X64 tweaks.

Use movapd instead of movsd to move between xmm registers. Avoids partial
register writes and potential stalls.

Use xorps instead of xorpd to zero a register. Same effect and latency, but
one byte smaller.

Review URL: http://codereview.chromium.org/6873006

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@7631 ce2b1a6d-e550-0410-aec6-3dcde31c8c00

src/x64/assembler-x64.cc

@@ -2540,6 +2540,24 @@ void Assembler::movq(Register dst, XMMRegister src) {
 }
 
 
+void Assembler::movq(XMMRegister dst, XMMRegister src) {
+  EnsureSpace ensure_space(this);
+  if (dst.low_bits() == 4) {
+    // Avoid unnecessary SIB byte.
+    emit(0xf3);
+    emit_optional_rex_32(dst, src);
+    emit(0x0F);
+    emit(0x7e);
+    emit_sse_operand(dst, src);
+  } else {
+    emit(0x66);
+    emit_optional_rex_32(src, dst);
+    emit(0x0F);
+    emit(0xD6);
+    emit_sse_operand(src, dst);
+  }
+}
+
 void Assembler::movdqa(const Operand& dst, XMMRegister src) {
   EnsureSpace ensure_space(this);
   emit(0x66);
@@ -2603,6 +2621,42 @@ void Assembler::movsd(XMMRegister dst, const Operand& src) {
 }
 
 
+void Assembler::movaps(XMMRegister dst, XMMRegister src) {
+  EnsureSpace ensure_space(this);
+  if (src.low_bits() == 4) {
+    // Try to avoid an unnecessary SIB byte.
+    emit_optional_rex_32(src, dst);
+    emit(0x0F);
+    emit(0x29);
+    emit_sse_operand(src, dst);
+  } else {
+    emit_optional_rex_32(dst, src);
+    emit(0x0F);
+    emit(0x28);
+    emit_sse_operand(dst, src);
+  }
+}
+
+
+void Assembler::movapd(XMMRegister dst, XMMRegister src) {
+  EnsureSpace ensure_space(this);
+  if (src.low_bits() == 4) {
+    // Try to avoid an unnecessary SIB byte.
+    emit(0x66);
+    emit_optional_rex_32(src, dst);
+    emit(0x0F);
+    emit(0x29);
+    emit_sse_operand(src, dst);
+  } else {
+    emit(0x66);
+    emit_optional_rex_32(dst, src);
+    emit(0x0F);
+    emit(0x28);
+    emit_sse_operand(dst, src);
+  }
+}
+
+
 void Assembler::movss(XMMRegister dst, const Operand& src) {
   EnsureSpace ensure_space(this);
   emit(0xF3);  // single
@@ -2833,6 +2887,15 @@ void Assembler::xorpd(XMMRegister dst, XMMRegister src) {
 }
 
 
+void Assembler::xorps(XMMRegister dst, XMMRegister src) {
+  EnsureSpace ensure_space(this);
+  emit_optional_rex_32(dst, src);
+  emit(0x0F);
+  emit(0x57);
+  emit_sse_operand(dst, src);
+}
+
+
 void Assembler::sqrtsd(XMMRegister dst, XMMRegister src) {
   EnsureSpace ensure_space(this);
   emit(0xF2);

src/x64/assembler-x64.h

@@ -1291,15 +1291,24 @@ class Assembler : public AssemblerBase {
   void movd(Register dst, XMMRegister src);
   void movq(XMMRegister dst, Register src);
   void movq(Register dst, XMMRegister src);
+  void movq(XMMRegister dst, XMMRegister src);
   void extractps(Register dst, XMMRegister src, byte imm8);
 
-  void movsd(const Operand& dst, XMMRegister src);
+  // Don't use this unless it's important to keep the
+  // top half of the destination register unchanged.
+  // Used movaps when moving double values and movq for integer
+  // values in xmm registers.
   void movsd(XMMRegister dst, XMMRegister src);
+
+  void movsd(const Operand& dst, XMMRegister src);
   void movsd(XMMRegister dst, const Operand& src);
 
   void movdqa(const Operand& dst, XMMRegister src);
   void movdqa(XMMRegister dst, const Operand& src);
 
+  void movapd(XMMRegister dst, XMMRegister src);
+  void movaps(XMMRegister dst, XMMRegister src);
+
   void movss(XMMRegister dst, const Operand& src);
   void movss(const Operand& dst, XMMRegister src);
 
@@ -1331,6 +1340,7 @@ class Assembler : public AssemblerBase {
   void andpd(XMMRegister dst, XMMRegister src);
   void orpd(XMMRegister dst, XMMRegister src);
   void xorpd(XMMRegister dst, XMMRegister src);
+  void xorps(XMMRegister dst, XMMRegister src);
   void sqrtsd(XMMRegister dst, XMMRegister src);
 
   void ucomisd(XMMRegister dst, XMMRegister src);

src/x64/code-stubs-x64.cc

@@ -266,7 +266,7 @@ void ToBooleanStub::Generate(MacroAssembler* masm) {
   __ j(not_equal, &true_result);
   // HeapNumber => false iff +0, -0, or NaN.
   // These three cases set the zero flag when compared to zero using ucomisd.
-  __ xorpd(xmm0, xmm0);
+  __ xorps(xmm0, xmm0);
   __ ucomisd(xmm0, FieldOperand(rax, HeapNumber::kValueOffset));
   __ j(zero, &false_result);
   // Fall through to |true_result|.
@@ -1602,7 +1602,7 @@ void MathPowStub::Generate(MacroAssembler* masm) {
   __ bind(&no_neg);
 
   // Load xmm1 with 1.
-  __ movsd(xmm1, xmm3);
+  __ movaps(xmm1, xmm3);
   NearLabel while_true;
   NearLabel no_multiply;
 
@@ -1620,8 +1620,8 @@ void MathPowStub::Generate(MacroAssembler* masm) {
   __ j(positive, &allocate_return);
   // Special case if xmm1 has reached infinity.
   __ divsd(xmm3, xmm1);
-  __ movsd(xmm1, xmm3);
-  __ xorpd(xmm0, xmm0);
+  __ movaps(xmm1, xmm3);
+  __ xorps(xmm0, xmm0);
   __ ucomisd(xmm0, xmm1);
   __ j(equal, &call_runtime);
 
@@ -1669,11 +1669,11 @@ void MathPowStub::Generate(MacroAssembler* masm) {
 
   // Calculates reciprocal of square root.
   // sqrtsd returns -0 when input is -0.  ECMA spec requires +0.
-  __ xorpd(xmm1, xmm1);
+  __ xorps(xmm1, xmm1);
   __ addsd(xmm1, xmm0);
   __ sqrtsd(xmm1, xmm1);
   __ divsd(xmm3, xmm1);
-  __ movsd(xmm1, xmm3);
+  __ movaps(xmm1, xmm3);
   __ jmp(&allocate_return);
 
   // Test for 0.5.
@@ -1686,8 +1686,8 @@ void MathPowStub::Generate(MacroAssembler* masm) {
   __ j(not_equal, &call_runtime);
   // Calculates square root.
   // sqrtsd returns -0 when input is -0.  ECMA spec requires +0.
-  __ xorpd(xmm1, xmm1);
-  __ addsd(xmm1, xmm0);
+  __ xorps(xmm1, xmm1);
+  __ addsd(xmm1, xmm0);  // Convert -0 to 0.
   __ sqrtsd(xmm1, xmm1);
 
   __ bind(&allocate_return);

src/x64/disasm-x64.cc

@@ -1033,7 +1033,14 @@ int DisassemblerX64::TwoByteOpcodeInstruction(byte* data) {
       }
     } else {
       get_modrm(*current, &mod, &regop, &rm);
-      if (opcode == 0x6E) {
+      if (opcode == 0x28) {
+        AppendToBuffer("movapd %s, ", NameOfXMMRegister(regop));
+        current += PrintRightXMMOperand(current);
+      } else if (opcode == 0x29) {
+        AppendToBuffer("movapd ");
+        current += PrintRightXMMOperand(current);
+        AppendToBuffer(", %s", NameOfXMMRegister(regop));
+      } else if (opcode == 0x6E) {
         AppendToBuffer("mov%c %s,",
                        rex_w() ? 'q' : 'd',
                        NameOfXMMRegister(regop));
@@ -1051,6 +1058,10 @@ int DisassemblerX64::TwoByteOpcodeInstruction(byte* data) {
         AppendToBuffer("movdqa ");
         current += PrintRightXMMOperand(current);
         AppendToBuffer(", %s", NameOfXMMRegister(regop));
+      } else if (opcode == 0xD6) {
+        AppendToBuffer("movq ");
+        current += PrintRightXMMOperand(current);
+        AppendToBuffer(", %s", NameOfXMMRegister(regop));
       } else {
         const char* mnemonic = "?";
         if (opcode == 0x50) {
@@ -1152,6 +1163,11 @@ int DisassemblerX64::TwoByteOpcodeInstruction(byte* data) {
       get_modrm(*current, &mod, &regop, &rm);
       AppendToBuffer("cvtss2sd %s,", NameOfXMMRegister(regop));
       current += PrintRightXMMOperand(current);
+    } else if (opcode == 0x7E) {
+      int mod, regop, rm;
+      get_modrm(*current, &mod, &regop, &rm);
+      AppendToBuffer("movq %s, ", NameOfXMMRegister(regop));
+      current += PrintRightXMMOperand(current);
     } else {
       UnimplementedInstruction();
     }
@@ -1169,6 +1185,22 @@ int DisassemblerX64::TwoByteOpcodeInstruction(byte* data) {
       current += 4;
     }  // else no immediate displacement.
     AppendToBuffer("nop");
+
+  } else if (opcode == 28) {
+    // movaps xmm, xmm/m128
+    int mod, regop, rm;
+    get_modrm(*current, &mod, &regop, &rm);
+    AppendToBuffer("movaps %s, ", NameOfXMMRegister(regop));
+    current += PrintRightXMMOperand(current);
+
+  } else if (opcode == 29) {
+    // movaps xmm/m128, xmm
+    int mod, regop, rm;
+    get_modrm(*current, &mod, &regop, &rm);
+    AppendToBuffer("movaps");
+    current += PrintRightXMMOperand(current);
+    AppendToBuffer(", %s", NameOfXMMRegister(regop));
+
   } else if (opcode == 0xA2 || opcode == 0x31) {
     // RDTSC or CPUID
     AppendToBuffer("%s", mnemonic);
@@ -1180,6 +1212,13 @@ int DisassemblerX64::TwoByteOpcodeInstruction(byte* data) {
     byte_size_operand_ = idesc.byte_size_operation;
     current += PrintOperands(idesc.mnem, idesc.op_order_, current);
 
+  } else if (opcode == 57) {
+    // xoprps xmm, xmm/m128
+    int mod, regop, rm;
+    get_modrm(*current, &mod, &regop, &rm);
+    AppendToBuffer("xorps %s, ", NameOfXMMRegister(regop));
+    current += PrintRightXMMOperand(current);
+
   } else if ((opcode & 0xF0) == 0x80) {
     // Jcc: Conditional jump (branch).
     current = data + JumpConditional(data);

src/x64/full-codegen-x64.cc

@@ -2758,7 +2758,7 @@ void FullCodeGenerator::EmitRandomHeapNumber(ZoneList<Expression*>* args) {
   __ movd(xmm1, rcx);
   __ movd(xmm0, rax);
   __ cvtss2sd(xmm1, xmm1);
-  __ xorpd(xmm0, xmm1);
+  __ xorps(xmm0, xmm1);
   __ subsd(xmm0, xmm1);
   __ movsd(FieldOperand(rbx, HeapNumber::kValueOffset), xmm0);
 

src/x64/lithium-codegen-x64.cc

@@ -1111,7 +1111,7 @@ void LCodeGen::DoConstantD(LConstantD* instr) {
   // Use xor to produce +0.0 in a fast and compact way, but avoid to
   // do so if the constant is -0.0.
   if (int_val == 0) {
-    __ xorpd(res, res);
+    __ xorps(res, res);
   } else {
     Register tmp = ToRegister(instr->TempAt(0));
     __ Set(tmp, int_val);
@@ -1223,12 +1223,12 @@ void LCodeGen::DoArithmeticD(LArithmeticD* instr) {
       break;
     case Token::MOD:
       __ PrepareCallCFunction(2);
-      __ movsd(xmm0, left);
+      __ movaps(xmm0, left);
       ASSERT(right.is(xmm1));
       __ CallCFunction(
           ExternalReference::double_fp_operation(Token::MOD, isolate()), 2);
       __ movq(rsi, Operand(rbp, StandardFrameConstants::kContextOffset));
-      __ movsd(result, xmm0);
+      __ movaps(result, xmm0);
       break;
     default:
       UNREACHABLE();
@@ -1287,7 +1287,7 @@ void LCodeGen::DoBranch(LBranch* instr) {
     EmitBranch(true_block, false_block, not_zero);
   } else if (r.IsDouble()) {
     XMMRegister reg = ToDoubleRegister(instr->InputAt(0));
-    __ xorpd(xmm0, xmm0);
+    __ xorps(xmm0, xmm0);
     __ ucomisd(reg, xmm0);
     EmitBranch(true_block, false_block, not_equal);
   } else {
@@ -1322,7 +1322,7 @@ void LCodeGen::DoBranch(LBranch* instr) {
 
       // HeapNumber => false iff +0, -0, or NaN. These three cases set the
       // zero flag when compared to zero using ucomisd.
-      __ xorpd(xmm0, xmm0);
+      __ xorps(xmm0, xmm0);
       __ ucomisd(xmm0, FieldOperand(reg, HeapNumber::kValueOffset));
       __ j(zero, false_label);
       __ jmp(true_label);
@@ -2671,7 +2671,7 @@ void LCodeGen::DoMathAbs(LUnaryMathOperation* instr) {
   if (r.IsDouble()) {
     XMMRegister scratch = xmm0;
     XMMRegister input_reg = ToDoubleRegister(instr->InputAt(0));
-    __ xorpd(scratch, scratch);
+    __ xorps(scratch, scratch);
     __ subsd(scratch, input_reg);
     __ andpd(input_reg, scratch);
   } else if (r.IsInteger32()) {
@@ -2708,7 +2708,7 @@ void LCodeGen::DoMathFloor(LUnaryMathOperation* instr) {
     __ cmpl(output_reg, Immediate(0x80000000));
     DeoptimizeIf(equal, instr->environment());
   } else {
-    __ xorpd(xmm_scratch, xmm_scratch);  // Zero the register.
+    __ xorps(xmm_scratch, xmm_scratch);  // Zero the register.
     __ ucomisd(input_reg, xmm_scratch);
 
     if (instr->hydrogen()->CheckFlag(HValue::kBailoutOnMinusZero)) {
@@ -2784,7 +2784,7 @@ void LCodeGen::DoMathPowHalf(LUnaryMathOperation* instr) {
   XMMRegister xmm_scratch = xmm0;
   XMMRegister input_reg = ToDoubleRegister(instr->InputAt(0));
   ASSERT(ToDoubleRegister(instr->result()).is(input_reg));
-  __ xorpd(xmm_scratch, xmm_scratch);
+  __ xorps(xmm_scratch, xmm_scratch);
   __ addsd(input_reg, xmm_scratch);  // Convert -0 to +0.
   __ sqrtsd(input_reg, input_reg);
 }
@@ -2800,7 +2800,7 @@ void LCodeGen::DoPower(LPower* instr) {
   if (exponent_type.IsDouble()) {
     __ PrepareCallCFunction(2);
     // Move arguments to correct registers
-    __ movsd(xmm0, left_reg);
+    __ movaps(xmm0, left_reg);
     ASSERT(ToDoubleRegister(right).is(xmm1));
     __ CallCFunction(
         ExternalReference::power_double_double_function(isolate()), 2);
@@ -2808,7 +2808,7 @@ void LCodeGen::DoPower(LPower* instr) {
     __ PrepareCallCFunction(2);
     // Move arguments to correct registers: xmm0 and edi (not rdi).
     // On Windows, the registers are xmm0 and edx.
-    __ movsd(xmm0, left_reg);
+    __ movaps(xmm0, left_reg);
 #ifdef _WIN64
     ASSERT(ToRegister(right).is(rdx));
 #else
@@ -2834,13 +2834,13 @@ void LCodeGen::DoPower(LPower* instr) {
     __ bind(&call);
     __ PrepareCallCFunction(2);
     // Move arguments to correct registers xmm0 and xmm1.
-    __ movsd(xmm0, left_reg);
+    __ movaps(xmm0, left_reg);
     // Right argument is already in xmm1.
     __ CallCFunction(
         ExternalReference::power_double_double_function(isolate()), 2);
   }
   // Return value is in xmm0.
-  __ movsd(result_reg, xmm0);
+  __ movaps(result_reg, xmm0);
   // Restore context register.
   __ movq(rsi, Operand(rbp, StandardFrameConstants::kContextOffset));
 }
@@ -3422,7 +3422,7 @@ void LCodeGen::EmitNumberUntagD(Register input_reg,
   DeoptimizeIf(not_equal, env);
 
   // Convert undefined to NaN. Compute NaN as 0/0.
-  __ xorpd(result_reg, result_reg);
+  __ xorps(result_reg, result_reg);
   __ divsd(result_reg, result_reg);
   __ jmp(&done);
 

src/x64/lithium-gap-resolver-x64.cc

@@ -214,7 +214,7 @@ void LGapResolver::EmitMove(int index) {
   } else if (source->IsDoubleRegister()) {
     XMMRegister src = cgen_->ToDoubleRegister(source);
     if (destination->IsDoubleRegister()) {
-      __ movsd(cgen_->ToDoubleRegister(destination), src);
+      __ movaps(cgen_->ToDoubleRegister(destination), src);
     } else {
       ASSERT(destination->IsDoubleStackSlot());
       __ movsd(cgen_->ToOperand(destination), src);
@@ -273,9 +273,9 @@ void LGapResolver::EmitSwap(int index) {
     // Swap two double registers.
     XMMRegister source_reg = cgen_->ToDoubleRegister(source);
     XMMRegister destination_reg = cgen_->ToDoubleRegister(destination);
-    __ movsd(xmm0, source_reg);
-    __ movsd(source_reg, destination_reg);
-    __ movsd(destination_reg, xmm0);
+    __ movaps(xmm0, source_reg);
+    __ movaps(source_reg, destination_reg);
+    __ movaps(destination_reg, xmm0);
 
   } else if (source->IsDoubleRegister() || destination->IsDoubleRegister()) {
     // Swap a double register and a double stack slot.

src/x64/macro-assembler-x64.cc

@@ -1248,12 +1248,17 @@ void MacroAssembler::SmiAdd(Register dst,
                             Register src2) {
   // No overflow checking. Use only when it's known that
   // overflowing is impossible.
-  ASSERT(!dst.is(src2));
   if (!dst.is(src1)) {
-    movq(dst, src1);
+    if (emit_debug_code()) {
+      movq(kScratchRegister, src1);
+      addq(kScratchRegister, src2);
+      Check(no_overflow, "Smi addition overflow");
+    }
+    lea(dst, Operand(src1, src2, times_1, 0));
+  } else {
+    addq(dst, src2);
+    Assert(no_overflow, "Smi addition overflow");
   }
-  addq(dst, src2);
-  Assert(no_overflow, "Smi addition overflow");
 }