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

[MatzeB]

Summary

This is a revised version of http://reviews.llvm.org/D8232.

It turned out the logic in isConjunctionDisjunctionTree() and emitConjunctionDisjunctionTree() is trickier when it comes to OR-expressions so I think another review is warranted for the update. There is a detailed explanation of the types of expressions that can and cannot be matched now. The patch also adds more testcases for the tricky situations. Everything else is the same as in the previous version.

[MatzeB]

ping

[rengolin]

Hi Matthias,

I couldn't find anything terribly wrong with it, but then again, I'm not terribly familiar with CCMP and others. Apart from the goto usage that can easily be replaced by a local flag, I have no reservations. But please wait for Tim, James or someone else with more knowledge on AArch64 to approve.

cheers,
--renato

lib/Target/AArch64/AArch64ISelLowering.cpp

1480

I don't think this is a really good use case for a goto.

[MatzeB]

ping

[jmolloy]

Apart from the goto comment which Renato already mentioned, this looks OK to me.

I've run it through all our tests and this time there were no conformance regressions (woohoo!). LNT diff looks like:

Performance Regressions - Execution Time	Delta
MultiSource/Applications/sqlite3/sqlite3	4.53%
MultiSource/Benchmarks/Olden/bh/bh	3.85%
SingleSource/Benchmarks/Shootout-C++/lists	3.20%
MultiSource/Applications/lua/lua	2.70%

Performance Improvements - Execution Time	Delta
MultiSource/Benchmarks/ASC_Sequoia/AMGmk/AMGmk	-10.44%
MultiSource/Benchmarks/VersaBench/bmm/bmm	-1.15%

That's on a Cortex-A57.

lib/Target/AArch64/AArch64ISelLowering.cpp

1480

Yeah I really don't like the goto here, sorry.

lib/Target/AArch64/AArch64ISelLowering.h

enum NodeType : unsigned {
 
   // 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::FCCMP:             return "AArch64ISD::FCCMP";
   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";
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);
 }
 
+/// \defgroup AArch64CCMP CMP;CCMP matching
+///
+/// These functions deal with the formation of CMP;CCMP;... sequences.
+/// The CCMP/CCMN/FCCMP/FCCMPE instructions allow the conditional execution of
+/// a comparison. They set the NZCV flags to a predefined value if their
+/// predicate is false. This allows to express arbitrary conjunctions, for
+/// example "cmp 0 (and (setCA (cmp A)) (setCB (cmp B))))"
+/// expressed as:
+///   cmp A
+///   ccmp B, inv(CB), CA
+///   check for CB flags
+///
+/// In general we can create code for arbitrary "... (and (and A B) C)"
+/// sequences. We can also implement some "or" expressions, because "(or A B)"
+/// is equivalent to "not (and (not A) (not B))" and we can implement some
+/// negation operations:
+/// We can negate the results of a single comparison by inverting the flags
+/// used when the predicate fails and inverting the flags tested in the next
+/// instruction; We can also negate the results of the whole previous
+/// conditional compare sequence by inverting the flags tested in the next
+/// instruction. However there is no way to negate the result of a partial
+/// sequence.
+///
+/// Therefore on encountering an "or" expression we can negate the subtree on
+/// one side and have to be able to push the negate to the leafs of the subtree
+/// on the other side (see also the comments in code). As complete example:
+/// "or (or (setCA (cmp A)) (setCB (cmp B)))
+///     (and (setCC (cmp C)) (setCD (cmp D)))"
+/// is transformed to
+/// "not (and (not (and (setCC (cmp C)) (setCC (cmp D))))
+///           (and (not (setCA (cmp A)) (not (setCB (cmp B))))))"
+/// and implemented as:
+///   cmp C
+///   ccmp D, inv(CD), CC
+///   ccmp A, CA, inv(CD)
+///   ccmp B, CB, inv(CA)
+///   check for CB flags
+/// A counterexample is "or (and A B) (and C D)" which cannot be implemented
+/// by conditional compare sequences.
+/// @{
+
+/// Create a conditional comparison; Use CCMP, CCMN or FCCMP as apropriate.
+static SDValue emitConditionalComparison(SDValue LHS, SDValue RHS,
+                                         ISD::CondCode CC, SDValue CCOp,
+                                         SDValue Condition, unsigned NZCV,
+                                         SDLoc DL, SelectionDAG &DAG) {
+  unsigned Opcode = 0;
+  if (LHS.getValueType().isFloatingPoint())
+    Opcode = AArch64ISD::FCCMP;
+  else if (RHS.getOpcode() == ISD::SUB) {
+    SDValue SubOp0 = RHS.getOperand(0);
+    if (const ConstantSDNode *SubOp0C = dyn_cast<ConstantSDNode>(SubOp0))
+      if (SubOp0C->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);
+      }
+  }
+  if (Opcode == 0)
+    Opcode = AArch64ISD::CCMP;
+
+  SDValue NZCVOp = DAG.getConstant(NZCV, DL, 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.
+/// CanPushNegate is set to true if we can push a negate operation through
+/// the tree in a was that we are left with AND operations and negate operations
+/// at the leafs only. i.e. "not (or (or x y) z)" can be changed to
+/// "and (and (not x) (not y)) (not z)"; "not (or (and x y) z)" cannot be
+/// brought into such a form.
+static bool isConjunctionDisjunctionTree(const SDValue Val, bool &CanPushNegate,
+                                         unsigned Depth = 0) {
+  if (!Val.hasOneUse())
+    return false;
+  unsigned Opcode = Val->getOpcode();
+  if (Opcode == ISD::SETCC) {
+    CanPushNegate = true;
+    return true;
+  }
+  // Protect against stack overflow.
+  if (Depth > 15)
+    return false;
+  if (Opcode == ISD::AND || Opcode == ISD::OR) {
+    SDValue O0 = Val->getOperand(0);
+    SDValue O1 = Val->getOperand(1);
+    bool CanPushNegateL;
+    if (!isConjunctionDisjunctionTree(O0, CanPushNegateL, Depth+1))
+      return false;
+    bool CanPushNegateR;
+    if (!isConjunctionDisjunctionTree(O1, CanPushNegateR, Depth+1))
+      return false;
+    // We cannot push a negate through an AND operation (it would become an OR),
+    // we can however change a (not (or x y)) to (and (not x) (not y)) if we can
+    // push the negate through the x/y subtrees.
+    CanPushNegate = (Opcode == ISD::OR) && CanPushNegateL && CanPushNegateR;
+    return true;
+  }
+  return false;
+}
+
+/// Emit conjunction or disjunction tree with the CMP/FCMP followed by a chain
+/// of CCMP/CFCMP ops. See @ref AArch64CCMP.
+/// Tries to transform the given i1 producing node @p Val to a series compare
+/// and conditional compare operations. @returns an NZCV flags producing node
+/// and sets @p OutCC to the flags that should be tested or returns SDValue() if
+/// transformation was not possible.
+/// On recursive invocations @p PushNegate may be set to true to have negation
+/// effects pushed to the tree leafs; @p Predicate is an NZCV flag predicate
+/// for the comparisons in the current subtree; @p Depth limits the search
+/// depth to avoid stack overflow.
+static SDValue emitConjunctionDisjunctionTree(SelectionDAG &DAG, SDValue Val,
+    AArch64CC::CondCode &OutCC, bool PushNegate = false,
+    SDValue CCOp = SDValue(), AArch64CC::CondCode Predicate = AArch64CC::AL,
+    unsigned Depth = 0) {
+  // We're at a tree leaf, produce a conditional comparison operation.
+  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 (PushNegate)
+      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 floating point 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(Predicate, DL, 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;
+        Predicate = 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(Predicate, DL, MVT_CC);
+    AArch64CC::CondCode InvOutCC = AArch64CC::getInvertedCondCode(OutCC);
+    unsigned NZCV = AArch64CC::getNZCVToSatisfyCondCode(InvOutCC);
+    return emitConditionalComparison(LHS, RHS, CC, CCOp, ConditionOp, NZCV, DL,
+                                     DAG);
+  } else if (Opcode != ISD::AND && Opcode != ISD::OR)
+    return SDValue();
+
+  assert((Opcode == ISD::OR || !PushNegate)
+         && "Can only push negate through OR operation");
+
+  // Check if both sides can be transformed.
+  SDValue LHS = Val->getOperand(0);
+  SDValue RHS = Val->getOperand(1);
+  bool CanPushNegateL;
+  if (!isConjunctionDisjunctionTree(LHS, CanPushNegateL, Depth+1))
+    return SDValue();
+  bool CanPushNegateR;
+  if (!isConjunctionDisjunctionTree(RHS, CanPushNegateR, Depth+1))
+    return SDValue();
+
+  // Do we need to negate our operands?
+  bool NegateOperands = Opcode == ISD::OR;
+  // We can negate the results of all previous operations by inverting the
+  // predicate flags giving us a free negation for one side. For the other side
+  // we need to be able to push the negation to the leafs of the tree.
+  if (NegateOperands) {
+    if (!CanPushNegateL && !CanPushNegateR)
+      return SDValue();
+    // Order the side where we can push the negate through to LHS.
+    if (!CanPushNegateL && CanPushNegateR) {
+      std::swap(LHS, RHS);
+      CanPushNegateL = true;
+    }
+  }
+
+  // Emit RHS. If we want to negate the tree we only need to push a negate
+  // through if we are already in a PushNegate case, otherwise we can negate
+  // the "flags to test" afterwards.
+  AArch64CC::CondCode RHSCC;
+  SDValue CmpR = emitConjunctionDisjunctionTree(DAG, RHS, RHSCC, PushNegate,
+                                                CCOp, Predicate, Depth+1);
+  if (NegateOperands && !PushNegate)
+    RHSCC = AArch64CC::getInvertedCondCode(RHSCC);
+  // Emit LHS. We must push the negate through if we need to negate it.
+  SDValue CmpL = emitConjunctionDisjunctionTree(DAG, LHS, OutCC, NegateOperands,
+                                                CmpR, RHSCC, Depth+1);
+  // If we transformed an OR to and AND then we have to negate the result
+  // (or absorb a PushNegate resulting in a double negation).
+  if (Opcode == ISD::OR && !PushNegate)
+    OutCC = AArch64CC::getInvertedCondCode(OutCC);
+  return CmpL;
+}
+
+/// @}
+
 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 (!isLegalArithImmed(C)) {
           CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
           C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
           RHS = DAG.getConstant(C, dl, VT);
         }
         break;
       }
     }
   }
+  if ((CC == ISD::SETEQ || CC == ISD::SETNE) && isa<ConstantSDNode>(RHS)) {
+    const ConstantSDNode *RHSC = cast<ConstantSDNode>(RHS);
+
     // The imm operand of ADDS is an unsigned immediate, in the range 0 to 4095.
     // For the i8 operand, the largest immediate is 255, so this can be easily
     // encoded in the compare instruction. For the i16 operand, however, the
     // largest immediate cannot be encoded in the compare.
-  // Therefore, use a sign extending load and cmn to avoid materializing the -1
-  // constant. For example,
+    // Therefore, use a sign extending load and cmn to avoid materializing the
+    // -1 constant. For example,
     // movz w1, #65535
     // ldrh w0, [x0, #0]
     // cmp w0, w1
     // >
     // ldrsh w0, [x0, #0]
     // cmn w0, #1
     // Fundamental, we're relying on the property that (zext LHS) == (zext RHS)
-  // if and only if (sext LHS) == (sext RHS). The checks are in place to ensure
-  // both the LHS and RHS are truely zero extended and to make sure the
+    // if and only if (sext LHS) == (sext RHS). The checks are in place to
+    // ensure both the LHS and RHS are truely zero extended and to make sure the
     // transformation is profitable.
-  if ((CC == ISD::SETEQ || CC == ISD::SETNE) && isa<ConstantSDNode>(RHS)) {
-    if ((cast<ConstantSDNode>(RHS)->getZExtValue() >> 16 == 0) &&
-        isa<LoadSDNode>(LHS)) {
-      if (cast<LoadSDNode>(LHS)->getExtensionType() == ISD::ZEXTLOAD &&
-          cast<LoadSDNode>(LHS)->getMemoryVT() == MVT::i16 &&
-          LHS.getNode()->hasNUsesOfValue(1, 0)) {
-        int16_t ValueofRHS = cast<ConstantSDNode>(RHS)->getZExtValue();
-        if (ValueofRHS < 0 && isLegalArithImmed(-ValueofRHS)) {
-          SDValue SExt =
-              DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, LHS.getValueType(), LHS,
-                          DAG.getValueType(MVT::i16));
-          Cmp = emitComparison(SExt,
-                               DAG.getConstant(ValueofRHS, dl,
+    if ((RHSC->getZExtValue() >> 16 == 0) && isa<LoadSDNode>(LHS) &&
+        cast<LoadSDNode>(LHS)->getExtensionType() == ISD::ZEXTLOAD &&
+        Cmp = emitComparison(SExt, DAG.getConstant(ValueofRHS, dl,
                                                    RHS.getValueType()),
-                               CC, dl, DAG);
-          AArch64CC = changeIntCCToAArch64CC(CC);
-          AArch64cc = DAG.getConstant(AArch64CC, dl, MVT::i32);
-          return Cmp;
-        }
-      }
+        goto CreateCCNode;
+
+    if (RHSC->isNullValue() || RHSC->isOne()) {
+      if ((Cmp = emitConjunctionDisjunctionTree(DAG, LHS, AArch64CC))) {
+        if ((CC == ISD::SETNE) ^ RHSC->isNullValue())
+          AArch64CC = AArch64CC::getInvertedCondCode(AArch64CC);
+        goto CreateCCNode;
       }
     }
+  }
+
   Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
   AArch64CC = changeIntCCToAArch64CC(CC);
-  AArch64cc = DAG.getConstant(AArch64CC, dl, MVT::i32);
+
+CreateCCNode:
+  AArch64cc = DAG.getConstant(AArch64CC, dl, 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.CreateCall(Stxr,
                                  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.
+// Conditional comparison 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 %w0, i32 %w1, i64 %x2, i64 %x3) {
+; CHECK: cmp w1, #5
+; CHECK-NEXT: ccmp w0, w1, #0, ne
+; CHECK-NEXT: csel x0, x2, x3, lt
+; CHECK-NEXT: ret
+  %1 = icmp slt i32 %w0, %w1
+  %2 = icmp ne i32 5, %w1
+  %3 = and i1 %1, %2
+  %sel = select i1 %3, i64 %x2, i64 %x3
+  ret i64 %sel
+}
+
+; CHECK-LABEL: select_or
+define i64 @select_or(i32 %w0, i32 %w1, i64 %x2, i64 %x3) {
+; CHECK: cmp w1, #5
+; CHECK-NEXT: ccmp w0, w1, #8, eq
+; CHECK-NEXT: csel x0, x2, x3, lt
+; CHECK-NEXT: ret
+  %1 = icmp slt i32 %w0, %w1
+  %2 = icmp ne i32 5, %w1
+  %3 = or i1 %1, %2
+  %sel = select i1 %3, i64 %x2, i64 %x3
+  ret i64 %sel
+}
+
+; CHECK-LABEL: select_complicated
+define i16 @select_complicated(double %v1, double %v2, i16 %a, i16 %b) {
+; CHECK: ldr [[REG:d[0-9]+]],
+; CHECK: fcmp d0, d2
+; CHECK-NEXT: fmov d2, #13.00000000
+; CHECK-NEXT: fccmp d1, d2, #4, ne
+; CHECK-NEXT: fccmp d0, d1, #1, ne
+; CHECK-NEXT: fccmp d0, d1, #4, vc
+; CEHCK-NEXT: csel w0, w0, w1, 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
+}
+
+; CHECK-LABEL: gccbug
+define i64 @gccbug(i64 %x0, i64 %x1) {
+; CHECK: cmp x1, #0
+; CHECK-NEXT: ccmp x0, #2, #0, eq
+; CHECK-NEXT: ccmp x0, #4, #4, ne
+; CHECK-NEXT: orr w[[REGNUM:[0-9]+]], wzr, #0x1
+; CHECK-NEXT: cinc x0, x[[REGNUM]], eq
+; CHECK-NEXT: ret
+  %cmp0 = icmp eq i64 %x1, 0
+  %cmp1 = icmp eq i64 %x0, 2
+  %cmp2 = icmp eq i64 %x0, 4
+
+  %or = or i1 %cmp2, %cmp1
+  %and = and i1 %or, %cmp0
+
+  %sel = select i1 %and, i64 2, i64 1
+  ret i64 %sel
+}
+
+; CHECK-LABEL: select_ororand
+define i32 @select_ororand(i32 %w0, i32 %w1, i32 %w2, i32 %w3) {
+; CHECK: cmp w3, #4
+; CHECK-NEXT: ccmp w2, #2, #0, gt
+; CHECK-NEXT: ccmp w1, #13, #2, ge
+; CHECK-NEXT: ccmp w0, #0, #4, ls
+; CHECK-NEXT: csel w0, w3, wzr, eq
+; CHECK-NEXT: ret
+  %c0 = icmp eq i32 %w0, 0
+  %c1 = icmp ugt i32 %w1, 13
+  %c2 = icmp slt i32 %w2, 2
+  %c4 = icmp sgt i32 %w3, 4
+  %or = or i1 %c0, %c1
+  %and = and i1 %c2, %c4
+  %or1 = or i1 %or, %and
+  %sel = select i1 %or1, i32 %w3, i32 0
+  ret i32 %sel
+}
+
+; CHECK-LABEL: select_noccmp
+define i64 @select_noccmp(i64 %v1, i64 %v2, i64 %v3, i64 %r) {
+; CHECK-NOT: CCMP
+  %c0 = icmp slt i64 %v1, 0
+  %c1 = icmp sgt i64 %v1, 13
+  %c2 = icmp slt i64 %v3, 2
+  %c4 = icmp sgt i64 %v3, 4
+  %and0 = and i1 %c0, %c1
+  %and1 = and i1 %c2, %c4
+  %or = or i1 %and0, %and1
+  %sel = select i1 %or, i64 0, i64 %r
+  ret i64 %sel
+}