ARM: Add optimization for constant RHS in DoMulI.

Patch by ARM Ltd.

Review URL: http://codereview.chromium.org/7027033

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@8172 ce2b1a6d-e550-0410-aec6-3dcde31c8c00

src/arm/code-stubs-arm.cc

@@ -2666,9 +2666,11 @@ void BinaryOpStub::GenerateInt32Stub(MacroAssembler* masm) {
           // DIV just falls through to allocating a heap number.
         }
 
+        __ bind(&return_heap_number);
+        // Return a heap number, or fall through to type transition or runtime
+        // call if we can't.
         if (result_type_ >= (op_ == Token::DIV) ? BinaryOpIC::HEAP_NUMBER
                                                 : BinaryOpIC::INT32) {
-          __ bind(&return_heap_number);
           // We are using vfp registers so r5 is available.
           heap_number_result = r5;
           GenerateHeapResultAllocation(masm,
@@ -2842,7 +2844,11 @@ void BinaryOpStub::GenerateInt32Stub(MacroAssembler* masm) {
       UNREACHABLE();
   }
 
-  if (transition.is_linked()) {
+  // We never expect DIV to yield an integer result, so we always generate
+  // type transition code for DIV operations expecting an integer result: the
+  // code will fall through to this type transition.
+  if (transition.is_linked() ||
+      ((op_ == Token::DIV) && (result_type_ <= BinaryOpIC::INT32))) {
     __ bind(&transition);
     GenerateTypeTransition(masm);
   }

src/arm/lithium-arm.cc

@@ -1412,13 +1412,16 @@ LInstruction* LChunkBuilder::DoMul(HMul* instr) {
     LOperand* left = UseRegisterAtStart(instr->LeastConstantOperand());
     LOperand* right = UseOrConstant(instr->MostConstantOperand());
     LOperand* temp = NULL;
-    if (instr->CheckFlag(HValue::kBailoutOnMinusZero)) {
+    if (instr->CheckFlag(HValue::kBailoutOnMinusZero) &&
+        (instr->CheckFlag(HValue::kCanOverflow) ||
+        !right->IsConstantOperand())) {
       temp = TempRegister();
     }
-    LMulI* mul = new LMulI(left, right, temp);
-    return AssignEnvironment(DefineSameAsFirst(mul));
+    return AssignEnvironment(DefineSameAsFirst(new LMulI(left, right, temp)));
+
   } else if (instr->representation().IsDouble()) {
     return DoArithmeticD(Token::MUL, instr);
+
   } else {
     return DoArithmeticT(Token::MUL, instr);
   }

src/arm/lithium-codegen-arm.cc

@@ -1118,40 +1118,98 @@ void LCodeGen::DoDeferredBinaryOpStub(LTemplateInstruction<1, 2, T>* instr,
 
 
 void LCodeGen::DoMulI(LMulI* instr) {
+  ASSERT(instr->result()->Equals(instr->InputAt(0)));
   Register scratch = scratch0();
+  Register result = ToRegister(instr->result());
   Register left = ToRegister(instr->InputAt(0));
-  Register right = EmitLoadRegister(instr->InputAt(1), scratch);
+  LOperand* right_op = instr->InputAt(1);
 
-  if (instr->hydrogen()->CheckFlag(HValue::kBailoutOnMinusZero) &&
-      !instr->InputAt(1)->IsConstantOperand()) {
-    __ orr(ToRegister(instr->TempAt(0)), left, right);
-  }
+  bool can_overflow = instr->hydrogen()->CheckFlag(HValue::kCanOverflow);
+  bool bailout_on_minus_zero =
+    instr->hydrogen()->CheckFlag(HValue::kBailoutOnMinusZero);
+
+  if (right_op->IsConstantOperand() && !can_overflow) {
+    // Use optimized code for specific constants.
+    int32_t constant = ToInteger32(LConstantOperand::cast(right_op));
+
+    if (bailout_on_minus_zero && (constant < 0)) {
+      // The case of a null constant will be handled separately.
+      // If constant is negative and left is null, the result should be -0.
+      __ cmp(left, Operand(0));
+      DeoptimizeIf(eq, instr->environment());
+    }
+
+    switch (constant) {
+      case -1:
+        __ rsb(result, left, Operand(0));
+        break;
+      case 0:
+        if (bailout_on_minus_zero) {
+          // If left is strictly negative and the constant is null, the
+          // result is -0. Deoptimize if required, otherwise return 0.
+          __ cmp(left, Operand(0));
+          DeoptimizeIf(mi, instr->environment());
+        }
+        __ mov(result, Operand(0));
+        break;
+      case 1:
+        // Nothing to do.
+        break;
+      default:
+        // Multiplying by powers of two and powers of two plus or minus
+        // one can be done faster with shifted operands.
+        // For other constants we emit standard code.
+        int32_t mask = constant >> 31;
+        uint32_t constant_abs = (constant + mask) ^ mask;
+
+        if (IsPowerOf2(constant_abs) ||
+            IsPowerOf2(constant_abs - 1) ||
+            IsPowerOf2(constant_abs + 1)) {
+          if (IsPowerOf2(constant_abs)) {
+            int32_t shift = WhichPowerOf2(constant_abs);
+            __ mov(result, Operand(left, LSL, shift));
+          } else if (IsPowerOf2(constant_abs - 1)) {
+            int32_t shift = WhichPowerOf2(constant_abs - 1);
+            __ add(result, left, Operand(left, LSL, shift));
+          } else if (IsPowerOf2(constant_abs + 1)) {
+            int32_t shift = WhichPowerOf2(constant_abs + 1);
+            __ rsb(result, left, Operand(left, LSL, shift));
+          }
+
+          // Correct the sign of the result is the constant is negative.
+          if (constant < 0)  __ rsb(result, result, Operand(0));
+
+        } else {
+          // Generate standard code.
+          __ mov(ip, Operand(constant));
+          __ mul(result, left, ip);
+        }
+    }
 
-  if (instr->hydrogen()->CheckFlag(HValue::kCanOverflow)) {
-    // scratch:left = left * right.
-    __ smull(left, scratch, left, right);
-    __ mov(ip, Operand(left, ASR, 31));
-    __ cmp(ip, Operand(scratch));
-    DeoptimizeIf(ne, instr->environment());
   } else {
-    __ mul(left, left, right);
-  }
+    Register right = EmitLoadRegister(right_op, scratch);
+    if (bailout_on_minus_zero) {
+      __ orr(ToRegister(instr->TempAt(0)), left, right);
+    }
 
-  if (instr->hydrogen()->CheckFlag(HValue::kBailoutOnMinusZero)) {
-    // Bail out if the result is supposed to be negative zero.
-    Label done;
-    __ cmp(left, Operand(0));
-    __ b(ne, &done);
-    if (instr->InputAt(1)->IsConstantOperand()) {
-      if (ToInteger32(LConstantOperand::cast(instr->InputAt(1))) <= 0) {
-        DeoptimizeIf(al, instr->environment());
-      }
+    if (can_overflow) {
+      // scratch:result = left * right.
+      __ smull(result, scratch, left, right);
+      __ cmp(scratch, Operand(result, ASR, 31));
+      DeoptimizeIf(ne, instr->environment());
     } else {
-      // Test the non-zero operand for negative sign.
+      __ mul(result, left, right);
+    }
+
+    if (bailout_on_minus_zero) {
+      // Bail out if the result is supposed to be negative zero.
+      Label done;
+      __ cmp(result, Operand(0));
+      __ b(ne, &done);
       __ cmp(ToRegister(instr->TempAt(0)), Operand(0));
       DeoptimizeIf(mi, instr->environment());
+      __ bind(&done);
     }
-    __ bind(&done);
   }
 }
 

src/arm/macro-assembler-arm.cc

@@ -330,7 +330,8 @@ void MacroAssembler::And(Register dst, Register src1, const Operand& src2,
              !src2.must_use_constant_pool() &&
              CpuFeatures::IsSupported(ARMv7) &&
              IsPowerOf2(src2.immediate() + 1)) {
-    ubfx(dst, src1, 0, WhichPowerOf2(src2.immediate() + 1), cond);
+    ubfx(dst, src1, 0,
+        WhichPowerOf2(static_cast<uint32_t>(src2.immediate()) + 1), cond);
 
   } else {
     and_(dst, src1, src2, LeaveCC, cond);

src/utils.h

@@ -52,11 +52,9 @@ static inline bool IsPowerOf2(T x) {
 
 
 // X must be a power of 2.  Returns the number of trailing zeros.
-template <typename T>
-static inline int WhichPowerOf2(T x) {
+static inline int WhichPowerOf2(uint32_t x) {
   ASSERT(IsPowerOf2(x));
   ASSERT(x != 0);
-  if (x < 0) return 31;
   int bits = 0;
 #ifdef DEBUG
   int original_x = x;