MIPS: Branch delay slot and other optimizations.

List of changes:
-added a minor optimization to the Simulator that quickly skips nops in the delay slot
-slightly re-worked CEntryStub to save a few instructions
 CEntryStub now expects the following values:
  -s0: number of arguments including receiver
  -s1: size of arguments excluding receiver
  -s2: pointer to builtin function
 Two new MacroAssembler functions were added to make usage more convenient:
  -PrepareCEntryArgs(int num_args) to set up s0 and s1
  -PrepareCEntryFunction(const ExternalReference&) to set up s2
-removed branch delay slot nops from the most frequently used code areas
-reorganized some code to execute fewer instructions
-utilized the delay slot of most Ret instructions
 This does not cover all Rets, only the most obvious cases.
 Also added a special version of DropAndRet that utilizes the delay slot.
-added some comments to code areas where explanation of the register/delay slot usage may be needed
-added an optimization to Jump so it doesn't always pre-load the target register

BUG=
TEST=

Review URL: https://chromiumcodereview.appspot.com/9699071
Patch from Daniel Kalmar <kalmard@homejinni.com>.

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@11099 ce2b1a6d-e550-0410-aec6-3dcde31c8c00

src/mips/builtins-mips.cc

@@ -67,9 +67,11 @@ void Builtins::Generate_Adaptor(MacroAssembler* masm,
     ASSERT(extra_args == NO_EXTRA_ARGUMENTS);
   }
 
-  // JumpToExternalReference expects a0 to contain the number of arguments
+  // JumpToExternalReference expects s0 to contain the number of arguments
   // including the receiver and the extra arguments.
-  __ Addu(a0, a0, Operand(num_extra_args + 1));
+  __ Addu(s0, a0, num_extra_args + 1);
+  __ sll(s1, s0, kPointerSizeLog2);
+  __ Subu(s1, s1, kPointerSize);
   __ JumpToExternalReference(ExternalReference(id, masm->isolate()));
 }
 

src/mips/code-stubs-mips.cc

@@ -70,13 +70,13 @@ void ToNumberStub::Generate(MacroAssembler* masm) {
   // The ToNumber stub takes one argument in a0.
   Label check_heap_number, call_builtin;
   __ JumpIfNotSmi(a0, &check_heap_number);
+  __ Ret(USE_DELAY_SLOT);
   __ mov(v0, a0);
-  __ Ret();
 
   __ bind(&check_heap_number);
   EmitCheckForHeapNumber(masm, a0, a1, t0, &call_builtin);
+  __ Ret(USE_DELAY_SLOT);
   __ mov(v0, a0);
-  __ Ret();
 
   __ bind(&call_builtin);
   __ push(a0);
@@ -128,9 +128,9 @@ void FastNewClosureStub::Generate(MacroAssembler* masm) {
   // found in the shared function info object.
   __ lw(a3, FieldMemOperand(a3, SharedFunctionInfo::kCodeOffset));
   __ Addu(a3, a3, Operand(Code::kHeaderSize - kHeapObjectTag));
-  __ sw(a3, FieldMemOperand(v0, JSFunction::kCodeEntryOffset));
 
   // Return result. The argument function info has been popped already.
+  __ sw(a3, FieldMemOperand(v0, JSFunction::kCodeEntryOffset));
   __ Ret();
 
   // Create a new closure through the slower runtime call.
@@ -179,8 +179,7 @@ void FastNewContextStub::Generate(MacroAssembler* masm) {
 
   // Remove the on-stack argument and return.
   __ mov(cp, v0);
-  __ Pop();
-  __ Ret();
+  __ DropAndRet(1);
 
   // Need to collect. Call into runtime system.
   __ bind(&gc);
@@ -242,8 +241,7 @@ void FastNewBlockContextStub::Generate(MacroAssembler* masm) {
 
   // Remove the on-stack argument and return.
   __ mov(cp, v0);
-  __ Addu(sp, sp, Operand(2 * kPointerSize));
-  __ Ret();
+  __ DropAndRet(2);
 
   // Need to collect. Call into runtime system.
   __ bind(&gc);
@@ -368,8 +366,7 @@ void FastCloneShallowArrayStub::Generate(MacroAssembler* masm) {
   GenerateFastCloneShallowArrayCommon(masm, length_, mode, &slow_case);
 
   // Return and remove the on-stack parameters.
-  __ Addu(sp, sp, Operand(3 * kPointerSize));
-  __ Ret();
+  __ DropAndRet(3);
 
   __ bind(&slow_case);
   __ TailCallRuntime(Runtime::kCreateArrayLiteralShallow, 3, 1);
@@ -405,16 +402,14 @@ void FastCloneShallowObjectStub::Generate(MacroAssembler* masm) {
 
   // Allocate the JS object and copy header together with all in-object
   // properties from the boilerplate.
-  __ AllocateInNewSpace(size, a0, a1, a2, &slow_case, TAG_OBJECT);
+  __ AllocateInNewSpace(size, v0, a1, a2, &slow_case, TAG_OBJECT);
   for (int i = 0; i < size; i += kPointerSize) {
     __ lw(a1, FieldMemOperand(a3, i));
-    __ sw(a1, FieldMemOperand(a0, i));
+    __ sw(a1, FieldMemOperand(v0, i));
   }
 
   // Return and remove the on-stack parameters.
-  __ Drop(4);
-  __ Ret(USE_DELAY_SLOT);
-  __ mov(v0, a0);
+  __ DropAndRet(4);
 
   __ bind(&slow_case);
   __ TailCallRuntime(Runtime::kCreateObjectLiteralShallow, 4, 1);
@@ -492,8 +487,8 @@ void ConvertToDoubleStub::Generate(MacroAssembler* masm) {
   __ Or(at, exponent, Operand(exponent_word_for_1));
   __ Movn(exponent, at, source_);  // Write exp when source not 0.
   // 1, 0 and -1 all have 0 for the second word.
+  __ Ret(USE_DELAY_SLOT);
   __ mov(mantissa, zero_reg);
-  __ Ret();
 
   __ bind(&not_special);
   // Count leading zeros.
@@ -514,9 +509,9 @@ void ConvertToDoubleStub::Generate(MacroAssembler* masm) {
   __ sll(mantissa, source_, HeapNumber::kMantissaBitsInTopWord);
   // And the top (top 20 bits).
   __ srl(source_, source_, 32 - HeapNumber::kMantissaBitsInTopWord);
-  __ or_(exponent, exponent, source_);
 
-  __ Ret();
+  __ Ret(USE_DELAY_SLOT);
+  __ or_(exponent, exponent, source_);
 }
 
 
@@ -1025,9 +1020,9 @@ void FloatingPointHelper::CallCCodeForDoubleOperation(
     __ sw(v0, FieldMemOperand(heap_number_result, HeapNumber::kMantissaOffset));
   }
   // Place heap_number_result in v0 and return to the pushed return address.
-  __ mov(v0, heap_number_result);
   __ pop(ra);
-  __ Ret();
+  __ Ret(USE_DELAY_SLOT);
+  __ mov(v0, heap_number_result);
 }
 
 
@@ -1163,6 +1158,7 @@ static void EmitIdenticalObjectComparison(MacroAssembler* masm,
   }
 
   __ bind(&return_equal);
+
   if (cc == less) {
     __ li(v0, Operand(GREATER));  // Things aren't less than themselves.
   } else if (cc == greater) {
@@ -1234,8 +1230,8 @@ static void EmitSmiNonsmiComparison(MacroAssembler* masm,
   if (strict) {
     // If lhs was not a number and rhs was a Smi then strict equality cannot
     // succeed. Return non-equal (lhs is already not zero).
+    __ Ret(USE_DELAY_SLOT, ne, t4, Operand(HEAP_NUMBER_TYPE));
     __ mov(v0, lhs);
-    __ Ret(ne, t4, Operand(HEAP_NUMBER_TYPE));
   } else {
     // Smi compared non-strictly with a non-Smi non-heap-number. Call
     // the runtime.
@@ -1273,8 +1269,8 @@ static void EmitSmiNonsmiComparison(MacroAssembler* masm,
   if (strict) {
     // If lhs was not a number and rhs was a Smi then strict equality cannot
     // succeed. Return non-equal.
+    __ Ret(USE_DELAY_SLOT, ne, t4, Operand(HEAP_NUMBER_TYPE));
     __ li(v0, Operand(1));
-    __ Ret(ne, t4, Operand(HEAP_NUMBER_TYPE));
   } else {
     // Smi compared non-strictly with a non-Smi non-heap-number. Call
     // the runtime.
@@ -1354,12 +1350,13 @@ void EmitNanCheck(MacroAssembler* masm, Condition cc) {
   __ bind(&one_is_nan);
   // NaN comparisons always fail.
   // Load whatever we need in v0 to make the comparison fail.
+
   if (cc == lt || cc == le) {
     __ li(v0, Operand(GREATER));
   } else {
     __ li(v0, Operand(LESS));
   }
-  __ Ret();  // Return.
+  __ Ret();
 
   __ bind(&neither_is_nan);
 }
@@ -1408,6 +1405,7 @@ static void EmitTwoNonNanDoubleComparison(MacroAssembler* masm, Condition cc) {
     __ Branch(&return_result_not_equal, ne, t4, Operand(zero_reg));
 
     __ bind(&return_result_equal);
+
     __ li(v0, Operand(EQUAL));
     __ Ret();
   }
@@ -1439,6 +1437,7 @@ static void EmitTwoNonNanDoubleComparison(MacroAssembler* masm, Condition cc) {
     __ BranchF(&less_than, NULL, lt, f12, f14);
 
     // Not equal, not less, not NaN, must be greater.
+
     __ li(v0, Operand(GREATER));
     __ Ret();
 
@@ -1469,8 +1468,8 @@ static void EmitStrictTwoHeapObjectCompare(MacroAssembler* masm,
     // Return non-zero.
     Label return_not_equal;
     __ bind(&return_not_equal);
+    __ Ret(USE_DELAY_SLOT);
     __ li(v0, Operand(1));
-    __ Ret();
 
     __ bind(&first_non_object);
     // Check for oddballs: true, false, null, undefined.
@@ -1549,8 +1548,8 @@ static void EmitCheckForSymbolsOrObjects(MacroAssembler* masm,
 
   // Both are symbols. We already checked they weren't the same pointer
   // so they are not equal.
+  __ Ret(USE_DELAY_SLOT);
   __ li(v0, Operand(1));   // Non-zero indicates not equal.
-  __ Ret();
 
   __ bind(&object_test);
   __ Branch(not_both_strings, lt, a2, Operand(FIRST_SPEC_OBJECT_TYPE));
@@ -1565,8 +1564,8 @@ static void EmitCheckForSymbolsOrObjects(MacroAssembler* masm,
   __ lbu(a3, FieldMemOperand(a3, Map::kBitFieldOffset));
   __ and_(a0, a2, a3);
   __ And(a0, a0, Operand(1 << Map::kIsUndetectable));
-  __ Xor(v0, a0, Operand(1 << Map::kIsUndetectable));
-  __ Ret();
+  __ Ret(USE_DELAY_SLOT);
+  __ xori(v0, a0, 1 << Map::kIsUndetectable);
 }
 
 
@@ -1673,8 +1672,7 @@ void NumberToStringStub::Generate(MacroAssembler* masm) {
 
   // Generate code to lookup number in the number string cache.
   GenerateLookupNumberStringCache(masm, a1, v0, a2, a3, t0, false, &runtime);
-  __ Addu(sp, sp, Operand(1 * kPointerSize));
-  __ Ret();
+  __ DropAndRet(1);
 
   __ bind(&runtime);
   // Handle number to string in the runtime system if not found in the cache.
@@ -1696,8 +1694,8 @@ void CompareStub::Generate(MacroAssembler* masm) {
     __ JumpIfNotSmi(a2, &not_two_smis);
     __ sra(a1, a1, 1);
     __ sra(a0, a0, 1);
-    __ Subu(v0, a1, a0);
-    __ Ret();
+    __ Ret(USE_DELAY_SLOT);
+    __ subu(v0, a1, a0);
     __ bind(&not_two_smis);
   } else if (FLAG_debug_code) {
     __ Or(a2, a1, a0);
@@ -1916,8 +1914,8 @@ void ToBooleanStub::Generate(MacroAssembler* masm) {
     __ lbu(at, FieldMemOperand(map, Map::kInstanceTypeOffset));
     Label skip;
     __ Branch(&skip, ge, at, Operand(FIRST_NONSTRING_TYPE));
+    __ Ret(USE_DELAY_SLOT);  // the string length is OK as the return value
     __ lw(tos_, FieldMemOperand(tos_, String::kLengthOffset));
-    __ Ret();  // the string length is OK as the return value
     __ bind(&skip);
   }
 
@@ -2092,8 +2090,8 @@ void UnaryOpStub::GenerateSmiCodeSub(MacroAssembler* masm,
   __ Branch(slow, eq, t0, Operand(zero_reg));
 
   // Return '0 - value'.
-  __ Subu(v0, zero_reg, a0);
-  __ Ret();
+  __ Ret(USE_DELAY_SLOT);
+  __ subu(v0, zero_reg, a0);
 }
 
 
@@ -2423,8 +2421,8 @@ void BinaryOpStub::GenerateSmiSmiOperation(MacroAssembler* masm) {
       // Negating it results in 'lt'.
       __ Branch(&skip, lt, scratch2, Operand(zero_reg));
       ASSERT(Smi::FromInt(0) == 0);
-      __ mov(v0, zero_reg);
-      __ Ret();  // Return smi 0 if the non-zero one was positive.
+      __ Ret(USE_DELAY_SLOT);
+      __ mov(v0, zero_reg);  // Return smi 0 if the non-zero one was positive.
       __ bind(&skip);
       // We fall through here if we multiplied a negative number with 0, because
       // that would mean we should produce -0.
@@ -2479,23 +2477,23 @@ void BinaryOpStub::GenerateSmiSmiOperation(MacroAssembler* masm) {
       }
       break;
     case Token::BIT_OR:
-      __ Or(v0, left, Operand(right));
-      __ Ret();
+      __ Ret(USE_DELAY_SLOT);
+      __ or_(v0, left, right);
       break;
     case Token::BIT_AND:
-      __ And(v0, left, Operand(right));
-      __ Ret();
+      __ Ret(USE_DELAY_SLOT);
+      __ and_(v0, left, right);
       break;
     case Token::BIT_XOR:
-      __ Xor(v0, left, Operand(right));
-      __ Ret();
+      __ Ret(USE_DELAY_SLOT);
+      __ xor_(v0, left, right);
       break;
     case Token::SAR:
       // Remove tags from right operand.
       __ GetLeastBitsFromSmi(scratch1, right, 5);
       __ srav(scratch1, left, scratch1);
       // Smi tag result.
-      __ And(v0, scratch1, Operand(~kSmiTagMask));
+      __ And(v0, scratch1, ~kSmiTagMask);
       __ Ret();
       break;
     case Token::SHR:
@@ -2607,8 +2605,8 @@ void BinaryOpStub::GenerateFPOperation(MacroAssembler* masm,
         // kValueOffset. On MIPS this workaround is built into sdc1 so
         // there's no point in generating even more instructions.
         __ sdc1(f10, FieldMemOperand(result, HeapNumber::kValueOffset));
+        __ Ret(USE_DELAY_SLOT);
         __ mov(v0, result);
-        __ Ret();
       } else {
         // Call the C function to handle the double operation.
         FloatingPointHelper::CallCCodeForDoubleOperation(masm,
@@ -3482,8 +3480,8 @@ void TranscendentalCacheStub::Generate(MacroAssembler* masm) {
     __ sw(a3, MemOperand(cache_entry, 1 * kPointerSize));
     __ sw(t2, MemOperand(cache_entry, 2 * kPointerSize));
 
+    __ Ret(USE_DELAY_SLOT);
     __ mov(v0, cache_entry);
-    __ Ret();
 
     __ bind(&invalid_cache);
     // The cache is invalid. Call runtime which will recreate the
@@ -3662,7 +3660,7 @@ void MathPowStub::Generate(MacroAssembler* masm) {
                  ne,
                  double_exponent,
                  double_scratch);
-
+      // double_scratch can be overwritten in the delay slot.
       // Calculates square root of base.  Check for the special case of
       // Math.pow(-Infinity, 0.5) == Infinity (ECMA spec, 15.8.2.13).
       __ Move(double_scratch, -V8_INFINITY);
@@ -3682,7 +3680,7 @@ void MathPowStub::Generate(MacroAssembler* masm) {
                  ne,
                  double_exponent,
                  double_scratch);
-
+      // double_scratch can be overwritten in the delay slot.
       // Calculates square root of base.  Check for the special case of
       // Math.pow(-Infinity, -0.5) == 0 (ECMA spec, 15.8.2.13).
       __ Move(double_scratch, -V8_INFINITY);
@@ -3866,9 +3864,10 @@ void CEntryStub::GenerateCore(MacroAssembler* masm,
     __ sw(a1, MemOperand(a0));
   }
 
-  // Prepare arguments for C routine: a0 = argc, a1 = argv
+  // Prepare arguments for C routine.
+  // a0 = argc
   __ mov(a0, s0);
-  __ mov(a1, s1);
+  // a1 = argv (set in the delay slot after find_ra below).
 
   // We are calling compiled C/C++ code. a0 and a1 hold our two arguments. We
   // also need to reserve the 4 argument slots on the stack.
@@ -3888,30 +3887,28 @@ void CEntryStub::GenerateCore(MacroAssembler* masm,
     // coverage code can interfere with the proper calculation of ra.
     Label find_ra;
     masm->bal(&find_ra);  // bal exposes branch delay slot.
-    masm->nop();  // Branch delay slot nop.
+    masm->mov(a1, s1);
     masm->bind(&find_ra);
 
     // Adjust the value in ra to point to the correct return location, 2nd
     // instruction past the real call into C code (the jalr(t9)), and push it.
     // This is the return address of the exit frame.
-    const int kNumInstructionsToJump = 6;
+    const int kNumInstructionsToJump = 5;
     masm->Addu(ra, ra, kNumInstructionsToJump * kPointerSize);
     masm->sw(ra, MemOperand(sp));  // This spot was reserved in EnterExitFrame.
-    masm->Subu(sp, sp, kCArgsSlotsSize);
+    // Stack space reservation moved to the branch delay slot below.
     // Stack is still aligned.
 
     // Call the C routine.
     masm->mov(t9, s2);  // Function pointer to t9 to conform to ABI for PIC.
     masm->jalr(t9);
-    masm->nop();    // Branch delay slot nop.
+    // Set up sp in the delay slot.
+    masm->addiu(sp, sp, -kCArgsSlotsSize);
     // Make sure the stored 'ra' points to this position.
     ASSERT_EQ(kNumInstructionsToJump,
               masm->InstructionsGeneratedSince(&find_ra));
   }
 
-  // Restore stack (remove arg slots).
-  __ Addu(sp, sp, kCArgsSlotsSize);
-
   if (always_allocate) {
     // It's okay to clobber a2 and a3 here. v0 & v1 contain result.
     __ li(a2, Operand(scope_depth));
@@ -3925,14 +3922,16 @@ void CEntryStub::GenerateCore(MacroAssembler* masm,
   STATIC_ASSERT(((kFailureTag + 1) & kFailureTagMask) == 0);
   __ addiu(a2, v0, 1);
   __ andi(t0, a2, kFailureTagMask);
-  __ Branch(&failure_returned, eq, t0, Operand(zero_reg));
+  __ Branch(USE_DELAY_SLOT, &failure_returned, eq, t0, Operand(zero_reg));
+  // Restore stack (remove arg slots) in branch delay slot.
+  __ addiu(sp, sp, kCArgsSlotsSize);
+
 
   // Exit C frame and return.
   // v0:v1: result
   // sp: stack pointer
   // fp: frame pointer
-  __ LeaveExitFrame(save_doubles_, s0);
-  __ Ret();
+  __ LeaveExitFrame(save_doubles_, s0, true);
 
   // Check if we should retry or throw exception.
   Label retry;
@@ -3943,8 +3942,10 @@ void CEntryStub::GenerateCore(MacroAssembler* masm,
 
   // Special handling of out of memory exceptions.
   Failure* out_of_memory = Failure::OutOfMemoryException();
-  __ Branch(throw_out_of_memory_exception, eq,
+  __ Branch(USE_DELAY_SLOT, throw_out_of_memory_exception, eq,
             v0, Operand(reinterpret_cast<int32_t>(out_of_memory)));
+  // If we throw the OOM exception, the value of a3 doesn't matter.
+  // Any instruction can be in the delay slot that's not a jump.
 
   // Retrieve the pending exception and clear the variable.
   __ li(a3, Operand(isolate->factory()->the_hole_value()));
@@ -3968,8 +3969,9 @@ void CEntryStub::GenerateCore(MacroAssembler* masm,
 
 void CEntryStub::Generate(MacroAssembler* masm) {
   // Called from JavaScript; parameters are on stack as if calling JS function
-  // a0: number of arguments including receiver
-  // a1: pointer to builtin function
+  // s0: number of arguments including receiver
+  // s1: size of arguments excluding receiver
+  // s2: pointer to builtin function
   // fp: frame pointer    (restored after C call)
   // sp: stack pointer    (restored as callee's sp after C call)
   // cp: current context  (C callee-saved)
@@ -3979,19 +3981,18 @@ void CEntryStub::Generate(MacroAssembler* masm) {
   // this by performing a garbage collection and retrying the
   // builtin once.
 
+  // NOTE: s0-s2 hold the arguments of this function instead of a0-a2.
+  // The reason for this is that these arguments would need to be saved anyway
+  // so it's faster to set them up directly.
+  // See MacroAssembler::PrepareCEntryArgs and PrepareCEntryFunction.
+
   // Compute the argv pointer in a callee-saved register.
-  __ sll(s1, a0, kPointerSizeLog2);
   __ Addu(s1, sp, s1);
-  __ Subu(s1, s1, Operand(kPointerSize));
 
   // Enter the exit frame that transitions from JavaScript to C++.
   FrameScope scope(masm, StackFrame::MANUAL);
   __ EnterExitFrame(save_doubles_);
 
-  // Set up argc and the builtin function in callee-saved registers.
-  __ mov(s0, a0);
-  __ mov(s2, a1);
-
   // s0: number of arguments (C callee-saved)
   // s1: pointer to first argument (C callee-saved)
   // s2: pointer to builtin function (C callee-saved)
@@ -4693,8 +4694,7 @@ void ArgumentsAccessStub::GenerateNewNonStrictFast(MacroAssembler* masm) {
   __ Branch(&arguments_loop, lt, t5, Operand(a2));
 
   // Return and remove the on-stack parameters.
-  __ Addu(sp, sp, Operand(3 * kPointerSize));
-  __ Ret();
+  __ DropAndRet(3);
 
   // Do the runtime call to allocate the arguments object.
   // a2 = argument count (tagged)
@@ -4799,8 +4799,7 @@ void ArgumentsAccessStub::GenerateNewStrict(MacroAssembler* masm) {
 
   // Return and remove the on-stack parameters.
   __ bind(&done);
-  __ Addu(sp, sp, Operand(3 * kPointerSize));
-  __ Ret();
+  __ DropAndRet(3);
 
   // Do the runtime call to allocate the arguments object.
   __ bind(&runtime);
@@ -5149,8 +5148,7 @@ void RegExpExecStub::Generate(MacroAssembler* masm) {
   __ bind(&failure);
   // For failure and exception return null.
   __ li(v0, Operand(isolate->factory()->null_value()));
-  __ Addu(sp, sp, Operand(4 * kPointerSize));
-  __ Ret();
+  __ DropAndRet(4);
 
   // Process the result from the native regexp code.
   __ bind(&success);
@@ -5217,8 +5215,7 @@ void RegExpExecStub::Generate(MacroAssembler* masm) {
 
   // Return last match info.
   __ lw(v0, MemOperand(sp, kLastMatchInfoOffset));
-  __ Addu(sp, sp, Operand(4 * kPointerSize));
-  __ Ret();
+  __ DropAndRet(4);
 
   // External string.  Short external strings have already been ruled out.
   // a0: scratch
@@ -5330,8 +5327,7 @@ void RegExpConstructResultStub::Generate(MacroAssembler* masm) {
   __ addiu(a3, a3, kPointerSize);  // In branch delay slot.
 
   __ bind(&done);
-  __ Addu(sp, sp, Operand(3 * kPointerSize));
-  __ Ret();
+  __ DropAndRet(3);
 
   __ bind(&slowcase);
   __ TailCallRuntime(Runtime::kRegExpConstructResult, 3, 1);
@@ -6136,7 +6132,7 @@ void SubStringStub::Generate(MacroAssembler* masm) {
   STATIC_ASSERT(kIsIndirectStringMask != 0);
   __ And(t0, a1, Operand(kIsIndirectStringMask));
   __ Branch(USE_DELAY_SLOT, &seq_or_external_string, eq, t0, Operand(zero_reg));
-
+  // t0 is used as a scratch register and can be overwritten in either case.
   __ And(t0, a1, Operand(kSlicedNotConsMask));
   __ Branch(&sliced_string, ne, t0, Operand(zero_reg));
   // Cons string.  Check whether it is flat, then fetch first part.
@@ -6409,8 +6405,7 @@ void StringCompareStub::Generate(MacroAssembler* masm) {
   STATIC_ASSERT(kSmiTag == 0);
   __ li(v0, Operand(Smi::FromInt(EQUAL)));
   __ IncrementCounter(counters->string_compare_native(), 1, a1, a2);
-  __ Addu(sp, sp, Operand(2 * kPointerSize));
-  __ Ret();
+  __ DropAndRet(2);
 
   __ bind(&not_same);
 
@@ -6815,16 +6810,16 @@ void ICCompareStub::GenerateHeapNumbers(MacroAssembler* masm) {
     __ BranchF(&fpu_lt, NULL, lt, f0, f2);
 
     // Otherwise it's greater, so just fall thru, and return.
-    __ Ret(USE_DELAY_SLOT);
-    __ li(v0, Operand(GREATER));  // In delay slot.
+    __ li(v0, Operand(GREATER));
+    __ Ret();
 
     __ bind(&fpu_eq);
-    __ Ret(USE_DELAY_SLOT);
-    __ li(v0, Operand(EQUAL));  // In delay slot.
+    __ li(v0, Operand(EQUAL));
+    __ Ret();
 
     __ bind(&fpu_lt);
-    __ Ret(USE_DELAY_SLOT);
-    __ li(v0, Operand(LESS));  // In delay slot.
+    __ li(v0, Operand(LESS));
+    __ Ret();
   }
 
   __ bind(&unordered);
@@ -6924,9 +6919,9 @@ void ICCompareStub::GenerateStrings(MacroAssembler* masm) {
   Label left_ne_right;
   STATIC_ASSERT(EQUAL == 0);
   STATIC_ASSERT(kSmiTag == 0);
-  __ Branch(&left_ne_right, ne, left, Operand(right), USE_DELAY_SLOT);
+  __ Branch(&left_ne_right, ne, left, Operand(right));
+  __ Ret(USE_DELAY_SLOT);
   __ mov(v0, zero_reg);  // In the delay slot.
-  __ Ret();
   __ bind(&left_ne_right);
 
   // Handle not identical strings.
@@ -6939,12 +6934,12 @@ void ICCompareStub::GenerateStrings(MacroAssembler* masm) {
     __ And(tmp3, tmp1, Operand(tmp2));
     __ And(tmp5, tmp3, Operand(kIsSymbolMask));
     Label is_symbol;
-    __ Branch(&is_symbol, eq, tmp5, Operand(zero_reg), USE_DELAY_SLOT);
-    __ mov(v0, a0);  // In the delay slot.
+    __ Branch(&is_symbol, eq, tmp5, Operand(zero_reg));
     // Make sure a0 is non-zero. At this point input operands are
     // guaranteed to be non-zero.
     ASSERT(right.is(a0));
-    __ Ret();
+    __ Ret(USE_DELAY_SLOT);
+    __ mov(v0, a0);  // In the delay slot.
     __ bind(&is_symbol);
   }
 
@@ -6988,8 +6983,8 @@ void ICCompareStub::GenerateObjects(MacroAssembler* masm) {
   __ Branch(&miss, ne, a2, Operand(JS_OBJECT_TYPE));
 
   ASSERT(GetCondition() == eq);
-  __ Subu(v0, a0, Operand(a1));
-  __ Ret();
+  __ Ret(USE_DELAY_SLOT);
+  __ subu(v0, a0, a1);
 
   __ bind(&miss);
   GenerateMiss(masm);
@@ -7022,8 +7017,9 @@ void ICCompareStub::GenerateMiss(MacroAssembler* masm) {
     __ push(ra);
     __ Push(a1, a0);
     __ li(t0, Operand(Smi::FromInt(op_)));
-    __ push(t0);
-    __ CallExternalReference(miss, 3);
+    __ addiu(sp, sp, -kPointerSize);
+    __ CallExternalReference(miss, 3, USE_DELAY_SLOT);
+    __ sw(t0, MemOperand(sp));  // In the delay slot.
     // Compute the entry point of the rewritten stub.
     __ Addu(a2, v0, Operand(Code::kHeaderSize - kHeapObjectTag));
     // Restore registers.
@@ -7333,17 +7329,17 @@ void StringDictionaryLookupStub::Generate(MacroAssembler* masm) {
   // treated as a lookup success. For positive lookup probing failure
   // should be treated as lookup failure.
   if (mode_ == POSITIVE_LOOKUP) {
+    __ Ret(USE_DELAY_SLOT);
     __ mov(result, zero_reg);
-    __ Ret();
   }
 
   __ bind(&in_dictionary);
+  __ Ret(USE_DELAY_SLOT);
   __ li(result, 1);
-  __ Ret();
 
   __ bind(&not_in_dictionary);
+  __ Ret(USE_DELAY_SLOT);
   __ mov(result, zero_reg);
-  __ Ret();
 }
 
 

src/mips/debug-mips.cc

@@ -152,8 +152,8 @@ static void Generate_DebugBreakCallHelper(MacroAssembler* masm,
 #ifdef DEBUG
     __ RecordComment("// Calling from debug break to runtime - come in - over");
 #endif
-    __ mov(a0, zero_reg);  // No arguments.
-    __ li(a1, Operand(ExternalReference::debug_break(masm->isolate())));
+    __ PrepareCEntryArgs(0);  // No arguments.
+    __ PrepareCEntryFunction(ExternalReference::debug_break(masm->isolate()));
 
     CEntryStub ceb(1);
     __ CallStub(&ceb);

src/mips/ic-mips.cc

@@ -512,8 +512,8 @@ void CallICBase::GenerateMiss(MacroAssembler* masm,
     __ Push(a3, a2);
 
     // Call the entry.
-    __ li(a0, Operand(2));
-    __ li(a1, Operand(ExternalReference(IC_Utility(id), isolate)));
+    __ PrepareCEntryArgs(2);
+    __ PrepareCEntryFunction(ExternalReference(IC_Utility(id), isolate));
 
     CEntryStub stub(1);
     __ CallStub(&stub);
@@ -844,8 +844,8 @@ void KeyedLoadIC::GenerateNonStrictArguments(MacroAssembler* masm) {
   Label slow, notin;
   MemOperand mapped_location =
       GenerateMappedArgumentsLookup(masm, a1, a0, a2, a3, t0, &notin, &slow);
+  __ Ret(USE_DELAY_SLOT);
   __ lw(v0, mapped_location);
-  __ Ret();
   __ bind(&notin);
   // The unmapped lookup expects that the parameter map is in a2.
   MemOperand unmapped_location =
@@ -853,8 +853,8 @@ void KeyedLoadIC::GenerateNonStrictArguments(MacroAssembler* masm) {
   __ lw(a2, unmapped_location);
   __ LoadRoot(a3, Heap::kTheHoleValueRootIndex);
   __ Branch(&slow, eq, a2, Operand(a3));
+  __ Ret(USE_DELAY_SLOT);
   __ mov(v0, a2);
-  __ Ret();
   __ bind(&slow);
   GenerateMiss(masm, false);
 }

src/mips/lithium-codegen-mips.cc

@@ -634,13 +634,9 @@ void LCodeGen::DeoptimizeIf(Condition cc,
     __ bind(&skip);
   }
 
-  if (cc == al) {
-    __ Jump(entry, RelocInfo::RUNTIME_ENTRY);
-  } else {
   // TODO(plind): The Arm port is a little different here, due to their
   // DeOpt jump table, which is not used for Mips yet.
   __ Jump(entry, RelocInfo::RUNTIME_ENTRY, cc, src1, src2);
-  }
 }
 
 

src/mips/macro-assembler-mips.cc

@@ -2438,8 +2438,15 @@ void MacroAssembler::Jump(intptr_t target,
                           Register rs,
                           const Operand& rt,
                           BranchDelaySlot bd) {
+  Label skip;
+  if (cond != cc_always) {
+    Branch(USE_DELAY_SLOT, &skip, NegateCondition(cond), rs, rt);
+  }
+  // The first instruction of 'li' may be placed in the delay slot.
+  // This is not an issue, t9 is expected to be clobbered anyway.
   li(t9, Operand(target, rmode));
-  Jump(t9, cond, rs, rt, bd);
+  Jump(t9, al, zero_reg, Operand(zero_reg), bd);
+  bind(&skip);
 }
 
 
@@ -2569,7 +2576,7 @@ void MacroAssembler::Call(Handle<Code> code,
     rmode = RelocInfo::CODE_TARGET_WITH_ID;
   }
   Call(reinterpret_cast<Address>(code.location()), rmode, cond, rs, rt, bd);
-  ASSERT_EQ(CallSize(code, rmode, ast_id, cond, rs, rt),
+  ASSERT_EQ(CallSize(code, rmode, ast_id, cond, rs, rt, bd),
             SizeOfCodeGeneratedSince(&start));
 }
 
@@ -2639,14 +2646,16 @@ void MacroAssembler::Jalr(Label* L, BranchDelaySlot bdslot) {
     nop();
 }
 
+void MacroAssembler::DropAndRet(int drop) {
+  Ret(USE_DELAY_SLOT);
+  addiu(sp, sp, drop * kPointerSize);
+}
 
 void MacroAssembler::DropAndRet(int drop,
                                 Condition cond,
                                 Register r1,
                                 const Operand& r2) {
-  // This is a workaround to make sure only one branch instruction is
-  // generated. It relies on Drop and Ret not creating branches if
-  // cond == cc_always.
+  // Both Drop and Ret need to be conditional.
   Label skip;
   if (cond != cc_always) {
     Branch(&skip, NegateCondition(cond), r1, r2);
@@ -2713,8 +2722,8 @@ void MacroAssembler::Push(Handle<Object> handle) {
 #ifdef ENABLE_DEBUGGER_SUPPORT
 
 void MacroAssembler::DebugBreak() {
-  mov(a0, zero_reg);
-  li(a1, Operand(ExternalReference(Runtime::kDebugBreak, isolate())));
+  PrepareCEntryArgs(0);
+  PrepareCEntryFunction(ExternalReference(Runtime::kDebugBreak, isolate()));
   CEntryStub ces(1);
   ASSERT(AllowThisStubCall(&ces));
   Call(ces.GetCode(), RelocInfo::DEBUG_BREAK);
@@ -3876,10 +3885,13 @@ void MacroAssembler::GetObjectType(Register object,
 // -----------------------------------------------------------------------------
 // Runtime calls.
 
-void MacroAssembler::CallStub(CodeStub* stub, Condition cond,
-                              Register r1, const Operand& r2) {
+void MacroAssembler::CallStub(CodeStub* stub,
+                              Condition cond,
+                              Register r1,
+                              const Operand& r2,
+                              BranchDelaySlot bd) {
   ASSERT(AllowThisStubCall(stub));  // Stub calls are not allowed in some stubs.
-  Call(stub->GetCode(), RelocInfo::CODE_TARGET, kNoASTId, cond, r1, r2);
+  Call(stub->GetCode(), RelocInfo::CODE_TARGET, kNoASTId, cond, r1, r2, bd);
 }
 
 
@@ -3962,8 +3974,7 @@ void MacroAssembler::CallApiFunctionAndReturn(ExternalReference function,
   lw(t1, MemOperand(at));
   Branch(&promote_scheduled_exception, ne, t0, Operand(t1));
   li(s0, Operand(stack_space));
-  LeaveExitFrame(false, s0);
-  Ret();
+  LeaveExitFrame(false, s0, true);
 
   bind(&promote_scheduled_exception);
   TailCallExternalReference(
@@ -4161,8 +4172,8 @@ void MacroAssembler::CallRuntime(const Runtime::Function* f,
   // arguments passed in because it is constant. At some point we
   // should remove this need and make the runtime routine entry code
   // smarter.
-  li(a0, num_arguments);
-  li(a1, Operand(ExternalReference(f, isolate())));
+  PrepareCEntryArgs(num_arguments);
+  PrepareCEntryFunction(ExternalReference(f, isolate()));
   CEntryStub stub(1);
   CallStub(&stub);
 }
@@ -4170,8 +4181,8 @@ void MacroAssembler::CallRuntime(const Runtime::Function* f,
 
 void MacroAssembler::CallRuntimeSaveDoubles(Runtime::FunctionId id) {
   const Runtime::Function* function = Runtime::FunctionForId(id);
-  li(a0, Operand(function->nargs));
-  li(a1, Operand(ExternalReference(function, isolate())));
+  PrepareCEntryArgs(function->nargs);
+  PrepareCEntryFunction(ExternalReference(function, isolate()));
   CEntryStub stub(1, kSaveFPRegs);
   CallStub(&stub);
 }
@@ -4183,12 +4194,13 @@ void MacroAssembler::CallRuntime(Runtime::FunctionId fid, int num_arguments) {
 
 
 void MacroAssembler::CallExternalReference(const ExternalReference& ext,
-                                           int num_arguments) {
-  li(a0, Operand(num_arguments));
-  li(a1, Operand(ext));
+                                           int num_arguments,
+                                           BranchDelaySlot bd) {
+  PrepareCEntryArgs(num_arguments);
+  PrepareCEntryFunction(ext);
 
   CEntryStub stub(1);
-  CallStub(&stub);
+  CallStub(&stub, al, zero_reg, Operand(zero_reg), bd);
 }
 
 
@@ -4199,7 +4211,7 @@ void MacroAssembler::TailCallExternalReference(const ExternalReference& ext,
   // arguments passed in because it is constant. At some point we
   // should remove this need and make the runtime routine entry code
   // smarter.
-  li(a0, Operand(num_arguments));
+  PrepareCEntryArgs(num_arguments);
   JumpToExternalReference(ext);
 }
 
@@ -4213,10 +4225,16 @@ void MacroAssembler::TailCallRuntime(Runtime::FunctionId fid,
 }
 
 
-void MacroAssembler::JumpToExternalReference(const ExternalReference& builtin) {
-  li(a1, Operand(builtin));
+void MacroAssembler::JumpToExternalReference(const ExternalReference& builtin,
+                                             BranchDelaySlot bd) {
+  PrepareCEntryFunction(builtin);
   CEntryStub stub(1);
-  Jump(stub.GetCode(), RelocInfo::CODE_TARGET);
+  Jump(stub.GetCode(),
+       RelocInfo::CODE_TARGET,
+       al,
+       zero_reg,
+       Operand(zero_reg),
+       bd);
 }
 
 
@@ -4563,7 +4581,8 @@ void MacroAssembler::EnterExitFrame(bool save_doubles,
 
 
 void MacroAssembler::LeaveExitFrame(bool save_doubles,
-                                    Register argument_count) {
+                                    Register argument_count,
+                                    bool do_return) {
   // Optionally restore all double registers.
   if (save_doubles) {
     // Remember: we only need to restore every 2nd double FPU value.
@@ -4589,11 +4608,17 @@ void MacroAssembler::LeaveExitFrame(bool save_doubles,
   mov(sp, fp);  // Respect ABI stack constraint.
   lw(fp, MemOperand(sp, ExitFrameConstants::kCallerFPOffset));
   lw(ra, MemOperand(sp, ExitFrameConstants::kCallerPCOffset));
-  addiu(sp, sp, 8);
+
   if (argument_count.is_valid()) {
     sll(t8, argument_count, kPointerSizeLog2);
     addu(sp, sp, t8);
   }
+
+  if (do_return) {
+    Ret(USE_DELAY_SLOT);
+    // If returning, the instruction in the delay slot will be the addiu below.
+  }
+  addiu(sp, sp, 8);
 }
 
 

src/mips/macro-assembler-mips.h

@@ -193,10 +193,14 @@ class MacroAssembler: public Assembler {
             Register reg = no_reg,
             const Operand& op = Operand(no_reg));
 
-  void DropAndRet(int drop = 0,
-                  Condition cond = cc_always,
-                  Register reg = no_reg,
-                  const Operand& op = Operand(no_reg));
+  // Trivial case of DropAndRet that utilizes the delay slot and only emits
+  // 2 instructions.
+  void DropAndRet(int drop);
+
+  void DropAndRet(int drop,
+                  Condition cond,
+                  Register reg,
+                  const Operand& op);
 
   // Swap two registers.  If the scratch register is omitted then a slightly
   // less efficient form using xor instead of mov is emitted.
@@ -773,7 +777,9 @@ class MacroAssembler: public Assembler {
                       int stack_space = 0);
 
   // Leave the current exit frame.
-  void LeaveExitFrame(bool save_doubles, Register arg_count);
+  void LeaveExitFrame(bool save_doubles,
+                      Register arg_count,
+                      bool do_return = false);
 
   // Get the actual activation frame alignment for target environment.
   static int ActivationFrameAlignment();
@@ -1084,9 +1090,22 @@ class MacroAssembler: public Assembler {
   // -------------------------------------------------------------------------
   // Runtime calls.
 
+  // See comments at the beginning of CEntryStub::Generate.
+  inline void PrepareCEntryArgs(int num_args) {
+    li(s0, num_args);
+    li(s1, (num_args - 1) * kPointerSize);
+  }
+
+  inline void PrepareCEntryFunction(const ExternalReference& ref) {
+    li(s2, Operand(ref));
+  }
+
   // Call a code stub.
-  void CallStub(CodeStub* stub, Condition cond = cc_always,
-                Register r1 = zero_reg, const Operand& r2 = Operand(zero_reg));
+  void CallStub(CodeStub* stub,
+                Condition cond = cc_always,
+                Register r1 = zero_reg,
+                const Operand& r2 = Operand(zero_reg),
+                BranchDelaySlot bd = PROTECT);
 
   // Tail call a code stub (jump).
   void TailCallStub(CodeStub* stub);
@@ -1102,7 +1121,8 @@ class MacroAssembler: public Assembler {
 
   // Convenience function: call an external reference.
   void CallExternalReference(const ExternalReference& ext,
-                             int num_arguments);
+                             int num_arguments,
+                             BranchDelaySlot bd = PROTECT);
 
   // Tail call of a runtime routine (jump).
   // Like JumpToExternalReference, but also takes care of passing the number
@@ -1168,7 +1188,8 @@ class MacroAssembler: public Assembler {
   void CallApiFunctionAndReturn(ExternalReference function, int stack_space);
 
   // Jump to the builtin routine.
-  void JumpToExternalReference(const ExternalReference& builtin);
+  void JumpToExternalReference(const ExternalReference& builtin,
+                               BranchDelaySlot bd = PROTECT);
 
   // Invoke specified builtin JavaScript function. Adds an entry to
   // the unresolved list if the name does not resolve.

src/mips/simulator-mips.h

@@ -309,6 +309,14 @@ class Simulator {
   void InstructionDecode(Instruction* instr);
   // Execute one instruction placed in a branch delay slot.
   void BranchDelayInstructionDecode(Instruction* instr) {
+    if (instr->InstructionBits() == nopInstr) {
+      // Short-cut generic nop instructions. They are always valid and they
+      // never change the simulator state.
+      set_register(pc, reinterpret_cast<int32_t>(instr) +
+                       Instruction::kInstrSize);
+      return;
+    }
+
     if (instr->IsForbiddenInBranchDelay()) {
       V8_Fatal(__FILE__, __LINE__,
                "Eror:Unexpected %i opcode in a branch delay slot.",

src/mips/stub-cache-mips.cc

@@ -577,8 +577,8 @@ static void CompileCallLoadPropertyWithInterceptor(
   ExternalReference ref =
       ExternalReference(IC_Utility(IC::kLoadPropertyWithInterceptorOnly),
           masm->isolate());
-  __ li(a0, Operand(5));
-  __ li(a1, Operand(ref));
+  __ PrepareCEntryArgs(5);
+  __ PrepareCEntryFunction(ref);
 
   CEntryStub stub(1);
   __ CallStub(&stub);
@@ -4107,7 +4107,8 @@ void KeyedLoadStubCompiler::GenerateLoadFastElement(MacroAssembler* masm) {
   // have been verified by the caller to not be a smi.
 
   // Check that the key is a smi.
-  __ JumpIfNotSmi(a0, &miss_force_generic);
+  __ JumpIfNotSmi(a0, &miss_force_generic, at, USE_DELAY_SLOT);
+  // The delay slot can be safely used here, a1 is an object pointer.
 
   // Get the elements array.
   __ lw(a2, FieldMemOperand(a1, JSObject::kElementsOffset));
@@ -4115,7 +4116,7 @@ void KeyedLoadStubCompiler::GenerateLoadFastElement(MacroAssembler* masm) {
 
   // Check that the key is within bounds.
   __ lw(a3, FieldMemOperand(a2, FixedArray::kLengthOffset));
-  __ Branch(&miss_force_generic, hs, a0, Operand(a3));
+  __ Branch(USE_DELAY_SLOT, &miss_force_generic, hs, a0, Operand(a3));
 
   // Load the result and make sure it's not the hole.
   __ Addu(a3, a2, Operand(FixedArray::kHeaderSize - kHeapObjectTag));
@@ -4125,8 +4126,8 @@ void KeyedLoadStubCompiler::GenerateLoadFastElement(MacroAssembler* masm) {
   __ lw(t0, MemOperand(t0));
   __ LoadRoot(t1, Heap::kTheHoleValueRootIndex);
   __ Branch(&miss_force_generic, eq, t0, Operand(t1));
+  __ Ret(USE_DELAY_SLOT);
   __ mov(v0, t0);
-  __ Ret();
 
   __ bind(&miss_force_generic);
   Handle<Code> stub =