStorage r_src,
-                       OpSize size) OVERRIDE;
+                       OpSize size, VolatileKind is_volatile) OVERRIDE;
+    LIR* StoreRefDisp(RegStorage r_base, int displacement, RegStorage r_src,
+                      VolatileKind is_volatile) OVERRIDE;
     LIR* StoreBaseIndexed(RegStorage r_base, RegStorage r_index, RegStorage r_src, int scale,
                           OpSize size) OVERRIDE;
+    LIR* StoreRefIndexed(RegStorage r_base, RegStorage r_index, RegStorage r_src) OVERRIDE;
     LIR* StoreBaseIndexedDisp(RegStorage r_base, RegStorage r_index, int scale, int displacement,
                               RegStorage r_src, OpSize size) OVERRIDE;
     void MarkGCCard(RegStorage val_reg, RegStorage tgt_addr_reg) OVERRIDE;
@@ -283,8 +291,15 @@
      * @see As64BitReg
      */
     RegStorage As32BitReg(RegStorage reg) {
-      DCHECK(reg.Is64Bit());
       DCHECK(!reg.IsPair());
+      if ((kFailOnSizeError || kReportSizeError) && !reg.Is64Bit()) {
+        if (kFailOnSizeError) {
+          LOG(FATAL) << "Expected 64b register";
+        } else {
+          LOG(WARNING) << "Expected 64b register";
+          return reg;
+        }
+      }
       RegStorage ret_val = RegStorage(RegStorage::k32BitSolo,
                                       reg.GetRawBits() & RegStorage::kRegTypeMask);
       DCHECK_EQ(GetRegInfo(reg)->FindMatchingView(RegisterInfo::k32SoloStorageMask)
@@ -293,6 +308,18 @@
       return ret_val;
     }
 
+    RegStorage Check32BitReg(RegStorage reg) {
+      if ((kFailOnSizeError || kReportSizeError) && !reg.Is32Bit()) {
+        if (kFailOnSizeError) {
+          LOG(FATAL) << "Checked for 32b register";
+        } else {
+          LOG(WARNING) << "Checked for 32b register";
+          return As32BitReg(reg);
+        }
+      }
+      return reg;
+    }
+
     /**
      * @brief Given register wNN (sNN), returns register xNN (dNN).
      * @param reg #RegStorage containing a Solo32 input register (e.g. @c w1 or @c s2).
@@ -300,8 +327,15 @@
      * @see As32BitReg
      */
     RegStorage As64BitReg(RegStorage reg) {
-      DCHECK(reg.Is32Bit());
       DCHECK(!reg.IsPair());
+      if ((kFailOnSizeError || kReportSizeError) && !reg.Is32Bit()) {
+        if (kFailOnSizeError) {
+          LOG(FATAL) << "Expected 32b register";
+        } else {
+          LOG(WARNING) << "Expected 32b register";
+          return reg;
+        }
+      }
       RegStorage ret_val = RegStorage(RegStorage::k64BitSolo,
                                       reg.GetRawBits() & RegStorage::kRegTypeMask);
       DCHECK_EQ(GetRegInfo(reg)->FindMatchingView(RegisterInfo::k64SoloStorageMask)
@@ -310,6 +344,18 @@
       return ret_val;
     }
 
+    RegStorage Check64BitReg(RegStorage reg) {
+      if ((kFailOnSizeError || kReportSizeError) && !reg.Is64Bit()) {
+        if (kFailOnSizeError) {
+          LOG(FATAL) << "Checked for 64b register";
+        } else {
+          LOG(WARNING) << "Checked for 64b register";
+          return As64BitReg(reg);
+        }
+      }
+      return reg;
+    }
+
     LIR* LoadFPConstantValue(RegStorage r_dest, int32_t value);
     LIR* LoadFPConstantValueWide(RegStorage r_dest, int64_t value);
     void ReplaceFixup(LIR* prev_lir, LIR* orig_lir, LIR* new_lir);
diff --git a/compiler/dex/quick/arm64/int_arm64.cc b/compiler/dex/quick/arm64/int_arm64.cc
index 2ac4adb..1fdbe2d 100644
--- a/compiler/dex/quick/arm64/int_arm64.cc
+++ b/compiler/dex/quick/arm64/int_arm64.cc
@@ -410,7 +410,7 @@
   RegLocation rl_address = LoadValue(rl_src_address, kCoreReg);   // kRefReg
   RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true);
 
-  LoadBaseDisp(rl_address.reg, 0, rl_result.reg, size);
+  LoadBaseDisp(rl_address.reg, 0, rl_result.reg, size, kNotVolatile);
   if (size == k64) {
     StoreValueWide(rl_dest, rl_result);
   } else {
@@ -433,7 +433,7 @@
     DCHECK(size == kSignedByte || size == kSignedHalf || size == k32);
     rl_value = LoadValue(rl_src_value, kCoreReg);
   }
-  StoreBaseDisp(rl_address.reg, 0, rl_value.reg, size);
+  StoreBaseDisp(rl_address.reg, 0, rl_value.reg, size, kNotVolatile);
   return true;
 }
 
@@ -747,7 +747,11 @@
       }
       FreeTemp(reg_len);
     }
-    LoadBaseDisp(reg_ptr, data_offset, rl_result.reg, size);
+    if (rl_result.ref) {
+      LoadRefDisp(reg_ptr, data_offset, rl_result.reg, kNotVolatile);
+    } else {
+      LoadBaseDisp(reg_ptr, data_offset, rl_result.reg, size, kNotVolatile);
+    }
     MarkPossibleNullPointerException(opt_flags);
     if (!constant_index) {
       FreeTemp(reg_ptr);
@@ -768,7 +772,11 @@
       GenArrayBoundsCheck(rl_index.reg, reg_len);
       FreeTemp(reg_len);
     }
-    LoadBaseIndexed(reg_ptr, As64BitReg(rl_index.reg), rl_result.reg, scale, size);
+    if (rl_result.ref) {
+      LoadRefIndexed(reg_ptr, As64BitReg(rl_index.reg), rl_result.reg);
+    } else {
+      LoadBaseIndexed(reg_ptr, As64BitReg(rl_index.reg), rl_result.reg, scale, size);
+    }
     MarkPossibleNullPointerException(opt_flags);
     FreeTemp(reg_ptr);
     StoreValue(rl_dest, rl_result);
@@ -847,8 +855,11 @@
       }
       FreeTemp(reg_len);
     }
-
-    StoreBaseDisp(reg_ptr, data_offset, rl_src.reg, size);
+    if (rl_src.ref) {
+      StoreRefDisp(reg_ptr, data_offset, rl_src.reg, kNotVolatile);
+    } else {
+      StoreBaseDisp(reg_ptr, data_offset, rl_src.reg, size, kNotVolatile);
+    }
     MarkPossibleNullPointerException(opt_flags);
   } else {
     /* reg_ptr -> array data */
@@ -858,7 +869,11 @@
       GenArrayBoundsCheck(rl_index.reg, reg_len);
       FreeTemp(reg_len);
     }
-    StoreBaseIndexed(reg_ptr, As64BitReg(rl_index.reg), rl_src.reg, scale, size);
+    if (rl_src.ref) {
+      StoreRefIndexed(reg_ptr, As64BitReg(rl_index.reg), rl_src.reg);
+    } else {
+      StoreBaseIndexed(reg_ptr, As64BitReg(rl_index.reg), rl_src.reg, scale, size);
+    }
     MarkPossibleNullPointerException(opt_flags);
   }
   if (allocated_reg_ptr_temp) {
diff --git a/compiler/dex/quick/arm64/target_arm64.cc b/compiler/dex/quick/arm64/target_arm64.cc
index 06e1cda..dfaa483 100644
--- a/compiler/dex/quick/arm64/target_arm64.cc
+++ b/compiler/dex/quick/arm64/target_arm64.cc
@@ -789,7 +789,7 @@
 RegStorage Arm64Mir2Lir::LoadHelper(ThreadOffset<8> offset) {
   // TODO(Arm64): use LoadWordDisp instead.
   //   e.g. LoadWordDisp(rs_rA64_SELF, offset.Int32Value(), rs_rA64_LR);
-  LoadBaseDisp(rs_rA64_SELF, offset.Int32Value(), rs_rA64_LR, k64);
+  LoadBaseDisp(rs_rA64_SELF, offset.Int32Value(), rs_rA64_LR, k64, kNotVolatile);
   return rs_rA64_LR;
 }
 
@@ -949,7 +949,7 @@
   StoreValue(rl_method, rl_src);
   // If Method* has been promoted, explicitly flush
   if (rl_method.location == kLocPhysReg) {
-    StoreRefDisp(TargetReg(kSp), 0, TargetReg(kArg0));
+    StoreRefDisp(TargetReg(kSp), 0, TargetReg(kArg0), kNotVolatile);
   }
 
   if (cu_->num_ins == 0) {
@@ -971,7 +971,7 @@
       } else if ((v_map->fp_location == kLocPhysReg) && t_loc->fp) {
         OpRegCopy(RegStorage::Solo32(v_map->FpReg), reg);
       } else {
-        StoreBaseDisp(TargetReg(kSp), SRegOffset(start_vreg + i), reg, op_size);
+        StoreBaseDisp(TargetReg(kSp), SRegOffset(start_vreg + i), reg, op_size, kNotVolatile);
         if (reg.Is64Bit()) {
           if (SRegOffset(start_vreg + i) + 4 != SRegOffset(start_vreg + i + 1)) {
             LOG(FATAL) << "64 bit value stored in non-consecutive 4 bytes slots";
@@ -1057,14 +1057,14 @@
         loc = UpdateLocWide(loc);
         if (loc.location == kLocPhysReg) {
           ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg);
-          StoreBaseDisp(TargetReg(kSp), SRegOffset(loc.s_reg_low), loc.reg, k64);
+          StoreBaseDisp(TargetReg(kSp), SRegOffset(loc.s_reg_low), loc.reg, k64, kNotVolatile);
         }
         next_arg += 2;
       } else {
         loc = UpdateLoc(loc);
         if (loc.location == kLocPhysReg) {
           ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg);
-          StoreBaseDisp(TargetReg(kSp), SRegOffset(loc.s_reg_low), loc.reg, k32);
+          StoreBaseDisp(TargetReg(kSp), SRegOffset(loc.s_reg_low), loc.reg, k32, kNotVolatile);
         }
         next_arg++;
       }
@@ -1122,18 +1122,27 @@
           ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg);
           if (rl_arg.wide) {
             if (rl_arg.location == kLocPhysReg) {
-              StoreBaseDisp(TargetReg(kSp), out_offset, rl_arg.reg, k64);
+              StoreBaseDisp(TargetReg(kSp), out_offset, rl_arg.reg, k64, kNotVolatile);
             } else {
               LoadValueDirectWideFixed(rl_arg, regWide);
-              StoreBaseDisp(TargetReg(kSp), out_offset, regWide, k64);
+              StoreBaseDisp(TargetReg(kSp), out_offset, regWide, k64, kNotVolatile);
             }
             i++;
           } else {
             if (rl_arg.location == kLocPhysReg) {
-              StoreBaseDisp(TargetReg(kSp), out_offset, rl_arg.reg, k32);
+              if (rl_arg.ref) {
+                StoreRefDisp(TargetReg(kSp), out_offset, rl_arg.reg, kNotVolatile);
+              } else {
+                StoreBaseDisp(TargetReg(kSp), out_offset, rl_arg.reg, k32, kNotVolatile);
+              }
             } else {
-              LoadValueDirectFixed(rl_arg, regSingle);
-              StoreBaseDisp(TargetReg(kSp), out_offset, regSingle, k32);
+              if (rl_arg.ref) {
+                LoadValueDirectFixed(rl_arg, regSingle);
+                StoreRefDisp(TargetReg(kSp), out_offset, regSingle, kNotVolatile);
+              } else {
+                LoadValueDirectFixed(rl_arg, As32BitReg(regSingle));
+                StoreBaseDisp(TargetReg(kSp), out_offset, As32BitReg(regSingle), k32, kNotVolatile);
+              }
             }
           }
         }
diff --git a/compiler/dex/quick/arm64/utility_arm64.cc b/compiler/dex/quick/arm64/utility_arm64.cc
index 672aa88..12c2f41 100644
--- a/compiler/dex/quick/arm64/utility_arm64.cc
+++ b/compiler/dex/quick/arm64/utility_arm64.cc
@@ -893,9 +893,7 @@
   ArmOpcode opcode = kA64Brk1d;
   DCHECK(r_base.Is64Bit());
   // TODO: need a cleaner handling of index registers here and throughout.
-  if (r_index.Is32Bit()) {
-    r_index = As64BitReg(r_index);
-  }
+  r_index = Check32BitReg(r_index);
 
   if (r_dest.IsFloat()) {
     if (r_dest.IsDouble()) {
@@ -918,12 +916,14 @@
     case kDouble:
     case kWord:
     case k64:
+      r_dest = Check64BitReg(r_dest);
       opcode = WIDE(kA64Ldr4rXxG);
       expected_scale = 3;
       break;
     case kSingle:
     case k32:
     case kReference:
+      r_dest = Check32BitReg(r_dest);
       opcode = kA64Ldr4rXxG;
       expected_scale = 2;
       break;
@@ -959,6 +959,10 @@
   return load;
 }
 
+LIR* Arm64Mir2Lir::LoadRefIndexed(RegStorage r_base, RegStorage r_index, RegStorage r_dest) {
+  return LoadBaseIndexed(r_base, r_index, As32BitReg(r_dest), 2, kReference);
+}
+
 LIR* Arm64Mir2Lir::StoreBaseIndexed(RegStorage r_base, RegStorage r_index, RegStorage r_src,
                                     int scale, OpSize size) {
   LIR* store;
@@ -966,9 +970,7 @@
   ArmOpcode opcode = kA64Brk1d;
   DCHECK(r_base.Is64Bit());
   // TODO: need a cleaner handling of index registers here and throughout.
-  if (r_index.Is32Bit()) {
-    r_index = As64BitReg(r_index);
-  }
+  r_index = Check32BitReg(r_index);
 
   if (r_src.IsFloat()) {
     if (r_src.IsDouble()) {
@@ -991,12 +993,14 @@
     case kDouble:     // Intentional fall-trough.
     case kWord:       // Intentional fall-trough.
     case k64:
+      r_src = Check64BitReg(r_src);
       opcode = WIDE(kA64Str4rXxG);
       expected_scale = 3;
       break;
     case kSingle:     // Intentional fall-trough.
     case k32:         // Intentional fall-trough.
     case kReference:
+      r_src = Check32BitReg(r_src);
       opcode = kA64Str4rXxG;
       expected_scale = 2;
       break;
@@ -1026,6 +1030,10 @@
   return store;
 }
 
+LIR* Arm64Mir2Lir::StoreRefIndexed(RegStorage r_base, RegStorage r_index, RegStorage r_src) {
+  return StoreBaseIndexed(r_base, r_index, As32BitReg(r_src), 2, kReference);
+}
+
 /*
  * Load value from base + displacement.  Optionally perform null check
  * on base (which must have an associated s_reg and MIR).  If not
@@ -1042,6 +1050,7 @@
     case kDouble:     // Intentional fall-through.
     case kWord:       // Intentional fall-through.
     case k64:
+      r_dest = Check64BitReg(r_dest);
       scale = 3;
       if (r_dest.IsFloat()) {
         DCHECK(r_dest.IsDouble());
@@ -1055,6 +1064,7 @@
     case kSingle:     // Intentional fall-through.
     case k32:         // Intentional fall-trough.
     case kReference:
+      r_dest = Check32BitReg(r_dest);
       scale = 2;
       if (r_dest.IsFloat()) {
         DCHECK(r_dest.IsSingle());
@@ -1106,19 +1116,28 @@
   return load;
 }
 
-LIR* Arm64Mir2Lir::LoadBaseDispVolatile(RegStorage r_base, int displacement, RegStorage r_dest,
-                                        OpSize size) {
+LIR* Arm64Mir2Lir::LoadBaseDisp(RegStorage r_base, int displacement, RegStorage r_dest,
+                                OpSize size, VolatileKind is_volatile) {
   // LoadBaseDisp() will emit correct insn for atomic load on arm64
   // assuming r_dest is correctly prepared using RegClassForFieldLoadStore().
-  return LoadBaseDisp(r_base, displacement, r_dest, size);
+
+  LIR* load = LoadBaseDispBody(r_base, displacement, r_dest, size);
+
+  if (UNLIKELY(is_volatile == kVolatile)) {
+    // Without context sensitive analysis, we must issue the most conservative barriers.
+    // In this case, either a load or store may follow so we issue both barriers.
+    GenMemBarrier(kLoadLoad);
+    GenMemBarrier(kLoadStore);
+  }
+
+  return load;
 }
 
-LIR* Arm64Mir2Lir::LoadBaseDisp(RegStorage r_base, int displacement, RegStorage r_dest,
-                                OpSize size) {
-  return LoadBaseDispBody(r_base, displacement, r_dest, size);
+LIR* Arm64Mir2Lir::LoadRefDisp(RegStorage r_base, int displacement, RegStorage r_dest,
+                               VolatileKind is_volatile) {
+  return LoadBaseDisp(r_base, displacement, As32BitReg(r_dest), kReference, is_volatile);
 }
 
-
 LIR* Arm64Mir2Lir::StoreBaseDispBody(RegStorage r_base, int displacement, RegStorage r_src,
                                      OpSize size) {
   LIR* store = NULL;
@@ -1130,6 +1149,7 @@
     case kDouble:     // Intentional fall-through.
     case kWord:       // Intentional fall-through.
     case k64:
+      r_src = Check64BitReg(r_src);
       scale = 3;
       if (r_src.IsFloat()) {
         DCHECK(r_src.IsDouble());
@@ -1143,6 +1163,7 @@
     case kSingle:     // Intentional fall-through.
     case k32:         // Intentional fall-trough.
     case kReference:
+      r_src = Check32BitReg(r_src);
       scale = 2;
       if (r_src.IsFloat()) {
         DCHECK(r_src.IsSingle());
@@ -1188,16 +1209,29 @@
   return store;
 }
 
-LIR* Arm64Mir2Lir::StoreBaseDispVolatile(RegStorage r_base, int displacement, RegStorage r_src,
-                                         OpSize size) {
+LIR* Arm64Mir2Lir::StoreBaseDisp(RegStorage r_base, int displacement, RegStorage r_src,
+                                 OpSize size, VolatileKind is_volatile) {
+  if (UNLIKELY(is_volatile == kVolatile)) {
+    // There might have been a store before this volatile one so insert StoreStore barrier.
+    GenMemBarrier(kStoreStore);
+  }
+
   // StoreBaseDisp() will emit correct insn for atomic store on arm64
   // assuming r_dest is correctly prepared using RegClassForFieldLoadStore().
-  return StoreBaseDisp(r_base, displacement, r_src, size);
+
+  LIR* store = StoreBaseDispBody(r_base, displacement, r_src, size);
+
+  if (UNLIKELY(is_volatile == kVolatile)) {
+    // A load might follow the volatile store so insert a StoreLoad barrier.
+    GenMemBarrier(kStoreLoad);
+  }
+
+  return store;
 }
 
-LIR* Arm64Mir2Lir::StoreBaseDisp(RegStorage r_base, int displacement, RegStorage r_src,
-                                 OpSize size) {
-  return StoreBaseDispBody(r_base, displacement, r_src, size);
+LIR* Arm64Mir2Lir::StoreRefDisp(RegStorage r_base, int displacement, RegStorage r_src,
+                                VolatileKind is_volatile) {
+  return StoreBaseDisp(r_base, displacement, As32BitReg(r_src), kReference, is_volatile);
 }
 
 LIR* Arm64Mir2Lir::OpFpRegCopy(RegStorage r_dest, RegStorage r_src) {
diff --git a/compiler/dex/quick/codegen_util.cc b/compiler/dex/quick/codegen_util.cc
index ec0fb43..f31b670 100644
--- a/compiler/dex/quick/codegen_util.cc
+++ b/compiler/dex/quick/codegen_util.cc
@@ -79,6 +79,20 @@
   DCHECK(safepoint_pc->u.m.def_mask->Equals(kEncodeAll));
 }
 
+void Mir2Lir::MarkSafepointPCAfter(LIR* after) {
+  DCHECK(!after->flags.use_def_invalid);
+  after->u.m.def_mask = &kEncodeAll;
+  // As NewLIR0 uses Append, we need to create the LIR by hand.
+  LIR* safepoint_pc = RawLIR(current_dalvik_offset_, kPseudoSafepointPC);
+  if (after->next == nullptr) {
+    DCHECK_EQ(after, last_lir_insn_);
+    AppendLIR(safepoint_pc);
+  } else {
+    InsertLIRAfter(after, safepoint_pc);
+  }
+  DCHECK(safepoint_pc->u.m.def_mask->Equals(kEncodeAll));
+}
+
 /* Remove a LIR from the list. */
 void Mir2Lir::UnlinkLIR(LIR* lir) {
   if (UNLIKELY(lir == first_lir_insn_)) {
@@ -1112,7 +1126,7 @@
 
 /*
  * Insert an LIR instruction after the current instruction, which cannot be the
- * first instruction.
+ * last instruction.
  *
  * current_lir -> new_lir -> old_next
  */
diff --git a/compiler/dex/quick/gen_common.cc b/compiler/dex/quick/gen_common.cc
index e36b592..b00cbeb 100644
--- a/compiler/dex/quick/gen_common.cc
+++ b/compiler/dex/quick/gen_common.cc
@@ -196,6 +196,15 @@
   }
 }
 
+void Mir2Lir::MarkPossibleNullPointerExceptionAfter(int opt_flags, LIR* after) {
+  if (!cu_->compiler_driver->GetCompilerOptions().GetExplicitNullChecks()) {
+    if (!(cu_->disable_opt & (1 << kNullCheckElimination)) && (opt_flags & MIR_IGNORE_NULL_CHECK)) {
+      return;
+    }
+    MarkSafepointPCAfter(after);
+  }
+}
+
 void Mir2Lir::MarkPossibleStackOverflowException() {
   if (!cu_->compiler_driver->GetCompilerOptions().GetExplicitStackOverflowChecks()) {
     MarkSafepointPC(last_lir_insn_);
@@ -506,7 +515,7 @@
     for (int i = 0; i < elems; i++) {
       RegLocation rl_arg = LoadValue(info->args[i], kCoreReg);
       Store32Disp(TargetReg(kRet0),
-                    mirror::Array::DataOffset(component_size).Int32Value() + i * 4, rl_arg.reg);
+                  mirror::Array::DataOffset(component_size).Int32Value() + i * 4, rl_arg.reg);
       // If the LoadValue caused a temp to be allocated, free it
       if (IsTemp(rl_arg.reg)) {
         FreeTemp(rl_arg.reg);
@@ -575,7 +584,8 @@
       // Fast path, static storage base is this method's class
       RegLocation rl_method = LoadCurrMethod();
       r_base = AllocTempRef();
-      LoadRefDisp(rl_method.reg, mirror::ArtMethod::DeclaringClassOffset().Int32Value(), r_base);
+      LoadRefDisp(rl_method.reg, mirror::ArtMethod::DeclaringClassOffset().Int32Value(), r_base,
+                  kNotVolatile);
       if (IsTemp(rl_method.reg)) {
         FreeTemp(rl_method.reg);
       }
@@ -592,9 +602,10 @@
       LoadCurrMethodDirect(r_method);
       r_base = TargetReg(kArg0);
       LockTemp(r_base);
-      LoadRefDisp(r_method, mirror::ArtMethod::DexCacheResolvedTypesOffset().Int32Value(), r_base);
+      LoadRefDisp(r_method, mirror::ArtMethod::DexCacheResolvedTypesOffset().Int32Value(), r_base,
+                  kNotVolatile);
       int32_t offset_of_field = ObjArray::OffsetOfElement(field_info.StorageIndex()).Int32Value();
-      LoadRefDisp(r_base, offset_of_field, r_base);
+      LoadRefDisp(r_base, offset_of_field, r_base, kNotVolatile);
       // r_base now points at static storage (Class*) or NULL if the type is not yet resolved.
       if (!field_info.IsInitialized() &&
           (mir->optimization_flags & MIR_IGNORE_CLINIT_CHECK) == 0) {
@@ -626,14 +637,12 @@
     } else {
       rl_src = LoadValue(rl_src, reg_class);
     }
-    if (field_info.IsVolatile()) {
-      // There might have been a store before this volatile one so insert StoreStore barrier.
-      GenMemBarrier(kStoreStore);
-      StoreBaseDispVolatile(r_base, field_info.FieldOffset().Int32Value(), rl_src.reg, store_size);
-      // A load might follow the volatile store so insert a StoreLoad barrier.
-      GenMemBarrier(kStoreLoad);
+    if (is_object) {
+      StoreRefDisp(r_base, field_info.FieldOffset().Int32Value(), rl_src.reg,
+                   field_info.IsVolatile() ? kVolatile : kNotVolatile);
     } else {
-      StoreBaseDisp(r_base, field_info.FieldOffset().Int32Value(), rl_src.reg, store_size);
+      StoreBaseDisp(r_base, field_info.FieldOffset().Int32Value(), rl_src.reg, store_size,
+                    field_info.IsVolatile() ? kVolatile : kNotVolatile);
     }
     if (is_object && !mir_graph_->IsConstantNullRef(rl_src)) {
       MarkGCCard(rl_src.reg, r_base);
@@ -672,7 +681,8 @@
       // Fast path, static storage base is this method's class
       RegLocation rl_method  = LoadCurrMethod();
       r_base = AllocTempRef();
-      LoadRefDisp(rl_method.reg, mirror::ArtMethod::DeclaringClassOffset().Int32Value(), r_base);
+      LoadRefDisp(rl_method.reg, mirror::ArtMethod::DeclaringClassOffset().Int32Value(), r_base,
+                  kNotVolatile);
     } else {
       // Medium path, static storage base in a different class which requires checks that the other
       // class is initialized
@@ -685,9 +695,10 @@
       LoadCurrMethodDirect(r_method);
       r_base = TargetReg(kArg0);
       LockTemp(r_base);
-      LoadRefDisp(r_method, mirror::ArtMethod::DexCacheResolvedTypesOffset().Int32Value(), r_base);
+      LoadRefDisp(r_method, mirror::ArtMethod::DexCacheResolvedTypesOffset().Int32Value(), r_base,
+                  kNotVolatile);
       int32_t offset_of_field = ObjArray::OffsetOfElement(field_info.StorageIndex()).Int32Value();
-      LoadRefDisp(r_base, offset_of_field, r_base);
+      LoadRefDisp(r_base, offset_of_field, r_base, kNotVolatile);
       // r_base now points at static storage (Class*) or NULL if the type is not yet resolved.
       if (!field_info.IsInitialized() &&
           (mir->optimization_flags & MIR_IGNORE_CLINIT_CHECK) == 0) {
@@ -717,14 +728,12 @@
     RegLocation rl_result = EvalLoc(rl_dest, reg_class, true);
 
     int field_offset = field_info.FieldOffset().Int32Value();
-    if (field_info.IsVolatile()) {
-      LoadBaseDispVolatile(r_base, field_offset, rl_result.reg, load_size);
-      // Without context sensitive analysis, we must issue the most conservative barriers.
-      // In this case, either a load or store may follow so we issue both barriers.
-      GenMemBarrier(kLoadLoad);
-      GenMemBarrier(kLoadStore);
+    if (is_object) {
+      LoadRefDisp(r_base, field_offset, rl_result.reg, field_info.IsVolatile() ? kVolatile :
+          kNotVolatile);
     } else {
-      LoadBaseDisp(r_base, field_offset, rl_result.reg, load_size);
+      LoadBaseDisp(r_base, field_offset, rl_result.reg, load_size, field_info.IsVolatile() ?
+          kVolatile : kNotVolatile);
     }
     FreeTemp(r_base);
 
@@ -785,17 +794,15 @@
     GenNullCheck(rl_obj.reg, opt_flags);
     RegLocation rl_result = EvalLoc(rl_dest, reg_class, true);
     int field_offset = field_info.FieldOffset().Int32Value();
-    if (field_info.IsVolatile()) {
-      LoadBaseDispVolatile(rl_obj.reg, field_offset, rl_result.reg, load_size);
-      MarkPossibleNullPointerException(opt_flags);
-      // Without context sensitive analysis, we must issue the most conservative barriers.
-      // In this case, either a load or store may follow so we issue both barriers.
-      GenMemBarrier(kLoadLoad);
-      GenMemBarrier(kLoadStore);
+    LIR* load_lir;
+    if (is_object) {
+      load_lir = LoadRefDisp(rl_obj.reg, field_offset, rl_result.reg, field_info.IsVolatile() ?
+          kVolatile : kNotVolatile);
     } else {
-      LoadBaseDisp(rl_obj.reg, field_offset, rl_result.reg, load_size);
-      MarkPossibleNullPointerException(opt_flags);
+      load_lir = LoadBaseDisp(rl_obj.reg, field_offset, rl_result.reg, load_size,
+                              field_info.IsVolatile() ? kVolatile : kNotVolatile);
     }
+    MarkPossibleNullPointerExceptionAfter(opt_flags, load_lir);
     if (is_long_or_double) {
       StoreValueWide(rl_dest, rl_result);
     } else {
@@ -847,17 +854,15 @@
     }
     GenNullCheck(rl_obj.reg, opt_flags);
     int field_offset = field_info.FieldOffset().Int32Value();
-    if (field_info.IsVolatile()) {
-      // There might have been a store before this volatile one so insert StoreStore barrier.
-      GenMemBarrier(kStoreStore);
-      StoreBaseDispVolatile(rl_obj.reg, field_offset, rl_src.reg, store_size);
-      MarkPossibleNullPointerException(opt_flags);
-      // A load might follow the volatile store so insert a StoreLoad barrier.
-      GenMemBarrier(kStoreLoad);
+    LIR* store;
+    if (is_object) {
+      store = StoreRefDisp(rl_obj.reg, field_offset, rl_src.reg, field_info.IsVolatile() ?
+          kVolatile : kNotVolatile);
     } else {
-      StoreBaseDisp(rl_obj.reg, field_offset, rl_src.reg, store_size);
-      MarkPossibleNullPointerException(opt_flags);
+      store = StoreBaseDisp(rl_obj.reg, field_offset, rl_src.reg, store_size,
+                            field_info.IsVolatile() ? kVolatile : kNotVolatile);
     }
+    MarkPossibleNullPointerExceptionAfter(opt_flags, store);
     if (is_object && !mir_graph_->IsConstantNullRef(rl_src)) {
       MarkGCCard(rl_src.reg, rl_obj.reg);
     }
@@ -916,9 +921,9 @@
     // We're don't need access checks, load type from dex cache
     int32_t dex_cache_offset =
         mirror::ArtMethod::DexCacheResolvedTypesOffset().Int32Value();
-    LoadRefDisp(rl_method.reg, dex_cache_offset, res_reg);
+    LoadRefDisp(rl_method.reg, dex_cache_offset, res_reg, kNotVolatile);
     int32_t offset_of_type = ClassArray::OffsetOfElement(type_idx).Int32Value();
-    LoadRefDisp(res_reg, offset_of_type, rl_result.reg);
+    LoadRefDisp(res_reg, offset_of_type, rl_result.reg, kNotVolatile);
     if (!cu_->compiler_driver->CanAssumeTypeIsPresentInDexCache(*cu_->dex_file,
         type_idx) || SLOW_TYPE_PATH) {
       // Slow path, at runtime test if type is null and if so initialize
@@ -989,10 +994,10 @@
       LoadCurrMethodDirect(r_method);
     }
     LoadRefDisp(r_method, mirror::ArtMethod::DexCacheStringsOffset().Int32Value(),
-                TargetReg(kArg0));
+                TargetReg(kArg0), kNotVolatile);
 
     // Might call out to helper, which will return resolved string in kRet0
-    LoadRefDisp(TargetReg(kArg0), offset_of_string, TargetReg(kRet0));
+    LoadRefDisp(TargetReg(kArg0), offset_of_string, TargetReg(kRet0), kNotVolatile);
     LIR* fromfast = OpCmpImmBranch(kCondEq, TargetReg(kRet0), 0, NULL);
     LIR* cont = NewLIR0(kPseudoTargetLabel);
 
@@ -1031,8 +1036,9 @@
     RegLocation rl_method = LoadCurrMethod();
     RegStorage res_reg = AllocTempRef();
     RegLocation rl_result = EvalLoc(rl_dest, kRefReg, true);
-    LoadRefDisp(rl_method.reg, mirror::ArtMethod::DexCacheStringsOffset().Int32Value(), res_reg);
-    LoadRefDisp(res_reg, offset_of_string, rl_result.reg);
+    LoadRefDisp(rl_method.reg, mirror::ArtMethod::DexCacheStringsOffset().Int32Value(), res_reg,
+                kNotVolatile);
+    LoadRefDisp(res_reg, offset_of_string, rl_result.reg, kNotVolatile);
     StoreValue(rl_dest, rl_result);
   }
 }
@@ -1133,14 +1139,17 @@
 
   LoadCurrMethodDirect(check_class);
   if (use_declaring_class) {
-    LoadRefDisp(check_class, mirror::ArtMethod::DeclaringClassOffset().Int32Value(), check_class);
-    LoadRefDisp(object.reg,  mirror::Object::ClassOffset().Int32Value(), object_class);
+    LoadRefDisp(check_class, mirror::ArtMethod::DeclaringClassOffset().Int32Value(), check_class,
+                kNotVolatile);
+    LoadRefDisp(object.reg,  mirror::Object::ClassOffset().Int32Value(), object_class,
+                kNotVolatile);
   } else {
     LoadRefDisp(check_class, mirror::ArtMethod::DexCacheResolvedTypesOffset().Int32Value(),
-                check_class);
-    LoadRefDisp(object.reg,  mirror::Object::ClassOffset().Int32Value(), object_class);
+                check_class, kNotVolatile);
+    LoadRefDisp(object.reg,  mirror::Object::ClassOffset().Int32Value(), object_class,
+                kNotVolatile);
     int32_t offset_of_type = ClassArray::OffsetOfElement(type_idx).Int32Value();
-    LoadRefDisp(check_class, offset_of_type, check_class);
+    LoadRefDisp(check_class, offset_of_type, check_class, kNotVolatile);
   }
 
   LIR* ne_branchover = NULL;
@@ -1196,14 +1205,14 @@
   } else if (use_declaring_class) {
     LoadValueDirectFixed(rl_src, TargetReg(kArg0));  // kArg0 <= ref
     LoadRefDisp(TargetReg(kArg1), mirror::ArtMethod::DeclaringClassOffset().Int32Value(),
-                 class_reg);
+                class_reg, kNotVolatile);
   } else {
     // Load dex cache entry into class_reg (kArg2)
     LoadValueDirectFixed(rl_src, TargetReg(kArg0));  // kArg0 <= ref
     LoadRefDisp(TargetReg(kArg1), mirror::ArtMethod::DexCacheResolvedTypesOffset().Int32Value(),
-                class_reg);
+                class_reg, kNotVolatile);
     int32_t offset_of_type = ClassArray::OffsetOfElement(type_idx).Int32Value();
-    LoadRefDisp(class_reg, offset_of_type, class_reg);
+    LoadRefDisp(class_reg, offset_of_type, class_reg, kNotVolatile);
     if (!can_assume_type_is_in_dex_cache) {
       // Need to test presence of type in dex cache at runtime
       LIR* hop_branch = OpCmpImmBranch(kCondNe, class_reg, 0, NULL);
@@ -1231,7 +1240,8 @@
 
   /* load object->klass_ */
   DCHECK_EQ(mirror::Object::ClassOffset().Int32Value(), 0);
-  LoadRefDisp(TargetReg(kArg0),  mirror::Object::ClassOffset().Int32Value(), TargetReg(kArg1));
+  LoadRefDisp(TargetReg(kArg0), mirror::Object::ClassOffset().Int32Value(), TargetReg(kArg1),
+              kNotVolatile);
   /* kArg0 is ref, kArg1 is ref->klass_, kArg2 is class */
   LIR* branchover = NULL;
   if (type_known_final) {
@@ -1344,13 +1354,13 @@
     OpRegCopy(class_reg, TargetReg(kRet0));  // Align usage with fast path
   } else if (use_declaring_class) {
     LoadRefDisp(TargetReg(kArg1), mirror::ArtMethod::DeclaringClassOffset().Int32Value(),
-                class_reg);
+                class_reg, kNotVolatile);
   } else {
     // Load dex cache entry into class_reg (kArg2)
     LoadRefDisp(TargetReg(kArg1), mirror::ArtMethod::DexCacheResolvedTypesOffset().Int32Value(),
-                class_reg);
+                class_reg, kNotVolatile);
     int32_t offset_of_type = ClassArray::OffsetOfElement(type_idx).Int32Value();
-    LoadRefDisp(class_reg, offset_of_type, class_reg);
+    LoadRefDisp(class_reg, offset_of_type, class_reg, kNotVolatile);
     if (!cu_->compiler_driver->CanAssumeTypeIsPresentInDexCache(*cu_->dex_file, type_idx)) {
       // Need to test presence of type in dex cache at runtime
       LIR* hop_branch = OpCmpImmBranch(kCondEq, class_reg, 0, NULL);
@@ -1405,7 +1415,7 @@
 
       if (load_) {
         m2l_->LoadRefDisp(m2l_->TargetReg(kArg0), mirror::Object::ClassOffset().Int32Value(),
-                          m2l_->TargetReg(kArg1));
+                          m2l_->TargetReg(kArg1), kNotVolatile);
       }
       if (m2l_->cu_->target64) {
         m2l_->CallRuntimeHelperRegReg(QUICK_ENTRYPOINT_OFFSET(8, pCheckCast), m2l_->TargetReg(kArg2),
@@ -1436,7 +1446,8 @@
     LIR* branch1 = OpCmpImmBranch(kCondEq, TargetReg(kArg0), 0, NULL);
     /* load object->klass_ */
     DCHECK_EQ(mirror::Object::ClassOffset().Int32Value(), 0);
-    LoadRefDisp(TargetReg(kArg0), mirror::Object::ClassOffset().Int32Value(), TargetReg(kArg1));
+    LoadRefDisp(TargetReg(kArg0), mirror::Object::ClassOffset().Int32Value(), TargetReg(kArg1),
+                kNotVolatile);
 
     LIR* branch2 = OpCmpBranch(kCondNe, TargetReg(kArg1), class_reg, NULL);
     LIR* cont = NewLIR0(kPseudoTargetLabel);
diff --git a/compiler/dex/quick/gen_invoke.cc b/compiler/dex/quick/gen_invoke.cc
index 638c590..008ebfb 100644
--- a/compiler/dex/quick/gen_invoke.cc
+++ b/compiler/dex/quick/gen_invoke.cc
@@ -501,7 +501,7 @@
   StoreValue(rl_method, rl_src);
   // If Method* has been promoted, explicitly flush
   if (rl_method.location == kLocPhysReg) {
-    StoreRefDisp(TargetReg(kSp), 0, TargetReg(kArg0));
+    StoreRefDisp(TargetReg(kSp), 0, TargetReg(kArg0), kNotVolatile);
   }
 
   if (cu_->num_ins == 0) {
@@ -616,7 +616,8 @@
     case 1:  // Get method->dex_cache_resolved_methods_
       cg->LoadRefDisp(cg->TargetReg(kArg0),
                       mirror::ArtMethod::DexCacheResolvedMethodsOffset().Int32Value(),
-                      cg->TargetReg(kArg0));
+                      cg->TargetReg(kArg0),
+                      kNotVolatile);
       // Set up direct code if known.
       if (direct_code != 0) {
         if (direct_code != static_cast<uintptr_t>(-1)) {
@@ -631,7 +632,8 @@
       CHECK_EQ(cu->dex_file, target_method.dex_file);
       cg->LoadRefDisp(cg->TargetReg(kArg0),
                       ObjArray::OffsetOfElement(target_method.dex_method_index).Int32Value(),
-                      cg->TargetReg(kArg0));
+                      cg->TargetReg(kArg0),
+                      kNotVolatile);
       break;
     case 3:  // Grab the code from the method*
       if (cu->instruction_set != kX86 && cu->instruction_set != kX86_64) {
@@ -676,17 +678,20 @@
       cg->GenNullCheck(cg->TargetReg(kArg1), info->opt_flags);
       // get this->klass_ [use kArg1, set kInvokeTgt]
       cg->LoadRefDisp(cg->TargetReg(kArg1), mirror::Object::ClassOffset().Int32Value(),
-                      cg->TargetReg(kInvokeTgt));
+                      cg->TargetReg(kInvokeTgt),
+                      kNotVolatile);
       cg->MarkPossibleNullPointerException(info->opt_flags);
       break;
     case 2:  // Get this->klass_->vtable [usr kInvokeTgt, set kInvokeTgt]
       cg->LoadRefDisp(cg->TargetReg(kInvokeTgt), mirror::Class::VTableOffset().Int32Value(),
-                      cg->TargetReg(kInvokeTgt));
+                      cg->TargetReg(kInvokeTgt),
+                      kNotVolatile);
       break;
     case 3:  // Get target method [use kInvokeTgt, set kArg0]
       cg->LoadRefDisp(cg->TargetReg(kInvokeTgt),
                       ObjArray::OffsetOfElement(method_idx).Int32Value(),
-                      cg->TargetReg(kArg0));
+                      cg->TargetReg(kArg0),
+                      kNotVolatile);
       break;
     case 4:  // Get the compiled code address [uses kArg0, sets kInvokeTgt]
       if (cu->instruction_set != kX86 && cu->instruction_set != kX86_64) {
@@ -731,19 +736,22 @@
       cg->GenNullCheck(cg->TargetReg(kArg1), info->opt_flags);
       // Get this->klass_ [use kArg1, set kInvokeTgt]
       cg->LoadRefDisp(cg->TargetReg(kArg1), mirror::Object::ClassOffset().Int32Value(),
-                      cg->TargetReg(kInvokeTgt));
+                      cg->TargetReg(kInvokeTgt),
+                      kNotVolatile);
       cg->MarkPossibleNullPointerException(info->opt_flags);
       break;
     case 3:  // Get this->klass_->imtable [use kInvokeTgt, set kInvokeTgt]
       // NOTE: native pointer.
       cg->LoadRefDisp(cg->TargetReg(kInvokeTgt), mirror::Class::ImTableOffset().Int32Value(),
-                      cg->TargetReg(kInvokeTgt));
+                      cg->TargetReg(kInvokeTgt),
+                      kNotVolatile);
       break;
     case 4:  // Get target method [use kInvokeTgt, set kArg0]
       // NOTE: native pointer.
       cg->LoadRefDisp(cg->TargetReg(kInvokeTgt),
                        ObjArray::OffsetOfElement(method_idx % ClassLinker::kImtSize).Int32Value(),
-                       cg->TargetReg(kArg0));
+                       cg->TargetReg(kArg0),
+                       kNotVolatile);
       break;
     case 5:  // Get the compiled code address [use kArg0, set kInvokeTgt]
       if (cu->instruction_set != kX86 && cu->instruction_set != kX86_64) {
@@ -967,7 +975,7 @@
       {
         ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg);
         if (rl_arg.wide) {
-          StoreBaseDisp(TargetReg(kSp), outs_offset, arg_reg, k64);
+          StoreBaseDisp(TargetReg(kSp), outs_offset, arg_reg, k64, kNotVolatile);
           next_use += 2;
         } else {
           Store32Disp(TargetReg(kSp), outs_offset, arg_reg);
@@ -1037,7 +1045,7 @@
       loc = UpdateLocWide(loc);
       if ((next_arg >= 2) && (loc.location == kLocPhysReg)) {
         ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg);
-        StoreBaseDisp(TargetReg(kSp), SRegOffset(loc.s_reg_low), loc.reg, k64);
+        StoreBaseDisp(TargetReg(kSp), SRegOffset(loc.s_reg_low), loc.reg, k64, kNotVolatile);
       }
       next_arg += 2;
     } else {
@@ -1307,7 +1315,7 @@
     reg_off = AllocTemp();
     reg_ptr = AllocTempRef();
     Load32Disp(rl_obj.reg, offset_offset, reg_off);
-    LoadRefDisp(rl_obj.reg, value_offset, reg_ptr);
+    LoadRefDisp(rl_obj.reg, value_offset, reg_ptr, kNotVolatile);
   }
   if (rl_idx.is_const) {
     OpRegImm(kOpAdd, reg_off, mir_graph_->ConstantValue(rl_idx.orig_sreg));
@@ -1672,7 +1680,7 @@
     } else {
       RegStorage rl_temp_offset = AllocTemp();
       OpRegRegReg(kOpAdd, rl_temp_offset, rl_object.reg, rl_offset.reg);
-      LoadBaseDisp(rl_temp_offset, 0, rl_result.reg, k64);
+      LoadBaseDisp(rl_temp_offset, 0, rl_result.reg, k64, kNotVolatile);
       FreeTemp(rl_temp_offset);
     }
   } else {
@@ -1719,7 +1727,7 @@
     } else {
       RegStorage rl_temp_offset = AllocTemp();
       OpRegRegReg(kOpAdd, rl_temp_offset, rl_object.reg, rl_offset.reg);
-      StoreBaseDisp(rl_temp_offset, 0, rl_value.reg, k64);
+      StoreBaseDisp(rl_temp_offset, 0, rl_value.reg, k64, kNotVolatile);
       FreeTemp(rl_temp_offset);
     }
   } else {
diff --git a/compiler/dex/quick/gen_loadstore.cc b/compiler/dex/quick/gen_loadstore.cc
index d6f6ea1..bfb77fc 100644
--- a/compiler/dex/quick/gen_loadstore.cc
+++ b/compiler/dex/quick/gen_loadstore.cc
@@ -66,7 +66,7 @@
       } else {
         // Lives in the frame, need to store.
         ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg);
-        StoreBaseDisp(TargetReg(kSp), SRegOffset(rl_dest.s_reg_low), temp_reg, k32);
+        StoreBaseDisp(TargetReg(kSp), SRegOffset(rl_dest.s_reg_low), temp_reg, k32, kNotVolatile);
       }
       if (!zero_reg.Valid()) {
         FreeTemp(temp_reg);
@@ -93,7 +93,7 @@
            (rl_src.location == kLocCompilerTemp));
     ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg);
     if (rl_src.ref) {
-      LoadRefDisp(TargetReg(kSp), SRegOffset(rl_src.s_reg_low), r_dest);
+      LoadRefDisp(TargetReg(kSp), SRegOffset(rl_src.s_reg_low), r_dest, kNotVolatile);
     } else {
       Load32Disp(TargetReg(kSp), SRegOffset(rl_src.s_reg_low), r_dest);
     }
@@ -126,7 +126,7 @@
     DCHECK((rl_src.location == kLocDalvikFrame) ||
            (rl_src.location == kLocCompilerTemp));
     ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg);
-    LoadBaseDisp(TargetReg(kSp), SRegOffset(rl_src.s_reg_low), r_dest, k64);
+    LoadBaseDisp(TargetReg(kSp), SRegOffset(rl_src.s_reg_low), r_dest, k64, kNotVolatile);
   }
 }
 
@@ -215,7 +215,7 @@
     def_start = last_lir_insn_;
     ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg);
     if (rl_dest.ref) {
-      StoreRefDisp(TargetReg(kSp), SRegOffset(rl_dest.s_reg_low), rl_dest.reg);
+      StoreRefDisp(TargetReg(kSp), SRegOffset(rl_dest.s_reg_low), rl_dest.reg, kNotVolatile);
     } else {
       Store32Disp(TargetReg(kSp), SRegOffset(rl_dest.s_reg_low), rl_dest.reg);
     }
@@ -305,7 +305,7 @@
     DCHECK_EQ((mir_graph_->SRegToVReg(rl_dest.s_reg_low)+1),
               mir_graph_->SRegToVReg(GetSRegHi(rl_dest.s_reg_low)));
     ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg);
-    StoreBaseDisp(TargetReg(kSp), SRegOffset(rl_dest.s_reg_low), rl_dest.reg, k64);
+    StoreBaseDisp(TargetReg(kSp), SRegOffset(rl_dest.s_reg_low), rl_dest.reg, k64, kNotVolatile);
     MarkClean(rl_dest);
     def_end = last_lir_insn_;
     MarkDefWide(rl_dest, def_start, def_end);
@@ -369,7 +369,7 @@
     DCHECK_EQ((mir_graph_->SRegToVReg(rl_dest.s_reg_low)+1),
               mir_graph_->SRegToVReg(GetSRegHi(rl_dest.s_reg_low)));
     ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg);
-    StoreBaseDisp(TargetReg(kSp), SRegOffset(rl_dest.s_reg_low), rl_dest.reg, k64);
+    StoreBaseDisp(TargetReg(kSp), SRegOffset(rl_dest.s_reg_low), rl_dest.reg, k64, kNotVolatile);
     MarkClean(rl_dest);
     LIR *def_end = last_lir_insn_;
     MarkDefWide(rl_dest, def_start, def_end);
diff --git a/compiler/dex/quick/mips/call_mips.cc b/compiler/dex/quick/mips/call_mips.cc
index e53105f..26ea6a8 100644
--- a/compiler/dex/quick/mips/call_mips.cc
+++ b/compiler/dex/quick/mips/call_mips.cc
@@ -264,9 +264,9 @@
   int ex_offset = Thread::ExceptionOffset<4>().Int32Value();
   RegLocation rl_result = EvalLoc(rl_dest, kRefReg, true);
   RegStorage reset_reg = AllocTempRef();
-  LoadRefDisp(rs_rMIPS_SELF, ex_offset, rl_result.reg);
+  LoadRefDisp(rs_rMIPS_SELF, ex_offset, rl_result.reg, kNotVolatile);
   LoadConstant(reset_reg, 0);
-  StoreRefDisp(rs_rMIPS_SELF, ex_offset, reset_reg);
+  StoreRefDisp(rs_rMIPS_SELF, ex_offset, reset_reg, kNotVolatile);
   FreeTemp(reset_reg);
   StoreValue(rl_dest, rl_result);
 }
diff --git a/compiler/dex/quick/mips/codegen_mips.h b/compiler/dex/quick/mips/codegen_mips.h
index 571adac..c0ad916 100644
--- a/compiler/dex/quick/mips/codegen_mips.h
+++ b/compiler/dex/quick/mips/codegen_mips.h
@@ -33,20 +33,16 @@
     LIR* CheckSuspendUsingLoad() OVERRIDE;
     RegStorage LoadHelper(ThreadOffset<4> offset) OVERRIDE;
     RegStorage LoadHelper(ThreadOffset<8> offset) OVERRIDE;
-    LIR* LoadBaseDispVolatile(RegStorage r_base, int displacement, RegStorage r_dest,
-                              OpSize size) OVERRIDE;
     LIR* LoadBaseDisp(RegStorage r_base, int displacement, RegStorage r_dest,
-                      OpSize size) OVERRIDE;
+                      OpSize size, VolatileKind is_volatile) OVERRIDE;
     LIR* LoadBaseIndexed(RegStorage r_base, RegStorage r_index, RegStorage r_dest, int scale,
                          OpSize size) OVERRIDE;
     LIR* LoadBaseIndexedDisp(RegStorage r_base, RegStorage r_index, int scale, int displacement,
                              RegStorage r_dest, OpSize size) OVERRIDE;
     LIR* LoadConstantNoClobber(RegStorage r_dest, int value);
     LIR* LoadConstantWide(RegStorage r_dest, int64_t value);
-    LIR* StoreBaseDispVolatile(RegStorage r_base, int displacement, RegStorage r_src,
-                               OpSize size) OVERRIDE;
     LIR* StoreBaseDisp(RegStorage r_base, int displacement, RegStorage r_src,
-                       OpSize size) OVERRIDE;
+                       OpSize size, VolatileKind is_volatile) OVERRIDE;
     LIR* StoreBaseIndexed(RegStorage r_base, RegStorage r_index, RegStorage r_src, int scale,
                           OpSize size) OVERRIDE;
     LIR* StoreBaseIndexedDisp(RegStorage r_base, RegStorage r_index, int scale, int displacement,
diff --git a/compiler/dex/quick/mips/int_mips.cc b/compiler/dex/quick/mips/int_mips.cc
index beaf6bb..903a770 100644
--- a/compiler/dex/quick/mips/int_mips.cc
+++ b/compiler/dex/quick/mips/int_mips.cc
@@ -294,7 +294,7 @@
   RegLocation rl_address = LoadValue(rl_src_address, kCoreReg);
   RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true);
   DCHECK(size == kSignedByte);
-  LoadBaseDisp(rl_address.reg, 0, rl_result.reg, size);
+  LoadBaseDisp(rl_address.reg, 0, rl_result.reg, size, kNotVolatile);
   StoreValue(rl_dest, rl_result);
   return true;
 }
@@ -310,7 +310,7 @@
   RegLocation rl_address = LoadValue(rl_src_address, kCoreReg);
   DCHECK(size == kSignedByte);
   RegLocation rl_value = LoadValue(rl_src_value, kCoreReg);
-  StoreBaseDisp(rl_address.reg, 0, rl_value.reg, size);
+  StoreBaseDisp(rl_address.reg, 0, rl_value.reg, size, kNotVolatile);
   return true;
 }
 
@@ -524,7 +524,7 @@
       GenArrayBoundsCheck(rl_index.reg, reg_len);
       FreeTemp(reg_len);
     }
-    LoadBaseDisp(reg_ptr, 0, rl_result.reg, size);
+    LoadBaseDisp(reg_ptr, 0, rl_result.reg, size, kNotVolatile);
 
     FreeTemp(reg_ptr);
     StoreValueWide(rl_dest, rl_result);
@@ -602,7 +602,7 @@
       FreeTemp(reg_len);
     }
 
-    StoreBaseDisp(reg_ptr, 0, rl_src.reg, size);
+    StoreBaseDisp(reg_ptr, 0, rl_src.reg, size, kNotVolatile);
   } else {
     rl_src = LoadValue(rl_src, reg_class);
     if (needs_range_check) {
diff --git a/compiler/dex/quick/mips/utility_mips.cc b/compiler/dex/quick/mips/utility_mips.cc
index 01b25f92..b49f436 100644
--- a/compiler/dex/quick/mips/utility_mips.cc
+++ b/compiler/dex/quick/mips/utility_mips.cc
@@ -546,23 +546,31 @@
   return load;
 }
 
-LIR* MipsMir2Lir::LoadBaseDispVolatile(RegStorage r_base, int displacement, RegStorage r_dest,
-                                       OpSize size) {
-  DCHECK(size != k64 && size != kDouble);
-  return LoadBaseDisp(r_base, displacement, r_dest, size);
-}
-
 LIR* MipsMir2Lir::LoadBaseDisp(RegStorage r_base, int displacement, RegStorage r_dest,
-                               OpSize size) {
+                               OpSize size, VolatileKind is_volatile) {
+  if (is_volatile == kVolatile) {
+    DCHECK(size != k64 && size != kDouble);
+  }
+
   // TODO: base this on target.
   if (size == kWord) {
     size = k32;
   }
+  LIR* load;
   if (size == k64 || size == kDouble) {
-    return LoadBaseDispBody(r_base, displacement, r_dest.GetLow(), r_dest.GetHigh(), size);
+    load = LoadBaseDispBody(r_base, displacement, r_dest.GetLow(), r_dest.GetHigh(), size);
   } else {
-    return LoadBaseDispBody(r_base, displacement, r_dest, RegStorage::InvalidReg(), size);
+    load = LoadBaseDispBody(r_base, displacement, r_dest, RegStorage::InvalidReg(), size);
   }
+
+  if (UNLIKELY(is_volatile == kVolatile)) {
+    // Without context sensitive analysis, we must issue the most conservative barriers.
+    // In this case, either a load or store may follow so we issue both barriers.
+    GenMemBarrier(kLoadLoad);
+    GenMemBarrier(kLoadStore);
+  }
+
+  return load;
 }
 
 // FIXME: don't split r_dest into 2 containers.
@@ -648,23 +656,31 @@
   return res;
 }
 
-LIR* MipsMir2Lir::StoreBaseDispVolatile(RegStorage r_base, int displacement, RegStorage r_src,
-                                        OpSize size) {
-  DCHECK(size != k64 && size != kDouble);
-  return StoreBaseDisp(r_base, displacement, r_src, size);
-}
-
 LIR* MipsMir2Lir::StoreBaseDisp(RegStorage r_base, int displacement, RegStorage r_src,
-                                OpSize size) {
+                                OpSize size, VolatileKind is_volatile) {
+  if (is_volatile == kVolatile) {
+    DCHECK(size != k64 && size != kDouble);
+    // There might have been a store before this volatile one so insert StoreStore barrier.
+    GenMemBarrier(kStoreStore);
+  }
+
   // TODO: base this on target.
   if (size == kWord) {
     size = k32;
   }
+  LIR* store;
   if (size == k64 || size == kDouble) {
-    return StoreBaseDispBody(r_base, displacement, r_src.GetLow(), r_src.GetHigh(), size);
+    store = StoreBaseDispBody(r_base, displacement, r_src.GetLow(), r_src.GetHigh(), size);
   } else {
-    return StoreBaseDispBody(r_base, displacement, r_src, RegStorage::InvalidReg(), size);
+    store = StoreBaseDispBody(r_base, displacement, r_src, RegStorage::InvalidReg(), size);
   }
+
+  if (UNLIKELY(is_volatile == kVolatile)) {
+    // A load might follow the volatile store so insert a StoreLoad barrier.
+    GenMemBarrier(kStoreLoad);
+  }
+
+  return store;
 }
 
 LIR* MipsMir2Lir::OpThreadMem(OpKind op, ThreadOffset<4> thread_offset) {
diff --git a/compiler/dex/quick/mir_to_lir.cc b/compiler/dex/quick/mir_to_lir.cc
index 1fc4163..5d68187 100644
--- a/compiler/dex/quick/mir_to_lir.cc
+++ b/compiler/dex/quick/mir_to_lir.cc
@@ -92,7 +92,7 @@
     if (!reg_arg.Valid()) {
       RegStorage new_reg =
           wide ?  AllocTypedTempWide(false, reg_class) : AllocTypedTemp(false, reg_class);
-      LoadBaseDisp(TargetReg(kSp), offset, new_reg, wide ? k64 : k32);
+      LoadBaseDisp(TargetReg(kSp), offset, new_reg, wide ? k64 : k32, kNotVolatile);
       return new_reg;
     } else {
       // Check if we need to copy the arg to a different reg_class.
@@ -120,7 +120,7 @@
     // If the low part is not in a reg, we allocate a pair. Otherwise, we just load to high reg.
     if (!reg_arg_low.Valid()) {
       RegStorage new_regs = AllocTypedTempWide(false, reg_class);
-      LoadBaseDisp(TargetReg(kSp), offset, new_regs, k64);
+      LoadBaseDisp(TargetReg(kSp), offset, new_regs, k64, kNotVolatile);
       return new_regs;  // The reg_class is OK, we can return.
     } else {
       // Assume that no ABI allows splitting a wide fp reg between a narrow fp reg and memory,
@@ -193,7 +193,7 @@
       if (reg.Valid()) {
         OpRegCopy(rl_dest.reg, reg);
       } else {
-        LoadBaseDisp(TargetReg(kSp), offset, rl_dest.reg, k64);
+        LoadBaseDisp(TargetReg(kSp), offset, rl_dest.reg, k64, kNotVolatile);
       }
       return;
     }
@@ -211,7 +211,7 @@
       OpRegCopy(rl_dest.reg.GetHigh(), reg_arg_high);
       Load32Disp(TargetReg(kSp), offset, rl_dest.reg.GetLow());
     } else {
-      LoadBaseDisp(TargetReg(kSp), offset, rl_dest.reg, k64);
+      LoadBaseDisp(TargetReg(kSp), offset, rl_dest.reg, k64, kNotVolatile);
     }
   }
 }
@@ -243,14 +243,11 @@
     r_result = wide ? AllocTypedTempWide(rl_dest.fp, reg_class)
                     : AllocTypedTemp(rl_dest.fp, reg_class);
   }
-  if (data.is_volatile) {
-    LoadBaseDispVolatile(reg_obj, data.field_offset, r_result, size);
-    // Without context sensitive analysis, we must issue the most conservative barriers.
-    // In this case, either a load or store may follow so we issue both barriers.
-    GenMemBarrier(kLoadLoad);
-    GenMemBarrier(kLoadStore);
+  if (ref) {
+    LoadRefDisp(reg_obj, data.field_offset, r_result, data.is_volatile ? kVolatile : kNotVolatile);
   } else {
-    LoadBaseDisp(reg_obj, data.field_offset, r_result, size);
+    LoadBaseDisp(reg_obj, data.field_offset, r_result, size, data.is_volatile ? kVolatile :
+        kNotVolatile);
   }
   if (r_result != rl_dest.reg) {
     if (wide) {
@@ -288,14 +285,11 @@
   RegStorage reg_obj = LoadArg(data.object_arg, kRefReg);
   RegisterClass reg_class = RegClassForFieldLoadStore(size, data.is_volatile);
   RegStorage reg_src = LoadArg(data.src_arg, reg_class, wide);
-  if (data.is_volatile) {
-    // There might have been a store before this volatile one so insert StoreStore barrier.
-    GenMemBarrier(kStoreStore);
-    StoreBaseDispVolatile(reg_obj, data.field_offset, reg_src, size);
-    // A load might follow the volatile store so insert a StoreLoad barrier.
-    GenMemBarrier(kStoreLoad);
+  if (ref) {
+    StoreRefDisp(reg_obj, data.field_offset, reg_src, data.is_volatile ? kVolatile : kNotVolatile);
   } else {
-    StoreBaseDisp(reg_obj, data.field_offset, reg_src, size);
+    StoreBaseDisp(reg_obj, data.field_offset, reg_src, size, data.is_volatile ? kVolatile :
+        kNotVolatile);
   }
   if (ref) {
     MarkGCCard(reg_src, reg_obj);
diff --git a/compiler/dex/quick/mir_to_lir.h b/compiler/dex/quick/mir_to_lir.h
index f70087d..b07c85e 100644
--- a/compiler/dex/quick/mir_to_lir.h
+++ b/compiler/dex/quick/mir_to_lir.h
@@ -663,6 +663,7 @@
     virtual void Materialize();
     virtual CompiledMethod* GetCompiledMethod();
     void MarkSafepointPC(LIR* inst);
+    void MarkSafepointPCAfter(LIR* after);
     void SetupResourceMasks(LIR* lir);
     void SetMemRefType(LIR* lir, bool is_load, int mem_type);
     void AnnotateDalvikRegAccess(LIR* lir, int reg_id, bool is_load, bool is64bit);
@@ -830,6 +831,7 @@
     void GenArrayBoundsCheck(int32_t index, RegStorage length);
     LIR* GenNullCheck(RegStorage reg);
     void MarkPossibleNullPointerException(int opt_flags);
+    void MarkPossibleNullPointerExceptionAfter(int opt_flags, LIR* after);
     void MarkPossibleStackOverflowException();
     void ForceImplicitNullCheck(RegStorage reg, int opt_flags);
     LIR* GenImmedCheck(ConditionCode c_code, RegStorage reg, int imm_val, ThrowKind kind);
@@ -1007,15 +1009,20 @@
     virtual LIR* LoadConstant(RegStorage r_dest, int value);
     // Natural word size.
     virtual LIR* LoadWordDisp(RegStorage r_base, int displacement, RegStorage r_dest) {
-      return LoadBaseDisp(r_base, displacement, r_dest, kWord);
+      return LoadBaseDisp(r_base, displacement, r_dest, kWord, kNotVolatile);
     }
     // Load 32 bits, regardless of target.
     virtual LIR* Load32Disp(RegStorage r_base, int displacement, RegStorage r_dest)  {
-      return LoadBaseDisp(r_base, displacement, r_dest, k32);
+      return LoadBaseDisp(r_base, displacement, r_dest, k32, kNotVolatile);
     }
     // Load a reference at base + displacement and decompress into register.
-    virtual LIR* LoadRefDisp(RegStorage r_base, int displacement, RegStorage r_dest) {
-      return LoadBaseDisp(r_base, displacement, r_dest, kReference);
+    virtual LIR* LoadRefDisp(RegStorage r_base, int displacement, RegStorage r_dest,
+                             VolatileKind is_volatile) {
+      return LoadBaseDisp(r_base, displacement, r_dest, kReference, is_volatile);
+    }
+    // Load a reference at base + index and decompress into register.
+    virtual LIR* LoadRefIndexed(RegStorage r_base, RegStorage r_index, RegStorage r_dest) {
+      return LoadBaseIndexed(r_base, r_index, r_dest, 2, kReference);
     }
     // Load Dalvik value with 32-bit memory storage.  If compressed object reference, decompress.
     virtual RegLocation LoadValue(RegLocation rl_src, RegisterClass op_kind);
@@ -1033,15 +1040,20 @@
     virtual void LoadValueDirectWideFixed(RegLocation rl_src, RegStorage r_dest);
     // Store an item of natural word size.
     virtual LIR* StoreWordDisp(RegStorage r_base, int displacement, RegStorage r_src) {
-      return StoreBaseDisp(r_base, displacement, r_src, kWord);
+      return StoreBaseDisp(r_base, displacement, r_src, kWord, kNotVolatile);
     }
     // Store an uncompressed reference into a compressed 32-bit container.
-    virtual LIR* StoreRefDisp(RegStorage r_base, int displacement, RegStorage r_src) {
-      return StoreBaseDisp(r_base, displacement, r_src, kReference);
+    virtual LIR* StoreRefDisp(RegStorage r_base, int displacement, RegStorage r_src,
+                              VolatileKind is_volatile) {
+      return StoreBaseDisp(r_base, displacement, r_src, kReference, is_volatile);
+    }
+    // Store an uncompressed reference into a compressed 32-bit container by index.
+    virtual LIR* StoreRefIndexed(RegStorage r_base, RegStorage r_index, RegStorage r_src) {
+      return StoreBaseIndexed(r_base, r_index, r_src, 2, kReference);
     }
     // Store 32 bits, regardless of target.
     virtual LIR* Store32Disp(RegStorage r_base, int displacement, RegStorage r_src) {
-      return StoreBaseDisp(r_base, displacement, r_src, k32);
+      return StoreBaseDisp(r_base, displacement, r_src, k32, kNotVolatile);
     }
 
     /**
@@ -1144,20 +1156,16 @@
     virtual RegStorage LoadHelper(ThreadOffset<4> offset) = 0;
     virtual RegStorage LoadHelper(ThreadOffset<8> offset) = 0;
 
-    virtual LIR* LoadBaseDispVolatile(RegStorage r_base, int displacement, RegStorage r_dest,
-                                      OpSize size) = 0;
     virtual LIR* LoadBaseDisp(RegStorage r_base, int displacement, RegStorage r_dest,
-                              OpSize size) = 0;
+                              OpSize size, VolatileKind is_volatile) = 0;
     virtual LIR* LoadBaseIndexed(RegStorage r_base, RegStorage r_index, RegStorage r_dest,
                                  int scale, OpSize size) = 0;
     virtual LIR* LoadBaseIndexedDisp(RegStorage r_base, RegStorage r_index, int scale,
                                      int displacement, RegStorage r_dest, OpSize size) = 0;
     virtual LIR* LoadConstantNoClobber(RegStorage r_dest, int value) = 0;
     virtual LIR* LoadConstantWide(RegStorage r_dest, int64_t value) = 0;
-    virtual LIR* StoreBaseDispVolatile(RegStorage r_base, int displacement, RegStorage r_src,
-                                       OpSize size) = 0;
     virtual LIR* StoreBaseDisp(RegStorage r_base, int displacement, RegStorage r_src,
-                               OpSize size) = 0;
+                               OpSize size, VolatileKind is_volatile) = 0;
     virtual LIR* StoreBaseIndexed(RegStorage r_base, RegStorage r_index, RegStorage r_src,
                                   int scale, OpSize size) = 0;
     virtual LIR* StoreBaseIndexedDisp(RegStorage r_base, RegStorage r_index, int scale,
diff --git a/compiler/dex/quick/ralloc_util.cc b/compiler/dex/quick/ralloc_util.cc
index 5bb0ee0..60eebe4 100644
--- a/compiler/dex/quick/ralloc_util.cc
+++ b/compiler/dex/quick/ralloc_util.cc
@@ -735,7 +735,7 @@
       }
       int v_reg = mir_graph_->SRegToVReg(info1->SReg());
       ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg);
-      StoreBaseDisp(TargetReg(kSp), VRegOffset(v_reg), reg, k64);
+      StoreBaseDisp(TargetReg(kSp), VRegOffset(v_reg), reg, k64, kNotVolatile);
     }
   } else {
     RegisterInfo* info = GetRegInfo(reg);
@@ -743,7 +743,7 @@
       info->SetIsDirty(false);
       int v_reg = mir_graph_->SRegToVReg(info->SReg());
       ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg);
-      StoreBaseDisp(TargetReg(kSp), VRegOffset(v_reg), reg, k64);
+      StoreBaseDisp(TargetReg(kSp), VRegOffset(v_reg), reg, k64, kNotVolatile);
     }
   }
 }
@@ -755,7 +755,7 @@
     info->SetIsDirty(false);
     int v_reg = mir_graph_->SRegToVReg(info->SReg());
     ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg);
-    StoreBaseDisp(TargetReg(kSp), VRegOffset(v_reg), reg, kWord);
+    StoreBaseDisp(TargetReg(kSp), VRegOffset(v_reg), reg, kWord, kNotVolatile);
   }
 }
 
diff --git a/compiler/dex/quick/x86/call_x86.cc b/compiler/dex/quick/x86/call_x86.cc
index 28195ab..425caec 100644
--- a/compiler/dex/quick/x86/call_x86.cc
+++ b/compiler/dex/quick/x86/call_x86.cc
@@ -295,7 +295,8 @@
     setup_method_address_[0] = NewLIR1(kX86StartOfMethod, rs_rX86_ARG0.GetReg());
     int displacement = SRegOffset(base_of_code_->s_reg_low);
     // Native pointer - must be natural word size.
-    setup_method_address_[1] = StoreBaseDisp(rs_rX86_SP, displacement, rs_rX86_ARG0, Gen64Bit() ? k64 : k32);
+    setup_method_address_[1] = StoreBaseDisp(rs_rX86_SP, displacement, rs_rX86_ARG0,
+                                             Gen64Bit() ? k64 : k32, kNotVolatile);
   }
 
   FreeTemp(rs_rX86_ARG0);
diff --git a/compiler/dex/quick/x86/codegen_x86.h b/compiler/dex/quick/x86/codegen_x86.h
index d482e58..70382c7 100644
--- a/compiler/dex/quick/x86/codegen_x86.h
+++ b/compiler/dex/quick/x86/codegen_x86.h
@@ -68,20 +68,16 @@
   LIR* CheckSuspendUsingLoad() OVERRIDE;
   RegStorage LoadHelper(ThreadOffset<4> offset) OVERRIDE;
   RegStorage LoadHelper(ThreadOffset<8> offset) OVERRIDE;
-  LIR* LoadBaseDispVolatile(RegStorage r_base, int displacement, RegStorage r_dest,
-                            OpSize size) OVERRIDE;
   LIR* LoadBaseDisp(RegStorage r_base, int displacement, RegStorage r_dest,
-                    OpSize size) OVERRIDE;
+                    OpSize size, VolatileKind is_volatile) OVERRIDE;
   LIR* LoadBaseIndexed(RegStorage r_base, RegStorage r_index, RegStorage r_dest, int scale,
                        OpSize size) OVERRIDE;
   LIR* LoadBaseIndexedDisp(RegStorage r_base, RegStorage r_index, int scale, int displacement,
                            RegStorage r_dest, OpSize size) OVERRIDE;
   LIR* LoadConstantNoClobber(RegStorage r_dest, int value);
   LIR* LoadConstantWide(RegStorage r_dest, int64_t value);
-  LIR* StoreBaseDispVolatile(RegStorage r_base, int displacement, RegStorage r_src,
-                             OpSize size) OVERRIDE;
   LIR* StoreBaseDisp(RegStorage r_base, int displacement, RegStorage r_src,
-                     OpSize size) OVERRIDE;
+                     OpSize size, VolatileKind is_volatile) OVERRIDE;
   LIR* StoreBaseIndexed(RegStorage r_base, RegStorage r_index, RegStorage r_src, int scale,
                         OpSize size) OVERRIDE;
   LIR* StoreBaseIndexedDisp(RegStorage r_base, RegStorage r_index, int scale, int displacement,
diff --git a/compiler/dex/quick/x86/fp_x86.cc b/compiler/dex/quick/x86/fp_x86.cc
index 5082d60..f854adb 100644
--- a/compiler/dex/quick/x86/fp_x86.cc
+++ b/compiler/dex/quick/x86/fp_x86.cc
@@ -144,7 +144,7 @@
     } else {
       // It must have been register promoted if it is not a temp but is still in physical
       // register. Since we need it to be in memory to convert, we place it there now.
-      StoreBaseDisp(TargetReg(kSp), src_v_reg_offset, rl_src.reg, k64);
+      StoreBaseDisp(TargetReg(kSp), src_v_reg_offset, rl_src.reg, k64, kNotVolatile);
     }
   }
 
@@ -178,7 +178,7 @@
      */
     rl_result = EvalLoc(rl_dest, kFPReg, true);
     if (is_double) {
-      LoadBaseDisp(TargetReg(kSp), dest_v_reg_offset, rl_result.reg, k64);
+      LoadBaseDisp(TargetReg(kSp), dest_v_reg_offset, rl_result.reg, k64, kNotVolatile);
 
       StoreFinalValueWide(rl_dest, rl_result);
     } else {
@@ -363,7 +363,8 @@
     } else {
       // It must have been register promoted if it is not a temp but is still in physical
       // register. Since we need it to be in memory to convert, we place it there now.
-      StoreBaseDisp(TargetReg(kSp), src1_v_reg_offset, rl_src1.reg, is_double ? k64 : k32);
+      StoreBaseDisp(TargetReg(kSp), src1_v_reg_offset, rl_src1.reg, is_double ? k64 : k32,
+                    kNotVolatile);
     }
   }
 
@@ -373,7 +374,8 @@
       FlushSpecificReg(reg_info);
       ResetDef(rl_src2.reg);
     } else {
-      StoreBaseDisp(TargetReg(kSp), src2_v_reg_offset, rl_src2.reg, is_double ? k64 : k32);
+      StoreBaseDisp(TargetReg(kSp), src2_v_reg_offset, rl_src2.reg, is_double ? k64 : k32,
+                    kNotVolatile);
     }
   }
 
@@ -433,7 +435,7 @@
   if (rl_result.location == kLocPhysReg) {
     rl_result = EvalLoc(rl_dest, kFPReg, true);
     if (is_double) {
-      LoadBaseDisp(TargetReg(kSp), dest_v_reg_offset, rl_result.reg, k64);
+      LoadBaseDisp(TargetReg(kSp), dest_v_reg_offset, rl_result.reg, k64, kNotVolatile);
       StoreFinalValueWide(rl_dest, rl_result);
     } else {
       Load32Disp(TargetReg(kSp), dest_v_reg_offset, rl_result.reg);
diff --git a/compiler/dex/quick/x86/int_x86.cc b/compiler/dex/quick/x86/int_x86.cc
index 2f914c1..350cfb8 100644
--- a/compiler/dex/quick/x86/int_x86.cc
+++ b/compiler/dex/quick/x86/int_x86.cc
@@ -754,7 +754,7 @@
   RegLocation rl_address = LoadValue(rl_src_address, kCoreReg);
   RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true);
   // Unaligned access is allowed on x86.
-  LoadBaseDisp(rl_address.reg, 0, rl_result.reg, size);
+  LoadBaseDisp(rl_address.reg, 0, rl_result.reg, size, kNotVolatile);
   if (size == k64) {
     StoreValueWide(rl_dest, rl_result);
   } else {
@@ -772,12 +772,12 @@
   if (size == k64) {
     // Unaligned access is allowed on x86.
     RegLocation rl_value = LoadValueWide(rl_src_value, kCoreReg);
-    StoreBaseDisp(rl_address.reg, 0, rl_value.reg, size);
+    StoreBaseDisp(rl_address.reg, 0, rl_value.reg, size, kNotVolatile);
   } else {
     DCHECK(size == kSignedByte || size == kSignedHalf || size == k32);
     // Unaligned access is allowed on x86.
     RegLocation rl_value = LoadValue(rl_src_value, kCoreReg);
-    StoreBaseDisp(rl_address.reg, 0, rl_value.reg, size);
+    StoreBaseDisp(rl_address.reg, 0, rl_value.reg, size, kNotVolatile);
   }
   return true;
 }
@@ -1138,7 +1138,7 @@
       NewLIR2(kX86Xor32RR, dest.GetReg(), dest.GetReg());
       break;
     case 1:
-      LoadBaseDisp(rs_rX86_SP, displacement, dest, k32);
+      LoadBaseDisp(rs_rX86_SP, displacement, dest, k32, kNotVolatile);
       break;
     default:
       m = NewLIR4(IS_SIMM8(val) ? kX86Imul32RMI8 : kX86Imul32RMI, dest.GetReg(),
@@ -1294,7 +1294,8 @@
   if (src1_in_reg) {
     NewLIR2(kX86Mov32RR, rs_r1.GetReg(), rl_src1.reg.GetHighReg());
   } else {
-    LoadBaseDisp(rs_rX86_SP, SRegOffset(rl_src1.s_reg_low) + HIWORD_OFFSET, rs_r1, k32);
+    LoadBaseDisp(rs_rX86_SP, SRegOffset(rl_src1.s_reg_low) + HIWORD_OFFSET, rs_r1, k32,
+                 kNotVolatile);
   }
 
   if (is_square) {
@@ -1317,7 +1318,8 @@
     if (src2_in_reg) {
       NewLIR2(kX86Mov32RR, rs_r0.GetReg(), rl_src2.reg.GetHighReg());
     } else {
-      LoadBaseDisp(rs_rX86_SP, SRegOffset(rl_src2.s_reg_low) + HIWORD_OFFSET, rs_r0, k32);
+      LoadBaseDisp(rs_rX86_SP, SRegOffset(rl_src2.s_reg_low) + HIWORD_OFFSET, rs_r0, k32,
+                   kNotVolatile);
     }
 
     // EAX <- EAX * 1L  (2H * 1L)
@@ -1350,7 +1352,8 @@
   if (src2_in_reg) {
     NewLIR2(kX86Mov32RR, rs_r0.GetReg(), rl_src2.reg.GetLowReg());
   } else {
-    LoadBaseDisp(rs_rX86_SP, SRegOffset(rl_src2.s_reg_low) + LOWORD_OFFSET, rs_r0, k32);
+    LoadBaseDisp(rs_rX86_SP, SRegOffset(rl_src2.s_reg_low) + LOWORD_OFFSET, rs_r0, k32,
+                 kNotVolatile);
   }
 
   // EDX:EAX <- 2L * 1L (double precision)
@@ -2289,21 +2292,21 @@
   if (rl_method.location == kLocPhysReg) {
     if (use_declaring_class) {
       LoadRefDisp(rl_method.reg, mirror::ArtMethod::DeclaringClassOffset().Int32Value(),
-                   check_class);
+                  check_class, kNotVolatile);
     } else {
       LoadRefDisp(rl_method.reg, mirror::ArtMethod::DexCacheResolvedTypesOffset().Int32Value(),
-                   check_class);
-      LoadRefDisp(check_class, offset_of_type, check_class);
+                  check_class, kNotVolatile);
+      LoadRefDisp(check_class, offset_of_type, check_class, kNotVolatile);
     }
   } else {
     LoadCurrMethodDirect(check_class);
     if (use_declaring_class) {
       LoadRefDisp(check_class, mirror::ArtMethod::DeclaringClassOffset().Int32Value(),
-                   check_class);
+                  check_class, kNotVolatile);
     } else {
       LoadRefDisp(check_class, mirror::ArtMethod::DexCacheResolvedTypesOffset().Int32Value(),
-                   check_class);
-      LoadRefDisp(check_class, offset_of_type, check_class);
+                  check_class, kNotVolatile);
+      LoadRefDisp(check_class, offset_of_type, check_class, kNotVolatile);
     }
   }
 
@@ -2350,16 +2353,16 @@
   } else if (use_declaring_class) {
     LoadValueDirectFixed(rl_src, TargetReg(kArg0));
     LoadRefDisp(TargetReg(kArg1), mirror::ArtMethod::DeclaringClassOffset().Int32Value(),
-                 class_reg);
+                class_reg, kNotVolatile);
   } else {
     // Load dex cache entry into class_reg (kArg2).
     LoadValueDirectFixed(rl_src, TargetReg(kArg0));
     LoadRefDisp(TargetReg(kArg1), mirror::ArtMethod::DexCacheResolvedTypesOffset().Int32Value(),
-                 class_reg);
+                class_reg, kNotVolatile);
     int32_t offset_of_type =
         mirror::Array::DataOffset(sizeof(mirror::HeapReference<mirror::Class*>)).Int32Value() +
         (sizeof(mirror::HeapReference<mirror::Class*>) * type_idx);
-    LoadRefDisp(class_reg, offset_of_type, class_reg);
+    LoadRefDisp(class_reg, offset_of_type, class_reg, kNotVolatile);
     if (!can_assume_type_is_in_dex_cache) {
       // Need to test presence of type in dex cache at runtime.
       LIR* hop_branch = OpCmpImmBranch(kCondNe, class_reg, 0, NULL);
@@ -2392,7 +2395,8 @@
 
   /* Load object->klass_. */
   DCHECK_EQ(mirror::Object::ClassOffset().Int32Value(), 0);
-  LoadRefDisp(TargetReg(kArg0),  mirror::Object::ClassOffset().Int32Value(), TargetReg(kArg1));
+  LoadRefDisp(TargetReg(kArg0),  mirror::Object::ClassOffset().Int32Value(), TargetReg(kArg1),
+              kNotVolatile);
   /* kArg0 is ref, kArg1 is ref->klass_, kArg2 is class. */
   LIR* branchover = nullptr;
   if (type_known_final) {
diff --git a/compiler/dex/quick/x86/target_x86.cc b/compiler/dex/quick/x86/target_x86.cc
index 078dd5a..e369d26 100644
--- a/compiler/dex/quick/x86/target_x86.cc
+++ b/compiler/dex/quick/x86/target_x86.cc
@@ -1866,7 +1866,7 @@
   StoreValue(rl_method, rl_src);
   // If Method* has been promoted, explicitly flush
   if (rl_method.location == kLocPhysReg) {
-    StoreRefDisp(TargetReg(kSp), 0, TargetReg(kArg0));
+    StoreRefDisp(TargetReg(kSp), 0, TargetReg(kArg0), kNotVolatile);
   }
 
   if (cu_->num_ins == 0) {
@@ -1916,11 +1916,11 @@
       }
       if (need_flush) {
         if (t_loc->wide && t_loc->fp) {
-          StoreBaseDisp(TargetReg(kSp), SRegOffset(start_vreg + i), reg, k64);
+          StoreBaseDisp(TargetReg(kSp), SRegOffset(start_vreg + i), reg, k64, kNotVolatile);
           // Increment i to skip the next one
           i++;
         } else if (t_loc->wide && !t_loc->fp) {
-          StoreBaseDisp(TargetReg(kSp), SRegOffset(start_vreg + i), reg, k64);
+          StoreBaseDisp(TargetReg(kSp), SRegOffset(start_vreg + i), reg, k64, kNotVolatile);
           // Increment i to skip the next one
           i++;
         } else {
@@ -2018,14 +2018,14 @@
         loc = UpdateLocWide(loc);
         if (loc.location == kLocPhysReg) {
           ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg);
-          StoreBaseDisp(TargetReg(kSp), SRegOffset(loc.s_reg_low), loc.reg, k64);
+          StoreBaseDisp(TargetReg(kSp), SRegOffset(loc.s_reg_low), loc.reg, k64, kNotVolatile);
         }
         next_arg += 2;
       } else {
         loc = UpdateLoc(loc);
         if (loc.location == kLocPhysReg) {
           ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg);
-          StoreBaseDisp(TargetReg(kSp), SRegOffset(loc.s_reg_low), loc.reg, k32);
+          StoreBaseDisp(TargetReg(kSp), SRegOffset(loc.s_reg_low), loc.reg, k32, kNotVolatile);
         }
         next_arg++;
       }
@@ -2161,18 +2161,18 @@
           ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg);
           if (rl_arg.wide) {
             if (rl_arg.location == kLocPhysReg) {
-              StoreBaseDisp(TargetReg(kSp), out_offset, rl_arg.reg, k64);
+              StoreBaseDisp(TargetReg(kSp), out_offset, rl_arg.reg, k64, kNotVolatile);
             } else {
               LoadValueDirectWideFixed(rl_arg, regWide);
-              StoreBaseDisp(TargetReg(kSp), out_offset, regWide, k64);
+              StoreBaseDisp(TargetReg(kSp), out_offset, regWide, k64, kNotVolatile);
             }
             i++;
           } else {
             if (rl_arg.location == kLocPhysReg) {
-              StoreBaseDisp(TargetReg(kSp), out_offset, rl_arg.reg, k32);
+              StoreBaseDisp(TargetReg(kSp), out_offset, rl_arg.reg, k32, kNotVolatile);
             } else {
               LoadValueDirectFixed(rl_arg, regSingle);
-              StoreBaseDisp(TargetReg(kSp), out_offset, regSingle, k32);
+              StoreBaseDisp(TargetReg(kSp), out_offset, regSingle, k32, kNotVolatile);
             }
           }
         }
diff --git a/compiler/dex/quick/x86/utility_x86.cc b/compiler/dex/quick/x86/utility_x86.cc
index ac5162e..0352808 100644
--- a/compiler/dex/quick/x86/utility_x86.cc
+++ b/compiler/dex/quick/x86/utility_x86.cc
@@ -585,7 +585,7 @@
         // value.
         ScopedMemRefType mem_ref_type(this, ResourceMask::kLiteral);
         res = LoadBaseDisp(rl_method.reg, 256 /* bogus */, RegStorage::FloatSolo64(low_reg_val),
-                           kDouble);
+                           kDouble, kNotVolatile);
         res->target = data_target;
         res->flags.fixup = kFixupLoad;
         store_method_addr_used_ = true;
@@ -756,17 +756,22 @@
   return LoadBaseIndexedDisp(r_base, r_index, scale, 0, r_dest, size);
 }
 
-LIR* X86Mir2Lir::LoadBaseDispVolatile(RegStorage r_base, int displacement, RegStorage r_dest,
-                                      OpSize size) {
+LIR* X86Mir2Lir::LoadBaseDisp(RegStorage r_base, int displacement, RegStorage r_dest,
+                              OpSize size, VolatileKind is_volatile) {
   // LoadBaseDisp() will emit correct insn for atomic load on x86
   // assuming r_dest is correctly prepared using RegClassForFieldLoadStore().
-  return LoadBaseDisp(r_base, displacement, r_dest, size);
-}
 
-LIR* X86Mir2Lir::LoadBaseDisp(RegStorage r_base, int displacement, RegStorage r_dest,
-                              OpSize size) {
-  return LoadBaseIndexedDisp(r_base, RegStorage::InvalidReg(), 0, displacement, r_dest,
-                             size);
+  LIR* load = LoadBaseIndexedDisp(r_base, RegStorage::InvalidReg(), 0, displacement, r_dest,
+                                  size);
+
+  if (UNLIKELY(is_volatile == kVolatile)) {
+    // Without context sensitive analysis, we must issue the most conservative barriers.
+    // In this case, either a load or store may follow so we issue both barriers.
+    GenMemBarrier(kLoadLoad);
+    GenMemBarrier(kLoadStore);
+  }
+
+  return load;
 }
 
 LIR* X86Mir2Lir::StoreBaseIndexedDisp(RegStorage r_base, RegStorage r_index, int scale,
@@ -854,20 +859,28 @@
 
 /* store value base base + scaled index. */
 LIR* X86Mir2Lir::StoreBaseIndexed(RegStorage r_base, RegStorage r_index, RegStorage r_src,
-                      int scale, OpSize size) {
+                                  int scale, OpSize size) {
   return StoreBaseIndexedDisp(r_base, r_index, scale, 0, r_src, size);
 }
 
-LIR* X86Mir2Lir::StoreBaseDispVolatile(RegStorage r_base, int displacement,
-                                       RegStorage r_src, OpSize size) {
+LIR* X86Mir2Lir::StoreBaseDisp(RegStorage r_base, int displacement, RegStorage r_src, OpSize size,
+                               VolatileKind is_volatile) {
+  if (UNLIKELY(is_volatile == kVolatile)) {
+    // There might have been a store before this volatile one so insert StoreStore barrier.
+    GenMemBarrier(kStoreStore);
+  }
+
   // StoreBaseDisp() will emit correct insn for atomic store on x86
   // assuming r_dest is correctly prepared using RegClassForFieldLoadStore().
-  return StoreBaseDisp(r_base, displacement, r_src, size);
-}
 
-LIR* X86Mir2Lir::StoreBaseDisp(RegStorage r_base, int displacement,
-                               RegStorage r_src, OpSize size) {
-  return StoreBaseIndexedDisp(r_base, RegStorage::InvalidReg(), 0, displacement, r_src, size);
+  LIR* store = StoreBaseIndexedDisp(r_base, RegStorage::InvalidReg(), 0, displacement, r_src, size);
+
+  if (UNLIKELY(is_volatile == kVolatile)) {
+    // A load might follow the volatile store so insert a StoreLoad barrier.
+    GenMemBarrier(kStoreLoad);
+  }
+
+  return store;
 }
 
 LIR* X86Mir2Lir::OpCmpMemImmBranch(ConditionCode cond, RegStorage temp_reg, RegStorage base_reg,