Merge "MIPS64: Support short and long branches"
diff --git a/compiler/Android.mk b/compiler/Android.mk
index 42ddfd8..b05f479 100644
--- a/compiler/Android.mk
+++ b/compiler/Android.mk
@@ -219,7 +219,8 @@
   utils/mips/assembler_mips.h
 
 LIBART_COMPILER_ENUM_OPERATOR_OUT_HEADER_FILES_mips64 := \
-  $(LIBART_COMPILER_ENUM_OPERATOR_OUT_HEADER_FILES_mips)
+  $(LIBART_COMPILER_ENUM_OPERATOR_OUT_HEADER_FILES_mips) \
+  utils/mips64/assembler_mips64.h
 
 LIBART_COMPILER_ENUM_OPERATOR_OUT_HEADER_FILES_x86 :=
 LIBART_COMPILER_ENUM_OPERATOR_OUT_HEADER_FILES_x86_64 := \
diff --git a/compiler/optimizing/code_generator_mips64.cc b/compiler/optimizing/code_generator_mips64.cc
index 6100859..802c435 100644
--- a/compiler/optimizing/code_generator_mips64.cc
+++ b/compiler/optimizing/code_generator_mips64.cc
@@ -27,8 +27,8 @@
 #include "mirror/class-inl.h"
 #include "offsets.h"
 #include "thread.h"
-#include "utils/mips64/assembler_mips64.h"
 #include "utils/assembler.h"
+#include "utils/mips64/assembler_mips64.h"
 #include "utils/stack_checks.h"
 
 namespace art {
@@ -210,7 +210,7 @@
     }
 
     RestoreLiveRegisters(codegen, locations);
-    __ B(GetExitLabel());
+    __ Bc(GetExitLabel());
   }
 
   const char* GetDescription() const OVERRIDE { return "LoadClassSlowPathMIPS64"; }
@@ -257,7 +257,7 @@
                                  type);
 
     RestoreLiveRegisters(codegen, locations);
-    __ B(GetExitLabel());
+    __ Bc(GetExitLabel());
   }
 
   const char* GetDescription() const OVERRIDE { return "LoadStringSlowPathMIPS64"; }
@@ -312,13 +312,13 @@
     CheckEntrypointTypes<kQuickTestSuspend, void, void>();
     RestoreLiveRegisters(codegen, instruction_->GetLocations());
     if (successor_ == nullptr) {
-      __ B(GetReturnLabel());
+      __ Bc(GetReturnLabel());
     } else {
-      __ B(mips64_codegen->GetLabelOf(successor_));
+      __ Bc(mips64_codegen->GetLabelOf(successor_));
     }
   }
 
-  Label* GetReturnLabel() {
+  Mips64Label* GetReturnLabel() {
     DCHECK(successor_ == nullptr);
     return &return_label_;
   }
@@ -331,7 +331,7 @@
   HBasicBlock* const successor_;
 
   // If `successor_` is null, the label to branch to after the suspend check.
-  Label return_label_;
+  Mips64Label return_label_;
 
   DISALLOW_COPY_AND_ASSIGN(SuspendCheckSlowPathMIPS64);
 };
@@ -380,7 +380,7 @@
     }
 
     RestoreLiveRegisters(codegen, locations);
-    __ B(GetExitLabel());
+    __ Bc(GetExitLabel());
   }
 
   const char* GetDescription() const OVERRIDE { return "TypeCheckSlowPathMIPS64"; }
@@ -441,6 +441,32 @@
 #define QUICK_ENTRY_POINT(x) QUICK_ENTRYPOINT_OFFSET(kMips64WordSize, x).Int32Value()
 
 void CodeGeneratorMIPS64::Finalize(CodeAllocator* allocator) {
+  // Ensure that we fix up branches.
+  __ FinalizeCode();
+
+  // Adjust native pc offsets in stack maps.
+  for (size_t i = 0, num = stack_map_stream_.GetNumberOfStackMaps(); i != num; ++i) {
+    uint32_t old_position = stack_map_stream_.GetStackMap(i).native_pc_offset;
+    uint32_t new_position = __ GetAdjustedPosition(old_position);
+    DCHECK_GE(new_position, old_position);
+    stack_map_stream_.SetStackMapNativePcOffset(i, new_position);
+  }
+
+  // Adjust pc offsets for the disassembly information.
+  if (disasm_info_ != nullptr) {
+    GeneratedCodeInterval* frame_entry_interval = disasm_info_->GetFrameEntryInterval();
+    frame_entry_interval->start = __ GetAdjustedPosition(frame_entry_interval->start);
+    frame_entry_interval->end = __ GetAdjustedPosition(frame_entry_interval->end);
+    for (auto& it : *disasm_info_->GetInstructionIntervals()) {
+      it.second.start = __ GetAdjustedPosition(it.second.start);
+      it.second.end = __ GetAdjustedPosition(it.second.end);
+    }
+    for (auto& it : *disasm_info_->GetSlowPathIntervals()) {
+      it.code_interval.start = __ GetAdjustedPosition(it.code_interval.start);
+      it.code_interval.end = __ GetAdjustedPosition(it.code_interval.end);
+    }
+  }
+
   CodeGenerator::Finalize(allocator);
 }
 
@@ -603,6 +629,7 @@
   }
 
   __ Jr(RA);
+  __ Nop();
 
   __ cfi().RestoreState();
   __ cfi().DefCFAOffset(GetFrameSize());
@@ -939,7 +966,7 @@
 }
 
 void CodeGeneratorMIPS64::MarkGCCard(GpuRegister object, GpuRegister value) {
-  Label done;
+  Mips64Label done;
   GpuRegister card = AT;
   GpuRegister temp = TMP;
   __ Beqzc(value, &done);
@@ -1048,6 +1075,7 @@
   // TODO: anything related to T9/GP/GOT/PIC/.so's?
   __ LoadFromOffset(kLoadDoubleword, T9, TR, entry_point_offset);
   __ Jalr(T9);
+  __ Nop();
   RecordPcInfo(instruction, dex_pc, slow_path);
 }
 
@@ -1079,7 +1107,7 @@
     __ Bind(slow_path->GetReturnLabel());
   } else {
     __ Beqzc(TMP, codegen_->GetLabelOf(successor));
-    __ B(slow_path->GetEntryLabel());
+    __ Bc(slow_path->GetEntryLabel());
     // slow_path will return to GetLabelOf(successor).
   }
 }
@@ -1669,12 +1697,7 @@
   // length is limited by the maximum positive signed 32-bit integer.
   // Unsigned comparison of length and index checks for index < 0
   // and for length <= index simultaneously.
-  // Mips R6 requires lhs != rhs for compact branches.
-  if (index == length) {
-    __ B(slow_path->GetEntryLabel());
-  } else {
-    __ Bgeuc(index, length, slow_path->GetEntryLabel());
-  }
+  __ Bgeuc(index, length, slow_path->GetEntryLabel());
 }
 
 void LocationsBuilderMIPS64::VisitCheckCast(HCheckCast* instruction) {
@@ -2264,7 +2287,7 @@
   if (value.IsConstant()) {
     int64_t divisor = codegen_->GetInt64ValueOf(value.GetConstant()->AsConstant());
     if (divisor == 0) {
-      __ B(slow_path->GetEntryLabel());
+      __ Bc(slow_path->GetEntryLabel());
     } else {
       // A division by a non-null constant is valid. We don't need to perform
       // any check, so simply fall through.
@@ -2316,7 +2339,7 @@
     GenerateSuspendCheck(previous->AsSuspendCheck(), nullptr);
   }
   if (!codegen_->GoesToNextBlock(block, successor)) {
-    __ B(codegen_->GetLabelOf(successor));
+    __ Bc(codegen_->GetLabelOf(successor));
   }
 }
 
@@ -2341,8 +2364,8 @@
 
 void InstructionCodeGeneratorMIPS64::GenerateTestAndBranch(HInstruction* instruction,
                                                            size_t condition_input_index,
-                                                           Label* true_target,
-                                                           Label* false_target) {
+                                                           Mips64Label* true_target,
+                                                           Mips64Label* false_target) {
   HInstruction* cond = instruction->InputAt(condition_input_index);
 
   if (true_target == nullptr && false_target == nullptr) {
@@ -2352,12 +2375,12 @@
     // Constant condition, statically compared against 1.
     if (cond->AsIntConstant()->IsOne()) {
       if (true_target != nullptr) {
-        __ B(true_target);
+        __ Bc(true_target);
       }
     } else {
       DCHECK(cond->AsIntConstant()->IsZero());
       if (false_target != nullptr) {
-        __ B(false_target);
+        __ Bc(false_target);
       }
     }
     return;
@@ -2397,7 +2420,7 @@
     }
 
     IfCondition if_cond;
-    Label* non_fallthrough_target;
+    Mips64Label* non_fallthrough_target;
     if (true_target == nullptr) {
       if_cond = condition->GetOppositeCondition();
       non_fallthrough_target = false_target;
@@ -2435,7 +2458,7 @@
           __ Bnezc(lhs, non_fallthrough_target);  // > 0 if non-zero
           break;
         case kCondAE:
-          __ B(non_fallthrough_target);  // always true
+          __ Bc(non_fallthrough_target);  // always true
           break;
       }
     } else {
@@ -2443,60 +2466,37 @@
         rhs_reg = TMP;
         __ LoadConst32(rhs_reg, rhs_imm);
       }
-      // It looks like we can get here with lhs == rhs. Should that be possible at all?
-      // Mips R6 requires lhs != rhs for compact branches.
-      if (lhs == rhs_reg) {
-        DCHECK(!use_imm);
-        switch (if_cond) {
-          case kCondEQ:
-          case kCondGE:
-          case kCondLE:
-          case kCondBE:
-          case kCondAE:
-            // if lhs == rhs for a positive condition, then it is a branch
-            __ B(non_fallthrough_target);
-            break;
-          case kCondNE:
-          case kCondLT:
-          case kCondGT:
-          case kCondB:
-          case kCondA:
-            // if lhs == rhs for a negative condition, then it is a NOP
-            break;
-        }
-      } else {
-        switch (if_cond) {
-          case kCondEQ:
-            __ Beqc(lhs, rhs_reg, non_fallthrough_target);
-            break;
-          case kCondNE:
-            __ Bnec(lhs, rhs_reg, non_fallthrough_target);
-            break;
-          case kCondLT:
-            __ Bltc(lhs, rhs_reg, non_fallthrough_target);
-            break;
-          case kCondGE:
-            __ Bgec(lhs, rhs_reg, non_fallthrough_target);
-            break;
-          case kCondLE:
-            __ Bgec(rhs_reg, lhs, non_fallthrough_target);
-            break;
-          case kCondGT:
-            __ Bltc(rhs_reg, lhs, non_fallthrough_target);
-            break;
-          case kCondB:
-            __ Bltuc(lhs, rhs_reg, non_fallthrough_target);
-            break;
-          case kCondAE:
-            __ Bgeuc(lhs, rhs_reg, non_fallthrough_target);
-            break;
-          case kCondBE:
-            __ Bgeuc(rhs_reg, lhs, non_fallthrough_target);
-            break;
-          case kCondA:
-            __ Bltuc(rhs_reg, lhs, non_fallthrough_target);
-            break;
-        }
+      switch (if_cond) {
+        case kCondEQ:
+          __ Beqc(lhs, rhs_reg, non_fallthrough_target);
+          break;
+        case kCondNE:
+          __ Bnec(lhs, rhs_reg, non_fallthrough_target);
+          break;
+        case kCondLT:
+          __ Bltc(lhs, rhs_reg, non_fallthrough_target);
+          break;
+        case kCondGE:
+          __ Bgec(lhs, rhs_reg, non_fallthrough_target);
+          break;
+        case kCondLE:
+          __ Bgec(rhs_reg, lhs, non_fallthrough_target);
+          break;
+        case kCondGT:
+          __ Bltc(rhs_reg, lhs, non_fallthrough_target);
+          break;
+        case kCondB:
+          __ Bltuc(lhs, rhs_reg, non_fallthrough_target);
+          break;
+        case kCondAE:
+          __ Bgeuc(lhs, rhs_reg, non_fallthrough_target);
+          break;
+        case kCondBE:
+          __ Bgeuc(rhs_reg, lhs, non_fallthrough_target);
+          break;
+        case kCondA:
+          __ Bltuc(rhs_reg, lhs, non_fallthrough_target);
+          break;
       }
     }
   }
@@ -2504,7 +2504,7 @@
   // If neither branch falls through (case 3), the conditional branch to `true_target`
   // was already emitted (case 2) and we need to emit a jump to `false_target`.
   if (true_target != nullptr && false_target != nullptr) {
-    __ B(false_target);
+    __ Bc(false_target);
   }
 }
 
@@ -2518,9 +2518,9 @@
 void InstructionCodeGeneratorMIPS64::VisitIf(HIf* if_instr) {
   HBasicBlock* true_successor = if_instr->IfTrueSuccessor();
   HBasicBlock* false_successor = if_instr->IfFalseSuccessor();
-  Label* true_target = codegen_->GoesToNextBlock(if_instr->GetBlock(), true_successor) ?
+  Mips64Label* true_target = codegen_->GoesToNextBlock(if_instr->GetBlock(), true_successor) ?
       nullptr : codegen_->GetLabelOf(true_successor);
-  Label* false_target = codegen_->GoesToNextBlock(if_instr->GetBlock(), false_successor) ?
+  Mips64Label* false_target = codegen_->GoesToNextBlock(if_instr->GetBlock(), false_successor) ?
       nullptr : codegen_->GetLabelOf(false_successor);
   GenerateTestAndBranch(if_instr, /* condition_input_index */ 0, true_target, false_target);
 }
@@ -2695,7 +2695,7 @@
   GpuRegister cls = locations->InAt(1).AsRegister<GpuRegister>();
   GpuRegister out = locations->Out().AsRegister<GpuRegister>();
 
-  Label done;
+  Mips64Label done;
 
   // Return 0 if `obj` is null.
   // TODO: Avoid this check if we know `obj` is not null.
@@ -2790,6 +2790,7 @@
   __ LoadFromOffset(kLoadDoubleword, T9, temp, entry_point.Int32Value());
   // T9();
   __ Jalr(T9);
+  __ Nop();
   DCHECK(!codegen_->IsLeafMethod());
   codegen_->RecordPcInfo(invoke, invoke->GetDexPc());
 }
@@ -2924,13 +2925,14 @@
 
   switch (invoke->GetCodePtrLocation()) {
     case HInvokeStaticOrDirect::CodePtrLocation::kCallSelf:
-      __ Jalr(&frame_entry_label_, T9);
+      __ Jialc(&frame_entry_label_, T9);
       break;
     case HInvokeStaticOrDirect::CodePtrLocation::kCallDirect:
       // LR = invoke->GetDirectCodePtr();
       __ LoadConst64(T9, invoke->GetDirectCodePtr());
       // LR()
       __ Jalr(T9);
+      __ Nop();
       break;
     case HInvokeStaticOrDirect::CodePtrLocation::kCallDirectWithFixup:
     case HInvokeStaticOrDirect::CodePtrLocation::kCallPCRelative:
@@ -2947,6 +2949,7 @@
                             kMips64WordSize).Int32Value());
       // T9()
       __ Jalr(T9);
+      __ Nop();
       break;
   }
   DCHECK(!IsLeafMethod());
@@ -2988,6 +2991,7 @@
   __ LoadFromOffset(kLoadDoubleword, T9, temp, entry_point.Int32Value());
   // T9();
   __ Jalr(T9);
+  __ Nop();
 }
 
 void InstructionCodeGeneratorMIPS64::VisitInvokeVirtual(HInvokeVirtual* invoke) {
@@ -3926,7 +3930,7 @@
   const ArenaVector<HBasicBlock*>& successors = switch_instr->GetBlock()->GetSuccessors();
   for (int32_t i = 0; i < num_entries; i++) {
     int32_t case_value = lower_bound + i;
-    Label* succ = codegen_->GetLabelOf(successors[i]);
+    Mips64Label* succ = codegen_->GetLabelOf(successors[i]);
     if (case_value == 0) {
       __ Beqzc(value_reg, succ);
     } else {
@@ -3937,7 +3941,7 @@
 
   // And the default for any other value.
   if (!codegen_->GoesToNextBlock(switch_instr->GetBlock(), default_block)) {
-    __ B(codegen_->GetLabelOf(default_block));
+    __ Bc(codegen_->GetLabelOf(default_block));
   }
 }
 
diff --git a/compiler/optimizing/code_generator_mips64.h b/compiler/optimizing/code_generator_mips64.h
index a078dd1..85e3a4a 100644
--- a/compiler/optimizing/code_generator_mips64.h
+++ b/compiler/optimizing/code_generator_mips64.h
@@ -158,12 +158,12 @@
  public:
   SlowPathCodeMIPS64() : entry_label_(), exit_label_() {}
 
-  Label* GetEntryLabel() { return &entry_label_; }
-  Label* GetExitLabel() { return &exit_label_; }
+  Mips64Label* GetEntryLabel() { return &entry_label_; }
+  Mips64Label* GetExitLabel() { return &exit_label_; }
 
  private:
-  Label entry_label_;
-  Label exit_label_;
+  Mips64Label entry_label_;
+  Mips64Label exit_label_;
 
   DISALLOW_COPY_AND_ASSIGN(SlowPathCodeMIPS64);
 };
@@ -231,8 +231,8 @@
   void GenerateExplicitNullCheck(HNullCheck* instruction);
   void GenerateTestAndBranch(HInstruction* instruction,
                              size_t condition_input_index,
-                             Label* true_target,
-                             Label* false_target);
+                             Mips64Label* true_target,
+                             Mips64Label* false_target);
   void DivRemOneOrMinusOne(HBinaryOperation* instruction);
   void DivRemByPowerOfTwo(HBinaryOperation* instruction);
   void GenerateDivRemWithAnyConstant(HBinaryOperation* instruction);
@@ -265,7 +265,7 @@
   size_t GetFloatingPointSpillSlotSize() const OVERRIDE { return kMips64WordSize; }
 
   uintptr_t GetAddressOf(HBasicBlock* block) const OVERRIDE {
-    return GetLabelOf(block)->Position();
+    return assembler_.GetLabelLocation(GetLabelOf(block));
   }
 
   HGraphVisitor* GetLocationBuilder() OVERRIDE { return &location_builder_; }
@@ -298,12 +298,12 @@
     return isa_features_;
   }
 
-  Label* GetLabelOf(HBasicBlock* block) const {
-    return CommonGetLabelOf<Label>(block_labels_, block);
+  Mips64Label* GetLabelOf(HBasicBlock* block) const {
+    return CommonGetLabelOf<Mips64Label>(block_labels_, block);
   }
 
   void Initialize() OVERRIDE {
-    block_labels_ = CommonInitializeLabels<Label>();
+    block_labels_ = CommonInitializeLabels<Mips64Label>();
   }
 
   void Finalize(CodeAllocator* allocator) OVERRIDE;
@@ -349,8 +349,8 @@
 
  private:
   // Labels for each block that will be compiled.
-  Label* block_labels_;  // Indexed by block id.
-  Label frame_entry_label_;
+  Mips64Label* block_labels_;  // Indexed by block id.
+  Mips64Label frame_entry_label_;
   LocationsBuilderMIPS64 location_builder_;
   InstructionCodeGeneratorMIPS64 instruction_visitor_;
   ParallelMoveResolverMIPS64 move_resolver_;
diff --git a/compiler/optimizing/intrinsics_mips64.cc b/compiler/optimizing/intrinsics_mips64.cc
index 3654159..ecee11d 100644
--- a/compiler/optimizing/intrinsics_mips64.cc
+++ b/compiler/optimizing/intrinsics_mips64.cc
@@ -115,7 +115,7 @@
     }
 
     RestoreLiveRegisters(codegen, invoke_->GetLocations());
-    __ B(GetExitLabel());
+    __ Bc(GetExitLabel());
   }
 
   const char* GetDescription() const OVERRIDE { return "IntrinsicSlowPathMIPS64"; }
@@ -806,7 +806,7 @@
 
   DCHECK_NE(in, out);
 
-  Label done;
+  Mips64Label done;
 
   // double floor/ceil(double in) {
   //     if in.isNaN || in.isInfinite || in.isZero {
@@ -1256,7 +1256,7 @@
   // } while (tmp_value == 0 && failure([tmp_ptr] <- r_new_value));
   // result = tmp_value != 0;
 
-  Label loop_head, exit_loop;
+  Mips64Label loop_head, exit_loop;
   __ Daddu(TMP, base, offset);
   __ Sync(0);
   __ Bind(&loop_head);
@@ -1418,10 +1418,10 @@
   GpuRegister temp2 = locations->GetTemp(1).AsRegister<GpuRegister>();
   GpuRegister temp3 = locations->GetTemp(2).AsRegister<GpuRegister>();
 
-  Label loop;
-  Label end;
-  Label return_true;
-  Label return_false;
+  Mips64Label loop;
+  Mips64Label end;
+  Mips64Label return_true;
+  Mips64Label return_false;
 
   // Get offsets of count, value, and class fields within a string object.
   const int32_t count_offset = mirror::String::CountOffset().Int32Value();
@@ -1485,7 +1485,7 @@
   // If loop does not result in returning false, we return true.
   __ Bind(&return_true);
   __ LoadConst64(out, 1);
-  __ B(&end);
+  __ Bc(&end);
 
   // Return false and exit the function.
   __ Bind(&return_false);
@@ -1514,7 +1514,7 @@
       // full slow-path down and branch unconditionally.
       slow_path = new (allocator) IntrinsicSlowPathMIPS64(invoke);
       codegen->AddSlowPath(slow_path);
-      __ B(slow_path->GetEntryLabel());
+      __ Bc(slow_path->GetEntryLabel());
       __ Bind(slow_path->GetExitLabel());
       return;
     }
diff --git a/compiler/optimizing/optimizing_cfi_test.cc b/compiler/optimizing/optimizing_cfi_test.cc
index 34f1fe5..2b0d522 100644
--- a/compiler/optimizing/optimizing_cfi_test.cc
+++ b/compiler/optimizing/optimizing_cfi_test.cc
@@ -25,6 +25,7 @@
 #include "utils/assembler.h"
 #include "utils/arm/assembler_thumb2.h"
 #include "utils/mips/assembler_mips.h"
+#include "utils/mips64/assembler_mips64.h"
 
 #include "optimizing/optimizing_cfi_test_expected.inc"
 
@@ -212,6 +213,34 @@
   Check(kMips, "kMips_adjust", expected_asm, expected_cfi);
 }
 
+TEST_F(OptimizingCFITest, kMips64Adjust) {
+  // One NOP in forbidden slot, 1 << 15 NOPS have size 1 << 17 which exceeds 18-bit signed maximum.
+  static constexpr size_t kNumNops = 1u + (1u << 15);
+  std::vector<uint8_t> expected_asm(
+      expected_asm_kMips64_adjust_head,
+      expected_asm_kMips64_adjust_head + arraysize(expected_asm_kMips64_adjust_head));
+  expected_asm.resize(expected_asm.size() + kNumNops * 4u, 0u);
+  expected_asm.insert(
+      expected_asm.end(),
+      expected_asm_kMips64_adjust_tail,
+      expected_asm_kMips64_adjust_tail + arraysize(expected_asm_kMips64_adjust_tail));
+  std::vector<uint8_t> expected_cfi(
+      expected_cfi_kMips64_adjust,
+      expected_cfi_kMips64_adjust + arraysize(expected_cfi_kMips64_adjust));
+  SetUpFrame(kMips64);
+#define __ down_cast<mips64::Mips64Assembler*>(GetCodeGenerator()->GetAssembler())->
+  mips64::Mips64Label target;
+  __ Beqc(mips64::A1, mips64::A2, &target);
+  // Push the target out of range of BEQC.
+  for (size_t i = 0; i != kNumNops; ++i) {
+    __ Nop();
+  }
+  __ Bind(&target);
+#undef __
+  Finish();
+  Check(kMips64, "kMips64_adjust", expected_asm, expected_cfi);
+}
+
 #endif  // __ANDROID__
 
 }  // namespace art
diff --git a/compiler/optimizing/optimizing_cfi_test_expected.inc b/compiler/optimizing/optimizing_cfi_test_expected.inc
index 4571ebf..de85729 100644
--- a/compiler/optimizing/optimizing_cfi_test_expected.inc
+++ b/compiler/optimizing/optimizing_cfi_test_expected.inc
@@ -413,3 +413,57 @@
 // 0x0002007c: nop
 // 0x00020080: .cfi_restore_state
 // 0x00020080: .cfi_def_cfa_offset: 64
+
+static constexpr uint8_t expected_asm_kMips64_adjust_head[] = {
+    0xD8, 0xFF, 0xBD, 0x67, 0x20, 0x00, 0xBF, 0xFF, 0x18, 0x00, 0xB1, 0xFF,
+    0x10, 0x00, 0xB0, 0xFF, 0x08, 0x00, 0xB9, 0xF7, 0x00, 0x00, 0xB8, 0xF7,
+    0xE8, 0xFF, 0xBD, 0x67, 0x00, 0x00, 0xA4, 0xFF, 0x02, 0x00, 0xA6, 0x60,
+    0x02, 0x00, 0x3E, 0xEC, 0x0C, 0x00, 0x01, 0xD8,
+};
+static constexpr uint8_t expected_asm_kMips64_adjust_tail[] = {
+    0x18, 0x00, 0xBD, 0x67, 0x00, 0x00, 0xB8, 0xD7, 0x08, 0x00, 0xB9, 0xD7,
+    0x10, 0x00, 0xB0, 0xDF, 0x18, 0x00, 0xB1, 0xDF, 0x20, 0x00, 0xBF, 0xDF,
+    0x28, 0x00, 0xBD, 0x67, 0x09, 0x00, 0xE0, 0x03, 0x00, 0x00, 0x00, 0x00,
+};
+static constexpr uint8_t expected_cfi_kMips64_adjust[] = {
+    0x44, 0x0E, 0x28, 0x44, 0x9F, 0x02, 0x44, 0x91, 0x04, 0x44, 0x90, 0x06,
+    0x4C, 0x0E, 0x40, 0x04, 0x14, 0x00, 0x02, 0x00, 0x0A, 0x44, 0x0E, 0x28,
+    0x4C, 0xD0, 0x44, 0xD1, 0x44, 0xDF, 0x44, 0x0E, 0x00, 0x48, 0x0B, 0x0E,
+    0x40,
+};
+// 0x00000000: daddiu r29, r29, -40
+// 0x00000004: .cfi_def_cfa_offset: 40
+// 0x00000004: sd r31, +32(r29)
+// 0x00000008: .cfi_offset: r31 at cfa-8
+// 0x00000008: sd r17, +24(r29)
+// 0x0000000c: .cfi_offset: r17 at cfa-16
+// 0x0000000c: sd r16, +16(r29)
+// 0x00000010: .cfi_offset: r16 at cfa-24
+// 0x00000010: sdc1 f25, +8(r29)
+// 0x00000014: sdc1 f24, +0(r29)
+// 0x00000018: daddiu r29, r29, -24
+// 0x0000001c: .cfi_def_cfa_offset: 64
+// 0x0000001c: sd r4, +0(r29)
+// 0x00000020: bnec r5, r6, 0x0000002c ; +12
+// 0x00000024: auipc r1, 2
+// 0x00000028: jic r1, 12 ; b 0x00020030 ; +131080
+// 0x0000002c: nop
+//             ...
+// 0x0002002c: nop
+// 0x00020030: .cfi_remember_state
+// 0x00020030: daddiu r29, r29, 24
+// 0x00020034: .cfi_def_cfa_offset: 40
+// 0x00020034: ldc1 f24, +0(r29)
+// 0x00020038: ldc1 f25, +8(r29)
+// 0x0002003c: ld r16, +16(r29)
+// 0x00020040: .cfi_restore: r16
+// 0x00020040: ld r17, +24(r29)
+// 0x00020044: .cfi_restore: r17
+// 0x00020044: ld r31, +32(r29)
+// 0x00020048: .cfi_restore: r31
+// 0x00020048: daddiu r29, r29, 40
+// 0x0002004c: .cfi_def_cfa_offset: 0
+// 0x0002004c: jr r31
+// 0x00020050: nop
+// 0x00020054: .cfi_restore_state
+// 0x00020054: .cfi_def_cfa_offset: 64
diff --git a/compiler/utils/assembler_test.h b/compiler/utils/assembler_test.h
index f1233ca..9457da1 100644
--- a/compiler/utils/assembler_test.h
+++ b/compiler/utils/assembler_test.h
@@ -840,12 +840,17 @@
     return str;
   }
 
+  // Override this to pad the code with NOPs to a certain size if needed.
+  virtual void Pad(std::vector<uint8_t>& data ATTRIBUTE_UNUSED) {
+  }
+
   void DriverWrapper(std::string assembly_text, std::string test_name) {
     assembler_->FinalizeCode();
     size_t cs = assembler_->CodeSize();
     std::unique_ptr<std::vector<uint8_t>> data(new std::vector<uint8_t>(cs));
     MemoryRegion code(&(*data)[0], data->size());
     assembler_->FinalizeInstructions(code);
+    Pad(*data);
     test_helper_->Driver(*data, assembly_text, test_name);
   }
 
diff --git a/compiler/utils/mips64/assembler_mips64.cc b/compiler/utils/mips64/assembler_mips64.cc
index ba2525e..107d5bb 100644
--- a/compiler/utils/mips64/assembler_mips64.cc
+++ b/compiler/utils/mips64/assembler_mips64.cc
@@ -19,15 +19,73 @@
 #include "base/bit_utils.h"
 #include "base/casts.h"
 #include "entrypoints/quick/quick_entrypoints.h"
+#include "entrypoints/quick/quick_entrypoints_enum.h"
 #include "memory_region.h"
 #include "thread.h"
 
 namespace art {
 namespace mips64 {
 
+void Mips64Assembler::FinalizeCode() {
+  for (auto& exception_block : exception_blocks_) {
+    EmitExceptionPoll(&exception_block);
+  }
+  PromoteBranches();
+}
+
+void Mips64Assembler::FinalizeInstructions(const MemoryRegion& region) {
+  EmitBranches();
+  Assembler::FinalizeInstructions(region);
+  PatchCFI();
+}
+
+void Mips64Assembler::PatchCFI() {
+  if (cfi().NumberOfDelayedAdvancePCs() == 0u) {
+    return;
+  }
+
+  typedef DebugFrameOpCodeWriterForAssembler::DelayedAdvancePC DelayedAdvancePC;
+  const auto data = cfi().ReleaseStreamAndPrepareForDelayedAdvancePC();
+  const std::vector<uint8_t>& old_stream = data.first;
+  const std::vector<DelayedAdvancePC>& advances = data.second;
+
+  // Refill our data buffer with patched opcodes.
+  cfi().ReserveCFIStream(old_stream.size() + advances.size() + 16);
+  size_t stream_pos = 0;
+  for (const DelayedAdvancePC& advance : advances) {
+    DCHECK_GE(advance.stream_pos, stream_pos);
+    // Copy old data up to the point where advance was issued.
+    cfi().AppendRawData(old_stream, stream_pos, advance.stream_pos);
+    stream_pos = advance.stream_pos;
+    // Insert the advance command with its final offset.
+    size_t final_pc = GetAdjustedPosition(advance.pc);
+    cfi().AdvancePC(final_pc);
+  }
+  // Copy the final segment if any.
+  cfi().AppendRawData(old_stream, stream_pos, old_stream.size());
+}
+
+void Mips64Assembler::EmitBranches() {
+  CHECK(!overwriting_);
+  // Switch from appending instructions at the end of the buffer to overwriting
+  // existing instructions (branch placeholders) in the buffer.
+  overwriting_ = true;
+  for (auto& branch : branches_) {
+    EmitBranch(&branch);
+  }
+  overwriting_ = false;
+}
+
 void Mips64Assembler::Emit(uint32_t value) {
-  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
-  buffer_.Emit<uint32_t>(value);
+  if (overwriting_) {
+    // Branches to labels are emitted into their placeholders here.
+    buffer_.Store<uint32_t>(overwrite_location_, value);
+    overwrite_location_ += sizeof(uint32_t);
+  } else {
+    // Other instructions are simply appended at the end here.
+    AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+    buffer_.Emit<uint32_t>(value);
+  }
 }
 
 void Mips64Assembler::EmitR(int opcode, GpuRegister rs, GpuRegister rt, GpuRegister rd,
@@ -82,15 +140,16 @@
 
 void Mips64Assembler::EmitI21(int opcode, GpuRegister rs, uint32_t imm21) {
   CHECK_NE(rs, kNoGpuRegister);
+  CHECK(IsUint<21>(imm21)) << imm21;
   uint32_t encoding = static_cast<uint32_t>(opcode) << kOpcodeShift |
                       static_cast<uint32_t>(rs) << kRsShift |
-                      (imm21 & 0x1FFFFF);
+                      imm21;
   Emit(encoding);
 }
 
-void Mips64Assembler::EmitJ(int opcode, uint32_t addr26) {
-  uint32_t encoding = static_cast<uint32_t>(opcode) << kOpcodeShift |
-                      (addr26 & 0x3FFFFFF);
+void Mips64Assembler::EmitI26(int opcode, uint32_t imm26) {
+  CHECK(IsUint<26>(imm26)) << imm26;
+  uint32_t encoding = static_cast<uint32_t>(opcode) << kOpcodeShift | imm26;
   Emit(encoding);
 }
 
@@ -428,26 +487,6 @@
   EmitI(0xb, rs, rt, imm16);
 }
 
-void Mips64Assembler::Beq(GpuRegister rs, GpuRegister rt, uint16_t imm16) {
-  EmitI(0x4, rs, rt, imm16);
-  Nop();
-}
-
-void Mips64Assembler::Bne(GpuRegister rs, GpuRegister rt, uint16_t imm16) {
-  EmitI(0x5, rs, rt, imm16);
-  Nop();
-}
-
-void Mips64Assembler::J(uint32_t addr26) {
-  EmitJ(0x2, addr26);
-  Nop();
-}
-
-void Mips64Assembler::Jal(uint32_t addr26) {
-  EmitJ(0x3, addr26);
-  Nop();
-}
-
 void Mips64Assembler::Seleqz(GpuRegister rd, GpuRegister rs, GpuRegister rt) {
   EmitR(0, rs, rt, rd, 0, 0x35);
 }
@@ -474,7 +513,6 @@
 
 void Mips64Assembler::Jalr(GpuRegister rd, GpuRegister rs) {
   EmitR(0, rs, static_cast<GpuRegister>(0), rd, 0, 0x09);
-  Nop();
 }
 
 void Mips64Assembler::Jalr(GpuRegister rs) {
@@ -489,6 +527,15 @@
   EmitI(0x3B, rs, static_cast<GpuRegister>(0x1E), imm16);
 }
 
+void Mips64Assembler::Addiupc(GpuRegister rs, uint32_t imm19) {
+  CHECK(IsUint<19>(imm19)) << imm19;
+  EmitI21(0x3B, rs, imm19);
+}
+
+void Mips64Assembler::Bc(uint32_t imm26) {
+  EmitI26(0x32, imm26);
+}
+
 void Mips64Assembler::Jic(GpuRegister rt, uint16_t imm16) {
   EmitI(0x36, static_cast<GpuRegister>(0), rt, imm16);
 }
@@ -549,14 +596,14 @@
   CHECK_NE(rs, ZERO);
   CHECK_NE(rt, ZERO);
   CHECK_NE(rs, rt);
-  EmitI(0x8, (rs < rt) ? rs : rt, (rs < rt) ? rt : rs, imm16);
+  EmitI(0x8, std::min(rs, rt), std::max(rs, rt), imm16);
 }
 
 void Mips64Assembler::Bnec(GpuRegister rs, GpuRegister rt, uint16_t imm16) {
   CHECK_NE(rs, ZERO);
   CHECK_NE(rt, ZERO);
   CHECK_NE(rs, rt);
-  EmitI(0x18, (rs < rt) ? rs : rt, (rs < rt) ? rt : rs, imm16);
+  EmitI(0x18, std::min(rs, rt), std::max(rs, rt), imm16);
 }
 
 void Mips64Assembler::Beqzc(GpuRegister rs, uint32_t imm21) {
@@ -569,6 +616,65 @@
   EmitI21(0x3E, rs, imm21);
 }
 
+void Mips64Assembler::EmitBcondc(BranchCondition cond,
+                                 GpuRegister rs,
+                                 GpuRegister rt,
+                                 uint32_t imm16_21) {
+  switch (cond) {
+    case kCondLT:
+      Bltc(rs, rt, imm16_21);
+      break;
+    case kCondGE:
+      Bgec(rs, rt, imm16_21);
+      break;
+    case kCondLE:
+      Bgec(rt, rs, imm16_21);
+      break;
+    case kCondGT:
+      Bltc(rt, rs, imm16_21);
+      break;
+    case kCondLTZ:
+      CHECK_EQ(rt, ZERO);
+      Bltzc(rs, imm16_21);
+      break;
+    case kCondGEZ:
+      CHECK_EQ(rt, ZERO);
+      Bgezc(rs, imm16_21);
+      break;
+    case kCondLEZ:
+      CHECK_EQ(rt, ZERO);
+      Blezc(rs, imm16_21);
+      break;
+    case kCondGTZ:
+      CHECK_EQ(rt, ZERO);
+      Bgtzc(rs, imm16_21);
+      break;
+    case kCondEQ:
+      Beqc(rs, rt, imm16_21);
+      break;
+    case kCondNE:
+      Bnec(rs, rt, imm16_21);
+      break;
+    case kCondEQZ:
+      CHECK_EQ(rt, ZERO);
+      Beqzc(rs, imm16_21);
+      break;
+    case kCondNEZ:
+      CHECK_EQ(rt, ZERO);
+      Bnezc(rs, imm16_21);
+      break;
+    case kCondLTU:
+      Bltuc(rs, rt, imm16_21);
+      break;
+    case kCondGEU:
+      Bgeuc(rs, rt, imm16_21);
+      break;
+    case kUncond:
+      LOG(FATAL) << "Unexpected branch condition " << cond;
+      UNREACHABLE();
+  }
+}
+
 void Mips64Assembler::AddS(FpuRegister fd, FpuRegister fs, FpuRegister ft) {
   EmitFR(0x11, 0x10, ft, fs, fd, 0x0);
 }
@@ -925,15 +1031,6 @@
   }
 }
 
-void Mips64Assembler::Addiu32(GpuRegister rt, GpuRegister rs, int32_t value, GpuRegister rtmp) {
-  if (IsInt<16>(value)) {
-    Addiu(rt, rs, value);
-  } else {
-    LoadConst32(rtmp, value);
-    Addu(rt, rs, rtmp);
-  }
-}
-
 void Mips64Assembler::Daddiu64(GpuRegister rt, GpuRegister rs, int64_t value, GpuRegister rtmp) {
   if (IsInt<16>(value)) {
     Daddiu(rt, rs, value);
@@ -943,177 +1040,621 @@
   }
 }
 
-//
-// MIPS64R6 branches
-//
-//
-// Unconditional (pc + 32-bit signed offset):
-//
-//   auipc    at, ofs_high
-//   jic      at, ofs_low
-//   // no delay/forbidden slot
-//
-//
-// Conditional (pc + 32-bit signed offset):
-//
-//   b<cond>c   reg, +2      // skip next 2 instructions
-//   auipc      at, ofs_high
-//   jic        at, ofs_low
-//   // no delay/forbidden slot
-//
-//
-// Unconditional (pc + 32-bit signed offset) and link:
-//
-//   auipc    reg, ofs_high
-//   daddiu   reg, ofs_low
-//   jialc    reg, 0
-//   // no delay/forbidden slot
-//
-//
-// TODO: use shorter instruction sequences whenever possible.
-//
+void Mips64Assembler::Branch::InitShortOrLong(Mips64Assembler::Branch::OffsetBits offset_size,
+                                              Mips64Assembler::Branch::Type short_type,
+                                              Mips64Assembler::Branch::Type long_type) {
+  type_ = (offset_size <= branch_info_[short_type].offset_size) ? short_type : long_type;
+}
 
-void Mips64Assembler::Bind(Label* label) {
+void Mips64Assembler::Branch::InitializeType(bool is_call) {
+  OffsetBits offset_size = GetOffsetSizeNeeded(location_, target_);
+  if (is_call) {
+    InitShortOrLong(offset_size, kCall, kLongCall);
+  } else if (condition_ == kUncond) {
+    InitShortOrLong(offset_size, kUncondBranch, kLongUncondBranch);
+  } else {
+    if (condition_ == kCondEQZ || condition_ == kCondNEZ) {
+      // Special case for beqzc/bnezc with longer offset than in other b<cond>c instructions.
+      type_ = (offset_size <= kOffset23) ? kCondBranch : kLongCondBranch;
+    } else {
+      InitShortOrLong(offset_size, kCondBranch, kLongCondBranch);
+    }
+  }
+  old_type_ = type_;
+}
+
+bool Mips64Assembler::Branch::IsNop(BranchCondition condition, GpuRegister lhs, GpuRegister rhs) {
+  switch (condition) {
+    case kCondLT:
+    case kCondGT:
+    case kCondNE:
+    case kCondLTU:
+      return lhs == rhs;
+    default:
+      return false;
+  }
+}
+
+bool Mips64Assembler::Branch::IsUncond(BranchCondition condition,
+                                       GpuRegister lhs,
+                                       GpuRegister rhs) {
+  switch (condition) {
+    case kUncond:
+      return true;
+    case kCondGE:
+    case kCondLE:
+    case kCondEQ:
+    case kCondGEU:
+      return lhs == rhs;
+    default:
+      return false;
+  }
+}
+
+Mips64Assembler::Branch::Branch(uint32_t location, uint32_t target)
+    : old_location_(location),
+      location_(location),
+      target_(target),
+      lhs_reg_(ZERO),
+      rhs_reg_(ZERO),
+      condition_(kUncond) {
+  InitializeType(false);
+}
+
+Mips64Assembler::Branch::Branch(uint32_t location,
+                                uint32_t target,
+                                Mips64Assembler::BranchCondition condition,
+                                GpuRegister lhs_reg,
+                                GpuRegister rhs_reg)
+    : old_location_(location),
+      location_(location),
+      target_(target),
+      lhs_reg_(lhs_reg),
+      rhs_reg_(rhs_reg),
+      condition_(condition) {
+  CHECK_NE(condition, kUncond);
+  switch (condition) {
+    case kCondEQ:
+    case kCondNE:
+    case kCondLT:
+    case kCondGE:
+    case kCondLE:
+    case kCondGT:
+    case kCondLTU:
+    case kCondGEU:
+      CHECK_NE(lhs_reg, ZERO);
+      CHECK_NE(rhs_reg, ZERO);
+      break;
+    case kCondLTZ:
+    case kCondGEZ:
+    case kCondLEZ:
+    case kCondGTZ:
+    case kCondEQZ:
+    case kCondNEZ:
+      CHECK_NE(lhs_reg, ZERO);
+      CHECK_EQ(rhs_reg, ZERO);
+      break;
+    case kUncond:
+      UNREACHABLE();
+  }
+  CHECK(!IsNop(condition, lhs_reg, rhs_reg));
+  if (IsUncond(condition, lhs_reg, rhs_reg)) {
+    // Branch condition is always true, make the branch unconditional.
+    condition_ = kUncond;
+  }
+  InitializeType(false);
+}
+
+Mips64Assembler::Branch::Branch(uint32_t location, uint32_t target, GpuRegister indirect_reg)
+    : old_location_(location),
+      location_(location),
+      target_(target),
+      lhs_reg_(indirect_reg),
+      rhs_reg_(ZERO),
+      condition_(kUncond) {
+  CHECK_NE(indirect_reg, ZERO);
+  CHECK_NE(indirect_reg, AT);
+  InitializeType(true);
+}
+
+Mips64Assembler::BranchCondition Mips64Assembler::Branch::OppositeCondition(
+    Mips64Assembler::BranchCondition cond) {
+  switch (cond) {
+    case kCondLT:
+      return kCondGE;
+    case kCondGE:
+      return kCondLT;
+    case kCondLE:
+      return kCondGT;
+    case kCondGT:
+      return kCondLE;
+    case kCondLTZ:
+      return kCondGEZ;
+    case kCondGEZ:
+      return kCondLTZ;
+    case kCondLEZ:
+      return kCondGTZ;
+    case kCondGTZ:
+      return kCondLEZ;
+    case kCondEQ:
+      return kCondNE;
+    case kCondNE:
+      return kCondEQ;
+    case kCondEQZ:
+      return kCondNEZ;
+    case kCondNEZ:
+      return kCondEQZ;
+    case kCondLTU:
+      return kCondGEU;
+    case kCondGEU:
+      return kCondLTU;
+    case kUncond:
+      LOG(FATAL) << "Unexpected branch condition " << cond;
+  }
+  UNREACHABLE();
+}
+
+Mips64Assembler::Branch::Type Mips64Assembler::Branch::GetType() const {
+  return type_;
+}
+
+Mips64Assembler::BranchCondition Mips64Assembler::Branch::GetCondition() const {
+  return condition_;
+}
+
+GpuRegister Mips64Assembler::Branch::GetLeftRegister() const {
+  return lhs_reg_;
+}
+
+GpuRegister Mips64Assembler::Branch::GetRightRegister() const {
+  return rhs_reg_;
+}
+
+uint32_t Mips64Assembler::Branch::GetTarget() const {
+  return target_;
+}
+
+uint32_t Mips64Assembler::Branch::GetLocation() const {
+  return location_;
+}
+
+uint32_t Mips64Assembler::Branch::GetOldLocation() const {
+  return old_location_;
+}
+
+uint32_t Mips64Assembler::Branch::GetLength() const {
+  return branch_info_[type_].length;
+}
+
+uint32_t Mips64Assembler::Branch::GetOldLength() const {
+  return branch_info_[old_type_].length;
+}
+
+uint32_t Mips64Assembler::Branch::GetSize() const {
+  return GetLength() * sizeof(uint32_t);
+}
+
+uint32_t Mips64Assembler::Branch::GetOldSize() const {
+  return GetOldLength() * sizeof(uint32_t);
+}
+
+uint32_t Mips64Assembler::Branch::GetEndLocation() const {
+  return GetLocation() + GetSize();
+}
+
+uint32_t Mips64Assembler::Branch::GetOldEndLocation() const {
+  return GetOldLocation() + GetOldSize();
+}
+
+bool Mips64Assembler::Branch::IsLong() const {
+  switch (type_) {
+    // Short branches.
+    case kUncondBranch:
+    case kCondBranch:
+    case kCall:
+      return false;
+    // Long branches.
+    case kLongUncondBranch:
+    case kLongCondBranch:
+    case kLongCall:
+      return true;
+  }
+  UNREACHABLE();
+}
+
+bool Mips64Assembler::Branch::IsResolved() const {
+  return target_ != kUnresolved;
+}
+
+Mips64Assembler::Branch::OffsetBits Mips64Assembler::Branch::GetOffsetSize() const {
+  OffsetBits offset_size =
+      (type_ == kCondBranch && (condition_ == kCondEQZ || condition_ == kCondNEZ))
+          ? kOffset23
+          : branch_info_[type_].offset_size;
+  return offset_size;
+}
+
+Mips64Assembler::Branch::OffsetBits Mips64Assembler::Branch::GetOffsetSizeNeeded(uint32_t location,
+                                                                                 uint32_t target) {
+  // For unresolved targets assume the shortest encoding
+  // (later it will be made longer if needed).
+  if (target == kUnresolved)
+    return kOffset16;
+  int64_t distance = static_cast<int64_t>(target) - location;
+  // To simplify calculations in composite branches consisting of multiple instructions
+  // bump up the distance by a value larger than the max byte size of a composite branch.
+  distance += (distance >= 0) ? kMaxBranchSize : -kMaxBranchSize;
+  if (IsInt<kOffset16>(distance))
+    return kOffset16;
+  else if (IsInt<kOffset18>(distance))
+    return kOffset18;
+  else if (IsInt<kOffset21>(distance))
+    return kOffset21;
+  else if (IsInt<kOffset23>(distance))
+    return kOffset23;
+  else if (IsInt<kOffset28>(distance))
+    return kOffset28;
+  return kOffset32;
+}
+
+void Mips64Assembler::Branch::Resolve(uint32_t target) {
+  target_ = target;
+}
+
+void Mips64Assembler::Branch::Relocate(uint32_t expand_location, uint32_t delta) {
+  if (location_ > expand_location) {
+    location_ += delta;
+  }
+  if (!IsResolved()) {
+    return;  // Don't know the target yet.
+  }
+  if (target_ > expand_location) {
+    target_ += delta;
+  }
+}
+
+void Mips64Assembler::Branch::PromoteToLong() {
+  switch (type_) {
+    // Short branches.
+    case kUncondBranch:
+      type_ = kLongUncondBranch;
+      break;
+    case kCondBranch:
+      type_ = kLongCondBranch;
+      break;
+    case kCall:
+      type_ = kLongCall;
+      break;
+    default:
+      // Note: 'type_' is already long.
+      break;
+  }
+  CHECK(IsLong());
+}
+
+uint32_t Mips64Assembler::Branch::PromoteIfNeeded(uint32_t max_short_distance) {
+  // If the branch is still unresolved or already long, nothing to do.
+  if (IsLong() || !IsResolved()) {
+    return 0;
+  }
+  // Promote the short branch to long if the offset size is too small
+  // to hold the distance between location_ and target_.
+  if (GetOffsetSizeNeeded(location_, target_) > GetOffsetSize()) {
+    PromoteToLong();
+    uint32_t old_size = GetOldSize();
+    uint32_t new_size = GetSize();
+    CHECK_GT(new_size, old_size);
+    return new_size - old_size;
+  }
+  // The following logic is for debugging/testing purposes.
+  // Promote some short branches to long when it's not really required.
+  if (UNLIKELY(max_short_distance != std::numeric_limits<uint32_t>::max())) {
+    int64_t distance = static_cast<int64_t>(target_) - location_;
+    distance = (distance >= 0) ? distance : -distance;
+    if (distance >= max_short_distance) {
+      PromoteToLong();
+      uint32_t old_size = GetOldSize();
+      uint32_t new_size = GetSize();
+      CHECK_GT(new_size, old_size);
+      return new_size - old_size;
+    }
+  }
+  return 0;
+}
+
+uint32_t Mips64Assembler::Branch::GetOffsetLocation() const {
+  return location_ + branch_info_[type_].instr_offset * sizeof(uint32_t);
+}
+
+uint32_t Mips64Assembler::Branch::GetOffset() const {
+  CHECK(IsResolved());
+  uint32_t ofs_mask = 0xFFFFFFFF >> (32 - GetOffsetSize());
+  // Calculate the byte distance between instructions and also account for
+  // different PC-relative origins.
+  uint32_t offset = target_ - GetOffsetLocation() - branch_info_[type_].pc_org * sizeof(uint32_t);
+  // Prepare the offset for encoding into the instruction(s).
+  offset = (offset & ofs_mask) >> branch_info_[type_].offset_shift;
+  return offset;
+}
+
+Mips64Assembler::Branch* Mips64Assembler::GetBranch(uint32_t branch_id) {
+  CHECK_LT(branch_id, branches_.size());
+  return &branches_[branch_id];
+}
+
+const Mips64Assembler::Branch* Mips64Assembler::GetBranch(uint32_t branch_id) const {
+  CHECK_LT(branch_id, branches_.size());
+  return &branches_[branch_id];
+}
+
+void Mips64Assembler::Bind(Mips64Label* label) {
   CHECK(!label->IsBound());
-  int32_t bound_pc = buffer_.Size();
+  uint32_t bound_pc = buffer_.Size();
 
-  // Walk the list of the branches (auipc + jic pairs) referring to and preceding this label.
-  // Embed the previously unknown pc-relative addresses in them.
+  // Walk the list of branches referring to and preceding this label.
+  // Store the previously unknown target addresses in them.
   while (label->IsLinked()) {
-    int32_t position = label->Position();
-    // Extract the branch (instruction pair)
-    uint32_t auipc = buffer_.Load<uint32_t>(position);
-    uint32_t jic = buffer_.Load<uint32_t>(position + 4);  // actually, jic or daddiu
+    uint32_t branch_id = label->Position();
+    Branch* branch = GetBranch(branch_id);
+    branch->Resolve(bound_pc);
 
-    // Extract the location of the previous pair in the list (walking the list backwards;
-    // the previous pair location was stored in the immediate operands of the instructions)
-    int32_t prev = (auipc << 16) | (jic & 0xFFFF);
-
-    // Get the pc-relative address
-    uint32_t offset = bound_pc - position;
-    offset += (offset & 0x8000) << 1;  // account for sign extension in jic/daddiu
-
-    // Embed it in the two instructions
-    auipc = (auipc & 0xFFFF0000) | (offset >> 16);
-    jic = (jic & 0xFFFF0000) | (offset & 0xFFFF);
-
-    // Save the adjusted instructions
-    buffer_.Store<uint32_t>(position, auipc);
-    buffer_.Store<uint32_t>(position + 4, jic);
+    uint32_t branch_location = branch->GetLocation();
+    // Extract the location of the previous branch in the list (walking the list backwards;
+    // the previous branch ID was stored in the space reserved for this branch).
+    uint32_t prev = buffer_.Load<uint32_t>(branch_location);
 
     // On to the previous branch in the list...
     label->position_ = prev;
   }
 
-  // Now make the label object contain its own location
-  // (it will be used by the branches referring to and following this label)
+  // Now make the label object contain its own location (relative to the end of the preceding
+  // branch, if any; it will be used by the branches referring to and following this label).
+  label->prev_branch_id_plus_one_ = branches_.size();
+  if (label->prev_branch_id_plus_one_) {
+    uint32_t branch_id = label->prev_branch_id_plus_one_ - 1;
+    const Branch* branch = GetBranch(branch_id);
+    bound_pc -= branch->GetEndLocation();
+  }
   label->BindTo(bound_pc);
 }
 
-void Mips64Assembler::B(Label* label) {
-  if (label->IsBound()) {
-    // Branch backwards (to a preceding label), distance is known
-    uint32_t offset = label->Position() - buffer_.Size();
-    CHECK_LE(static_cast<int32_t>(offset), 0);
-    offset += (offset & 0x8000) << 1;  // account for sign extension in jic
-    Auipc(AT, offset >> 16);
-    Jic(AT, offset);
-  } else {
-    // Branch forward (to a following label), distance is unknown
-    int32_t position = buffer_.Size();
-    // The first branch forward will have 0 in its pc-relative address (copied from label's
-    // position). It will be the terminator of the list of forward-reaching branches.
-    uint32_t prev = label->position_;
-    Auipc(AT, prev >> 16);
-    Jic(AT, prev);
-    // Now make the link object point to the location of this branch
-    // (this forms a linked list of branches preceding this label)
-    label->LinkTo(position);
+uint32_t Mips64Assembler::GetLabelLocation(Mips64Label* label) const {
+  CHECK(label->IsBound());
+  uint32_t target = label->Position();
+  if (label->prev_branch_id_plus_one_) {
+    // Get label location based on the branch preceding it.
+    uint32_t branch_id = label->prev_branch_id_plus_one_ - 1;
+    const Branch* branch = GetBranch(branch_id);
+    target += branch->GetEndLocation();
+  }
+  return target;
+}
+
+uint32_t Mips64Assembler::GetAdjustedPosition(uint32_t old_position) {
+  // We can reconstruct the adjustment by going through all the branches from the beginning
+  // up to the old_position. Since we expect AdjustedPosition() to be called in a loop
+  // with increasing old_position, we can use the data from last AdjustedPosition() to
+  // continue where we left off and the whole loop should be O(m+n) where m is the number
+  // of positions to adjust and n is the number of branches.
+  if (old_position < last_old_position_) {
+    last_position_adjustment_ = 0;
+    last_old_position_ = 0;
+    last_branch_id_ = 0;
+  }
+  while (last_branch_id_ != branches_.size()) {
+    const Branch* branch = GetBranch(last_branch_id_);
+    if (branch->GetLocation() >= old_position + last_position_adjustment_) {
+      break;
+    }
+    last_position_adjustment_ += branch->GetSize() - branch->GetOldSize();
+    ++last_branch_id_;
+  }
+  last_old_position_ = old_position;
+  return old_position + last_position_adjustment_;
+}
+
+void Mips64Assembler::FinalizeLabeledBranch(Mips64Label* label) {
+  uint32_t length = branches_.back().GetLength();
+  if (!label->IsBound()) {
+    // Branch forward (to a following label), distance is unknown.
+    // The first branch forward will contain 0, serving as the terminator of
+    // the list of forward-reaching branches.
+    Emit(label->position_);
+    length--;
+    // Now make the label object point to this branch
+    // (this forms a linked list of branches preceding this label).
+    uint32_t branch_id = branches_.size() - 1;
+    label->LinkTo(branch_id);
+  }
+  // Reserve space for the branch.
+  while (length--) {
+    Nop();
   }
 }
 
-void Mips64Assembler::Jalr(Label* label, GpuRegister indirect_reg) {
-  if (label->IsBound()) {
-    // Branch backwards (to a preceding label), distance is known
-    uint32_t offset = label->Position() - buffer_.Size();
-    CHECK_LE(static_cast<int32_t>(offset), 0);
-    offset += (offset & 0x8000) << 1;  // account for sign extension in daddiu
-    Auipc(indirect_reg, offset >> 16);
-    Daddiu(indirect_reg, indirect_reg, offset);
-    Jialc(indirect_reg, 0);
-  } else {
-    // Branch forward (to a following label), distance is unknown
-    int32_t position = buffer_.Size();
-    // The first branch forward will have 0 in its pc-relative address (copied from label's
-    // position). It will be the terminator of the list of forward-reaching branches.
-    uint32_t prev = label->position_;
-    Auipc(indirect_reg, prev >> 16);
-    Daddiu(indirect_reg, indirect_reg, prev);
-    Jialc(indirect_reg, 0);
-    // Now make the link object point to the location of this branch
-    // (this forms a linked list of branches preceding this label)
-    label->LinkTo(position);
+void Mips64Assembler::Buncond(Mips64Label* label) {
+  uint32_t target = label->IsBound() ? GetLabelLocation(label) : Branch::kUnresolved;
+  branches_.emplace_back(buffer_.Size(), target);
+  FinalizeLabeledBranch(label);
+}
+
+void Mips64Assembler::Bcond(Mips64Label* label,
+                            BranchCondition condition,
+                            GpuRegister lhs,
+                            GpuRegister rhs) {
+  // If lhs = rhs, this can be a NOP.
+  if (Branch::IsNop(condition, lhs, rhs)) {
+    return;
+  }
+  uint32_t target = label->IsBound() ? GetLabelLocation(label) : Branch::kUnresolved;
+  branches_.emplace_back(buffer_.Size(), target, condition, lhs, rhs);
+  FinalizeLabeledBranch(label);
+}
+
+void Mips64Assembler::Call(Mips64Label* label, GpuRegister indirect_reg) {
+  uint32_t target = label->IsBound() ? GetLabelLocation(label) : Branch::kUnresolved;
+  branches_.emplace_back(buffer_.Size(), target, indirect_reg);
+  FinalizeLabeledBranch(label);
+}
+
+void Mips64Assembler::PromoteBranches() {
+  // Promote short branches to long as necessary.
+  bool changed;
+  do {
+    changed = false;
+    for (auto& branch : branches_) {
+      CHECK(branch.IsResolved());
+      uint32_t delta = branch.PromoteIfNeeded();
+      // If this branch has been promoted and needs to expand in size,
+      // relocate all branches by the expansion size.
+      if (delta) {
+        changed = true;
+        uint32_t expand_location = branch.GetLocation();
+        for (auto& branch2 : branches_) {
+          branch2.Relocate(expand_location, delta);
+        }
+      }
+    }
+  } while (changed);
+
+  // Account for branch expansion by resizing the code buffer
+  // and moving the code in it to its final location.
+  size_t branch_count = branches_.size();
+  if (branch_count > 0) {
+    // Resize.
+    Branch& last_branch = branches_[branch_count - 1];
+    uint32_t size_delta = last_branch.GetEndLocation() - last_branch.GetOldEndLocation();
+    uint32_t old_size = buffer_.Size();
+    buffer_.Resize(old_size + size_delta);
+    // Move the code residing between branch placeholders.
+    uint32_t end = old_size;
+    for (size_t i = branch_count; i > 0; ) {
+      Branch& branch = branches_[--i];
+      uint32_t size = end - branch.GetOldEndLocation();
+      buffer_.Move(branch.GetEndLocation(), branch.GetOldEndLocation(), size);
+      end = branch.GetOldLocation();
+    }
   }
 }
 
-void Mips64Assembler::Bltc(GpuRegister rs, GpuRegister rt, Label* label) {
-  Bgec(rs, rt, 2);
-  B(label);
+// Note: make sure branch_info_[] and EmitBranch() are kept synchronized.
+const Mips64Assembler::Branch::BranchInfo Mips64Assembler::Branch::branch_info_[] = {
+  // Short branches.
+  {  1, 0, 1, Mips64Assembler::Branch::kOffset28, 2 },  // kUncondBranch
+  {  2, 0, 1, Mips64Assembler::Branch::kOffset18, 2 },  // kCondBranch
+                                                        // Exception: kOffset23 for beqzc/bnezc
+  {  2, 0, 0, Mips64Assembler::Branch::kOffset21, 2 },  // kCall
+  // Long branches.
+  {  2, 0, 0, Mips64Assembler::Branch::kOffset32, 0 },  // kLongUncondBranch
+  {  3, 1, 0, Mips64Assembler::Branch::kOffset32, 0 },  // kLongCondBranch
+  {  3, 0, 0, Mips64Assembler::Branch::kOffset32, 0 },  // kLongCall
+};
+
+// Note: make sure branch_info_[] and EmitBranch() are kept synchronized.
+void Mips64Assembler::EmitBranch(Mips64Assembler::Branch* branch) {
+  CHECK(overwriting_);
+  overwrite_location_ = branch->GetLocation();
+  uint32_t offset = branch->GetOffset();
+  BranchCondition condition = branch->GetCondition();
+  GpuRegister lhs = branch->GetLeftRegister();
+  GpuRegister rhs = branch->GetRightRegister();
+  switch (branch->GetType()) {
+    // Short branches.
+    case Branch::kUncondBranch:
+      CHECK_EQ(overwrite_location_, branch->GetOffsetLocation());
+      Bc(offset);
+      break;
+    case Branch::kCondBranch:
+      CHECK_EQ(overwrite_location_, branch->GetOffsetLocation());
+      EmitBcondc(condition, lhs, rhs, offset);
+      Nop();  // TODO: improve by filling the forbidden slot.
+      break;
+    case Branch::kCall:
+      CHECK_EQ(overwrite_location_, branch->GetOffsetLocation());
+      Addiupc(lhs, offset);
+      Jialc(lhs, 0);
+      break;
+
+    // Long branches.
+    case Branch::kLongUncondBranch:
+      offset += (offset & 0x8000) << 1;  // Account for sign extension in jic.
+      CHECK_EQ(overwrite_location_, branch->GetOffsetLocation());
+      Auipc(AT, High16Bits(offset));
+      Jic(AT, Low16Bits(offset));
+      break;
+    case Branch::kLongCondBranch:
+      EmitBcondc(Branch::OppositeCondition(condition), lhs, rhs, 2);
+      offset += (offset & 0x8000) << 1;  // Account for sign extension in jic.
+      CHECK_EQ(overwrite_location_, branch->GetOffsetLocation());
+      Auipc(AT, High16Bits(offset));
+      Jic(AT, Low16Bits(offset));
+      break;
+    case Branch::kLongCall:
+      offset += (offset & 0x8000) << 1;  // Account for sign extension in daddiu.
+      CHECK_EQ(overwrite_location_, branch->GetOffsetLocation());
+      Auipc(lhs, High16Bits(offset));
+      Daddiu(lhs, lhs, Low16Bits(offset));
+      Jialc(lhs, 0);
+      break;
+  }
+  CHECK_EQ(overwrite_location_, branch->GetEndLocation());
+  CHECK_LT(branch->GetSize(), static_cast<uint32_t>(Branch::kMaxBranchSize));
 }
 
-void Mips64Assembler::Bltzc(GpuRegister rt, Label* label) {
-  Bgezc(rt, 2);
-  B(label);
+void Mips64Assembler::Bc(Mips64Label* label) {
+  Buncond(label);
 }
 
-void Mips64Assembler::Bgtzc(GpuRegister rt, Label* label) {
-  Blezc(rt, 2);
-  B(label);
+void Mips64Assembler::Jialc(Mips64Label* label, GpuRegister indirect_reg) {
+  Call(label, indirect_reg);
 }
 
-void Mips64Assembler::Bgec(GpuRegister rs, GpuRegister rt, Label* label) {
-  Bltc(rs, rt, 2);
-  B(label);
+void Mips64Assembler::Bltc(GpuRegister rs, GpuRegister rt, Mips64Label* label) {
+  Bcond(label, kCondLT, rs, rt);
 }
 
-void Mips64Assembler::Bgezc(GpuRegister rt, Label* label) {
-  Bltzc(rt, 2);
-  B(label);
+void Mips64Assembler::Bltzc(GpuRegister rt, Mips64Label* label) {
+  Bcond(label, kCondLTZ, rt);
 }
 
-void Mips64Assembler::Blezc(GpuRegister rt, Label* label) {
-  Bgtzc(rt, 2);
-  B(label);
+void Mips64Assembler::Bgtzc(GpuRegister rt, Mips64Label* label) {
+  Bcond(label, kCondGTZ, rt);
 }
 
-void Mips64Assembler::Bltuc(GpuRegister rs, GpuRegister rt, Label* label) {
-  Bgeuc(rs, rt, 2);
-  B(label);
+void Mips64Assembler::Bgec(GpuRegister rs, GpuRegister rt, Mips64Label* label) {
+  Bcond(label, kCondGE, rs, rt);
 }
 
-void Mips64Assembler::Bgeuc(GpuRegister rs, GpuRegister rt, Label* label) {
-  Bltuc(rs, rt, 2);
-  B(label);
+void Mips64Assembler::Bgezc(GpuRegister rt, Mips64Label* label) {
+  Bcond(label, kCondGEZ, rt);
 }
 
-void Mips64Assembler::Beqc(GpuRegister rs, GpuRegister rt, Label* label) {
-  Bnec(rs, rt, 2);
-  B(label);
+void Mips64Assembler::Blezc(GpuRegister rt, Mips64Label* label) {
+  Bcond(label, kCondLEZ, rt);
 }
 
-void Mips64Assembler::Bnec(GpuRegister rs, GpuRegister rt, Label* label) {
-  Beqc(rs, rt, 2);
-  B(label);
+void Mips64Assembler::Bltuc(GpuRegister rs, GpuRegister rt, Mips64Label* label) {
+  Bcond(label, kCondLTU, rs, rt);
 }
 
-void Mips64Assembler::Beqzc(GpuRegister rs, Label* label) {
-  Bnezc(rs, 2);
-  B(label);
+void Mips64Assembler::Bgeuc(GpuRegister rs, GpuRegister rt, Mips64Label* label) {
+  Bcond(label, kCondGEU, rs, rt);
 }
 
-void Mips64Assembler::Bnezc(GpuRegister rs, Label* label) {
-  Beqzc(rs, 2);
-  B(label);
+void Mips64Assembler::Beqc(GpuRegister rs, GpuRegister rt, Mips64Label* label) {
+  Bcond(label, kCondEQ, rs, rt);
+}
+
+void Mips64Assembler::Bnec(GpuRegister rs, GpuRegister rt, Mips64Label* label) {
+  Bcond(label, kCondNE, rs, rt);
+}
+
+void Mips64Assembler::Beqzc(GpuRegister rs, Mips64Label* label) {
+  Bcond(label, kCondEQZ, rs);
+}
+
+void Mips64Assembler::Bnezc(GpuRegister rs, Mips64Label* label) {
+  Bcond(label, kCondNEZ, rs);
 }
 
 void Mips64Assembler::LoadFromOffset(LoadOperandType type, GpuRegister reg, GpuRegister base,
@@ -1256,6 +1797,7 @@
                                  const std::vector<ManagedRegister>& callee_save_regs,
                                  const ManagedRegisterEntrySpills& entry_spills) {
   CHECK_ALIGNED(frame_size, kStackAlignment);
+  DCHECK(!overwriting_);
 
   // Increase frame to required size.
   IncreaseFrameSize(frame_size);
@@ -1298,6 +1840,7 @@
 void Mips64Assembler::RemoveFrame(size_t frame_size,
                                   const std::vector<ManagedRegister>& callee_save_regs) {
   CHECK_ALIGNED(frame_size, kStackAlignment);
+  DCHECK(!overwriting_);
   cfi_.RememberState();
 
   // Pop callee saves and return address
@@ -1316,6 +1859,7 @@
 
   // Then jump to the return address.
   Jr(RA);
+  Nop();
 
   // The CFI should be restored for any code that follows the exit block.
   cfi_.RestoreState();
@@ -1324,12 +1868,14 @@
 
 void Mips64Assembler::IncreaseFrameSize(size_t adjust) {
   CHECK_ALIGNED(adjust, kFramePointerSize);
+  DCHECK(!overwriting_);
   Daddiu64(SP, SP, static_cast<int32_t>(-adjust));
   cfi_.AdjustCFAOffset(adjust);
 }
 
 void Mips64Assembler::DecreaseFrameSize(size_t adjust) {
   CHECK_ALIGNED(adjust, kFramePointerSize);
+  DCHECK(!overwriting_);
   Daddiu64(SP, SP, static_cast<int32_t>(adjust));
   cfi_.AdjustCFAOffset(-adjust);
 }
@@ -1379,17 +1925,7 @@
   StoreToOffset(kStoreWord, scratch.AsGpuRegister(), SP, dest.Int32Value());
 }
 
-void Mips64Assembler::StoreImmediateToThread64(ThreadOffset<8> dest, uint32_t imm,
-                                               ManagedRegister mscratch) {
-  Mips64ManagedRegister scratch = mscratch.AsMips64();
-  CHECK(scratch.IsGpuRegister()) << scratch;
-  // TODO: it's unclear wether 32 or 64 bits need to be stored (Arm64 and x86/x64 disagree?).
-  // Is this function even referenced anywhere else in the code?
-  LoadConst32(scratch.AsGpuRegister(), imm);
-  StoreToOffset(kStoreDoubleword, scratch.AsGpuRegister(), S1, dest.Int32Value());
-}
-
-void Mips64Assembler::StoreStackOffsetToThread64(ThreadOffset<8> thr_offs,
+void Mips64Assembler::StoreStackOffsetToThread64(ThreadOffset<kMipsDoublewordSize> thr_offs,
                                                  FrameOffset fr_offs,
                                                  ManagedRegister mscratch) {
   Mips64ManagedRegister scratch = mscratch.AsMips64();
@@ -1398,7 +1934,7 @@
   StoreToOffset(kStoreDoubleword, scratch.AsGpuRegister(), S1, thr_offs.Int32Value());
 }
 
-void Mips64Assembler::StoreStackPointerToThread64(ThreadOffset<8> thr_offs) {
+void Mips64Assembler::StoreStackPointerToThread64(ThreadOffset<kMipsDoublewordSize> thr_offs) {
   StoreToOffset(kStoreDoubleword, SP, S1, thr_offs.Int32Value());
 }
 
@@ -1415,7 +1951,9 @@
   return EmitLoad(mdest, SP, src.Int32Value(), size);
 }
 
-void Mips64Assembler::LoadFromThread64(ManagedRegister mdest, ThreadOffset<8> src, size_t size) {
+void Mips64Assembler::LoadFromThread64(ManagedRegister mdest,
+                                       ThreadOffset<kMipsDoublewordSize> src,
+                                       size_t size) {
   return EmitLoad(mdest, S1, src.Int32Value(), size);
 }
 
@@ -1449,18 +1987,20 @@
 }
 
 void Mips64Assembler::LoadRawPtrFromThread64(ManagedRegister mdest,
-                                             ThreadOffset<8> offs) {
+                                             ThreadOffset<kMipsDoublewordSize> offs) {
   Mips64ManagedRegister dest = mdest.AsMips64();
   CHECK(dest.IsGpuRegister());
   LoadFromOffset(kLoadDoubleword, dest.AsGpuRegister(), S1, offs.Int32Value());
 }
 
-void Mips64Assembler::SignExtend(ManagedRegister /*mreg*/, size_t /*size*/) {
-  UNIMPLEMENTED(FATAL) << "no sign extension necessary for mips";
+void Mips64Assembler::SignExtend(ManagedRegister mreg ATTRIBUTE_UNUSED,
+                                 size_t size ATTRIBUTE_UNUSED) {
+  UNIMPLEMENTED(FATAL) << "No sign extension necessary for MIPS64";
 }
 
-void Mips64Assembler::ZeroExtend(ManagedRegister /*mreg*/, size_t /*size*/) {
-  UNIMPLEMENTED(FATAL) << "no zero extension necessary for mips";
+void Mips64Assembler::ZeroExtend(ManagedRegister mreg ATTRIBUTE_UNUSED,
+                                 size_t size ATTRIBUTE_UNUSED) {
+  UNIMPLEMENTED(FATAL) << "No zero extension necessary for MIPS64";
 }
 
 void Mips64Assembler::Move(ManagedRegister mdest, ManagedRegister msrc, size_t size) {
@@ -1492,7 +2032,7 @@
 }
 
 void Mips64Assembler::CopyRawPtrFromThread64(FrameOffset fr_offs,
-                                             ThreadOffset<8> thr_offs,
+                                             ThreadOffset<kMipsDoublewordSize> thr_offs,
                                              ManagedRegister mscratch) {
   Mips64ManagedRegister scratch = mscratch.AsMips64();
   CHECK(scratch.IsGpuRegister()) << scratch;
@@ -1500,7 +2040,7 @@
   StoreToOffset(kStoreDoubleword, scratch.AsGpuRegister(), SP, fr_offs.Int32Value());
 }
 
-void Mips64Assembler::CopyRawPtrToThread64(ThreadOffset<8> thr_offs,
+void Mips64Assembler::CopyRawPtrToThread64(ThreadOffset<kMipsDoublewordSize> thr_offs,
                                            FrameOffset fr_offs,
                                            ManagedRegister mscratch) {
   Mips64ManagedRegister scratch = mscratch.AsMips64();
@@ -1561,9 +2101,12 @@
   }
 }
 
-void Mips64Assembler::Copy(FrameOffset /*dest*/, FrameOffset /*src_base*/, Offset /*src_offset*/,
-                         ManagedRegister /*mscratch*/, size_t /*size*/) {
-  UNIMPLEMENTED(FATAL) << "no mips64 implementation";
+void Mips64Assembler::Copy(FrameOffset dest ATTRIBUTE_UNUSED,
+                           FrameOffset src_base ATTRIBUTE_UNUSED,
+                           Offset src_offset ATTRIBUTE_UNUSED,
+                           ManagedRegister mscratch ATTRIBUTE_UNUSED,
+                           size_t size ATTRIBUTE_UNUSED) {
+  UNIMPLEMENTED(FATAL) << "No MIPS64 implementation";
 }
 
 void Mips64Assembler::Copy(ManagedRegister dest, Offset dest_offset,
@@ -1584,15 +2127,18 @@
   }
 }
 
-void Mips64Assembler::Copy(FrameOffset /*dest*/, Offset /*dest_offset*/, FrameOffset /*src*/, Offset
-/*src_offset*/,
-                         ManagedRegister /*mscratch*/, size_t /*size*/) {
-  UNIMPLEMENTED(FATAL) << "no mips64 implementation";
+void Mips64Assembler::Copy(FrameOffset dest ATTRIBUTE_UNUSED,
+                           Offset dest_offset ATTRIBUTE_UNUSED,
+                           FrameOffset src ATTRIBUTE_UNUSED,
+                           Offset src_offset ATTRIBUTE_UNUSED,
+                           ManagedRegister mscratch ATTRIBUTE_UNUSED,
+                           size_t size ATTRIBUTE_UNUSED) {
+  UNIMPLEMENTED(FATAL) << "No MIPS64 implementation";
 }
 
-void Mips64Assembler::MemoryBarrier(ManagedRegister) {
+void Mips64Assembler::MemoryBarrier(ManagedRegister mreg ATTRIBUTE_UNUSED) {
   // TODO: sync?
-  UNIMPLEMENTED(FATAL) << "no mips64 implementation";
+  UNIMPLEMENTED(FATAL) << "No MIPS64 implementation";
 }
 
 void Mips64Assembler::CreateHandleScopeEntry(ManagedRegister mout_reg,
@@ -1604,7 +2150,7 @@
   CHECK(in_reg.IsNoRegister() || in_reg.IsGpuRegister()) << in_reg;
   CHECK(out_reg.IsGpuRegister()) << out_reg;
   if (null_allowed) {
-    Label null_arg;
+    Mips64Label null_arg;
     // Null values get a handle scope entry value of 0.  Otherwise, the handle scope entry is
     // the address in the handle scope holding the reference.
     // e.g. out_reg = (handle == 0) ? 0 : (SP+handle_offset)
@@ -1631,7 +2177,7 @@
   Mips64ManagedRegister scratch = mscratch.AsMips64();
   CHECK(scratch.IsGpuRegister()) << scratch;
   if (null_allowed) {
-    Label null_arg;
+    Mips64Label null_arg;
     LoadFromOffset(kLoadUnsignedWord, scratch.AsGpuRegister(), SP,
                    handle_scope_offset.Int32Value());
     // Null values get a handle scope entry value of 0.  Otherwise, the handle scope entry is
@@ -1653,7 +2199,7 @@
   Mips64ManagedRegister in_reg = min_reg.AsMips64();
   CHECK(out_reg.IsGpuRegister()) << out_reg;
   CHECK(in_reg.IsGpuRegister()) << in_reg;
-  Label null_arg;
+  Mips64Label null_arg;
   if (!out_reg.Equals(in_reg)) {
     LoadConst32(out_reg.AsGpuRegister(), 0);
   }
@@ -1663,11 +2209,13 @@
   Bind(&null_arg);
 }
 
-void Mips64Assembler::VerifyObject(ManagedRegister /*src*/, bool /*could_be_null*/) {
+void Mips64Assembler::VerifyObject(ManagedRegister src ATTRIBUTE_UNUSED,
+                                   bool could_be_null ATTRIBUTE_UNUSED) {
   // TODO: not validating references
 }
 
-void Mips64Assembler::VerifyObject(FrameOffset /*src*/, bool /*could_be_null*/) {
+void Mips64Assembler::VerifyObject(FrameOffset src ATTRIBUTE_UNUSED,
+                                   bool could_be_null ATTRIBUTE_UNUSED) {
   // TODO: not validating references
 }
 
@@ -1679,6 +2227,7 @@
   LoadFromOffset(kLoadDoubleword, scratch.AsGpuRegister(),
                  base.AsGpuRegister(), offset.Int32Value());
   Jalr(scratch.AsGpuRegister());
+  Nop();
   // TODO: place reference map on call
 }
 
@@ -1691,11 +2240,13 @@
   LoadFromOffset(kLoadDoubleword, scratch.AsGpuRegister(),
                  scratch.AsGpuRegister(), offset.Int32Value());
   Jalr(scratch.AsGpuRegister());
+  Nop();
   // TODO: place reference map on call
 }
 
-void Mips64Assembler::CallFromThread64(ThreadOffset<8> /*offset*/, ManagedRegister /*mscratch*/) {
-  UNIMPLEMENTED(FATAL) << "no mips64 implementation";
+void Mips64Assembler::CallFromThread64(ThreadOffset<kMipsDoublewordSize> offset ATTRIBUTE_UNUSED,
+                                       ManagedRegister mscratch ATTRIBUTE_UNUSED) {
+  UNIMPLEMENTED(FATAL) << "No MIPS64 implementation";
 }
 
 void Mips64Assembler::GetCurrentThread(ManagedRegister tr) {
@@ -1703,37 +2254,39 @@
 }
 
 void Mips64Assembler::GetCurrentThread(FrameOffset offset,
-                                       ManagedRegister /*mscratch*/) {
+                                       ManagedRegister mscratch ATTRIBUTE_UNUSED) {
   StoreToOffset(kStoreDoubleword, S1, SP, offset.Int32Value());
 }
 
 void Mips64Assembler::ExceptionPoll(ManagedRegister mscratch, size_t stack_adjust) {
   Mips64ManagedRegister scratch = mscratch.AsMips64();
-  Mips64ExceptionSlowPath* slow = new Mips64ExceptionSlowPath(scratch, stack_adjust);
-  buffer_.EnqueueSlowPath(slow);
-  LoadFromOffset(kLoadDoubleword, scratch.AsGpuRegister(),
-                 S1, Thread::ExceptionOffset<8>().Int32Value());
-  Bnezc(scratch.AsGpuRegister(), slow->Entry());
+  exception_blocks_.emplace_back(scratch, stack_adjust);
+  LoadFromOffset(kLoadDoubleword,
+                 scratch.AsGpuRegister(),
+                 S1,
+                 Thread::ExceptionOffset<kMipsDoublewordSize>().Int32Value());
+  Bnezc(scratch.AsGpuRegister(), exception_blocks_.back().Entry());
 }
 
-void Mips64ExceptionSlowPath::Emit(Assembler* sasm) {
-  Mips64Assembler* sp_asm = down_cast<Mips64Assembler*>(sasm);
-#define __ sp_asm->
-  __ Bind(&entry_);
-  if (stack_adjust_ != 0) {  // Fix up the frame.
-    __ DecreaseFrameSize(stack_adjust_);
+void Mips64Assembler::EmitExceptionPoll(Mips64ExceptionSlowPath* exception) {
+  Bind(exception->Entry());
+  if (exception->stack_adjust_ != 0) {  // Fix up the frame.
+    DecreaseFrameSize(exception->stack_adjust_);
   }
-  // Pass exception object as argument
-  // Don't care about preserving A0 as this call won't return
-  __ Move(A0, scratch_.AsGpuRegister());
+  // Pass exception object as argument.
+  // Don't care about preserving A0 as this call won't return.
+  CheckEntrypointTypes<kQuickDeliverException, void, mirror::Object*>();
+  Move(A0, exception->scratch_.AsGpuRegister());
   // Set up call to Thread::Current()->pDeliverException
-  __ LoadFromOffset(kLoadDoubleword, T9, S1,
-                    QUICK_ENTRYPOINT_OFFSET(8, pDeliverException).Int32Value());
-  // TODO: check T9 usage
-  __ Jr(T9);
+  LoadFromOffset(kLoadDoubleword,
+                 T9,
+                 S1,
+                 QUICK_ENTRYPOINT_OFFSET(kMipsDoublewordSize, pDeliverException).Int32Value());
+  Jr(T9);
+  Nop();
+
   // Call never returns
-  __ Break();
-#undef __
+  Break();
 }
 
 }  // namespace mips64
diff --git a/compiler/utils/mips64/assembler_mips64.h b/compiler/utils/mips64/assembler_mips64.h
index 42962bc..57fc19a 100644
--- a/compiler/utils/mips64/assembler_mips64.h
+++ b/compiler/utils/mips64/assembler_mips64.h
@@ -17,18 +17,22 @@
 #ifndef ART_COMPILER_UTILS_MIPS64_ASSEMBLER_MIPS64_H_
 #define ART_COMPILER_UTILS_MIPS64_ASSEMBLER_MIPS64_H_
 
+#include <utility>
 #include <vector>
 
 #include "base/macros.h"
 #include "constants_mips64.h"
 #include "globals.h"
 #include "managed_register_mips64.h"
-#include "utils/assembler.h"
 #include "offsets.h"
+#include "utils/assembler.h"
+#include "utils/label.h"
 
 namespace art {
 namespace mips64 {
 
+static constexpr size_t kMipsDoublewordSize = 8;
+
 enum LoadOperandType {
   kLoadSignedByte,
   kLoadUnsignedByte,
@@ -60,10 +64,57 @@
   kPositiveZero      = 0x200,
 };
 
+class Mips64Label : public Label {
+ public:
+  Mips64Label() : prev_branch_id_plus_one_(0) {}
+
+  Mips64Label(Mips64Label&& src)
+      : Label(std::move(src)), prev_branch_id_plus_one_(src.prev_branch_id_plus_one_) {}
+
+ private:
+  uint32_t prev_branch_id_plus_one_;  // To get distance from preceding branch, if any.
+
+  friend class Mips64Assembler;
+  DISALLOW_COPY_AND_ASSIGN(Mips64Label);
+};
+
+// Slowpath entered when Thread::Current()->_exception is non-null.
+class Mips64ExceptionSlowPath {
+ public:
+  explicit Mips64ExceptionSlowPath(Mips64ManagedRegister scratch, size_t stack_adjust)
+      : scratch_(scratch), stack_adjust_(stack_adjust) {}
+
+  Mips64ExceptionSlowPath(Mips64ExceptionSlowPath&& src)
+      : scratch_(src.scratch_),
+        stack_adjust_(src.stack_adjust_),
+        exception_entry_(std::move(src.exception_entry_)) {}
+
+ private:
+  Mips64Label* Entry() { return &exception_entry_; }
+  const Mips64ManagedRegister scratch_;
+  const size_t stack_adjust_;
+  Mips64Label exception_entry_;
+
+  friend class Mips64Assembler;
+  DISALLOW_COPY_AND_ASSIGN(Mips64ExceptionSlowPath);
+};
+
 class Mips64Assembler FINAL : public Assembler {
  public:
-  Mips64Assembler() {}
-  virtual ~Mips64Assembler() {}
+  Mips64Assembler()
+      : overwriting_(false),
+        overwrite_location_(0),
+        last_position_adjustment_(0),
+        last_old_position_(0),
+        last_branch_id_(0) {
+    cfi().DelayEmittingAdvancePCs();
+  }
+
+  virtual ~Mips64Assembler() {
+    for (auto& branch : branches_) {
+      CHECK(branch.IsResolved());
+    }
+  }
 
   // Emit Machine Instructions.
   void Addu(GpuRegister rd, GpuRegister rs, GpuRegister rt);
@@ -156,14 +207,12 @@
   void Dclz(GpuRegister rd, GpuRegister rs);
   void Dclo(GpuRegister rd, GpuRegister rs);
 
-  void Beq(GpuRegister rs, GpuRegister rt, uint16_t imm16);
-  void Bne(GpuRegister rs, GpuRegister rt, uint16_t imm16);
-  void J(uint32_t addr26);
-  void Jal(uint32_t addr26);
   void Jalr(GpuRegister rd, GpuRegister rs);
   void Jalr(GpuRegister rs);
   void Jr(GpuRegister rs);
   void Auipc(GpuRegister rs, uint16_t imm16);
+  void Addiupc(GpuRegister rs, uint32_t imm19);
+  void Bc(uint32_t imm26);
   void Jic(GpuRegister rt, uint16_t imm16);
   void Jialc(GpuRegister rt, uint16_t imm16);
   void Bltc(GpuRegister rs, GpuRegister rt, uint16_t imm16);
@@ -240,32 +289,34 @@
   void Clear(GpuRegister rd);
   void Not(GpuRegister rd, GpuRegister rs);
 
-  // Higher level composite instructions
+  // Higher level composite instructions.
   void LoadConst32(GpuRegister rd, int32_t value);
   void LoadConst64(GpuRegister rd, int64_t value);  // MIPS64
 
-  void Addiu32(GpuRegister rt, GpuRegister rs, int32_t value, GpuRegister rtmp = AT);
   void Daddiu64(GpuRegister rt, GpuRegister rs, int64_t value, GpuRegister rtmp = AT);  // MIPS64
 
-  void Bind(Label* label) OVERRIDE;
-  void Jump(Label* label) OVERRIDE {
-    B(label);
+  void Bind(Label* label) OVERRIDE {
+    Bind(down_cast<Mips64Label*>(label));
   }
-  void B(Label* label);
-  void Jalr(Label* label, GpuRegister indirect_reg = RA);
-  // TODO: implement common for R6 and non-R6 interface for conditional branches?
-  void Bltc(GpuRegister rs, GpuRegister rt, Label* label);
-  void Bltzc(GpuRegister rt, Label* label);
-  void Bgtzc(GpuRegister rt, Label* label);
-  void Bgec(GpuRegister rs, GpuRegister rt, Label* label);
-  void Bgezc(GpuRegister rt, Label* label);
-  void Blezc(GpuRegister rt, Label* label);
-  void Bltuc(GpuRegister rs, GpuRegister rt, Label* label);
-  void Bgeuc(GpuRegister rs, GpuRegister rt, Label* label);
-  void Beqc(GpuRegister rs, GpuRegister rt, Label* label);
-  void Bnec(GpuRegister rs, GpuRegister rt, Label* label);
-  void Beqzc(GpuRegister rs, Label* label);
-  void Bnezc(GpuRegister rs, Label* label);
+  void Jump(Label* label ATTRIBUTE_UNUSED) OVERRIDE {
+    UNIMPLEMENTED(FATAL) << "Do not use Jump for MIPS64";
+  }
+
+  void Bind(Mips64Label* label);
+  void Bc(Mips64Label* label);
+  void Jialc(Mips64Label* label, GpuRegister indirect_reg);
+  void Bltc(GpuRegister rs, GpuRegister rt, Mips64Label* label);
+  void Bltzc(GpuRegister rt, Mips64Label* label);
+  void Bgtzc(GpuRegister rt, Mips64Label* label);
+  void Bgec(GpuRegister rs, GpuRegister rt, Mips64Label* label);
+  void Bgezc(GpuRegister rt, Mips64Label* label);
+  void Blezc(GpuRegister rt, Mips64Label* label);
+  void Bltuc(GpuRegister rs, GpuRegister rt, Mips64Label* label);
+  void Bgeuc(GpuRegister rs, GpuRegister rt, Mips64Label* label);
+  void Beqc(GpuRegister rs, GpuRegister rt, Mips64Label* label);
+  void Bnec(GpuRegister rs, GpuRegister rt, Mips64Label* label);
+  void Beqzc(GpuRegister rs, Mips64Label* label);
+  void Bnezc(GpuRegister rs, Mips64Label* label);
 
   void EmitLoad(ManagedRegister m_dst, GpuRegister src_register, int32_t src_offset, size_t size);
   void LoadFromOffset(LoadOperandType type, GpuRegister reg, GpuRegister base, int32_t offset);
@@ -277,43 +328,42 @@
   void Emit(uint32_t value);
 
   //
-  // Overridden common assembler high-level functionality
+  // Overridden common assembler high-level functionality.
   //
 
-  // Emit code that will create an activation on the stack
+  // Emit code that will create an activation on the stack.
   void BuildFrame(size_t frame_size, ManagedRegister method_reg,
                   const std::vector<ManagedRegister>& callee_save_regs,
                   const ManagedRegisterEntrySpills& entry_spills) OVERRIDE;
 
-  // Emit code that will remove an activation from the stack
+  // Emit code that will remove an activation from the stack.
   void RemoveFrame(size_t frame_size,
                    const std::vector<ManagedRegister>& callee_save_regs) OVERRIDE;
 
   void IncreaseFrameSize(size_t adjust) OVERRIDE;
   void DecreaseFrameSize(size_t adjust) OVERRIDE;
 
-  // Store routines
+  // Store routines.
   void Store(FrameOffset offs, ManagedRegister msrc, size_t size) OVERRIDE;
   void StoreRef(FrameOffset dest, ManagedRegister msrc) OVERRIDE;
   void StoreRawPtr(FrameOffset dest, ManagedRegister msrc) OVERRIDE;
 
   void StoreImmediateToFrame(FrameOffset dest, uint32_t imm, ManagedRegister mscratch) OVERRIDE;
 
-  void StoreImmediateToThread64(ThreadOffset<8> dest, uint32_t imm,
-                                ManagedRegister mscratch) OVERRIDE;
-
-  void StoreStackOffsetToThread64(ThreadOffset<8> thr_offs, FrameOffset fr_offs,
+  void StoreStackOffsetToThread64(ThreadOffset<kMipsDoublewordSize> thr_offs, FrameOffset fr_offs,
                                   ManagedRegister mscratch) OVERRIDE;
 
-  void StoreStackPointerToThread64(ThreadOffset<8> thr_offs) OVERRIDE;
+  void StoreStackPointerToThread64(ThreadOffset<kMipsDoublewordSize> thr_offs) OVERRIDE;
 
   void StoreSpanning(FrameOffset dest, ManagedRegister msrc, FrameOffset in_off,
                      ManagedRegister mscratch) OVERRIDE;
 
-  // Load routines
+  // Load routines.
   void Load(ManagedRegister mdest, FrameOffset src, size_t size) OVERRIDE;
 
-  void LoadFromThread64(ManagedRegister mdest, ThreadOffset<8> src, size_t size) OVERRIDE;
+  void LoadFromThread64(ManagedRegister mdest,
+                        ThreadOffset<kMipsDoublewordSize> src,
+                        size_t size) OVERRIDE;
 
   void LoadRef(ManagedRegister dest, FrameOffset src) OVERRIDE;
 
@@ -322,15 +372,16 @@
 
   void LoadRawPtr(ManagedRegister mdest, ManagedRegister base, Offset offs) OVERRIDE;
 
-  void LoadRawPtrFromThread64(ManagedRegister mdest, ThreadOffset<8> offs) OVERRIDE;
+  void LoadRawPtrFromThread64(ManagedRegister mdest,
+                              ThreadOffset<kMipsDoublewordSize> offs) OVERRIDE;
 
-  // Copying routines
+  // Copying routines.
   void Move(ManagedRegister mdest, ManagedRegister msrc, size_t size) OVERRIDE;
 
-  void CopyRawPtrFromThread64(FrameOffset fr_offs, ThreadOffset<8> thr_offs,
+  void CopyRawPtrFromThread64(FrameOffset fr_offs, ThreadOffset<kMipsDoublewordSize> thr_offs,
                               ManagedRegister mscratch) OVERRIDE;
 
-  void CopyRawPtrToThread64(ThreadOffset<8> thr_offs, FrameOffset fr_offs,
+  void CopyRawPtrToThread64(ThreadOffset<kMipsDoublewordSize> thr_offs, FrameOffset fr_offs,
                             ManagedRegister mscratch) OVERRIDE;
 
   void CopyRef(FrameOffset dest, FrameOffset src, ManagedRegister mscratch) OVERRIDE;
@@ -354,13 +405,13 @@
 
   void MemoryBarrier(ManagedRegister) OVERRIDE;
 
-  // Sign extension
+  // Sign extension.
   void SignExtend(ManagedRegister mreg, size_t size) OVERRIDE;
 
-  // Zero extension
+  // Zero extension.
   void ZeroExtend(ManagedRegister mreg, size_t size) OVERRIDE;
 
-  // Exploit fast access in managed code to Thread::Current()
+  // Exploit fast access in managed code to Thread::Current().
   void GetCurrentThread(ManagedRegister tr) OVERRIDE;
   void GetCurrentThread(FrameOffset dest_offset, ManagedRegister mscratch) OVERRIDE;
 
@@ -376,7 +427,7 @@
   void CreateHandleScopeEntry(FrameOffset out_off, FrameOffset handlescope_offset, ManagedRegister
                               mscratch, bool null_allowed) OVERRIDE;
 
-  // src holds a handle scope entry (Object**) load this into dst
+  // src holds a handle scope entry (Object**) load this into dst.
   void LoadReferenceFromHandleScope(ManagedRegister dst, ManagedRegister src) OVERRIDE;
 
   // Heap::VerifyObject on src. In some cases (such as a reference to this) we
@@ -384,39 +435,255 @@
   void VerifyObject(ManagedRegister src, bool could_be_null) OVERRIDE;
   void VerifyObject(FrameOffset src, bool could_be_null) OVERRIDE;
 
-  // Call to address held at [base+offset]
+  // Call to address held at [base+offset].
   void Call(ManagedRegister base, Offset offset, ManagedRegister mscratch) OVERRIDE;
   void Call(FrameOffset base, Offset offset, ManagedRegister mscratch) OVERRIDE;
-  void CallFromThread64(ThreadOffset<8> offset, ManagedRegister mscratch) OVERRIDE;
+  void CallFromThread64(ThreadOffset<kMipsDoublewordSize> offset,
+                        ManagedRegister mscratch) OVERRIDE;
 
   // Generate code to check if Thread::Current()->exception_ is non-null
   // and branch to a ExceptionSlowPath if it is.
   void ExceptionPoll(ManagedRegister mscratch, size_t stack_adjust) OVERRIDE;
 
+  // Emit slow paths queued during assembly and promote short branches to long if needed.
+  void FinalizeCode() OVERRIDE;
+
+  // Emit branches and finalize all instructions.
+  void FinalizeInstructions(const MemoryRegion& region);
+
+  // Returns the (always-)current location of a label (can be used in class CodeGeneratorMIPS64,
+  // must be used instead of Mips64Label::GetPosition()).
+  uint32_t GetLabelLocation(Mips64Label* label) const;
+
+  // Get the final position of a label after local fixup based on the old position
+  // recorded before FinalizeCode().
+  uint32_t GetAdjustedPosition(uint32_t old_position);
+
+  enum BranchCondition {
+    kCondLT,
+    kCondGE,
+    kCondLE,
+    kCondGT,
+    kCondLTZ,
+    kCondGEZ,
+    kCondLEZ,
+    kCondGTZ,
+    kCondEQ,
+    kCondNE,
+    kCondEQZ,
+    kCondNEZ,
+    kCondLTU,
+    kCondGEU,
+    kUncond,
+  };
+  friend std::ostream& operator<<(std::ostream& os, const BranchCondition& rhs);
+
  private:
+  class Branch {
+   public:
+    enum Type {
+      // Short branches.
+      kUncondBranch,
+      kCondBranch,
+      kCall,
+      // Long branches.
+      kLongUncondBranch,
+      kLongCondBranch,
+      kLongCall,
+    };
+
+    // Bit sizes of offsets defined as enums to minimize chance of typos.
+    enum OffsetBits {
+      kOffset16 = 16,
+      kOffset18 = 18,
+      kOffset21 = 21,
+      kOffset23 = 23,
+      kOffset28 = 28,
+      kOffset32 = 32,
+    };
+
+    static constexpr uint32_t kUnresolved = 0xffffffff;  // Unresolved target_
+    static constexpr int32_t kMaxBranchLength = 32;
+    static constexpr int32_t kMaxBranchSize = kMaxBranchLength * sizeof(uint32_t);
+
+    struct BranchInfo {
+      // Branch length as a number of 4-byte-long instructions.
+      uint32_t length;
+      // Ordinal number (0-based) of the first (or the only) instruction that contains the branch's
+      // PC-relative offset (or its most significant 16-bit half, which goes first).
+      uint32_t instr_offset;
+      // Different MIPS instructions with PC-relative offsets apply said offsets to slightly
+      // different origins, e.g. to PC or PC+4. Encode the origin distance (as a number of 4-byte
+      // instructions) from the instruction containing the offset.
+      uint32_t pc_org;
+      // How large (in bits) a PC-relative offset can be for a given type of branch (kCondBranch is
+      // an exception: use kOffset23 for beqzc/bnezc).
+      OffsetBits offset_size;
+      // Some MIPS instructions with PC-relative offsets shift the offset by 2. Encode the shift
+      // count.
+      int offset_shift;
+    };
+    static const BranchInfo branch_info_[/* Type */];
+
+    // Unconditional branch.
+    Branch(uint32_t location, uint32_t target);
+    // Conditional branch.
+    Branch(uint32_t location,
+           uint32_t target,
+           BranchCondition condition,
+           GpuRegister lhs_reg,
+           GpuRegister rhs_reg = ZERO);
+    // Call (branch and link) that stores the target address in a given register (i.e. T9).
+    Branch(uint32_t location, uint32_t target, GpuRegister indirect_reg);
+
+    // Some conditional branches with lhs = rhs are effectively NOPs, while some
+    // others are effectively unconditional. MIPSR6 conditional branches require lhs != rhs.
+    // So, we need a way to identify such branches in order to emit no instructions for them
+    // or change them to unconditional.
+    static bool IsNop(BranchCondition condition, GpuRegister lhs, GpuRegister rhs);
+    static bool IsUncond(BranchCondition condition, GpuRegister lhs, GpuRegister rhs);
+
+    static BranchCondition OppositeCondition(BranchCondition cond);
+
+    Type GetType() const;
+    BranchCondition GetCondition() const;
+    GpuRegister GetLeftRegister() const;
+    GpuRegister GetRightRegister() const;
+    uint32_t GetTarget() const;
+    uint32_t GetLocation() const;
+    uint32_t GetOldLocation() const;
+    uint32_t GetLength() const;
+    uint32_t GetOldLength() const;
+    uint32_t GetSize() const;
+    uint32_t GetOldSize() const;
+    uint32_t GetEndLocation() const;
+    uint32_t GetOldEndLocation() const;
+    bool IsLong() const;
+    bool IsResolved() const;
+
+    // Returns the bit size of the signed offset that the branch instruction can handle.
+    OffsetBits GetOffsetSize() const;
+
+    // Calculates the distance between two byte locations in the assembler buffer and
+    // returns the number of bits needed to represent the distance as a signed integer.
+    //
+    // Branch instructions have signed offsets of 16, 19 (addiupc), 21 (beqzc/bnezc),
+    // and 26 (bc) bits, which are additionally shifted left 2 positions at run time.
+    //
+    // Composite branches (made of several instructions) with longer reach have 32-bit
+    // offsets encoded as 2 16-bit "halves" in two instructions (high half goes first).
+    // The composite branches cover the range of PC + ~+/-2GB. The range is not end-to-end,
+    // however. Consider the following implementation of a long unconditional branch, for
+    // example:
+    //
+    //   auipc at, offset_31_16  // at = pc + sign_extend(offset_31_16) << 16
+    //   jic   at, offset_15_0   // pc = at + sign_extend(offset_15_0)
+    //
+    // Both of the above instructions take 16-bit signed offsets as immediate operands.
+    // When bit 15 of offset_15_0 is 1, it effectively causes subtraction of 0x10000
+    // due to sign extension. This must be compensated for by incrementing offset_31_16
+    // by 1. offset_31_16 can only be incremented by 1 if it's not 0x7FFF. If it is
+    // 0x7FFF, adding 1 will overflow the positive offset into the negative range.
+    // Therefore, the long branch range is something like from PC - 0x80000000 to
+    // PC + 0x7FFF7FFF, IOW, shorter by 32KB on one side.
+    //
+    // The returned values are therefore: 18, 21, 23, 28 and 32. There's also a special
+    // case with the addiu instruction and a 16 bit offset.
+    static OffsetBits GetOffsetSizeNeeded(uint32_t location, uint32_t target);
+
+    // Resolve a branch when the target is known.
+    void Resolve(uint32_t target);
+
+    // Relocate a branch by a given delta if needed due to expansion of this or another
+    // branch at a given location by this delta (just changes location_ and target_).
+    void Relocate(uint32_t expand_location, uint32_t delta);
+
+    // If the branch is short, changes its type to long.
+    void PromoteToLong();
+
+    // If necessary, updates the type by promoting a short branch to a long branch
+    // based on the branch location and target. Returns the amount (in bytes) by
+    // which the branch size has increased.
+    // max_short_distance caps the maximum distance between location_ and target_
+    // that is allowed for short branches. This is for debugging/testing purposes.
+    // max_short_distance = 0 forces all short branches to become long.
+    // Use the implicit default argument when not debugging/testing.
+    uint32_t PromoteIfNeeded(uint32_t max_short_distance = std::numeric_limits<uint32_t>::max());
+
+    // Returns the location of the instruction(s) containing the offset.
+    uint32_t GetOffsetLocation() const;
+
+    // Calculates and returns the offset ready for encoding in the branch instruction(s).
+    uint32_t GetOffset() const;
+
+   private:
+    // Completes branch construction by determining and recording its type.
+    void InitializeType(bool is_call);
+    // Helper for the above.
+    void InitShortOrLong(OffsetBits ofs_size, Type short_type, Type long_type);
+
+    uint32_t old_location_;      // Offset into assembler buffer in bytes.
+    uint32_t location_;          // Offset into assembler buffer in bytes.
+    uint32_t target_;            // Offset into assembler buffer in bytes.
+
+    GpuRegister lhs_reg_;        // Left-hand side register in conditional branches or
+                                 // indirect call register.
+    GpuRegister rhs_reg_;        // Right-hand side register in conditional branches.
+    BranchCondition condition_;  // Condition for conditional branches.
+
+    Type type_;                  // Current type of the branch.
+    Type old_type_;              // Initial type of the branch.
+  };
+  friend std::ostream& operator<<(std::ostream& os, const Branch::Type& rhs);
+  friend std::ostream& operator<<(std::ostream& os, const Branch::OffsetBits& rhs);
+
   void EmitR(int opcode, GpuRegister rs, GpuRegister rt, GpuRegister rd, int shamt, int funct);
   void EmitRsd(int opcode, GpuRegister rs, GpuRegister rd, int shamt, int funct);
   void EmitRtd(int opcode, GpuRegister rt, GpuRegister rd, int shamt, int funct);
   void EmitI(int opcode, GpuRegister rs, GpuRegister rt, uint16_t imm);
   void EmitI21(int opcode, GpuRegister rs, uint32_t imm21);
-  void EmitJ(int opcode, uint32_t addr26);
+  void EmitI26(int opcode, uint32_t imm26);
   void EmitFR(int opcode, int fmt, FpuRegister ft, FpuRegister fs, FpuRegister fd, int funct);
   void EmitFI(int opcode, int fmt, FpuRegister rt, uint16_t imm);
+  void EmitBcondc(BranchCondition cond, GpuRegister rs, GpuRegister rt, uint32_t imm16_21);
+
+  void Buncond(Mips64Label* label);
+  void Bcond(Mips64Label* label,
+             BranchCondition condition,
+             GpuRegister lhs,
+             GpuRegister rhs = ZERO);
+  void Call(Mips64Label* label, GpuRegister indirect_reg);
+  void FinalizeLabeledBranch(Mips64Label* label);
+
+  Branch* GetBranch(uint32_t branch_id);
+  const Branch* GetBranch(uint32_t branch_id) const;
+
+  void PromoteBranches();
+  void EmitBranch(Branch* branch);
+  void EmitBranches();
+  void PatchCFI();
+
+  // Emits exception block.
+  void EmitExceptionPoll(Mips64ExceptionSlowPath* exception);
+
+  // List of exception blocks to generate at the end of the code cache.
+  std::vector<Mips64ExceptionSlowPath> exception_blocks_;
+
+  std::vector<Branch> branches_;
+
+  // Whether appending instructions at the end of the buffer or overwriting the existing ones.
+  bool overwriting_;
+  // The current overwrite location.
+  uint32_t overwrite_location_;
+
+  // Data for AdjustedPosition(), see the description there.
+  uint32_t last_position_adjustment_;
+  uint32_t last_old_position_;
+  uint32_t last_branch_id_;
 
   DISALLOW_COPY_AND_ASSIGN(Mips64Assembler);
 };
 
-// Slowpath entered when Thread::Current()->_exception is non-null
-class Mips64ExceptionSlowPath FINAL : public SlowPath {
- public:
-  Mips64ExceptionSlowPath(Mips64ManagedRegister scratch, size_t stack_adjust)
-      : scratch_(scratch), stack_adjust_(stack_adjust) {}
-  virtual void Emit(Assembler *sp_asm) OVERRIDE;
- private:
-  const Mips64ManagedRegister scratch_;
-  const size_t stack_adjust_;
-};
-
 }  // namespace mips64
 }  // namespace art
 
diff --git a/compiler/utils/mips64/assembler_mips64_test.cc b/compiler/utils/mips64/assembler_mips64_test.cc
index 4413906..29a5a88 100644
--- a/compiler/utils/mips64/assembler_mips64_test.cc
+++ b/compiler/utils/mips64/assembler_mips64_test.cc
@@ -24,6 +24,8 @@
 #include "base/stl_util.h"
 #include "utils/assembler_test.h"
 
+#define __ GetAssembler()->
+
 namespace art {
 
 struct MIPS64CpuRegisterCompare {
@@ -48,8 +50,26 @@
     return "mips64";
   }
 
+  std::string GetAssemblerCmdName() OVERRIDE {
+    // We assemble and link for MIPS64R6. See GetAssemblerParameters() for details.
+    return "gcc";
+  }
+
   std::string GetAssemblerParameters() OVERRIDE {
-    return " --no-warn -march=mips64r6";
+    // We assemble and link for MIPS64R6. The reason is that object files produced for MIPS64R6
+    // (and MIPS32R6) with the GNU assembler don't have correct final offsets in PC-relative
+    // branches in the .text section and so they require a relocation pass (there's a relocation
+    // section, .rela.text, that has the needed info to fix up the branches).
+    return " -march=mips64r6 -Wa,--no-warn -Wl,-Ttext=0 -Wl,-e0 -nostdlib";
+  }
+
+  void Pad(std::vector<uint8_t>& data) OVERRIDE {
+    // The GNU linker unconditionally pads the code segment with NOPs to a size that is a multiple
+    // of 16 and there doesn't appear to be a way to suppress this padding. Our assembler doesn't
+    // pad, so, in order for two assembler outputs to match, we need to match the padding as well.
+    // NOP is encoded as four zero bytes on MIPS.
+    size_t pad_size = RoundUp(data.size(), 16u) - data.size();
+    data.insert(data.end(), pad_size, 0);
   }
 
   std::string GetDisassembleParameters() OVERRIDE {
@@ -182,6 +202,71 @@
     return secondary_register_names_[reg];
   }
 
+  std::string RepeatInsn(size_t count, const std::string& insn) {
+    std::string result;
+    for (; count != 0u; --count) {
+      result += insn;
+    }
+    return result;
+  }
+
+  void BranchCondOneRegHelper(void (mips64::Mips64Assembler::*f)(mips64::GpuRegister,
+                                                                 mips64::Mips64Label*),
+                              std::string instr_name) {
+    mips64::Mips64Label label;
+    (Base::GetAssembler()->*f)(mips64::A0, &label);
+    constexpr size_t kAdduCount1 = 63;
+    for (size_t i = 0; i != kAdduCount1; ++i) {
+      __ Addu(mips64::ZERO, mips64::ZERO, mips64::ZERO);
+    }
+    __ Bind(&label);
+    constexpr size_t kAdduCount2 = 64;
+    for (size_t i = 0; i != kAdduCount2; ++i) {
+      __ Addu(mips64::ZERO, mips64::ZERO, mips64::ZERO);
+    }
+    (Base::GetAssembler()->*f)(mips64::A1, &label);
+
+    std::string expected =
+        ".set noreorder\n" +
+        instr_name + " $a0, 1f\n"
+        "nop\n" +
+        RepeatInsn(kAdduCount1, "addu $zero, $zero, $zero\n") +
+        "1:\n" +
+        RepeatInsn(kAdduCount2, "addu $zero, $zero, $zero\n") +
+        instr_name + " $a1, 1b\n"
+        "nop\n";
+    DriverStr(expected, instr_name);
+  }
+
+  void BranchCondTwoRegsHelper(void (mips64::Mips64Assembler::*f)(mips64::GpuRegister,
+                                                                  mips64::GpuRegister,
+                                                                  mips64::Mips64Label*),
+                               std::string instr_name) {
+    mips64::Mips64Label label;
+    (Base::GetAssembler()->*f)(mips64::A0, mips64::A1, &label);
+    constexpr size_t kAdduCount1 = 63;
+    for (size_t i = 0; i != kAdduCount1; ++i) {
+      __ Addu(mips64::ZERO, mips64::ZERO, mips64::ZERO);
+    }
+    __ Bind(&label);
+    constexpr size_t kAdduCount2 = 64;
+    for (size_t i = 0; i != kAdduCount2; ++i) {
+      __ Addu(mips64::ZERO, mips64::ZERO, mips64::ZERO);
+    }
+    (Base::GetAssembler()->*f)(mips64::A2, mips64::A3, &label);
+
+    std::string expected =
+        ".set noreorder\n" +
+        instr_name + " $a0, $a1, 1f\n"
+        "nop\n" +
+        RepeatInsn(kAdduCount1, "addu $zero, $zero, $zero\n") +
+        "1:\n" +
+        RepeatInsn(kAdduCount2, "addu $zero, $zero, $zero\n") +
+        instr_name + " $a2, $a3, 1b\n"
+        "nop\n";
+    DriverStr(expected, instr_name);
+  }
+
  private:
   std::vector<mips64::GpuRegister*> registers_;
   std::map<mips64::GpuRegister, std::string, MIPS64CpuRegisterCompare> secondary_register_names_;
@@ -194,7 +279,6 @@
   EXPECT_TRUE(CheckTools());
 }
 
-
 ///////////////////
 // FP Operations //
 ///////////////////
@@ -348,7 +432,203 @@
 ////////////////
 
 TEST_F(AssemblerMIPS64Test, Jalr) {
-  DriverStr(RepeatRRNoDupes(&mips64::Mips64Assembler::Jalr, "jalr ${reg1}, ${reg2}"), "jalr");
+  DriverStr(".set noreorder\n" +
+            RepeatRRNoDupes(&mips64::Mips64Assembler::Jalr, "jalr ${reg1}, ${reg2}"), "jalr");
+}
+
+TEST_F(AssemblerMIPS64Test, Jialc) {
+  mips64::Mips64Label label1, label2;
+  __ Jialc(&label1, mips64::T9);
+  constexpr size_t kAdduCount1 = 63;
+  for (size_t i = 0; i != kAdduCount1; ++i) {
+    __ Addu(mips64::ZERO, mips64::ZERO, mips64::ZERO);
+  }
+  __ Bind(&label1);
+  __ Jialc(&label2, mips64::T9);
+  constexpr size_t kAdduCount2 = 64;
+  for (size_t i = 0; i != kAdduCount2; ++i) {
+    __ Addu(mips64::ZERO, mips64::ZERO, mips64::ZERO);
+  }
+  __ Bind(&label2);
+  __ Jialc(&label1, mips64::T9);
+
+  std::string expected =
+      ".set noreorder\n"
+      "lapc $t9, 1f\n"
+      "jialc $t9, 0\n" +
+      RepeatInsn(kAdduCount1, "addu $zero, $zero, $zero\n") +
+      "1:\n"
+      "lapc $t9, 2f\n"
+      "jialc $t9, 0\n" +
+      RepeatInsn(kAdduCount2, "addu $zero, $zero, $zero\n") +
+      "2:\n"
+      "lapc $t9, 1b\n"
+      "jialc $t9, 0\n";
+  DriverStr(expected, "Jialc");
+}
+
+TEST_F(AssemblerMIPS64Test, LongJialc) {
+  mips64::Mips64Label label1, label2;
+  __ Jialc(&label1, mips64::T9);
+  constexpr uint32_t kAdduCount1 = (1u << 18) + 1;
+  for (uint32_t i = 0; i != kAdduCount1; ++i) {
+    __ Addu(mips64::ZERO, mips64::ZERO, mips64::ZERO);
+  }
+  __ Bind(&label1);
+  __ Jialc(&label2, mips64::T9);
+  constexpr uint32_t kAdduCount2 = (1u << 18) + 1;
+  for (uint32_t i = 0; i != kAdduCount2; ++i) {
+    __ Addu(mips64::ZERO, mips64::ZERO, mips64::ZERO);
+  }
+  __ Bind(&label2);
+  __ Jialc(&label1, mips64::T9);
+
+  uint32_t offset_forward1 = 3 + kAdduCount1;  // 3: account for auipc, daddiu and jic.
+  offset_forward1 <<= 2;
+  offset_forward1 += (offset_forward1 & 0x8000) << 1;  // Account for sign extension in daddiu.
+
+  uint32_t offset_forward2 = 3 + kAdduCount2;  // 3: account for auipc, daddiu and jic.
+  offset_forward2 <<= 2;
+  offset_forward2 += (offset_forward2 & 0x8000) << 1;  // Account for sign extension in daddiu.
+
+  uint32_t offset_back = -(3 + kAdduCount2);  // 3: account for auipc, daddiu and jic.
+  offset_back <<= 2;
+  offset_back += (offset_back & 0x8000) << 1;  // Account for sign extension in daddiu.
+
+  std::ostringstream oss;
+  oss <<
+      ".set noreorder\n"
+      "auipc $t9, 0x" << std::hex << High16Bits(offset_forward1) << "\n"
+      "daddiu $t9, 0x" << std::hex << Low16Bits(offset_forward1) << "\n"
+      "jialc $t9, 0\n" <<
+      RepeatInsn(kAdduCount1, "addu $zero, $zero, $zero\n") <<
+      "1:\n"
+      "auipc $t9, 0x" << std::hex << High16Bits(offset_forward2) << "\n"
+      "daddiu $t9, 0x" << std::hex << Low16Bits(offset_forward2) << "\n"
+      "jialc $t9, 0\n" <<
+      RepeatInsn(kAdduCount2, "addu $zero, $zero, $zero\n") <<
+      "2:\n"
+      "auipc $t9, 0x" << std::hex << High16Bits(offset_back) << "\n"
+      "daddiu $t9, 0x" << std::hex << Low16Bits(offset_back) << "\n"
+      "jialc $t9, 0\n";
+  std::string expected = oss.str();
+  DriverStr(expected, "LongJialc");
+}
+
+TEST_F(AssemblerMIPS64Test, Bc) {
+  mips64::Mips64Label label1, label2;
+  __ Bc(&label1);
+  constexpr size_t kAdduCount1 = 63;
+  for (size_t i = 0; i != kAdduCount1; ++i) {
+    __ Addu(mips64::ZERO, mips64::ZERO, mips64::ZERO);
+  }
+  __ Bind(&label1);
+  __ Bc(&label2);
+  constexpr size_t kAdduCount2 = 64;
+  for (size_t i = 0; i != kAdduCount2; ++i) {
+    __ Addu(mips64::ZERO, mips64::ZERO, mips64::ZERO);
+  }
+  __ Bind(&label2);
+  __ Bc(&label1);
+
+  std::string expected =
+      ".set noreorder\n"
+      "bc 1f\n" +
+      RepeatInsn(kAdduCount1, "addu $zero, $zero, $zero\n") +
+      "1:\n"
+      "bc 2f\n" +
+      RepeatInsn(kAdduCount2, "addu $zero, $zero, $zero\n") +
+      "2:\n"
+      "bc 1b\n";
+  DriverStr(expected, "Bc");
+}
+
+TEST_F(AssemblerMIPS64Test, Beqzc) {
+  BranchCondOneRegHelper(&mips64::Mips64Assembler::Beqzc, "Beqzc");
+}
+
+TEST_F(AssemblerMIPS64Test, Bnezc) {
+  BranchCondOneRegHelper(&mips64::Mips64Assembler::Bnezc, "Bnezc");
+}
+
+TEST_F(AssemblerMIPS64Test, Bltzc) {
+  BranchCondOneRegHelper(&mips64::Mips64Assembler::Bltzc, "Bltzc");
+}
+
+TEST_F(AssemblerMIPS64Test, Bgezc) {
+  BranchCondOneRegHelper(&mips64::Mips64Assembler::Bgezc, "Bgezc");
+}
+
+TEST_F(AssemblerMIPS64Test, Blezc) {
+  BranchCondOneRegHelper(&mips64::Mips64Assembler::Blezc, "Blezc");
+}
+
+TEST_F(AssemblerMIPS64Test, Bgtzc) {
+  BranchCondOneRegHelper(&mips64::Mips64Assembler::Bgtzc, "Bgtzc");
+}
+
+TEST_F(AssemblerMIPS64Test, Beqc) {
+  BranchCondTwoRegsHelper(&mips64::Mips64Assembler::Beqc, "Beqc");
+}
+
+TEST_F(AssemblerMIPS64Test, Bnec) {
+  BranchCondTwoRegsHelper(&mips64::Mips64Assembler::Bnec, "Bnec");
+}
+
+TEST_F(AssemblerMIPS64Test, Bltc) {
+  BranchCondTwoRegsHelper(&mips64::Mips64Assembler::Bltc, "Bltc");
+}
+
+TEST_F(AssemblerMIPS64Test, Bgec) {
+  BranchCondTwoRegsHelper(&mips64::Mips64Assembler::Bgec, "Bgec");
+}
+
+TEST_F(AssemblerMIPS64Test, Bltuc) {
+  BranchCondTwoRegsHelper(&mips64::Mips64Assembler::Bltuc, "Bltuc");
+}
+
+TEST_F(AssemblerMIPS64Test, Bgeuc) {
+  BranchCondTwoRegsHelper(&mips64::Mips64Assembler::Bgeuc, "Bgeuc");
+}
+
+TEST_F(AssemblerMIPS64Test, LongBeqc) {
+  mips64::Mips64Label label;
+  __ Beqc(mips64::A0, mips64::A1, &label);
+  constexpr uint32_t kAdduCount1 = (1u << 15) + 1;
+  for (uint32_t i = 0; i != kAdduCount1; ++i) {
+    __ Addu(mips64::ZERO, mips64::ZERO, mips64::ZERO);
+  }
+  __ Bind(&label);
+  constexpr uint32_t kAdduCount2 = (1u << 15) + 1;
+  for (uint32_t i = 0; i != kAdduCount2; ++i) {
+    __ Addu(mips64::ZERO, mips64::ZERO, mips64::ZERO);
+  }
+  __ Beqc(mips64::A2, mips64::A3, &label);
+
+  uint32_t offset_forward = 2 + kAdduCount1;  // 2: account for auipc and jic.
+  offset_forward <<= 2;
+  offset_forward += (offset_forward & 0x8000) << 1;  // Account for sign extension in jic.
+
+  uint32_t offset_back = -(kAdduCount2 + 1);  // 1: account for bnec.
+  offset_back <<= 2;
+  offset_back += (offset_back & 0x8000) << 1;  // Account for sign extension in jic.
+
+  std::ostringstream oss;
+  oss <<
+      ".set noreorder\n"
+      "bnec $a0, $a1, 1f\n"
+      "auipc $at, 0x" << std::hex << High16Bits(offset_forward) << "\n"
+      "jic $at, 0x" << std::hex << Low16Bits(offset_forward) << "\n"
+      "1:\n" <<
+      RepeatInsn(kAdduCount1, "addu $zero, $zero, $zero\n") <<
+      "2:\n" <<
+      RepeatInsn(kAdduCount2, "addu $zero, $zero, $zero\n") <<
+      "bnec $a2, $a3, 3f\n"
+      "auipc $at, 0x" << std::hex << High16Bits(offset_back) << "\n"
+      "jic $at, 0x" << std::hex << Low16Bits(offset_back) << "\n"
+      "3:\n";
+  std::string expected = oss.str();
+  DriverStr(expected, "LongBeqc");
 }
 
 //////////