AArch64: Use CMP;CCMP sequences for and/or/setcc trees.

[MatzeB]

Summary

Previously CCMP/FCCMP instructions were only used by the
AArch64ConditionalCompares pass for control flow. This patch uses them
for SELECT like instructions as well by matching patterns in ISelLowering.

[MatzeB]

I am currently performing llvm-testsuite runs with this change. I'm seeing a couple of swings in the positive and negative direction. I'll investigate these further.

[MatzeB]

I investigated the performance changes in depth showing the change improves things:

• Most of the changes were noise. I did not consider anything changing less than 0.3%
• 9 Benchmarks improve (between 0.3% and 1.7%)
• 2 Regressions I have seen stem from the fact that we generate code for the left operand of the or first while previously we would first produce it for the right operand of the OR, this led to a case of MachineGVN not firing and another case where the slightly more limited range of immediates on ccmp compared to cmp mattered. Changing the code to handle the right side of the AND/OR first improved out luck...
• There is a regression 464.h264ref (0.5%) and 400.perlbench(0.4%). I couldn't see any obvious reasons when comparing the generated code, in fact many places improved. So I'm blaming cache effects and/or measuring noise.

[MatzeB]

ping

[MatzeB]

ping

[MatzeB]

ping 3

[jmolloy]

Hi Matthias,

Thanks for working on this, this looks great!

I have some comments below. The patch is also a bit big and touches different areas (ISel & tablegen) but I see there is no real way to test just the tablegen changes in isolation so I think this is OK.

I'd really like Tim to cast an eye over this before it goes in, so LGTM but hold off for his.

Cheers,

James

lib/Target/AArch64/AArch64ISelLowering.cpp

771

Have you missed FCCMP out here?

1094

Was this broken before, or are you changing something, or are you just making something more explicit?

1140

This would probably look nicer with the cast to ConstantSDNode taken out of the if:

auto *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0));
if (...)
else if (RHS.getOpcode() == ISD::SUB && C && C->isNullValue() ...

1154

I'm wondering whether this should have a maximum recursion depth. I can imagine it entirely possible that some auto-generated code could produce a *huge* conjunction ladder, and might cause us to overflow here.

1195

Need brackets around this one-liner as the else case has brackets.

lib/Target/AArch64/AArch64ISelLowering.h

503

& binds to the right.

[MatzeB]

Thanks for the review James.

The attached version is updated to ToT and addresses your comments. I backed out the unrelated cleanup of the node type.

[MatzeB]

ping

[MatzeB]

ping

[t.p.northover]

Hi Matthias,

Sorry this took so long. I couldn't spot any problems with the code either (after James's suggested cleanups).

Tim.

lib/Target/AArch64/AArch64ISelLowering.h

enum {
 
   // Arithmetic instructions which write flags.
   ADDS,
   SUBS,
   ADCS,
   SBCS,
   ANDS,
 
+  // Conditional compares. Operands: left,right,falsecc,cc,flags
+  CCMP,
+  CCMN,
+  FCCMP,
+
   // Floating point comparison
   FCMP,
 
   // Floating point max and min instructions.
   FMAX,
   FMIN,
 
   // Scalar extract
bool getPostIndexedAddressParts(SDNode *N, SDNode *Op, SDValue &Base,
                                   SelectionDAG &DAG) const override;
 
   void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue> &Results,
                           SelectionDAG &DAG) const override;
 
   bool functionArgumentNeedsConsecutiveRegisters(Type *Ty,
                                                  CallingConv::ID CallConv,
                                                  bool isVarArg) const override;
+
+  bool shouldNormalizeToSelectSequence(LLVMContext&, EVT) const override;
 };
 
 namespace AArch64 {
 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
                          const TargetLibraryInfo *libInfo);
 } // end namespace AArch64
 
 } // end namespace llvm
 
 #endif

lib/Target/AArch64/AArch64ISelLowering.cpp

 // FIXME: The necessary dtprel relocations don't seem to be supported
 // well in the GNU bfd and gold linkers at the moment. Therefore, by
 // default, for now, fall back to GeneralDynamic code generation.
 cl::opt<bool> EnableAArch64ELFLocalDynamicTLSGeneration(
     "aarch64-elf-ldtls-generation", cl::Hidden,
     cl::desc("Allow AArch64 Local Dynamic TLS code generation"),
     cl::init(false));
 
+/// Value type used for condition codes.
+static const MVT MVT_CC = MVT::i32;
+
 AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM,
                                              const AArch64Subtarget &STI)
     : TargetLowering(TM), Subtarget(&STI) {
 
   // AArch64 doesn't have comparisons which set GPRs or setcc instructions, so
   // we have to make something up. Arbitrarily, choose ZeroOrOne.
   setBooleanContents(ZeroOrOneBooleanContent);
   // When comparing vectors the result sets the different elements in the
const char *AArch64TargetLowering::getTargetNodeName(unsigned Opcode) const {
   case AArch64ISD::TLSDESC_CALLSEQ:   return "AArch64ISD::TLSDESC_CALLSEQ";
   case AArch64ISD::ADC:               return "AArch64ISD::ADC";
   case AArch64ISD::SBC:               return "AArch64ISD::SBC";
   case AArch64ISD::ADDS:              return "AArch64ISD::ADDS";
   case AArch64ISD::SUBS:              return "AArch64ISD::SUBS";
   case AArch64ISD::ADCS:              return "AArch64ISD::ADCS";
   case AArch64ISD::SBCS:              return "AArch64ISD::SBCS";
   case AArch64ISD::ANDS:              return "AArch64ISD::ANDS";
+  case AArch64ISD::CCMP:              return "AArch64ISD::CCMP";
+  case AArch64ISD::CCMN:              return "AArch64ISD::CCMN";
   case AArch64ISD::FCMP:              return "AArch64ISD::FCMP";
   case AArch64ISD::FMIN:              return "AArch64ISD::FMIN";
   case AArch64ISD::FMAX:              return "AArch64ISD::FMAX";
   case AArch64ISD::DUP:               return "AArch64ISD::DUP";
   case AArch64ISD::DUPLANE8:          return "AArch64ISD::DUPLANE8";
   case AArch64ISD::DUPLANE16:         return "AArch64ISD::DUPLANE16";
   case AArch64ISD::DUPLANE32:         return "AArch64ISD::DUPLANE32";
   case AArch64ISD::DUPLANE64:         return "AArch64ISD::DUPLANE64";
static bool isLegalArithImmed(uint64_t C) {
   return (C >> 12 == 0) || ((C & 0xFFFULL) == 0 && C >> 24 == 0);
 }
 
 static SDValue emitComparison(SDValue LHS, SDValue RHS, ISD::CondCode CC,
                               SDLoc dl, SelectionDAG &DAG) {
   EVT VT = LHS.getValueType();
 
   if (VT.isFloatingPoint())
-    return DAG.getNode(AArch64ISD::FCMP, dl, VT, LHS, RHS);
+    return DAG.getNode(AArch64ISD::FCMP, dl, MVT_CC, LHS, RHS);
 
   // The CMP instruction is just an alias for SUBS, and representing it as
   // SUBS means that it's possible to get CSE with subtract operations.
   // A later phase can perform the optimization of setting the destination
   // register to WZR/XZR if it ends up being unused.
   unsigned Opcode = AArch64ISD::SUBS;
 
   if (RHS.getOpcode() == ISD::SUB && isa<ConstantSDNode>(RHS.getOperand(0)) &&
if (RHS.getOpcode() == ISD::SUB && isa<ConstantSDNode>(RHS.getOperand(0)) &&
     // Similarly, (CMP (and X, Y), 0) can be implemented with a TST
     // (a.k.a. ANDS) except that the flags are only guaranteed to work for one
     // of the signed comparisons.
     Opcode = AArch64ISD::ANDS;
     RHS = LHS.getOperand(1);
     LHS = LHS.getOperand(0);
   }
 
-  return DAG.getNode(Opcode, dl, DAG.getVTList(VT, MVT::i32), LHS, RHS)
+  return DAG.getNode(Opcode, dl, DAG.getVTList(VT, MVT_CC), LHS, RHS)
       .getValue(1);
 }
 
+static SDValue emitConditionalComparison(SDValue LHS, SDValue RHS,
+                                         ISD::CondCode CC, SDValue CCOp,
+                                         SDValue Condition, unsigned NZCV,
+                                         SDLoc DL, SelectionDAG &DAG) {
+  unsigned Opcode;
+  if (LHS.getValueType().isFloatingPoint())
+    Opcode = AArch64ISD::FCCMP;
+  else if (RHS.getOpcode() == ISD::SUB &&
+           isa<ConstantSDNode>(RHS.getOperand(0)) &&
+           cast<ConstantSDNode>(RHS.getOperand(0))->isNullValue() &&
+           (CC == ISD::SETEQ || CC == ISD::SETNE)) {
+    // See emitComparison on why we can only do this for SETEQ and SETNE
+    Opcode = AArch64ISD::CCMN;
+    RHS = RHS.getOperand(1);
+  } else
+    Opcode = AArch64ISD::CCMP;
+
+  SDValue NZCVOp = DAG.getConstant(NZCV, MVT::i32);
+  return DAG.getNode(Opcode, DL, MVT_CC, LHS, RHS, NZCVOp, Condition, CCOp);
+}
+
+/// Returns true if @p Val is a tree of AND/OR/SETCC operations.
+static bool isConjunctionDisjunctionTree(const SDValue Val) {
+  if (!Val.hasOneUse())
+    return false;
+  if (Val->getOpcode() == ISD::SETCC)
+    return true;
+  if (Val->getOpcode() == ISD::AND || Val->getOpcode() == ISD::OR) {
+    SDValue O0 = Val->getOperand(0);
+    SDValue O1 = Val->getOperand(1);
+    return isConjunctionDisjunctionTree(O0) && isConjunctionDisjunctionTree(O1);
+  }
+  return false;
+}
+
+/// Emit conjunction or disjunction tree with the CMP/FCMP followed by a chain
+/// of CCMP/CFCMP ops. For example (SETCC_0 & SETCC_1) with condition cond0 and
+/// cond1 can be transformed into "CMP; CCMP" with CCMP executing on cond_0
+/// and setting flags to inversed(cond_1) otherwise.
+/// This recursive function produces DAG nodes that produce condition flags
+/// suitable to determine the truth value of @p Val (which is AND/OR/SETCC)
+/// by testing the result for the condition set to @p OutCC. If @p Negate is
+/// set the opposite truth value is produced. If @p CCOp and @p Condition are
+/// given then conditional comparison are created so that false is reported
+/// when they are false.
+static SDValue emitConjunctionDisjunctionTree(SelectionDAG &DAG, SDValue Val,
+                                              AArch64CC::CondCode &OutCC,
+                                              bool Negate,
+                                              SDValue CCOp = SDValue(),
+                                              AArch64CC::CondCode Condition =
+                                              AArch64CC::AL) {
+  assert(isConjunctionDisjunctionTree(Val));
+  // We're at a tree leaf, produce a c?f?cmp.
+  unsigned Opcode = Val->getOpcode();
+  if (Opcode == ISD::SETCC) {
+    SDValue LHS = Val->getOperand(0);
+    SDValue RHS = Val->getOperand(1);
+    ISD::CondCode CC = cast<CondCodeSDNode>(Val->getOperand(2))->get();
+    bool isInteger = LHS.getValueType().isInteger();
+    if (Negate)
+      CC = getSetCCInverse(CC, isInteger);
+    SDLoc DL(Val);
+    // Determine OutCC and handle FP special case.
+    if (isInteger)
+      OutCC = changeIntCCToAArch64CC(CC);
+    else {
+      assert(LHS.getValueType().isFloatingPoint());
+      AArch64CC::CondCode ExtraCC;
+      changeFPCCToAArch64CC(CC, OutCC, ExtraCC);
+      // Surpisingly some floatingpoint conditions can't be tested with a single
+      // condition code. Construct an additional comparison in this case.
+      // See comment below on how we deal with OR conditions.
+      if (ExtraCC != AArch64CC::AL) {
+        SDValue ExtraCmp;
+        if (!CCOp.getNode())
+          ExtraCmp = emitComparison(LHS, RHS, CC, DL, DAG);
+        else {
+          SDValue ConditionOp = DAG.getConstant(Condition, MVT_CC);
+          // Note that we want the inverse of ExtraCC, so NZCV is not inversed.
+          unsigned NZCV = AArch64CC::getNZCVToSatisfyCondCode(ExtraCC);
+          ExtraCmp = emitConditionalComparison(LHS, RHS, CC, CCOp, ConditionOp,
+                                               NZCV, DL, DAG);
+        }
+        CCOp = ExtraCmp;
+        Condition = AArch64CC::getInvertedCondCode(ExtraCC);
+        OutCC = AArch64CC::getInvertedCondCode(OutCC);
+      }
+    }
+
+    // Produce a normal comparison if we are first in the chain
+    if (!CCOp.getNode())
+      return emitComparison(LHS, RHS, CC, DL, DAG);
+    // Otherwise produce a ccmp.
+    SDValue ConditionOp = DAG.getConstant(Condition, MVT_CC);
+    AArch64CC::CondCode InvOutCC = AArch64CC::getInvertedCondCode(OutCC);
+    unsigned NZCV = AArch64CC::getNZCVToSatisfyCondCode(InvOutCC);
+    return emitConditionalComparison(LHS, RHS, CC, CCOp, ConditionOp, NZCV, DL,
+                                     DAG);
+  }
+
+  // Construct comparison sequence for the left hand side.
+  SDValue LHS = Val->getOperand(0);
+  SDValue RHS = Val->getOperand(1);
+
+  // We can only implement AND-like behaviour here, but negation is free. So we
+  // use (not (and (not x) (not y))) to implement (or x y).
+  bool isOr = Val->getOpcode() == ISD::OR;
+  assert((isOr || Val->getOpcode() == ISD::AND) && "Should have AND or OR.");
+  Negate ^= isOr;
+
+  AArch64CC::CondCode LHSCC;
+  SDValue CmpL = emitConjunctionDisjunctionTree(DAG, LHS, LHSCC, isOr, CCOp,
+                                                Condition);
+  SDValue CmpR = emitConjunctionDisjunctionTree(DAG, RHS, OutCC, isOr, CmpL,
+                                                LHSCC);
+  if (Negate)
+    OutCC = AArch64CC::getInvertedCondCode(OutCC);
+  return CmpR;
+}
+
 static SDValue getAArch64Cmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
                              SDValue &AArch64cc, SelectionDAG &DAG, SDLoc dl) {
   SDValue Cmp;
   AArch64CC::CondCode AArch64CC;
   if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
     EVT VT = RHS.getValueType();
     uint64_t C = RHSC->getZExtValue();
     if (!isLegalArithImmed(C)) {
if ((cast<ConstantSDNode>(RHS)->getZExtValue() >> 16 == 0) &&
                                CC, dl, DAG);
           AArch64CC = changeIntCCToAArch64CC(CC);
           AArch64cc = DAG.getConstant(AArch64CC, MVT::i32);
           return Cmp;
         }
       }
     }
   }
+
+  if ((CC == ISD::SETEQ || CC == ISD::SETNE) && isa<ConstantSDNode>(RHS) &&
+      cast<ConstantSDNode>(RHS)->isNullValue() &&
+      isConjunctionDisjunctionTree(LHS)) {
+    bool Negate = CC == ISD::SETEQ;
+    Cmp = emitConjunctionDisjunctionTree(DAG, LHS, AArch64CC, Negate);
+  } else {
     Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
     AArch64CC = changeIntCCToAArch64CC(CC);
-  AArch64cc = DAG.getConstant(AArch64CC, MVT::i32);
+  }
+
+  AArch64cc = DAG.getConstant(AArch64CC, MVT_CC);
   return Cmp;
 }
 
 static std::pair<SDValue, SDValue>
 getAArch64XALUOOp(AArch64CC::CondCode &CC, SDValue Op, SelectionDAG &DAG) {
   assert((Op.getValueType() == MVT::i32 || Op.getValueType() == MVT::i64) &&
          "Unsupported value type");
   SDValue Value, Overflow;
return Builder.CreateCall2(
                 Val, Stxr->getFunctionType()->getParamType(0)),
       Addr);
 }
 
 bool AArch64TargetLowering::functionArgumentNeedsConsecutiveRegisters(
     Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
   return Ty->isArrayTy();
 }
+
+bool AArch64TargetLowering::shouldNormalizeToSelectSequence(LLVMContext &,
+                                                            EVT) const {
+  return false;
+}

lib/Target/AArch64/AArch64InstrFormats.td

 
 // imm0_31 predicate - True if the immediate is in the range [0,31]
 def imm0_31 : Operand<i64>, ImmLeaf<i64, [{
   return ((uint64_t)Imm) < 32;
 }]> {
   let ParserMatchClass = Imm0_31Operand;
 }
 
+// True if the 32-bit immediate is in the range [0,31]
+def imm32_0_31 : Operand<i32>, ImmLeaf<i32, [{
+  return ((uint64_t)Imm) < 32;
+}]> {
+  let ParserMatchClass = Imm0_31Operand;
+}
+
 // imm0_15 predicate - True if the immediate is in the range [0,15]
 def imm0_15 : Operand<i64>, ImmLeaf<i64, [{
   return ((uint64_t)Imm) < 16;
 }]> {
   let ParserMatchClass = Imm0_15Operand;
 }
 
 // imm0_7 predicate - True if the immediate is in the range [0,7]
 def imm0_7 : Operand<i64>, ImmLeaf<i64, [{
   return ((uint64_t)Imm) < 8;
 }]> {
   let ParserMatchClass = Imm0_7Operand;
 }
 
 // imm32_0_15 predicate - True if the 32-bit immediate is in the range [0,15]
 def imm32_0_15 : Operand<i32>, ImmLeaf<i32, [{
   return ((uint32_t)Imm) < 16;
-}]>;
+}]> {
+  let ParserMatchClass = Imm0_15Operand;
+}
 
 // An arithmetic shifter operand:
 //  {7-6} - shift type: 00 = lsl, 01 = lsr, 10 = asr
 //  {5-0} - imm6
 class arith_shift<ValueType Ty, int width> : Operand<Ty> {
   let PrintMethod = "printShifter";
   let ParserMatchClass = !cast<AsmOperandClass>(
                          "ArithmeticShifterOperand" # width);
def : LogicalRegAlias<mnemonic,
                         !cast<Instruction>(NAME#"Xrs"), GPR64>;
 }
 
 //---
 // Conditionally set flags
 //---
 
 let mayLoad = 0, mayStore = 0, hasSideEffects = 0 in
-class BaseCondSetFlagsImm<bit op, RegisterClass regtype, string asm>
-    : I<(outs), (ins regtype:$Rn, imm0_31:$imm, imm0_15:$nzcv, ccode:$cond),
-         asm, "\t$Rn, $imm, $nzcv, $cond", "", []>,
+class BaseCondComparisonImm<bit op, RegisterClass regtype, ImmLeaf immtype,
+                            string mnemonic, SDNode OpNode>
+    : I<(outs), (ins regtype:$Rn, immtype:$imm, imm32_0_15:$nzcv, ccode:$cond),
+         mnemonic, "\t$Rn, $imm, $nzcv, $cond", "",
+         [(set NZCV, (OpNode regtype:$Rn, immtype:$imm, (i32 imm:$nzcv),
+                             (i32 imm:$cond), NZCV))]>,
       Sched<[WriteI, ReadI]> {
   let Uses = [NZCV];
   let Defs = [NZCV];
 
   bits<5> Rn;
   bits<5> imm;
   bits<4> nzcv;
   bits<4> cond;
 
   let Inst{30}    = op;
   let Inst{29-21} = 0b111010010;
   let Inst{20-16} = imm;
   let Inst{15-12} = cond;
   let Inst{11-10} = 0b10;
   let Inst{9-5}   = Rn;
   let Inst{4}     = 0b0;
   let Inst{3-0}   = nzcv;
 }
 
-multiclass CondSetFlagsImm<bit op, string asm> {
-  def Wi : BaseCondSetFlagsImm<op, GPR32, asm> {
-    let Inst{31} = 0;
-  }
-  def Xi : BaseCondSetFlagsImm<op, GPR64, asm> {
-    let Inst{31} = 1;
-  }
-}
-
 let mayLoad = 0, mayStore = 0, hasSideEffects = 0 in
-class BaseCondSetFlagsReg<bit op, RegisterClass regtype, string asm>
-    : I<(outs), (ins regtype:$Rn, regtype:$Rm, imm0_15:$nzcv, ccode:$cond),
-         asm, "\t$Rn, $Rm, $nzcv, $cond", "", []>,
+class BaseCondComparisonReg<bit op, RegisterClass regtype, string mnemonic,
+                            SDNode OpNode>
+    : I<(outs), (ins regtype:$Rn, regtype:$Rm, imm32_0_15:$nzcv, ccode:$cond),
+         mnemonic, "\t$Rn, $Rm, $nzcv, $cond", "",
+         [(set NZCV, (OpNode regtype:$Rn, regtype:$Rm, (i32 imm:$nzcv),
+                             (i32 imm:$cond), NZCV))]>,
       Sched<[WriteI, ReadI, ReadI]> {
   let Uses = [NZCV];
   let Defs = [NZCV];
 
   bits<5> Rn;
   bits<5> Rm;
   bits<4> nzcv;
   bits<4> cond;
 
   let Inst{30}    = op;
   let Inst{29-21} = 0b111010010;
   let Inst{20-16} = Rm;
   let Inst{15-12} = cond;
   let Inst{11-10} = 0b00;
   let Inst{9-5}   = Rn;
   let Inst{4}     = 0b0;
   let Inst{3-0}   = nzcv;
 }
 
-multiclass CondSetFlagsReg<bit op, string asm> {
-  def Wr : BaseCondSetFlagsReg<op, GPR32, asm> {
+multiclass CondComparison<bit op, string mnemonic, SDNode OpNode> {
+  // immediate operand variants
+  def Wi : BaseCondComparisonImm<op, GPR32, imm32_0_31, mnemonic, OpNode> {
+    let Inst{31} = 0;
+  }
+  def Xi : BaseCondComparisonImm<op, GPR64, imm0_31, mnemonic, OpNode> {
+    let Inst{31} = 1;
+  }
+  // register operand variants
+  def Wr : BaseCondComparisonReg<op, GPR32, mnemonic, OpNode> {
     let Inst{31} = 0;
   }
-  def Xr : BaseCondSetFlagsReg<op, GPR64, asm> {
+  def Xr : BaseCondComparisonReg<op, GPR64, mnemonic, OpNode> {
     let Inst{31} = 1;
   }
 }
 
 //---
 // Conditional select
 //---
 
multiclass FPComparison<bit signalAllNans, string asm,
   } // Defs = [NZCV]
 }
 
 //---
 // Floating point conditional comparisons
 //---
 
 let mayLoad = 0, mayStore = 0, hasSideEffects = 0 in
-class BaseFPCondComparison<bit signalAllNans,
-                              RegisterClass regtype, string asm>
-    : I<(outs), (ins regtype:$Rn, regtype:$Rm, imm0_15:$nzcv, ccode:$cond),
-         asm, "\t$Rn, $Rm, $nzcv, $cond", "", []>,
+class BaseFPCondComparison<bit signalAllNans, RegisterClass regtype,
+                           string mnemonic, list<dag> pat>
+    : I<(outs), (ins regtype:$Rn, regtype:$Rm, imm32_0_15:$nzcv, ccode:$cond),
+         mnemonic, "\t$Rn, $Rm, $nzcv, $cond", "", pat>,
       Sched<[WriteFCmp]> {
+  let Uses = [NZCV];
+  let Defs = [NZCV];
+
   bits<5> Rn;
   bits<5> Rm;
   bits<4> nzcv;
   bits<4> cond;
 
   let Inst{31-23} = 0b000111100;
   let Inst{21}    = 1;
   let Inst{20-16} = Rm;
   let Inst{15-12} = cond;
   let Inst{11-10} = 0b01;
   let Inst{9-5}   = Rn;
   let Inst{4}     = signalAllNans;
   let Inst{3-0}   = nzcv;
 }
 
-multiclass FPCondComparison<bit signalAllNans, string asm> {
-  let Defs = [NZCV], Uses = [NZCV] in {
-  def Srr : BaseFPCondComparison<signalAllNans, FPR32, asm> {
+multiclass FPCondComparison<bit signalAllNans, string mnemonic,
+                            SDPatternOperator OpNode = null_frag> {
+  def Srr : BaseFPCondComparison<signalAllNans, FPR32, mnemonic,
+      [(set NZCV, (OpNode (f32 FPR32:$Rn), (f32 FPR32:$Rm), (i32 imm:$nzcv),
+                          (i32 imm:$cond), NZCV))]> {
     let Inst{22} = 0;
   }
-
-  def Drr : BaseFPCondComparison<signalAllNans, FPR64, asm> {
+  def Drr : BaseFPCondComparison<signalAllNans, FPR64, mnemonic,
+      [(set NZCV, (OpNode (f64 FPR64:$Rn), (f64 FPR64:$Rm), (i32 imm:$nzcv),
+                          (i32 imm:$cond), NZCV))]> {
     let Inst{22} = 1;
   }
-  } // Defs = [NZCV], Uses = [NZCV]
 }
 
 //---
 // Floating point conditional select
 //---
 
 class BaseFPCondSelect<RegisterClass regtype, ValueType vt, string asm>
     : I<(outs regtype:$Rd), (ins regtype:$Rn, regtype:$Rm, ccode:$cond),

lib/Target/AArch64/AArch64InstrInfo.td

def SDT_AArch64tbz : SDTypeProfile<0, 3, [SDTCisInt<0>, SDTCisInt<1>,
                                         SDTCisVT<2, OtherVT>]>;
 
 
 def SDT_AArch64CSel  : SDTypeProfile<1, 4,
                                    [SDTCisSameAs<0, 1>,
                                     SDTCisSameAs<0, 2>,
                                     SDTCisInt<3>,
                                     SDTCisVT<4, i32>]>;
+def SDT_AArch64CCMP : SDTypeProfile<1, 5,
+                                    [SDTCisVT<0, i32>,
+                                     SDTCisInt<1>,
+                                     SDTCisSameAs<1, 2>,
+                                     SDTCisInt<3>,
+                                     SDTCisInt<4>,
+                                     SDTCisVT<5, i32>]>;
+def SDT_AArch64FCCMP : SDTypeProfile<1, 5,
+                                     [SDTCisVT<0, i32>,
+                                      SDTCisFP<1>,
+                                      SDTCisSameAs<1, 2>,
+                                      SDTCisInt<3>,
+                                      SDTCisInt<4>,
+                                      SDTCisVT<5, i32>]>;
 def SDT_AArch64FCmp   : SDTypeProfile<0, 2,
                                    [SDTCisFP<0>,
                                     SDTCisSameAs<0, 1>]>;
 def SDT_AArch64Dup   : SDTypeProfile<1, 1, [SDTCisVec<0>]>;
 def SDT_AArch64DupLane   : SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisInt<2>]>;
 def SDT_AArch64Zip   : SDTypeProfile<1, 2, [SDTCisVec<0>,
                                           SDTCisSameAs<0, 1>,
                                           SDTCisSameAs<0, 2>]>;

 def AArch64add_flag  : SDNode<"AArch64ISD::ADDS",  SDTBinaryArithWithFlagsOut,
                             [SDNPCommutative]>;
 def AArch64sub_flag  : SDNode<"AArch64ISD::SUBS",  SDTBinaryArithWithFlagsOut>;
 def AArch64and_flag  : SDNode<"AArch64ISD::ANDS",  SDTBinaryArithWithFlagsOut,
                             [SDNPCommutative]>;
 def AArch64adc_flag  : SDNode<"AArch64ISD::ADCS",  SDTBinaryArithWithFlagsInOut>;
 def AArch64sbc_flag  : SDNode<"AArch64ISD::SBCS",  SDTBinaryArithWithFlagsInOut>;
 
+def AArch64ccmp      : SDNode<"AArch64ISD::CCMP",  SDT_AArch64CCMP>;
+def AArch64ccmn      : SDNode<"AArch64ISD::CCMN",  SDT_AArch64CCMP>;
+def AArch64fccmp     : SDNode<"AArch64ISD::FCCMP", SDT_AArch64FCCMP>;
+
 def AArch64threadpointer : SDNode<"AArch64ISD::THREAD_POINTER", SDTPtrLeaf>;
 
 def AArch64fcmp      : SDNode<"AArch64ISD::FCMP", SDT_AArch64FCmp>;
 
 def AArch64fmax      : SDNode<"AArch64ISD::FMAX", SDTFPBinOp>;
 def AArch64fmin      : SDNode<"AArch64ISD::FMIN", SDTFPBinOp>;
 
 def AArch64dup       : SDNode<"AArch64ISD::DUP", SDT_AArch64Dup>;
def : InstAlias<"lsr $dst, $src, $shift",
                 (UBFMXri GPR64:$dst, GPR64:$src, imm0_63:$shift, 63)>;
 def : InstAlias<"uxtb $dst, $src", (UBFMWri GPR32:$dst, GPR32:$src, 0, 7)>;
 def : InstAlias<"uxtb $dst, $src", (UBFMXri GPR64:$dst, GPR64:$src, 0, 7)>;
 def : InstAlias<"uxth $dst, $src", (UBFMWri GPR32:$dst, GPR32:$src, 0, 15)>;
 def : InstAlias<"uxth $dst, $src", (UBFMXri GPR64:$dst, GPR64:$src, 0, 15)>;
 def : InstAlias<"uxtw $dst, $src", (UBFMXri GPR64:$dst, GPR64:$src, 0, 31)>;
 
 //===----------------------------------------------------------------------===//
-// Conditionally set flags instructions.
+// Conditionally compare instructions.
 //===----------------------------------------------------------------------===//
-defm CCMN : CondSetFlagsImm<0, "ccmn">;
-defm CCMP : CondSetFlagsImm<1, "ccmp">;
-
-defm CCMN : CondSetFlagsReg<0, "ccmn">;
-defm CCMP : CondSetFlagsReg<1, "ccmp">;
+defm CCMN : CondComparison<0, "ccmn", AArch64ccmn>;
+defm CCMP : CondComparison<1, "ccmp", AArch64ccmp>;
 
 //===----------------------------------------------------------------------===//
 // Conditional select instructions.
 //===----------------------------------------------------------------------===//
 defm CSEL  : CondSelect<0, 0b00, "csel">;
 
 def inc : PatFrag<(ops node:$in), (add node:$in, 1)>;
 defm CSINC : CondSelectOp<0, 0b01, "csinc", inc>;

 defm FCMPE : FPComparison<1, "fcmpe">;
 defm FCMP  : FPComparison<0, "fcmp", AArch64fcmp>;
 
 //===----------------------------------------------------------------------===//
 // Floating point conditional comparison instructions.
 //===----------------------------------------------------------------------===//
 
 defm FCCMPE : FPCondComparison<1, "fccmpe">;
-defm FCCMP  : FPCondComparison<0, "fccmp">;
+defm FCCMP  : FPCondComparison<0, "fccmp", AArch64fccmp>;
 
 //===----------------------------------------------------------------------===//
 // Floating point conditional select instruction.
 //===----------------------------------------------------------------------===//
 
 defm FCSEL : FPCondSelect<"fcsel">;
 
 // CSEL instructions providing f128 types need to be handled by a

test/CodeGen/AArch64/arm64-ccmp.ll

 sw.bb.i.i:
   %ref.tr.i.i = phi %str1* [ %0, %sw.bb.i.i ], [ undef, %entry ]
   %operands.i.i = getelementptr inbounds %str1, %str1* %ref.tr.i.i, i64 0, i32 0, i32 2
   %arrayidx.i.i = bitcast i32* %operands.i.i to %str1**
   %0 = load %str1*, %str1** %arrayidx.i.i, align 8
   %code1.i.i.phi.trans.insert = getelementptr inbounds %str1, %str1* %0, i64 0, i32 0, i32 0, i64 16
   br label %sw.bb.i.i
 }
+
+; CHECK-LABEL: select_and
+define i64 @select_and(i32 %v1, i32 %v2, i64 %a, i64 %b) {
+; CHECK: cmp
+; CHECK: ccmp{{.*}}, #4, lt
+; CHECK: csel{{.*}}, ne
+  %1 = icmp slt i32 %v1, %v2
+  %2 = icmp ne i32 5, %v2
+  %3 = and i1 %1, %2
+  %sel = select i1 %3, i64 %a, i64 %b
+  ret i64 %sel
+}
+
+; CHECK-LABEL: select_or
+define i64 @select_or(i32 %v1, i32 %v2, i64 %a, i64 %b) {
+; CHECK: cmp
+; CHECK: ccmp{{.*}}, #0, ge
+; CHECK: csel{{.*}}, ne
+  %1 = icmp slt i32 %v1, %v2
+  %2 = icmp ne i32 5, %v2
+  %3 = or i1 %1, %2
+  %sel = select i1 %3, i64 %a, i64 %b
+  ret i64 %sel
+}
+
+; CHECK-LABEL: select_complicated
+define i16 @select_complicated(double %v1, double %v2, i16 %a, i16 %b) {
+; CHECK: fcmp
+; CHECK: fccmp{{.*}}, #4, vc
+; CHECK: fccmp{{.*}}, #4, ne
+; CHECK: fccmp{{.*}}, #4, ne
+; CEHCK: csel{{.*}}, eq
+  %1 = fcmp one double %v1, %v2
+  %2 = fcmp oeq double %v2, 13.0
+  %3 = fcmp oeq double %v1, 42.0
+  %or0 = or i1 %2, %3
+  %or1 = or i1 %1, %or0
+  %sel = select i1 %or1, i16 %a, i16 %b
+  ret i16 %sel
+}