diff options
Diffstat (limited to 'lib/Target/Hexagon')
44 files changed, 3547 insertions, 6135 deletions
diff --git a/lib/Target/Hexagon/CMakeLists.txt b/lib/Target/Hexagon/CMakeLists.txt index b5b887e..2b79791 100644 --- a/lib/Target/Hexagon/CMakeLists.txt +++ b/lib/Target/Hexagon/CMakeLists.txt @@ -9,8 +9,6 @@ tablegen(LLVM HexagonGenSubtargetInfo.inc -gen-subtarget) tablegen(LLVM HexagonGenDFAPacketizer.inc -gen-dfa-packetizer) add_public_tablegen_target(HexagonCommonTableGen) -set(LLVM_COMMON_DEPENDS intrinsics_gen) - add_llvm_target(HexagonCodeGen HexagonAsmPrinter.cpp HexagonCallingConvLower.cpp @@ -28,14 +26,18 @@ add_llvm_target(HexagonCodeGen HexagonRegisterInfo.cpp HexagonRemoveSZExtArgs.cpp HexagonSelectionDAGInfo.cpp + HexagonSplitConst32AndConst64.cpp HexagonSplitTFRCondSets.cpp HexagonSubtarget.cpp HexagonTargetMachine.cpp HexagonTargetObjectFile.cpp HexagonVLIWPacketizer.cpp HexagonNewValueJump.cpp + HexagonCopyToCombine.cpp ) +add_dependencies(LLVMHexagonCodeGen HexagonCommonTableGen intrinsics_gen) + add_subdirectory(TargetInfo) add_subdirectory(InstPrinter) add_subdirectory(MCTargetDesc) diff --git a/lib/Target/Hexagon/Hexagon.h b/lib/Target/Hexagon/Hexagon.h index dfbefc8..5467ee3 100644 --- a/lib/Target/Hexagon/Hexagon.h +++ b/lib/Target/Hexagon/Hexagon.h @@ -31,24 +31,27 @@ namespace llvm { FunctionPass *createHexagonISelDag(HexagonTargetMachine &TM, CodeGenOpt::Level OptLevel); - FunctionPass *createHexagonDelaySlotFillerPass(TargetMachine &TM); - FunctionPass *createHexagonFPMoverPass(TargetMachine &TM); - FunctionPass *createHexagonRemoveExtendOps(HexagonTargetMachine &TM); - FunctionPass *createHexagonCFGOptimizer(HexagonTargetMachine &TM); - - FunctionPass *createHexagonSplitTFRCondSets(HexagonTargetMachine &TM); - FunctionPass *createHexagonExpandPredSpillCode(HexagonTargetMachine &TM); + FunctionPass *createHexagonDelaySlotFillerPass(const TargetMachine &TM); + FunctionPass *createHexagonFPMoverPass(const TargetMachine &TM); + FunctionPass *createHexagonRemoveExtendArgs(const HexagonTargetMachine &TM); + FunctionPass *createHexagonCFGOptimizer(const HexagonTargetMachine &TM); + FunctionPass *createHexagonSplitTFRCondSets(const HexagonTargetMachine &TM); + FunctionPass *createHexagonSplitConst32AndConst64( + const HexagonTargetMachine &TM); + FunctionPass *createHexagonExpandPredSpillCode( + const HexagonTargetMachine &TM); FunctionPass *createHexagonHardwareLoops(); FunctionPass *createHexagonPeephole(); FunctionPass *createHexagonFixupHwLoops(); + FunctionPass *createHexagonNewValueJump(); + FunctionPass *createHexagonCopyToCombine(); FunctionPass *createHexagonPacketizer(); FunctionPass *createHexagonNewValueJump(); - /* TODO: object output. MCCodeEmitter *createHexagonMCCodeEmitter(const Target &, - TargetMachine &TM, + const TargetMachine &TM, MCContext &Ctx); */ /* TODO: assembler input. diff --git a/lib/Target/Hexagon/Hexagon.td b/lib/Target/Hexagon/Hexagon.td index 8a5ee40..568798c 100644 --- a/lib/Target/Hexagon/Hexagon.td +++ b/lib/Target/Hexagon/Hexagon.td @@ -84,27 +84,75 @@ def getPredOpcode : InstrMapping { } //===----------------------------------------------------------------------===// +// Generate mapping table to relate predicate-true instructions with their +// predicate-false forms +// +def getFalsePredOpcode : InstrMapping { + let FilterClass = "PredRel"; + let RowFields = ["BaseOpcode", "PNewValue", "isNVStore", "isBrTaken"]; + let ColFields = ["PredSense"]; + let KeyCol = ["true"]; + let ValueCols = [["false"]]; +} + +//===----------------------------------------------------------------------===// +// Generate mapping table to relate predicate-false instructions with their +// predicate-true forms +// +def getTruePredOpcode : InstrMapping { + let FilterClass = "PredRel"; + let RowFields = ["BaseOpcode", "PNewValue", "isNVStore", "isBrTaken"]; + let ColFields = ["PredSense"]; + let KeyCol = ["false"]; + let ValueCols = [["true"]]; +} + +//===----------------------------------------------------------------------===// // Generate mapping table to relate predicated instructions with their .new // format. // def getPredNewOpcode : InstrMapping { let FilterClass = "PredNewRel"; - let RowFields = ["BaseOpcode", "PredSense", "isNVStore"]; + let RowFields = ["BaseOpcode", "PredSense", "isNVStore", "isBrTaken"]; let ColFields = ["PNewValue"]; let KeyCol = [""]; let ValueCols = [["new"]]; } //===----------------------------------------------------------------------===// +// Generate mapping table to relate .new predicated instructions with their old +// format. +// +def getPredOldOpcode : InstrMapping { + let FilterClass = "PredNewRel"; + let RowFields = ["BaseOpcode", "PredSense", "isNVStore"]; + let ColFields = ["PNewValue"]; + let KeyCol = ["new"]; + let ValueCols = [[""]]; +} + +//===----------------------------------------------------------------------===// // Generate mapping table to relate store instructions with their new-value // format. // def getNewValueOpcode : InstrMapping { let FilterClass = "NewValueRel"; let RowFields = ["BaseOpcode", "PredSense", "PNewValue"]; - let ColFields = ["isNVStore"]; - let KeyCol = ["0"]; - let ValueCols = [["1"]]; + let ColFields = ["NValueST"]; + let KeyCol = ["false"]; + let ValueCols = [["true"]]; +} + +//===----------------------------------------------------------------------===// +// Generate mapping table to relate new-value store instructions with their old +// format. +// +def getNonNVStore : InstrMapping { + let FilterClass = "NewValueRel"; + let RowFields = ["BaseOpcode", "PredSense", "PNewValue"]; + let ColFields = ["NValueST"]; + let KeyCol = ["true"]; + let ValueCols = [["false"]]; } def getBasedWithImmOffset : InstrMapping { diff --git a/lib/Target/Hexagon/HexagonCFGOptimizer.cpp b/lib/Target/Hexagon/HexagonCFGOptimizer.cpp index d4078ad..8597f11 100644 --- a/lib/Target/Hexagon/HexagonCFGOptimizer.cpp +++ b/lib/Target/Hexagon/HexagonCFGOptimizer.cpp @@ -26,21 +26,27 @@ using namespace llvm; +namespace llvm { + void initializeHexagonCFGOptimizerPass(PassRegistry&); +} + + namespace { class HexagonCFGOptimizer : public MachineFunctionPass { private: - HexagonTargetMachine& QTM; + const HexagonTargetMachine& QTM; const HexagonSubtarget &QST; void InvertAndChangeJumpTarget(MachineInstr*, MachineBasicBlock*); public: static char ID; - HexagonCFGOptimizer(HexagonTargetMachine& TM) : MachineFunctionPass(ID), - QTM(TM), - QST(*TM.getSubtargetImpl()) {} + HexagonCFGOptimizer(const HexagonTargetMachine& TM) + : MachineFunctionPass(ID), QTM(TM), QST(*TM.getSubtargetImpl()) { + initializeHexagonCFGOptimizerPass(*PassRegistry::getPassRegistry()); + } const char *getPassName() const { return "Hexagon CFG Optimizer"; @@ -52,8 +58,8 @@ private: char HexagonCFGOptimizer::ID = 0; static bool IsConditionalBranch(int Opc) { - return (Opc == Hexagon::JMP_c) || (Opc == Hexagon::JMP_cNot) - || (Opc == Hexagon::JMP_cdnPt) || (Opc == Hexagon::JMP_cdnNotPt); + return (Opc == Hexagon::JMP_t) || (Opc == Hexagon::JMP_f) + || (Opc == Hexagon::JMP_tnew_t) || (Opc == Hexagon::JMP_fnew_t); } @@ -68,20 +74,20 @@ HexagonCFGOptimizer::InvertAndChangeJumpTarget(MachineInstr* MI, const HexagonInstrInfo *QII = QTM.getInstrInfo(); int NewOpcode = 0; switch(MI->getOpcode()) { - case Hexagon::JMP_c: - NewOpcode = Hexagon::JMP_cNot; + case Hexagon::JMP_t: + NewOpcode = Hexagon::JMP_f; break; - case Hexagon::JMP_cNot: - NewOpcode = Hexagon::JMP_c; + case Hexagon::JMP_f: + NewOpcode = Hexagon::JMP_t; break; - case Hexagon::JMP_cdnPt: - NewOpcode = Hexagon::JMP_cdnNotPt; + case Hexagon::JMP_tnew_t: + NewOpcode = Hexagon::JMP_fnew_t; break; - case Hexagon::JMP_cdnNotPt: - NewOpcode = Hexagon::JMP_cdnPt; + case Hexagon::JMP_fnew_t: + NewOpcode = Hexagon::JMP_tnew_t; break; default: @@ -156,8 +162,8 @@ bool HexagonCFGOptimizer::runOnMachineFunction(MachineFunction &Fn) { // The target of the unconditional branch must be JumpAroundTarget. // TODO: If not, we should not invert the unconditional branch. MachineBasicBlock* CondBranchTarget = NULL; - if ((MI->getOpcode() == Hexagon::JMP_c) || - (MI->getOpcode() == Hexagon::JMP_cNot)) { + if ((MI->getOpcode() == Hexagon::JMP_t) || + (MI->getOpcode() == Hexagon::JMP_f)) { CondBranchTarget = MI->getOperand(1).getMBB(); } @@ -231,6 +237,16 @@ bool HexagonCFGOptimizer::runOnMachineFunction(MachineFunction &Fn) { // Public Constructor Functions //===----------------------------------------------------------------------===// -FunctionPass *llvm::createHexagonCFGOptimizer(HexagonTargetMachine &TM) { +static void initializePassOnce(PassRegistry &Registry) { + PassInfo *PI = new PassInfo("Hexagon CFG Optimizer", "hexagon-cfg", + &HexagonCFGOptimizer::ID, 0, false, false); + Registry.registerPass(*PI, true); +} + +void llvm::initializeHexagonCFGOptimizerPass(PassRegistry &Registry) { + CALL_ONCE_INITIALIZATION(initializePassOnce) +} + +FunctionPass *llvm::createHexagonCFGOptimizer(const HexagonTargetMachine &TM) { return new HexagonCFGOptimizer(TM); } diff --git a/lib/Target/Hexagon/HexagonCallingConvLower.cpp b/lib/Target/Hexagon/HexagonCallingConvLower.cpp index 2c93d04..f5f958c 100644 --- a/lib/Target/Hexagon/HexagonCallingConvLower.cpp +++ b/lib/Target/Hexagon/HexagonCallingConvLower.cpp @@ -25,14 +25,13 @@ using namespace llvm; Hexagon_CCState::Hexagon_CCState(CallingConv::ID CC, bool isVarArg, const TargetMachine &tm, - SmallVector<CCValAssign, 16> &locs, + SmallVectorImpl<CCValAssign> &locs, LLVMContext &c) - : CallingConv(CC), IsVarArg(isVarArg), TM(tm), - TRI(*TM.getRegisterInfo()), Locs(locs), Context(c) { + : CallingConv(CC), IsVarArg(isVarArg), TM(tm), Locs(locs), Context(c) { // No stack is used. StackOffset = 0; - UsedRegs.resize((TRI.getNumRegs()+31)/32); + UsedRegs.resize((TM.getRegisterInfo()->getNumRegs()+31)/32); } // HandleByVal - Allocate a stack slot large enough to pass an argument by @@ -56,6 +55,7 @@ void Hexagon_CCState::HandleByVal(unsigned ValNo, EVT ValVT, /// MarkAllocated - Mark a register and all of its aliases as allocated. void Hexagon_CCState::MarkAllocated(unsigned Reg) { + const TargetRegisterInfo &TRI = *TM.getRegisterInfo(); for (MCRegAliasIterator AI(Reg, &TRI, true); AI.isValid(); ++AI) UsedRegs[*AI/32] |= 1 << (*AI&31); } diff --git a/lib/Target/Hexagon/HexagonCallingConvLower.h b/lib/Target/Hexagon/HexagonCallingConvLower.h index 489b3a3..33c8306 100644 --- a/lib/Target/Hexagon/HexagonCallingConvLower.h +++ b/lib/Target/Hexagon/HexagonCallingConvLower.h @@ -48,15 +48,14 @@ class Hexagon_CCState { CallingConv::ID CallingConv; bool IsVarArg; const TargetMachine &TM; - const TargetRegisterInfo &TRI; - SmallVector<CCValAssign, 16> &Locs; + SmallVectorImpl<CCValAssign> &Locs; LLVMContext &Context; unsigned StackOffset; SmallVector<uint32_t, 16> UsedRegs; public: Hexagon_CCState(CallingConv::ID CC, bool isVarArg, const TargetMachine &TM, - SmallVector<CCValAssign, 16> &locs, LLVMContext &c); + SmallVectorImpl<CCValAssign> &locs, LLVMContext &c); void addLoc(const CCValAssign &V) { Locs.push_back(V); diff --git a/lib/Target/Hexagon/HexagonCopyToCombine.cpp b/lib/Target/Hexagon/HexagonCopyToCombine.cpp new file mode 100644 index 0000000..dc440cb --- /dev/null +++ b/lib/Target/Hexagon/HexagonCopyToCombine.cpp @@ -0,0 +1,677 @@ +//===------- HexagonCopyToCombine.cpp - Hexagon Copy-To-Combine Pass ------===// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +// This pass replaces transfer instructions by combine instructions. +// We walk along a basic block and look for two combinable instructions and try +// to move them together. If we can move them next to each other we do so and +// replace them with a combine instruction. +//===----------------------------------------------------------------------===// +#define DEBUG_TYPE "hexagon-copy-combine" + +#include "llvm/PassSupport.h" +#include "llvm/ADT/DenseSet.h" +#include "llvm/ADT/DenseMap.h" +#include "llvm/CodeGen/Passes.h" +#include "llvm/CodeGen/MachineBasicBlock.h" +#include "llvm/CodeGen/MachineFunction.h" +#include "llvm/CodeGen/MachineFunctionPass.h" +#include "llvm/CodeGen/MachineInstr.h" +#include "llvm/CodeGen/MachineInstrBuilder.h" +#include "llvm/Target/TargetRegisterInfo.h" +#include "llvm/Support/CodeGen.h" +#include "llvm/Support/CommandLine.h" +#include "llvm/Support/Debug.h" +#include "llvm/Support/raw_ostream.h" + +#include "Hexagon.h" +#include "HexagonInstrInfo.h" +#include "HexagonRegisterInfo.h" +#include "HexagonSubtarget.h" +#include "HexagonTargetMachine.h" +#include "HexagonMachineFunctionInfo.h" + +using namespace llvm; + +static +cl::opt<bool> IsCombinesDisabled("disable-merge-into-combines", + cl::Hidden, cl::ZeroOrMore, + cl::init(false), + cl::desc("Disable merging into combines")); +static +cl::opt<unsigned> +MaxNumOfInstsBetweenNewValueStoreAndTFR("max-num-inst-between-tfr-and-nv-store", + cl::Hidden, cl::init(4), + cl::desc("Maximum distance between a tfr feeding a store we " + "consider the store still to be newifiable")); + +namespace llvm { + void initializeHexagonCopyToCombinePass(PassRegistry&); +} + + +namespace { + +class HexagonCopyToCombine : public MachineFunctionPass { + const HexagonInstrInfo *TII; + const TargetRegisterInfo *TRI; + bool ShouldCombineAggressively; + + DenseSet<MachineInstr *> PotentiallyNewifiableTFR; +public: + static char ID; + + HexagonCopyToCombine() : MachineFunctionPass(ID) { + initializeHexagonCopyToCombinePass(*PassRegistry::getPassRegistry()); + } + + virtual void getAnalysisUsage(AnalysisUsage &AU) const { + MachineFunctionPass::getAnalysisUsage(AU); + } + + const char *getPassName() const { + return "Hexagon Copy-To-Combine Pass"; + } + + virtual bool runOnMachineFunction(MachineFunction &Fn); + +private: + MachineInstr *findPairable(MachineInstr *I1, bool &DoInsertAtI1); + + void findPotentialNewifiableTFRs(MachineBasicBlock &); + + void combine(MachineInstr *I1, MachineInstr *I2, + MachineBasicBlock::iterator &MI, bool DoInsertAtI1); + + bool isSafeToMoveTogether(MachineInstr *I1, MachineInstr *I2, + unsigned I1DestReg, unsigned I2DestReg, + bool &DoInsertAtI1); + + void emitCombineRR(MachineBasicBlock::iterator &Before, unsigned DestReg, + MachineOperand &HiOperand, MachineOperand &LoOperand); + + void emitCombineRI(MachineBasicBlock::iterator &Before, unsigned DestReg, + MachineOperand &HiOperand, MachineOperand &LoOperand); + + void emitCombineIR(MachineBasicBlock::iterator &Before, unsigned DestReg, + MachineOperand &HiOperand, MachineOperand &LoOperand); + + void emitCombineII(MachineBasicBlock::iterator &Before, unsigned DestReg, + MachineOperand &HiOperand, MachineOperand &LoOperand); +}; + +} // End anonymous namespace. + +char HexagonCopyToCombine::ID = 0; + +INITIALIZE_PASS(HexagonCopyToCombine, "hexagon-copy-combine", + "Hexagon Copy-To-Combine Pass", false, false) + +static bool isCombinableInstType(MachineInstr *MI, + const HexagonInstrInfo *TII, + bool ShouldCombineAggressively) { + switch(MI->getOpcode()) { + case Hexagon::TFR: { + // A COPY instruction can be combined if its arguments are IntRegs (32bit). + assert(MI->getOperand(0).isReg() && MI->getOperand(1).isReg()); + + unsigned DestReg = MI->getOperand(0).getReg(); + unsigned SrcReg = MI->getOperand(1).getReg(); + return Hexagon::IntRegsRegClass.contains(DestReg) && + Hexagon::IntRegsRegClass.contains(SrcReg); + } + + case Hexagon::TFRI: { + // A transfer-immediate can be combined if its argument is a signed 8bit + // value. + assert(MI->getOperand(0).isReg() && MI->getOperand(1).isImm()); + unsigned DestReg = MI->getOperand(0).getReg(); + + // Only combine constant extended TFRI if we are in aggressive mode. + return Hexagon::IntRegsRegClass.contains(DestReg) && + (ShouldCombineAggressively || isInt<8>(MI->getOperand(1).getImm())); + } + + case Hexagon::TFRI_V4: { + if (!ShouldCombineAggressively) + return false; + assert(MI->getOperand(0).isReg() && MI->getOperand(1).isGlobal()); + + // Ensure that TargetFlags are MO_NO_FLAG for a global. This is a + // workaround for an ABI bug that prevents GOT relocations on combine + // instructions + if (MI->getOperand(1).getTargetFlags() != HexagonII::MO_NO_FLAG) + return false; + + unsigned DestReg = MI->getOperand(0).getReg(); + return Hexagon::IntRegsRegClass.contains(DestReg); + } + + default: + break; + } + + return false; +} + +static bool isGreaterThan8BitTFRI(MachineInstr *I) { + return I->getOpcode() == Hexagon::TFRI && + !isInt<8>(I->getOperand(1).getImm()); +} +static bool isGreaterThan6BitTFRI(MachineInstr *I) { + return I->getOpcode() == Hexagon::TFRI && + !isUInt<6>(I->getOperand(1).getImm()); +} + +/// areCombinableOperations - Returns true if the two instruction can be merge +/// into a combine (ignoring register constraints). +static bool areCombinableOperations(const TargetRegisterInfo *TRI, + MachineInstr *HighRegInst, + MachineInstr *LowRegInst) { + assert((HighRegInst->getOpcode() == Hexagon::TFR || + HighRegInst->getOpcode() == Hexagon::TFRI || + HighRegInst->getOpcode() == Hexagon::TFRI_V4) && + (LowRegInst->getOpcode() == Hexagon::TFR || + LowRegInst->getOpcode() == Hexagon::TFRI || + LowRegInst->getOpcode() == Hexagon::TFRI_V4) && + "Assume individual instructions are of a combinable type"); + + const HexagonRegisterInfo *QRI = + static_cast<const HexagonRegisterInfo *>(TRI); + + // V4 added some combine variations (mixed immediate and register source + // operands), if we are on < V4 we can only combine 2 register-to-register + // moves and 2 immediate-to-register moves. We also don't have + // constant-extenders. + if (!QRI->Subtarget.hasV4TOps()) + return HighRegInst->getOpcode() == LowRegInst->getOpcode() && + !isGreaterThan8BitTFRI(HighRegInst) && + !isGreaterThan6BitTFRI(LowRegInst); + + // There is no combine of two constant extended values. + if ((HighRegInst->getOpcode() == Hexagon::TFRI_V4 || + isGreaterThan8BitTFRI(HighRegInst)) && + (LowRegInst->getOpcode() == Hexagon::TFRI_V4 || + isGreaterThan6BitTFRI(LowRegInst))) + return false; + + return true; +} + +static bool isEvenReg(unsigned Reg) { + assert(TargetRegisterInfo::isPhysicalRegister(Reg) && + Hexagon::IntRegsRegClass.contains(Reg)); + return (Reg - Hexagon::R0) % 2 == 0; +} + +static void removeKillInfo(MachineInstr *MI, unsigned RegNotKilled) { + for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) { + MachineOperand &Op = MI->getOperand(I); + if (!Op.isReg() || Op.getReg() != RegNotKilled || !Op.isKill()) + continue; + Op.setIsKill(false); + } +} + +/// isUnsafeToMoveAcross - Returns true if it is unsafe to move a copy +/// instruction from \p UseReg to \p DestReg over the instruction \p I. +static bool isUnsafeToMoveAcross(MachineInstr *I, unsigned UseReg, + unsigned DestReg, + const TargetRegisterInfo *TRI) { + return (UseReg && (I->modifiesRegister(UseReg, TRI))) || + I->modifiesRegister(DestReg, TRI) || + I->readsRegister(DestReg, TRI) || + I->hasUnmodeledSideEffects() || + I->isInlineAsm() || I->isDebugValue(); +} + +/// isSafeToMoveTogether - Returns true if it is safe to move I1 next to I2 such +/// that the two instructions can be paired in a combine. +bool HexagonCopyToCombine::isSafeToMoveTogether(MachineInstr *I1, + MachineInstr *I2, + unsigned I1DestReg, + unsigned I2DestReg, + bool &DoInsertAtI1) { + + bool IsImmUseReg = I2->getOperand(1).isImm() || I2->getOperand(1).isGlobal(); + unsigned I2UseReg = IsImmUseReg ? 0 : I2->getOperand(1).getReg(); + + // It is not safe to move I1 and I2 into one combine if I2 has a true + // dependence on I1. + if (I2UseReg && I1->modifiesRegister(I2UseReg, TRI)) + return false; + + bool isSafe = true; + + // First try to move I2 towards I1. + { + // A reverse_iterator instantiated like below starts before I2, and I1 + // respectively. + // Look at instructions I in between I2 and (excluding) I1. + MachineBasicBlock::reverse_iterator I(I2), + End = --(MachineBasicBlock::reverse_iterator(I1)); + // At 03 we got better results (dhrystone!) by being more conservative. + if (!ShouldCombineAggressively) + End = MachineBasicBlock::reverse_iterator(I1); + // If I2 kills its operand and we move I2 over an instruction that also + // uses I2's use reg we need to modify that (first) instruction to now kill + // this reg. + unsigned KilledOperand = 0; + if (I2->killsRegister(I2UseReg)) + KilledOperand = I2UseReg; + MachineInstr *KillingInstr = 0; + + for (; I != End; ++I) { + // If the intervening instruction I: + // * modifies I2's use reg + // * modifies I2's def reg + // * reads I2's def reg + // * or has unmodelled side effects + // we can't move I2 across it. + if (isUnsafeToMoveAcross(&*I, I2UseReg, I2DestReg, TRI)) { + isSafe = false; + break; + } + + // Update first use of the killed operand. + if (!KillingInstr && KilledOperand && + I->readsRegister(KilledOperand, TRI)) + KillingInstr = &*I; + } + if (isSafe) { + // Update the intermediate instruction to with the kill flag. + if (KillingInstr) { + bool Added = KillingInstr->addRegisterKilled(KilledOperand, TRI, true); + (void)Added; // supress compiler warning + assert(Added && "Must successfully update kill flag"); + removeKillInfo(I2, KilledOperand); + } + DoInsertAtI1 = true; + return true; + } + } + + // Try to move I1 towards I2. + { + // Look at instructions I in between I1 and (excluding) I2. + MachineBasicBlock::iterator I(I1), End(I2); + // At O3 we got better results (dhrystone) by being more conservative here. + if (!ShouldCombineAggressively) + End = llvm::next(MachineBasicBlock::iterator(I2)); + IsImmUseReg = I1->getOperand(1).isImm() || I1->getOperand(1).isGlobal(); + unsigned I1UseReg = IsImmUseReg ? 0 : I1->getOperand(1).getReg(); + // Track killed operands. If we move across an instruction that kills our + // operand, we need to update the kill information on the moved I1. It kills + // the operand now. + MachineInstr *KillingInstr = 0; + unsigned KilledOperand = 0; + + while(++I != End) { + // If the intervening instruction I: + // * modifies I1's use reg + // * modifies I1's def reg + // * reads I1's def reg + // * or has unmodelled side effects + // We introduce this special case because llvm has no api to remove a + // kill flag for a register (a removeRegisterKilled() analogous to + // addRegisterKilled) that handles aliased register correctly. + // * or has a killed aliased register use of I1's use reg + // %D4<def> = TFRI64 16 + // %R6<def> = TFR %R9 + // %R8<def> = KILL %R8, %D4<imp-use,kill> + // If we want to move R6 = across the KILL instruction we would have + // to remove the %D4<imp-use,kill> operand. For now, we are + // conservative and disallow the move. + // we can't move I1 across it. + if (isUnsafeToMoveAcross(I, I1UseReg, I1DestReg, TRI) || + // Check for an aliased register kill. Bail out if we see one. + (!I->killsRegister(I1UseReg) && I->killsRegister(I1UseReg, TRI))) + return false; + + // Check for an exact kill (registers match). + if (I1UseReg && I->killsRegister(I1UseReg)) { + assert(KillingInstr == 0 && "Should only see one killing instruction"); + KilledOperand = I1UseReg; + KillingInstr = &*I; + } + } + if (KillingInstr) { + removeKillInfo(KillingInstr, KilledOperand); + // Update I1 to set the kill flag. This flag will later be picked up by + // the new COMBINE instruction. + bool Added = I1->addRegisterKilled(KilledOperand, TRI); + (void)Added; // supress compiler warning + assert(Added && "Must successfully update kill flag"); + } + DoInsertAtI1 = false; + } + + return true; +} + +/// findPotentialNewifiableTFRs - Finds tranfers that feed stores that could be +/// newified. (A use of a 64 bit register define can not be newified) +void +HexagonCopyToCombine::findPotentialNewifiableTFRs(MachineBasicBlock &BB) { + DenseMap<unsigned, MachineInstr *> LastDef; + for (MachineBasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) { + MachineInstr *MI = I; + // Mark TFRs that feed a potential new value store as such. + if(TII->mayBeNewStore(MI)) { + // Look for uses of TFR instructions. + for (unsigned OpdIdx = 0, OpdE = MI->getNumOperands(); OpdIdx != OpdE; + ++OpdIdx) { + MachineOperand &Op = MI->getOperand(OpdIdx); + + // Skip over anything except register uses. + if (!Op.isReg() || !Op.isUse() || !Op.getReg()) + continue; + + // Look for the defining instruction. + unsigned Reg = Op.getReg(); + MachineInstr *DefInst = LastDef[Reg]; + if (!DefInst) + continue; + if (!isCombinableInstType(DefInst, TII, ShouldCombineAggressively)) + continue; + + // Only close newifiable stores should influence the decision. + MachineBasicBlock::iterator It(DefInst); + unsigned NumInstsToDef = 0; + while (&*It++ != MI) + ++NumInstsToDef; + + if (NumInstsToDef > MaxNumOfInstsBetweenNewValueStoreAndTFR) + continue; + + PotentiallyNewifiableTFR.insert(DefInst); + } + // Skip to next instruction. + continue; + } + + // Put instructions that last defined integer or double registers into the + // map. + for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) { + MachineOperand &Op = MI->getOperand(I); + if (!Op.isReg() || !Op.isDef() || !Op.getReg()) + continue; + unsigned Reg = Op.getReg(); + if (Hexagon::DoubleRegsRegClass.contains(Reg)) { + for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs) { + LastDef[*SubRegs] = MI; + } + } else if (Hexagon::IntRegsRegClass.contains(Reg)) + LastDef[Reg] = MI; + } + } +} + +bool HexagonCopyToCombine::runOnMachineFunction(MachineFunction &MF) { + + if (IsCombinesDisabled) return false; + + bool HasChanged = false; + + // Get target info. + TRI = MF.getTarget().getRegisterInfo(); + TII = static_cast<const HexagonInstrInfo *>(MF.getTarget().getInstrInfo()); + + // Combine aggressively (for code size) + ShouldCombineAggressively = + MF.getTarget().getOptLevel() <= CodeGenOpt::Default; + + // Traverse basic blocks. + for (MachineFunction::iterator BI = MF.begin(), BE = MF.end(); BI != BE; + ++BI) { + PotentiallyNewifiableTFR.clear(); + findPotentialNewifiableTFRs(*BI); + + // Traverse instructions in basic block. + for(MachineBasicBlock::iterator MI = BI->begin(), End = BI->end(); + MI != End;) { + MachineInstr *I1 = MI++; + // Don't combine a TFR whose user could be newified (instructions that + // define double registers can not be newified - Programmer's Ref Manual + // 5.4.2 New-value stores). + if (ShouldCombineAggressively && PotentiallyNewifiableTFR.count(I1)) + continue; + + // Ignore instructions that are not combinable. + if (!isCombinableInstType(I1, TII, ShouldCombineAggressively)) + continue; + + // Find a second instruction that can be merged into a combine + // instruction. + bool DoInsertAtI1 = false; + MachineInstr *I2 = findPairable(I1, DoInsertAtI1); + if (I2) { + HasChanged = true; + combine(I1, I2, MI, DoInsertAtI1); + } + } + } + + return HasChanged; +} + +/// findPairable - Returns an instruction that can be merged with \p I1 into a +/// COMBINE instruction or 0 if no such instruction can be found. Returns true +/// in \p DoInsertAtI1 if the combine must be inserted at instruction \p I1 +/// false if the combine must be inserted at the returned instruction. +MachineInstr *HexagonCopyToCombine::findPairable(MachineInstr *I1, + bool &DoInsertAtI1) { + MachineBasicBlock::iterator I2 = llvm::next(MachineBasicBlock::iterator(I1)); + unsigned I1DestReg = I1->getOperand(0).getReg(); + + for (MachineBasicBlock::iterator End = I1->getParent()->end(); I2 != End; + ++I2) { + // Bail out early if we see a second definition of I1DestReg. + if (I2->modifiesRegister(I1DestReg, TRI)) + break; + + // Ignore non-combinable instructions. + if (!isCombinableInstType(I2, TII, ShouldCombineAggressively)) + continue; + + // Don't combine a TFR whose user could be newified. + if (ShouldCombineAggressively && PotentiallyNewifiableTFR.count(I2)) + continue; + + unsigned I2DestReg = I2->getOperand(0).getReg(); + + // Check that registers are adjacent and that the first destination register + // is even. + bool IsI1LowReg = (I2DestReg - I1DestReg) == 1; + bool IsI2LowReg = (I1DestReg - I2DestReg) == 1; + unsigned FirstRegIndex = IsI1LowReg ? I1DestReg : I2DestReg; + if ((!IsI1LowReg && !IsI2LowReg) || !isEvenReg(FirstRegIndex)) + continue; + + // Check that the two instructions are combinable. V4 allows more + // instructions to be merged into a combine. + // The order matters because in a TFRI we might can encode a int8 as the + // hi reg operand but only a uint6 as the low reg operand. + if ((IsI2LowReg && !areCombinableOperations(TRI, I1, I2)) || + (IsI1LowReg && !areCombinableOperations(TRI, I2, I1))) + break; + + if (isSafeToMoveTogether(I1, I2, I1DestReg, I2DestReg, + DoInsertAtI1)) + return I2; + + // Not safe. Stop searching. + break; + } + return 0; +} + +void HexagonCopyToCombine::combine(MachineInstr *I1, MachineInstr *I2, + MachineBasicBlock::iterator &MI, + bool DoInsertAtI1) { + // We are going to delete I2. If MI points to I2 advance it to the next + // instruction. + if ((MachineInstr *)MI == I2) ++MI; + + // Figure out whether I1 or I2 goes into the lowreg part. + unsigned I1DestReg = I1->getOperand(0).getReg(); + unsigned I2DestReg = I2->getOperand(0).getReg(); + bool IsI1Loreg = (I2DestReg - I1DestReg) == 1; + unsigned LoRegDef = IsI1Loreg ? I1DestReg : I2DestReg; + + // Get the double word register. + unsigned DoubleRegDest = + TRI->getMatchingSuperReg(LoRegDef, Hexagon::subreg_loreg, + &Hexagon::DoubleRegsRegClass); + assert(DoubleRegDest != 0 && "Expect a valid register"); + + + // Setup source operands. + MachineOperand &LoOperand = IsI1Loreg ? I1->getOperand(1) : + I2->getOperand(1); + MachineOperand &HiOperand = IsI1Loreg ? I2->getOperand(1) : + I1->getOperand(1); + + // Figure out which source is a register and which a constant. + bool IsHiReg = HiOperand.isReg(); + bool IsLoReg = LoOperand.isReg(); + + MachineBasicBlock::iterator InsertPt(DoInsertAtI1 ? I1 : I2); + // Emit combine. + if (IsHiReg && IsLoReg) + emitCombineRR(InsertPt, DoubleRegDest, HiOperand, LoOperand); + else if (IsHiReg) + emitCombineRI(InsertPt, DoubleRegDest, HiOperand, LoOperand); + else if (IsLoReg) + emitCombineIR(InsertPt, DoubleRegDest, HiOperand, LoOperand); + else + emitCombineII(InsertPt, DoubleRegDest, HiOperand, LoOperand); + + I1->eraseFromParent(); + I2->eraseFromParent(); +} + +void HexagonCopyToCombine::emitCombineII(MachineBasicBlock::iterator &InsertPt, + unsigned DoubleDestReg, + MachineOperand &HiOperand, + MachineOperand &LoOperand) { + DebugLoc DL = InsertPt->getDebugLoc(); + MachineBasicBlock *BB = InsertPt->getParent(); + + // Handle globals. + if (HiOperand.isGlobal()) { + BuildMI(*BB, InsertPt, DL, TII->get(Hexagon::COMBINE_Ii), DoubleDestReg) + .addGlobalAddress(HiOperand.getGlobal(), HiOperand.getOffset(), + HiOperand.getTargetFlags()) + .addImm(LoOperand.getImm()); + return; + } + if (LoOperand.isGlobal()) { + BuildMI(*BB, InsertPt, DL, TII->get(Hexagon::COMBINE_iI_V4), DoubleDestReg) + .addImm(HiOperand.getImm()) + .addGlobalAddress(LoOperand.getGlobal(), LoOperand.getOffset(), + LoOperand.getTargetFlags()); + return; + } + + // Handle constant extended immediates. + if (!isInt<8>(HiOperand.getImm())) { + assert(isInt<8>(LoOperand.getImm())); + BuildMI(*BB, InsertPt, DL, TII->get(Hexagon::COMBINE_Ii), DoubleDestReg) + .addImm(HiOperand.getImm()) + .addImm(LoOperand.getImm()); + return; + } + + if (!isUInt<6>(LoOperand.getImm())) { + assert(isInt<8>(HiOperand.getImm())); + BuildMI(*BB, InsertPt, DL, TII->get(Hexagon::COMBINE_iI_V4), DoubleDestReg) + .addImm(HiOperand.getImm()) + .addImm(LoOperand.getImm()); + return; + } + + // Insert new combine instruction. + // DoubleRegDest = combine #HiImm, #LoImm + BuildMI(*BB, InsertPt, DL, TII->get(Hexagon::COMBINE_Ii), DoubleDestReg) + .addImm(HiOperand.getImm()) + .addImm(LoOperand.getImm()); +} + +void HexagonCopyToCombine::emitCombineIR(MachineBasicBlock::iterator &InsertPt, + unsigned DoubleDestReg, + MachineOperand &HiOperand, + MachineOperand &LoOperand) { + unsigned LoReg = LoOperand.getReg(); + unsigned LoRegKillFlag = getKillRegState(LoOperand.isKill()); + + DebugLoc DL = InsertPt->getDebugLoc(); + MachineBasicBlock *BB = InsertPt->getParent(); + + // Handle global. + if (HiOperand.isGlobal()) { + BuildMI(*BB, InsertPt, DL, TII->get(Hexagon::COMBINE_Ir_V4), DoubleDestReg) + .addGlobalAddress(HiOperand.getGlobal(), HiOperand.getOffset(), + HiOperand.getTargetFlags()) + .addReg(LoReg, LoRegKillFlag); + return; + } + // Insert new combine instruction. + // DoubleRegDest = combine #HiImm, LoReg + BuildMI(*BB, InsertPt, DL, TII->get(Hexagon::COMBINE_Ir_V4), DoubleDestReg) + .addImm(HiOperand.getImm()) + .addReg(LoReg, LoRegKillFlag); +} + +void HexagonCopyToCombine::emitCombineRI(MachineBasicBlock::iterator &InsertPt, + unsigned DoubleDestReg, + MachineOperand &HiOperand, + MachineOperand &LoOperand) { + unsigned HiRegKillFlag = getKillRegState(HiOperand.isKill()); + unsigned HiReg = HiOperand.getReg(); + + DebugLoc DL = InsertPt->getDebugLoc(); + MachineBasicBlock *BB = InsertPt->getParent(); + + // Handle global. + if (LoOperand.isGlobal()) { + BuildMI(*BB, InsertPt, DL, TII->get(Hexagon::COMBINE_rI_V4), DoubleDestReg) + .addReg(HiReg, HiRegKillFlag) + .addGlobalAddress(LoOperand.getGlobal(), LoOperand.getOffset(), + LoOperand.getTargetFlags()); + return; + } + + // Insert new combine instruction. + // DoubleRegDest = combine HiReg, #LoImm + BuildMI(*BB, InsertPt, DL, TII->get(Hexagon::COMBINE_rI_V4), DoubleDestReg) + .addReg(HiReg, HiRegKillFlag) + .addImm(LoOperand.getImm()); +} + +void HexagonCopyToCombine::emitCombineRR(MachineBasicBlock::iterator &InsertPt, + unsigned DoubleDestReg, + MachineOperand &HiOperand, + MachineOperand &LoOperand) { + unsigned LoRegKillFlag = getKillRegState(LoOperand.isKill()); + unsigned HiRegKillFlag = getKillRegState(HiOperand.isKill()); + unsigned LoReg = LoOperand.getReg(); + unsigned HiReg = HiOperand.getReg(); + + DebugLoc DL = InsertPt->getDebugLoc(); + MachineBasicBlock *BB = InsertPt->getParent(); + + // Insert new combine instruction. + // DoubleRegDest = combine HiReg, LoReg + BuildMI(*BB, InsertPt, DL, TII->get(Hexagon::COMBINE_rr), DoubleDestReg) + .addReg(HiReg, HiRegKillFlag) + .addReg(LoReg, LoRegKillFlag); +} + +FunctionPass *llvm::createHexagonCopyToCombine() { + return new HexagonCopyToCombine(); +} diff --git a/lib/Target/Hexagon/HexagonExpandPredSpillCode.cpp b/lib/Target/Hexagon/HexagonExpandPredSpillCode.cpp index 0814421..8a5991f 100644 --- a/lib/Target/Hexagon/HexagonExpandPredSpillCode.cpp +++ b/lib/Target/Hexagon/HexagonExpandPredSpillCode.cpp @@ -41,16 +41,24 @@ using namespace llvm; +namespace llvm { + void initializeHexagonExpandPredSpillCodePass(PassRegistry&); +} + + namespace { class HexagonExpandPredSpillCode : public MachineFunctionPass { - HexagonTargetMachine& QTM; + const HexagonTargetMachine& QTM; const HexagonSubtarget &QST; public: static char ID; - HexagonExpandPredSpillCode(HexagonTargetMachine& TM) : - MachineFunctionPass(ID), QTM(TM), QST(*TM.getSubtargetImpl()) {} + HexagonExpandPredSpillCode(const HexagonTargetMachine& TM) : + MachineFunctionPass(ID), QTM(TM), QST(*TM.getSubtargetImpl()) { + PassRegistry &Registry = *PassRegistry::getPassRegistry(); + initializeHexagonExpandPredSpillCodePass(Registry); + } const char *getPassName() const { return "Hexagon Expand Predicate Spill Code"; @@ -175,6 +183,19 @@ bool HexagonExpandPredSpillCode::runOnMachineFunction(MachineFunction &Fn) { // Public Constructor Functions //===----------------------------------------------------------------------===// -FunctionPass *llvm::createHexagonExpandPredSpillCode(HexagonTargetMachine &TM) { +static void initializePassOnce(PassRegistry &Registry) { + const char *Name = "Hexagon Expand Predicate Spill Code"; + PassInfo *PI = new PassInfo(Name, "hexagon-spill-pred", + &HexagonExpandPredSpillCode::ID, + 0, false, false); + Registry.registerPass(*PI, true); +} + +void llvm::initializeHexagonExpandPredSpillCodePass(PassRegistry &Registry) { + CALL_ONCE_INITIALIZATION(initializePassOnce) +} + +FunctionPass* +llvm::createHexagonExpandPredSpillCode(const HexagonTargetMachine &TM) { return new HexagonExpandPredSpillCode(TM); } diff --git a/lib/Target/Hexagon/HexagonFrameLowering.cpp b/lib/Target/Hexagon/HexagonFrameLowering.cpp index d6a9329..2b04f25 100644 --- a/lib/Target/Hexagon/HexagonFrameLowering.cpp +++ b/lib/Target/Hexagon/HexagonFrameLowering.cpp @@ -76,17 +76,12 @@ void HexagonFrameLowering::determineFrameLayout(MachineFunction &MF) const { void HexagonFrameLowering::emitPrologue(MachineFunction &MF) const { MachineBasicBlock &MBB = MF.front(); MachineFrameInfo *MFI = MF.getFrameInfo(); - MachineModuleInfo &MMI = MF.getMMI(); MachineBasicBlock::iterator MBBI = MBB.begin(); const HexagonRegisterInfo *QRI = static_cast<const HexagonRegisterInfo *>(MF.getTarget().getRegisterInfo()); DebugLoc dl = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc(); determineFrameLayout(MF); - // Check if frame moves are needed for EH. - bool needsFrameMoves = MMI.hasDebugInfo() || - !MF.getFunction()->needsUnwindTableEntry(); - // Get the number of bytes to allocate from the FrameInfo. int NumBytes = (int) MFI->getStackSize(); @@ -113,28 +108,6 @@ void HexagonFrameLowering::emitPrologue(MachineFunction &MF) const { MO.setImm(MFI->getMaxCallFrameSize()); } - std::vector<MachineMove> &Moves = MMI.getFrameMoves(); - - if (needsFrameMoves) { - // Advance CFA. DW_CFA_def_cfa - unsigned FPReg = QRI->getFrameRegister(); - unsigned RAReg = QRI->getRARegister(); - - MachineLocation Dst(MachineLocation::VirtualFP); - MachineLocation Src(FPReg, -8); - Moves.push_back(MachineMove(0, Dst, Src)); - - // R31 = (R31 - #4) - MachineLocation LRDst(RAReg, -4); - MachineLocation LRSrc(RAReg); - Moves.push_back(MachineMove(0, LRDst, LRSrc)); - - // R30 = (R30 - #8) - MachineLocation SPDst(FPReg, -8); - MachineLocation SPSrc(FPReg); - Moves.push_back(MachineMove(0, SPDst, SPSrc)); - } - // // Only insert ALLOCFRAME if we need to. // @@ -174,30 +147,55 @@ void HexagonFrameLowering::emitEpilogue(MachineFunction &MF, MachineBasicBlock::iterator MBBI = prior(MBB.end()); DebugLoc dl = MBBI->getDebugLoc(); // - // Only insert deallocframe if we need to. + // Only insert deallocframe if we need to. Also at -O0. See comment + // in emitPrologue above. // - if (hasFP(MF)) { + if (hasFP(MF) || MF.getTarget().getOptLevel() == CodeGenOpt::None) { MachineBasicBlock::iterator MBBI = prior(MBB.end()); MachineBasicBlock::iterator MBBI_end = MBB.end(); - // - // For Hexagon, we don't need the frame size. - // - MachineFrameInfo *MFI = MF.getFrameInfo(); - int NumBytes = (int) MFI->getStackSize(); const TargetInstrInfo &TII = *MF.getTarget().getInstrInfo(); - + // Handle EH_RETURN. + if (MBBI->getOpcode() == Hexagon::EH_RETURN_JMPR) { + assert(MBBI->getOperand(0).isReg() && "Offset should be in register!"); + BuildMI(MBB, MBBI, dl, TII.get(Hexagon::DEALLOCFRAME)); + BuildMI(MBB, MBBI, dl, TII.get(Hexagon::ADD_rr), + Hexagon::R29).addReg(Hexagon::R29).addReg(Hexagon::R28); + return; + } // Replace 'jumpr r31' instruction with dealloc_return for V4 and higher // versions. - if (STI.hasV4TOps() && MBBI->getOpcode() == Hexagon::JMPR + if (STI.hasV4TOps() && MBBI->getOpcode() == Hexagon::JMPret && !DisableDeallocRet) { - // Remove jumpr node. - MBB.erase(MBBI); + // Check for RESTORE_DEALLOC_RET_JMP_V4 call. Don't emit an extra DEALLOC + // instruction if we encounter it. + MachineBasicBlock::iterator BeforeJMPR = + MBB.begin() == MBBI ? MBBI : prior(MBBI); + if (BeforeJMPR != MBBI && + BeforeJMPR->getOpcode() == Hexagon::RESTORE_DEALLOC_RET_JMP_V4) { + // Remove the JMPR node. + MBB.erase(MBBI); + return; + } + // Add dealloc_return. - BuildMI(MBB, MBBI_end, dl, TII.get(Hexagon::DEALLOC_RET_V4)) - .addImm(NumBytes); - } else { // Add deallocframe for V2 and V3. - BuildMI(MBB, MBBI, dl, TII.get(Hexagon::DEALLOCFRAME)).addImm(NumBytes); + MachineInstrBuilder MIB = + BuildMI(MBB, MBBI_end, dl, TII.get(Hexagon::DEALLOC_RET_V4)); + // Transfer the function live-out registers. + MIB->copyImplicitOps(*MBB.getParent(), &*MBBI); + // Remove the JUMPR node. + MBB.erase(MBBI); + } else { // Add deallocframe for V2 and V3, and V4 tail calls. + // Check for RESTORE_DEALLOC_BEFORE_TAILCALL_V4. We don't need an extra + // DEALLOCFRAME instruction after it. + MachineBasicBlock::iterator Term = MBB.getFirstTerminator(); + MachineBasicBlock::iterator I = + Term == MBB.begin() ? MBB.end() : prior(Term); + if (I != MBB.end() && + I->getOpcode() == Hexagon::RESTORE_DEALLOC_BEFORE_TAILCALL_V4) + return; + + BuildMI(MBB, MBBI, dl, TII.get(Hexagon::DEALLOCFRAME)); } } } diff --git a/lib/Target/Hexagon/HexagonHardwareLoops.cpp b/lib/Target/Hexagon/HexagonHardwareLoops.cpp index 62aed13..3c4ca0f 100644 --- a/lib/Target/Hexagon/HexagonHardwareLoops.cpp +++ b/lib/Target/Hexagon/HexagonHardwareLoops.cpp @@ -134,7 +134,7 @@ namespace { /// has a computable trip count and, if so, return a value that represents /// the trip count expression. CountValue *getLoopTripCount(MachineLoop *L, - SmallVector<MachineInstr*, 2> &OldInsts); + SmallVectorImpl<MachineInstr *> &OldInsts); /// \brief Return the expression that represents the number of times /// a loop iterates. The function takes the operands that represent the @@ -164,7 +164,7 @@ namespace { /// \brief Return true if the instruction is now dead. bool isDead(const MachineInstr *MI, - SmallVector<MachineInstr*, 1> &DeadPhis) const; + SmallVectorImpl<MachineInstr *> &DeadPhis) const; /// \brief Remove the instruction if it is now dead. void removeIfDead(MachineInstr *MI); @@ -428,7 +428,7 @@ bool HexagonHardwareLoops::findInductionRegister(MachineLoop *L, /// induction variable patterns that are used in the calculation for /// the number of time the loop is executed. CountValue *HexagonHardwareLoops::getLoopTripCount(MachineLoop *L, - SmallVector<MachineInstr*, 2> &OldInsts) { + SmallVectorImpl<MachineInstr *> &OldInsts) { MachineBasicBlock *TopMBB = L->getTopBlock(); MachineBasicBlock::pred_iterator PI = TopMBB->pred_begin(); assert(PI != TopMBB->pred_end() && @@ -541,12 +541,6 @@ CountValue *HexagonHardwareLoops::getLoopTripCount(MachineLoop *L, case Hexagon::CMPEQrr: Cmp = !Negated ? Comparison::EQ : Comparison::NE; break; - case Hexagon::CMPLTrr: - Cmp = !Negated ? Comparison::LTs : Comparison::GEs; - break; - case Hexagon::CMPLTUrr: - Cmp = !Negated ? Comparison::LTu : Comparison::GEu; - break; case Hexagon::CMPGTUri: case Hexagon::CMPGTUrr: Cmp = !Negated ? Comparison::GTu : Comparison::LEu; @@ -701,7 +695,7 @@ CountValue *HexagonHardwareLoops::computeCount(MachineLoop *Loop, // If the induction variable bump is not a power of 2, quit. // Othwerise we'd need a general integer division. - if (!isPowerOf2_64(abs(IVBump))) + if (!isPowerOf2_64(abs64(IVBump))) return 0; MachineBasicBlock *PH = Loop->getLoopPreheader(); @@ -896,7 +890,7 @@ bool HexagonHardwareLoops::containsInvalidInstruction(MachineLoop *L) const { /// for inline asm, physical registers and instructions with side effects /// removed. bool HexagonHardwareLoops::isDead(const MachineInstr *MI, - SmallVector<MachineInstr*, 1> &DeadPhis) const { + SmallVectorImpl<MachineInstr *> &DeadPhis) const { // Examine each operand. for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { const MachineOperand &MO = MI->getOperand(i); @@ -1125,8 +1119,8 @@ bool HexagonHardwareLoops::convertToHardwareLoop(MachineLoop *L) { // The loop ends with either: // - a conditional branch followed by an unconditional branch, or // - a conditional branch to the loop start. - if (LastI->getOpcode() == Hexagon::JMP_c || - LastI->getOpcode() == Hexagon::JMP_cNot) { + if (LastI->getOpcode() == Hexagon::JMP_t || + LastI->getOpcode() == Hexagon::JMP_f) { // Delete one and change/add an uncond. branch to out of the loop. MachineBasicBlock *BranchTarget = LastI->getOperand(1).getMBB(); LastI = LastMBB->erase(LastI); @@ -1430,7 +1424,6 @@ MachineBasicBlock *HexagonHardwareLoops::createPreheaderForLoop( return 0; typedef MachineBasicBlock::instr_iterator instr_iterator; - typedef MachineBasicBlock::pred_iterator pred_iterator; // Verify that all existing predecessors have analyzable branches // (or no branches at all). diff --git a/lib/Target/Hexagon/HexagonISelDAGToDAG.cpp b/lib/Target/Hexagon/HexagonISelDAGToDAG.cpp index 3a1c48b..9e78e51 100644 --- a/lib/Target/Hexagon/HexagonISelDAGToDAG.cpp +++ b/lib/Target/Hexagon/HexagonISelDAGToDAG.cpp @@ -49,16 +49,14 @@ class HexagonDAGToDAGISel : public SelectionDAGISel { const HexagonSubtarget &Subtarget; // Keep a reference to HexagonTargetMachine. - HexagonTargetMachine& TM; - const HexagonInstrInfo *TII; + const HexagonTargetMachine& TM; DenseMap<const GlobalValue *, unsigned> GlobalAddressUseCountMap; public: explicit HexagonDAGToDAGISel(HexagonTargetMachine &targetmachine, CodeGenOpt::Level OptLevel) : SelectionDAGISel(targetmachine, OptLevel), Subtarget(targetmachine.getSubtarget<HexagonSubtarget>()), - TM(targetmachine), - TII(static_cast<const HexagonInstrInfo*>(TM.getInstrInfo())) { + TM(targetmachine) { initializeHexagonDAGToDAGISelPass(*PassRegistry::getPassRegistry()); } bool hasNumUsesBelowThresGA(SDNode *N) const; @@ -92,14 +90,14 @@ public: bool SelectAddr(SDNode *Op, SDValue Addr, SDValue &Base, SDValue &Offset); SDNode *SelectLoad(SDNode *N); - SDNode *SelectBaseOffsetLoad(LoadSDNode *LD, DebugLoc dl); - SDNode *SelectIndexedLoad(LoadSDNode *LD, DebugLoc dl); + SDNode *SelectBaseOffsetLoad(LoadSDNode *LD, SDLoc dl); + SDNode *SelectIndexedLoad(LoadSDNode *LD, SDLoc dl); SDNode *SelectIndexedLoadZeroExtend64(LoadSDNode *LD, unsigned Opcode, - DebugLoc dl); + SDLoc dl); SDNode *SelectIndexedLoadSignExtend64(LoadSDNode *LD, unsigned Opcode, - DebugLoc dl); - SDNode *SelectBaseOffsetStore(StoreSDNode *ST, DebugLoc dl); - SDNode *SelectIndexedStore(StoreSDNode *ST, DebugLoc dl); + SDLoc dl); + SDNode *SelectBaseOffsetStore(StoreSDNode *ST, SDLoc dl); + SDNode *SelectIndexedStore(StoreSDNode *ST, SDLoc dl); SDNode *SelectStore(SDNode *N); SDNode *SelectSHL(SDNode *N); SDNode *SelectSelect(SDNode *N); @@ -113,6 +111,46 @@ public: SDNode *SelectAdd(SDNode *N); bool isConstExtProfitable(SDNode *N) const; +// XformMskToBitPosU5Imm - Returns the bit position which +// the single bit 32 bit mask represents. +// Used in Clr and Set bit immediate memops. +SDValue XformMskToBitPosU5Imm(uint32_t Imm) { + int32_t bitPos; + bitPos = Log2_32(Imm); + assert(bitPos >= 0 && bitPos < 32 && + "Constant out of range for 32 BitPos Memops"); + return CurDAG->getTargetConstant(bitPos, MVT::i32); +} + +// XformMskToBitPosU4Imm - Returns the bit position which the single bit 16 bit +// mask represents. Used in Clr and Set bit immediate memops. +SDValue XformMskToBitPosU4Imm(uint16_t Imm) { + return XformMskToBitPosU5Imm(Imm); +} + +// XformMskToBitPosU3Imm - Returns the bit position which the single bit 8 bit +// mask represents. Used in Clr and Set bit immediate memops. +SDValue XformMskToBitPosU3Imm(uint8_t Imm) { + return XformMskToBitPosU5Imm(Imm); +} + +// Return true if there is exactly one bit set in V, i.e., if V is one of the +// following integers: 2^0, 2^1, ..., 2^31. +bool ImmIsSingleBit(uint32_t v) const { + uint32_t c = CountPopulation_64(v); + // Only return true if we counted 1 bit. + return c == 1; +} + +// XformM5ToU5Imm - Return a target constant with the specified value, of type +// i32 where the negative literal is transformed into a positive literal for +// use in -= memops. +inline SDValue XformM5ToU5Imm(signed Imm) { + assert( (Imm >= -31 && Imm <= -1) && "Constant out of range for Memops"); + return CurDAG->getTargetConstant( - Imm, MVT::i32); +} + + // XformU7ToU7M1Imm - Return a target constant decremented by 1, in range // [1..128], used in cmpb.gtu instructions. inline SDValue XformU7ToU7M1Imm(signed Imm) { @@ -120,6 +158,17 @@ inline SDValue XformU7ToU7M1Imm(signed Imm) { return CurDAG->getTargetConstant(Imm - 1, MVT::i8); } +// XformS8ToS8M1Imm - Return a target constant decremented by 1. +inline SDValue XformSToSM1Imm(signed Imm) { + return CurDAG->getTargetConstant(Imm - 1, MVT::i32); +} + +// XformU8ToU8M1Imm - Return a target constant decremented by 1. +inline SDValue XformUToUM1Imm(unsigned Imm) { + assert((Imm >= 1) && "Cannot decrement unsigned int less than 1"); + return CurDAG->getTargetConstant(Imm - 1, MVT::i32); +} + // Include the pieces autogenerated from the target description. #include "HexagonGenDAGISel.inc" }; @@ -334,7 +383,7 @@ static bool OffsetFitsS11(EVT MemType, int64_t Offset) { // lowering for GlobalAddress nodes has already turned it into a // CONST32. // -SDNode *HexagonDAGToDAGISel::SelectBaseOffsetLoad(LoadSDNode *LD, DebugLoc dl) { +SDNode *HexagonDAGToDAGISel::SelectBaseOffsetLoad(LoadSDNode *LD, SDLoc dl) { SDValue Chain = LD->getChain(); SDNode* Const32 = LD->getBasePtr().getNode(); unsigned Opcode = 0; @@ -345,7 +394,7 @@ SDNode *HexagonDAGToDAGISel::SelectBaseOffsetLoad(LoadSDNode *LD, DebugLoc dl) { EVT LoadedVT = LD->getMemoryVT(); int64_t Offset = cast<GlobalAddressSDNode>(Base)->getOffset(); if (Offset != 0 && OffsetFitsS11(LoadedVT, Offset)) { - MVT PointerTy = TLI.getPointerTy(); + MVT PointerTy = getTargetLowering()->getPointerTy(); const GlobalValue* GV = cast<GlobalAddressSDNode>(Base)->getGlobal(); SDValue TargAddr = @@ -382,7 +431,7 @@ SDNode *HexagonDAGToDAGISel::SelectBaseOffsetLoad(LoadSDNode *LD, DebugLoc dl) { SDNode *HexagonDAGToDAGISel::SelectIndexedLoadSignExtend64(LoadSDNode *LD, unsigned Opcode, - DebugLoc dl) + SDLoc dl) { SDValue Chain = LD->getChain(); EVT LoadedVT = LD->getMemoryVT(); @@ -393,8 +442,11 @@ SDNode *HexagonDAGToDAGISel::SelectIndexedLoadSignExtend64(LoadSDNode *LD, SDValue N1 = LD->getOperand(1); SDValue CPTmpN1_0; SDValue CPTmpN1_1; + if (SelectADDRriS11_2(N1, CPTmpN1_0, CPTmpN1_1) && N1.getNode()->getValueType(0) == MVT::i32) { + const HexagonInstrInfo *TII = + static_cast<const HexagonInstrInfo*>(TM.getInstrInfo()); if (TII->isValidAutoIncImm(LoadedVT, Val)) { SDValue TargetConst = CurDAG->getTargetConstant(Val, MVT::i32); SDNode *Result_1 = CurDAG->getMachineNode(Opcode, dl, MVT::i32, MVT::i32, @@ -446,7 +498,7 @@ SDNode *HexagonDAGToDAGISel::SelectIndexedLoadSignExtend64(LoadSDNode *LD, SDNode *HexagonDAGToDAGISel::SelectIndexedLoadZeroExtend64(LoadSDNode *LD, unsigned Opcode, - DebugLoc dl) + SDLoc dl) { SDValue Chain = LD->getChain(); EVT LoadedVT = LD->getMemoryVT(); @@ -457,8 +509,11 @@ SDNode *HexagonDAGToDAGISel::SelectIndexedLoadZeroExtend64(LoadSDNode *LD, SDValue N1 = LD->getOperand(1); SDValue CPTmpN1_0; SDValue CPTmpN1_1; + if (SelectADDRriS11_2(N1, CPTmpN1_0, CPTmpN1_1) && N1.getNode()->getValueType(0) == MVT::i32) { + const HexagonInstrInfo *TII = + static_cast<const HexagonInstrInfo*>(TM.getInstrInfo()); if (TII->isValidAutoIncImm(LoadedVT, Val)) { SDValue TargetConstVal = CurDAG->getTargetConstant(Val, MVT::i32); SDValue TargetConst0 = CurDAG->getTargetConstant(0, MVT::i32); @@ -521,7 +576,7 @@ SDNode *HexagonDAGToDAGISel::SelectIndexedLoadZeroExtend64(LoadSDNode *LD, } -SDNode *HexagonDAGToDAGISel::SelectIndexedLoad(LoadSDNode *LD, DebugLoc dl) { +SDNode *HexagonDAGToDAGISel::SelectIndexedLoad(LoadSDNode *LD, SDLoc dl) { SDValue Chain = LD->getChain(); SDValue Base = LD->getBasePtr(); SDValue Offset = LD->getOffset(); @@ -535,6 +590,8 @@ SDNode *HexagonDAGToDAGISel::SelectIndexedLoad(LoadSDNode *LD, DebugLoc dl) { bool zextval = (LD->getExtensionType() == ISD::ZEXTLOAD); // Figure out the opcode. + const HexagonInstrInfo *TII = + static_cast<const HexagonInstrInfo*>(TM.getInstrInfo()); if (LoadedVT == MVT::i64) { if (TII->isValidAutoIncImm(LoadedVT, Val)) Opcode = Hexagon::POST_LDrid; @@ -616,7 +673,7 @@ SDNode *HexagonDAGToDAGISel::SelectIndexedLoad(LoadSDNode *LD, DebugLoc dl) { SDNode *HexagonDAGToDAGISel::SelectLoad(SDNode *N) { SDNode *result; - DebugLoc dl = N->getDebugLoc(); + SDLoc dl(N); LoadSDNode *LD = cast<LoadSDNode>(N); ISD::MemIndexedMode AM = LD->getAddressingMode(); @@ -631,7 +688,7 @@ SDNode *HexagonDAGToDAGISel::SelectLoad(SDNode *N) { } -SDNode *HexagonDAGToDAGISel::SelectIndexedStore(StoreSDNode *ST, DebugLoc dl) { +SDNode *HexagonDAGToDAGISel::SelectIndexedStore(StoreSDNode *ST, SDLoc dl) { SDValue Chain = ST->getChain(); SDValue Base = ST->getBasePtr(); SDValue Offset = ST->getOffset(); @@ -643,6 +700,8 @@ SDNode *HexagonDAGToDAGISel::SelectIndexedStore(StoreSDNode *ST, DebugLoc dl) { // Offset value must be within representable range // and must have correct alignment properties. + const HexagonInstrInfo *TII = + static_cast<const HexagonInstrInfo*>(TM.getInstrInfo()); if (TII->isValidAutoIncImm(StoredVT, Val)) { SDValue Ops[] = {Base, CurDAG->getTargetConstant(Val, MVT::i32), Value, Chain}; @@ -657,7 +716,7 @@ SDNode *HexagonDAGToDAGISel::SelectIndexedStore(StoreSDNode *ST, DebugLoc dl) { // Build post increment store. SDNode* Result = CurDAG->getMachineNode(Opcode, dl, MVT::i32, - MVT::Other, Ops, 4); + MVT::Other, Ops); MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1); MemOp[0] = ST->getMemOperand(); cast<MachineSDNode>(Result)->setMemRefs(MemOp, MemOp + 1); @@ -683,8 +742,7 @@ SDNode *HexagonDAGToDAGISel::SelectIndexedStore(StoreSDNode *ST, DebugLoc dl) { // Build regular store. SDValue TargetConstVal = CurDAG->getTargetConstant(Val, MVT::i32); - SDNode* Result_1 = CurDAG->getMachineNode(Opcode, dl, MVT::Other, Ops, - 4); + SDNode* Result_1 = CurDAG->getMachineNode(Opcode, dl, MVT::Other, Ops); // Build splitted incriment instruction. SDNode* Result_2 = CurDAG->getMachineNode(Hexagon::ADD_ri, dl, MVT::i32, Base, @@ -701,7 +759,7 @@ SDNode *HexagonDAGToDAGISel::SelectIndexedStore(StoreSDNode *ST, DebugLoc dl) { SDNode *HexagonDAGToDAGISel::SelectBaseOffsetStore(StoreSDNode *ST, - DebugLoc dl) { + SDLoc dl) { SDValue Chain = ST->getChain(); SDNode* Const32 = ST->getBasePtr().getNode(); SDValue Value = ST->getValue(); @@ -719,7 +777,7 @@ SDNode *HexagonDAGToDAGISel::SelectBaseOffsetStore(StoreSDNode *ST, EVT StoredVT = ST->getMemoryVT(); int64_t Offset = cast<GlobalAddressSDNode>(Base)->getOffset(); if (Offset != 0 && OffsetFitsS11(StoredVT, Offset)) { - MVT PointerTy = TLI.getPointerTy(); + MVT PointerTy = getTargetLowering()->getPointerTy(); const GlobalValue* GV = cast<GlobalAddressSDNode>(Base)->getGlobal(); SDValue TargAddr = @@ -740,7 +798,7 @@ SDNode *HexagonDAGToDAGISel::SelectBaseOffsetStore(StoreSDNode *ST, Value, Chain}; // build indexed store SDNode* Result = CurDAG->getMachineNode(Opcode, dl, - MVT::Other, Ops, 4); + MVT::Other, Ops); MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1); MemOp[0] = ST->getMemOperand(); cast<MachineSDNode>(Result)->setMemRefs(MemOp, MemOp + 1); @@ -755,7 +813,7 @@ SDNode *HexagonDAGToDAGISel::SelectBaseOffsetStore(StoreSDNode *ST, SDNode *HexagonDAGToDAGISel::SelectStore(SDNode *N) { - DebugLoc dl = N->getDebugLoc(); + SDLoc dl(N); StoreSDNode *ST = cast<StoreSDNode>(N); ISD::MemIndexedMode AM = ST->getAddressingMode(); @@ -768,7 +826,7 @@ SDNode *HexagonDAGToDAGISel::SelectStore(SDNode *N) { } SDNode *HexagonDAGToDAGISel::SelectMul(SDNode *N) { - DebugLoc dl = N->getDebugLoc(); + SDLoc dl(N); // // %conv.i = sext i32 %tmp1 to i64 @@ -852,7 +910,7 @@ SDNode *HexagonDAGToDAGISel::SelectMul(SDNode *N) { SDNode *HexagonDAGToDAGISel::SelectSelect(SDNode *N) { - DebugLoc dl = N->getDebugLoc(); + SDLoc dl(N); SDValue N0 = N->getOperand(0); if (N0.getOpcode() == ISD::SETCC) { SDValue N00 = N0.getOperand(0); @@ -919,7 +977,7 @@ SDNode *HexagonDAGToDAGISel::SelectSelect(SDNode *N) { SDNode *HexagonDAGToDAGISel::SelectTruncate(SDNode *N) { - DebugLoc dl = N->getDebugLoc(); + SDLoc dl(N); SDValue Shift = N->getOperand(0); // @@ -1032,7 +1090,7 @@ SDNode *HexagonDAGToDAGISel::SelectTruncate(SDNode *N) { SDNode *HexagonDAGToDAGISel::SelectSHL(SDNode *N) { - DebugLoc dl = N->getDebugLoc(); + SDLoc dl(N); if (N->getValueType(0) == MVT::i32) { SDValue Shl_0 = N->getOperand(0); SDValue Shl_1 = N->getOperand(1); @@ -1108,7 +1166,7 @@ SDNode *HexagonDAGToDAGISel::SelectSHL(SDNode *N) { // We want to preserve all the lower 8-bits and, not just 1 LSB bit. // SDNode *HexagonDAGToDAGISel::SelectZeroExtend(SDNode *N) { - DebugLoc dl = N->getDebugLoc(); + SDLoc dl(N); SDNode *IsIntrinsic = N->getOperand(0).getNode(); if ((IsIntrinsic->getOpcode() == ISD::INTRINSIC_WO_CHAIN)) { unsigned ID = @@ -1151,7 +1209,7 @@ SDNode *HexagonDAGToDAGISel::SelectZeroExtend(SDNode *N) { // and lowering to the actual intrinsic. // SDNode *HexagonDAGToDAGISel::SelectIntrinsicWOChain(SDNode *N) { - DebugLoc dl = N->getDebugLoc(); + SDLoc dl(N); unsigned ID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue(); unsigned IntrinsicWithPred = doesIntrinsicContainPredicate(ID); @@ -1159,6 +1217,8 @@ SDNode *HexagonDAGToDAGISel::SelectIntrinsicWOChain(SDNode *N) { // as at least one of the operands. if (IntrinsicWithPred) { SmallVector<SDValue, 8> Ops; + const HexagonInstrInfo *TII = + static_cast<const HexagonInstrInfo*>(TM.getInstrInfo()); const MCInstrDesc &MCID = TII->get(IntrinsicWithPred); const TargetRegisterInfo *TRI = TM.getRegisterInfo(); @@ -1190,8 +1250,7 @@ SDNode *HexagonDAGToDAGISel::SelectIntrinsicWOChain(SDNode *N) { } EVT ReturnValueVT = N->getValueType(0); SDNode *Result = CurDAG->getMachineNode(IntrinsicWithPred, dl, - ReturnValueVT, - Ops.data(), Ops.size()); + ReturnValueVT, Ops); ReplaceUses(N, Result); return Result; } @@ -1202,7 +1261,7 @@ SDNode *HexagonDAGToDAGISel::SelectIntrinsicWOChain(SDNode *N) { // Map floating point constant values. // SDNode *HexagonDAGToDAGISel::SelectConstantFP(SDNode *N) { - DebugLoc dl = N->getDebugLoc(); + SDLoc dl(N); ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(N); APFloat APF = CN->getValueAPF(); if (N->getValueType(0) == MVT::f32) { @@ -1222,7 +1281,7 @@ SDNode *HexagonDAGToDAGISel::SelectConstantFP(SDNode *N) { // Map predicate true (encoded as -1 in LLVM) to a XOR. // SDNode *HexagonDAGToDAGISel::SelectConstant(SDNode *N) { - DebugLoc dl = N->getDebugLoc(); + SDLoc dl(N); if (N->getValueType(0) == MVT::i1) { SDNode* Result; int32_t Val = cast<ConstantSDNode>(N)->getSExtValue(); @@ -1261,7 +1320,7 @@ SDNode *HexagonDAGToDAGISel::SelectConstant(SDNode *N) { // Map add followed by a asr -> asr +=. // SDNode *HexagonDAGToDAGISel::SelectAdd(SDNode *N) { - DebugLoc dl = N->getDebugLoc(); + SDLoc dl(N); if (N->getValueType(0) != MVT::i32) { return SelectCode(N); } @@ -1611,7 +1670,7 @@ bool HexagonDAGToDAGISel::foldGlobalAddressImpl(SDValue &N, SDValue &R, !hasNumUsesBelowThresGA(GA)) return false; R = CurDAG->getTargetGlobalAddress(GA->getGlobal(), - Const->getDebugLoc(), + SDLoc(Const), N.getValueType(), GA->getOffset() + (uint64_t)Const->getSExtValue()); diff --git a/lib/Target/Hexagon/HexagonISelLowering.cpp b/lib/Target/Hexagon/HexagonISelLowering.cpp index 0a8b1af..567faca 100644 --- a/lib/Target/Hexagon/HexagonISelLowering.cpp +++ b/lib/Target/Hexagon/HexagonISelLowering.cpp @@ -285,7 +285,7 @@ const { static SDValue CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain, ISD::ArgFlagsTy Flags, SelectionDAG &DAG, - DebugLoc dl) { + SDLoc dl) { SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32); return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(), @@ -302,7 +302,7 @@ HexagonTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool isVarArg, const SmallVectorImpl<ISD::OutputArg> &Outs, const SmallVectorImpl<SDValue> &OutVals, - DebugLoc dl, SelectionDAG &DAG) const { + SDLoc dl, SelectionDAG &DAG) const { // CCValAssign - represent the assignment of the return value to locations. SmallVector<CCValAssign, 16> RVLocs; @@ -351,7 +351,7 @@ HexagonTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg, const SmallVectorImpl<ISD::InputArg> &Ins, - DebugLoc dl, SelectionDAG &DAG, + SDLoc dl, SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals, const SmallVectorImpl<SDValue> &OutVals, SDValue Callee) const { @@ -382,10 +382,10 @@ SDValue HexagonTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, SmallVectorImpl<SDValue> &InVals) const { SelectionDAG &DAG = CLI.DAG; - DebugLoc &dl = CLI.DL; - SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs; - SmallVector<SDValue, 32> &OutVals = CLI.OutVals; - SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins; + SDLoc &dl = CLI.DL; + SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs; + SmallVectorImpl<SDValue> &OutVals = CLI.OutVals; + SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins; SDValue Chain = CLI.Chain; SDValue Callee = CLI.Callee; bool &isTailCall = CLI.IsTailCall; @@ -513,7 +513,8 @@ HexagonTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, if (!isTailCall) Chain = DAG.getCALLSEQ_START(Chain, DAG.getConstant(NumBytes, - getPointerTy(), true)); + getPointerTy(), true), + dl); // Build a sequence of copy-to-reg nodes chained together with token // chain and flag operands which copy the outgoing args into registers. @@ -588,7 +589,7 @@ HexagonTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, // Create the CALLSEQ_END node. Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true), - DAG.getIntPtrConstant(0, true), InFlag); + DAG.getIntPtrConstant(0, true), InFlag, dl); InFlag = Chain.getValue(1); // Handle result values, copying them out of physregs into vregs that we @@ -730,7 +731,7 @@ LowerBR_JT(SDValue Op, SelectionDAG &DAG) const SDValue Chain = Op.getOperand(0); SDValue Table = Op.getOperand(1); SDValue Index = Op.getOperand(2); - DebugLoc dl = Op.getDebugLoc(); + SDLoc dl(Op); JumpTableSDNode *JT = cast<JumpTableSDNode>(Table); unsigned JTI = JT->getIndex(); MachineFunction &MF = DAG.getMachineFunction(); @@ -766,7 +767,7 @@ HexagonTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const { SDValue Chain = Op.getOperand(0); SDValue Size = Op.getOperand(1); - DebugLoc dl = Op.getDebugLoc(); + SDLoc dl(Op); unsigned SPReg = getStackPointerRegisterToSaveRestore(); @@ -812,7 +813,7 @@ HexagonTargetLowering::LowerFormalArguments(SDValue Chain, bool isVarArg, const SmallVectorImpl<ISD::InputArg> &Ins, - DebugLoc dl, SelectionDAG &DAG, + SDLoc dl, SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { @@ -925,7 +926,7 @@ HexagonTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const { HexagonMachineFunctionInfo *QFI = MF.getInfo<HexagonMachineFunctionInfo>(); SDValue Addr = DAG.getFrameIndex(QFI->getVarArgsFrameIndex(), MVT::i32); const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); - return DAG.getStore(Op.getOperand(0), Op.getDebugLoc(), Addr, + return DAG.getStore(Op.getOperand(0), SDLoc(Op), Addr, Op.getOperand(1), MachinePointerInfo(SV), false, false, 0); } @@ -937,7 +938,7 @@ HexagonTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const { SDValue CC = Op.getOperand(4); SDValue TrueVal = Op.getOperand(2); SDValue FalseVal = Op.getOperand(3); - DebugLoc dl = Op.getDebugLoc(); + SDLoc dl(Op); SDNode* OpNode = Op.getNode(); EVT SVT = OpNode->getValueType(0); @@ -948,8 +949,7 @@ HexagonTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const { SDValue HexagonTargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const { EVT ValTy = Op.getValueType(); - - DebugLoc dl = Op.getDebugLoc(); + SDLoc dl(Op); ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op); SDValue Res; if (CP->isMachineConstantPoolEntry()) @@ -969,7 +969,7 @@ HexagonTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const { MFI->setReturnAddressIsTaken(true); EVT VT = Op.getValueType(); - DebugLoc dl = Op.getDebugLoc(); + SDLoc dl(Op); unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); if (Depth) { SDValue FrameAddr = LowerFRAMEADDR(Op, DAG); @@ -991,7 +991,7 @@ HexagonTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const { MFI->setFrameAddressIsTaken(true); EVT VT = Op.getValueType(); - DebugLoc dl = Op.getDebugLoc(); + SDLoc dl(Op); unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, TRI->getFrameRegister(), VT); @@ -1002,17 +1002,9 @@ HexagonTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const { return FrameAddr; } - -SDValue HexagonTargetLowering::LowerMEMBARRIER(SDValue Op, - SelectionDAG& DAG) const { - DebugLoc dl = Op.getDebugLoc(); - return DAG.getNode(HexagonISD::BARRIER, dl, MVT::Other, Op.getOperand(0)); -} - - SDValue HexagonTargetLowering::LowerATOMIC_FENCE(SDValue Op, SelectionDAG& DAG) const { - DebugLoc dl = Op.getDebugLoc(); + SDLoc dl(Op); return DAG.getNode(HexagonISD::BARRIER, dl, MVT::Other, Op.getOperand(0)); } @@ -1022,7 +1014,7 @@ SDValue HexagonTargetLowering::LowerGLOBALADDRESS(SDValue Op, SDValue Result; const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal(); int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset(); - DebugLoc dl = Op.getDebugLoc(); + SDLoc dl(Op); Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset); const HexagonTargetObjectFile &TLOF = @@ -1038,7 +1030,7 @@ SDValue HexagonTargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const { const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress(); SDValue BA_SD = DAG.getTargetBlockAddress(BA, MVT::i32); - DebugLoc dl = Op.getDebugLoc(); + SDLoc dl(Op); return DAG.getNode(HexagonISD::CONST32_GP, dl, getPointerTy(), BA_SD); } @@ -1361,7 +1353,6 @@ HexagonTargetLowering::HexagonTargetLowering(HexagonTargetMachine } - setOperationAction(ISD::BRIND, MVT::Other, Expand); if (EmitJumpTables) { setOperationAction(ISD::BR_JT, MVT::Other, Custom); } else { @@ -1374,8 +1365,8 @@ HexagonTargetLowering::HexagonTargetLowering(HexagonTargetMachine setOperationAction(ISD::BR_CC, MVT::f64, Expand); setOperationAction(ISD::BR_CC, MVT::i1, Expand); setOperationAction(ISD::BR_CC, MVT::i32, Expand); + setOperationAction(ISD::BR_CC, MVT::i64, Expand); - setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom); setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom); setOperationAction(ISD::FSIN , MVT::f64, Expand); @@ -1437,12 +1428,7 @@ HexagonTargetLowering::HexagonTargetLowering(HexagonTargetMachine setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand); setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand); - setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand); - setOperationAction(ISD::EHSELECTION, MVT::i64, Expand); - setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand); - setOperationAction(ISD::EHSELECTION, MVT::i32, Expand); - - setOperationAction(ISD::EH_RETURN, MVT::Other, Expand); + setOperationAction(ISD::EH_RETURN, MVT::Other, Custom); if (TM.getSubtargetImpl()->isSubtargetV2()) { setExceptionPointerRegister(Hexagon::R20); @@ -1497,6 +1483,7 @@ HexagonTargetLowering::getTargetNodeName(unsigned Opcode) const { case HexagonISD::RET_FLAG: return "HexagonISD::RET_FLAG"; case HexagonISD::BR_JT: return "HexagonISD::BR_JT"; case HexagonISD::TC_RETURN: return "HexagonISD::TC_RETURN"; + case HexagonISD::EH_RETURN: return "HexagonISD::EH_RETURN"; } } @@ -1517,17 +1504,57 @@ bool HexagonTargetLowering::isTruncateFree(EVT VT1, EVT VT2) const { return ((VT1.getSimpleVT() == MVT::i64) && (VT2.getSimpleVT() == MVT::i32)); } +bool +HexagonTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const { + // Assuming the caller does not have either a signext or zeroext modifier, and + // only one value is accepted, any reasonable truncation is allowed. + if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy()) + return false; + + // FIXME: in principle up to 64-bit could be made safe, but it would be very + // fragile at the moment: any support for multiple value returns would be + // liable to disallow tail calls involving i64 -> iN truncation in many cases. + return Ty1->getPrimitiveSizeInBits() <= 32; +} + +SDValue +HexagonTargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const { + SDValue Chain = Op.getOperand(0); + SDValue Offset = Op.getOperand(1); + SDValue Handler = Op.getOperand(2); + SDLoc dl(Op); + + // Mark function as containing a call to EH_RETURN. + HexagonMachineFunctionInfo *FuncInfo = + DAG.getMachineFunction().getInfo<HexagonMachineFunctionInfo>(); + FuncInfo->setHasEHReturn(); + + unsigned OffsetReg = Hexagon::R28; + + SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), + DAG.getRegister(Hexagon::R30, getPointerTy()), + DAG.getIntPtrConstant(4)); + Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(), + false, false, 0); + Chain = DAG.getCopyToReg(Chain, dl, OffsetReg, Offset); + + // Not needed we already use it as explict input to EH_RETURN. + // MF.getRegInfo().addLiveOut(OffsetReg); + + return DAG.getNode(HexagonISD::EH_RETURN, dl, MVT::Other, Chain); +} + SDValue HexagonTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const { switch (Op.getOpcode()) { default: llvm_unreachable("Should not custom lower this!"); case ISD::ConstantPool: return LowerConstantPool(Op, DAG); + case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG); // Frame & Return address. Currently unimplemented. case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG); case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG); case ISD::GlobalTLSAddress: llvm_unreachable("TLS not implemented for Hexagon."); - case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG); case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG); case ISD::GlobalAddress: return LowerGLOBALADDRESS(Op, DAG); case ISD::BlockAddress: return LowerBlockAddress(Op, DAG); @@ -1571,11 +1598,11 @@ const { std::pair<unsigned, const TargetRegisterClass*> HexagonTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint, - EVT VT) const { + MVT VT) const { if (Constraint.size() == 1) { switch (Constraint[0]) { case 'r': // R0-R31 - switch (VT.getSimpleVT().SimpleTy) { + switch (VT.SimpleTy) { default: llvm_unreachable("getRegForInlineAsmConstraint Unhandled data type"); case MVT::i32: diff --git a/lib/Target/Hexagon/HexagonISelLowering.h b/lib/Target/Hexagon/HexagonISelLowering.h index 3279cc6..4fe0107 100644 --- a/lib/Target/Hexagon/HexagonISelLowering.h +++ b/lib/Target/Hexagon/HexagonISelLowering.h @@ -62,7 +62,8 @@ namespace llvm { WrapperShuffEH, WrapperShuffOB, WrapperShuffOH, - TC_RETURN + TC_RETURN, + EH_RETURN }; } @@ -94,6 +95,8 @@ namespace llvm { virtual bool isTruncateFree(Type *Ty1, Type *Ty2) const; virtual bool isTruncateFree(EVT VT1, EVT VT2) const; + virtual bool allowTruncateForTailCall(Type *Ty1, Type *Ty2) const; + virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const; virtual const char *getTargetNodeName(unsigned Opcode) const; @@ -101,10 +104,11 @@ namespace llvm { SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const; SDValue LowerINLINEASM(SDValue Op, SelectionDAG &DAG) const; SDValue LowerEH_LABEL(SDValue Op, SelectionDAG &DAG) const; + SDValue LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const; SDValue LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, bool isVarArg, const SmallVectorImpl<ISD::InputArg> &Ins, - DebugLoc dl, SelectionDAG &DAG, + SDLoc dl, SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const; SDValue LowerGLOBALADDRESS(SDValue Op, SelectionDAG &DAG) const; SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const; @@ -115,14 +119,13 @@ namespace llvm { SDValue LowerCallResult(SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg, const SmallVectorImpl<ISD::InputArg> &Ins, - DebugLoc dl, SelectionDAG &DAG, + SDLoc dl, SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals, const SmallVectorImpl<SDValue> &OutVals, SDValue Callee) const; SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const; SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const; - SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG& DAG) const; SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG& DAG) const; SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const; @@ -130,7 +133,7 @@ namespace llvm { CallingConv::ID CallConv, bool isVarArg, const SmallVectorImpl<ISD::OutputArg> &Outs, const SmallVectorImpl<SDValue> &OutVals, - DebugLoc dl, SelectionDAG &DAG) const; + SDLoc dl, SelectionDAG &DAG) const; virtual MachineBasicBlock *EmitInstrWithCustomInserter(MachineInstr *MI, @@ -138,7 +141,7 @@ namespace llvm { SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const; SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const; - virtual EVT getSetCCResultType(EVT VT) const { + virtual EVT getSetCCResultType(LLVMContext &, EVT) const { return MVT::i1; } @@ -149,7 +152,7 @@ namespace llvm { std::pair<unsigned, const TargetRegisterClass*> getRegForInlineAsmConstraint(const std::string &Constraint, - EVT VT) const; + MVT VT) const; // Intrinsics virtual SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, diff --git a/lib/Target/Hexagon/HexagonInstrFormats.td b/lib/Target/Hexagon/HexagonInstrFormats.td index 587fa7d..e71386a 100644 --- a/lib/Target/Hexagon/HexagonInstrFormats.td +++ b/lib/Target/Hexagon/HexagonInstrFormats.td @@ -54,6 +54,7 @@ def AbsoluteSet : AddrModeType<2>; // Absolute set addressing mode def BaseImmOffset : AddrModeType<3>; // Indirect with offset def BaseLongOffset : AddrModeType<4>; // Indirect with long offset def BaseRegOffset : AddrModeType<5>; // Indirect with register offset +def PostInc : AddrModeType<6>; // Post increment addressing mode class MemAccessSize<bits<3> value> { bits<3> Value = value; @@ -157,6 +158,7 @@ class InstHexagon<dag outs, dag ins, string asmstr, list<dag> pattern, string CextOpcode = ""; string PredSense = ""; string PNewValue = ""; + string NValueST = ""; // Set to "true" for new-value stores. string InputType = ""; // Input is "imm" or "reg" type. string isMEMri = "false"; // Set to "true" for load/store with MEMri operand. string isFloat = "false"; // Set to "true" for the floating-point load/store. @@ -165,6 +167,7 @@ class InstHexagon<dag outs, dag ins, string asmstr, list<dag> pattern, let PredSense = !if(isPredicated, !if(isPredicatedFalse, "false", "true"), ""); let PNewValue = !if(isPredicatedNew, "new", ""); + let NValueST = !if(isNVStore, "true", "false"); // *** Must match MCTargetDesc/HexagonBaseInfo.h *** } diff --git a/lib/Target/Hexagon/HexagonInstrInfo.cpp b/lib/Target/Hexagon/HexagonInstrInfo.cpp index 96a252e..5af645c 100644 --- a/lib/Target/Hexagon/HexagonInstrInfo.cpp +++ b/lib/Target/Hexagon/HexagonInstrInfo.cpp @@ -23,7 +23,9 @@ #include "llvm/CodeGen/MachineMemOperand.h" #include "llvm/CodeGen/MachineRegisterInfo.h" #include "llvm/CodeGen/PseudoSourceValue.h" +#include "llvm/Support/Debug.h" #include "llvm/Support/MathExtras.h" +#include "llvm/Support/raw_ostream.h" #define GET_INSTRINFO_CTOR #define GET_INSTRMAP_INFO #include "HexagonGenInstrInfo.inc" @@ -56,7 +58,7 @@ const int Hexagon_MEMB_AUTOINC_MIN = -8; HexagonInstrInfo::HexagonInstrInfo(HexagonSubtarget &ST) : HexagonGenInstrInfo(Hexagon::ADJCALLSTACKDOWN, Hexagon::ADJCALLSTACKUP), - RI(ST, *this), Subtarget(ST) { + RI(ST), Subtarget(ST) { } @@ -118,16 +120,16 @@ HexagonInstrInfo::InsertBranch(MachineBasicBlock &MBB,MachineBasicBlock *TBB, DebugLoc DL) const{ int BOpc = Hexagon::JMP; - int BccOpc = Hexagon::JMP_c; + int BccOpc = Hexagon::JMP_t; assert(TBB && "InsertBranch must not be told to insert a fallthrough"); int regPos = 0; // Check if ReverseBranchCondition has asked to reverse this branch // If we want to reverse the branch an odd number of times, we want - // JMP_cNot. + // JMP_f. if (!Cond.empty() && Cond[0].isImm() && Cond[0].getImm() == 0) { - BccOpc = Hexagon::JMP_cNot; + BccOpc = Hexagon::JMP_f; regPos = 1; } @@ -174,8 +176,8 @@ bool HexagonInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB, FBB = NULL; // If the block has no terminators, it just falls into the block after it. - MachineBasicBlock::iterator I = MBB.end(); - if (I == MBB.begin()) + MachineBasicBlock::instr_iterator I = MBB.instr_end(); + if (I == MBB.instr_begin()) return false; // A basic block may looks like this: @@ -194,13 +196,24 @@ bool HexagonInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB, --I; if (I->isEHLabel()) return true; - } while (I != MBB.begin()); + } while (I != MBB.instr_begin()); - I = MBB.end(); + I = MBB.instr_end(); --I; while (I->isDebugValue()) { - if (I == MBB.begin()) + if (I == MBB.instr_begin()) + return false; + --I; + } + + // Delete the JMP if it's equivalent to a fall-through. + if (AllowModify && I->getOpcode() == Hexagon::JMP && + MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) { + DEBUG(dbgs()<< "\nErasing the jump to successor block\n";); + I->eraseFromParent(); + I = MBB.instr_end(); + if (I == MBB.instr_begin()) return false; --I; } @@ -209,23 +222,42 @@ bool HexagonInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB, // Get the last instruction in the block. MachineInstr *LastInst = I; + MachineInstr *SecondLastInst = NULL; + // Find one more terminator if present. + do { + if (&*I != LastInst && !I->isBundle() && isUnpredicatedTerminator(I)) { + if (!SecondLastInst) + SecondLastInst = I; + else + // This is a third branch. + return true; + } + if (I == MBB.instr_begin()) + break; + --I; + } while(I); + + int LastOpcode = LastInst->getOpcode(); + + bool LastOpcodeHasJMP_c = PredOpcodeHasJMP_c(LastOpcode); + bool LastOpcodeHasNot = PredOpcodeHasNot(LastOpcode); // If there is only one terminator instruction, process it. - if (I == MBB.begin() || !isUnpredicatedTerminator(--I)) { - if (LastInst->getOpcode() == Hexagon::JMP) { + if (LastInst && !SecondLastInst) { + if (LastOpcode == Hexagon::JMP) { TBB = LastInst->getOperand(0).getMBB(); return false; } - if (LastInst->getOpcode() == Hexagon::JMP_c) { - // Block ends with fall-through true condbranch. - TBB = LastInst->getOperand(1).getMBB(); + if (LastOpcode == Hexagon::ENDLOOP0) { + TBB = LastInst->getOperand(0).getMBB(); Cond.push_back(LastInst->getOperand(0)); return false; } - if (LastInst->getOpcode() == Hexagon::JMP_cNot) { - // Block ends with fall-through false condbranch. + if (LastOpcodeHasJMP_c) { TBB = LastInst->getOperand(1).getMBB(); - Cond.push_back(MachineOperand::CreateImm(0)); + if (LastOpcodeHasNot) { + Cond.push_back(MachineOperand::CreateImm(0)); + } Cond.push_back(LastInst->getOperand(0)); return false; } @@ -233,29 +265,14 @@ bool HexagonInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB, return true; } - // Get the instruction before it if it's a terminator. - MachineInstr *SecondLastInst = I; - - // If there are three terminators, we don't know what sort of block this is. - if (SecondLastInst && I != MBB.begin() && - isUnpredicatedTerminator(--I)) - return true; + int SecLastOpcode = SecondLastInst->getOpcode(); - // If the block ends with Hexagon::BRCOND and Hexagon:JMP, handle it. - if (((SecondLastInst->getOpcode() == Hexagon::BRCOND) || - (SecondLastInst->getOpcode() == Hexagon::JMP_c)) && - LastInst->getOpcode() == Hexagon::JMP) { + bool SecLastOpcodeHasJMP_c = PredOpcodeHasJMP_c(SecLastOpcode); + bool SecLastOpcodeHasNot = PredOpcodeHasNot(SecLastOpcode); + if (SecLastOpcodeHasJMP_c && (LastOpcode == Hexagon::JMP)) { TBB = SecondLastInst->getOperand(1).getMBB(); - Cond.push_back(SecondLastInst->getOperand(0)); - FBB = LastInst->getOperand(0).getMBB(); - return false; - } - - // If the block ends with Hexagon::JMP_cNot and Hexagon:JMP, handle it. - if ((SecondLastInst->getOpcode() == Hexagon::JMP_cNot) && - LastInst->getOpcode() == Hexagon::JMP) { - TBB = SecondLastInst->getOperand(1).getMBB(); - Cond.push_back(MachineOperand::CreateImm(0)); + if (SecLastOpcodeHasNot) + Cond.push_back(MachineOperand::CreateImm(0)); Cond.push_back(SecondLastInst->getOperand(0)); FBB = LastInst->getOperand(0).getMBB(); return false; @@ -263,8 +280,7 @@ bool HexagonInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB, // If the block ends with two Hexagon:JMPs, handle it. The second one is not // executed, so remove it. - if (SecondLastInst->getOpcode() == Hexagon::JMP && - LastInst->getOpcode() == Hexagon::JMP) { + if (SecLastOpcode == Hexagon::JMP && LastOpcode == Hexagon::JMP) { TBB = SecondLastInst->getOperand(0).getMBB(); I = LastInst; if (AllowModify) @@ -272,6 +288,15 @@ bool HexagonInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB, return false; } + // If the block ends with an ENDLOOP, and JMP, handle it. + if (SecLastOpcode == Hexagon::ENDLOOP0 && + LastOpcode == Hexagon::JMP) { + TBB = SecondLastInst->getOperand(0).getMBB(); + Cond.push_back(SecondLastInst->getOperand(0)); + FBB = LastInst->getOperand(0).getMBB(); + return false; + } + // Otherwise, can't handle this. return true; } @@ -279,8 +304,8 @@ bool HexagonInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB, unsigned HexagonInstrInfo::RemoveBranch(MachineBasicBlock &MBB) const { int BOpc = Hexagon::JMP; - int BccOpc = Hexagon::JMP_c; - int BccOpcNot = Hexagon::JMP_cNot; + int BccOpc = Hexagon::JMP_t; + int BccOpcNot = Hexagon::JMP_f; MachineBasicBlock::iterator I = MBB.end(); if (I == MBB.begin()) return 0; @@ -325,8 +350,6 @@ bool HexagonInstrInfo::analyzeCompare(const MachineInstr *MI, case Hexagon::CMPGTUrr: case Hexagon::CMPGTri: case Hexagon::CMPGTrr: - case Hexagon::CMPLTUrr: - case Hexagon::CMPLTrr: SrcReg = MI->getOperand(1).getReg(); Mask = ~0; break; @@ -366,8 +389,6 @@ bool HexagonInstrInfo::analyzeCompare(const MachineInstr *MI, case Hexagon::CMPhEQrr_xor_V4: case Hexagon::CMPhGTUrr_V4: case Hexagon::CMPhGTrr_shl_V4: - case Hexagon::CMPLTUrr: - case Hexagon::CMPLTrr: SrcReg2 = MI->getOperand(2).getReg(); return true; @@ -537,7 +558,6 @@ MachineInstr *HexagonInstrInfo::foldMemoryOperandImpl(MachineFunction &MF, return(0); } - unsigned HexagonInstrInfo::createVR(MachineFunction* MF, MVT VT) const { MachineRegisterInfo &RegInfo = MF->getRegInfo(); @@ -596,307 +616,8 @@ bool HexagonInstrInfo::isExtended(const MachineInstr *MI) const { return false; } -bool HexagonInstrInfo::isNewValueJump(const MachineInstr *MI) const { - switch (MI->getOpcode()) { - default: return false; - // JMP_EQri - case Hexagon::JMP_EQriPt_nv_V4: - case Hexagon::JMP_EQriPnt_nv_V4: - case Hexagon::JMP_EQriNotPt_nv_V4: - case Hexagon::JMP_EQriNotPnt_nv_V4: - case Hexagon::JMP_EQriPt_ie_nv_V4: - case Hexagon::JMP_EQriPnt_ie_nv_V4: - case Hexagon::JMP_EQriNotPt_ie_nv_V4: - case Hexagon::JMP_EQriNotPnt_ie_nv_V4: - - // JMP_EQri - with -1 - case Hexagon::JMP_EQriPtneg_nv_V4: - case Hexagon::JMP_EQriPntneg_nv_V4: - case Hexagon::JMP_EQriNotPtneg_nv_V4: - case Hexagon::JMP_EQriNotPntneg_nv_V4: - case Hexagon::JMP_EQriPtneg_ie_nv_V4: - case Hexagon::JMP_EQriPntneg_ie_nv_V4: - case Hexagon::JMP_EQriNotPtneg_ie_nv_V4: - case Hexagon::JMP_EQriNotPntneg_ie_nv_V4: - - // JMP_EQrr - case Hexagon::JMP_EQrrPt_nv_V4: - case Hexagon::JMP_EQrrPnt_nv_V4: - case Hexagon::JMP_EQrrNotPt_nv_V4: - case Hexagon::JMP_EQrrNotPnt_nv_V4: - case Hexagon::JMP_EQrrPt_ie_nv_V4: - case Hexagon::JMP_EQrrPnt_ie_nv_V4: - case Hexagon::JMP_EQrrNotPt_ie_nv_V4: - case Hexagon::JMP_EQrrNotPnt_ie_nv_V4: - - // JMP_GTri - case Hexagon::JMP_GTriPt_nv_V4: - case Hexagon::JMP_GTriPnt_nv_V4: - case Hexagon::JMP_GTriNotPt_nv_V4: - case Hexagon::JMP_GTriNotPnt_nv_V4: - case Hexagon::JMP_GTriPt_ie_nv_V4: - case Hexagon::JMP_GTriPnt_ie_nv_V4: - case Hexagon::JMP_GTriNotPt_ie_nv_V4: - case Hexagon::JMP_GTriNotPnt_ie_nv_V4: - - // JMP_GTri - with -1 - case Hexagon::JMP_GTriPtneg_nv_V4: - case Hexagon::JMP_GTriPntneg_nv_V4: - case Hexagon::JMP_GTriNotPtneg_nv_V4: - case Hexagon::JMP_GTriNotPntneg_nv_V4: - case Hexagon::JMP_GTriPtneg_ie_nv_V4: - case Hexagon::JMP_GTriPntneg_ie_nv_V4: - case Hexagon::JMP_GTriNotPtneg_ie_nv_V4: - case Hexagon::JMP_GTriNotPntneg_ie_nv_V4: - - // JMP_GTrr - case Hexagon::JMP_GTrrPt_nv_V4: - case Hexagon::JMP_GTrrPnt_nv_V4: - case Hexagon::JMP_GTrrNotPt_nv_V4: - case Hexagon::JMP_GTrrNotPnt_nv_V4: - case Hexagon::JMP_GTrrPt_ie_nv_V4: - case Hexagon::JMP_GTrrPnt_ie_nv_V4: - case Hexagon::JMP_GTrrNotPt_ie_nv_V4: - case Hexagon::JMP_GTrrNotPnt_ie_nv_V4: - - // JMP_GTrrdn - case Hexagon::JMP_GTrrdnPt_nv_V4: - case Hexagon::JMP_GTrrdnPnt_nv_V4: - case Hexagon::JMP_GTrrdnNotPt_nv_V4: - case Hexagon::JMP_GTrrdnNotPnt_nv_V4: - case Hexagon::JMP_GTrrdnPt_ie_nv_V4: - case Hexagon::JMP_GTrrdnPnt_ie_nv_V4: - case Hexagon::JMP_GTrrdnNotPt_ie_nv_V4: - case Hexagon::JMP_GTrrdnNotPnt_ie_nv_V4: - - // JMP_GTUri - case Hexagon::JMP_GTUriPt_nv_V4: - case Hexagon::JMP_GTUriPnt_nv_V4: - case Hexagon::JMP_GTUriNotPt_nv_V4: - case Hexagon::JMP_GTUriNotPnt_nv_V4: - case Hexagon::JMP_GTUriPt_ie_nv_V4: - case Hexagon::JMP_GTUriPnt_ie_nv_V4: - case Hexagon::JMP_GTUriNotPt_ie_nv_V4: - case Hexagon::JMP_GTUriNotPnt_ie_nv_V4: - - // JMP_GTUrr - case Hexagon::JMP_GTUrrPt_nv_V4: - case Hexagon::JMP_GTUrrPnt_nv_V4: - case Hexagon::JMP_GTUrrNotPt_nv_V4: - case Hexagon::JMP_GTUrrNotPnt_nv_V4: - case Hexagon::JMP_GTUrrPt_ie_nv_V4: - case Hexagon::JMP_GTUrrPnt_ie_nv_V4: - case Hexagon::JMP_GTUrrNotPt_ie_nv_V4: - case Hexagon::JMP_GTUrrNotPnt_ie_nv_V4: - - // JMP_GTUrrdn - case Hexagon::JMP_GTUrrdnPt_nv_V4: - case Hexagon::JMP_GTUrrdnPnt_nv_V4: - case Hexagon::JMP_GTUrrdnNotPt_nv_V4: - case Hexagon::JMP_GTUrrdnNotPnt_nv_V4: - case Hexagon::JMP_GTUrrdnPt_ie_nv_V4: - case Hexagon::JMP_GTUrrdnPnt_ie_nv_V4: - case Hexagon::JMP_GTUrrdnNotPt_ie_nv_V4: - case Hexagon::JMP_GTUrrdnNotPnt_ie_nv_V4: - return true; - } -} - -bool HexagonInstrInfo::isNewValueStore(const MachineInstr *MI) const { - switch (MI->getOpcode()) { - default: return false; - // Store Byte - case Hexagon::STrib_nv_V4: - case Hexagon::STrib_indexed_nv_V4: - case Hexagon::STrib_indexed_shl_nv_V4: - case Hexagon::STrib_shl_nv_V4: - case Hexagon::STb_GP_nv_V4: - case Hexagon::POST_STbri_nv_V4: - case Hexagon::STrib_cPt_nv_V4: - case Hexagon::STrib_cdnPt_nv_V4: - case Hexagon::STrib_cNotPt_nv_V4: - case Hexagon::STrib_cdnNotPt_nv_V4: - case Hexagon::STrib_indexed_cPt_nv_V4: - case Hexagon::STrib_indexed_cdnPt_nv_V4: - case Hexagon::STrib_indexed_cNotPt_nv_V4: - case Hexagon::STrib_indexed_cdnNotPt_nv_V4: - case Hexagon::STrib_indexed_shl_cPt_nv_V4: - case Hexagon::STrib_indexed_shl_cdnPt_nv_V4: - case Hexagon::STrib_indexed_shl_cNotPt_nv_V4: - case Hexagon::STrib_indexed_shl_cdnNotPt_nv_V4: - case Hexagon::POST_STbri_cPt_nv_V4: - case Hexagon::POST_STbri_cdnPt_nv_V4: - case Hexagon::POST_STbri_cNotPt_nv_V4: - case Hexagon::POST_STbri_cdnNotPt_nv_V4: - case Hexagon::STb_GP_cPt_nv_V4: - case Hexagon::STb_GP_cNotPt_nv_V4: - case Hexagon::STb_GP_cdnPt_nv_V4: - case Hexagon::STb_GP_cdnNotPt_nv_V4: - case Hexagon::STrib_abs_nv_V4: - case Hexagon::STrib_abs_cPt_nv_V4: - case Hexagon::STrib_abs_cdnPt_nv_V4: - case Hexagon::STrib_abs_cNotPt_nv_V4: - case Hexagon::STrib_abs_cdnNotPt_nv_V4: - case Hexagon::STrib_imm_abs_nv_V4: - case Hexagon::STrib_imm_abs_cPt_nv_V4: - case Hexagon::STrib_imm_abs_cdnPt_nv_V4: - case Hexagon::STrib_imm_abs_cNotPt_nv_V4: - case Hexagon::STrib_imm_abs_cdnNotPt_nv_V4: - - // Store Halfword - case Hexagon::STrih_nv_V4: - case Hexagon::STrih_indexed_nv_V4: - case Hexagon::STrih_indexed_shl_nv_V4: - case Hexagon::STrih_shl_nv_V4: - case Hexagon::STh_GP_nv_V4: - case Hexagon::POST_SThri_nv_V4: - case Hexagon::STrih_cPt_nv_V4: - case Hexagon::STrih_cdnPt_nv_V4: - case Hexagon::STrih_cNotPt_nv_V4: - case Hexagon::STrih_cdnNotPt_nv_V4: - case Hexagon::STrih_indexed_cPt_nv_V4: - case Hexagon::STrih_indexed_cdnPt_nv_V4: - case Hexagon::STrih_indexed_cNotPt_nv_V4: - case Hexagon::STrih_indexed_cdnNotPt_nv_V4: - case Hexagon::STrih_indexed_shl_cPt_nv_V4: - case Hexagon::STrih_indexed_shl_cdnPt_nv_V4: - case Hexagon::STrih_indexed_shl_cNotPt_nv_V4: - case Hexagon::STrih_indexed_shl_cdnNotPt_nv_V4: - case Hexagon::POST_SThri_cPt_nv_V4: - case Hexagon::POST_SThri_cdnPt_nv_V4: - case Hexagon::POST_SThri_cNotPt_nv_V4: - case Hexagon::POST_SThri_cdnNotPt_nv_V4: - case Hexagon::STh_GP_cPt_nv_V4: - case Hexagon::STh_GP_cNotPt_nv_V4: - case Hexagon::STh_GP_cdnPt_nv_V4: - case Hexagon::STh_GP_cdnNotPt_nv_V4: - case Hexagon::STrih_abs_nv_V4: - case Hexagon::STrih_abs_cPt_nv_V4: - case Hexagon::STrih_abs_cdnPt_nv_V4: - case Hexagon::STrih_abs_cNotPt_nv_V4: - case Hexagon::STrih_abs_cdnNotPt_nv_V4: - case Hexagon::STrih_imm_abs_nv_V4: - case Hexagon::STrih_imm_abs_cPt_nv_V4: - case Hexagon::STrih_imm_abs_cdnPt_nv_V4: - case Hexagon::STrih_imm_abs_cNotPt_nv_V4: - case Hexagon::STrih_imm_abs_cdnNotPt_nv_V4: - - // Store Word - case Hexagon::STriw_nv_V4: - case Hexagon::STriw_indexed_nv_V4: - case Hexagon::STriw_indexed_shl_nv_V4: - case Hexagon::STriw_shl_nv_V4: - case Hexagon::STw_GP_nv_V4: - case Hexagon::POST_STwri_nv_V4: - case Hexagon::STriw_cPt_nv_V4: - case Hexagon::STriw_cdnPt_nv_V4: - case Hexagon::STriw_cNotPt_nv_V4: - case Hexagon::STriw_cdnNotPt_nv_V4: - case Hexagon::STriw_indexed_cPt_nv_V4: - case Hexagon::STriw_indexed_cdnPt_nv_V4: - case Hexagon::STriw_indexed_cNotPt_nv_V4: - case Hexagon::STriw_indexed_cdnNotPt_nv_V4: - case Hexagon::STriw_indexed_shl_cPt_nv_V4: - case Hexagon::STriw_indexed_shl_cdnPt_nv_V4: - case Hexagon::STriw_indexed_shl_cNotPt_nv_V4: - case Hexagon::STriw_indexed_shl_cdnNotPt_nv_V4: - case Hexagon::POST_STwri_cPt_nv_V4: - case Hexagon::POST_STwri_cdnPt_nv_V4: - case Hexagon::POST_STwri_cNotPt_nv_V4: - case Hexagon::POST_STwri_cdnNotPt_nv_V4: - case Hexagon::STw_GP_cPt_nv_V4: - case Hexagon::STw_GP_cNotPt_nv_V4: - case Hexagon::STw_GP_cdnPt_nv_V4: - case Hexagon::STw_GP_cdnNotPt_nv_V4: - case Hexagon::STriw_abs_nv_V4: - case Hexagon::STriw_abs_cPt_nv_V4: - case Hexagon::STriw_abs_cdnPt_nv_V4: - case Hexagon::STriw_abs_cNotPt_nv_V4: - case Hexagon::STriw_abs_cdnNotPt_nv_V4: - case Hexagon::STriw_imm_abs_nv_V4: - case Hexagon::STriw_imm_abs_cPt_nv_V4: - case Hexagon::STriw_imm_abs_cdnPt_nv_V4: - case Hexagon::STriw_imm_abs_cNotPt_nv_V4: - case Hexagon::STriw_imm_abs_cdnNotPt_nv_V4: - return true; - } -} - -bool HexagonInstrInfo::isPostIncrement (const MachineInstr* MI) const { - switch (MI->getOpcode()) - { - default: return false; - // Load Byte - case Hexagon::POST_LDrib: - case Hexagon::POST_LDrib_cPt: - case Hexagon::POST_LDrib_cNotPt: - case Hexagon::POST_LDrib_cdnPt_V4: - case Hexagon::POST_LDrib_cdnNotPt_V4: - - // Load unsigned byte - case Hexagon::POST_LDriub: - case Hexagon::POST_LDriub_cPt: - case Hexagon::POST_LDriub_cNotPt: - case Hexagon::POST_LDriub_cdnPt_V4: - case Hexagon::POST_LDriub_cdnNotPt_V4: - - // Load halfword - case Hexagon::POST_LDrih: - case Hexagon::POST_LDrih_cPt: - case Hexagon::POST_LDrih_cNotPt: - case Hexagon::POST_LDrih_cdnPt_V4: - case Hexagon::POST_LDrih_cdnNotPt_V4: - - // Load unsigned halfword - case Hexagon::POST_LDriuh: - case Hexagon::POST_LDriuh_cPt: - case Hexagon::POST_LDriuh_cNotPt: - case Hexagon::POST_LDriuh_cdnPt_V4: - case Hexagon::POST_LDriuh_cdnNotPt_V4: - - // Load word - case Hexagon::POST_LDriw: - case Hexagon::POST_LDriw_cPt: - case Hexagon::POST_LDriw_cNotPt: - case Hexagon::POST_LDriw_cdnPt_V4: - case Hexagon::POST_LDriw_cdnNotPt_V4: - - // Load double word - case Hexagon::POST_LDrid: - case Hexagon::POST_LDrid_cPt: - case Hexagon::POST_LDrid_cNotPt: - case Hexagon::POST_LDrid_cdnPt_V4: - case Hexagon::POST_LDrid_cdnNotPt_V4: - - // Store byte - case Hexagon::POST_STbri: - case Hexagon::POST_STbri_cPt: - case Hexagon::POST_STbri_cNotPt: - case Hexagon::POST_STbri_cdnPt_V4: - case Hexagon::POST_STbri_cdnNotPt_V4: - - // Store halfword - case Hexagon::POST_SThri: - case Hexagon::POST_SThri_cPt: - case Hexagon::POST_SThri_cNotPt: - case Hexagon::POST_SThri_cdnPt_V4: - case Hexagon::POST_SThri_cdnNotPt_V4: - - // Store word - case Hexagon::POST_STwri: - case Hexagon::POST_STwri_cPt: - case Hexagon::POST_STwri_cNotPt: - case Hexagon::POST_STwri_cdnPt_V4: - case Hexagon::POST_STwri_cdnNotPt_V4: - - // Store double word - case Hexagon::POST_STdri: - case Hexagon::POST_STdri_cPt: - case Hexagon::POST_STdri_cNotPt: - case Hexagon::POST_STdri_cdnPt_V4: - case Hexagon::POST_STdri_cdnNotPt_V4: - return true; - } +bool HexagonInstrInfo::isBranch (const MachineInstr *MI) const { + return MI->getDesc().isBranch(); } bool HexagonInstrInfo::isNewValueInst(const MachineInstr *MI) const { @@ -994,9 +715,6 @@ bool HexagonInstrInfo::isPredicable(MachineInstr *MI) const { case Hexagon::ZXTB: case Hexagon::ZXTH: return Subtarget.hasV4TOps(); - - case Hexagon::JMPR: - return false; } return true; @@ -1007,466 +725,40 @@ bool HexagonInstrInfo::isPredicable(MachineInstr *MI) const { // cPt ---> cNotPt // cNotPt ---> cPt // -// however, these inversiones are NOT included: -// -// cdnPt -X-> cdnNotPt -// cdnNotPt -X-> cdnPt -// cPt_nv -X-> cNotPt_nv (new value stores) -// cNotPt_nv -X-> cPt_nv (new value stores) -// -// because only the following transformations are allowed: -// -// cNotPt ---> cdnNotPt -// cPt ---> cdnPt -// cNotPt ---> cNotPt_nv -// cPt ---> cPt_nv unsigned HexagonInstrInfo::getInvertedPredicatedOpcode(const int Opc) const { + int InvPredOpcode; + InvPredOpcode = isPredicatedTrue(Opc) ? Hexagon::getFalsePredOpcode(Opc) + : Hexagon::getTruePredOpcode(Opc); + if (InvPredOpcode >= 0) // Valid instruction with the inverted predicate. + return InvPredOpcode; + switch(Opc) { default: llvm_unreachable("Unexpected predicated instruction"); - case Hexagon::TFR_cPt: - return Hexagon::TFR_cNotPt; - case Hexagon::TFR_cNotPt: - return Hexagon::TFR_cPt; - - case Hexagon::TFRI_cPt: - return Hexagon::TFRI_cNotPt; - case Hexagon::TFRI_cNotPt: - return Hexagon::TFRI_cPt; - - case Hexagon::JMP_c: - return Hexagon::JMP_cNot; - case Hexagon::JMP_cNot: - return Hexagon::JMP_c; - - case Hexagon::ADD_ri_cPt: - return Hexagon::ADD_ri_cNotPt; - case Hexagon::ADD_ri_cNotPt: - return Hexagon::ADD_ri_cPt; - - case Hexagon::ADD_rr_cPt: - return Hexagon::ADD_rr_cNotPt; - case Hexagon::ADD_rr_cNotPt: - return Hexagon::ADD_rr_cPt; - - case Hexagon::XOR_rr_cPt: - return Hexagon::XOR_rr_cNotPt; - case Hexagon::XOR_rr_cNotPt: - return Hexagon::XOR_rr_cPt; - - case Hexagon::AND_rr_cPt: - return Hexagon::AND_rr_cNotPt; - case Hexagon::AND_rr_cNotPt: - return Hexagon::AND_rr_cPt; - - case Hexagon::OR_rr_cPt: - return Hexagon::OR_rr_cNotPt; - case Hexagon::OR_rr_cNotPt: - return Hexagon::OR_rr_cPt; - - case Hexagon::SUB_rr_cPt: - return Hexagon::SUB_rr_cNotPt; - case Hexagon::SUB_rr_cNotPt: - return Hexagon::SUB_rr_cPt; - case Hexagon::COMBINE_rr_cPt: return Hexagon::COMBINE_rr_cNotPt; case Hexagon::COMBINE_rr_cNotPt: return Hexagon::COMBINE_rr_cPt; - case Hexagon::ASLH_cPt_V4: - return Hexagon::ASLH_cNotPt_V4; - case Hexagon::ASLH_cNotPt_V4: - return Hexagon::ASLH_cPt_V4; - - case Hexagon::ASRH_cPt_V4: - return Hexagon::ASRH_cNotPt_V4; - case Hexagon::ASRH_cNotPt_V4: - return Hexagon::ASRH_cPt_V4; - - case Hexagon::SXTB_cPt_V4: - return Hexagon::SXTB_cNotPt_V4; - case Hexagon::SXTB_cNotPt_V4: - return Hexagon::SXTB_cPt_V4; - - case Hexagon::SXTH_cPt_V4: - return Hexagon::SXTH_cNotPt_V4; - case Hexagon::SXTH_cNotPt_V4: - return Hexagon::SXTH_cPt_V4; - - case Hexagon::ZXTB_cPt_V4: - return Hexagon::ZXTB_cNotPt_V4; - case Hexagon::ZXTB_cNotPt_V4: - return Hexagon::ZXTB_cPt_V4; - - case Hexagon::ZXTH_cPt_V4: - return Hexagon::ZXTH_cNotPt_V4; - case Hexagon::ZXTH_cNotPt_V4: - return Hexagon::ZXTH_cPt_V4; - - - case Hexagon::JMPR_cPt: - return Hexagon::JMPR_cNotPt; - case Hexagon::JMPR_cNotPt: - return Hexagon::JMPR_cPt; - - // V4 indexed+scaled load. - case Hexagon::LDrid_indexed_shl_cPt_V4: - return Hexagon::LDrid_indexed_shl_cNotPt_V4; - case Hexagon::LDrid_indexed_shl_cNotPt_V4: - return Hexagon::LDrid_indexed_shl_cPt_V4; - - case Hexagon::LDrib_indexed_shl_cPt_V4: - return Hexagon::LDrib_indexed_shl_cNotPt_V4; - case Hexagon::LDrib_indexed_shl_cNotPt_V4: - return Hexagon::LDrib_indexed_shl_cPt_V4; - - case Hexagon::LDriub_indexed_shl_cPt_V4: - return Hexagon::LDriub_indexed_shl_cNotPt_V4; - case Hexagon::LDriub_indexed_shl_cNotPt_V4: - return Hexagon::LDriub_indexed_shl_cPt_V4; - - case Hexagon::LDrih_indexed_shl_cPt_V4: - return Hexagon::LDrih_indexed_shl_cNotPt_V4; - case Hexagon::LDrih_indexed_shl_cNotPt_V4: - return Hexagon::LDrih_indexed_shl_cPt_V4; - - case Hexagon::LDriuh_indexed_shl_cPt_V4: - return Hexagon::LDriuh_indexed_shl_cNotPt_V4; - case Hexagon::LDriuh_indexed_shl_cNotPt_V4: - return Hexagon::LDriuh_indexed_shl_cPt_V4; - - case Hexagon::LDriw_indexed_shl_cPt_V4: - return Hexagon::LDriw_indexed_shl_cNotPt_V4; - case Hexagon::LDriw_indexed_shl_cNotPt_V4: - return Hexagon::LDriw_indexed_shl_cPt_V4; - - // Byte. - case Hexagon::POST_STbri_cPt: - return Hexagon::POST_STbri_cNotPt; - case Hexagon::POST_STbri_cNotPt: - return Hexagon::POST_STbri_cPt; - - case Hexagon::STrib_cPt: - return Hexagon::STrib_cNotPt; - case Hexagon::STrib_cNotPt: - return Hexagon::STrib_cPt; - - case Hexagon::STrib_indexed_cPt: - return Hexagon::STrib_indexed_cNotPt; - case Hexagon::STrib_indexed_cNotPt: - return Hexagon::STrib_indexed_cPt; - - case Hexagon::STrib_imm_cPt_V4: - return Hexagon::STrib_imm_cNotPt_V4; - case Hexagon::STrib_imm_cNotPt_V4: - return Hexagon::STrib_imm_cPt_V4; - - case Hexagon::STrib_indexed_shl_cPt_V4: - return Hexagon::STrib_indexed_shl_cNotPt_V4; - case Hexagon::STrib_indexed_shl_cNotPt_V4: - return Hexagon::STrib_indexed_shl_cPt_V4; - - // Halfword. - case Hexagon::POST_SThri_cPt: - return Hexagon::POST_SThri_cNotPt; - case Hexagon::POST_SThri_cNotPt: - return Hexagon::POST_SThri_cPt; - - case Hexagon::STrih_cPt: - return Hexagon::STrih_cNotPt; - case Hexagon::STrih_cNotPt: - return Hexagon::STrih_cPt; - - case Hexagon::STrih_indexed_cPt: - return Hexagon::STrih_indexed_cNotPt; - case Hexagon::STrih_indexed_cNotPt: - return Hexagon::STrih_indexed_cPt; - - case Hexagon::STrih_imm_cPt_V4: - return Hexagon::STrih_imm_cNotPt_V4; - case Hexagon::STrih_imm_cNotPt_V4: - return Hexagon::STrih_imm_cPt_V4; - - case Hexagon::STrih_indexed_shl_cPt_V4: - return Hexagon::STrih_indexed_shl_cNotPt_V4; - case Hexagon::STrih_indexed_shl_cNotPt_V4: - return Hexagon::STrih_indexed_shl_cPt_V4; - - // Word. - case Hexagon::POST_STwri_cPt: - return Hexagon::POST_STwri_cNotPt; - case Hexagon::POST_STwri_cNotPt: - return Hexagon::POST_STwri_cPt; - - case Hexagon::STriw_cPt: - return Hexagon::STriw_cNotPt; - case Hexagon::STriw_cNotPt: - return Hexagon::STriw_cPt; - - case Hexagon::STriw_indexed_cPt: - return Hexagon::STriw_indexed_cNotPt; - case Hexagon::STriw_indexed_cNotPt: - return Hexagon::STriw_indexed_cPt; - - case Hexagon::STriw_indexed_shl_cPt_V4: - return Hexagon::STriw_indexed_shl_cNotPt_V4; - case Hexagon::STriw_indexed_shl_cNotPt_V4: - return Hexagon::STriw_indexed_shl_cPt_V4; - - case Hexagon::STriw_imm_cPt_V4: - return Hexagon::STriw_imm_cNotPt_V4; - case Hexagon::STriw_imm_cNotPt_V4: - return Hexagon::STriw_imm_cPt_V4; - - // Double word. - case Hexagon::POST_STdri_cPt: - return Hexagon::POST_STdri_cNotPt; - case Hexagon::POST_STdri_cNotPt: - return Hexagon::POST_STdri_cPt; - - case Hexagon::STrid_cPt: - return Hexagon::STrid_cNotPt; - case Hexagon::STrid_cNotPt: - return Hexagon::STrid_cPt; - - case Hexagon::STrid_indexed_cPt: - return Hexagon::STrid_indexed_cNotPt; - case Hexagon::STrid_indexed_cNotPt: - return Hexagon::STrid_indexed_cPt; - - case Hexagon::STrid_indexed_shl_cPt_V4: - return Hexagon::STrid_indexed_shl_cNotPt_V4; - case Hexagon::STrid_indexed_shl_cNotPt_V4: - return Hexagon::STrid_indexed_shl_cPt_V4; - - // V4 Store to global address. - case Hexagon::STd_GP_cPt_V4: - return Hexagon::STd_GP_cNotPt_V4; - case Hexagon::STd_GP_cNotPt_V4: - return Hexagon::STd_GP_cPt_V4; - - case Hexagon::STb_GP_cPt_V4: - return Hexagon::STb_GP_cNotPt_V4; - case Hexagon::STb_GP_cNotPt_V4: - return Hexagon::STb_GP_cPt_V4; - - case Hexagon::STh_GP_cPt_V4: - return Hexagon::STh_GP_cNotPt_V4; - case Hexagon::STh_GP_cNotPt_V4: - return Hexagon::STh_GP_cPt_V4; - - case Hexagon::STw_GP_cPt_V4: - return Hexagon::STw_GP_cNotPt_V4; - case Hexagon::STw_GP_cNotPt_V4: - return Hexagon::STw_GP_cPt_V4; - - // Load. - case Hexagon::LDrid_cPt: - return Hexagon::LDrid_cNotPt; - case Hexagon::LDrid_cNotPt: - return Hexagon::LDrid_cPt; - - case Hexagon::LDriw_cPt: - return Hexagon::LDriw_cNotPt; - case Hexagon::LDriw_cNotPt: - return Hexagon::LDriw_cPt; - - case Hexagon::LDrih_cPt: - return Hexagon::LDrih_cNotPt; - case Hexagon::LDrih_cNotPt: - return Hexagon::LDrih_cPt; - - case Hexagon::LDriuh_cPt: - return Hexagon::LDriuh_cNotPt; - case Hexagon::LDriuh_cNotPt: - return Hexagon::LDriuh_cPt; - - case Hexagon::LDrib_cPt: - return Hexagon::LDrib_cNotPt; - case Hexagon::LDrib_cNotPt: - return Hexagon::LDrib_cPt; - - case Hexagon::LDriub_cPt: - return Hexagon::LDriub_cNotPt; - case Hexagon::LDriub_cNotPt: - return Hexagon::LDriub_cPt; - - // Load Indexed. - case Hexagon::LDrid_indexed_cPt: - return Hexagon::LDrid_indexed_cNotPt; - case Hexagon::LDrid_indexed_cNotPt: - return Hexagon::LDrid_indexed_cPt; - - case Hexagon::LDriw_indexed_cPt: - return Hexagon::LDriw_indexed_cNotPt; - case Hexagon::LDriw_indexed_cNotPt: - return Hexagon::LDriw_indexed_cPt; - - case Hexagon::LDrih_indexed_cPt: - return Hexagon::LDrih_indexed_cNotPt; - case Hexagon::LDrih_indexed_cNotPt: - return Hexagon::LDrih_indexed_cPt; - - case Hexagon::LDriuh_indexed_cPt: - return Hexagon::LDriuh_indexed_cNotPt; - case Hexagon::LDriuh_indexed_cNotPt: - return Hexagon::LDriuh_indexed_cPt; - - case Hexagon::LDrib_indexed_cPt: - return Hexagon::LDrib_indexed_cNotPt; - case Hexagon::LDrib_indexed_cNotPt: - return Hexagon::LDrib_indexed_cPt; - - case Hexagon::LDriub_indexed_cPt: - return Hexagon::LDriub_indexed_cNotPt; - case Hexagon::LDriub_indexed_cNotPt: - return Hexagon::LDriub_indexed_cPt; - - // Post Inc Load. - case Hexagon::POST_LDrid_cPt: - return Hexagon::POST_LDrid_cNotPt; - case Hexagon::POST_LDriw_cNotPt: - return Hexagon::POST_LDriw_cPt; - - case Hexagon::POST_LDrih_cPt: - return Hexagon::POST_LDrih_cNotPt; - case Hexagon::POST_LDrih_cNotPt: - return Hexagon::POST_LDrih_cPt; - - case Hexagon::POST_LDriuh_cPt: - return Hexagon::POST_LDriuh_cNotPt; - case Hexagon::POST_LDriuh_cNotPt: - return Hexagon::POST_LDriuh_cPt; - - case Hexagon::POST_LDrib_cPt: - return Hexagon::POST_LDrib_cNotPt; - case Hexagon::POST_LDrib_cNotPt: - return Hexagon::POST_LDrib_cPt; - - case Hexagon::POST_LDriub_cPt: - return Hexagon::POST_LDriub_cNotPt; - case Hexagon::POST_LDriub_cNotPt: - return Hexagon::POST_LDriub_cPt; - - // Dealloc_return. + // Dealloc_return. case Hexagon::DEALLOC_RET_cPt_V4: return Hexagon::DEALLOC_RET_cNotPt_V4; case Hexagon::DEALLOC_RET_cNotPt_V4: return Hexagon::DEALLOC_RET_cPt_V4; - - // New Value Jump. - // JMPEQ_ri - with -1. - case Hexagon::JMP_EQriPtneg_nv_V4: - return Hexagon::JMP_EQriNotPtneg_nv_V4; - case Hexagon::JMP_EQriNotPtneg_nv_V4: - return Hexagon::JMP_EQriPtneg_nv_V4; - - case Hexagon::JMP_EQriPntneg_nv_V4: - return Hexagon::JMP_EQriNotPntneg_nv_V4; - case Hexagon::JMP_EQriNotPntneg_nv_V4: - return Hexagon::JMP_EQriPntneg_nv_V4; - - // JMPEQ_ri. - case Hexagon::JMP_EQriPt_nv_V4: - return Hexagon::JMP_EQriNotPt_nv_V4; - case Hexagon::JMP_EQriNotPt_nv_V4: - return Hexagon::JMP_EQriPt_nv_V4; - - case Hexagon::JMP_EQriPnt_nv_V4: - return Hexagon::JMP_EQriNotPnt_nv_V4; - case Hexagon::JMP_EQriNotPnt_nv_V4: - return Hexagon::JMP_EQriPnt_nv_V4; - - // JMPEQ_rr. - case Hexagon::JMP_EQrrPt_nv_V4: - return Hexagon::JMP_EQrrNotPt_nv_V4; - case Hexagon::JMP_EQrrNotPt_nv_V4: - return Hexagon::JMP_EQrrPt_nv_V4; - - case Hexagon::JMP_EQrrPnt_nv_V4: - return Hexagon::JMP_EQrrNotPnt_nv_V4; - case Hexagon::JMP_EQrrNotPnt_nv_V4: - return Hexagon::JMP_EQrrPnt_nv_V4; - - // JMPGT_ri - with -1. - case Hexagon::JMP_GTriPtneg_nv_V4: - return Hexagon::JMP_GTriNotPtneg_nv_V4; - case Hexagon::JMP_GTriNotPtneg_nv_V4: - return Hexagon::JMP_GTriPtneg_nv_V4; - - case Hexagon::JMP_GTriPntneg_nv_V4: - return Hexagon::JMP_GTriNotPntneg_nv_V4; - case Hexagon::JMP_GTriNotPntneg_nv_V4: - return Hexagon::JMP_GTriPntneg_nv_V4; - - // JMPGT_ri. - case Hexagon::JMP_GTriPt_nv_V4: - return Hexagon::JMP_GTriNotPt_nv_V4; - case Hexagon::JMP_GTriNotPt_nv_V4: - return Hexagon::JMP_GTriPt_nv_V4; - - case Hexagon::JMP_GTriPnt_nv_V4: - return Hexagon::JMP_GTriNotPnt_nv_V4; - case Hexagon::JMP_GTriNotPnt_nv_V4: - return Hexagon::JMP_GTriPnt_nv_V4; - - // JMPGT_rr. - case Hexagon::JMP_GTrrPt_nv_V4: - return Hexagon::JMP_GTrrNotPt_nv_V4; - case Hexagon::JMP_GTrrNotPt_nv_V4: - return Hexagon::JMP_GTrrPt_nv_V4; - - case Hexagon::JMP_GTrrPnt_nv_V4: - return Hexagon::JMP_GTrrNotPnt_nv_V4; - case Hexagon::JMP_GTrrNotPnt_nv_V4: - return Hexagon::JMP_GTrrPnt_nv_V4; - - // JMPGT_rrdn. - case Hexagon::JMP_GTrrdnPt_nv_V4: - return Hexagon::JMP_GTrrdnNotPt_nv_V4; - case Hexagon::JMP_GTrrdnNotPt_nv_V4: - return Hexagon::JMP_GTrrdnPt_nv_V4; - - case Hexagon::JMP_GTrrdnPnt_nv_V4: - return Hexagon::JMP_GTrrdnNotPnt_nv_V4; - case Hexagon::JMP_GTrrdnNotPnt_nv_V4: - return Hexagon::JMP_GTrrdnPnt_nv_V4; - - // JMPGTU_ri. - case Hexagon::JMP_GTUriPt_nv_V4: - return Hexagon::JMP_GTUriNotPt_nv_V4; - case Hexagon::JMP_GTUriNotPt_nv_V4: - return Hexagon::JMP_GTUriPt_nv_V4; - - case Hexagon::JMP_GTUriPnt_nv_V4: - return Hexagon::JMP_GTUriNotPnt_nv_V4; - case Hexagon::JMP_GTUriNotPnt_nv_V4: - return Hexagon::JMP_GTUriPnt_nv_V4; - - // JMPGTU_rr. - case Hexagon::JMP_GTUrrPt_nv_V4: - return Hexagon::JMP_GTUrrNotPt_nv_V4; - case Hexagon::JMP_GTUrrNotPt_nv_V4: - return Hexagon::JMP_GTUrrPt_nv_V4; - - case Hexagon::JMP_GTUrrPnt_nv_V4: - return Hexagon::JMP_GTUrrNotPnt_nv_V4; - case Hexagon::JMP_GTUrrNotPnt_nv_V4: - return Hexagon::JMP_GTUrrPnt_nv_V4; - - // JMPGTU_rrdn. - case Hexagon::JMP_GTUrrdnPt_nv_V4: - return Hexagon::JMP_GTUrrdnNotPt_nv_V4; - case Hexagon::JMP_GTUrrdnNotPt_nv_V4: - return Hexagon::JMP_GTUrrdnPt_nv_V4; - - case Hexagon::JMP_GTUrrdnPnt_nv_V4: - return Hexagon::JMP_GTUrrdnNotPnt_nv_V4; - case Hexagon::JMP_GTUrrdnNotPnt_nv_V4: - return Hexagon::JMP_GTUrrdnPnt_nv_V4; } } +// New Value Store instructions. +bool HexagonInstrInfo::isNewValueStore(const MachineInstr *MI) const { + const uint64_t F = MI->getDesc().TSFlags; + + return ((F >> HexagonII::NVStorePos) & HexagonII::NVStoreMask); +} + +bool HexagonInstrInfo::isNewValueStore(unsigned Opcode) const { + const uint64_t F = get(Opcode).TSFlags; + + return ((F >> HexagonII::NVStorePos) & HexagonII::NVStoreMask); +} int HexagonInstrInfo:: getMatchingCondBranchOpcode(int Opc, bool invertPredicate) const { @@ -1480,223 +772,21 @@ getMatchingCondBranchOpcode(int Opc, bool invertPredicate) const { // This switch case will be removed once all the instructions have been // modified to use relation maps. switch(Opc) { - case Hexagon::TFR: - return !invertPredicate ? Hexagon::TFR_cPt : - Hexagon::TFR_cNotPt; case Hexagon::TFRI_f: return !invertPredicate ? Hexagon::TFRI_cPt_f : Hexagon::TFRI_cNotPt_f; - case Hexagon::TFRI: - return !invertPredicate ? Hexagon::TFRI_cPt : - Hexagon::TFRI_cNotPt; - case Hexagon::JMP: - return !invertPredicate ? Hexagon::JMP_c : - Hexagon::JMP_cNot; - case Hexagon::JMP_EQrrPt_nv_V4: - return !invertPredicate ? Hexagon::JMP_EQrrPt_nv_V4 : - Hexagon::JMP_EQrrNotPt_nv_V4; - case Hexagon::JMP_EQriPt_nv_V4: - return !invertPredicate ? Hexagon::JMP_EQriPt_nv_V4 : - Hexagon::JMP_EQriNotPt_nv_V4; case Hexagon::COMBINE_rr: return !invertPredicate ? Hexagon::COMBINE_rr_cPt : Hexagon::COMBINE_rr_cNotPt; - case Hexagon::ASLH: - return !invertPredicate ? Hexagon::ASLH_cPt_V4 : - Hexagon::ASLH_cNotPt_V4; - case Hexagon::ASRH: - return !invertPredicate ? Hexagon::ASRH_cPt_V4 : - Hexagon::ASRH_cNotPt_V4; - case Hexagon::SXTB: - return !invertPredicate ? Hexagon::SXTB_cPt_V4 : - Hexagon::SXTB_cNotPt_V4; - case Hexagon::SXTH: - return !invertPredicate ? Hexagon::SXTH_cPt_V4 : - Hexagon::SXTH_cNotPt_V4; - case Hexagon::ZXTB: - return !invertPredicate ? Hexagon::ZXTB_cPt_V4 : - Hexagon::ZXTB_cNotPt_V4; - case Hexagon::ZXTH: - return !invertPredicate ? Hexagon::ZXTH_cPt_V4 : - Hexagon::ZXTH_cNotPt_V4; - - case Hexagon::JMPR: - return !invertPredicate ? Hexagon::JMPR_cPt : - Hexagon::JMPR_cNotPt; - - // V4 indexed+scaled load. - case Hexagon::LDrid_indexed_shl_V4: - return !invertPredicate ? Hexagon::LDrid_indexed_shl_cPt_V4 : - Hexagon::LDrid_indexed_shl_cNotPt_V4; - case Hexagon::LDrib_indexed_shl_V4: - return !invertPredicate ? Hexagon::LDrib_indexed_shl_cPt_V4 : - Hexagon::LDrib_indexed_shl_cNotPt_V4; - case Hexagon::LDriub_indexed_shl_V4: - return !invertPredicate ? Hexagon::LDriub_indexed_shl_cPt_V4 : - Hexagon::LDriub_indexed_shl_cNotPt_V4; - case Hexagon::LDrih_indexed_shl_V4: - return !invertPredicate ? Hexagon::LDrih_indexed_shl_cPt_V4 : - Hexagon::LDrih_indexed_shl_cNotPt_V4; - case Hexagon::LDriuh_indexed_shl_V4: - return !invertPredicate ? Hexagon::LDriuh_indexed_shl_cPt_V4 : - Hexagon::LDriuh_indexed_shl_cNotPt_V4; - case Hexagon::LDriw_indexed_shl_V4: - return !invertPredicate ? Hexagon::LDriw_indexed_shl_cPt_V4 : - Hexagon::LDriw_indexed_shl_cNotPt_V4; - - // V4 Load from global address - case Hexagon::LDd_GP_V4: - return !invertPredicate ? Hexagon::LDd_GP_cPt_V4 : - Hexagon::LDd_GP_cNotPt_V4; - case Hexagon::LDb_GP_V4: - return !invertPredicate ? Hexagon::LDb_GP_cPt_V4 : - Hexagon::LDb_GP_cNotPt_V4; - case Hexagon::LDub_GP_V4: - return !invertPredicate ? Hexagon::LDub_GP_cPt_V4 : - Hexagon::LDub_GP_cNotPt_V4; - case Hexagon::LDh_GP_V4: - return !invertPredicate ? Hexagon::LDh_GP_cPt_V4 : - Hexagon::LDh_GP_cNotPt_V4; - case Hexagon::LDuh_GP_V4: - return !invertPredicate ? Hexagon::LDuh_GP_cPt_V4 : - Hexagon::LDuh_GP_cNotPt_V4; - case Hexagon::LDw_GP_V4: - return !invertPredicate ? Hexagon::LDw_GP_cPt_V4 : - Hexagon::LDw_GP_cNotPt_V4; - - // Byte. - case Hexagon::POST_STbri: - return !invertPredicate ? Hexagon::POST_STbri_cPt : - Hexagon::POST_STbri_cNotPt; - case Hexagon::STrib: - return !invertPredicate ? Hexagon::STrib_cPt : - Hexagon::STrib_cNotPt; - case Hexagon::STrib_indexed: - return !invertPredicate ? Hexagon::STrib_indexed_cPt : - Hexagon::STrib_indexed_cNotPt; - case Hexagon::STrib_imm_V4: - return !invertPredicate ? Hexagon::STrib_imm_cPt_V4 : - Hexagon::STrib_imm_cNotPt_V4; - case Hexagon::STrib_indexed_shl_V4: - return !invertPredicate ? Hexagon::STrib_indexed_shl_cPt_V4 : - Hexagon::STrib_indexed_shl_cNotPt_V4; - // Halfword. - case Hexagon::POST_SThri: - return !invertPredicate ? Hexagon::POST_SThri_cPt : - Hexagon::POST_SThri_cNotPt; - case Hexagon::STrih: - return !invertPredicate ? Hexagon::STrih_cPt : - Hexagon::STrih_cNotPt; - case Hexagon::STrih_indexed: - return !invertPredicate ? Hexagon::STrih_indexed_cPt : - Hexagon::STrih_indexed_cNotPt; - case Hexagon::STrih_imm_V4: - return !invertPredicate ? Hexagon::STrih_imm_cPt_V4 : - Hexagon::STrih_imm_cNotPt_V4; - case Hexagon::STrih_indexed_shl_V4: - return !invertPredicate ? Hexagon::STrih_indexed_shl_cPt_V4 : - Hexagon::STrih_indexed_shl_cNotPt_V4; + // Word. - case Hexagon::POST_STwri: - return !invertPredicate ? Hexagon::POST_STwri_cPt : - Hexagon::POST_STwri_cNotPt; - case Hexagon::STriw: + case Hexagon::STriw_f: return !invertPredicate ? Hexagon::STriw_cPt : Hexagon::STriw_cNotPt; - case Hexagon::STriw_indexed: + case Hexagon::STriw_indexed_f: return !invertPredicate ? Hexagon::STriw_indexed_cPt : Hexagon::STriw_indexed_cNotPt; - case Hexagon::STriw_indexed_shl_V4: - return !invertPredicate ? Hexagon::STriw_indexed_shl_cPt_V4 : - Hexagon::STriw_indexed_shl_cNotPt_V4; - case Hexagon::STriw_imm_V4: - return !invertPredicate ? Hexagon::STriw_imm_cPt_V4 : - Hexagon::STriw_imm_cNotPt_V4; - // Double word. - case Hexagon::POST_STdri: - return !invertPredicate ? Hexagon::POST_STdri_cPt : - Hexagon::POST_STdri_cNotPt; - case Hexagon::STrid: - return !invertPredicate ? Hexagon::STrid_cPt : - Hexagon::STrid_cNotPt; - case Hexagon::STrid_indexed: - return !invertPredicate ? Hexagon::STrid_indexed_cPt : - Hexagon::STrid_indexed_cNotPt; - case Hexagon::STrid_indexed_shl_V4: - return !invertPredicate ? Hexagon::STrid_indexed_shl_cPt_V4 : - Hexagon::STrid_indexed_shl_cNotPt_V4; - - // V4 Store to global address - case Hexagon::STd_GP_V4: - return !invertPredicate ? Hexagon::STd_GP_cPt_V4 : - Hexagon::STd_GP_cNotPt_V4; - case Hexagon::STb_GP_V4: - return !invertPredicate ? Hexagon::STb_GP_cPt_V4 : - Hexagon::STb_GP_cNotPt_V4; - case Hexagon::STh_GP_V4: - return !invertPredicate ? Hexagon::STh_GP_cPt_V4 : - Hexagon::STh_GP_cNotPt_V4; - case Hexagon::STw_GP_V4: - return !invertPredicate ? Hexagon::STw_GP_cPt_V4 : - Hexagon::STw_GP_cNotPt_V4; - - // Load. - case Hexagon::LDrid: - return !invertPredicate ? Hexagon::LDrid_cPt : - Hexagon::LDrid_cNotPt; - case Hexagon::LDriw: - return !invertPredicate ? Hexagon::LDriw_cPt : - Hexagon::LDriw_cNotPt; - case Hexagon::LDrih: - return !invertPredicate ? Hexagon::LDrih_cPt : - Hexagon::LDrih_cNotPt; - case Hexagon::LDriuh: - return !invertPredicate ? Hexagon::LDriuh_cPt : - Hexagon::LDriuh_cNotPt; - case Hexagon::LDrib: - return !invertPredicate ? Hexagon::LDrib_cPt : - Hexagon::LDrib_cNotPt; - case Hexagon::LDriub: - return !invertPredicate ? Hexagon::LDriub_cPt : - Hexagon::LDriub_cNotPt; - // Load Indexed. - case Hexagon::LDrid_indexed: - return !invertPredicate ? Hexagon::LDrid_indexed_cPt : - Hexagon::LDrid_indexed_cNotPt; - case Hexagon::LDriw_indexed: - return !invertPredicate ? Hexagon::LDriw_indexed_cPt : - Hexagon::LDriw_indexed_cNotPt; - case Hexagon::LDrih_indexed: - return !invertPredicate ? Hexagon::LDrih_indexed_cPt : - Hexagon::LDrih_indexed_cNotPt; - case Hexagon::LDriuh_indexed: - return !invertPredicate ? Hexagon::LDriuh_indexed_cPt : - Hexagon::LDriuh_indexed_cNotPt; - case Hexagon::LDrib_indexed: - return !invertPredicate ? Hexagon::LDrib_indexed_cPt : - Hexagon::LDrib_indexed_cNotPt; - case Hexagon::LDriub_indexed: - return !invertPredicate ? Hexagon::LDriub_indexed_cPt : - Hexagon::LDriub_indexed_cNotPt; - // Post Increment Load. - case Hexagon::POST_LDrid: - return !invertPredicate ? Hexagon::POST_LDrid_cPt : - Hexagon::POST_LDrid_cNotPt; - case Hexagon::POST_LDriw: - return !invertPredicate ? Hexagon::POST_LDriw_cPt : - Hexagon::POST_LDriw_cNotPt; - case Hexagon::POST_LDrih: - return !invertPredicate ? Hexagon::POST_LDrih_cPt : - Hexagon::POST_LDrih_cNotPt; - case Hexagon::POST_LDriuh: - return !invertPredicate ? Hexagon::POST_LDriuh_cPt : - Hexagon::POST_LDriuh_cNotPt; - case Hexagon::POST_LDrib: - return !invertPredicate ? Hexagon::POST_LDrib_cPt : - Hexagon::POST_LDrib_cNotPt; - case Hexagon::POST_LDriub: - return !invertPredicate ? Hexagon::POST_LDriub_cPt : - Hexagon::POST_LDriub_cNotPt; + // DEALLOC_RETURN. case Hexagon::DEALLOC_RET_V4: return !invertPredicate ? Hexagon::DEALLOC_RET_cPt_V4 : @@ -1821,11 +911,15 @@ PredicateInstruction(MachineInstr *MI, // It is better to have an assert here to check this. But I don't know how // to write this assert because findFirstPredOperandIdx() would return -1 if (oper < -1) oper = -1; + MI->getOperand(oper+1).ChangeToRegister(PredMO.getReg(), PredMO.isDef(), - PredMO.isImplicit(), PredMO.isKill(), + PredMO.isImplicit(), false, PredMO.isDead(), PredMO.isUndef(), PredMO.isDebug()); + MachineRegisterInfo &RegInfo = MI->getParent()->getParent()->getRegInfo(); + RegInfo.clearKillFlags(PredMO.getReg()); + if (hasGAOpnd) { unsigned int i; @@ -1874,13 +968,65 @@ isProfitableToIfCvt(MachineBasicBlock &TMBB, return true; } - +// Returns true if an instruction is predicated irrespective of the predicate +// sense. For example, all of the following will return true. +// if (p0) R1 = add(R2, R3) +// if (!p0) R1 = add(R2, R3) +// if (p0.new) R1 = add(R2, R3) +// if (!p0.new) R1 = add(R2, R3) bool HexagonInstrInfo::isPredicated(const MachineInstr *MI) const { const uint64_t F = MI->getDesc().TSFlags; return ((F >> HexagonII::PredicatedPos) & HexagonII::PredicatedMask); } +bool HexagonInstrInfo::isPredicated(unsigned Opcode) const { + const uint64_t F = get(Opcode).TSFlags; + + return ((F >> HexagonII::PredicatedPos) & HexagonII::PredicatedMask); +} + +bool HexagonInstrInfo::isPredicatedTrue(const MachineInstr *MI) const { + const uint64_t F = MI->getDesc().TSFlags; + + assert(isPredicated(MI)); + return (!((F >> HexagonII::PredicatedFalsePos) & + HexagonII::PredicatedFalseMask)); +} + +bool HexagonInstrInfo::isPredicatedTrue(unsigned Opcode) const { + const uint64_t F = get(Opcode).TSFlags; + + // Make sure that the instruction is predicated. + assert((F>> HexagonII::PredicatedPos) & HexagonII::PredicatedMask); + return (!((F >> HexagonII::PredicatedFalsePos) & + HexagonII::PredicatedFalseMask)); +} + +bool HexagonInstrInfo::isPredicatedNew(const MachineInstr *MI) const { + const uint64_t F = MI->getDesc().TSFlags; + + assert(isPredicated(MI)); + return ((F >> HexagonII::PredicatedNewPos) & HexagonII::PredicatedNewMask); +} + +bool HexagonInstrInfo::isPredicatedNew(unsigned Opcode) const { + const uint64_t F = get(Opcode).TSFlags; + + assert(isPredicated(Opcode)); + return ((F >> HexagonII::PredicatedNewPos) & HexagonII::PredicatedNewMask); +} + +// Returns true, if a ST insn can be promoted to a new-value store. +bool HexagonInstrInfo::mayBeNewStore(const MachineInstr *MI) const { + const HexagonRegisterInfo& QRI = getRegisterInfo(); + const uint64_t F = MI->getDesc().TSFlags; + + return ((F >> HexagonII::mayNVStorePos) & + HexagonII::mayNVStoreMask & + QRI.Subtarget.hasV4TOps()); +} + bool HexagonInstrInfo::DefinesPredicate(MachineInstr *MI, std::vector<MachineOperand> &Pred) const { @@ -1991,46 +1137,28 @@ isValidOffset(const int Opcode, const int Offset) const { return (Offset >= Hexagon_ADDI_OFFSET_MIN) && (Offset <= Hexagon_ADDI_OFFSET_MAX); - case Hexagon::MEMw_ADDi_indexed_MEM_V4 : - case Hexagon::MEMw_SUBi_indexed_MEM_V4 : - case Hexagon::MEMw_ADDr_indexed_MEM_V4 : - case Hexagon::MEMw_SUBr_indexed_MEM_V4 : - case Hexagon::MEMw_ANDr_indexed_MEM_V4 : - case Hexagon::MEMw_ORr_indexed_MEM_V4 : - case Hexagon::MEMw_ADDi_MEM_V4 : - case Hexagon::MEMw_SUBi_MEM_V4 : - case Hexagon::MEMw_ADDr_MEM_V4 : - case Hexagon::MEMw_SUBr_MEM_V4 : - case Hexagon::MEMw_ANDr_MEM_V4 : - case Hexagon::MEMw_ORr_MEM_V4 : + case Hexagon::MemOPw_ADDi_V4 : + case Hexagon::MemOPw_SUBi_V4 : + case Hexagon::MemOPw_ADDr_V4 : + case Hexagon::MemOPw_SUBr_V4 : + case Hexagon::MemOPw_ANDr_V4 : + case Hexagon::MemOPw_ORr_V4 : return (0 <= Offset && Offset <= 255); - case Hexagon::MEMh_ADDi_indexed_MEM_V4 : - case Hexagon::MEMh_SUBi_indexed_MEM_V4 : - case Hexagon::MEMh_ADDr_indexed_MEM_V4 : - case Hexagon::MEMh_SUBr_indexed_MEM_V4 : - case Hexagon::MEMh_ANDr_indexed_MEM_V4 : - case Hexagon::MEMh_ORr_indexed_MEM_V4 : - case Hexagon::MEMh_ADDi_MEM_V4 : - case Hexagon::MEMh_SUBi_MEM_V4 : - case Hexagon::MEMh_ADDr_MEM_V4 : - case Hexagon::MEMh_SUBr_MEM_V4 : - case Hexagon::MEMh_ANDr_MEM_V4 : - case Hexagon::MEMh_ORr_MEM_V4 : + case Hexagon::MemOPh_ADDi_V4 : + case Hexagon::MemOPh_SUBi_V4 : + case Hexagon::MemOPh_ADDr_V4 : + case Hexagon::MemOPh_SUBr_V4 : + case Hexagon::MemOPh_ANDr_V4 : + case Hexagon::MemOPh_ORr_V4 : return (0 <= Offset && Offset <= 127); - case Hexagon::MEMb_ADDi_indexed_MEM_V4 : - case Hexagon::MEMb_SUBi_indexed_MEM_V4 : - case Hexagon::MEMb_ADDr_indexed_MEM_V4 : - case Hexagon::MEMb_SUBr_indexed_MEM_V4 : - case Hexagon::MEMb_ANDr_indexed_MEM_V4 : - case Hexagon::MEMb_ORr_indexed_MEM_V4 : - case Hexagon::MEMb_ADDi_MEM_V4 : - case Hexagon::MEMb_SUBi_MEM_V4 : - case Hexagon::MEMb_ADDr_MEM_V4 : - case Hexagon::MEMb_SUBr_MEM_V4 : - case Hexagon::MEMb_ANDr_MEM_V4 : - case Hexagon::MEMb_ORr_MEM_V4 : + case Hexagon::MemOPb_ADDi_V4 : + case Hexagon::MemOPb_SUBi_V4 : + case Hexagon::MemOPb_ADDr_V4 : + case Hexagon::MemOPb_SUBr_V4 : + case Hexagon::MemOPb_ANDr_V4 : + case Hexagon::MemOPb_ORr_V4 : return (0 <= Offset && Offset <= 63); // LDri_pred and STriw_pred are pseudo operations, so it has to take offset of @@ -2083,47 +1211,38 @@ isValidAutoIncImm(const EVT VT, const int Offset) const { bool HexagonInstrInfo:: isMemOp(const MachineInstr *MI) const { +// return MI->getDesc().mayLoad() && MI->getDesc().mayStore(); + switch (MI->getOpcode()) { default: return false; - case Hexagon::MEMw_ADDi_indexed_MEM_V4 : - case Hexagon::MEMw_SUBi_indexed_MEM_V4 : - case Hexagon::MEMw_ADDr_indexed_MEM_V4 : - case Hexagon::MEMw_SUBr_indexed_MEM_V4 : - case Hexagon::MEMw_ANDr_indexed_MEM_V4 : - case Hexagon::MEMw_ORr_indexed_MEM_V4 : - case Hexagon::MEMw_ADDi_MEM_V4 : - case Hexagon::MEMw_SUBi_MEM_V4 : - case Hexagon::MEMw_ADDr_MEM_V4 : - case Hexagon::MEMw_SUBr_MEM_V4 : - case Hexagon::MEMw_ANDr_MEM_V4 : - case Hexagon::MEMw_ORr_MEM_V4 : - case Hexagon::MEMh_ADDi_indexed_MEM_V4 : - case Hexagon::MEMh_SUBi_indexed_MEM_V4 : - case Hexagon::MEMh_ADDr_indexed_MEM_V4 : - case Hexagon::MEMh_SUBr_indexed_MEM_V4 : - case Hexagon::MEMh_ANDr_indexed_MEM_V4 : - case Hexagon::MEMh_ORr_indexed_MEM_V4 : - case Hexagon::MEMh_ADDi_MEM_V4 : - case Hexagon::MEMh_SUBi_MEM_V4 : - case Hexagon::MEMh_ADDr_MEM_V4 : - case Hexagon::MEMh_SUBr_MEM_V4 : - case Hexagon::MEMh_ANDr_MEM_V4 : - case Hexagon::MEMh_ORr_MEM_V4 : - case Hexagon::MEMb_ADDi_indexed_MEM_V4 : - case Hexagon::MEMb_SUBi_indexed_MEM_V4 : - case Hexagon::MEMb_ADDr_indexed_MEM_V4 : - case Hexagon::MEMb_SUBr_indexed_MEM_V4 : - case Hexagon::MEMb_ANDr_indexed_MEM_V4 : - case Hexagon::MEMb_ORr_indexed_MEM_V4 : - case Hexagon::MEMb_ADDi_MEM_V4 : - case Hexagon::MEMb_SUBi_MEM_V4 : - case Hexagon::MEMb_ADDr_MEM_V4 : - case Hexagon::MEMb_SUBr_MEM_V4 : - case Hexagon::MEMb_ANDr_MEM_V4 : - case Hexagon::MEMb_ORr_MEM_V4 : - return true; + case Hexagon::MemOPw_ADDi_V4 : + case Hexagon::MemOPw_SUBi_V4 : + case Hexagon::MemOPw_ADDr_V4 : + case Hexagon::MemOPw_SUBr_V4 : + case Hexagon::MemOPw_ANDr_V4 : + case Hexagon::MemOPw_ORr_V4 : + case Hexagon::MemOPh_ADDi_V4 : + case Hexagon::MemOPh_SUBi_V4 : + case Hexagon::MemOPh_ADDr_V4 : + case Hexagon::MemOPh_SUBr_V4 : + case Hexagon::MemOPh_ANDr_V4 : + case Hexagon::MemOPh_ORr_V4 : + case Hexagon::MemOPb_ADDi_V4 : + case Hexagon::MemOPb_SUBi_V4 : + case Hexagon::MemOPb_ADDr_V4 : + case Hexagon::MemOPb_SUBr_V4 : + case Hexagon::MemOPb_ANDr_V4 : + case Hexagon::MemOPb_ORr_V4 : + case Hexagon::MemOPb_SETBITi_V4: + case Hexagon::MemOPh_SETBITi_V4: + case Hexagon::MemOPw_SETBITi_V4: + case Hexagon::MemOPb_CLRBITi_V4: + case Hexagon::MemOPh_CLRBITi_V4: + case Hexagon::MemOPw_CLRBITi_V4: + return true; } + return false; } @@ -2142,14 +1261,10 @@ bool HexagonInstrInfo::isNewValueJumpCandidate(const MachineInstr *MI) const { default: return false; case Hexagon::CMPEQrr: case Hexagon::CMPEQri: - case Hexagon::CMPLTrr: case Hexagon::CMPGTrr: case Hexagon::CMPGTri: - case Hexagon::CMPLTUrr: case Hexagon::CMPGTUrr: case Hexagon::CMPGTUri: - case Hexagon::CMPGEri: - case Hexagon::CMPGEUri: return true; } } @@ -2382,6 +1497,120 @@ isConditionalStore (const MachineInstr* MI) const { } } + +bool HexagonInstrInfo::isNewValueJump(const MachineInstr *MI) const { + if (isNewValue(MI) && isBranch(MI)) + return true; + return false; +} + +bool HexagonInstrInfo::isPostIncrement (const MachineInstr* MI) const { + return (getAddrMode(MI) == HexagonII::PostInc); +} + +bool HexagonInstrInfo::isNewValue(const MachineInstr* MI) const { + const uint64_t F = MI->getDesc().TSFlags; + return ((F >> HexagonII::NewValuePos) & HexagonII::NewValueMask); +} + +// Returns true, if any one of the operands is a dot new +// insn, whether it is predicated dot new or register dot new. +bool HexagonInstrInfo::isDotNewInst (const MachineInstr* MI) const { + return (isNewValueInst(MI) || + (isPredicated(MI) && isPredicatedNew(MI))); +} + +// Returns the most basic instruction for the .new predicated instructions and +// new-value stores. +// For example, all of the following instructions will be converted back to the +// same instruction: +// 1) if (p0.new) memw(R0+#0) = R1.new ---> +// 2) if (p0) memw(R0+#0)= R1.new -------> if (p0) memw(R0+#0) = R1 +// 3) if (p0.new) memw(R0+#0) = R1 ---> +// + +int HexagonInstrInfo::GetDotOldOp(const int opc) const { + int NewOp = opc; + if (isPredicated(NewOp) && isPredicatedNew(NewOp)) { // Get predicate old form + NewOp = Hexagon::getPredOldOpcode(NewOp); + if (NewOp < 0) + assert(0 && "Couldn't change predicate new instruction to its old form."); + } + + if (isNewValueStore(NewOp)) { // Convert into non new-value format + NewOp = Hexagon::getNonNVStore(NewOp); + if (NewOp < 0) + assert(0 && "Couldn't change new-value store to its old form."); + } + return NewOp; +} + +// Return the new value instruction for a given store. +int HexagonInstrInfo::GetDotNewOp(const MachineInstr* MI) const { + int NVOpcode = Hexagon::getNewValueOpcode(MI->getOpcode()); + if (NVOpcode >= 0) // Valid new-value store instruction. + return NVOpcode; + + switch (MI->getOpcode()) { + default: llvm_unreachable("Unknown .new type"); + // store new value byte + case Hexagon::STrib_shl_V4: + return Hexagon::STrib_shl_nv_V4; + + case Hexagon::STrih_shl_V4: + return Hexagon::STrih_shl_nv_V4; + + case Hexagon::STriw_f: + return Hexagon::STriw_nv_V4; + + case Hexagon::STriw_indexed_f: + return Hexagon::STriw_indexed_nv_V4; + + case Hexagon::STriw_shl_V4: + return Hexagon::STriw_shl_nv_V4; + + } + return 0; +} + +// Return .new predicate version for an instruction. +int HexagonInstrInfo::GetDotNewPredOp(MachineInstr *MI, + const MachineBranchProbabilityInfo + *MBPI) const { + + int NewOpcode = Hexagon::getPredNewOpcode(MI->getOpcode()); + if (NewOpcode >= 0) // Valid predicate new instruction + return NewOpcode; + + switch (MI->getOpcode()) { + default: llvm_unreachable("Unknown .new type"); + // Condtional Jumps + case Hexagon::JMP_t: + case Hexagon::JMP_f: + return getDotNewPredJumpOp(MI, MBPI); + + case Hexagon::JMPR_t: + return Hexagon::JMPR_tnew_tV3; + + case Hexagon::JMPR_f: + return Hexagon::JMPR_fnew_tV3; + + case Hexagon::JMPret_t: + return Hexagon::JMPret_tnew_tV3; + + case Hexagon::JMPret_f: + return Hexagon::JMPret_fnew_tV3; + + + // Conditional combine + case Hexagon::COMBINE_rr_cPt : + return Hexagon::COMBINE_rr_cdnPt; + case Hexagon::COMBINE_rr_cNotPt : + return Hexagon::COMBINE_rr_cdnNotPt; + } +} + + unsigned HexagonInstrInfo::getAddrMode(const MachineInstr* MI) const { const uint64_t F = MI->getDesc().TSFlags; @@ -2476,6 +1705,34 @@ bool HexagonInstrInfo::isConstExtended(MachineInstr *MI) const { return (ImmValue < MinValue || ImmValue > MaxValue); } +// Returns the opcode to use when converting MI, which is a conditional jump, +// into a conditional instruction which uses the .new value of the predicate. +// We also use branch probabilities to add a hint to the jump. +int +HexagonInstrInfo::getDotNewPredJumpOp(MachineInstr *MI, + const + MachineBranchProbabilityInfo *MBPI) const { + + // We assume that block can have at most two successors. + bool taken = false; + MachineBasicBlock *Src = MI->getParent(); + MachineOperand *BrTarget = &MI->getOperand(1); + MachineBasicBlock *Dst = BrTarget->getMBB(); + + const BranchProbability Prediction = MBPI->getEdgeProbability(Src, Dst); + if (Prediction >= BranchProbability(1,2)) + taken = true; + + switch (MI->getOpcode()) { + case Hexagon::JMP_t: + return taken ? Hexagon::JMP_tnew_t : Hexagon::JMP_tnew_nt; + case Hexagon::JMP_f: + return taken ? Hexagon::JMP_fnew_t : Hexagon::JMP_fnew_nt; + + default: + llvm_unreachable("Unexpected jump instruction."); + } +} // Returns true if a particular operand is extendable for an instruction. bool HexagonInstrInfo::isOperandExtended(const MachineInstr *MI, unsigned short OperandNum) const { @@ -2580,3 +1837,18 @@ short HexagonInstrInfo::getNonExtOpcode (const MachineInstr *MI) const { } return -1; } + +bool HexagonInstrInfo::PredOpcodeHasJMP_c(Opcode_t Opcode) const { + return (Opcode == Hexagon::JMP_t) || + (Opcode == Hexagon::JMP_f) || + (Opcode == Hexagon::JMP_tnew_t) || + (Opcode == Hexagon::JMP_fnew_t) || + (Opcode == Hexagon::JMP_tnew_nt) || + (Opcode == Hexagon::JMP_fnew_nt); +} + +bool HexagonInstrInfo::PredOpcodeHasNot(Opcode_t Opcode) const { + return (Opcode == Hexagon::JMP_f) || + (Opcode == Hexagon::JMP_fnew_t) || + (Opcode == Hexagon::JMP_fnew_nt); +} diff --git a/lib/Target/Hexagon/HexagonInstrInfo.h b/lib/Target/Hexagon/HexagonInstrInfo.h index d2f059a..3c28df4 100644 --- a/lib/Target/Hexagon/HexagonInstrInfo.h +++ b/lib/Target/Hexagon/HexagonInstrInfo.h @@ -16,9 +16,9 @@ #include "HexagonRegisterInfo.h" #include "MCTargetDesc/HexagonBaseInfo.h" -#include "llvm/Target/TargetFrameLowering.h" #include "llvm/Target/TargetInstrInfo.h" - +#include "llvm/Target/TargetFrameLowering.h" +#include "llvm/CodeGen/MachineBranchProbabilityInfo.h" #define GET_INSTRINFO_HEADER #include "HexagonGenInstrInfo.inc" @@ -27,7 +27,9 @@ namespace llvm { class HexagonInstrInfo : public HexagonGenInstrInfo { const HexagonRegisterInfo RI; - const HexagonSubtarget& Subtarget; + const HexagonSubtarget &Subtarget; + typedef unsigned Opcode_t; + public: explicit HexagonInstrInfo(HexagonSubtarget &ST); @@ -111,6 +113,7 @@ public: unsigned createVR(MachineFunction* MF, MVT VT) const; + virtual bool isBranch(const MachineInstr *MI) const; virtual bool isPredicable(MachineInstr *MI) const; virtual bool PredicateInstruction(MachineInstr *MI, @@ -127,6 +130,11 @@ public: const BranchProbability &Probability) const; virtual bool isPredicated(const MachineInstr *MI) const; + virtual bool isPredicated(unsigned Opcode) const; + virtual bool isPredicatedTrue(const MachineInstr *MI) const; + virtual bool isPredicatedTrue(unsigned Opcode) const; + virtual bool isPredicatedNew(const MachineInstr *MI) const; + virtual bool isPredicatedNew(unsigned Opcode) const; virtual bool DefinesPredicate(MachineInstr *MI, std::vector<MachineOperand> &Pred) const; virtual bool @@ -170,18 +178,29 @@ public: bool isConditionalLoad (const MachineInstr* MI) const; bool isConditionalStore(const MachineInstr* MI) const; bool isNewValueInst(const MachineInstr* MI) const; + bool isNewValue(const MachineInstr* MI) const; + bool isDotNewInst(const MachineInstr* MI) const; + int GetDotOldOp(const int opc) const; + int GetDotNewOp(const MachineInstr* MI) const; + int GetDotNewPredOp(MachineInstr *MI, + const MachineBranchProbabilityInfo + *MBPI) const; + bool mayBeNewStore(const MachineInstr* MI) const; bool isDeallocRet(const MachineInstr *MI) const; unsigned getInvertedPredicatedOpcode(const int Opc) const; bool isExtendable(const MachineInstr* MI) const; bool isExtended(const MachineInstr* MI) const; bool isPostIncrement(const MachineInstr* MI) const; bool isNewValueStore(const MachineInstr* MI) const; + bool isNewValueStore(unsigned Opcode) const; bool isNewValueJump(const MachineInstr* MI) const; bool isNewValueJumpCandidate(const MachineInstr *MI) const; void immediateExtend(MachineInstr *MI) const; bool isConstExtended(MachineInstr *MI) const; + int getDotNewPredJumpOp(MachineInstr *MI, + const MachineBranchProbabilityInfo *MBPI) const; unsigned getAddrMode(const MachineInstr* MI) const; bool isOperandExtended(const MachineInstr *MI, unsigned short OperandNum) const; @@ -190,6 +209,9 @@ public: int getMaxValue(const MachineInstr *MI) const; bool NonExtEquivalentExists (const MachineInstr *MI) const; short getNonExtOpcode(const MachineInstr *MI) const; + bool PredOpcodeHasJMP_c(Opcode_t Opcode) const; + bool PredOpcodeHasNot(Opcode_t Opcode) const; + private: int getMatchingCondBranchOpcode(int Opc, bool sense) const; diff --git a/lib/Target/Hexagon/HexagonInstrInfo.td b/lib/Target/Hexagon/HexagonInstrInfo.td index d7bab20..c96aaca 100644 --- a/lib/Target/Hexagon/HexagonInstrInfo.td +++ b/lib/Target/Hexagon/HexagonInstrInfo.td @@ -14,6 +14,8 @@ include "HexagonInstrFormats.td" include "HexagonOperands.td" +//===----------------------------------------------------------------------===// + // Multi-class for logical operators. multiclass ALU32_rr_ri<string OpcStr, SDNode OpNode> { def rr : ALU32_rr<(outs IntRegs:$dst), (ins IntRegs:$b, IntRegs:$c), @@ -34,12 +36,6 @@ multiclass CMP64_rr<string OpcStr, PatFrag OpNode> { [(set (i1 PredRegs:$dst), (OpNode (i64 DoubleRegs:$b), (i64 DoubleRegs:$c)))]>; } -multiclass CMP32_rr<string OpcStr, PatFrag OpNode> { - def rr : ALU32_rr<(outs PredRegs:$dst), (ins IntRegs:$b, IntRegs:$c), - !strconcat("$dst = ", !strconcat(OpcStr, "($b, $c)")), - [(set (i1 PredRegs:$dst), - (OpNode (i32 IntRegs:$b), (i32 IntRegs:$c)))]>; -} multiclass CMP32_rr_ri_s10<string OpcStr, string CextOp, PatFrag OpNode> { let CextOpcode = CextOp in { @@ -75,14 +71,6 @@ multiclass CMP32_rr_ri_u9<string OpcStr, string CextOp, PatFrag OpNode> { } } -multiclass CMP32_ri_u8<string OpcStr, PatFrag OpNode> { -let isExtendable = 1, opExtendable = 2, isExtentSigned = 0, opExtentBits = 8 in - def ri : ALU32_ri<(outs PredRegs:$dst), (ins IntRegs:$b, u8Ext:$c), - !strconcat("$dst = ", !strconcat(OpcStr, "($b, #$c)")), - [(set (i1 PredRegs:$dst), (OpNode (i32 IntRegs:$b), - u8ExtPred:$c))]>; -} - multiclass CMP32_ri_s8<string OpcStr, PatFrag OpNode> { let isExtendable = 1, opExtendable = 2, isExtentSigned = 1, opExtentBits = 8 in def ri : ALU32_ri<(outs PredRegs:$dst), (ins IntRegs:$b, s8Ext:$c), @@ -95,22 +83,30 @@ let isExtendable = 1, opExtendable = 2, isExtentSigned = 1, opExtentBits = 8 in //===----------------------------------------------------------------------===// // ALU32/ALU (Instructions with register-register form) //===----------------------------------------------------------------------===// -multiclass ALU32_Pbase<string mnemonic, bit isNot, - bit isPredNew> { +def SDTHexagonI64I32I32 : SDTypeProfile<1, 2, + [SDTCisVT<0, i64>, SDTCisVT<1, i32>, SDTCisSameAs<1, 2>]>; + +def HexagonWrapperCombineII : + SDNode<"HexagonISD::WrapperCombineII", SDTHexagonI64I32I32>; - let PNewValue = !if(isPredNew, "new", "") in - def NAME : ALU32_rr<(outs IntRegs:$dst), +def HexagonWrapperCombineRR : + SDNode<"HexagonISD::WrapperCombineRR", SDTHexagonI64I32I32>; + +multiclass ALU32_Pbase<string mnemonic, RegisterClass RC, bit isNot, + bit isPredNew> { + let isPredicatedNew = isPredNew in + def NAME : ALU32_rr<(outs RC:$dst), (ins PredRegs:$src1, IntRegs:$src2, IntRegs: $src3), !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew,".new) $dst = ", ") $dst = ")#mnemonic#"($src2, $src3)", []>; } -multiclass ALU32_Pred<string mnemonic, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { - defm _c#NAME : ALU32_Pbase<mnemonic, PredNot, 0>; +multiclass ALU32_Pred<string mnemonic, RegisterClass RC, bit PredNot> { + let isPredicatedFalse = PredNot in { + defm _c#NAME : ALU32_Pbase<mnemonic, RC, PredNot, 0>; // Predicate new - defm _cdn#NAME : ALU32_Pbase<mnemonic, PredNot, 1>; + defm _cdn#NAME : ALU32_Pbase<mnemonic, RC, PredNot, 1>; } } @@ -125,8 +121,8 @@ multiclass ALU32_base<string mnemonic, string CextOp, SDNode OpNode> { (i32 IntRegs:$src2)))]>; let neverHasSideEffects = 1, isPredicated = 1 in { - defm Pt : ALU32_Pred<mnemonic, 0>; - defm NotPt : ALU32_Pred<mnemonic, 1>; + defm Pt : ALU32_Pred<mnemonic, IntRegs, 0>; + defm NotPt : ALU32_Pred<mnemonic, IntRegs, 1>; } } } @@ -140,11 +136,42 @@ let isCommutable = 1 in { defm SUB_rr : ALU32_base<"sub", "SUB", sub>, ImmRegRel, PredNewRel; +// Combines the two integer registers SRC1 and SRC2 into a double register. +let isPredicable = 1 in +class T_Combine : ALU32_rr<(outs DoubleRegs:$dst), + (ins IntRegs:$src1, IntRegs:$src2), + "$dst = combine($src1, $src2)", + [(set (i64 DoubleRegs:$dst), + (i64 (HexagonWrapperCombineRR (i32 IntRegs:$src1), + (i32 IntRegs:$src2))))]>; + +multiclass Combine_base { + let BaseOpcode = "combine" in { + def NAME : T_Combine; + let neverHasSideEffects = 1, isPredicated = 1 in { + defm Pt : ALU32_Pred<"combine", DoubleRegs, 0>; + defm NotPt : ALU32_Pred<"combine", DoubleRegs, 1>; + } + } +} + +defm COMBINE_rr : Combine_base, PredNewRel; + +// Combines the two immediates SRC1 and SRC2 into a double register. +class COMBINE_imm<Operand imm1, Operand imm2, PatLeaf pat1, PatLeaf pat2> : + ALU32_ii<(outs DoubleRegs:$dst), (ins imm1:$src1, imm2:$src2), + "$dst = combine(#$src1, #$src2)", + [(set (i64 DoubleRegs:$dst), + (i64 (HexagonWrapperCombineII (i32 pat1:$src1), (i32 pat2:$src2))))]>; + +let isExtendable = 1, opExtendable = 1, isExtentSigned = 1, opExtentBits = 8 in +def COMBINE_Ii : COMBINE_imm<s8Ext, s8Imm, s8ExtPred, s8ImmPred>; + //===----------------------------------------------------------------------===// // ALU32/ALU (ADD with register-immediate form) //===----------------------------------------------------------------------===// multiclass ALU32ri_Pbase<string mnemonic, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME : ALU32_ri<(outs IntRegs:$dst), (ins PredRegs:$src1, IntRegs:$src2, s8Ext: $src3), !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew,".new) $dst = ", @@ -153,7 +180,7 @@ multiclass ALU32ri_Pbase<string mnemonic, bit isNot, bit isPredNew> { } multiclass ALU32ri_Pred<string mnemonic, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { defm _c#NAME : ALU32ri_Pbase<mnemonic, PredNot, 0>; // Predicate new defm _cdn#NAME : ALU32ri_Pbase<mnemonic, PredNot, 1>; @@ -189,11 +216,6 @@ def OR_ri : ALU32_ri<(outs IntRegs:$dst), [(set (i32 IntRegs:$dst), (or (i32 IntRegs:$src1), s10ExtPred:$src2))]>, ImmRegRel; -def NOT_rr : ALU32_rr<(outs IntRegs:$dst), - (ins IntRegs:$src1), - "$dst = not($src1)", - [(set (i32 IntRegs:$dst), (not (i32 IntRegs:$src1)))]>; - let isExtendable = 1, opExtendable = 2, isExtentSigned = 1, opExtentBits = 10, InputType = "imm", CextOpcode = "AND" in def AND_ri : ALU32_ri<(outs IntRegs:$dst), @@ -201,10 +223,7 @@ def AND_ri : ALU32_ri<(outs IntRegs:$dst), "$dst = and($src1, #$src2)", [(set (i32 IntRegs:$dst), (and (i32 IntRegs:$src1), s10ExtPred:$src2))]>, ImmRegRel; -// Negate. -def NEG : ALU32_rr<(outs IntRegs:$dst), (ins IntRegs:$src1), - "$dst = neg($src1)", - [(set (i32 IntRegs:$dst), (ineg (i32 IntRegs:$src1)))]>; + // Nop. let neverHasSideEffects = 1 in def NOP : ALU32_rr<(outs), (ins), @@ -220,15 +239,21 @@ def SUB_ri : ALU32_ri<(outs IntRegs:$dst), [(set IntRegs:$dst, (sub s10ExtPred:$src1, IntRegs:$src2))]>, ImmRegRel; +// Rd = not(Rs) gets mapped to Rd=sub(#-1, Rs). +def : Pat<(not (i32 IntRegs:$src1)), + (SUB_ri -1, (i32 IntRegs:$src1))>; + +// Rd = neg(Rs) gets mapped to Rd=sub(#0, Rs). +// Pattern definition for 'neg' was not necessary. multiclass TFR_Pred<bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { def _c#NAME : ALU32_rr<(outs IntRegs:$dst), (ins PredRegs:$src1, IntRegs:$src2), !if(PredNot, "if (!$src1", "if ($src1")#") $dst = $src2", []>; // Predicate new - let PNewValue = "new" in + let isPredicatedNew = 1 in def _cdn#NAME : ALU32_rr<(outs IntRegs:$dst), (ins PredRegs:$src1, IntRegs:$src2), !if(PredNot, "if (!$src1", "if ($src1")#".new) $dst = $src2", @@ -274,10 +299,10 @@ class T_TFR64_Pred<bit PredNot, bit isPredNew> } multiclass TFR64_Pred<bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { def _c#NAME : T_TFR64_Pred<PredNot, 0>; - let PNewValue = "new" in + let isPredicatedNew = 1 in def _cdn#NAME : T_TFR64_Pred<PredNot, 1>; // Predicate new } } @@ -309,14 +334,14 @@ multiclass TFR64_base<string BaseName> { } multiclass TFRI_Pred<bit PredNot> { - let isMoveImm = 1, PredSense = !if(PredNot, "false", "true") in { + let isMoveImm = 1, isPredicatedFalse = PredNot in { def _c#NAME : ALU32_ri<(outs IntRegs:$dst), (ins PredRegs:$src1, s12Ext:$src2), !if(PredNot, "if (!$src1", "if ($src1")#") $dst = #$src2", []>; // Predicate new - let PNewValue = "new" in + let isPredicatedNew = 1 in def _cdn#NAME : ALU32_rr<(outs IntRegs:$dst), (ins PredRegs:$src1, s12Ext:$src2), !if(PredNot, "if (!$src1", "if ($src1")#".new) $dst = #$src2", @@ -359,51 +384,11 @@ def TFCR : CRInst<(outs CRRegs:$dst), (ins IntRegs:$src1), // ALU32/PERM + //===----------------------------------------------------------------------===// -// Combine. - -def SDTHexagonI64I32I32 : SDTypeProfile<1, 2, - [SDTCisVT<0, i64>, SDTCisVT<1, i32>, SDTCisSameAs<1, 2>]>; - -def HexagonWrapperCombineII : - SDNode<"HexagonISD::WrapperCombineII", SDTHexagonI64I32I32>; -def HexagonWrapperCombineRR : - SDNode<"HexagonISD::WrapperCombineRR", SDTHexagonI64I32I32>; - -// Combines the two integer registers SRC1 and SRC2 into a double register. -let isPredicable = 1 in -def COMBINE_rr : ALU32_rr<(outs DoubleRegs:$dst), (ins IntRegs:$src1, - IntRegs:$src2), - "$dst = combine($src1, $src2)", - [(set (i64 DoubleRegs:$dst), - (i64 (HexagonWrapperCombineRR (i32 IntRegs:$src1), - (i32 IntRegs:$src2))))]>; - -// Rd=combine(Rt.[HL], Rs.[HL]) -class COMBINE_halves<string A, string B>: ALU32_rr<(outs IntRegs:$dst), - (ins IntRegs:$src1, - IntRegs:$src2), - "$dst = combine($src1."# A #", $src2."# B #")", []>; - -let isPredicable = 1 in { - def COMBINE_hh : COMBINE_halves<"H", "H">; - def COMBINE_hl : COMBINE_halves<"H", "L">; - def COMBINE_lh : COMBINE_halves<"L", "H">; - def COMBINE_ll : COMBINE_halves<"L", "L">; -} - -def : Pat<(i32 (trunc (i64 (srl (i64 DoubleRegs:$a), (i32 16))))), - (COMBINE_lh (EXTRACT_SUBREG (i64 DoubleRegs:$a), subreg_hireg), - (EXTRACT_SUBREG (i64 DoubleRegs:$a), subreg_loreg))>; - -// Combines the two immediates SRC1 and SRC2 into a double register. -class COMBINE_imm<Operand imm1, Operand imm2, PatLeaf pat1, PatLeaf pat2> : - ALU32_ii<(outs DoubleRegs:$dst), (ins imm1:$src1, imm2:$src2), - "$dst = combine(#$src1, #$src2)", - [(set (i64 DoubleRegs:$dst), - (i64 (HexagonWrapperCombineII (i32 pat1:$src1), (i32 pat2:$src2))))]>; - -let isExtendable = 1, opExtendable = 1, isExtentSigned = 1, opExtentBits = 8 in -def COMBINE_Ii : COMBINE_imm<s8Ext, s8Imm, s8ExtPred, s8ImmPred>; +let neverHasSideEffects = 1 in +def COMBINE_ii : ALU32_ii<(outs DoubleRegs:$dst), + (ins s8Imm:$src1, s8Imm:$src2), + "$dst = combine(#$src1, #$src2)", + []>; // Mux. def VMUX_prr64 : ALU64_rr<(outs DoubleRegs:$dst), (ins PredRegs:$src1, @@ -446,38 +431,58 @@ def MUX_ii : ALU32_ii<(outs IntRegs:$dst), (ins PredRegs:$src1, s8Ext:$src2, s8ExtPred:$src2, s8ImmPred:$src3)))]>; -// Shift halfword. -let isPredicable = 1 in -def ASLH : ALU32_rr<(outs IntRegs:$dst), (ins IntRegs:$src1), - "$dst = aslh($src1)", - [(set (i32 IntRegs:$dst), (shl 16, (i32 IntRegs:$src1)))]>; +// ALU32 - aslh, asrh, sxtb, sxth, zxtb, zxth +multiclass ALU32_2op_Pbase<string mnemonic, bit isNot, bit isPredNew> { + let isPredicatedNew = isPredNew in + def NAME : ALU32Inst<(outs IntRegs:$dst), + (ins PredRegs:$src1, IntRegs:$src2), + !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew,".new) $dst = ", + ") $dst = ")#mnemonic#"($src2)">, + Requires<[HasV4T]>; +} -let isPredicable = 1 in -def ASRH : ALU32_rr<(outs IntRegs:$dst), (ins IntRegs:$src1), - "$dst = asrh($src1)", - [(set (i32 IntRegs:$dst), (sra 16, (i32 IntRegs:$src1)))]>; +multiclass ALU32_2op_Pred<string mnemonic, bit PredNot> { + let isPredicatedFalse = PredNot in { + defm _c#NAME : ALU32_2op_Pbase<mnemonic, PredNot, 0>; + // Predicate new + defm _cdn#NAME : ALU32_2op_Pbase<mnemonic, PredNot, 1>; + } +} -// Sign extend. -let isPredicable = 1 in -def SXTB : ALU32_rr<(outs IntRegs:$dst), (ins IntRegs:$src1), - "$dst = sxtb($src1)", - [(set (i32 IntRegs:$dst), (sext_inreg (i32 IntRegs:$src1), i8))]>; +multiclass ALU32_2op_base<string mnemonic> { + let BaseOpcode = mnemonic in { + let isPredicable = 1, neverHasSideEffects = 1 in + def NAME : ALU32Inst<(outs IntRegs:$dst), + (ins IntRegs:$src1), + "$dst = "#mnemonic#"($src1)">; -let isPredicable = 1 in -def SXTH : ALU32_rr<(outs IntRegs:$dst), (ins IntRegs:$src1), - "$dst = sxth($src1)", - [(set (i32 IntRegs:$dst), (sext_inreg (i32 IntRegs:$src1), i16))]>; - -// Zero extend. -let isPredicable = 1, neverHasSideEffects = 1 in -def ZXTB : ALU32_rr<(outs IntRegs:$dst), (ins IntRegs:$src1), - "$dst = zxtb($src1)", - []>; + let Predicates = [HasV4T], validSubTargets = HasV4SubT, isPredicated = 1, + neverHasSideEffects = 1 in { + defm Pt_V4 : ALU32_2op_Pred<mnemonic, 0>; + defm NotPt_V4 : ALU32_2op_Pred<mnemonic, 1>; + } + } +} + +defm ASLH : ALU32_2op_base<"aslh">, PredNewRel; +defm ASRH : ALU32_2op_base<"asrh">, PredNewRel; +defm SXTB : ALU32_2op_base<"sxtb">, PredNewRel; +defm SXTH : ALU32_2op_base<"sxth">, PredNewRel; +defm ZXTB : ALU32_2op_base<"zxtb">, PredNewRel; +defm ZXTH : ALU32_2op_base<"zxth">, PredNewRel; + +def : Pat <(shl (i32 IntRegs:$src1), (i32 16)), + (ASLH IntRegs:$src1)>; + +def : Pat <(sra (i32 IntRegs:$src1), (i32 16)), + (ASRH IntRegs:$src1)>; + +def : Pat <(sext_inreg (i32 IntRegs:$src1), i8), + (SXTB IntRegs:$src1)>; + +def : Pat <(sext_inreg (i32 IntRegs:$src1), i16), + (SXTH IntRegs:$src1)>; -let isPredicable = 1, neverHasSideEffects = 1 in -def ZXTH : ALU32_rr<(outs IntRegs:$dst), (ins IntRegs:$src1), - "$dst = zxth($src1)", - []>; //===----------------------------------------------------------------------===// // ALU32/PERM - //===----------------------------------------------------------------------===// @@ -487,39 +492,24 @@ def ZXTH : ALU32_rr<(outs IntRegs:$dst), (ins IntRegs:$src1), // ALU32/PRED + //===----------------------------------------------------------------------===// -// Conditional combine. -let neverHasSideEffects = 1, isPredicated = 1 in -def COMBINE_rr_cPt : ALU32_rr<(outs DoubleRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2, IntRegs:$src3), - "if ($src1) $dst = combine($src2, $src3)", - []>; - -let neverHasSideEffects = 1, isPredicated = 1 in -def COMBINE_rr_cNotPt : ALU32_rr<(outs DoubleRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2, IntRegs:$src3), - "if (!$src1) $dst = combine($src2, $src3)", - []>; - -let neverHasSideEffects = 1, isPredicated = 1 in -def COMBINE_rr_cdnPt : ALU32_rr<(outs DoubleRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2, IntRegs:$src3), - "if ($src1.new) $dst = combine($src2, $src3)", - []>; - -let neverHasSideEffects = 1, isPredicated = 1 in -def COMBINE_rr_cdnNotPt : ALU32_rr<(outs DoubleRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2, IntRegs:$src3), - "if (!$src1.new) $dst = combine($src2, $src3)", - []>; - // Compare. defm CMPGTU : CMP32_rr_ri_u9<"cmp.gtu", "CMPGTU", setugt>, ImmRegRel; defm CMPGT : CMP32_rr_ri_s10<"cmp.gt", "CMPGT", setgt>, ImmRegRel; -defm CMPLT : CMP32_rr<"cmp.lt", setlt>; -defm CMPLTU : CMP32_rr<"cmp.ltu", setult>; defm CMPEQ : CMP32_rr_ri_s10<"cmp.eq", "CMPEQ", seteq>, ImmRegRel; -defm CMPGE : CMP32_ri_s8<"cmp.ge", setge>; -defm CMPGEU : CMP32_ri_u8<"cmp.geu", setuge>; + +// SDNode for converting immediate C to C-1. +def DEC_CONST_SIGNED : SDNodeXForm<imm, [{ + // Return the byte immediate const-1 as an SDNode. + int32_t imm = N->getSExtValue(); + return XformSToSM1Imm(imm); +}]>; + +// SDNode for converting immediate C to C-1. +def DEC_CONST_UNSIGNED : SDNodeXForm<imm, [{ + // Return the byte immediate const-1 as an SDNode. + uint32_t imm = N->getZExtValue(); + return XformUToUM1Imm(imm); +}]>; def CTLZ_rr : SInst<(outs IntRegs:$dst), (ins IntRegs:$src1), "$dst = cl0($src1)", @@ -753,112 +743,153 @@ def XOR_pp : SInst<(outs PredRegs:$dst), (ins PredRegs:$src1, PredRegs:$src2), // CR - //===----------------------------------------------------------------------===// +def retflag : SDNode<"HexagonISD::RET_FLAG", SDTNone, + [SDNPHasChain, SDNPOptInGlue, SDNPVariadic]>; +def eh_return: SDNode<"HexagonISD::EH_RETURN", SDTNone, + [SDNPHasChain]>; -//===----------------------------------------------------------------------===// -// J + -//===----------------------------------------------------------------------===// -// Jump to address. -let isBranch = 1, isTerminator=1, isBarrier = 1, isPredicable = 1 in { - def JMP : JInst< (outs), - (ins brtarget:$offset), - "jump $offset", - [(br bb:$offset)]>; -} +def SDHexagonBR_JT: SDTypeProfile<0, 1, [SDTCisVT<0, i32>]>; +def HexagonBR_JT: SDNode<"HexagonISD::BR_JT", SDHexagonBR_JT, [SDNPHasChain]>; -// if (p0) jump -let isBranch = 1, isTerminator=1, Defs = [PC], - isPredicated = 1 in { - def JMP_c : JInst< (outs), - (ins PredRegs:$src, brtarget:$offset), - "if ($src) jump $offset", - [(brcond (i1 PredRegs:$src), bb:$offset)]>; -} +let InputType = "imm", isBarrier = 1, isPredicable = 1, +Defs = [PC], isExtendable = 1, opExtendable = 0, isExtentSigned = 1, +opExtentBits = 24 in +class T_JMP <dag InsDag, list<dag> JumpList = []> + : JInst<(outs), InsDag, + "jump $dst" , JumpList> { + bits<24> dst; + + let IClass = 0b0101; + + let Inst{27-25} = 0b100; + let Inst{24-16} = dst{23-15}; + let Inst{13-1} = dst{14-2}; +} + +let InputType = "imm", isExtendable = 1, opExtendable = 1, isExtentSigned = 1, +Defs = [PC], isPredicated = 1, opExtentBits = 17 in +class T_JMP_c <bit PredNot, bit isPredNew, bit isTaken>: + JInst<(outs ), (ins PredRegs:$src, brtarget:$dst), + !if(PredNot, "if (!$src", "if ($src")# + !if(isPredNew, ".new) ", ") ")#"jump"# + !if(isPredNew, !if(isTaken, ":t ", ":nt "), " ")#"$dst"> { + + let isBrTaken = !if(isPredNew, !if(isTaken, "true", "false"), ""); + let isPredicatedFalse = PredNot; + let isPredicatedNew = isPredNew; + bits<2> src; + bits<17> dst; + + let IClass = 0b0101; + + let Inst{27-24} = 0b1100; + let Inst{21} = PredNot; + let Inst{12} = !if(isPredNew, isTaken, zero); + let Inst{11} = isPredNew; + let Inst{9-8} = src; + let Inst{23-22} = dst{16-15}; + let Inst{20-16} = dst{14-10}; + let Inst{13} = dst{9}; + let Inst{7-1} = dst{8-2}; + } -// if (!p0) jump -let isBranch = 1, isTerminator=1, neverHasSideEffects = 1, Defs = [PC], - isPredicated = 1 in { - def JMP_cNot : JInst< (outs), - (ins PredRegs:$src, brtarget:$offset), - "if (!$src) jump $offset", - []>; +let isBarrier = 1, Defs = [PC], isPredicable = 1, InputType = "reg" in +class T_JMPr<dag InsDag = (ins IntRegs:$dst)> + : JRInst<(outs ), InsDag, + "jumpr $dst" , + []> { + bits<5> dst; + + let IClass = 0b0101; + let Inst{27-21} = 0b0010100; + let Inst{20-16} = dst; } -let isTerminator = 1, isBranch = 1, neverHasSideEffects = 1, Defs = [PC], - isPredicated = 1 in { - def BRCOND : JInst < (outs), (ins PredRegs:$pred, brtarget:$dst), - "if ($pred) jump $dst", - []>; +let Defs = [PC], isPredicated = 1, InputType = "reg" in +class T_JMPr_c <bit PredNot, bit isPredNew, bit isTaken>: + JRInst <(outs ), (ins PredRegs:$src, IntRegs:$dst), + !if(PredNot, "if (!$src", "if ($src")# + !if(isPredNew, ".new) ", ") ")#"jumpr"# + !if(isPredNew, !if(isTaken, ":t ", ":nt "), " ")#"$dst"> { + + let isBrTaken = !if(isPredNew, !if(isTaken, "true", "false"), ""); + let isPredicatedFalse = PredNot; + let isPredicatedNew = isPredNew; + bits<2> src; + bits<5> dst; + + let IClass = 0b0101; + + let Inst{27-22} = 0b001101; + let Inst{21} = PredNot; + let Inst{20-16} = dst; + let Inst{12} = !if(isPredNew, isTaken, zero); + let Inst{11} = isPredNew; + let Inst{9-8} = src; + let Predicates = !if(isPredNew, [HasV3T], [HasV2T]); + let validSubTargets = !if(isPredNew, HasV3SubT, HasV2SubT); } -// Jump to address conditioned on new predicate. -// if (p0) jump:t -let isBranch = 1, isTerminator=1, neverHasSideEffects = 1, Defs = [PC], - isPredicated = 1 in { - def JMP_cdnPt : JInst< (outs), - (ins PredRegs:$src, brtarget:$offset), - "if ($src.new) jump:t $offset", - []>; +multiclass JMP_Pred<bit PredNot> { + def _#NAME : T_JMP_c<PredNot, 0, 0>; + // Predicate new + def _#NAME#new_t : T_JMP_c<PredNot, 1, 1>; // taken + def _#NAME#new_nt : T_JMP_c<PredNot, 1, 0>; // not taken } -// if (!p0) jump:t -let isBranch = 1, isTerminator=1, neverHasSideEffects = 1, Defs = [PC], - isPredicated = 1 in { - def JMP_cdnNotPt : JInst< (outs), - (ins PredRegs:$src, brtarget:$offset), - "if (!$src.new) jump:t $offset", - []>; +multiclass JMP_base<string BaseOp> { + let BaseOpcode = BaseOp in { + def NAME : T_JMP<(ins brtarget:$dst), [(br bb:$dst)]>; + defm t : JMP_Pred<0>; + defm f : JMP_Pred<1>; + } } -// Not taken. -let isBranch = 1, isTerminator=1, neverHasSideEffects = 1, Defs = [PC], - isPredicated = 1 in { - def JMP_cdnPnt : JInst< (outs), - (ins PredRegs:$src, brtarget:$offset), - "if ($src.new) jump:nt $offset", - []>; +multiclass JMPR_Pred<bit PredNot> { + def NAME: T_JMPr_c<PredNot, 0, 0>; + // Predicate new + def NAME#new_tV3 : T_JMPr_c<PredNot, 1, 1>; // taken + def NAME#new_ntV3 : T_JMPr_c<PredNot, 1, 0>; // not taken } -// Not taken. -let isBranch = 1, isTerminator=1, neverHasSideEffects = 1, Defs = [PC], - isPredicated = 1 in { - def JMP_cdnNotPnt : JInst< (outs), - (ins PredRegs:$src, brtarget:$offset), - "if (!$src.new) jump:nt $offset", - []>; +multiclass JMPR_base<string BaseOp> { + let BaseOpcode = BaseOp in { + def NAME : T_JMPr; + defm _t : JMPR_Pred<0>; + defm _f : JMPR_Pred<1>; + } } -//===----------------------------------------------------------------------===// -// J - -//===----------------------------------------------------------------------===// -//===----------------------------------------------------------------------===// -// JR + -//===----------------------------------------------------------------------===// -def retflag : SDNode<"HexagonISD::RET_FLAG", SDTNone, - [SDNPHasChain, SDNPOptInGlue, SDNPVariadic]>; +let isTerminator = 1, neverHasSideEffects = 1 in { +let isBranch = 1 in +defm JMP : JMP_base<"JMP">, PredNewRel; -// Jump to address from register. -let isPredicable =1, isReturn = 1, isTerminator = 1, isBarrier = 1, - Defs = [PC], Uses = [R31] in { - def JMPR: JRInst<(outs), (ins), - "jumpr r31", - [(retflag)]>; -} +let isBranch = 1, isIndirectBranch = 1 in +defm JMPR : JMPR_base<"JMPr">, PredNewRel; -// Jump to address from register. -let isReturn = 1, isTerminator = 1, isBarrier = 1, isPredicated = 1, - Defs = [PC], Uses = [R31] in { - def JMPR_cPt: JRInst<(outs), (ins PredRegs:$src1), - "if ($src1) jumpr r31", - []>; +let isReturn = 1, isCodeGenOnly = 1 in +defm JMPret : JMPR_base<"JMPret">, PredNewRel; } -// Jump to address from register. -let isReturn = 1, isTerminator = 1, isBarrier = 1, isPredicated = 1, - Defs = [PC], Uses = [R31] in { - def JMPR_cNotPt: JRInst<(outs), (ins PredRegs:$src1), - "if (!$src1) jumpr r31", - []>; -} +def : Pat<(retflag), + (JMPret (i32 R31))>; + +def : Pat <(brcond (i1 PredRegs:$src1), bb:$offset), + (JMP_t (i1 PredRegs:$src1), bb:$offset)>; + +// A return through builtin_eh_return. +let isReturn = 1, isTerminator = 1, isBarrier = 1, neverHasSideEffects = 1, +isCodeGenOnly = 1, Defs = [PC], Uses = [R28], isPredicable = 0 in +def EH_RETURN_JMPR : T_JMPr; + +def : Pat<(eh_return), + (EH_RETURN_JMPR (i32 R31))>; + +def : Pat<(HexagonBR_JT (i32 IntRegs:$dst)), + (JMPR (i32 IntRegs:$dst))>; + +def : Pat<(brind (i32 IntRegs:$dst)), + (JMPR (i32 IntRegs:$dst))>; //===----------------------------------------------------------------------===// // JR - @@ -871,7 +902,7 @@ let isReturn = 1, isTerminator = 1, isBarrier = 1, isPredicated = 1, // Load -- MEMri operand multiclass LD_MEMri_Pbase<string mnemonic, RegisterClass RC, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME : LDInst2<(outs RC:$dst), (ins PredRegs:$src1, MEMri:$addr), !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ", @@ -880,7 +911,7 @@ multiclass LD_MEMri_Pbase<string mnemonic, RegisterClass RC, } multiclass LD_MEMri_Pred<string mnemonic, RegisterClass RC, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { defm _c#NAME : LD_MEMri_Pbase<mnemonic, RC, PredNot, 0>; // Predicate new defm _cdn#NAME : LD_MEMri_Pbase<mnemonic, RC, PredNot, 1>; @@ -907,12 +938,21 @@ multiclass LD_MEMri<string mnemonic, string CextOp, RegisterClass RC, } let addrMode = BaseImmOffset, isMEMri = "true" in { - defm LDrib: LD_MEMri < "memb", "LDrib", IntRegs, 11, 6>, AddrModeRel; - defm LDriub: LD_MEMri < "memub" , "LDriub", IntRegs, 11, 6>, AddrModeRel; - defm LDrih: LD_MEMri < "memh", "LDrih", IntRegs, 12, 7>, AddrModeRel; - defm LDriuh: LD_MEMri < "memuh", "LDriuh", IntRegs, 12, 7>, AddrModeRel; - defm LDriw: LD_MEMri < "memw", "LDriw", IntRegs, 13, 8>, AddrModeRel; - defm LDrid: LD_MEMri < "memd", "LDrid", DoubleRegs, 14, 9>, AddrModeRel; + let accessSize = ByteAccess in { + defm LDrib: LD_MEMri < "memb", "LDrib", IntRegs, 11, 6>, AddrModeRel; + defm LDriub: LD_MEMri < "memub" , "LDriub", IntRegs, 11, 6>, AddrModeRel; + } + + let accessSize = HalfWordAccess in { + defm LDrih: LD_MEMri < "memh", "LDrih", IntRegs, 12, 7>, AddrModeRel; + defm LDriuh: LD_MEMri < "memuh", "LDriuh", IntRegs, 12, 7>, AddrModeRel; + } + + let accessSize = WordAccess in + defm LDriw: LD_MEMri < "memw", "LDriw", IntRegs, 13, 8>, AddrModeRel; + + let accessSize = DoubleWordAccess in + defm LDrid: LD_MEMri < "memd", "LDrid", DoubleRegs, 14, 9>, AddrModeRel; } def : Pat < (i32 (sextloadi8 ADDRriS11_0:$addr)), @@ -937,7 +977,7 @@ def : Pat < (i64 (load ADDRriS11_3:$addr)), // Load - Base with Immediate offset addressing mode multiclass LD_Idxd_Pbase<string mnemonic, RegisterClass RC, Operand predImmOp, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME : LDInst2<(outs RC:$dst), (ins PredRegs:$src1, IntRegs:$src2, predImmOp:$src3), !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ", @@ -947,7 +987,7 @@ multiclass LD_Idxd_Pbase<string mnemonic, RegisterClass RC, Operand predImmOp, multiclass LD_Idxd_Pred<string mnemonic, RegisterClass RC, Operand predImmOp, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { defm _c#NAME : LD_Idxd_Pbase<mnemonic, RC, predImmOp, PredNot, 0>; // Predicate new defm _cdn#NAME : LD_Idxd_Pbase<mnemonic, RC, predImmOp, PredNot, 1>; @@ -975,18 +1015,25 @@ multiclass LD_Idxd<string mnemonic, string CextOp, RegisterClass RC, } let addrMode = BaseImmOffset in { - defm LDrib_indexed: LD_Idxd <"memb", "LDrib", IntRegs, s11_0Ext, u6_0Ext, - 11, 6>, AddrModeRel; - defm LDriub_indexed: LD_Idxd <"memub" , "LDriub", IntRegs, s11_0Ext, u6_0Ext, - 11, 6>, AddrModeRel; - defm LDrih_indexed: LD_Idxd <"memh", "LDrih", IntRegs, s11_1Ext, u6_1Ext, - 12, 7>, AddrModeRel; - defm LDriuh_indexed: LD_Idxd <"memuh", "LDriuh", IntRegs, s11_1Ext, u6_1Ext, - 12, 7>, AddrModeRel; - defm LDriw_indexed: LD_Idxd <"memw", "LDriw", IntRegs, s11_2Ext, u6_2Ext, - 13, 8>, AddrModeRel; - defm LDrid_indexed: LD_Idxd <"memd", "LDrid", DoubleRegs, s11_3Ext, u6_3Ext, - 14, 9>, AddrModeRel; + let accessSize = ByteAccess in { + defm LDrib_indexed: LD_Idxd <"memb", "LDrib", IntRegs, s11_0Ext, u6_0Ext, + 11, 6>, AddrModeRel; + defm LDriub_indexed: LD_Idxd <"memub" , "LDriub", IntRegs, s11_0Ext, u6_0Ext, + 11, 6>, AddrModeRel; + } + let accessSize = HalfWordAccess in { + defm LDrih_indexed: LD_Idxd <"memh", "LDrih", IntRegs, s11_1Ext, u6_1Ext, + 12, 7>, AddrModeRel; + defm LDriuh_indexed: LD_Idxd <"memuh", "LDriuh", IntRegs, s11_1Ext, u6_1Ext, + 12, 7>, AddrModeRel; + } + let accessSize = WordAccess in + defm LDriw_indexed: LD_Idxd <"memw", "LDriw", IntRegs, s11_2Ext, u6_2Ext, + 13, 8>, AddrModeRel; + + let accessSize = DoubleWordAccess in + defm LDrid_indexed: LD_Idxd <"memd", "LDrid", DoubleRegs, s11_3Ext, u6_3Ext, + 14, 9>, AddrModeRel; } let AddedComplexity = 20 in { @@ -1009,29 +1056,13 @@ def : Pat < (i64 (load (add IntRegs:$src1, s11_3ExtPred:$offset))), (LDrid_indexed IntRegs:$src1, s11_3ExtPred:$offset) >; } -let neverHasSideEffects = 1 in -def LDrid_GP : LDInst2<(outs DoubleRegs:$dst), - (ins globaladdress:$global, u16Imm:$offset), - "$dst = memd(#$global+$offset)", - []>, - Requires<[NoV4T]>; - -let neverHasSideEffects = 1, validSubTargets = NoV4SubT in -def LDd_GP : LDInst2<(outs DoubleRegs:$dst), - (ins globaladdress:$global), - "$dst = memd(#$global)", - []>, - Requires<[NoV4T]>; - //===----------------------------------------------------------------------===// // Post increment load -// Make sure that in post increment load, the first operand is always the post -// increment operand. //===----------------------------------------------------------------------===// multiclass LD_PostInc_Pbase<string mnemonic, RegisterClass RC, Operand ImmOp, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME : LDInst2PI<(outs RC:$dst, IntRegs:$dst2), (ins PredRegs:$src1, IntRegs:$src2, ImmOp:$offset), !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ", @@ -1042,7 +1073,7 @@ multiclass LD_PostInc_Pbase<string mnemonic, RegisterClass RC, Operand ImmOp, multiclass LD_PostInc_Pred<string mnemonic, RegisterClass RC, Operand ImmOp, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { defm _c#NAME : LD_PostInc_Pbase<mnemonic, RC, ImmOp, PredNot, 0>; // Predicate new let Predicates = [HasV4T], validSubTargets = HasV4SubT in @@ -1068,7 +1099,7 @@ multiclass LD_PostInc<string mnemonic, string BaseOp, RegisterClass RC, } } -let hasCtrlDep = 1, neverHasSideEffects = 1 in { +let hasCtrlDep = 1, neverHasSideEffects = 1, addrMode = PostInc in { defm POST_LDrib : LD_PostInc<"memb", "LDrib", IntRegs, s4_0Imm>, PredNewRel; defm POST_LDriub : LD_PostInc<"memub", "LDriub", IntRegs, s4_0Imm>, @@ -1095,27 +1126,6 @@ let AddedComplexity = 20 in def : Pat < (i32 (extloadi8 (add IntRegs:$src1, s11_0ImmPred:$offset))), (i32 (LDrib_indexed IntRegs:$src1, s11_0ImmPred:$offset)) >; -let neverHasSideEffects = 1 in -def LDrib_GP : LDInst2<(outs IntRegs:$dst), - (ins globaladdress:$global, u16Imm:$offset), - "$dst = memb(#$global+$offset)", - []>, - Requires<[NoV4T]>; - -let neverHasSideEffects = 1, validSubTargets = NoV4SubT in -def LDb_GP : LDInst2<(outs IntRegs:$dst), - (ins globaladdress:$global), - "$dst = memb(#$global)", - []>, - Requires<[NoV4T]>; - -let neverHasSideEffects = 1, validSubTargets = NoV4SubT in -def LDub_GP : LDInst2<(outs IntRegs:$dst), - (ins globaladdress:$global), - "$dst = memub(#$global)", - []>, - Requires<[NoV4T]>; - def : Pat < (i32 (extloadi16 ADDRriS11_1:$addr)), (i32 (LDrih ADDRriS11_1:$addr))>; @@ -1123,27 +1133,6 @@ let AddedComplexity = 20 in def : Pat < (i32 (extloadi16 (add IntRegs:$src1, s11_1ImmPred:$offset))), (i32 (LDrih_indexed IntRegs:$src1, s11_1ImmPred:$offset)) >; -let neverHasSideEffects = 1 in -def LDrih_GP : LDInst2<(outs IntRegs:$dst), - (ins globaladdress:$global, u16Imm:$offset), - "$dst = memh(#$global+$offset)", - []>, - Requires<[NoV4T]>; - -let neverHasSideEffects = 1, validSubTargets = NoV4SubT in -def LDh_GP : LDInst2<(outs IntRegs:$dst), - (ins globaladdress:$global), - "$dst = memh(#$global)", - []>, - Requires<[NoV4T]>; - -let neverHasSideEffects = 1, validSubTargets = NoV4SubT in -def LDuh_GP : LDInst2<(outs IntRegs:$dst), - (ins globaladdress:$global), - "$dst = memuh(#$global)", - []>, - Requires<[NoV4T]>; - let AddedComplexity = 10 in def : Pat < (i32 (zextloadi1 ADDRriS11_0:$addr)), (i32 (LDriub ADDRriS11_0:$addr))>; @@ -1152,21 +1141,6 @@ let AddedComplexity = 20 in def : Pat < (i32 (zextloadi1 (add IntRegs:$src1, s11_0ImmPred:$offset))), (i32 (LDriub_indexed IntRegs:$src1, s11_0ImmPred:$offset))>; -let neverHasSideEffects = 1 in -def LDriub_GP : LDInst2<(outs IntRegs:$dst), - (ins globaladdress:$global, u16Imm:$offset), - "$dst = memub(#$global+$offset)", - []>, - Requires<[NoV4T]>; - -// Load unsigned halfword. -let neverHasSideEffects = 1 in -def LDriuh_GP : LDInst2<(outs IntRegs:$dst), - (ins globaladdress:$global, u16Imm:$offset), - "$dst = memuh(#$global+$offset)", - []>, - Requires<[NoV4T]>; - // Load predicate. let isExtendable = 1, opExtendable = 2, isExtentSigned = 1, opExtentBits = 13, isPseudo = 1, Defs = [R10,R11,D5], neverHasSideEffects = 1 in @@ -1175,21 +1149,6 @@ def LDriw_pred : LDInst2<(outs PredRegs:$dst), "Error; should not emit", []>; -// Indexed load. -let neverHasSideEffects = 1 in -def LDriw_GP : LDInst2<(outs IntRegs:$dst), - (ins globaladdress:$global, u16Imm:$offset), - "$dst = memw(#$global+$offset)", - []>, - Requires<[NoV4T]>; - -let neverHasSideEffects = 1, validSubTargets = NoV4SubT in -def LDw_GP : LDInst2<(outs IntRegs:$dst), - (ins globaladdress:$global), - "$dst = memw(#$global)", - []>, - Requires<[NoV4T]>; - // Deallocate stack frame. let Defs = [R29, R30, R31], Uses = [R29], neverHasSideEffects = 1 in { def DEALLOCFRAME : LDInst2<(outs), (ins), @@ -1423,35 +1382,15 @@ def SUBri_acc : MInst_acc<(outs IntRegs: $dst), (ins IntRegs:$src1, // ST + //===----------------------------------------------------------------------===// /// -/// Assumptions::: ****** DO NOT IGNORE ******** -/// 1. Make sure that in post increment store, the zero'th operand is always the -/// post increment operand. -/// 2. Make sure that the store value operand(Rt/Rtt) in a store is always the -/// last operand. -/// // Store doubleword. -let neverHasSideEffects = 1 in -def STrid_GP : STInst2<(outs), - (ins globaladdress:$global, u16Imm:$offset, DoubleRegs:$src), - "memd(#$global+$offset) = $src", - []>, - Requires<[NoV4T]>; - -let neverHasSideEffects = 1, validSubTargets = NoV4SubT in -def STd_GP : STInst2<(outs), - (ins globaladdress:$global, DoubleRegs:$src), - "memd(#$global) = $src", - []>, - Requires<[NoV4T]>; - //===----------------------------------------------------------------------===// // Post increment store //===----------------------------------------------------------------------===// multiclass ST_PostInc_Pbase<string mnemonic, RegisterClass RC, Operand ImmOp, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME : STInst2PI<(outs IntRegs:$dst), (ins PredRegs:$src1, IntRegs:$src2, ImmOp:$offset, RC:$src3), !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ", @@ -1462,8 +1401,8 @@ multiclass ST_PostInc_Pbase<string mnemonic, RegisterClass RC, Operand ImmOp, multiclass ST_PostInc_Pred<string mnemonic, RegisterClass RC, Operand ImmOp, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { - defm _c#NAME# : ST_PostInc_Pbase<mnemonic, RC, ImmOp, PredNot, 0>; + let isPredicatedFalse = PredNot in { + defm _c#NAME : ST_PostInc_Pbase<mnemonic, RC, ImmOp, PredNot, 0>; // Predicate new let Predicates = [HasV4T], validSubTargets = HasV4SubT in defm _cdn#NAME#_V4 : ST_PostInc_Pbase<mnemonic, RC, ImmOp, PredNot, 1>; @@ -1478,7 +1417,7 @@ multiclass ST_PostInc<string mnemonic, string BaseOp, RegisterClass RC, let isPredicable = 1 in def NAME : STInst2PI<(outs IntRegs:$dst), (ins IntRegs:$src1, ImmOp:$offset, RC:$src2), - #mnemonic#"($src1++#$offset) = $src2", + mnemonic#"($src1++#$offset) = $src2", [], "$src1 = $dst">; @@ -1516,7 +1455,7 @@ def : Pat<(post_store (i64 DoubleRegs:$src1), IntRegs:$src2, //===----------------------------------------------------------------------===// multiclass ST_MEMri_Pbase<string mnemonic, RegisterClass RC, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME : STInst2<(outs), (ins PredRegs:$src1, MEMri:$addr, RC: $src2), !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ", @@ -1525,7 +1464,7 @@ multiclass ST_MEMri_Pbase<string mnemonic, RegisterClass RC, bit isNot, } multiclass ST_MEMri_Pred<string mnemonic, RegisterClass RC, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { defm _c#NAME : ST_MEMri_Pbase<mnemonic, RC, PredNot, 0>; // Predicate new @@ -1555,12 +1494,17 @@ multiclass ST_MEMri<string mnemonic, string CextOp, RegisterClass RC, } let addrMode = BaseImmOffset, isMEMri = "true" in { - defm STrib: ST_MEMri < "memb", "STrib", IntRegs, 11, 6>, AddrModeRel; - defm STrih: ST_MEMri < "memh", "STrih", IntRegs, 12, 7>, AddrModeRel; - defm STriw: ST_MEMri < "memw", "STriw", IntRegs, 13, 8>, AddrModeRel; + let accessSize = ByteAccess in + defm STrib: ST_MEMri < "memb", "STrib", IntRegs, 11, 6>, AddrModeRel; - let isNVStorable = 0 in - defm STrid: ST_MEMri < "memd", "STrid", DoubleRegs, 14, 9>, AddrModeRel; + let accessSize = HalfWordAccess in + defm STrih: ST_MEMri < "memh", "STrih", IntRegs, 12, 7>, AddrModeRel; + + let accessSize = WordAccess in + defm STriw: ST_MEMri < "memw", "STriw", IntRegs, 13, 8>, AddrModeRel; + + let accessSize = DoubleWordAccess, isNVStorable = 0 in + defm STrid: ST_MEMri < "memd", "STrid", DoubleRegs, 14, 9>, AddrModeRel; } def : Pat<(truncstorei8 (i32 IntRegs:$src1), ADDRriS11_0:$addr), @@ -1582,7 +1526,7 @@ def : Pat<(store (i64 DoubleRegs:$src1), ADDRriS11_3:$addr), //===----------------------------------------------------------------------===// multiclass ST_Idxd_Pbase<string mnemonic, RegisterClass RC, Operand predImmOp, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME : STInst2<(outs), (ins PredRegs:$src1, IntRegs:$src2, predImmOp:$src3, RC: $src4), !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ", @@ -1592,7 +1536,7 @@ multiclass ST_Idxd_Pbase<string mnemonic, RegisterClass RC, Operand predImmOp, multiclass ST_Idxd_Pred<string mnemonic, RegisterClass RC, Operand predImmOp, bit PredNot> { - let PredSense = !if(PredNot, "false", "true"), isPredicated = 1 in { + let isPredicatedFalse = PredNot, isPredicated = 1 in { defm _c#NAME : ST_Idxd_Pbase<mnemonic, RC, predImmOp, PredNot, 0>; // Predicate new @@ -1622,15 +1566,21 @@ multiclass ST_Idxd<string mnemonic, string CextOp, RegisterClass RC, } let addrMode = BaseImmOffset, InputType = "reg" in { - defm STrib_indexed: ST_Idxd < "memb", "STrib", IntRegs, s11_0Ext, - u6_0Ext, 11, 6>, AddrModeRel, ImmRegRel; - defm STrih_indexed: ST_Idxd < "memh", "STrih", IntRegs, s11_1Ext, - u6_1Ext, 12, 7>, AddrModeRel, ImmRegRel; - defm STriw_indexed: ST_Idxd < "memw", "STriw", IntRegs, s11_2Ext, - u6_2Ext, 13, 8>, AddrModeRel, ImmRegRel; - let isNVStorable = 0 in - defm STrid_indexed: ST_Idxd < "memd", "STrid", DoubleRegs, s11_3Ext, - u6_3Ext, 14, 9>, AddrModeRel; + let accessSize = ByteAccess in + defm STrib_indexed: ST_Idxd < "memb", "STrib", IntRegs, s11_0Ext, + u6_0Ext, 11, 6>, AddrModeRel, ImmRegRel; + + let accessSize = HalfWordAccess in + defm STrih_indexed: ST_Idxd < "memh", "STrih", IntRegs, s11_1Ext, + u6_1Ext, 12, 7>, AddrModeRel, ImmRegRel; + + let accessSize = WordAccess in + defm STriw_indexed: ST_Idxd < "memw", "STriw", IntRegs, s11_2Ext, + u6_2Ext, 13, 8>, AddrModeRel, ImmRegRel; + + let accessSize = DoubleWordAccess, isNVStorable = 0 in + defm STrid_indexed: ST_Idxd < "memd", "STrid", DoubleRegs, s11_3Ext, + u6_3Ext, 14, 9>, AddrModeRel; } let AddedComplexity = 10 in { @@ -1655,36 +1605,6 @@ def : Pat<(store (i64 DoubleRegs:$src1), (add IntRegs:$src2, (i64 DoubleRegs:$src1))>; } -// memb(gp+#u16:0)=Rt -let neverHasSideEffects = 1 in -def STrib_GP : STInst2<(outs), - (ins globaladdress:$global, u16Imm:$offset, IntRegs:$src), - "memb(#$global+$offset) = $src", - []>, - Requires<[NoV4T]>; - -// memb(#global)=Rt -let neverHasSideEffects = 1, validSubTargets = NoV4SubT in -def STb_GP : STInst2<(outs), - (ins globaladdress:$global, IntRegs:$src), - "memb(#$global) = $src", - []>, - Requires<[NoV4T]>; - -let neverHasSideEffects = 1 in -def STrih_GP : STInst2<(outs), - (ins globaladdress:$global, u16Imm:$offset, IntRegs:$src), - "memh(#$global+$offset) = $src", - []>, - Requires<[NoV4T]>; - -let neverHasSideEffects = 1, validSubTargets = NoV4SubT in -def STh_GP : STInst2<(outs), - (ins globaladdress:$global, IntRegs:$src), - "memh(#$global) = $src", - []>, - Requires<[NoV4T]>; - // memh(Rx++#s4:1)=Rt.H // Store word. @@ -1695,20 +1615,6 @@ def STriw_pred : STInst2<(outs), "Error; should not emit", []>; -let neverHasSideEffects = 1 in -def STriw_GP : STInst2<(outs), - (ins globaladdress:$global, u16Imm:$offset, IntRegs:$src), - "memw(#$global+$offset) = $src", - []>, - Requires<[NoV4T]>; - -let neverHasSideEffects = 1, validSubTargets = NoV4SubT in -def STw_GP : STInst2<(outs), - (ins globaladdress:$global, IntRegs:$src), - "memw(#$global) = $src", - []>, - Requires<[NoV4T]>; - // Allocate stack frame. let Defs = [R29, R30], Uses = [R31, R30], neverHasSideEffects = 1 in { def ALLOCFRAME : STInst2<(outs), @@ -2152,20 +2058,18 @@ let isCall = 1, neverHasSideEffects = 1, []>; } -// Tail Calls. -let isCall = 1, isBarrier = 1, isReturn = 1, isTerminator = 1 in { - def TCRETURNtg : JInst<(outs), (ins calltarget:$dst), - "jump $dst // TAILCALL", []>; -} -let isCall = 1, isBarrier = 1, isReturn = 1, isTerminator = 1 in { - def TCRETURNtext : JInst<(outs), (ins calltarget:$dst), - "jump $dst // TAILCALL", []>; -} -let isCall = 1, isBarrier = 1, isReturn = 1, isTerminator = 1 in { - def TCRETURNR : JInst<(outs), (ins IntRegs:$dst), - "jumpr $dst // TAILCALL", []>; +// Indirect tail-call. +let isCodeGenOnly = 1, isCall = 1, isReturn = 1 in +def TCRETURNR : T_JMPr; + +// Direct tail-calls. +let isCall = 1, isReturn = 1, isBarrier = 1, isPredicable = 0, +isTerminator = 1, isCodeGenOnly = 1 in { + def TCRETURNtg : T_JMP<(ins calltarget:$dst)>; + def TCRETURNtext : T_JMP<(ins calltarget:$dst)>; } + // Map call instruction. def : Pat<(call (i32 IntRegs:$dst)), (CALLR (i32 IntRegs:$dst))>, Requires<[HasV2TOnly]>; @@ -2183,68 +2087,26 @@ def : Pat<(HexagonTCRet (i32 IntRegs:$dst)), // Atomic load and store support // 8 bit atomic load -def : Pat<(atomic_load_8 (HexagonCONST32_GP tglobaladdr:$global)), - (i32 (LDub_GP tglobaladdr:$global))>, - Requires<[NoV4T]>; - -def : Pat<(atomic_load_8 (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset)), - (i32 (LDriub_GP tglobaladdr:$global, u16ImmPred:$offset))>, - Requires<[NoV4T]>; - def : Pat<(atomic_load_8 ADDRriS11_0:$src1), (i32 (LDriub ADDRriS11_0:$src1))>; def : Pat<(atomic_load_8 (add (i32 IntRegs:$src1), s11_0ImmPred:$offset)), (i32 (LDriub_indexed (i32 IntRegs:$src1), s11_0ImmPred:$offset))>; - - // 16 bit atomic load -def : Pat<(atomic_load_16 (HexagonCONST32_GP tglobaladdr:$global)), - (i32 (LDuh_GP tglobaladdr:$global))>, - Requires<[NoV4T]>; - -def : Pat<(atomic_load_16 (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset)), - (i32 (LDriuh_GP tglobaladdr:$global, u16ImmPred:$offset))>, - Requires<[NoV4T]>; - def : Pat<(atomic_load_16 ADDRriS11_1:$src1), (i32 (LDriuh ADDRriS11_1:$src1))>; def : Pat<(atomic_load_16 (add (i32 IntRegs:$src1), s11_1ImmPred:$offset)), (i32 (LDriuh_indexed (i32 IntRegs:$src1), s11_1ImmPred:$offset))>; - - -// 32 bit atomic load -def : Pat<(atomic_load_32 (HexagonCONST32_GP tglobaladdr:$global)), - (i32 (LDw_GP tglobaladdr:$global))>, - Requires<[NoV4T]>; - -def : Pat<(atomic_load_32 (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset)), - (i32 (LDriw_GP tglobaladdr:$global, u16ImmPred:$offset))>, - Requires<[NoV4T]>; - def : Pat<(atomic_load_32 ADDRriS11_2:$src1), (i32 (LDriw ADDRriS11_2:$src1))>; def : Pat<(atomic_load_32 (add (i32 IntRegs:$src1), s11_2ImmPred:$offset)), (i32 (LDriw_indexed (i32 IntRegs:$src1), s11_2ImmPred:$offset))>; - // 64 bit atomic load -def : Pat<(atomic_load_64 (HexagonCONST32_GP tglobaladdr:$global)), - (i64 (LDd_GP tglobaladdr:$global))>, - Requires<[NoV4T]>; - -def : Pat<(atomic_load_64 (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset)), - (i64 (LDrid_GP tglobaladdr:$global, u16ImmPred:$offset))>, - Requires<[NoV4T]>; - def : Pat<(atomic_load_64 ADDRriS11_3:$src1), (i64 (LDrid ADDRriS11_3:$src1))>; @@ -2252,30 +2114,6 @@ def : Pat<(atomic_load_64 (add (i32 IntRegs:$src1), s11_3ImmPred:$offset)), (i64 (LDrid_indexed (i32 IntRegs:$src1), s11_3ImmPred:$offset))>; -// 64 bit atomic store -def : Pat<(atomic_store_64 (HexagonCONST32_GP tglobaladdr:$global), - (i64 DoubleRegs:$src1)), - (STd_GP tglobaladdr:$global, (i64 DoubleRegs:$src1))>, - Requires<[NoV4T]>; - -def : Pat<(atomic_store_64 (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset), - (i64 DoubleRegs:$src1)), - (STrid_GP tglobaladdr:$global, u16ImmPred:$offset, - (i64 DoubleRegs:$src1))>, Requires<[NoV4T]>; - -// 8 bit atomic store -def : Pat<(atomic_store_8 (HexagonCONST32_GP tglobaladdr:$global), - (i32 IntRegs:$src1)), - (STb_GP tglobaladdr:$global, (i32 IntRegs:$src1))>, - Requires<[NoV4T]>; - -def : Pat<(atomic_store_8 (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset), - (i32 IntRegs:$src1)), - (STrib_GP tglobaladdr:$global, u16ImmPred:$offset, - (i32 IntRegs:$src1))>, Requires<[NoV4T]>; - def : Pat<(atomic_store_8 ADDRriS11_0:$src2, (i32 IntRegs:$src1)), (STrib ADDRriS11_0:$src2, (i32 IntRegs:$src1))>; @@ -2285,18 +2123,6 @@ def : Pat<(atomic_store_8 (add (i32 IntRegs:$src2), s11_0ImmPred:$offset), (i32 IntRegs:$src1))>; -// 16 bit atomic store -def : Pat<(atomic_store_16 (HexagonCONST32_GP tglobaladdr:$global), - (i32 IntRegs:$src1)), - (STh_GP tglobaladdr:$global, (i32 IntRegs:$src1))>, - Requires<[NoV4T]>; - -def : Pat<(atomic_store_16 (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset), - (i32 IntRegs:$src1)), - (STrih_GP tglobaladdr:$global, u16ImmPred:$offset, - (i32 IntRegs:$src1))>, Requires<[NoV4T]>; - def : Pat<(atomic_store_16 ADDRriS11_1:$src2, (i32 IntRegs:$src1)), (STrih ADDRriS11_1:$src2, (i32 IntRegs:$src1))>; @@ -2305,20 +2131,6 @@ def : Pat<(atomic_store_16 (i32 IntRegs:$src1), (STrih_indexed (i32 IntRegs:$src2), s11_1ImmPred:$offset, (i32 IntRegs:$src1))>; - -// 32 bit atomic store -def : Pat<(atomic_store_32 (HexagonCONST32_GP tglobaladdr:$global), - (i32 IntRegs:$src1)), - (STw_GP tglobaladdr:$global, (i32 IntRegs:$src1))>, - Requires<[NoV4T]>; - -def : Pat<(atomic_store_32 (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset), - (i32 IntRegs:$src1)), - (STriw_GP tglobaladdr:$global, u16ImmPred:$offset, - (i32 IntRegs:$src1))>, - Requires<[NoV4T]>; - def : Pat<(atomic_store_32 ADDRriS11_2:$src2, (i32 IntRegs:$src1)), (STriw ADDRriS11_2:$src2, (i32 IntRegs:$src1))>; @@ -2354,10 +2166,11 @@ def : Pat <(add (i1 PredRegs:$src1), -1), // Map from p0 = setlt(r0, r1) r2 = mux(p0, r3, r4) => // p0 = cmp.lt(r0, r1), r0 = mux(p0, r2, r1). +// cmp.lt(r0, r1) -> cmp.gt(r1, r0) def : Pat <(select (i1 (setlt (i32 IntRegs:$src1), (i32 IntRegs:$src2))), (i32 IntRegs:$src3), (i32 IntRegs:$src4)), - (i32 (TFR_condset_rr (CMPLTrr (i32 IntRegs:$src1), (i32 IntRegs:$src2)), + (i32 (TFR_condset_rr (CMPGTrr (i32 IntRegs:$src2), (i32 IntRegs:$src1)), (i32 IntRegs:$src4), (i32 IntRegs:$src3)))>, Requires<[HasV2TOnly]>; @@ -2375,210 +2188,27 @@ def : Pat <(select (not (i1 PredRegs:$src1)), s12ImmPred:$src2, // Map from p0 = pnot(p0); r0 = mux(p0, r1, #i) // => r0 = TFR_condset_ir(p0, #i, r1) -def : Pat <(select (not PredRegs:$src1), IntRegs:$src2, s12ImmPred:$src3), +def : Pat <(select (not (i1 PredRegs:$src1)), IntRegs:$src2, s12ImmPred:$src3), (i32 (TFR_condset_ir (i1 PredRegs:$src1), s12ImmPred:$src3, (i32 IntRegs:$src2)))>; // Map from p0 = pnot(p0); if (p0) jump => if (!p0) jump. -def : Pat <(brcond (not PredRegs:$src1), bb:$offset), - (JMP_cNot (i1 PredRegs:$src1), bb:$offset)>; +def : Pat <(brcond (not (i1 PredRegs:$src1)), bb:$offset), + (JMP_f (i1 PredRegs:$src1), bb:$offset)>; // Map from p2 = pnot(p2); p1 = and(p0, p2) => p1 = and(p0, !p2). -def : Pat <(and PredRegs:$src1, (not PredRegs:$src2)), +def : Pat <(and (i1 PredRegs:$src1), (not (i1 PredRegs:$src2))), (i1 (AND_pnotp (i1 PredRegs:$src1), (i1 PredRegs:$src2)))>; -// Map from store(globaladdress + x) -> memd(#foo + x). -let AddedComplexity = 100 in -def : Pat <(store (i64 DoubleRegs:$src1), - (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset)), - (STrid_GP tglobaladdr:$global, u16ImmPred:$offset, - (i64 DoubleRegs:$src1))>, Requires<[NoV4T]>; - -// Map from store(globaladdress) -> memd(#foo). -let AddedComplexity = 100 in -def : Pat <(store (i64 DoubleRegs:$src1), - (HexagonCONST32_GP tglobaladdr:$global)), - (STd_GP tglobaladdr:$global, (i64 DoubleRegs:$src1))>, - Requires<[NoV4T]>; -// Map from store(globaladdress + x) -> memw(#foo + x). let AddedComplexity = 100 in -def : Pat <(store (i32 IntRegs:$src1), - (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset)), - (STriw_GP tglobaladdr:$global, u16ImmPred:$offset, (i32 IntRegs:$src1))>, - Requires<[NoV4T]>; - -// Map from store(globaladdress) -> memw(#foo + 0). -let AddedComplexity = 100 in -def : Pat <(store (i32 IntRegs:$src1), (HexagonCONST32_GP tglobaladdr:$global)), - (STriw_GP tglobaladdr:$global, 0, (i32 IntRegs:$src1))>; - -// Map from store(globaladdress) -> memw(#foo). -let AddedComplexity = 100 in -def : Pat <(store (i32 IntRegs:$src1), (HexagonCONST32_GP tglobaladdr:$global)), - (STriw_GP tglobaladdr:$global, 0, (i32 IntRegs:$src1))>, - Requires<[NoV4T]>; - -// Map from store(globaladdress + x) -> memh(#foo + x). -let AddedComplexity = 100 in -def : Pat <(truncstorei16 (i32 IntRegs:$src1), - (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset)), - (STrih_GP tglobaladdr:$global, u16ImmPred:$offset, (i32 IntRegs:$src1))>, - Requires<[NoV4T]>; - -// Map from store(globaladdress) -> memh(#foo). -let AddedComplexity = 100 in -def : Pat <(truncstorei16 (i32 IntRegs:$src1), - (HexagonCONST32_GP tglobaladdr:$global)), - (STh_GP tglobaladdr:$global, (i32 IntRegs:$src1))>, - Requires<[NoV4T]>; - -// Map from store(globaladdress + x) -> memb(#foo + x). -let AddedComplexity = 100 in -def : Pat <(truncstorei8 (i32 IntRegs:$src1), - (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset)), - (STrib_GP tglobaladdr:$global, u16ImmPred:$offset, (i32 IntRegs:$src1))>, - Requires<[NoV4T]>; - -// Map from store(globaladdress) -> memb(#foo). -let AddedComplexity = 100 in -def : Pat <(truncstorei8 (i32 IntRegs:$src1), - (HexagonCONST32_GP tglobaladdr:$global)), - (STb_GP tglobaladdr:$global, (i32 IntRegs:$src1))>, - Requires<[NoV4T]>; - -// Map from load(globaladdress + x) -> memw(#foo + x). -let AddedComplexity = 100 in -def : Pat <(i32 (load (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset))), - (i32 (LDriw_GP tglobaladdr:$global, u16ImmPred:$offset))>, - Requires<[NoV4T]>; - -// Map from load(globaladdress) -> memw(#foo). -let AddedComplexity = 100 in -def : Pat <(i32 (load (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDw_GP tglobaladdr:$global))>, - Requires<[NoV4T]>; - -// Map from load(globaladdress + x) -> memd(#foo + x). -let AddedComplexity = 100 in -def : Pat <(i64 (load (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset))), - (i64 (LDrid_GP tglobaladdr:$global, u16ImmPred:$offset))>, - Requires<[NoV4T]>; - -// Map from load(globaladdress) -> memw(#foo + 0). -let AddedComplexity = 100 in -def : Pat <(i64 (load (HexagonCONST32_GP tglobaladdr:$global))), - (i64 (LDd_GP tglobaladdr:$global))>, - Requires<[NoV4T]>; - -// Map from Pd = load(globaladdress) -> Rd = memb(globaladdress), Pd = Rd. -let AddedComplexity = 100 in -def : Pat <(i1 (load (HexagonCONST32_GP tglobaladdr:$global))), - (i1 (TFR_PdRs (i32 (LDb_GP tglobaladdr:$global))))>, - Requires<[NoV4T]>; - -// Map from load(globaladdress + x) -> memh(#foo + x). -let AddedComplexity = 100 in -def : Pat <(i32 (extloadi16 (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset))), - (i32 (LDrih_GP tglobaladdr:$global, u16ImmPred:$offset))>, - Requires<[NoV4T]>; - -// Map from load(globaladdress + x) -> memh(#foo + x). -let AddedComplexity = 100 in -def : Pat <(i32 (sextloadi16 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDrih_GP tglobaladdr:$global, 0))>, - Requires<[NoV4T]>; - -// Map from load(globaladdress + x) -> memuh(#foo + x). -let AddedComplexity = 100 in -def : Pat <(i32 (zextloadi16 (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset))), - (i32 (LDriuh_GP tglobaladdr:$global, u16ImmPred:$offset))>, - Requires<[NoV4T]>; - -// Map from load(globaladdress) -> memuh(#foo). -let AddedComplexity = 100 in -def : Pat <(i32 (zextloadi16 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDriuh_GP tglobaladdr:$global, 0))>, - Requires<[NoV4T]>; - -// Map from load(globaladdress) -> memh(#foo). -let AddedComplexity = 100 in -def : Pat <(i32 (sextloadi16 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDh_GP tglobaladdr:$global))>, - Requires<[NoV4T]>; - -// Map from load(globaladdress) -> memuh(#foo). -let AddedComplexity = 100 in -def : Pat <(i32 (zextloadi16 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDuh_GP tglobaladdr:$global))>, - Requires<[NoV4T]>; - -// Map from load(globaladdress + x) -> memb(#foo + x). -let AddedComplexity = 100 in -def : Pat <(i32 (extloadi8 (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset))), - (i32 (LDrib_GP tglobaladdr:$global, u16ImmPred:$offset))>, - Requires<[NoV4T]>; - -// Map from load(globaladdress + x) -> memb(#foo + x). -let AddedComplexity = 100 in -def : Pat <(i32 (sextloadi8 (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset))), - (i32 (LDrib_GP tglobaladdr:$global, u16ImmPred:$offset))>, - Requires<[NoV4T]>; - -// Map from load(globaladdress + x) -> memub(#foo + x). -let AddedComplexity = 100 in -def : Pat <(i32 (zextloadi8 (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset))), - (i32 (LDriub_GP tglobaladdr:$global, u16ImmPred:$offset))>, - Requires<[NoV4T]>; - -// Map from load(globaladdress) -> memb(#foo). -let AddedComplexity = 100 in -def : Pat <(i32 (extloadi8 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDb_GP tglobaladdr:$global))>, - Requires<[NoV4T]>; - -// Map from load(globaladdress) -> memb(#foo). -let AddedComplexity = 100 in -def : Pat <(i32 (sextloadi8 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDb_GP tglobaladdr:$global))>, - Requires<[NoV4T]>; - -// Map from load(globaladdress) -> memub(#foo). -let AddedComplexity = 100 in -def : Pat <(i32 (zextloadi8 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDub_GP tglobaladdr:$global))>, - Requires<[NoV4T]>; - -// When the Interprocedural Global Variable optimizer realizes that a -// certain global variable takes only two constant values, it shrinks the -// global to a boolean. Catch those loads here in the following 3 patterns. -let AddedComplexity = 100 in -def : Pat <(i32 (extloadi1 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDb_GP tglobaladdr:$global))>, - Requires<[NoV4T]>; - -let AddedComplexity = 100 in -def : Pat <(i32 (sextloadi1 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDb_GP tglobaladdr:$global))>, - Requires<[NoV4T]>; - -let AddedComplexity = 100 in -def : Pat <(i32 (zextloadi1 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDub_GP tglobaladdr:$global))>, +def : Pat <(i64 (zextloadi1 (HexagonCONST32 tglobaladdr:$global))), + (i64 (COMBINE_rr (TFRI 0), + (LDriub_indexed (CONST32_set tglobaladdr:$global), 0)))>, Requires<[NoV4T]>; // Map from i1 loads to 32 bits. This assumes that the i1* is byte aligned. +let AddedComplexity = 10 in def : Pat <(i32 (zextloadi1 ADDRriS11_0:$addr)), (i32 (AND_rr (i32 (LDrib ADDRriS11_0:$addr)), (TFRI 0x1)))>; @@ -2597,43 +2227,46 @@ def : Pat <(i64 (sext_inreg (i64 DoubleRegs:$src1), i8)), subreg_loreg))))))>; // We want to prevent emitting pnot's as much as possible. -// Map brcond with an unsupported setcc to a JMP_cNot. +// Map brcond with an unsupported setcc to a JMP_f. def : Pat <(brcond (i1 (setne (i32 IntRegs:$src1), (i32 IntRegs:$src2))), bb:$offset), - (JMP_cNot (CMPEQrr (i32 IntRegs:$src1), (i32 IntRegs:$src2)), + (JMP_f (CMPEQrr (i32 IntRegs:$src1), (i32 IntRegs:$src2)), bb:$offset)>; def : Pat <(brcond (i1 (setne (i32 IntRegs:$src1), s10ImmPred:$src2)), bb:$offset), - (JMP_cNot (CMPEQri (i32 IntRegs:$src1), s10ImmPred:$src2), bb:$offset)>; + (JMP_f (CMPEQri (i32 IntRegs:$src1), s10ImmPred:$src2), bb:$offset)>; def : Pat <(brcond (i1 (setne (i1 PredRegs:$src1), (i1 -1))), bb:$offset), - (JMP_cNot (i1 PredRegs:$src1), bb:$offset)>; + (JMP_f (i1 PredRegs:$src1), bb:$offset)>; def : Pat <(brcond (i1 (setne (i1 PredRegs:$src1), (i1 0))), bb:$offset), - (JMP_c (i1 PredRegs:$src1), bb:$offset)>; + (JMP_t (i1 PredRegs:$src1), bb:$offset)>; +// cmp.lt(Rs, Imm) -> !cmp.ge(Rs, Imm) -> !cmp.gt(Rs, Imm-1) def : Pat <(brcond (i1 (setlt (i32 IntRegs:$src1), s8ImmPred:$src2)), bb:$offset), - (JMP_cNot (CMPGEri (i32 IntRegs:$src1), s8ImmPred:$src2), bb:$offset)>; + (JMP_f (CMPGTri (i32 IntRegs:$src1), + (DEC_CONST_SIGNED s8ImmPred:$src2)), bb:$offset)>; +// cmp.lt(r0, r1) -> cmp.gt(r1, r0) def : Pat <(brcond (i1 (setlt (i32 IntRegs:$src1), (i32 IntRegs:$src2))), bb:$offset), - (JMP_c (CMPLTrr (i32 IntRegs:$src1), (i32 IntRegs:$src2)), bb:$offset)>; + (JMP_t (CMPGTrr (i32 IntRegs:$src2), (i32 IntRegs:$src1)), bb:$offset)>; def : Pat <(brcond (i1 (setuge (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2))), bb:$offset), - (JMP_cNot (CMPGTU64rr (i64 DoubleRegs:$src2), (i64 DoubleRegs:$src1)), + (JMP_f (CMPGTU64rr (i64 DoubleRegs:$src2), (i64 DoubleRegs:$src1)), bb:$offset)>; def : Pat <(brcond (i1 (setule (i32 IntRegs:$src1), (i32 IntRegs:$src2))), bb:$offset), - (JMP_cNot (CMPGTUrr (i32 IntRegs:$src1), (i32 IntRegs:$src2)), + (JMP_f (CMPGTUrr (i32 IntRegs:$src1), (i32 IntRegs:$src2)), bb:$offset)>; def : Pat <(brcond (i1 (setule (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2))), bb:$offset), - (JMP_cNot (CMPGTU64rr (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2)), + (JMP_f (CMPGTU64rr (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2)), bb:$offset)>; // Map from a 64-bit select to an emulated 64-bit mux. @@ -2694,12 +2327,6 @@ def : Pat<(truncstorei32 (i64 DoubleRegs:$src), ADDRriS11_0:$addr), def : Pat<(store (i1 -1), ADDRriS11_2:$addr), (STrib ADDRriS11_2:$addr, (TFRI 1))>; -let AddedComplexity = 100 in -// Map from i1 = constant<-1>; memw(CONST32(#foo)) = i1 -> r0 = 1; -// memw(#foo) = r0 -def : Pat<(store (i1 -1), (HexagonCONST32_GP tglobaladdr:$global)), - (STb_GP tglobaladdr:$global, (TFRI 1))>, - Requires<[NoV4T]>; // Map from i1 = constant<-1>; store i1 -> r0 = 1; store r0. def : Pat<(store (i1 -1), ADDRriS11_2:$addr), @@ -2717,8 +2344,8 @@ def : Pat<(i64 (anyext (i32 IntRegs:$src1))), // Map cmple -> cmpgt. // rs <= rt -> !(rs > rt). -def : Pat<(i1 (setle (i32 IntRegs:$src1), s10ImmPred:$src2)), - (i1 (NOT_p (CMPGTri (i32 IntRegs:$src1), s10ImmPred:$src2)))>; +def : Pat<(i1 (setle (i32 IntRegs:$src1), s10ExtPred:$src2)), + (i1 (NOT_p (CMPGTri (i32 IntRegs:$src1), s10ExtPred:$src2)))>; // rs <= rt -> !(rs > rt). def : Pat<(i1 (setle (i32 IntRegs:$src1), (i32 IntRegs:$src2))), @@ -2731,8 +2358,8 @@ def : Pat<(i1 (setle (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2))), // Map cmpne -> cmpeq. // Hexagon_TODO: We should improve on this. // rs != rt -> !(rs == rt). -def : Pat <(i1 (setne (i32 IntRegs:$src1), s10ImmPred:$src2)), - (i1 (NOT_p(i1 (CMPEQri (i32 IntRegs:$src1), s10ImmPred:$src2))))>; +def : Pat <(i1 (setne (i32 IntRegs:$src1), s10ExtPred:$src2)), + (i1 (NOT_p(i1 (CMPEQri (i32 IntRegs:$src1), s10ExtPred:$src2))))>; // Map cmpne(Rs) -> !cmpeqe(Rs). // rs != rt -> !(rs == rt). @@ -2754,8 +2381,9 @@ def : Pat <(i1 (setne (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2))), def : Pat <(i1 (setge (i32 IntRegs:$src1), (i32 IntRegs:$src2))), (i1 (NOT_p (i1 (CMPGTrr (i32 IntRegs:$src2), (i32 IntRegs:$src1)))))>; -def : Pat <(i1 (setge (i32 IntRegs:$src1), s8ImmPred:$src2)), - (i1 (CMPGEri (i32 IntRegs:$src1), s8ImmPred:$src2))>; +// cmpge(Rs, Imm) -> cmpgt(Rs, Imm-1) +def : Pat <(i1 (setge (i32 IntRegs:$src1), s8ExtPred:$src2)), + (i1 (CMPGTri (i32 IntRegs:$src1), (DEC_CONST_SIGNED s8ExtPred:$src2)))>; // Map cmpge(Rss, Rtt) -> !cmpgt(Rtt, Rss). // rss >= rtt -> !(rtt > rss). @@ -2764,9 +2392,10 @@ def : Pat <(i1 (setge (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2))), (i64 DoubleRegs:$src1)))))>; // Map cmplt(Rs, Imm) -> !cmpge(Rs, Imm). +// !cmpge(Rs, Imm) -> !cmpgt(Rs, Imm-1). // rs < rt -> !(rs >= rt). -def : Pat <(i1 (setlt (i32 IntRegs:$src1), s8ImmPred:$src2)), - (i1 (NOT_p (CMPGEri (i32 IntRegs:$src1), s8ImmPred:$src2)))>; +def : Pat <(i1 (setlt (i32 IntRegs:$src1), s8ExtPred:$src2)), + (i1 (NOT_p (CMPGTri (i32 IntRegs:$src1), (DEC_CONST_SIGNED s8ExtPred:$src2))))>; // Map cmplt(Rs, Rt) -> cmpgt(Rt, Rs). // rs < rt -> rt > rs. @@ -2790,13 +2419,17 @@ def : Pat <(i1 (setult (i32 IntRegs:$src1), (i32 IntRegs:$src2))), def : Pat <(i1 (setult (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2))), (i1 (CMPGTU64rr (i64 DoubleRegs:$src2), (i64 DoubleRegs:$src1)))>; -// Generate cmpgeu(Rs, #u8) -def : Pat <(i1 (setuge (i32 IntRegs:$src1), u8ImmPred:$src2)), - (i1 (CMPGEUri (i32 IntRegs:$src1), u8ImmPred:$src2))>; +// Generate cmpgeu(Rs, #0) -> cmpeq(Rs, Rs) +def : Pat <(i1 (setuge (i32 IntRegs:$src1), 0)), + (i1 (CMPEQrr (i32 IntRegs:$src1), (i32 IntRegs:$src1)))>; + +// Generate cmpgeu(Rs, #u8) -> cmpgtu(Rs, #u8 -1) +def : Pat <(i1 (setuge (i32 IntRegs:$src1), u8ExtPred:$src2)), + (i1 (CMPGTUri (i32 IntRegs:$src1), (DEC_CONST_UNSIGNED u8ExtPred:$src2)))>; // Generate cmpgtu(Rs, #u9) -def : Pat <(i1 (setugt (i32 IntRegs:$src1), u9ImmPred:$src2)), - (i1 (CMPGTUri (i32 IntRegs:$src1), u9ImmPred:$src2))>; +def : Pat <(i1 (setugt (i32 IntRegs:$src1), u9ExtPred:$src2)), + (i1 (CMPGTUri (i32 IntRegs:$src1), u9ExtPred:$src2))>; // Map from Rs >= Rt -> !(Rt > Rs). // rs >= rt -> !(rt > rs). @@ -2808,7 +2441,7 @@ def : Pat <(i1 (setuge (i32 IntRegs:$src1), (i32 IntRegs:$src2))), def : Pat <(i1 (setuge (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2))), (i1 (NOT_p (CMPGTU64rr (i64 DoubleRegs:$src2), (i64 DoubleRegs:$src1))))>; -// Map from cmpleu(Rs, Rs) -> !cmpgtu(Rs, Rs). +// Map from cmpleu(Rs, Rt) -> !cmpgtu(Rs, Rt). // Map from (Rs <= Rt) -> !(Rs > Rt). def : Pat <(i1 (setule (i32 IntRegs:$src1), (i32 IntRegs:$src2))), (i1 (NOT_p (CMPGTUrr (i32 IntRegs:$src1), (i32 IntRegs:$src2))))>; @@ -2904,6 +2537,13 @@ def: Pat <(i64 (zextloadi32 ADDRriS11_2:$src1)), (i64 (COMBINE_rr (TFRI 0), (LDriw ADDRriS11_2:$src1)))>, Requires<[NoV4T]>; +let AddedComplexity = 100 in +def: Pat <(i64 (zextloadi32 (i32 (add IntRegs:$src1, s11_2ExtPred:$offset)))), + (i64 (COMBINE_rr (TFRI 0), (LDriw_indexed IntRegs:$src1, + s11_2ExtPred:$offset)))>, + Requires<[NoV4T]>; + +let AddedComplexity = 10 in def: Pat <(i32 (zextloadi1 ADDRriS11_0:$src1)), (i32 (LDriw ADDRriS11_0:$src1))>; @@ -2920,6 +2560,48 @@ def : Pat <(i64 (anyext (i1 PredRegs:$src1))), (i64 (SXTW (i32 (MUX_ii (i1 PredRegs:$src1), 1, 0))))>; +let AddedComplexity = 100 in +def: Pat<(i64 (or (i64 (shl (i64 DoubleRegs:$srcHigh), + (i32 32))), + (i64 (zextloadi32 (i32 (add IntRegs:$src2, + s11_2ExtPred:$offset2)))))), + (i64 (COMBINE_rr (EXTRACT_SUBREG (i64 DoubleRegs:$srcHigh), subreg_loreg), + (LDriw_indexed IntRegs:$src2, + s11_2ExtPred:$offset2)))>; + +def: Pat<(i64 (or (i64 (shl (i64 DoubleRegs:$srcHigh), + (i32 32))), + (i64 (zextloadi32 ADDRriS11_2:$srcLow)))), + (i64 (COMBINE_rr (EXTRACT_SUBREG (i64 DoubleRegs:$srcHigh), subreg_loreg), + (LDriw ADDRriS11_2:$srcLow)))>; + +def: Pat<(i64 (or (i64 (shl (i64 DoubleRegs:$srcHigh), + (i32 32))), + (i64 (zext (i32 IntRegs:$srcLow))))), + (i64 (COMBINE_rr (EXTRACT_SUBREG (i64 DoubleRegs:$srcHigh), subreg_loreg), + IntRegs:$srcLow))>; + +let AddedComplexity = 100 in +def: Pat<(i64 (or (i64 (shl (i64 DoubleRegs:$srcHigh), + (i32 32))), + (i64 (zextloadi32 (i32 (add IntRegs:$src2, + s11_2ExtPred:$offset2)))))), + (i64 (COMBINE_rr (EXTRACT_SUBREG (i64 DoubleRegs:$srcHigh), subreg_loreg), + (LDriw_indexed IntRegs:$src2, + s11_2ExtPred:$offset2)))>; + +def: Pat<(i64 (or (i64 (shl (i64 DoubleRegs:$srcHigh), + (i32 32))), + (i64 (zextloadi32 ADDRriS11_2:$srcLow)))), + (i64 (COMBINE_rr (EXTRACT_SUBREG (i64 DoubleRegs:$srcHigh), subreg_loreg), + (LDriw ADDRriS11_2:$srcLow)))>; + +def: Pat<(i64 (or (i64 (shl (i64 DoubleRegs:$srcHigh), + (i32 32))), + (i64 (zext (i32 IntRegs:$srcLow))))), + (i64 (COMBINE_rr (EXTRACT_SUBREG (i64 DoubleRegs:$srcHigh), subreg_loreg), + IntRegs:$srcLow))>; + // Any extended 64-bit load. // anyext i32 -> i64 def: Pat <(i64 (extloadi32 ADDRriS11_2:$src1)), @@ -3054,19 +2736,6 @@ let AddedComplexity = 100 in def : Pat<(i32 (sext_inreg (Hexagon_ARGEXTEND (i32 IntRegs:$src1)), i16)), (COPY (i32 IntRegs:$src1))>; -def SDHexagonBR_JT: SDTypeProfile<0, 1, [SDTCisVT<0, i32>]>; -def HexagonBR_JT: SDNode<"HexagonISD::BR_JT", SDHexagonBR_JT, [SDNPHasChain]>; - -let isBranch=1, isIndirectBranch=1, isTerminator=1, isBarrier = 1 in -def BR_JT : JRInst<(outs), (ins IntRegs:$src), - "jumpr $src", - [(HexagonBR_JT (i32 IntRegs:$src))]>; - -let isBranch=1, isIndirectBranch=1, isTerminator=1 in -def BRIND : JRInst<(outs), (ins IntRegs:$src), - "jumpr $src", - [(brind (i32 IntRegs:$src))]>; - def HexagonWrapperJT: SDNode<"HexagonISD::WrapperJT", SDTIntUnaryOp>; def : Pat<(HexagonWrapperJT tjumptable:$dst), diff --git a/lib/Target/Hexagon/HexagonInstrInfoV3.td b/lib/Target/Hexagon/HexagonInstrInfoV3.td index 157ab3d..7e75554 100644 --- a/lib/Target/Hexagon/HexagonInstrInfoV3.td +++ b/lib/Target/Hexagon/HexagonInstrInfoV3.td @@ -11,6 +11,11 @@ // //===----------------------------------------------------------------------===// +def callv3 : SDNode<"HexagonISD::CALLv3", SDT_SPCall, + [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue, SDNPVariadic]>; + +def callv3nr : SDNode<"HexagonISD::CALLv3nr", SDT_SPCall, + [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue, SDNPVariadic]>; //===----------------------------------------------------------------------===// // J + @@ -40,41 +45,6 @@ let isCall = 1, neverHasSideEffects = 1, []>, Requires<[HasV3TOnly]>; } - -// Jump to address from register -// if(p?.new) jumpr:t r? -let isReturn = 1, isTerminator = 1, isBarrier = 1, - Defs = [PC], Uses = [R31] in { - def JMPR_cdnPt_V3: JRInst<(outs), (ins PredRegs:$src1, IntRegs:$src2), - "if ($src1.new) jumpr:t $src2", - []>, Requires<[HasV3T]>; -} - -// if (!p?.new) jumpr:t r? -let isReturn = 1, isTerminator = 1, isBarrier = 1, - Defs = [PC], Uses = [R31] in { - def JMPR_cdnNotPt_V3: JRInst<(outs), (ins PredRegs:$src1, IntRegs:$src2), - "if (!$src1.new) jumpr:t $src2", - []>, Requires<[HasV3T]>; -} - -// Not taken. -// if(p?.new) jumpr:nt r? -let isReturn = 1, isTerminator = 1, isBarrier = 1, - Defs = [PC], Uses = [R31] in { - def JMPR_cdnPnt: JRInst<(outs), (ins PredRegs:$src1, IntRegs:$src2), - "if ($src1.new) jumpr:nt $src2", - []>, Requires<[HasV3T]>; -} - -// if (!p?.new) jumpr:nt r? -let isReturn = 1, isTerminator = 1, isBarrier = 1, - Defs = [PC], Uses = [R31] in { - def JMPR_cdnNotPnt: JRInst<(outs), (ins PredRegs:$src1, IntRegs:$src2), - "if (!$src1.new) jumpr:nt $src2", - []>, Requires<[HasV3T]>; -} - //===----------------------------------------------------------------------===// // JR - //===----------------------------------------------------------------------===// diff --git a/lib/Target/Hexagon/HexagonInstrInfoV4.td b/lib/Target/Hexagon/HexagonInstrInfoV4.td index 1d0643d..fee83fb 100644 --- a/lib/Target/Hexagon/HexagonInstrInfoV4.td +++ b/lib/Target/Hexagon/HexagonInstrInfoV4.td @@ -95,164 +95,6 @@ def NumUsesBelowThresCONST32 : PatFrag<(ops node:$addr), //===----------------------------------------------------------------------===// // ALU32 + //===----------------------------------------------------------------------===// - -// Shift halfword. - -let isPredicated = 1, neverHasSideEffects = 1, validSubTargets = HasV4SubT in { -def ASLH_cPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if ($src1) $dst = aslh($src2)", - []>, - Requires<[HasV4T]>; - -def ASLH_cNotPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if (!$src1) $dst = aslh($src2)", - []>, - Requires<[HasV4T]>; - -def ASLH_cdnPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if ($src1.new) $dst = aslh($src2)", - []>, - Requires<[HasV4T]>; - -def ASLH_cdnNotPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if (!$src1.new) $dst = aslh($src2)", - []>, - Requires<[HasV4T]>; - -def ASRH_cPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if ($src1) $dst = asrh($src2)", - []>, - Requires<[HasV4T]>; - -def ASRH_cNotPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if (!$src1) $dst = asrh($src2)", - []>, - Requires<[HasV4T]>; - -def ASRH_cdnPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if ($src1.new) $dst = asrh($src2)", - []>, - Requires<[HasV4T]>; - -def ASRH_cdnNotPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if (!$src1.new) $dst = asrh($src2)", - []>, - Requires<[HasV4T]>; -} - -// Sign extend. - -let isPredicated = 1, neverHasSideEffects = 1, validSubTargets = HasV4SubT in { -def SXTB_cPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if ($src1) $dst = sxtb($src2)", - []>, - Requires<[HasV4T]>; - -def SXTB_cNotPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if (!$src1) $dst = sxtb($src2)", - []>, - Requires<[HasV4T]>; - -def SXTB_cdnPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if ($src1.new) $dst = sxtb($src2)", - []>, - Requires<[HasV4T]>; - -def SXTB_cdnNotPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if (!$src1.new) $dst = sxtb($src2)", - []>, - Requires<[HasV4T]>; - - -def SXTH_cPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if ($src1) $dst = sxth($src2)", - []>, - Requires<[HasV4T]>; - -def SXTH_cNotPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if (!$src1) $dst = sxth($src2)", - []>, - Requires<[HasV4T]>; - -def SXTH_cdnPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if ($src1.new) $dst = sxth($src2)", - []>, - Requires<[HasV4T]>; - -def SXTH_cdnNotPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if (!$src1.new) $dst = sxth($src2)", - []>, - Requires<[HasV4T]>; -} - -// Zero exten. - -let neverHasSideEffects = 1, isPredicated = 1, validSubTargets = HasV4SubT in { -def ZXTB_cPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if ($src1) $dst = zxtb($src2)", - []>, - Requires<[HasV4T]>; - -def ZXTB_cNotPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if (!$src1) $dst = zxtb($src2)", - []>, - Requires<[HasV4T]>; - -def ZXTB_cdnPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if ($src1.new) $dst = zxtb($src2)", - []>, - Requires<[HasV4T]>; - -def ZXTB_cdnNotPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if (!$src1.new) $dst = zxtb($src2)", - []>, - Requires<[HasV4T]>; - -def ZXTH_cPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if ($src1) $dst = zxth($src2)", - []>, - Requires<[HasV4T]>; - -def ZXTH_cNotPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if (!$src1) $dst = zxth($src2)", - []>, - Requires<[HasV4T]>; - -def ZXTH_cdnPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if ($src1.new) $dst = zxth($src2)", - []>, - Requires<[HasV4T]>; - -def ZXTH_cdnNotPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if (!$src1.new) $dst = zxth($src2)", - []>, - Requires<[HasV4T]>; -} - // Generate frame index addresses. let neverHasSideEffects = 1, isReMaterializable = 1, isExtended = 1, opExtendable = 2, validSubTargets = HasV4SubT in @@ -367,105 +209,31 @@ def COMBINE_iI_V4 : ALU32_ii<(outs DoubleRegs:$dst), //===----------------------------------------------------------------------===// // LD + //===----------------------------------------------------------------------===// -// -// These absolute set addressing mode instructions accept immediate as -// an operand. We have duplicated these patterns to take global address. - +//===----------------------------------------------------------------------===// +// Template class for load instructions with Absolute set addressing mode. +//===----------------------------------------------------------------------===// let isExtended = 1, opExtendable = 2, neverHasSideEffects = 1, -validSubTargets = HasV4SubT in { -def LDrid_abs_setimm_V4 : LDInst2<(outs DoubleRegs:$dst1, IntRegs:$dst2), +validSubTargets = HasV4SubT, addrMode = AbsoluteSet in +class T_LD_abs_set<string mnemonic, RegisterClass RC>: + LDInst2<(outs RC:$dst1, IntRegs:$dst2), (ins u0AlwaysExt:$addr), - "$dst1 = memd($dst2=##$addr)", + "$dst1 = "#mnemonic#"($dst2=##$addr)", []>, Requires<[HasV4T]>; -// Rd=memb(Re=#U6) -def LDrib_abs_setimm_V4 : LDInst2<(outs IntRegs:$dst1, IntRegs:$dst2), - (ins u0AlwaysExt:$addr), - "$dst1 = memb($dst2=##$addr)", - []>, - Requires<[HasV4T]>; - -// Rd=memh(Re=#U6) -def LDrih_abs_setimm_V4 : LDInst2<(outs IntRegs:$dst1, IntRegs:$dst2), - (ins u0AlwaysExt:$addr), - "$dst1 = memh($dst2=##$addr)", - []>, - Requires<[HasV4T]>; - -// Rd=memub(Re=#U6) -def LDriub_abs_setimm_V4 : LDInst2<(outs IntRegs:$dst1, IntRegs:$dst2), - (ins u0AlwaysExt:$addr), - "$dst1 = memub($dst2=##$addr)", - []>, - Requires<[HasV4T]>; - -// Rd=memuh(Re=#U6) -def LDriuh_abs_setimm_V4 : LDInst2<(outs IntRegs:$dst1, IntRegs:$dst2), - (ins u0AlwaysExt:$addr), - "$dst1 = memuh($dst2=##$addr)", - []>, - Requires<[HasV4T]>; - -// Rd=memw(Re=#U6) -def LDriw_abs_setimm_V4 : LDInst2<(outs IntRegs:$dst1, IntRegs:$dst2), - (ins u0AlwaysExt:$addr), - "$dst1 = memw($dst2=##$addr)", - []>, - Requires<[HasV4T]>; -} - -// Following patterns are defined for absolute set addressing mode -// instruction which take global address as operand. -let isExtended = 1, opExtendable = 2, neverHasSideEffects = 1, -validSubTargets = HasV4SubT in { -def LDrid_abs_set_V4 : LDInst2<(outs DoubleRegs:$dst1, IntRegs:$dst2), - (ins globaladdressExt:$addr), - "$dst1 = memd($dst2=##$addr)", - []>, - Requires<[HasV4T]>; +def LDrid_abs_set_V4 : T_LD_abs_set <"memd", DoubleRegs>; +def LDrib_abs_set_V4 : T_LD_abs_set <"memb", IntRegs>; +def LDriub_abs_set_V4 : T_LD_abs_set <"memub", IntRegs>; +def LDrih_abs_set_V4 : T_LD_abs_set <"memh", IntRegs>; +def LDriw_abs_set_V4 : T_LD_abs_set <"memw", IntRegs>; +def LDriuh_abs_set_V4 : T_LD_abs_set <"memuh", IntRegs>; -// Rd=memb(Re=#U6) -def LDrib_abs_set_V4 : LDInst2<(outs IntRegs:$dst1, IntRegs:$dst2), - (ins globaladdressExt:$addr), - "$dst1 = memb($dst2=##$addr)", - []>, - Requires<[HasV4T]>; - -// Rd=memh(Re=#U6) -def LDrih_abs_set_V4 : LDInst2<(outs IntRegs:$dst1, IntRegs:$dst2), - (ins globaladdressExt:$addr), - "$dst1 = memh($dst2=##$addr)", - []>, - Requires<[HasV4T]>; - -// Rd=memub(Re=#U6) -def LDriub_abs_set_V4 : LDInst2<(outs IntRegs:$dst1, IntRegs:$dst2), - (ins globaladdressExt:$addr), - "$dst1 = memub($dst2=##$addr)", - []>, - Requires<[HasV4T]>; - -// Rd=memuh(Re=#U6) -def LDriuh_abs_set_V4 : LDInst2<(outs IntRegs:$dst1, IntRegs:$dst2), - (ins globaladdressExt:$addr), - "$dst1 = memuh($dst2=##$addr)", - []>, - Requires<[HasV4T]>; - -// Rd=memw(Re=#U6) -def LDriw_abs_set_V4 : LDInst2<(outs IntRegs:$dst1, IntRegs:$dst2), - (ins globaladdressExt:$addr), - "$dst1 = memw($dst2=##$addr)", - []>, - Requires<[HasV4T]>; -} // multiclass for load instructions with base + register offset // addressing mode multiclass ld_idxd_shl_pbase<string mnemonic, RegisterClass RC, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME : LDInst2<(outs RC:$dst), (ins PredRegs:$src1, IntRegs:$src2, IntRegs:$src3, u2Imm:$offset), !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ", @@ -474,7 +242,7 @@ multiclass ld_idxd_shl_pbase<string mnemonic, RegisterClass RC, bit isNot, } multiclass ld_idxd_shl_pred<string mnemonic, RegisterClass RC, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { defm _c#NAME : ld_idxd_shl_pbase<mnemonic, RC, PredNot, 0>; // Predicate new defm _cdn#NAME : ld_idxd_shl_pbase<mnemonic, RC, PredNot, 1>; @@ -498,12 +266,23 @@ multiclass ld_idxd_shl<string mnemonic, string CextOp, RegisterClass RC> { } let addrMode = BaseRegOffset in { - defm LDrib_indexed_shl: ld_idxd_shl<"memb", "LDrib", IntRegs>, AddrModeRel; - defm LDriub_indexed_shl: ld_idxd_shl<"memub", "LDriub", IntRegs>, AddrModeRel; - defm LDrih_indexed_shl: ld_idxd_shl<"memh", "LDrih", IntRegs>, AddrModeRel; - defm LDriuh_indexed_shl: ld_idxd_shl<"memuh", "LDriuh", IntRegs>, AddrModeRel; - defm LDriw_indexed_shl: ld_idxd_shl<"memw", "LDriw", IntRegs>, AddrModeRel; - defm LDrid_indexed_shl: ld_idxd_shl<"memd", "LDrid", DoubleRegs>, AddrModeRel; + let accessSize = ByteAccess in { + defm LDrib_indexed_shl: ld_idxd_shl<"memb", "LDrib", IntRegs>, + AddrModeRel; + defm LDriub_indexed_shl: ld_idxd_shl<"memub", "LDriub", IntRegs>, + AddrModeRel; + } + let accessSize = HalfWordAccess in { + defm LDrih_indexed_shl: ld_idxd_shl<"memh", "LDrih", IntRegs>, AddrModeRel; + defm LDriuh_indexed_shl: ld_idxd_shl<"memuh", "LDriuh", IntRegs>, + AddrModeRel; + } + let accessSize = WordAccess in + defm LDriw_indexed_shl: ld_idxd_shl<"memw", "LDriw", IntRegs>, AddrModeRel; + + let accessSize = DoubleWordAccess in + defm LDrid_indexed_shl: ld_idxd_shl<"memd", "LDrid", DoubleRegs>, + AddrModeRel; } // 'def pats' for load instructions with base + register offset and non-zero @@ -596,329 +375,6 @@ def : Pat <(i32 (load (add IntRegs:$src1, IntRegs:$src2))), Requires<[HasV4T]>; } -let isPredicable = 1, neverHasSideEffects = 1, validSubTargets = HasV4SubT in -def LDd_GP_V4 : LDInst2<(outs DoubleRegs:$dst), - (ins globaladdress:$global), - "$dst=memd(#$global)", - []>, - Requires<[HasV4T]>; - -// if (Pv) Rtt=memd(##global) -let neverHasSideEffects = 1, isPredicated = 1, isExtended = 1, opExtendable = 2, -validSubTargets = HasV4SubT in { -def LDd_GP_cPt_V4 : LDInst2<(outs DoubleRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if ($src1) $dst=memd(##$global)", - []>, - Requires<[HasV4T]>; - - -// if (!Pv) Rtt=memd(##global) -def LDd_GP_cNotPt_V4 : LDInst2<(outs DoubleRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if (!$src1) $dst=memd(##$global)", - []>, - Requires<[HasV4T]>; - -// if (Pv) Rtt=memd(##global) -def LDd_GP_cdnPt_V4 : LDInst2<(outs DoubleRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if ($src1.new) $dst=memd(##$global)", - []>, - Requires<[HasV4T]>; - - -// if (!Pv) Rtt=memd(##global) -def LDd_GP_cdnNotPt_V4 : LDInst2<(outs DoubleRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if (!$src1.new) $dst=memd(##$global)", - []>, - Requires<[HasV4T]>; -} - -let isPredicable = 1, neverHasSideEffects = 1, validSubTargets = HasV4SubT in -def LDb_GP_V4 : LDInst2<(outs IntRegs:$dst), - (ins globaladdress:$global), - "$dst=memb(#$global)", - []>, - Requires<[HasV4T]>; - -// if (Pv) Rt=memb(##global) -let neverHasSideEffects = 1, isPredicated = 1, isExtended = 1, opExtendable = 2, -validSubTargets = HasV4SubT in { -def LDb_GP_cPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if ($src1) $dst=memb(##$global)", - []>, - Requires<[HasV4T]>; - -// if (!Pv) Rt=memb(##global) -def LDb_GP_cNotPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if (!$src1) $dst=memb(##$global)", - []>, - Requires<[HasV4T]>; - -// if (Pv) Rt=memb(##global) -def LDb_GP_cdnPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if ($src1.new) $dst=memb(##$global)", - []>, - Requires<[HasV4T]>; - -// if (!Pv) Rt=memb(##global) -def LDb_GP_cdnNotPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if (!$src1.new) $dst=memb(##$global)", - []>, - Requires<[HasV4T]>; -} - -let isPredicable = 1, neverHasSideEffects = 1, validSubTargets = HasV4SubT in -def LDub_GP_V4 : LDInst2<(outs IntRegs:$dst), - (ins globaladdress:$global), - "$dst=memub(#$global)", - []>, - Requires<[HasV4T]>; - -// if (Pv) Rt=memub(##global) -let neverHasSideEffects = 1, isPredicated = 1, isExtended = 1, opExtendable = 2, -validSubTargets = HasV4SubT in { -def LDub_GP_cPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if ($src1) $dst=memub(##$global)", - []>, - Requires<[HasV4T]>; - - -// if (!Pv) Rt=memub(##global) -def LDub_GP_cNotPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if (!$src1) $dst=memub(##$global)", - []>, - Requires<[HasV4T]>; - -// if (Pv) Rt=memub(##global) -def LDub_GP_cdnPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if ($src1.new) $dst=memub(##$global)", - []>, - Requires<[HasV4T]>; - - -// if (!Pv) Rt=memub(##global) -def LDub_GP_cdnNotPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if (!$src1.new) $dst=memub(##$global)", - []>, - Requires<[HasV4T]>; -} - -let isPredicable = 1, neverHasSideEffects = 1, validSubTargets = HasV4SubT in -def LDh_GP_V4 : LDInst2<(outs IntRegs:$dst), - (ins globaladdress:$global), - "$dst=memh(#$global)", - []>, - Requires<[HasV4T]>; - -// if (Pv) Rt=memh(##global) -let neverHasSideEffects = 1, isPredicated = 1, isExtended = 1, opExtendable = 2, -validSubTargets = HasV4SubT in { -def LDh_GP_cPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if ($src1) $dst=memh(##$global)", - []>, - Requires<[HasV4T]>; - -// if (!Pv) Rt=memh(##global) -def LDh_GP_cNotPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if (!$src1) $dst=memh(##$global)", - []>, - Requires<[HasV4T]>; - -// if (Pv) Rt=memh(##global) -def LDh_GP_cdnPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if ($src1.new) $dst=memh(##$global)", - []>, - Requires<[HasV4T]>; - -// if (!Pv) Rt=memh(##global) -def LDh_GP_cdnNotPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if (!$src1.new) $dst=memh(##$global)", - []>, - Requires<[HasV4T]>; -} - -let isPredicable = 1, neverHasSideEffects = 1, validSubTargets = HasV4SubT in -def LDuh_GP_V4 : LDInst2<(outs IntRegs:$dst), - (ins globaladdress:$global), - "$dst=memuh(#$global)", - []>, - Requires<[HasV4T]>; - -// if (Pv) Rt=memuh(##global) -let neverHasSideEffects = 1, isPredicated = 1, isExtended = 1, opExtendable = 2, -validSubTargets = HasV4SubT in { -def LDuh_GP_cPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if ($src1) $dst=memuh(##$global)", - []>, - Requires<[HasV4T]>; - -// if (!Pv) Rt=memuh(##global) -def LDuh_GP_cNotPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if (!$src1) $dst=memuh(##$global)", - []>, - Requires<[HasV4T]>; - -// if (Pv) Rt=memuh(##global) -def LDuh_GP_cdnPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if ($src1.new) $dst=memuh(##$global)", - []>, - Requires<[HasV4T]>; - -// if (!Pv) Rt=memuh(##global) -def LDuh_GP_cdnNotPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if (!$src1.new) $dst=memuh(##$global)", - []>, - Requires<[HasV4T]>; -} - -let isPredicable = 1, neverHasSideEffects = 1, validSubTargets = HasV4SubT in -def LDw_GP_V4 : LDInst2<(outs IntRegs:$dst), - (ins globaladdress:$global), - "$dst=memw(#$global)", - []>, - Requires<[HasV4T]>; - -// if (Pv) Rt=memw(##global) -let neverHasSideEffects = 1, isPredicated = 1, isExtended = 1, opExtendable = 2, -validSubTargets = HasV4SubT in { -def LDw_GP_cPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if ($src1) $dst=memw(##$global)", - []>, - Requires<[HasV4T]>; - - -// if (!Pv) Rt=memw(##global) -def LDw_GP_cNotPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if (!$src1) $dst=memw(##$global)", - []>, - Requires<[HasV4T]>; - -// if (Pv) Rt=memw(##global) -def LDw_GP_cdnPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if ($src1.new) $dst=memw(##$global)", - []>, - Requires<[HasV4T]>; - - -// if (!Pv) Rt=memw(##global) -def LDw_GP_cdnNotPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if (!$src1.new) $dst=memw(##$global)", - []>, - Requires<[HasV4T]>; -} - - -def : Pat <(atomic_load_64 (HexagonCONST32_GP tglobaladdr:$global)), - (i64 (LDd_GP_V4 tglobaladdr:$global))>, - Requires<[HasV4T]>; - -def : Pat <(atomic_load_32 (HexagonCONST32_GP tglobaladdr:$global)), - (i32 (LDw_GP_V4 tglobaladdr:$global))>, - Requires<[HasV4T]>; - -def : Pat <(atomic_load_16 (HexagonCONST32_GP tglobaladdr:$global)), - (i32 (LDuh_GP_V4 tglobaladdr:$global))>, - Requires<[HasV4T]>; - -def : Pat <(atomic_load_8 (HexagonCONST32_GP tglobaladdr:$global)), - (i32 (LDub_GP_V4 tglobaladdr:$global))>, - Requires<[HasV4T]>; - -// Map from load(globaladdress) -> memw(#foo + 0) -let AddedComplexity = 100 in -def : Pat <(i64 (load (HexagonCONST32_GP tglobaladdr:$global))), - (i64 (LDd_GP_V4 tglobaladdr:$global))>, - Requires<[HasV4T]>; - -// Map from Pd = load(globaladdress) -> Rd = memb(globaladdress), Pd = Rd -let AddedComplexity = 100 in -def : Pat <(i1 (load (HexagonCONST32_GP tglobaladdr:$global))), - (i1 (TFR_PdRs (i32 (LDb_GP_V4 tglobaladdr:$global))))>, - Requires<[HasV4T]>; - -// When the Interprocedural Global Variable optimizer realizes that a certain -// global variable takes only two constant values, it shrinks the global to -// a boolean. Catch those loads here in the following 3 patterns. -let AddedComplexity = 100 in -def : Pat <(i32 (extloadi1 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDb_GP_V4 tglobaladdr:$global))>, - Requires<[HasV4T]>; - -let AddedComplexity = 100 in -def : Pat <(i32 (sextloadi1 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDb_GP_V4 tglobaladdr:$global))>, - Requires<[HasV4T]>; - -// Map from load(globaladdress) -> memb(#foo) -let AddedComplexity = 100 in -def : Pat <(i32 (extloadi8 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDb_GP_V4 tglobaladdr:$global))>, - Requires<[HasV4T]>; - -// Map from load(globaladdress) -> memb(#foo) -let AddedComplexity = 100 in -def : Pat <(i32 (sextloadi8 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDb_GP_V4 tglobaladdr:$global))>, - Requires<[HasV4T]>; - -let AddedComplexity = 100 in -def : Pat <(i32 (zextloadi1 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDub_GP_V4 tglobaladdr:$global))>, - Requires<[HasV4T]>; - -// Map from load(globaladdress) -> memub(#foo) -let AddedComplexity = 100 in -def : Pat <(i32 (zextloadi8 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDub_GP_V4 tglobaladdr:$global))>, - Requires<[HasV4T]>; - -// Map from load(globaladdress) -> memh(#foo) -let AddedComplexity = 100 in -def : Pat <(i32 (extloadi16 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDh_GP_V4 tglobaladdr:$global))>, - Requires<[HasV4T]>; - -// Map from load(globaladdress) -> memh(#foo) -let AddedComplexity = 100 in -def : Pat <(i32 (sextloadi16 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDh_GP_V4 tglobaladdr:$global))>, - Requires<[HasV4T]>; - -// Map from load(globaladdress) -> memuh(#foo) -let AddedComplexity = 100 in -def : Pat <(i32 (zextloadi16 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDuh_GP_V4 tglobaladdr:$global))>, - Requires<[HasV4T]>; - -// Map from load(globaladdress) -> memw(#foo) -let AddedComplexity = 100 in -def : Pat <(i32 (load (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDw_GP_V4 tglobaladdr:$global))>, - Requires<[HasV4T]>; - // zext i1->i64 def : Pat <(i64 (zext (i1 PredRegs:$src1))), (i64 (COMBINE_Ir_V4 0, (MUX_ii (i1 PredRegs:$src1), 1, 0)))>, @@ -1008,78 +464,30 @@ def: Pat <(i64 (extloadi32 (i32 (add IntRegs:$src1, s11_2ExtPred:$offset)))), // ST + //===----------------------------------------------------------------------===// /// -/// Assumptions::: ****** DO NOT IGNORE ******** -/// 1. Make sure that in post increment store, the zero'th operand is always the -/// post increment operand. -/// 2. Make sure that the store value operand(Rt/Rtt) in a store is always the -/// last operand. -/// - -// memd(Re=#U)=Rtt -let isExtended = 1, opExtendable = 2, validSubTargets = HasV4SubT in { -def STrid_abs_setimm_V4 : STInst2<(outs IntRegs:$dst1), - (ins DoubleRegs:$src1, u0AlwaysExt:$src2), - "memd($dst1=##$src2) = $src1", - []>, - Requires<[HasV4T]>; - -// memb(Re=#U)=Rs -def STrib_abs_setimm_V4 : STInst2<(outs IntRegs:$dst1), - (ins IntRegs:$src1, u0AlwaysExt:$src2), - "memb($dst1=##$src2) = $src1", - []>, - Requires<[HasV4T]>; - -// memh(Re=#U)=Rs -def STrih_abs_setimm_V4 : STInst2<(outs IntRegs:$dst1), - (ins IntRegs:$src1, u0AlwaysExt:$src2), - "memh($dst1=##$src2) = $src1", - []>, - Requires<[HasV4T]>; - -// memw(Re=#U)=Rs -def STriw_abs_setimm_V4 : STInst2<(outs IntRegs:$dst1), - (ins IntRegs:$src1, u0AlwaysExt:$src2), - "memw($dst1=##$src2) = $src1", - []>, - Requires<[HasV4T]>; -} - -// memd(Re=#U)=Rtt -let isExtended = 1, opExtendable = 2, validSubTargets = HasV4SubT in { -def STrid_abs_set_V4 : STInst2<(outs IntRegs:$dst1), - (ins DoubleRegs:$src1, globaladdressExt:$src2), - "memd($dst1=##$src2) = $src1", - []>, - Requires<[HasV4T]>; - -// memb(Re=#U)=Rs -def STrib_abs_set_V4 : STInst2<(outs IntRegs:$dst1), - (ins IntRegs:$src1, globaladdressExt:$src2), - "memb($dst1=##$src2) = $src1", - []>, - Requires<[HasV4T]>; - -// memh(Re=#U)=Rs -def STrih_abs_set_V4 : STInst2<(outs IntRegs:$dst1), - (ins IntRegs:$src1, globaladdressExt:$src2), - "memh($dst1=##$src2) = $src1", +//===----------------------------------------------------------------------===// +// Template class for store instructions with Absolute set addressing mode. +//===----------------------------------------------------------------------===// +let isExtended = 1, opExtendable = 2, validSubTargets = HasV4SubT, +addrMode = AbsoluteSet in +class T_ST_abs_set<string mnemonic, RegisterClass RC>: + STInst2<(outs IntRegs:$dst1), + (ins RC:$src1, u0AlwaysExt:$src2), + mnemonic#"($dst1=##$src2) = $src1", []>, Requires<[HasV4T]>; -// memw(Re=#U)=Rs -def STriw_abs_set_V4 : STInst2<(outs IntRegs:$dst1), - (ins IntRegs:$src1, globaladdressExt:$src2), - "memw($dst1=##$src2) = $src1", - []>, - Requires<[HasV4T]>; -} +def STrid_abs_set_V4 : T_ST_abs_set <"memd", DoubleRegs>; +def STrib_abs_set_V4 : T_ST_abs_set <"memb", IntRegs>; +def STrih_abs_set_V4 : T_ST_abs_set <"memh", IntRegs>; +def STriw_abs_set_V4 : T_ST_abs_set <"memw", IntRegs>; +//===----------------------------------------------------------------------===// // multiclass for store instructions with base + register offset addressing // mode +//===----------------------------------------------------------------------===// multiclass ST_Idxd_shl_Pbase<string mnemonic, RegisterClass RC, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME : STInst2<(outs), (ins PredRegs:$src1, IntRegs:$src2, IntRegs:$src3, u2Imm:$src4, RC:$src5), @@ -1090,7 +498,7 @@ multiclass ST_Idxd_shl_Pbase<string mnemonic, RegisterClass RC, bit isNot, } multiclass ST_Idxd_shl_Pred<string mnemonic, RegisterClass RC, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { defm _c#NAME : ST_Idxd_shl_Pbase<mnemonic, RC, PredNot, 0>; // Predicate new defm _cdn#NAME : ST_Idxd_shl_Pbase<mnemonic, RC, PredNot, 1>; @@ -1118,7 +526,7 @@ multiclass ST_Idxd_shl<string mnemonic, string CextOp, RegisterClass RC> { // addressing mode. multiclass ST_Idxd_shl_Pbase_nv<string mnemonic, RegisterClass RC, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME#_nv_V4 : NVInst_V4<(outs), (ins PredRegs:$src1, IntRegs:$src2, IntRegs:$src3, u2Imm:$src4, RC:$src5), @@ -1129,7 +537,7 @@ multiclass ST_Idxd_shl_Pbase_nv<string mnemonic, RegisterClass RC, bit isNot, } multiclass ST_Idxd_shl_Pred_nv<string mnemonic, RegisterClass RC, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { defm _c#NAME : ST_Idxd_shl_Pbase_nv<mnemonic, RC, PredNot, 0>; // Predicate new defm _cdn#NAME : ST_Idxd_shl_Pbase_nv<mnemonic, RC, PredNot, 1>; @@ -1155,17 +563,20 @@ multiclass ST_Idxd_shl_nv<string mnemonic, string CextOp, RegisterClass RC> { let addrMode = BaseRegOffset, neverHasSideEffects = 1, validSubTargets = HasV4SubT in { - defm STrib_indexed_shl: ST_Idxd_shl<"memb", "STrib", IntRegs>, - ST_Idxd_shl_nv<"memb", "STrib", IntRegs>, AddrModeRel; + let accessSize = ByteAccess in + defm STrib_indexed_shl: ST_Idxd_shl<"memb", "STrib", IntRegs>, + ST_Idxd_shl_nv<"memb", "STrib", IntRegs>, AddrModeRel; - defm STrih_indexed_shl: ST_Idxd_shl<"memh", "STrih", IntRegs>, - ST_Idxd_shl_nv<"memh", "STrih", IntRegs>, AddrModeRel; + let accessSize = HalfWordAccess in + defm STrih_indexed_shl: ST_Idxd_shl<"memh", "STrih", IntRegs>, + ST_Idxd_shl_nv<"memh", "STrih", IntRegs>, AddrModeRel; - defm STriw_indexed_shl: ST_Idxd_shl<"memw", "STriw", IntRegs>, - ST_Idxd_shl_nv<"memw", "STriw", IntRegs>, AddrModeRel; + let accessSize = WordAccess in + defm STriw_indexed_shl: ST_Idxd_shl<"memw", "STriw", IntRegs>, + ST_Idxd_shl_nv<"memw", "STriw", IntRegs>, AddrModeRel; - let isNVStorable = 0 in - defm STrid_indexed_shl: ST_Idxd_shl<"memd", "STrid", DoubleRegs>, AddrModeRel; + let isNVStorable = 0, accessSize = DoubleWordAccess in + defm STrid_indexed_shl: ST_Idxd_shl<"memd", "STrid", DoubleRegs>, AddrModeRel; } let Predicates = [HasV4T], AddedComplexity = 10 in { @@ -1192,17 +603,59 @@ def : Pat<(store (i64 DoubleRegs:$src4), u2ImmPred:$src3, DoubleRegs:$src4)>; } -// memd(Ru<<#u2+#U6)=Rtt -let isExtended = 1, opExtendable = 2, AddedComplexity = 10, -validSubTargets = HasV4SubT in -def STrid_shl_V4 : STInst<(outs), - (ins IntRegs:$src1, u2Imm:$src2, u0AlwaysExt:$src3, DoubleRegs:$src4), - "memd($src1<<#$src2+#$src3) = $src4", - [(store (i64 DoubleRegs:$src4), +let isExtended = 1, opExtendable = 2 in +class T_ST_LongOff <string mnemonic, PatFrag stOp, RegisterClass RC, ValueType VT> : + STInst<(outs), + (ins IntRegs:$src1, u2Imm:$src2, u0AlwaysExt:$src3, RC:$src4), + mnemonic#"($src1<<#$src2+##$src3) = $src4", + [(stOp (VT RC:$src4), (add (shl (i32 IntRegs:$src1), u2ImmPred:$src2), u0AlwaysExtPred:$src3))]>, Requires<[HasV4T]>; +let isExtended = 1, opExtendable = 2, mayStore = 1, isNVStore = 1 in +class T_ST_LongOff_nv <string mnemonic> : + NVInst_V4<(outs), + (ins IntRegs:$src1, u2Imm:$src2, u0AlwaysExt:$src3, IntRegs:$src4), + mnemonic#"($src1<<#$src2+##$src3) = $src4.new", + []>, + Requires<[HasV4T]>; + +multiclass ST_LongOff <string mnemonic, string BaseOp, PatFrag stOp> { + let BaseOpcode = BaseOp#"_shl" in { + let isNVStorable = 1 in + def NAME#_V4 : T_ST_LongOff<mnemonic, stOp, IntRegs, i32>; + + def NAME#_nv_V4 : T_ST_LongOff_nv<mnemonic>; + } +} + +let AddedComplexity = 10, validSubTargets = HasV4SubT in { + def STrid_shl_V4 : T_ST_LongOff<"memd", store, DoubleRegs, i64>; + defm STrib_shl : ST_LongOff <"memb", "STrib", truncstorei8>, NewValueRel; + defm STrih_shl : ST_LongOff <"memh", "Strih", truncstorei16>, NewValueRel; + defm STriw_shl : ST_LongOff <"memw", "STriw", store>, NewValueRel; +} + +let AddedComplexity = 40 in +multiclass T_ST_LOff_Pats <InstHexagon I, RegisterClass RC, ValueType VT, + PatFrag stOp> { + def : Pat<(stOp (VT RC:$src4), + (add (shl IntRegs:$src1, u2ImmPred:$src2), + (NumUsesBelowThresCONST32 tglobaladdr:$src3))), + (I IntRegs:$src1, u2ImmPred:$src2, tglobaladdr:$src3, RC:$src4)>; + + def : Pat<(stOp (VT RC:$src4), + (add IntRegs:$src1, + (NumUsesBelowThresCONST32 tglobaladdr:$src3))), + (I IntRegs:$src1, 0, tglobaladdr:$src3, RC:$src4)>; +} + +defm : T_ST_LOff_Pats<STrid_shl_V4, DoubleRegs, i64, store>; +defm : T_ST_LOff_Pats<STriw_shl_V4, IntRegs, i32, store>; +defm : T_ST_LOff_Pats<STrib_shl_V4, IntRegs, i32, truncstorei8>; +defm : T_ST_LOff_Pats<STrih_shl_V4, IntRegs, i32, truncstorei16>; + // memd(Rx++#s4:3)=Rtt // memd(Rx++#s4:3:circ(Mu))=Rtt // memd(Rx++I:circ(Mu))=Rtt @@ -1222,7 +675,7 @@ def STrid_shl_V4 : STInst<(outs), //===----------------------------------------------------------------------===// multiclass ST_Imm_Pbase<string mnemonic, Operand OffsetOp, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME : STInst2<(outs), (ins PredRegs:$src1, IntRegs:$src2, OffsetOp:$src3, s6Ext:$src4), !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ", @@ -1232,7 +685,7 @@ multiclass ST_Imm_Pbase<string mnemonic, Operand OffsetOp, bit isNot, } multiclass ST_Imm_Pred<string mnemonic, Operand OffsetOp, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { defm _c#NAME : ST_Imm_Pbase<mnemonic, OffsetOp, PredNot, 0>; // Predicate new defm _cdn#NAME : ST_Imm_Pbase<mnemonic, OffsetOp, PredNot, 1>; @@ -1257,10 +710,15 @@ multiclass ST_Imm<string mnemonic, string CextOp, Operand OffsetOp> { } let addrMode = BaseImmOffset, InputType = "imm", - validSubTargets = HasV4SubT in { - defm STrib_imm : ST_Imm<"memb", "STrib", u6_0Imm>, ImmRegRel, PredNewRel; - defm STrih_imm : ST_Imm<"memh", "STrih", u6_1Imm>, ImmRegRel, PredNewRel; - defm STriw_imm : ST_Imm<"memw", "STriw", u6_2Imm>, ImmRegRel, PredNewRel; +validSubTargets = HasV4SubT in { + let accessSize = ByteAccess in + defm STrib_imm : ST_Imm<"memb", "STrib", u6_0Imm>, ImmRegRel, PredNewRel; + + let accessSize = HalfWordAccess in + defm STrih_imm : ST_Imm<"memh", "STrih", u6_1Imm>, ImmRegRel, PredNewRel; + + let accessSize = WordAccess in + defm STriw_imm : ST_Imm<"memw", "STriw", u6_2Imm>, ImmRegRel, PredNewRel; } let Predicates = [HasV4T], AddedComplexity = 10 in { @@ -1280,17 +738,6 @@ def : Pat <(truncstorei8 s8ExtPred:$src2, (i32 IntRegs:$src1)), (STrib_imm_V4 IntRegs:$src1, 0, s8ExtPred:$src2)>, Requires<[HasV4T]>; -// memb(Ru<<#u2+#U6)=Rt -let isExtended = 1, opExtendable = 2, AddedComplexity = 10, isNVStorable = 1, -validSubTargets = HasV4SubT in -def STrib_shl_V4 : STInst<(outs), - (ins IntRegs:$src1, u2Imm:$src2, u0AlwaysExt:$src3, IntRegs:$src4), - "memb($src1<<#$src2+#$src3) = $src4", - [(truncstorei8 (i32 IntRegs:$src4), - (add (shl (i32 IntRegs:$src1), u2ImmPred:$src2), - u0AlwaysExtPred:$src3))]>, - Requires<[HasV4T]>; - // memb(Rx++#s4:0:circ(Mu))=Rt // memb(Rx++I:circ(Mu))=Rt // memb(Rx++Mu)=Rt @@ -1311,17 +758,6 @@ def : Pat <(truncstorei16 s8ExtPred:$src2, (i32 IntRegs:$src1)), // TODO: needs to be implemented. // memh(Ru<<#u2+#U6)=Rt.H -// memh(Ru<<#u2+#U6)=Rt -let isExtended = 1, opExtendable = 2, AddedComplexity = 10, isNVStorable = 1, -validSubTargets = HasV4SubT in -def STrih_shl_V4 : STInst<(outs), - (ins IntRegs:$src1, u2Imm:$src2, u0AlwaysExt:$src3, IntRegs:$src4), - "memh($src1<<#$src2+#$src3) = $src4", - [(truncstorei16 (i32 IntRegs:$src4), - (add (shl (i32 IntRegs:$src1), u2ImmPred:$src2), - u0AlwaysExtPred:$src3))]>, - Requires<[HasV4T]>; - // memh(Rx++#s4:1:circ(Mu))=Rt.H // memh(Rx++#s4:1:circ(Mu))=Rt // memh(Rx++I:circ(Mu))=Rt.H @@ -1358,241 +794,11 @@ def : Pat <(store s8ExtPred:$src2, (i32 IntRegs:$src1)), (STriw_imm_V4 IntRegs:$src1, 0, s8ExtPred:$src2)>, Requires<[HasV4T]>; -// memw(Ru<<#u2+#U6)=Rt -let isExtended = 1, opExtendable = 2, AddedComplexity = 10, isNVStorable = 1, -validSubTargets = HasV4SubT in -def STriw_shl_V4 : STInst<(outs), - (ins IntRegs:$src1, u2Imm:$src2, u0AlwaysExt:$src3, IntRegs:$src4), - "memw($src1<<#$src2+#$src3) = $src4", - [(store (i32 IntRegs:$src4), - (add (shl (i32 IntRegs:$src1), u2ImmPred:$src2), - u0AlwaysExtPred:$src3))]>, - Requires<[HasV4T]>; - // memw(Rx++#s4:2)=Rt // memw(Rx++#s4:2:circ(Mu))=Rt // memw(Rx++I:circ(Mu))=Rt // memw(Rx++Mu)=Rt // memw(Rx++Mu:brev)=Rt -// memw(gp+#u16:2)=Rt - - -// memd(#global)=Rtt -let isPredicable = 1, mayStore = 1, neverHasSideEffects = 1, -validSubTargets = HasV4SubT in -def STd_GP_V4 : STInst2<(outs), - (ins globaladdress:$global, DoubleRegs:$src), - "memd(#$global) = $src", - []>, - Requires<[HasV4T]>; - -// if (Pv) memd(##global) = Rtt -let mayStore = 1, neverHasSideEffects = 1, isPredicated = 1, -isExtended = 1, opExtendable = 1, validSubTargets = HasV4SubT in { -def STd_GP_cPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdress:$global, DoubleRegs:$src2), - "if ($src1) memd(##$global) = $src2", - []>, - Requires<[HasV4T]>; - -// if (!Pv) memd(##global) = Rtt -def STd_GP_cNotPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdress:$global, DoubleRegs:$src2), - "if (!$src1) memd(##$global) = $src2", - []>, - Requires<[HasV4T]>; - -// if (Pv) memd(##global) = Rtt -def STd_GP_cdnPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdress:$global, DoubleRegs:$src2), - "if ($src1.new) memd(##$global) = $src2", - []>, - Requires<[HasV4T]>; - -// if (!Pv) memd(##global) = Rtt -def STd_GP_cdnNotPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdress:$global, DoubleRegs:$src2), - "if (!$src1.new) memd(##$global) = $src2", - []>, - Requires<[HasV4T]>; -} - -// memb(#global)=Rt -let isPredicable = 1, neverHasSideEffects = 1, isNVStorable = 1, -validSubTargets = HasV4SubT in -def STb_GP_V4 : STInst2<(outs), - (ins globaladdress:$global, IntRegs:$src), - "memb(#$global) = $src", - []>, - Requires<[HasV4T]>; - -// if (Pv) memb(##global) = Rt -let neverHasSideEffects = 1, isPredicated = 1, isNVStorable = 1, -isExtended = 1, opExtendable = 1, validSubTargets = HasV4SubT in { -def STb_GP_cPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if ($src1) memb(##$global) = $src2", - []>, - Requires<[HasV4T]>; - -// if (!Pv) memb(##global) = Rt -def STb_GP_cNotPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if (!$src1) memb(##$global) = $src2", - []>, - Requires<[HasV4T]>; - -// if (Pv) memb(##global) = Rt -def STb_GP_cdnPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if ($src1.new) memb(##$global) = $src2", - []>, - Requires<[HasV4T]>; - -// if (!Pv) memb(##global) = Rt -def STb_GP_cdnNotPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if (!$src1.new) memb(##$global) = $src2", - []>, - Requires<[HasV4T]>; -} - -// memh(#global)=Rt -let isPredicable = 1, neverHasSideEffects = 1, isNVStorable = 1, -validSubTargets = HasV4SubT in -def STh_GP_V4 : STInst2<(outs), - (ins globaladdress:$global, IntRegs:$src), - "memh(#$global) = $src", - []>, - Requires<[HasV4T]>; - -// if (Pv) memh(##global) = Rt -let neverHasSideEffects = 1, isPredicated = 1, isNVStorable = 1, -isExtended = 1, opExtendable = 1, validSubTargets = HasV4SubT in { -def STh_GP_cPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if ($src1) memh(##$global) = $src2", - []>, - Requires<[HasV4T]>; - -// if (!Pv) memh(##global) = Rt -def STh_GP_cNotPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if (!$src1) memh(##$global) = $src2", - []>, - Requires<[HasV4T]>; - -// if (Pv) memh(##global) = Rt -def STh_GP_cdnPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if ($src1.new) memh(##$global) = $src2", - []>, - Requires<[HasV4T]>; - -// if (!Pv) memh(##global) = Rt -def STh_GP_cdnNotPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if (!$src1.new) memh(##$global) = $src2", - []>, - Requires<[HasV4T]>; -} - -// memw(#global)=Rt -let isPredicable = 1, neverHasSideEffects = 1, isNVStorable = 1, -validSubTargets = HasV4SubT in -def STw_GP_V4 : STInst2<(outs), - (ins globaladdress:$global, IntRegs:$src), - "memw(#$global) = $src", - []>, - Requires<[HasV4T]>; - -// if (Pv) memw(##global) = Rt -let neverHasSideEffects = 1, isPredicated = 1, isNVStorable = 1, -isExtended = 1, opExtendable = 1, validSubTargets = HasV4SubT in { -def STw_GP_cPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if ($src1) memw(##$global) = $src2", - []>, - Requires<[HasV4T]>; - -// if (!Pv) memw(##global) = Rt -def STw_GP_cNotPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if (!$src1) memw(##$global) = $src2", - []>, - Requires<[HasV4T]>; - -// if (Pv) memw(##global) = Rt -def STw_GP_cdnPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if ($src1.new) memw(##$global) = $src2", - []>, - Requires<[HasV4T]>; - -// if (!Pv) memw(##global) = Rt -def STw_GP_cdnNotPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if (!$src1.new) memw(##$global) = $src2", - []>, - Requires<[HasV4T]>; -} - -// 64 bit atomic store -def : Pat <(atomic_store_64 (HexagonCONST32_GP tglobaladdr:$global), - (i64 DoubleRegs:$src1)), - (STd_GP_V4 tglobaladdr:$global, (i64 DoubleRegs:$src1))>, - Requires<[HasV4T]>; - -// Map from store(globaladdress) -> memd(#foo) -let AddedComplexity = 100 in -def : Pat <(store (i64 DoubleRegs:$src1), - (HexagonCONST32_GP tglobaladdr:$global)), - (STd_GP_V4 tglobaladdr:$global, (i64 DoubleRegs:$src1))>, - Requires<[HasV4T]>; - -// 8 bit atomic store -def : Pat < (atomic_store_8 (HexagonCONST32_GP tglobaladdr:$global), - (i32 IntRegs:$src1)), - (STb_GP_V4 tglobaladdr:$global, (i32 IntRegs:$src1))>, - Requires<[HasV4T]>; - -// Map from store(globaladdress) -> memb(#foo) -let AddedComplexity = 100 in -def : Pat<(truncstorei8 (i32 IntRegs:$src1), - (HexagonCONST32_GP tglobaladdr:$global)), - (STb_GP_V4 tglobaladdr:$global, (i32 IntRegs:$src1))>, - Requires<[HasV4T]>; - -// Map from "i1 = constant<-1>; memw(CONST32(#foo)) = i1" -// to "r0 = 1; memw(#foo) = r0" -let AddedComplexity = 100 in -def : Pat<(store (i1 -1), (HexagonCONST32_GP tglobaladdr:$global)), - (STb_GP_V4 tglobaladdr:$global, (TFRI 1))>, - Requires<[HasV4T]>; - -def : Pat<(atomic_store_16 (HexagonCONST32_GP tglobaladdr:$global), - (i32 IntRegs:$src1)), - (STh_GP_V4 tglobaladdr:$global, (i32 IntRegs:$src1))>, - Requires<[HasV4T]>; - -// Map from store(globaladdress) -> memh(#foo) -let AddedComplexity = 100 in -def : Pat<(truncstorei16 (i32 IntRegs:$src1), - (HexagonCONST32_GP tglobaladdr:$global)), - (STh_GP_V4 tglobaladdr:$global, (i32 IntRegs:$src1))>, - Requires<[HasV4T]>; - -// 32 bit atomic store -def : Pat<(atomic_store_32 (HexagonCONST32_GP tglobaladdr:$global), - (i32 IntRegs:$src1)), - (STw_GP_V4 tglobaladdr:$global, (i32 IntRegs:$src1))>, - Requires<[HasV4T]>; - -// Map from store(globaladdress) -> memw(#foo) -let AddedComplexity = 100 in -def : Pat<(store (i32 IntRegs:$src1), (HexagonCONST32_GP tglobaladdr:$global)), - (STw_GP_V4 tglobaladdr:$global, (i32 IntRegs:$src1))>, - Requires<[HasV4T]>; //===----------------------------------------------------------------------=== // ST - @@ -1607,7 +813,7 @@ def : Pat<(store (i32 IntRegs:$src1), (HexagonCONST32_GP tglobaladdr:$global)), // multiclass ST_Idxd_Pbase_nv<string mnemonic, RegisterClass RC, Operand predImmOp, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME#_nv_V4 : NVInst_V4<(outs), (ins PredRegs:$src1, IntRegs:$src2, predImmOp:$src3, RC: $src4), !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ", @@ -1618,7 +824,7 @@ multiclass ST_Idxd_Pbase_nv<string mnemonic, RegisterClass RC, multiclass ST_Idxd_Pred_nv<string mnemonic, RegisterClass RC, Operand predImmOp, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { defm _c#NAME : ST_Idxd_Pbase_nv<mnemonic, RC, predImmOp, PredNot, 0>; // Predicate new defm _cdn#NAME : ST_Idxd_Pbase_nv<mnemonic, RC, predImmOp, PredNot, 1>; @@ -1648,19 +854,24 @@ multiclass ST_Idxd_nv<string mnemonic, string CextOp, RegisterClass RC, } let addrMode = BaseImmOffset, validSubTargets = HasV4SubT in { - defm STrib_indexed: ST_Idxd_nv<"memb", "STrib", IntRegs, s11_0Ext, - u6_0Ext, 11, 6>, AddrModeRel; - defm STrih_indexed: ST_Idxd_nv<"memh", "STrih", IntRegs, s11_1Ext, - u6_1Ext, 12, 7>, AddrModeRel; - defm STriw_indexed: ST_Idxd_nv<"memw", "STriw", IntRegs, s11_2Ext, - u6_2Ext, 13, 8>, AddrModeRel; + let accessSize = ByteAccess in + defm STrib_indexed: ST_Idxd_nv<"memb", "STrib", IntRegs, s11_0Ext, + u6_0Ext, 11, 6>, AddrModeRel; + + let accessSize = HalfWordAccess in + defm STrih_indexed: ST_Idxd_nv<"memh", "STrih", IntRegs, s11_1Ext, + u6_1Ext, 12, 7>, AddrModeRel; + + let accessSize = WordAccess in + defm STriw_indexed: ST_Idxd_nv<"memw", "STriw", IntRegs, s11_2Ext, + u6_2Ext, 13, 8>, AddrModeRel; } // multiclass for new-value store instructions with base + immediate offset. // and MEMri operand. multiclass ST_MEMri_Pbase_nv<string mnemonic, RegisterClass RC, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME#_nv_V4 : NVInst_V4<(outs), (ins PredRegs:$src1, MEMri:$addr, RC: $src2), !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ", @@ -1670,7 +881,7 @@ multiclass ST_MEMri_Pbase_nv<string mnemonic, RegisterClass RC, bit isNot, } multiclass ST_MEMri_Pred_nv<string mnemonic, RegisterClass RC, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { defm _c#NAME : ST_MEMri_Pbase_nv<mnemonic, RC, PredNot, 0>; // Predicate new @@ -1701,19 +912,15 @@ multiclass ST_MEMri_nv<string mnemonic, string CextOp, RegisterClass RC, let addrMode = BaseImmOffset, isMEMri = "true", validSubTargets = HasV4SubT, mayStore = 1 in { - defm STrib: ST_MEMri_nv<"memb", "STrib", IntRegs, 11, 6>, AddrModeRel; - defm STrih: ST_MEMri_nv<"memh", "STrih", IntRegs, 12, 7>, AddrModeRel; - defm STriw: ST_MEMri_nv<"memw", "STriw", IntRegs, 13, 8>, AddrModeRel; -} + let accessSize = ByteAccess in + defm STrib: ST_MEMri_nv<"memb", "STrib", IntRegs, 11, 6>, AddrModeRel; -// memb(Ru<<#u2+#U6)=Nt.new -let isExtended = 1, opExtendable = 2, mayStore = 1, AddedComplexity = 10, -isNVStore = 1, validSubTargets = HasV4SubT in -def STrib_shl_nv_V4 : NVInst_V4<(outs), - (ins IntRegs:$src1, u2Imm:$src2, u0AlwaysExt:$src3, IntRegs:$src4), - "memb($src1<<#$src2+#$src3) = $src4.new", - []>, - Requires<[HasV4T]>; + let accessSize = HalfWordAccess in + defm STrih: ST_MEMri_nv<"memh", "STrih", IntRegs, 12, 7>, AddrModeRel; + + let accessSize = WordAccess in + defm STriw: ST_MEMri_nv<"memw", "STriw", IntRegs, 13, 8>, AddrModeRel; +} //===----------------------------------------------------------------------===// // Post increment store @@ -1722,7 +929,7 @@ def STrib_shl_nv_V4 : NVInst_V4<(outs), multiclass ST_PostInc_Pbase_nv<string mnemonic, RegisterClass RC, Operand ImmOp, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME#_nv_V4 : NVInstPI_V4<(outs IntRegs:$dst), (ins PredRegs:$src1, IntRegs:$src2, ImmOp:$offset, RC:$src3), !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ", @@ -1734,7 +941,7 @@ multiclass ST_PostInc_Pbase_nv<string mnemonic, RegisterClass RC, Operand ImmOp, multiclass ST_PostInc_Pred_nv<string mnemonic, RegisterClass RC, Operand ImmOp, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { defm _c#NAME : ST_PostInc_Pbase_nv<mnemonic, RC, ImmOp, PredNot, 0>; // Predicate new let Predicates = [HasV4T], validSubTargets = HasV4SubT in @@ -1762,7 +969,7 @@ multiclass ST_PostInc_nv<string mnemonic, string BaseOp, RegisterClass RC, } } -let validSubTargets = HasV4SubT in { +let addrMode = PostInc, validSubTargets = HasV4SubT in { defm POST_STbri: ST_PostInc_nv <"memb", "STrib", IntRegs, s4_0Imm>, AddrModeRel; defm POST_SThri: ST_PostInc_nv <"memh", "STrih", IntRegs, s4_1Imm>, AddrModeRel; defm POST_STwri: ST_PostInc_nv <"memw", "STriw", IntRegs, s4_2Imm>, AddrModeRel; @@ -1772,146 +979,15 @@ defm POST_STwri: ST_PostInc_nv <"memw", "STriw", IntRegs, s4_2Imm>, AddrModeRel; // memb(Rx++I:circ(Mu))=Nt.new // memb(Rx++Mu)=Nt.new // memb(Rx++Mu:brev)=Nt.new - -// memb(#global)=Nt.new -let mayStore = 1, neverHasSideEffects = 1 in -def STb_GP_nv_V4 : NVInst_V4<(outs), - (ins globaladdress:$global, IntRegs:$src), - "memb(#$global) = $src.new", - []>, - Requires<[HasV4T]>; - -// memh(Ru<<#u2+#U6)=Nt.new -let isExtended = 1, opExtendable = 2, mayStore = 1, AddedComplexity = 10, -isNVStore = 1, validSubTargets = HasV4SubT in -def STrih_shl_nv_V4 : NVInst_V4<(outs), - (ins IntRegs:$src1, u2Imm:$src2, u0AlwaysExt:$src3, IntRegs:$src4), - "memh($src1<<#$src2+#$src3) = $src4.new", - []>, - Requires<[HasV4T]>; - // memh(Rx++#s4:1:circ(Mu))=Nt.new // memh(Rx++I:circ(Mu))=Nt.new // memh(Rx++Mu)=Nt.new // memh(Rx++Mu:brev)=Nt.new -// memh(#global)=Nt.new -let mayStore = 1, neverHasSideEffects = 1 in -def STh_GP_nv_V4 : NVInst_V4<(outs), - (ins globaladdress:$global, IntRegs:$src), - "memh(#$global) = $src.new", - []>, - Requires<[HasV4T]>; - -// memw(Ru<<#u2+#U6)=Nt.new -let isExtended = 1, opExtendable = 2, mayStore = 1, AddedComplexity = 10, -isNVStore = 1, validSubTargets = HasV4SubT in -def STriw_shl_nv_V4 : NVInst_V4<(outs), - (ins IntRegs:$src1, u2Imm:$src2, u0AlwaysExt:$src3, IntRegs:$src4), - "memw($src1<<#$src2+#$src3) = $src4.new", - []>, - Requires<[HasV4T]>; - // memw(Rx++#s4:2:circ(Mu))=Nt.new // memw(Rx++I:circ(Mu))=Nt.new // memw(Rx++Mu)=Nt.new // memw(Rx++Mu:brev)=Nt.new -// memw(gp+#u16:2)=Nt.new - -let mayStore = 1, neverHasSideEffects = 1, isNVStore = 1, -validSubTargets = HasV4SubT in -def STw_GP_nv_V4 : NVInst_V4<(outs), - (ins globaladdress:$global, IntRegs:$src), - "memw(#$global) = $src.new", - []>, - Requires<[HasV4T]>; - -// if (Pv) memb(##global) = Rt -let mayStore = 1, neverHasSideEffects = 1, isNVStore = 1, -isExtended = 1, opExtendable = 1, validSubTargets = HasV4SubT in { -def STb_GP_cPt_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if ($src1) memb(##$global) = $src2.new", - []>, - Requires<[HasV4T]>; - -// if (!Pv) memb(##global) = Rt -def STb_GP_cNotPt_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if (!$src1) memb(##$global) = $src2.new", - []>, - Requires<[HasV4T]>; - -// if (Pv) memb(##global) = Rt -def STb_GP_cdnPt_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if ($src1.new) memb(##$global) = $src2.new", - []>, - Requires<[HasV4T]>; - -// if (!Pv) memb(##global) = Rt -def STb_GP_cdnNotPt_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if (!$src1.new) memb(##$global) = $src2.new", - []>, - Requires<[HasV4T]>; - -// if (Pv) memh(##global) = Rt -def STh_GP_cPt_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if ($src1) memh(##$global) = $src2.new", - []>, - Requires<[HasV4T]>; - -// if (!Pv) memh(##global) = Rt -def STh_GP_cNotPt_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if (!$src1) memh(##$global) = $src2.new", - []>, - Requires<[HasV4T]>; - -// if (Pv) memh(##global) = Rt -def STh_GP_cdnPt_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if ($src1.new) memh(##$global) = $src2.new", - []>, - Requires<[HasV4T]>; - -// if (!Pv) memh(##global) = Rt -def STh_GP_cdnNotPt_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if (!$src1.new) memh(##$global) = $src2.new", - []>, - Requires<[HasV4T]>; - -// if (Pv) memw(##global) = Rt -def STw_GP_cPt_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if ($src1) memw(##$global) = $src2.new", - []>, - Requires<[HasV4T]>; - -// if (!Pv) memw(##global) = Rt -def STw_GP_cNotPt_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if (!$src1) memw(##$global) = $src2.new", - []>, - Requires<[HasV4T]>; - -// if (Pv) memw(##global) = Rt -def STw_GP_cdnPt_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if ($src1.new) memw(##$global) = $src2.new", - []>, - Requires<[HasV4T]>; - -// if (!Pv) memw(##global) = Rt -def STw_GP_cdnNotPt_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if (!$src1.new) memw(##$global) = $src2.new", - []>, - Requires<[HasV4T]>; -} //===----------------------------------------------------------------------===// // NV/ST - @@ -1921,181 +997,195 @@ def STw_GP_cdnNotPt_nv_V4 : NVInst_V4<(outs), // NV/J + //===----------------------------------------------------------------------===// -multiclass NVJ_type_basic_reg<string NotStr, string OpcStr, string TakenStr> { - def _ie_nv_V4 : NVInst_V4<(outs), - (ins IntRegs:$src1, IntRegs:$src2, brtarget:$offset), - !strconcat("if (", !strconcat(NotStr, !strconcat(OpcStr, - !strconcat("($src1.new, $src2)) jump:", - !strconcat(TakenStr, " $offset"))))), - []>, - Requires<[HasV4T]>; - - def _nv_V4 : NVInst_V4<(outs), - (ins IntRegs:$src1, IntRegs:$src2, brtarget:$offset), - !strconcat("if (", !strconcat(NotStr, !strconcat(OpcStr, - !strconcat("($src1.new, $src2)) jump:", - !strconcat(TakenStr, " $offset"))))), - []>, - Requires<[HasV4T]>; -} - -multiclass NVJ_type_basic_2ndDotNew<string NotStr, string OpcStr, - string TakenStr> { - def _ie_nv_V4 : NVInst_V4<(outs), - (ins IntRegs:$src1, IntRegs:$src2, brtarget:$offset), - !strconcat("if (", !strconcat(NotStr, !strconcat(OpcStr, - !strconcat("($src1, $src2.new)) jump:", - !strconcat(TakenStr, " $offset"))))), - []>, - Requires<[HasV4T]>; - - def _nv_V4 : NVInst_V4<(outs), - (ins IntRegs:$src1, IntRegs:$src2, brtarget:$offset), - !strconcat("if (", !strconcat(NotStr, !strconcat(OpcStr, - !strconcat("($src1, $src2.new)) jump:", - !strconcat(TakenStr, " $offset"))))), - []>, - Requires<[HasV4T]>; -} - -multiclass NVJ_type_basic_imm<string NotStr, string OpcStr, string TakenStr> { - def _ie_nv_V4 : NVInst_V4<(outs), - (ins IntRegs:$src1, u5Imm:$src2, brtarget:$offset), - !strconcat("if (", !strconcat(NotStr, !strconcat(OpcStr, - !strconcat("($src1.new, #$src2)) jump:", - !strconcat(TakenStr, " $offset"))))), - []>, - Requires<[HasV4T]>; - - def _nv_V4 : NVInst_V4<(outs), - (ins IntRegs:$src1, u5Imm:$src2, brtarget:$offset), - !strconcat("if (", !strconcat(NotStr, !strconcat(OpcStr, - !strconcat("($src1.new, #$src2)) jump:", - !strconcat(TakenStr, " $offset"))))), - []>, - Requires<[HasV4T]>; -} - -multiclass NVJ_type_basic_neg<string NotStr, string OpcStr, string TakenStr> { - def _ie_nv_V4 : NVInst_V4<(outs), - (ins IntRegs:$src1, nOneImm:$src2, brtarget:$offset), - !strconcat("if (", !strconcat(NotStr, !strconcat(OpcStr, - !strconcat("($src1.new, #$src2)) jump:", - !strconcat(TakenStr, " $offset"))))), - []>, - Requires<[HasV4T]>; - - def _nv_V4 : NVInst_V4<(outs), - (ins IntRegs:$src1, nOneImm:$src2, brtarget:$offset), - !strconcat("if (", !strconcat(NotStr, !strconcat(OpcStr, - !strconcat("($src1.new, #$src2)) jump:", - !strconcat(TakenStr, " $offset"))))), - []>, - Requires<[HasV4T]>; -} - -multiclass NVJ_type_basic_tstbit<string NotStr, string OpcStr, - string TakenStr> { - def _ie_nv_V4 : NVInst_V4<(outs), - (ins IntRegs:$src1, u1Imm:$src2, brtarget:$offset), - !strconcat("if (", !strconcat(NotStr, !strconcat(OpcStr, - !strconcat("($src1.new, #$src2)) jump:", - !strconcat(TakenStr, " $offset"))))), - []>, - Requires<[HasV4T]>; +//===----------------------------------------------------------------------===// +// multiclass/template class for the new-value compare jumps with the register +// operands. +//===----------------------------------------------------------------------===// - def _nv_V4 : NVInst_V4<(outs), - (ins IntRegs:$src1, u1Imm:$src2, brtarget:$offset), - !strconcat("if (", !strconcat(NotStr, !strconcat(OpcStr, - !strconcat("($src1.new, #$src2)) jump:", - !strconcat(TakenStr, " $offset"))))), - []>, - Requires<[HasV4T]>; +let isExtendable = 1, opExtendable = 2, isExtentSigned = 1, opExtentBits = 11 in +class NVJrr_template<string mnemonic, bits<3> majOp, bit NvOpNum, + bit isNegCond, bit isTaken> + : NVInst_V4<(outs), + (ins IntRegs:$src1, IntRegs:$src2, brtarget:$offset), + "if ("#!if(isNegCond, "!","")#mnemonic# + "($src1"#!if(!eq(NvOpNum, 0),".new, ",", ")# + "$src2"#!if(!eq(NvOpNum, 1),".new))","))")#" jump:" + #!if(isTaken, "t","nt")#" $offset", + []>, Requires<[HasV4T]> { + + bits<5> src1; + bits<5> src2; + bits<3> Ns; // New-Value Operand + bits<5> RegOp; // Non New-Value Operand + bits<11> offset; + + let isBrTaken = !if(isTaken, "true", "false"); + let isPredicatedFalse = isNegCond; + + let Ns = !if(!eq(NvOpNum, 0), src1{2-0}, src2{2-0}); + let RegOp = !if(!eq(NvOpNum, 0), src2, src1); + + let IClass = 0b0010; + let Inst{26} = 0b0; + let Inst{25-23} = majOp; + let Inst{22} = isNegCond; + let Inst{18-16} = Ns; + let Inst{13} = isTaken; + let Inst{12-8} = RegOp; + let Inst{21-20} = offset{10-9}; + let Inst{7-1} = offset{8-2}; +} + + +multiclass NVJrr_cond<string mnemonic, bits<3> majOp, bit NvOpNum, + bit isNegCond> { + // Branch not taken: + def _nt_V4: NVJrr_template<mnemonic, majOp, NvOpNum, isNegCond, 0>; + // Branch taken: + def _t_V4: NVJrr_template<mnemonic, majOp, NvOpNum, isNegCond, 1>; +} + +// NvOpNum = 0 -> First Operand is a new-value Register +// NvOpNum = 1 -> Second Operand is a new-value Register + +multiclass NVJrr_base<string mnemonic, string BaseOp, bits<3> majOp, + bit NvOpNum> { + let BaseOpcode = BaseOp#_NVJ in { + defm _t_Jumpnv : NVJrr_cond<mnemonic, majOp, NvOpNum, 0>; // True cond + defm _f_Jumpnv : NVJrr_cond<mnemonic, majOp, NvOpNum, 1>; // False cond + } } -// Multiclass for regular dot new of Ist operand register. -multiclass NVJ_type_br_pred_reg<string NotStr, string OpcStr> { - defm Pt : NVJ_type_basic_reg<NotStr, OpcStr, "t">; - defm Pnt : NVJ_type_basic_reg<NotStr, OpcStr, "nt">; -} +// if ([!]cmp.eq(Ns.new,Rt)) jump:[n]t #r9:2 +// if ([!]cmp.gt(Ns.new,Rt)) jump:[n]t #r9:2 +// if ([!]cmp.gtu(Ns.new,Rt)) jump:[n]t #r9:2 +// if ([!]cmp.gt(Rt,Ns.new)) jump:[n]t #r9:2 +// if ([!]cmp.gtu(Rt,Ns.new)) jump:[n]t #r9:2 -// Multiclass for dot new of 2nd operand register. -multiclass NVJ_type_br_pred_2ndDotNew<string NotStr, string OpcStr> { - defm Pt : NVJ_type_basic_2ndDotNew<NotStr, OpcStr, "t">; - defm Pnt : NVJ_type_basic_2ndDotNew<NotStr, OpcStr, "nt">; +let isPredicated = 1, isBranch = 1, isNewValue = 1, isTerminator = 1, + Defs = [PC], neverHasSideEffects = 1, validSubTargets = HasV4SubT in { + defm CMPEQrr : NVJrr_base<"cmp.eq", "CMPEQ", 0b000, 0>, PredRel; + defm CMPGTrr : NVJrr_base<"cmp.gt", "CMPGT", 0b001, 0>, PredRel; + defm CMPGTUrr : NVJrr_base<"cmp.gtu", "CMPGTU", 0b010, 0>, PredRel; + defm CMPLTrr : NVJrr_base<"cmp.gt", "CMPLT", 0b011, 1>, PredRel; + defm CMPLTUrr : NVJrr_base<"cmp.gtu", "CMPLTU", 0b100, 1>, PredRel; } -// Multiclass for 2nd operand immediate, including -1. -multiclass NVJ_type_br_pred_imm<string NotStr, string OpcStr> { - defm Pt : NVJ_type_basic_imm<NotStr, OpcStr, "t">; - defm Pnt : NVJ_type_basic_imm<NotStr, OpcStr, "nt">; - defm Ptneg : NVJ_type_basic_neg<NotStr, OpcStr, "t">; - defm Pntneg : NVJ_type_basic_neg<NotStr, OpcStr, "nt">; -} +//===----------------------------------------------------------------------===// +// multiclass/template class for the new-value compare jumps instruction +// with a register and an unsigned immediate (U5) operand. +//===----------------------------------------------------------------------===// -// Multiclass for 2nd operand immediate, excluding -1. -multiclass NVJ_type_br_pred_imm_only<string NotStr, string OpcStr> { - defm Pt : NVJ_type_basic_imm<NotStr, OpcStr, "t">; - defm Pnt : NVJ_type_basic_imm<NotStr, OpcStr, "nt">; +let isExtendable = 1, opExtendable = 2, isExtentSigned = 1, opExtentBits = 11 in +class NVJri_template<string mnemonic, bits<3> majOp, bit isNegCond, + bit isTaken> + : NVInst_V4<(outs), + (ins IntRegs:$src1, u5Imm:$src2, brtarget:$offset), + "if ("#!if(isNegCond, "!","")#mnemonic#"($src1.new, #$src2)) jump:" + #!if(isTaken, "t","nt")#" $offset", + []>, Requires<[HasV4T]> { + + let isPredicatedFalse = isNegCond; + let isBrTaken = !if(isTaken, "true", "false"); + + bits<3> src1; + bits<5> src2; + bits<11> offset; + + let IClass = 0b0010; + let Inst{26} = 0b1; + let Inst{25-23} = majOp; + let Inst{22} = isNegCond; + let Inst{18-16} = src1; + let Inst{13} = isTaken; + let Inst{12-8} = src2; + let Inst{21-20} = offset{10-9}; + let Inst{7-1} = offset{8-2}; +} + +multiclass NVJri_cond<string mnemonic, bits<3> majOp, bit isNegCond> { + // Branch not taken: + def _nt_V4: NVJri_template<mnemonic, majOp, isNegCond, 0>; + // Branch taken: + def _t_V4: NVJri_template<mnemonic, majOp, isNegCond, 1>; +} + +multiclass NVJri_base<string mnemonic, string BaseOp, bits<3> majOp> { + let BaseOpcode = BaseOp#_NVJri in { + defm _t_Jumpnv : NVJri_cond<mnemonic, majOp, 0>; // True Cond + defm _f_Jumpnv : NVJri_cond<mnemonic, majOp, 1>; // False cond + } } -// Multiclass for tstbit, where 2nd operand is always #0. -multiclass NVJ_type_br_pred_tstbit<string NotStr, string OpcStr> { - defm Pt : NVJ_type_basic_tstbit<NotStr, OpcStr, "t">; - defm Pnt : NVJ_type_basic_tstbit<NotStr, OpcStr, "nt">; -} +// if ([!]cmp.eq(Ns.new,#U5)) jump:[n]t #r9:2 +// if ([!]cmp.gt(Ns.new,#U5)) jump:[n]t #r9:2 +// if ([!]cmp.gtu(Ns.new,#U5)) jump:[n]t #r9:2 -// Multiclass for GT. -multiclass NVJ_type_rr_ri<string OpcStr> { - defm rrNot : NVJ_type_br_pred_reg<"!", OpcStr>; - defm rr : NVJ_type_br_pred_reg<"", OpcStr>; - defm rrdnNot : NVJ_type_br_pred_2ndDotNew<"!", OpcStr>; - defm rrdn : NVJ_type_br_pred_2ndDotNew<"", OpcStr>; - defm riNot : NVJ_type_br_pred_imm<"!", OpcStr>; - defm ri : NVJ_type_br_pred_imm<"", OpcStr>; +let isPredicated = 1, isBranch = 1, isNewValue = 1, isTerminator = 1, + Defs = [PC], neverHasSideEffects = 1, validSubTargets = HasV4SubT in { + defm CMPEQri : NVJri_base<"cmp.eq", "CMPEQ", 0b000>, PredRel; + defm CMPGTri : NVJri_base<"cmp.gt", "CMPGT", 0b001>, PredRel; + defm CMPGTUri : NVJri_base<"cmp.gtu", "CMPGTU", 0b010>, PredRel; } -// Multiclass for EQ. -multiclass NVJ_type_rr_ri_no_2ndDotNew<string OpcStr> { - defm rrNot : NVJ_type_br_pred_reg<"!", OpcStr>; - defm rr : NVJ_type_br_pred_reg<"", OpcStr>; - defm riNot : NVJ_type_br_pred_imm<"!", OpcStr>; - defm ri : NVJ_type_br_pred_imm<"", OpcStr>; -} +//===----------------------------------------------------------------------===// +// multiclass/template class for the new-value compare jumps instruction +// with a register and an hardcoded 0/-1 immediate value. +//===----------------------------------------------------------------------===// -// Multiclass for GTU. -multiclass NVJ_type_rr_ri_no_nOne<string OpcStr> { - defm rrNot : NVJ_type_br_pred_reg<"!", OpcStr>; - defm rr : NVJ_type_br_pred_reg<"", OpcStr>; - defm rrdnNot : NVJ_type_br_pred_2ndDotNew<"!", OpcStr>; - defm rrdn : NVJ_type_br_pred_2ndDotNew<"", OpcStr>; - defm riNot : NVJ_type_br_pred_imm_only<"!", OpcStr>; - defm ri : NVJ_type_br_pred_imm_only<"", OpcStr>; +let isExtendable = 1, opExtendable = 1, isExtentSigned = 1, opExtentBits = 11 in +class NVJ_ConstImm_template<string mnemonic, bits<3> majOp, string ImmVal, + bit isNegCond, bit isTaken> + : NVInst_V4<(outs), + (ins IntRegs:$src1, brtarget:$offset), + "if ("#!if(isNegCond, "!","")#mnemonic + #"($src1.new, #"#ImmVal#")) jump:" + #!if(isTaken, "t","nt")#" $offset", + []>, Requires<[HasV4T]> { + + let isPredicatedFalse = isNegCond; + let isBrTaken = !if(isTaken, "true", "false"); + + bits<3> src1; + bits<11> offset; + let IClass = 0b0010; + let Inst{26} = 0b1; + let Inst{25-23} = majOp; + let Inst{22} = isNegCond; + let Inst{18-16} = src1; + let Inst{13} = isTaken; + let Inst{21-20} = offset{10-9}; + let Inst{7-1} = offset{8-2}; +} + +multiclass NVJ_ConstImm_cond<string mnemonic, bits<3> majOp, string ImmVal, + bit isNegCond> { + // Branch not taken: + def _nt_V4: NVJ_ConstImm_template<mnemonic, majOp, ImmVal, isNegCond, 0>; + // Branch taken: + def _t_V4: NVJ_ConstImm_template<mnemonic, majOp, ImmVal, isNegCond, 1>; +} + +multiclass NVJ_ConstImm_base<string mnemonic, string BaseOp, bits<3> majOp, + string ImmVal> { + let BaseOpcode = BaseOp#_NVJ_ConstImm in { + defm _t_Jumpnv : NVJ_ConstImm_cond<mnemonic, majOp, ImmVal, 0>; // True cond + defm _f_Jumpnv : NVJ_ConstImm_cond<mnemonic, majOp, ImmVal, 1>; // False Cond + } } -// Multiclass for tstbit. -multiclass NVJ_type_r0<string OpcStr> { - defm r0Not : NVJ_type_br_pred_tstbit<"!", OpcStr>; - defm r0 : NVJ_type_br_pred_tstbit<"", OpcStr>; - } - -// Base Multiclass for New Value Jump. -multiclass NVJ_type { - defm GT : NVJ_type_rr_ri<"cmp.gt">; - defm EQ : NVJ_type_rr_ri_no_2ndDotNew<"cmp.eq">; - defm GTU : NVJ_type_rr_ri_no_nOne<"cmp.gtu">; - defm TSTBIT : NVJ_type_r0<"tstbit">; -} +// if ([!]tstbit(Ns.new,#0)) jump:[n]t #r9:2 +// if ([!]cmp.eq(Ns.new,#-1)) jump:[n]t #r9:2 +// if ([!]cmp.gt(Ns.new,#-1)) jump:[n]t #r9:2 -let isBranch = 1, isTerminator=1, neverHasSideEffects = 1, Defs = [PC] in { - defm JMP_ : NVJ_type; +let isPredicated = 1, isBranch = 1, isNewValue = 1, isTerminator=1, + Defs = [PC], neverHasSideEffects = 1 in { + defm TSTBIT0 : NVJ_ConstImm_base<"tstbit", "TSTBIT", 0b011, "0">, PredRel; + defm CMPEQn1 : NVJ_ConstImm_base<"cmp.eq", "CMPEQ", 0b100, "-1">, PredRel; + defm CMPGTn1 : NVJ_ConstImm_base<"cmp.gt", "CMPGT", 0b101, "-1">, PredRel; } //===----------------------------------------------------------------------===// -// NV/J - -//===----------------------------------------------------------------------===// - -//===----------------------------------------------------------------------===// // XTYPE/ALU + //===----------------------------------------------------------------------===// @@ -2658,414 +1748,367 @@ def LSRd_rr_xor_V4 : MInst_acc<(outs DoubleRegs:$dst), // MEMOP: Word, Half, Byte //===----------------------------------------------------------------------===// +def MEMOPIMM : SDNodeXForm<imm, [{ + // Call the transformation function XformM5ToU5Imm to get the negative + // immediate's positive counterpart. + int32_t imm = N->getSExtValue(); + return XformM5ToU5Imm(imm); +}]>; + +def MEMOPIMM_HALF : SDNodeXForm<imm, [{ + // -1 .. -31 represented as 65535..65515 + // assigning to a short restores our desired signed value. + // Call the transformation function XformM5ToU5Imm to get the negative + // immediate's positive counterpart. + int16_t imm = N->getSExtValue(); + return XformM5ToU5Imm(imm); +}]>; + +def MEMOPIMM_BYTE : SDNodeXForm<imm, [{ + // -1 .. -31 represented as 255..235 + // assigning to a char restores our desired signed value. + // Call the transformation function XformM5ToU5Imm to get the negative + // immediate's positive counterpart. + int8_t imm = N->getSExtValue(); + return XformM5ToU5Imm(imm); +}]>; + +def SETMEMIMM : SDNodeXForm<imm, [{ + // Return the bit position we will set [0-31]. + // As an SDNode. + int32_t imm = N->getSExtValue(); + return XformMskToBitPosU5Imm(imm); +}]>; + +def CLRMEMIMM : SDNodeXForm<imm, [{ + // Return the bit position we will clear [0-31]. + // As an SDNode. + // we bit negate the value first + int32_t imm = ~(N->getSExtValue()); + return XformMskToBitPosU5Imm(imm); +}]>; + +def SETMEMIMM_SHORT : SDNodeXForm<imm, [{ + // Return the bit position we will set [0-15]. + // As an SDNode. + int16_t imm = N->getSExtValue(); + return XformMskToBitPosU4Imm(imm); +}]>; + +def CLRMEMIMM_SHORT : SDNodeXForm<imm, [{ + // Return the bit position we will clear [0-15]. + // As an SDNode. + // we bit negate the value first + int16_t imm = ~(N->getSExtValue()); + return XformMskToBitPosU4Imm(imm); +}]>; + +def SETMEMIMM_BYTE : SDNodeXForm<imm, [{ + // Return the bit position we will set [0-7]. + // As an SDNode. + int8_t imm = N->getSExtValue(); + return XformMskToBitPosU3Imm(imm); +}]>; + +def CLRMEMIMM_BYTE : SDNodeXForm<imm, [{ + // Return the bit position we will clear [0-7]. + // As an SDNode. + // we bit negate the value first + int8_t imm = ~(N->getSExtValue()); + return XformMskToBitPosU3Imm(imm); +}]>; + //===----------------------------------------------------------------------===// -// MEMOP: Word -// -// Implemented: -// MEMw_ADDi_indexed_V4 : memw(Rs+#u6:2)+=#U5 -// MEMw_SUBi_indexed_V4 : memw(Rs+#u6:2)-=#U5 -// MEMw_ADDr_indexed_V4 : memw(Rs+#u6:2)+=Rt -// MEMw_SUBr_indexed_V4 : memw(Rs+#u6:2)-=Rt -// MEMw_CLRr_indexed_V4 : memw(Rs+#u6:2)&=Rt -// MEMw_SETr_indexed_V4 : memw(Rs+#u6:2)|=Rt -// MEMw_ADDi_V4 : memw(Rs+#u6:2)+=#U5 -// MEMw_SUBi_V4 : memw(Rs+#u6:2)-=#U5 -// MEMw_ADDr_V4 : memw(Rs+#u6:2)+=Rt -// MEMw_SUBr_V4 : memw(Rs+#u6:2)-=Rt -// MEMw_CLRr_V4 : memw(Rs+#u6:2)&=Rt -// MEMw_SETr_V4 : memw(Rs+#u6:2)|=Rt -// -// Not implemented: -// MEMw_CLRi_indexed_V4 : memw(Rs+#u6:2)=clrbit(#U5) -// MEMw_SETi_indexed_V4 : memw(Rs+#u6:2)=setbit(#U5) -// MEMw_CLRi_V4 : memw(Rs+#u6:2)=clrbit(#U5) -// MEMw_SETi_V4 : memw(Rs+#u6:2)=setbit(#U5) +// Template class for MemOp instructions with the register value. //===----------------------------------------------------------------------===// +class MemOp_rr_base <string opc, bits<2> opcBits, Operand ImmOp, + string memOp, bits<2> memOpBits> : + MEMInst_V4<(outs), + (ins IntRegs:$base, ImmOp:$offset, IntRegs:$delta), + opc#"($base+#$offset)"#memOp#"$delta", + []>, + Requires<[HasV4T, UseMEMOP]> { + + bits<5> base; + bits<5> delta; + bits<32> offset; + bits<6> offsetBits; // memb - u6:0 , memh - u6:1, memw - u6:2 + + let offsetBits = !if (!eq(opcBits, 0b00), offset{5-0}, + !if (!eq(opcBits, 0b01), offset{6-1}, + !if (!eq(opcBits, 0b10), offset{7-2},0))); + + let IClass = 0b0011; + let Inst{27-24} = 0b1110; + let Inst{22-21} = opcBits; + let Inst{20-16} = base; + let Inst{13} = 0b0; + let Inst{12-7} = offsetBits; + let Inst{6-5} = memOpBits; + let Inst{4-0} = delta; +} +//===----------------------------------------------------------------------===// +// Template class for MemOp instructions with the immediate value. +//===----------------------------------------------------------------------===// +class MemOp_ri_base <string opc, bits<2> opcBits, Operand ImmOp, + string memOp, bits<2> memOpBits> : + MEMInst_V4 <(outs), + (ins IntRegs:$base, ImmOp:$offset, u5Imm:$delta), + opc#"($base+#$offset)"#memOp#"#$delta" + #!if(memOpBits{1},")", ""), // clrbit, setbit - include ')' + []>, + Requires<[HasV4T, UseMEMOP]> { + bits<5> base; + bits<5> delta; + bits<32> offset; + bits<6> offsetBits; // memb - u6:0 , memh - u6:1, memw - u6:2 -// memw(Rs+#u6:2) += #U5 -let AddedComplexity = 30 in -def MEMw_ADDi_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_2Imm:$offset, u5Imm:$addend), - "memw($base+#$offset) += #$addend", - []>, - Requires<[HasV4T, UseMEMOP]>; + let offsetBits = !if (!eq(opcBits, 0b00), offset{5-0}, + !if (!eq(opcBits, 0b01), offset{6-1}, + !if (!eq(opcBits, 0b10), offset{7-2},0))); -// memw(Rs+#u6:2) -= #U5 -let AddedComplexity = 30 in -def MEMw_SUBi_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_2Imm:$offset, u5Imm:$subend), - "memw($base+#$offset) -= #$subend", - []>, - Requires<[HasV4T, UseMEMOP]>; - -// memw(Rs+#u6:2) += Rt -let AddedComplexity = 30 in -def MEMw_ADDr_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_2Imm:$offset, IntRegs:$addend), - "memw($base+#$offset) += $addend", - [(store (add (load (add (i32 IntRegs:$base), u6_2ImmPred:$offset)), - (i32 IntRegs:$addend)), - (add (i32 IntRegs:$base), u6_2ImmPred:$offset))]>, - Requires<[HasV4T, UseMEMOP]>; - -// memw(Rs+#u6:2) -= Rt -let AddedComplexity = 30 in -def MEMw_SUBr_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_2Imm:$offset, IntRegs:$subend), - "memw($base+#$offset) -= $subend", - [(store (sub (load (add (i32 IntRegs:$base), u6_2ImmPred:$offset)), - (i32 IntRegs:$subend)), - (add (i32 IntRegs:$base), u6_2ImmPred:$offset))]>, - Requires<[HasV4T, UseMEMOP]>; - -// memw(Rs+#u6:2) &= Rt -let AddedComplexity = 30 in -def MEMw_ANDr_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_2Imm:$offset, IntRegs:$andend), - "memw($base+#$offset) &= $andend", - [(store (and (load (add (i32 IntRegs:$base), u6_2ImmPred:$offset)), - (i32 IntRegs:$andend)), - (add (i32 IntRegs:$base), u6_2ImmPred:$offset))]>, - Requires<[HasV4T, UseMEMOP]>; - -// memw(Rs+#u6:2) |= Rt -let AddedComplexity = 30 in -def MEMw_ORr_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_2Imm:$offset, IntRegs:$orend), - "memw($base+#$offset) |= $orend", - [(store (or (load (add (i32 IntRegs:$base), u6_2ImmPred:$offset)), - (i32 IntRegs:$orend)), - (add (i32 IntRegs:$base), u6_2ImmPred:$offset))]>, - Requires<[HasV4T, UseMEMOP]>; - -// memw(Rs+#u6:2) += #U5 -let AddedComplexity = 30 in -def MEMw_ADDi_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, u5Imm:$addend), - "memw($addr) += $addend", - []>, - Requires<[HasV4T, UseMEMOP]>; + let IClass = 0b0011; + let Inst{27-24} = 0b1111; + let Inst{22-21} = opcBits; + let Inst{20-16} = base; + let Inst{13} = 0b0; + let Inst{12-7} = offsetBits; + let Inst{6-5} = memOpBits; + let Inst{4-0} = delta; +} -// memw(Rs+#u6:2) -= #U5 -let AddedComplexity = 30 in -def MEMw_SUBi_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, u5Imm:$subend), - "memw($addr) -= $subend", - []>, - Requires<[HasV4T, UseMEMOP]>; - -// memw(Rs+#u6:2) += Rt -let AddedComplexity = 30 in -def MEMw_ADDr_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, IntRegs:$addend), - "memw($addr) += $addend", - [(store (add (load ADDRriU6_2:$addr), (i32 IntRegs:$addend)), - ADDRriU6_2:$addr)]>, - Requires<[HasV4T, UseMEMOP]>; - -// memw(Rs+#u6:2) -= Rt -let AddedComplexity = 30 in -def MEMw_SUBr_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, IntRegs:$subend), - "memw($addr) -= $subend", - [(store (sub (load ADDRriU6_2:$addr), (i32 IntRegs:$subend)), - ADDRriU6_2:$addr)]>, - Requires<[HasV4T, UseMEMOP]>; - -// memw(Rs+#u6:2) &= Rt -let AddedComplexity = 30 in -def MEMw_ANDr_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, IntRegs:$andend), - "memw($addr) &= $andend", - [(store (and (load ADDRriU6_2:$addr), (i32 IntRegs:$andend)), - ADDRriU6_2:$addr)]>, - Requires<[HasV4T, UseMEMOP]>; - -// memw(Rs+#u6:2) |= Rt -let AddedComplexity = 30 in -def MEMw_ORr_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, IntRegs:$orend), - "memw($addr) |= $orend", - [(store (or (load ADDRriU6_2:$addr), (i32 IntRegs:$orend)), - ADDRriU6_2:$addr)]>, - Requires<[HasV4T, UseMEMOP]>; +// multiclass to define MemOp instructions with register operand. +multiclass MemOp_rr<string opc, bits<2> opcBits, Operand ImmOp> { + def _ADD#NAME#_V4 : MemOp_rr_base <opc, opcBits, ImmOp, " += ", 0b00>; // add + def _SUB#NAME#_V4 : MemOp_rr_base <opc, opcBits, ImmOp, " -= ", 0b01>; // sub + def _AND#NAME#_V4 : MemOp_rr_base <opc, opcBits, ImmOp, " &= ", 0b10>; // and + def _OR#NAME#_V4 : MemOp_rr_base <opc, opcBits, ImmOp, " |= ", 0b11>; // or +} + +// multiclass to define MemOp instructions with immediate Operand. +multiclass MemOp_ri<string opc, bits<2> opcBits, Operand ImmOp> { + def _ADD#NAME#_V4 : MemOp_ri_base <opc, opcBits, ImmOp, " += ", 0b00 >; + def _SUB#NAME#_V4 : MemOp_ri_base <opc, opcBits, ImmOp, " -= ", 0b01 >; + def _CLRBIT#NAME#_V4 : MemOp_ri_base<opc, opcBits, ImmOp, " =clrbit(", 0b10>; + def _SETBIT#NAME#_V4 : MemOp_ri_base<opc, opcBits, ImmOp, " =setbit(", 0b11>; +} + +multiclass MemOp_base <string opc, bits<2> opcBits, Operand ImmOp> { + defm r : MemOp_rr <opc, opcBits, ImmOp>; + defm i : MemOp_ri <opc, opcBits, ImmOp>; +} + +// Define MemOp instructions. +let isExtendable = 1, opExtendable = 1, isExtentSigned = 0, +validSubTargets =HasV4SubT in { + let opExtentBits = 6, accessSize = ByteAccess in + defm MemOPb : MemOp_base <"memb", 0b00, u6_0Ext>; + + let opExtentBits = 7, accessSize = HalfWordAccess in + defm MemOPh : MemOp_base <"memh", 0b01, u6_1Ext>; + + let opExtentBits = 8, accessSize = WordAccess in + defm MemOPw : MemOp_base <"memw", 0b10, u6_2Ext>; +} //===----------------------------------------------------------------------===// -// MEMOP: Halfword -// -// Implemented: -// MEMh_ADDi_indexed_V4 : memw(Rs+#u6:2)+=#U5 -// MEMh_SUBi_indexed_V4 : memw(Rs+#u6:2)-=#U5 -// MEMh_ADDr_indexed_V4 : memw(Rs+#u6:2)+=Rt -// MEMh_SUBr_indexed_V4 : memw(Rs+#u6:2)-=Rt -// MEMh_CLRr_indexed_V4 : memw(Rs+#u6:2)&=Rt -// MEMh_SETr_indexed_V4 : memw(Rs+#u6:2)|=Rt -// MEMh_ADDi_V4 : memw(Rs+#u6:2)+=#U5 -// MEMh_SUBi_V4 : memw(Rs+#u6:2)-=#U5 -// MEMh_ADDr_V4 : memw(Rs+#u6:2)+=Rt -// MEMh_SUBr_V4 : memw(Rs+#u6:2)-=Rt -// MEMh_CLRr_V4 : memw(Rs+#u6:2)&=Rt -// MEMh_SETr_V4 : memw(Rs+#u6:2)|=Rt -// -// Not implemented: -// MEMh_CLRi_indexed_V4 : memw(Rs+#u6:2)=clrbit(#U5) -// MEMh_SETi_indexed_V4 : memw(Rs+#u6:2)=setbit(#U5) -// MEMh_CLRi_V4 : memw(Rs+#u6:2)=clrbit(#U5) -// MEMh_SETi_V4 : memw(Rs+#u6:2)=setbit(#U5) +// Multiclass to define 'Def Pats' for ALU operations on the memory +// Here value used for the ALU operation is an immediate value. +// mem[bh](Rs+#0) += #U5 +// mem[bh](Rs+#u6) += #U5 //===----------------------------------------------------------------------===// +multiclass MemOpi_u5Pats <PatFrag ldOp, PatFrag stOp, PatLeaf ExtPred, + InstHexagon MI, SDNode OpNode> { + let AddedComplexity = 180 in + def : Pat < (stOp (OpNode (ldOp IntRegs:$addr), u5ImmPred:$addend), + IntRegs:$addr), + (MI IntRegs:$addr, #0, u5ImmPred:$addend )>; -// memh(Rs+#u6:1) += #U5 -let AddedComplexity = 30 in -def MEMh_ADDi_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_1Imm:$offset, u5Imm:$addend), - "memh($base+#$offset) += $addend", - []>, - Requires<[HasV4T, UseMEMOP]>; + let AddedComplexity = 190 in + def : Pat <(stOp (OpNode (ldOp (add IntRegs:$base, ExtPred:$offset)), + u5ImmPred:$addend), + (add IntRegs:$base, ExtPred:$offset)), + (MI IntRegs:$base, ExtPred:$offset, u5ImmPred:$addend)>; +} -// memh(Rs+#u6:1) -= #U5 -let AddedComplexity = 30 in -def MEMh_SUBi_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_1Imm:$offset, u5Imm:$subend), - "memh($base+#$offset) -= $subend", - []>, - Requires<[HasV4T, UseMEMOP]>; - -// memh(Rs+#u6:1) += Rt -let AddedComplexity = 30 in -def MEMh_ADDr_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_1Imm:$offset, IntRegs:$addend), - "memh($base+#$offset) += $addend", - [(truncstorei16 (add (sextloadi16 (add (i32 IntRegs:$base), - u6_1ImmPred:$offset)), - (i32 IntRegs:$addend)), - (add (i32 IntRegs:$base), u6_1ImmPred:$offset))]>, - Requires<[HasV4T, UseMEMOP]>; - -// memh(Rs+#u6:1) -= Rt -let AddedComplexity = 30 in -def MEMh_SUBr_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_1Imm:$offset, IntRegs:$subend), - "memh($base+#$offset) -= $subend", - [(truncstorei16 (sub (sextloadi16 (add (i32 IntRegs:$base), - u6_1ImmPred:$offset)), - (i32 IntRegs:$subend)), - (add (i32 IntRegs:$base), u6_1ImmPred:$offset))]>, - Requires<[HasV4T, UseMEMOP]>; - -// memh(Rs+#u6:1) &= Rt -let AddedComplexity = 30 in -def MEMh_ANDr_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_1Imm:$offset, IntRegs:$andend), - "memh($base+#$offset) += $andend", - [(truncstorei16 (and (sextloadi16 (add (i32 IntRegs:$base), - u6_1ImmPred:$offset)), - (i32 IntRegs:$andend)), - (add (i32 IntRegs:$base), u6_1ImmPred:$offset))]>, - Requires<[HasV4T, UseMEMOP]>; - -// memh(Rs+#u6:1) |= Rt -let AddedComplexity = 30 in -def MEMh_ORr_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_1Imm:$offset, IntRegs:$orend), - "memh($base+#$offset) |= $orend", - [(truncstorei16 (or (sextloadi16 (add (i32 IntRegs:$base), - u6_1ImmPred:$offset)), - (i32 IntRegs:$orend)), - (add (i32 IntRegs:$base), u6_1ImmPred:$offset))]>, - Requires<[HasV4T, UseMEMOP]>; - -// memh(Rs+#u6:1) += #U5 -let AddedComplexity = 30 in -def MEMh_ADDi_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, u5Imm:$addend), - "memh($addr) += $addend", - []>, - Requires<[HasV4T, UseMEMOP]>; +multiclass MemOpi_u5ALUOp<PatFrag ldOp, PatFrag stOp, PatLeaf ExtPred, + InstHexagon addMI, InstHexagon subMI> { + defm : MemOpi_u5Pats<ldOp, stOp, ExtPred, addMI, add>; + defm : MemOpi_u5Pats<ldOp, stOp, ExtPred, subMI, sub>; +} -// memh(Rs+#u6:1) -= #U5 -let AddedComplexity = 30 in -def MEMh_SUBi_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, u5Imm:$subend), - "memh($addr) -= $subend", - []>, - Requires<[HasV4T, UseMEMOP]>; - -// memh(Rs+#u6:1) += Rt -let AddedComplexity = 30 in -def MEMh_ADDr_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, IntRegs:$addend), - "memh($addr) += $addend", - [(truncstorei16 (add (sextloadi16 ADDRriU6_1:$addr), - (i32 IntRegs:$addend)), ADDRriU6_1:$addr)]>, - Requires<[HasV4T, UseMEMOP]>; - -// memh(Rs+#u6:1) -= Rt -let AddedComplexity = 30 in -def MEMh_SUBr_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, IntRegs:$subend), - "memh($addr) -= $subend", - [(truncstorei16 (sub (sextloadi16 ADDRriU6_1:$addr), - (i32 IntRegs:$subend)), ADDRriU6_1:$addr)]>, - Requires<[HasV4T, UseMEMOP]>; - -// memh(Rs+#u6:1) &= Rt -let AddedComplexity = 30 in -def MEMh_ANDr_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, IntRegs:$andend), - "memh($addr) &= $andend", - [(truncstorei16 (and (sextloadi16 ADDRriU6_1:$addr), - (i32 IntRegs:$andend)), ADDRriU6_1:$addr)]>, - Requires<[HasV4T, UseMEMOP]>; - -// memh(Rs+#u6:1) |= Rt -let AddedComplexity = 30 in -def MEMh_ORr_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, IntRegs:$orend), - "memh($addr) |= $orend", - [(truncstorei16 (or (sextloadi16 ADDRriU6_1:$addr), - (i32 IntRegs:$orend)), ADDRriU6_1:$addr)]>, - Requires<[HasV4T, UseMEMOP]>; +multiclass MemOpi_u5ExtType<PatFrag ldOpByte, PatFrag ldOpHalf > { + // Half Word + defm : MemOpi_u5ALUOp <ldOpHalf, truncstorei16, u6_1ExtPred, + MemOPh_ADDi_V4, MemOPh_SUBi_V4>; + // Byte + defm : MemOpi_u5ALUOp <ldOpByte, truncstorei8, u6ExtPred, + MemOPb_ADDi_V4, MemOPb_SUBi_V4>; +} + +let Predicates = [HasV4T, UseMEMOP] in { + defm : MemOpi_u5ExtType<zextloadi8, zextloadi16>; // zero extend + defm : MemOpi_u5ExtType<sextloadi8, sextloadi16>; // sign extend + defm : MemOpi_u5ExtType<extloadi8, extloadi16>; // any extend + // Word + defm : MemOpi_u5ALUOp <load, store, u6_2ExtPred, MemOPw_ADDi_V4, + MemOPw_SUBi_V4>; +} //===----------------------------------------------------------------------===// -// MEMOP: Byte -// -// Implemented: -// MEMb_ADDi_indexed_V4 : memb(Rs+#u6:0)+=#U5 -// MEMb_SUBi_indexed_V4 : memb(Rs+#u6:0)-=#U5 -// MEMb_ADDr_indexed_V4 : memb(Rs+#u6:0)+=Rt -// MEMb_SUBr_indexed_V4 : memb(Rs+#u6:0)-=Rt -// MEMb_CLRr_indexed_V4 : memb(Rs+#u6:0)&=Rt -// MEMb_SETr_indexed_V4 : memb(Rs+#u6:0)|=Rt -// MEMb_ADDi_V4 : memb(Rs+#u6:0)+=#U5 -// MEMb_SUBi_V4 : memb(Rs+#u6:0)-=#U5 -// MEMb_ADDr_V4 : memb(Rs+#u6:0)+=Rt -// MEMb_SUBr_V4 : memb(Rs+#u6:0)-=Rt -// MEMb_CLRr_V4 : memb(Rs+#u6:0)&=Rt -// MEMb_SETr_V4 : memb(Rs+#u6:0)|=Rt -// -// Not implemented: -// MEMb_CLRi_indexed_V4 : memb(Rs+#u6:0)=clrbit(#U5) -// MEMb_SETi_indexed_V4 : memb(Rs+#u6:0)=setbit(#U5) -// MEMb_CLRi_V4 : memb(Rs+#u6:0)=clrbit(#U5) -// MEMb_SETi_V4 : memb(Rs+#u6:0)=setbit(#U5) +// multiclass to define 'Def Pats' for ALU operations on the memory. +// Here value used for the ALU operation is a negative value. +// mem[bh](Rs+#0) += #m5 +// mem[bh](Rs+#u6) += #m5 //===----------------------------------------------------------------------===// -// memb(Rs+#u6:0) += #U5 -let AddedComplexity = 30 in -def MEMb_ADDi_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_0Imm:$offset, u5Imm:$addend), - "memb($base+#$offset) += $addend", - []>, - Requires<[HasV4T, UseMEMOP]>; +multiclass MemOpi_m5Pats <PatFrag ldOp, PatFrag stOp, PatLeaf extPred, + PatLeaf immPred, ComplexPattern addrPred, + SDNodeXForm xformFunc, InstHexagon MI> { + let AddedComplexity = 190 in + def : Pat <(stOp (add (ldOp IntRegs:$addr), immPred:$subend), + IntRegs:$addr), + (MI IntRegs:$addr, #0, (xformFunc immPred:$subend) )>; -// memb(Rs+#u6:0) -= #U5 -let AddedComplexity = 30 in -def MEMb_SUBi_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_0Imm:$offset, u5Imm:$subend), - "memb($base+#$offset) -= $subend", - []>, - Requires<[HasV4T, UseMEMOP]>; - -// memb(Rs+#u6:0) += Rt -let AddedComplexity = 30 in -def MEMb_ADDr_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_0Imm:$offset, IntRegs:$addend), - "memb($base+#$offset) += $addend", - [(truncstorei8 (add (sextloadi8 (add (i32 IntRegs:$base), - u6_0ImmPred:$offset)), - (i32 IntRegs:$addend)), - (add (i32 IntRegs:$base), u6_0ImmPred:$offset))]>, - Requires<[HasV4T, UseMEMOP]>; - -// memb(Rs+#u6:0) -= Rt -let AddedComplexity = 30 in -def MEMb_SUBr_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_0Imm:$offset, IntRegs:$subend), - "memb($base+#$offset) -= $subend", - [(truncstorei8 (sub (sextloadi8 (add (i32 IntRegs:$base), - u6_0ImmPred:$offset)), - (i32 IntRegs:$subend)), - (add (i32 IntRegs:$base), u6_0ImmPred:$offset))]>, - Requires<[HasV4T, UseMEMOP]>; - -// memb(Rs+#u6:0) &= Rt -let AddedComplexity = 30 in -def MEMb_ANDr_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_0Imm:$offset, IntRegs:$andend), - "memb($base+#$offset) += $andend", - [(truncstorei8 (and (sextloadi8 (add (i32 IntRegs:$base), - u6_0ImmPred:$offset)), - (i32 IntRegs:$andend)), - (add (i32 IntRegs:$base), u6_0ImmPred:$offset))]>, - Requires<[HasV4T, UseMEMOP]>; - -// memb(Rs+#u6:0) |= Rt -let AddedComplexity = 30 in -def MEMb_ORr_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_0Imm:$offset, IntRegs:$orend), - "memb($base+#$offset) |= $orend", - [(truncstorei8 (or (sextloadi8 (add (i32 IntRegs:$base), - u6_0ImmPred:$offset)), - (i32 IntRegs:$orend)), - (add (i32 IntRegs:$base), u6_0ImmPred:$offset))]>, - Requires<[HasV4T, UseMEMOP]>; - -// memb(Rs+#u6:0) += #U5 -let AddedComplexity = 30 in -def MEMb_ADDi_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, u5Imm:$addend), - "memb($addr) += $addend", - []>, - Requires<[HasV4T, UseMEMOP]>; + let AddedComplexity = 195 in + def : Pat<(stOp (add (ldOp (add IntRegs:$base, extPred:$offset)), + immPred:$subend), + (add IntRegs:$base, extPred:$offset)), + (MI IntRegs:$base, extPred:$offset, (xformFunc immPred:$subend))>; +} -// memb(Rs+#u6:0) -= #U5 -let AddedComplexity = 30 in -def MEMb_SUBi_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, u5Imm:$subend), - "memb($addr) -= $subend", - []>, - Requires<[HasV4T, UseMEMOP]>; - -// memb(Rs+#u6:0) += Rt -let AddedComplexity = 30 in -def MEMb_ADDr_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, IntRegs:$addend), - "memb($addr) += $addend", - [(truncstorei8 (add (sextloadi8 ADDRriU6_0:$addr), - (i32 IntRegs:$addend)), ADDRriU6_0:$addr)]>, - Requires<[HasV4T, UseMEMOP]>; - -// memb(Rs+#u6:0) -= Rt -let AddedComplexity = 30 in -def MEMb_SUBr_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, IntRegs:$subend), - "memb($addr) -= $subend", - [(truncstorei8 (sub (sextloadi8 ADDRriU6_0:$addr), - (i32 IntRegs:$subend)), ADDRriU6_0:$addr)]>, - Requires<[HasV4T, UseMEMOP]>; - -// memb(Rs+#u6:0) &= Rt -let AddedComplexity = 30 in -def MEMb_ANDr_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, IntRegs:$andend), - "memb($addr) &= $andend", - [(truncstorei8 (and (sextloadi8 ADDRriU6_0:$addr), - (i32 IntRegs:$andend)), ADDRriU6_0:$addr)]>, - Requires<[HasV4T, UseMEMOP]>; - -// memb(Rs+#u6:0) |= Rt -let AddedComplexity = 30 in -def MEMb_ORr_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, IntRegs:$orend), - "memb($addr) |= $orend", - [(truncstorei8 (or (sextloadi8 ADDRriU6_0:$addr), - (i32 IntRegs:$orend)), ADDRriU6_0:$addr)]>, - Requires<[HasV4T, UseMEMOP]>; +multiclass MemOpi_m5ExtType<PatFrag ldOpByte, PatFrag ldOpHalf > { + // Half Word + defm : MemOpi_m5Pats <ldOpHalf, truncstorei16, u6_1ExtPred, m5HImmPred, + ADDRriU6_1, MEMOPIMM_HALF, MemOPh_SUBi_V4>; + // Byte + defm : MemOpi_m5Pats <ldOpByte, truncstorei8, u6ExtPred, m5BImmPred, + ADDRriU6_0, MEMOPIMM_BYTE, MemOPb_SUBi_V4>; +} + +let Predicates = [HasV4T, UseMEMOP] in { + defm : MemOpi_m5ExtType<zextloadi8, zextloadi16>; // zero extend + defm : MemOpi_m5ExtType<sextloadi8, sextloadi16>; // sign extend + defm : MemOpi_m5ExtType<extloadi8, extloadi16>; // any extend + + // Word + defm : MemOpi_m5Pats <load, store, u6_2ExtPred, m5ImmPred, + ADDRriU6_2, MEMOPIMM, MemOPw_SUBi_V4>; +} + +//===----------------------------------------------------------------------===// +// Multiclass to define 'def Pats' for bit operations on the memory. +// mem[bhw](Rs+#0) = [clrbit|setbit](#U5) +// mem[bhw](Rs+#u6) = [clrbit|setbit](#U5) +//===----------------------------------------------------------------------===// + +multiclass MemOpi_bitPats <PatFrag ldOp, PatFrag stOp, PatLeaf immPred, + PatLeaf extPred, ComplexPattern addrPred, + SDNodeXForm xformFunc, InstHexagon MI, SDNode OpNode> { + + // mem[bhw](Rs+#u6:[012]) = [clrbit|setbit](#U5) + let AddedComplexity = 250 in + def : Pat<(stOp (OpNode (ldOp (add IntRegs:$base, extPred:$offset)), + immPred:$bitend), + (add IntRegs:$base, extPred:$offset)), + (MI IntRegs:$base, extPred:$offset, (xformFunc immPred:$bitend))>; + + // mem[bhw](Rs+#0) = [clrbit|setbit](#U5) + let AddedComplexity = 225 in + def : Pat <(stOp (OpNode (ldOp addrPred:$addr), immPred:$bitend), + addrPred:$addr), + (MI IntRegs:$addr, #0, (xformFunc immPred:$bitend))>; +} + +multiclass MemOpi_bitExtType<PatFrag ldOpByte, PatFrag ldOpHalf > { + // Byte - clrbit + defm : MemOpi_bitPats<ldOpByte, truncstorei8, Clr3ImmPred, u6ExtPred, + ADDRriU6_0, CLRMEMIMM_BYTE, MemOPb_CLRBITi_V4, and>; + // Byte - setbit + defm : MemOpi_bitPats<ldOpByte, truncstorei8, Set3ImmPred, u6ExtPred, + ADDRriU6_0, SETMEMIMM_BYTE, MemOPb_SETBITi_V4, or>; + // Half Word - clrbit + defm : MemOpi_bitPats<ldOpHalf, truncstorei16, Clr4ImmPred, u6_1ExtPred, + ADDRriU6_1, CLRMEMIMM_SHORT, MemOPh_CLRBITi_V4, and>; + // Half Word - setbit + defm : MemOpi_bitPats<ldOpHalf, truncstorei16, Set4ImmPred, u6_1ExtPred, + ADDRriU6_1, SETMEMIMM_SHORT, MemOPh_SETBITi_V4, or>; +} + +let Predicates = [HasV4T, UseMEMOP] in { + // mem[bh](Rs+#0) = [clrbit|setbit](#U5) + // mem[bh](Rs+#u6:[01]) = [clrbit|setbit](#U5) + defm : MemOpi_bitExtType<zextloadi8, zextloadi16>; // zero extend + defm : MemOpi_bitExtType<sextloadi8, sextloadi16>; // sign extend + defm : MemOpi_bitExtType<extloadi8, extloadi16>; // any extend + + // memw(Rs+#0) = [clrbit|setbit](#U5) + // memw(Rs+#u6:2) = [clrbit|setbit](#U5) + defm : MemOpi_bitPats<load, store, Clr5ImmPred, u6_2ExtPred, ADDRriU6_2, + CLRMEMIMM, MemOPw_CLRBITi_V4, and>; + defm : MemOpi_bitPats<load, store, Set5ImmPred, u6_2ExtPred, ADDRriU6_2, + SETMEMIMM, MemOPw_SETBITi_V4, or>; +} +//===----------------------------------------------------------------------===// +// Multiclass to define 'def Pats' for ALU operations on the memory +// where addend is a register. +// mem[bhw](Rs+#0) [+-&|]= Rt +// mem[bhw](Rs+#U6:[012]) [+-&|]= Rt +//===----------------------------------------------------------------------===// + +multiclass MemOpr_Pats <PatFrag ldOp, PatFrag stOp, ComplexPattern addrPred, + PatLeaf extPred, InstHexagon MI, SDNode OpNode> { + let AddedComplexity = 141 in + // mem[bhw](Rs+#0) [+-&|]= Rt + def : Pat <(stOp (OpNode (ldOp addrPred:$addr), (i32 IntRegs:$addend)), + addrPred:$addr), + (MI IntRegs:$addr, #0, (i32 IntRegs:$addend) )>; + + // mem[bhw](Rs+#U6:[012]) [+-&|]= Rt + let AddedComplexity = 150 in + def : Pat <(stOp (OpNode (ldOp (add IntRegs:$base, extPred:$offset)), + (i32 IntRegs:$orend)), + (add IntRegs:$base, extPred:$offset)), + (MI IntRegs:$base, extPred:$offset, (i32 IntRegs:$orend) )>; +} + +multiclass MemOPr_ALUOp<PatFrag ldOp, PatFrag stOp, + ComplexPattern addrPred, PatLeaf extPred, + InstHexagon addMI, InstHexagon subMI, + InstHexagon andMI, InstHexagon orMI > { + + defm : MemOpr_Pats <ldOp, stOp, addrPred, extPred, addMI, add>; + defm : MemOpr_Pats <ldOp, stOp, addrPred, extPred, subMI, sub>; + defm : MemOpr_Pats <ldOp, stOp, addrPred, extPred, andMI, and>; + defm : MemOpr_Pats <ldOp, stOp, addrPred, extPred, orMI, or>; +} + +multiclass MemOPr_ExtType<PatFrag ldOpByte, PatFrag ldOpHalf > { + // Half Word + defm : MemOPr_ALUOp <ldOpHalf, truncstorei16, ADDRriU6_1, u6_1ExtPred, + MemOPh_ADDr_V4, MemOPh_SUBr_V4, + MemOPh_ANDr_V4, MemOPh_ORr_V4>; + // Byte + defm : MemOPr_ALUOp <ldOpByte, truncstorei8, ADDRriU6_0, u6ExtPred, + MemOPb_ADDr_V4, MemOPb_SUBr_V4, + MemOPb_ANDr_V4, MemOPb_ORr_V4>; +} + +// Define 'def Pats' for MemOps with register addend. +let Predicates = [HasV4T, UseMEMOP] in { + // Byte, Half Word + defm : MemOPr_ExtType<zextloadi8, zextloadi16>; // zero extend + defm : MemOPr_ExtType<sextloadi8, sextloadi16>; // sign extend + defm : MemOPr_ExtType<extloadi8, extloadi16>; // any extend + // Word + defm : MemOPr_ALUOp <load, store, ADDRriU6_2, u6_2ExtPred, MemOPw_ADDr_V4, + MemOPw_SUBr_V4, MemOPw_ANDr_V4, MemOPw_ORr_V4 >; +} //===----------------------------------------------------------------------===// // XTYPE/PRED + @@ -3146,7 +2189,7 @@ def CMPbEQri_V4 : MInst<(outs PredRegs:$dst), def : Pat <(brcond (i1 (setne (and (i32 IntRegs:$src1), 255), u8ImmPred:$src2)), bb:$offset), - (JMP_cNot (CMPbEQri_V4 (i32 IntRegs:$src1), u8ImmPred:$src2), + (JMP_f (CMPbEQri_V4 (i32 IntRegs:$src1), u8ImmPred:$src2), bb:$offset)>, Requires<[HasV4T]>; @@ -3521,8 +2564,9 @@ def NTSTBIT_ri : SInst<(outs PredRegs:$dst), (ins IntRegs:$src1, u5Imm:$src2), //Deallocate frame and return. // dealloc_return let isReturn = 1, isTerminator = 1, isBarrier = 1, isPredicable = 1, - Defs = [R29, R30, R31, PC], Uses = [R29, R31], neverHasSideEffects = 1 in { - def DEALLOC_RET_V4 : NVInst_V4<(outs), (ins i32imm:$amt1), + Defs = [R29, R30, R31, PC], Uses = [R30], neverHasSideEffects = 1 in { +let validSubTargets = HasV4SubT in + def DEALLOC_RET_V4 : LD0Inst<(outs), (ins), "dealloc_return", []>, Requires<[HasV4T]>; @@ -3531,9 +2575,10 @@ let isReturn = 1, isTerminator = 1, isBarrier = 1, isPredicable = 1, // Restore registers and dealloc return function call. let isCall = 1, isBarrier = 1, isReturn = 1, isTerminator = 1, Defs = [R29, R30, R31, PC] in { +let validSubTargets = HasV4SubT in def RESTORE_DEALLOC_RET_JMP_V4 : JInst<(outs), (ins calltarget:$dst), - "jump $dst // Restore_and_dealloc_return", + "jump $dst", []>, Requires<[HasV4T]>; } @@ -3541,9 +2586,10 @@ let isCall = 1, isBarrier = 1, isReturn = 1, isTerminator = 1, // Restore registers and dealloc frame before a tail call. let isCall = 1, isBarrier = 1, Defs = [R29, R30, R31, PC] in { +let validSubTargets = HasV4SubT in def RESTORE_DEALLOC_BEFORE_TAILCALL_V4 : JInst<(outs), (ins calltarget:$dst), - "call $dst // Restore_and_dealloc_before_tailcall", + "call $dst", []>, Requires<[HasV4T]>; } @@ -3560,10 +2606,11 @@ let isCall = 1, isBarrier = 1, // if (Ps) dealloc_return let isReturn = 1, isTerminator = 1, - Defs = [R29, R30, R31, PC], Uses = [R29, R31], neverHasSideEffects = 1, + Defs = [R29, R30, R31, PC], Uses = [R30], neverHasSideEffects = 1, isPredicated = 1 in { - def DEALLOC_RET_cPt_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, i32imm:$amt1), +let validSubTargets = HasV4SubT in + def DEALLOC_RET_cPt_V4 : LD0Inst<(outs), + (ins PredRegs:$src1), "if ($src1) dealloc_return", []>, Requires<[HasV4T]>; @@ -3571,10 +2618,10 @@ let isReturn = 1, isTerminator = 1, // if (!Ps) dealloc_return let isReturn = 1, isTerminator = 1, - Defs = [R29, R30, R31, PC], Uses = [R29, R31], neverHasSideEffects = 1, - isPredicated = 1 in { - def DEALLOC_RET_cNotPt_V4 : NVInst_V4<(outs), (ins PredRegs:$src1, - i32imm:$amt1), + Defs = [R29, R30, R31, PC], Uses = [R30], neverHasSideEffects = 1, + isPredicated = 1, isPredicatedFalse = 1 in { +let validSubTargets = HasV4SubT in + def DEALLOC_RET_cNotPt_V4 : LD0Inst<(outs), (ins PredRegs:$src1), "if (!$src1) dealloc_return", []>, Requires<[HasV4T]>; @@ -3582,10 +2629,10 @@ let isReturn = 1, isTerminator = 1, // if (Ps.new) dealloc_return:nt let isReturn = 1, isTerminator = 1, - Defs = [R29, R30, R31, PC], Uses = [R29, R31], neverHasSideEffects = 1, + Defs = [R29, R30, R31, PC], Uses = [R30], neverHasSideEffects = 1, isPredicated = 1 in { - def DEALLOC_RET_cdnPnt_V4 : NVInst_V4<(outs), (ins PredRegs:$src1, - i32imm:$amt1), +let validSubTargets = HasV4SubT in + def DEALLOC_RET_cdnPnt_V4 : LD0Inst<(outs), (ins PredRegs:$src1), "if ($src1.new) dealloc_return:nt", []>, Requires<[HasV4T]>; @@ -3593,10 +2640,10 @@ let isReturn = 1, isTerminator = 1, // if (!Ps.new) dealloc_return:nt let isReturn = 1, isTerminator = 1, - Defs = [R29, R30, R31, PC], Uses = [R29, R31], neverHasSideEffects = 1, - isPredicated = 1 in { - def DEALLOC_RET_cNotdnPnt_V4 : NVInst_V4<(outs), (ins PredRegs:$src1, - i32imm:$amt1), + Defs = [R29, R30, R31, PC], Uses = [R30], neverHasSideEffects = 1, + isPredicated = 1, isPredicatedFalse = 1 in { +let validSubTargets = HasV4SubT in + def DEALLOC_RET_cNotdnPnt_V4 : LD0Inst<(outs), (ins PredRegs:$src1), "if (!$src1.new) dealloc_return:nt", []>, Requires<[HasV4T]>; @@ -3604,21 +2651,21 @@ let isReturn = 1, isTerminator = 1, // if (Ps.new) dealloc_return:t let isReturn = 1, isTerminator = 1, - Defs = [R29, R30, R31, PC], Uses = [R29, R31], neverHasSideEffects = 1, + Defs = [R29, R30, R31, PC], Uses = [R30], neverHasSideEffects = 1, isPredicated = 1 in { - def DEALLOC_RET_cdnPt_V4 : NVInst_V4<(outs), (ins PredRegs:$src1, - i32imm:$amt1), +let validSubTargets = HasV4SubT in + def DEALLOC_RET_cdnPt_V4 : LD0Inst<(outs), (ins PredRegs:$src1), "if ($src1.new) dealloc_return:t", []>, Requires<[HasV4T]>; } -// if (!Ps.new) dealloc_return:nt +// if (!Ps.new) dealloc_return:nt let isReturn = 1, isTerminator = 1, - Defs = [R29, R30, R31, PC], Uses = [R29, R31], neverHasSideEffects = 1, - isPredicated = 1 in { - def DEALLOC_RET_cNotdnPt_V4 : NVInst_V4<(outs), (ins PredRegs:$src1, - i32imm:$amt1), + Defs = [R29, R30, R31, PC], Uses = [R30], neverHasSideEffects = 1, + isPredicated = 1, isPredicatedFalse = 1 in { +let validSubTargets = HasV4SubT in + def DEALLOC_RET_cNotdnPt_V4 : LD0Inst<(outs), (ins PredRegs:$src1), "if (!$src1.new) dealloc_return:t", []>, Requires<[HasV4T]>; @@ -3629,9 +2676,9 @@ let isReturn = 1, isTerminator = 1, multiclass ST_Abs_Predbase<string mnemonic, RegisterClass RC, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME#_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdressExt:$absaddr, RC: $src2), + (ins PredRegs:$src1, u0AlwaysExt:$absaddr, RC: $src2), !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ", ") ")#mnemonic#"(##$absaddr) = $src2", []>, @@ -3639,7 +2686,7 @@ multiclass ST_Abs_Predbase<string mnemonic, RegisterClass RC, bit isNot, } multiclass ST_Abs_Pred<string mnemonic, RegisterClass RC, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { defm _c#NAME : ST_Abs_Predbase<mnemonic, RC, PredNot, 0>; // Predicate new defm _cdn#NAME : ST_Abs_Predbase<mnemonic, RC, PredNot, 1>; @@ -3651,7 +2698,7 @@ multiclass ST_Abs<string mnemonic, string CextOp, RegisterClass RC> { let CextOpcode = CextOp, BaseOpcode = CextOp#_abs in { let opExtendable = 0, isPredicable = 1 in def NAME#_V4 : STInst2<(outs), - (ins globaladdressExt:$absaddr, RC:$src), + (ins u0AlwaysExt:$absaddr, RC:$src), mnemonic#"(##$absaddr) = $src", []>, Requires<[HasV4T]>; @@ -3665,9 +2712,9 @@ multiclass ST_Abs<string mnemonic, string CextOp, RegisterClass RC> { multiclass ST_Abs_Predbase_nv<string mnemonic, RegisterClass RC, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME#_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, globaladdressExt:$absaddr, RC: $src2), + (ins PredRegs:$src1, u0AlwaysExt:$absaddr, RC: $src2), !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ", ") ")#mnemonic#"(##$absaddr) = $src2.new", []>, @@ -3675,7 +2722,7 @@ multiclass ST_Abs_Predbase_nv<string mnemonic, RegisterClass RC, bit isNot, } multiclass ST_Abs_Pred_nv<string mnemonic, RegisterClass RC, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { defm _c#NAME : ST_Abs_Predbase_nv<mnemonic, RC, PredNot, 0>; // Predicate new defm _cdn#NAME : ST_Abs_Predbase_nv<mnemonic, RC, PredNot, 1>; @@ -3687,7 +2734,7 @@ multiclass ST_Abs_nv<string mnemonic, string CextOp, RegisterClass RC> { let CextOpcode = CextOp, BaseOpcode = CextOp#_abs in { let opExtendable = 0, isPredicable = 1 in def NAME#_nv_V4 : NVInst_V4<(outs), - (ins globaladdressExt:$absaddr, RC:$src), + (ins u0AlwaysExt:$absaddr, RC:$src), mnemonic#"(##$absaddr) = $src.new", []>, Requires<[HasV4T]>; @@ -3700,16 +2747,19 @@ multiclass ST_Abs_nv<string mnemonic, string CextOp, RegisterClass RC> { } let addrMode = Absolute in { + let accessSize = ByteAccess in defm STrib_abs : ST_Abs<"memb", "STrib", IntRegs>, ST_Abs_nv<"memb", "STrib", IntRegs>, AddrModeRel; + let accessSize = HalfWordAccess in defm STrih_abs : ST_Abs<"memh", "STrih", IntRegs>, ST_Abs_nv<"memh", "STrih", IntRegs>, AddrModeRel; + let accessSize = WordAccess in defm STriw_abs : ST_Abs<"memw", "STriw", IntRegs>, ST_Abs_nv<"memw", "STriw", IntRegs>, AddrModeRel; - let isNVStorable = 0 in + let accessSize = DoubleWordAccess, isNVStorable = 0 in defm STrid_abs : ST_Abs<"memd", "STrid", DoubleRegs>, AddrModeRel; } @@ -3730,11 +2780,115 @@ def : Pat<(store (i64 DoubleRegs:$src1), (STrid_abs_V4 tglobaladdr: $absaddr, DoubleRegs: $src1)>; } +//===----------------------------------------------------------------------===// +// multiclass for store instructions with GP-relative addressing mode. +// mem[bhwd](#global)=Rt +// if ([!]Pv[.new]) mem[bhwd](##global) = Rt +//===----------------------------------------------------------------------===// +let mayStore = 1, isNVStorable = 1 in +multiclass ST_GP<string mnemonic, string BaseOp, RegisterClass RC> { + let BaseOpcode = BaseOp, isPredicable = 1 in + def NAME#_V4 : STInst2<(outs), + (ins globaladdress:$global, RC:$src), + mnemonic#"(#$global) = $src", + []>; + + // When GP-relative instructions are predicated, their addressing mode is + // changed to absolute and they are always constant extended. + let BaseOpcode = BaseOp, isExtended = 1, opExtendable = 1, + isPredicated = 1 in { + defm Pt : ST_Abs_Pred <mnemonic, RC, 0>; + defm NotPt : ST_Abs_Pred <mnemonic, RC, 1>; + } +} + +let mayStore = 1, isNVStore = 1 in +multiclass ST_GP_nv<string mnemonic, string BaseOp, RegisterClass RC> { + let BaseOpcode = BaseOp, isPredicable = 1 in + def NAME#_nv_V4 : NVInst_V4<(outs), + (ins u0AlwaysExt:$global, RC:$src), + mnemonic#"(#$global) = $src.new", + []>, + Requires<[HasV4T]>; + + // When GP-relative instructions are predicated, their addressing mode is + // changed to absolute and they are always constant extended. + let BaseOpcode = BaseOp, isExtended = 1, opExtendable = 1, + isPredicated = 1 in { + defm Pt : ST_Abs_Pred_nv<mnemonic, RC, 0>; + defm NotPt : ST_Abs_Pred_nv<mnemonic, RC, 1>; + } +} + +let validSubTargets = HasV4SubT, neverHasSideEffects = 1 in { + let isNVStorable = 0 in + defm STd_GP : ST_GP <"memd", "STd_GP", DoubleRegs>, PredNewRel; + + defm STb_GP : ST_GP<"memb", "STb_GP", IntRegs>, + ST_GP_nv<"memb", "STb_GP", IntRegs>, NewValueRel; + defm STh_GP : ST_GP<"memh", "STh_GP", IntRegs>, + ST_GP_nv<"memh", "STh_GP", IntRegs>, NewValueRel; + defm STw_GP : ST_GP<"memw", "STw_GP", IntRegs>, + ST_GP_nv<"memw", "STw_GP", IntRegs>, NewValueRel; +} + +// 64 bit atomic store +def : Pat <(atomic_store_64 (HexagonCONST32_GP tglobaladdr:$global), + (i64 DoubleRegs:$src1)), + (STd_GP_V4 tglobaladdr:$global, (i64 DoubleRegs:$src1))>, + Requires<[HasV4T]>; + +// Map from store(globaladdress) -> memd(#foo) +let AddedComplexity = 100 in +def : Pat <(store (i64 DoubleRegs:$src1), + (HexagonCONST32_GP tglobaladdr:$global)), + (STd_GP_V4 tglobaladdr:$global, (i64 DoubleRegs:$src1))>; + +// 8 bit atomic store +def : Pat < (atomic_store_8 (HexagonCONST32_GP tglobaladdr:$global), + (i32 IntRegs:$src1)), + (STb_GP_V4 tglobaladdr:$global, (i32 IntRegs:$src1))>; + +// Map from store(globaladdress) -> memb(#foo) +let AddedComplexity = 100 in +def : Pat<(truncstorei8 (i32 IntRegs:$src1), + (HexagonCONST32_GP tglobaladdr:$global)), + (STb_GP_V4 tglobaladdr:$global, (i32 IntRegs:$src1))>; + +// Map from "i1 = constant<-1>; memw(CONST32(#foo)) = i1" +// to "r0 = 1; memw(#foo) = r0" +let AddedComplexity = 100 in +def : Pat<(store (i1 -1), (HexagonCONST32_GP tglobaladdr:$global)), + (STb_GP_V4 tglobaladdr:$global, (TFRI 1))>; + +def : Pat<(atomic_store_16 (HexagonCONST32_GP tglobaladdr:$global), + (i32 IntRegs:$src1)), + (STh_GP_V4 tglobaladdr:$global, (i32 IntRegs:$src1))>; + +// Map from store(globaladdress) -> memh(#foo) +let AddedComplexity = 100 in +def : Pat<(truncstorei16 (i32 IntRegs:$src1), + (HexagonCONST32_GP tglobaladdr:$global)), + (STh_GP_V4 tglobaladdr:$global, (i32 IntRegs:$src1))>; + +// 32 bit atomic store +def : Pat<(atomic_store_32 (HexagonCONST32_GP tglobaladdr:$global), + (i32 IntRegs:$src1)), + (STw_GP_V4 tglobaladdr:$global, (i32 IntRegs:$src1))>; + +// Map from store(globaladdress) -> memw(#foo) +let AddedComplexity = 100 in +def : Pat<(store (i32 IntRegs:$src1), (HexagonCONST32_GP tglobaladdr:$global)), + (STw_GP_V4 tglobaladdr:$global, (i32 IntRegs:$src1))>; + +//===----------------------------------------------------------------------===// +// Multiclass for the load instructions with absolute addressing mode. +//===----------------------------------------------------------------------===// multiclass LD_Abs_Predbase<string mnemonic, RegisterClass RC, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME : LDInst2<(outs RC:$dst), - (ins PredRegs:$src1, globaladdressExt:$absaddr), + (ins PredRegs:$src1, u0AlwaysExt:$absaddr), !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ", ") ")#"$dst = "#mnemonic#"(##$absaddr)", []>, @@ -3742,7 +2896,7 @@ multiclass LD_Abs_Predbase<string mnemonic, RegisterClass RC, bit isNot, } multiclass LD_Abs_Pred<string mnemonic, RegisterClass RC, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { defm _c#NAME : LD_Abs_Predbase<mnemonic, RC, PredNot, 0>; // Predicate new defm _cdn#NAME : LD_Abs_Predbase<mnemonic, RC, PredNot, 1>; @@ -3754,7 +2908,7 @@ multiclass LD_Abs<string mnemonic, string CextOp, RegisterClass RC> { let CextOpcode = CextOp, BaseOpcode = CextOp#_abs in { let opExtendable = 1, isPredicable = 1 in def NAME#_V4 : LDInst2<(outs RC:$dst), - (ins globaladdressExt:$absaddr), + (ins u0AlwaysExt:$absaddr), "$dst = "#mnemonic#"(##$absaddr)", []>, Requires<[HasV4T]>; @@ -3767,38 +2921,144 @@ multiclass LD_Abs<string mnemonic, string CextOp, RegisterClass RC> { } let addrMode = Absolute in { + let accessSize = ByteAccess in { defm LDrib_abs : LD_Abs<"memb", "LDrib", IntRegs>, AddrModeRel; defm LDriub_abs : LD_Abs<"memub", "LDriub", IntRegs>, AddrModeRel; + } + let accessSize = HalfWordAccess in { defm LDrih_abs : LD_Abs<"memh", "LDrih", IntRegs>, AddrModeRel; defm LDriuh_abs : LD_Abs<"memuh", "LDriuh", IntRegs>, AddrModeRel; + } + let accessSize = WordAccess in defm LDriw_abs : LD_Abs<"memw", "LDriw", IntRegs>, AddrModeRel; + + let accessSize = DoubleWordAccess in defm LDrid_abs : LD_Abs<"memd", "LDrid", DoubleRegs>, AddrModeRel; } -let Predicates = [HasV4T], AddedComplexity = 30 in +let Predicates = [HasV4T], AddedComplexity = 30 in { def : Pat<(i32 (load (HexagonCONST32 tglobaladdr:$absaddr))), (LDriw_abs_V4 tglobaladdr: $absaddr)>; -let Predicates = [HasV4T], AddedComplexity=30 in def : Pat<(i32 (sextloadi8 (HexagonCONST32 tglobaladdr:$absaddr))), (LDrib_abs_V4 tglobaladdr:$absaddr)>; -let Predicates = [HasV4T], AddedComplexity=30 in def : Pat<(i32 (zextloadi8 (HexagonCONST32 tglobaladdr:$absaddr))), (LDriub_abs_V4 tglobaladdr:$absaddr)>; -let Predicates = [HasV4T], AddedComplexity=30 in def : Pat<(i32 (sextloadi16 (HexagonCONST32 tglobaladdr:$absaddr))), (LDrih_abs_V4 tglobaladdr:$absaddr)>; -let Predicates = [HasV4T], AddedComplexity=30 in def : Pat<(i32 (zextloadi16 (HexagonCONST32 tglobaladdr:$absaddr))), (LDriuh_abs_V4 tglobaladdr:$absaddr)>; +} + +//===----------------------------------------------------------------------===// +// multiclass for load instructions with GP-relative addressing mode. +// Rx=mem[bhwd](##global) +// if ([!]Pv[.new]) Rx=mem[bhwd](##global) +//===----------------------------------------------------------------------===// +let neverHasSideEffects = 1, validSubTargets = HasV4SubT in +multiclass LD_GP<string mnemonic, string BaseOp, RegisterClass RC> { + let BaseOpcode = BaseOp in { + let isPredicable = 1 in + def NAME#_V4 : LDInst2<(outs RC:$dst), + (ins globaladdress:$global), + "$dst = "#mnemonic#"(#$global)", + []>; + + let isExtended = 1, opExtendable = 2, isPredicated = 1 in { + defm Pt_V4 : LD_Abs_Pred<mnemonic, RC, 0>; + defm NotPt_V4 : LD_Abs_Pred<mnemonic, RC, 1>; + } + } +} + +defm LDd_GP : LD_GP<"memd", "LDd_GP", DoubleRegs>, PredNewRel; +defm LDb_GP : LD_GP<"memb", "LDb_GP", IntRegs>, PredNewRel; +defm LDub_GP : LD_GP<"memub", "LDub_GP", IntRegs>, PredNewRel; +defm LDh_GP : LD_GP<"memh", "LDh_GP", IntRegs>, PredNewRel; +defm LDuh_GP : LD_GP<"memuh", "LDuh_GP", IntRegs>, PredNewRel; +defm LDw_GP : LD_GP<"memw", "LDw_GP", IntRegs>, PredNewRel; + +def : Pat <(atomic_load_64 (HexagonCONST32_GP tglobaladdr:$global)), + (i64 (LDd_GP_V4 tglobaladdr:$global))>; + +def : Pat <(atomic_load_32 (HexagonCONST32_GP tglobaladdr:$global)), + (i32 (LDw_GP_V4 tglobaladdr:$global))>; + +def : Pat <(atomic_load_16 (HexagonCONST32_GP tglobaladdr:$global)), + (i32 (LDuh_GP_V4 tglobaladdr:$global))>; + +def : Pat <(atomic_load_8 (HexagonCONST32_GP tglobaladdr:$global)), + (i32 (LDub_GP_V4 tglobaladdr:$global))>; + +// Map from load(globaladdress) -> memw(#foo + 0) +let AddedComplexity = 100 in +def : Pat <(i64 (load (HexagonCONST32_GP tglobaladdr:$global))), + (i64 (LDd_GP_V4 tglobaladdr:$global))>; + +// Map from Pd = load(globaladdress) -> Rd = memb(globaladdress), Pd = Rd +let AddedComplexity = 100 in +def : Pat <(i1 (load (HexagonCONST32_GP tglobaladdr:$global))), + (i1 (TFR_PdRs (i32 (LDb_GP_V4 tglobaladdr:$global))))>; + +// When the Interprocedural Global Variable optimizer realizes that a certain +// global variable takes only two constant values, it shrinks the global to +// a boolean. Catch those loads here in the following 3 patterns. +let AddedComplexity = 100 in +def : Pat <(i32 (extloadi1 (HexagonCONST32_GP tglobaladdr:$global))), + (i32 (LDb_GP_V4 tglobaladdr:$global))>; + +let AddedComplexity = 100 in +def : Pat <(i32 (sextloadi1 (HexagonCONST32_GP tglobaladdr:$global))), + (i32 (LDb_GP_V4 tglobaladdr:$global))>; + +// Map from load(globaladdress) -> memb(#foo) +let AddedComplexity = 100 in +def : Pat <(i32 (extloadi8 (HexagonCONST32_GP tglobaladdr:$global))), + (i32 (LDb_GP_V4 tglobaladdr:$global))>; + +// Map from load(globaladdress) -> memb(#foo) +let AddedComplexity = 100 in +def : Pat <(i32 (sextloadi8 (HexagonCONST32_GP tglobaladdr:$global))), + (i32 (LDb_GP_V4 tglobaladdr:$global))>; + +let AddedComplexity = 100 in +def : Pat <(i32 (zextloadi1 (HexagonCONST32_GP tglobaladdr:$global))), + (i32 (LDub_GP_V4 tglobaladdr:$global))>; + +// Map from load(globaladdress) -> memub(#foo) +let AddedComplexity = 100 in +def : Pat <(i32 (zextloadi8 (HexagonCONST32_GP tglobaladdr:$global))), + (i32 (LDub_GP_V4 tglobaladdr:$global))>; + +// Map from load(globaladdress) -> memh(#foo) +let AddedComplexity = 100 in +def : Pat <(i32 (extloadi16 (HexagonCONST32_GP tglobaladdr:$global))), + (i32 (LDh_GP_V4 tglobaladdr:$global))>; + +// Map from load(globaladdress) -> memh(#foo) +let AddedComplexity = 100 in +def : Pat <(i32 (sextloadi16 (HexagonCONST32_GP tglobaladdr:$global))), + (i32 (LDh_GP_V4 tglobaladdr:$global))>; + +// Map from load(globaladdress) -> memuh(#foo) +let AddedComplexity = 100 in +def : Pat <(i32 (zextloadi16 (HexagonCONST32_GP tglobaladdr:$global))), + (i32 (LDuh_GP_V4 tglobaladdr:$global))>; + +// Map from load(globaladdress) -> memw(#foo) +let AddedComplexity = 100 in +def : Pat <(i32 (load (HexagonCONST32_GP tglobaladdr:$global))), + (i32 (LDw_GP_V4 tglobaladdr:$global))>; + // Transfer global address into a register -let AddedComplexity=50, isMoveImm = 1, isReMaterializable = 1 in -def TFRI_V4 : ALU32_ri<(outs IntRegs:$dst), (ins globaladdress:$src1), - "$dst = ##$src1", +let isExtended = 1, opExtendable = 1, AddedComplexity=50, isMoveImm = 1, +isAsCheapAsAMove = 1, isReMaterializable = 1, validSubTargets = HasV4SubT in +def TFRI_V4 : ALU32_ri<(outs IntRegs:$dst), (ins s16Ext:$src1), + "$dst = #$src1", [(set IntRegs:$dst, (HexagonCONST32 tglobaladdr:$src1))]>, Requires<[HasV4T]>; @@ -3807,54 +3067,61 @@ def : Pat<(HexagonCONST32_GP tblockaddress:$src1), (TFRI_V4 tblockaddress:$src1)>, Requires<[HasV4T]>; -let AddedComplexity=50, neverHasSideEffects = 1, isPredicated = 1 in +let isExtended = 1, opExtendable = 2, AddedComplexity=50, +neverHasSideEffects = 1, isPredicated = 1, validSubTargets = HasV4SubT in def TFRI_cPt_V4 : ALU32_ri<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$src2), - "if($src1) $dst = ##$src2", + (ins PredRegs:$src1, s16Ext:$src2), + "if($src1) $dst = #$src2", []>, Requires<[HasV4T]>; -let AddedComplexity=50, neverHasSideEffects = 1, isPredicated = 1 in +let isExtended = 1, opExtendable = 2, AddedComplexity=50, isPredicatedFalse = 1, +neverHasSideEffects = 1, isPredicated = 1, validSubTargets = HasV4SubT in def TFRI_cNotPt_V4 : ALU32_ri<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$src2), - "if(!$src1) $dst = ##$src2", + (ins PredRegs:$src1, s16Ext:$src2), + "if(!$src1) $dst = #$src2", []>, Requires<[HasV4T]>; -let AddedComplexity=50, neverHasSideEffects = 1, isPredicated = 1 in +let isExtended = 1, opExtendable = 2, AddedComplexity=50, +neverHasSideEffects = 1, isPredicated = 1, validSubTargets = HasV4SubT in def TFRI_cdnPt_V4 : ALU32_ri<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$src2), - "if($src1.new) $dst = ##$src2", + (ins PredRegs:$src1, s16Ext:$src2), + "if($src1.new) $dst = #$src2", []>, Requires<[HasV4T]>; -let AddedComplexity=50, neverHasSideEffects = 1, isPredicated = 1 in +let isExtended = 1, opExtendable = 2, AddedComplexity=50, isPredicatedFalse = 1, +neverHasSideEffects = 1, isPredicated = 1, validSubTargets = HasV4SubT in def TFRI_cdnNotPt_V4 : ALU32_ri<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$src2), - "if(!$src1.new) $dst = ##$src2", + (ins PredRegs:$src1, s16Ext:$src2), + "if(!$src1.new) $dst = #$src2", []>, Requires<[HasV4T]>; let AddedComplexity = 50, Predicates = [HasV4T] in def : Pat<(HexagonCONST32_GP tglobaladdr:$src1), - (TFRI_V4 tglobaladdr:$src1)>; + (TFRI_V4 tglobaladdr:$src1)>, + Requires<[HasV4T]>; // Load - Indirect with long offset: These instructions take global address // as an operand -let AddedComplexity = 10 in +let isExtended = 1, opExtendable = 3, AddedComplexity = 40, +validSubTargets = HasV4SubT in def LDrid_ind_lo_V4 : LDInst<(outs DoubleRegs:$dst), - (ins IntRegs:$src1, u2Imm:$src2, globaladdress:$offset), + (ins IntRegs:$src1, u2Imm:$src2, globaladdressExt:$offset), "$dst=memd($src1<<#$src2+##$offset)", [(set (i64 DoubleRegs:$dst), (load (add (shl IntRegs:$src1, u2ImmPred:$src2), (HexagonCONST32 tglobaladdr:$offset))))]>, Requires<[HasV4T]>; -let AddedComplexity = 10 in +let AddedComplexity = 40 in multiclass LD_indirect_lo<string OpcStr, PatFrag OpNode> { +let isExtended = 1, opExtendable = 3, validSubTargets = HasV4SubT in def _lo_V4 : LDInst<(outs IntRegs:$dst), - (ins IntRegs:$src1, u2Imm:$src2, globaladdress:$offset), + (ins IntRegs:$src1, u2Imm:$src2, globaladdressExt:$offset), !strconcat("$dst = ", !strconcat(OpcStr, "($src1<<#$src2+##$offset)")), [(set IntRegs:$dst, @@ -3865,202 +3132,53 @@ multiclass LD_indirect_lo<string OpcStr, PatFrag OpNode> { defm LDrib_ind : LD_indirect_lo<"memb", sextloadi8>; defm LDriub_ind : LD_indirect_lo<"memub", zextloadi8>; +defm LDriub_ind_anyext : LD_indirect_lo<"memub", extloadi8>; defm LDrih_ind : LD_indirect_lo<"memh", sextloadi16>; defm LDriuh_ind : LD_indirect_lo<"memuh", zextloadi16>; +defm LDriuh_ind_anyext : LD_indirect_lo<"memuh", extloadi16>; defm LDriw_ind : LD_indirect_lo<"memw", load>; -// Store - Indirect with long offset: These instructions take global address -// as an operand -let AddedComplexity = 10 in -def STrid_ind_lo_V4 : STInst<(outs), - (ins IntRegs:$src1, u2Imm:$src2, globaladdress:$src3, - DoubleRegs:$src4), - "memd($src1<<#$src2+#$src3) = $src4", - [(store (i64 DoubleRegs:$src4), - (add (shl IntRegs:$src1, u2ImmPred:$src2), - (HexagonCONST32 tglobaladdr:$src3)))]>, - Requires<[HasV4T]>; - -let AddedComplexity = 10 in -multiclass ST_indirect_lo<string OpcStr, PatFrag OpNode> { - def _lo_V4 : STInst<(outs), - (ins IntRegs:$src1, u2Imm:$src2, globaladdress:$src3, - IntRegs:$src4), - !strconcat(OpcStr, "($src1<<#$src2+##$src3) = $src4"), - [(OpNode (i32 IntRegs:$src4), - (add (shl IntRegs:$src1, u2ImmPred:$src2), - (HexagonCONST32 tglobaladdr:$src3)))]>, - Requires<[HasV4T]>; -} - -defm STrib_ind : ST_indirect_lo<"memb", truncstorei8>; -defm STrih_ind : ST_indirect_lo<"memh", truncstorei16>; -defm STriw_ind : ST_indirect_lo<"memw", store>; - -// Store - absolute addressing mode: These instruction take constant -// value as the extended operand. -multiclass ST_absimm<string OpcStr> { -let isExtended = 1, opExtendable = 0, isPredicable = 1, -validSubTargets = HasV4SubT in - def _abs_V4 : STInst2<(outs), - (ins u0AlwaysExt:$src1, IntRegs:$src2), - !strconcat(OpcStr, "(##$src1) = $src2"), - []>, - Requires<[HasV4T]>; - -let isExtended = 1, opExtendable = 1, isPredicated = 1, -validSubTargets = HasV4SubT in { - def _abs_cPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, u0AlwaysExt:$src2, IntRegs:$src3), - !strconcat("if ($src1)", !strconcat(OpcStr, "(##$src2) = $src3")), - []>, - Requires<[HasV4T]>; - - def _abs_cNotPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, u0AlwaysExt:$src2, IntRegs:$src3), - !strconcat("if (!$src1)", !strconcat(OpcStr, "(##$src2) = $src3")), - []>, - Requires<[HasV4T]>; - - def _abs_cdnPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, u0AlwaysExt:$src2, IntRegs:$src3), - !strconcat("if ($src1.new)", - !strconcat(OpcStr, "(##$src2) = $src3")), - []>, - Requires<[HasV4T]>; - - def _abs_cdnNotPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, u0AlwaysExt:$src2, IntRegs:$src3), - !strconcat("if (!$src1.new)", - !strconcat(OpcStr, "(##$src2) = $src3")), - []>, - Requires<[HasV4T]>; -} - -let isExtended = 1, opExtendable = 0, mayStore = 1, isNVStore = 1, -validSubTargets = HasV4SubT in - def _abs_nv_V4 : NVInst_V4<(outs), - (ins u0AlwaysExt:$src1, IntRegs:$src2), - !strconcat(OpcStr, "(##$src1) = $src2.new"), - []>, - Requires<[HasV4T]>; - -let isExtended = 1, opExtendable = 1, mayStore = 1, isPredicated = 1, -isNVStore = 1, validSubTargets = HasV4SubT in { - def _abs_cPt_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, u0AlwaysExt:$src2, IntRegs:$src3), - !strconcat("if ($src1)", - !strconcat(OpcStr, "(##$src2) = $src3.new")), - []>, - Requires<[HasV4T]>; - - def _abs_cNotPt_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, u0AlwaysExt:$src2, IntRegs:$src3), - !strconcat("if (!$src1)", - !strconcat(OpcStr, "(##$src2) = $src3.new")), - []>, - Requires<[HasV4T]>; - - def _abs_cdnPt_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, u0AlwaysExt:$src2, IntRegs:$src3), - !strconcat("if ($src1.new)", - !strconcat(OpcStr, "(##$src2) = $src3.new")), - []>, - Requires<[HasV4T]>; - - def _abs_cdnNotPt_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, u0AlwaysExt:$src2, IntRegs:$src3), - !strconcat("if (!$src1.new)", - !strconcat(OpcStr, "(##$src2) = $src3.new")), - []>, - Requires<[HasV4T]>; -} -} +let AddedComplexity = 40 in +def : Pat <(i32 (sextloadi8 (add IntRegs:$src1, + (NumUsesBelowThresCONST32 tglobaladdr:$offset)))), + (i32 (LDrib_ind_lo_V4 IntRegs:$src1, 0, tglobaladdr:$offset))>, + Requires<[HasV4T]>; -defm STrib_imm : ST_absimm<"memb">; -defm STrih_imm : ST_absimm<"memh">; -defm STriw_imm : ST_absimm<"memw">; +let AddedComplexity = 40 in +def : Pat <(i32 (zextloadi8 (add IntRegs:$src1, + (NumUsesBelowThresCONST32 tglobaladdr:$offset)))), + (i32 (LDriub_ind_lo_V4 IntRegs:$src1, 0, tglobaladdr:$offset))>, + Requires<[HasV4T]>; let Predicates = [HasV4T], AddedComplexity = 30 in { def : Pat<(truncstorei8 (i32 IntRegs:$src1), u0AlwaysExtPred:$src2), - (STrib_imm_abs_V4 u0AlwaysExtPred:$src2, IntRegs: $src1)>; + (STrib_abs_V4 u0AlwaysExtPred:$src2, IntRegs: $src1)>; def : Pat<(truncstorei16 (i32 IntRegs:$src1), u0AlwaysExtPred:$src2), - (STrih_imm_abs_V4 u0AlwaysExtPred:$src2, IntRegs: $src1)>; + (STrih_abs_V4 u0AlwaysExtPred:$src2, IntRegs: $src1)>; def : Pat<(store (i32 IntRegs:$src1), u0AlwaysExtPred:$src2), - (STriw_imm_abs_V4 u0AlwaysExtPred:$src2, IntRegs: $src1)>; -} - -// Load - absolute addressing mode: These instruction take constant -// value as the extended operand - -multiclass LD_absimm<string OpcStr> { -let isExtended = 1, opExtendable = 1, isPredicable = 1, -validSubTargets = HasV4SubT in - def _abs_V4 : LDInst2<(outs IntRegs:$dst), - (ins u0AlwaysExt:$src), - !strconcat("$dst = ", - !strconcat(OpcStr, "(##$src)")), - []>, - Requires<[HasV4T]>; - -let isExtended = 1, opExtendable = 2, isPredicated = 1, -validSubTargets = HasV4SubT in { - def _abs_cPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, u0AlwaysExt:$src2), - !strconcat("if ($src1) $dst = ", - !strconcat(OpcStr, "(##$src2)")), - []>, - Requires<[HasV4T]>; - - def _abs_cNotPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, u0AlwaysExt:$src2), - !strconcat("if (!$src1) $dst = ", - !strconcat(OpcStr, "(##$src2)")), - []>, - Requires<[HasV4T]>; - - def _abs_cdnPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, u0AlwaysExt:$src2), - !strconcat("if ($src1.new) $dst = ", - !strconcat(OpcStr, "(##$src2)")), - []>, - Requires<[HasV4T]>; - - def _abs_cdnNotPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, u0AlwaysExt:$src2), - !strconcat("if (!$src1.new) $dst = ", - !strconcat(OpcStr, "(##$src2)")), - []>, - Requires<[HasV4T]>; + (STriw_abs_V4 u0AlwaysExtPred:$src2, IntRegs: $src1)>; } -} - -defm LDrib_imm : LD_absimm<"memb">; -defm LDriub_imm : LD_absimm<"memub">; -defm LDrih_imm : LD_absimm<"memh">; -defm LDriuh_imm : LD_absimm<"memuh">; -defm LDriw_imm : LD_absimm<"memw">; let Predicates = [HasV4T], AddedComplexity = 30 in { def : Pat<(i32 (load u0AlwaysExtPred:$src)), - (LDriw_imm_abs_V4 u0AlwaysExtPred:$src)>; + (LDriw_abs_V4 u0AlwaysExtPred:$src)>; def : Pat<(i32 (sextloadi8 u0AlwaysExtPred:$src)), - (LDrib_imm_abs_V4 u0AlwaysExtPred:$src)>; + (LDrib_abs_V4 u0AlwaysExtPred:$src)>; def : Pat<(i32 (zextloadi8 u0AlwaysExtPred:$src)), - (LDriub_imm_abs_V4 u0AlwaysExtPred:$src)>; + (LDriub_abs_V4 u0AlwaysExtPred:$src)>; def : Pat<(i32 (sextloadi16 u0AlwaysExtPred:$src)), - (LDrih_imm_abs_V4 u0AlwaysExtPred:$src)>; + (LDrih_abs_V4 u0AlwaysExtPred:$src)>; def : Pat<(i32 (zextloadi16 u0AlwaysExtPred:$src)), - (LDriuh_imm_abs_V4 u0AlwaysExtPred:$src)>; + (LDriuh_abs_V4 u0AlwaysExtPred:$src)>; } -// Indexed store double word - global address. +// Indexed store word - global address. // memw(Rs+#u6:2)=#S8 let AddedComplexity = 10 in def STriw_offset_ext_V4 : STInst<(outs), @@ -4070,6 +3188,93 @@ def STriw_offset_ext_V4 : STInst<(outs), (add IntRegs:$src1, u6_2ImmPred:$src2))]>, Requires<[HasV4T]>; +def : Pat<(i64 (ctlz (i64 DoubleRegs:$src1))), + (i64 (COMBINE_Ir_V4 (i32 0), (i32 (CTLZ64_rr DoubleRegs:$src1))))>, + Requires<[HasV4T]>; + +def : Pat<(i64 (cttz (i64 DoubleRegs:$src1))), + (i64 (COMBINE_Ir_V4 (i32 0), (i32 (CTTZ64_rr DoubleRegs:$src1))))>, + Requires<[HasV4T]>; + + +// i8 -> i64 loads +// We need a complexity of 120 here to overide preceeding handling of +// zextloadi8. +let Predicates = [HasV4T], AddedComplexity = 120 in { +def: Pat <(i64 (extloadi8 (NumUsesBelowThresCONST32 tglobaladdr:$addr))), + (i64 (COMBINE_Ir_V4 0, (LDrib_abs_V4 tglobaladdr:$addr)))>; + +def: Pat <(i64 (zextloadi8 (NumUsesBelowThresCONST32 tglobaladdr:$addr))), + (i64 (COMBINE_Ir_V4 0, (LDriub_abs_V4 tglobaladdr:$addr)))>; + +def: Pat <(i64 (sextloadi8 (NumUsesBelowThresCONST32 tglobaladdr:$addr))), + (i64 (SXTW (LDrib_abs_V4 tglobaladdr:$addr)))>; + +def: Pat <(i64 (extloadi8 FoldGlobalAddr:$addr)), + (i64 (COMBINE_Ir_V4 0, (LDrib_abs_V4 FoldGlobalAddr:$addr)))>; + +def: Pat <(i64 (zextloadi8 FoldGlobalAddr:$addr)), + (i64 (COMBINE_Ir_V4 0, (LDriub_abs_V4 FoldGlobalAddr:$addr)))>; + +def: Pat <(i64 (sextloadi8 FoldGlobalAddr:$addr)), + (i64 (SXTW (LDrib_abs_V4 FoldGlobalAddr:$addr)))>; +} +// i16 -> i64 loads +// We need a complexity of 120 here to overide preceeding handling of +// zextloadi16. +let AddedComplexity = 120 in { +def: Pat <(i64 (extloadi16 (NumUsesBelowThresCONST32 tglobaladdr:$addr))), + (i64 (COMBINE_Ir_V4 0, (LDrih_abs_V4 tglobaladdr:$addr)))>, + Requires<[HasV4T]>; + +def: Pat <(i64 (zextloadi16 (NumUsesBelowThresCONST32 tglobaladdr:$addr))), + (i64 (COMBINE_Ir_V4 0, (LDriuh_abs_V4 tglobaladdr:$addr)))>, + Requires<[HasV4T]>; + +def: Pat <(i64 (sextloadi16 (NumUsesBelowThresCONST32 tglobaladdr:$addr))), + (i64 (SXTW (LDrih_abs_V4 tglobaladdr:$addr)))>, + Requires<[HasV4T]>; + +def: Pat <(i64 (extloadi16 FoldGlobalAddr:$addr)), + (i64 (COMBINE_Ir_V4 0, (LDrih_abs_V4 FoldGlobalAddr:$addr)))>, + Requires<[HasV4T]>; + +def: Pat <(i64 (zextloadi16 FoldGlobalAddr:$addr)), + (i64 (COMBINE_Ir_V4 0, (LDriuh_abs_V4 FoldGlobalAddr:$addr)))>, + Requires<[HasV4T]>; + +def: Pat <(i64 (sextloadi16 FoldGlobalAddr:$addr)), + (i64 (SXTW (LDrih_abs_V4 FoldGlobalAddr:$addr)))>, + Requires<[HasV4T]>; +} +// i32->i64 loads +// We need a complexity of 120 here to overide preceeding handling of +// zextloadi32. +let AddedComplexity = 120 in { +def: Pat <(i64 (extloadi32 (NumUsesBelowThresCONST32 tglobaladdr:$addr))), + (i64 (COMBINE_Ir_V4 0, (LDriw_abs_V4 tglobaladdr:$addr)))>, + Requires<[HasV4T]>; + +def: Pat <(i64 (zextloadi32 (NumUsesBelowThresCONST32 tglobaladdr:$addr))), + (i64 (COMBINE_Ir_V4 0, (LDriw_abs_V4 tglobaladdr:$addr)))>, + Requires<[HasV4T]>; + +def: Pat <(i64 (sextloadi32 (NumUsesBelowThresCONST32 tglobaladdr:$addr))), + (i64 (SXTW (LDriw_abs_V4 tglobaladdr:$addr)))>, + Requires<[HasV4T]>; + +def: Pat <(i64 (extloadi32 FoldGlobalAddr:$addr)), + (i64 (COMBINE_Ir_V4 0, (LDriw_abs_V4 FoldGlobalAddr:$addr)))>, + Requires<[HasV4T]>; + +def: Pat <(i64 (zextloadi32 FoldGlobalAddr:$addr)), + (i64 (COMBINE_Ir_V4 0, (LDriw_abs_V4 FoldGlobalAddr:$addr)))>, + Requires<[HasV4T]>; + +def: Pat <(i64 (sextloadi32 FoldGlobalAddr:$addr)), + (i64 (SXTW (LDriw_abs_V4 FoldGlobalAddr:$addr)))>, + Requires<[HasV4T]>; +} // Indexed store double word - global address. // memw(Rs+#u6:2)=#S8 diff --git a/lib/Target/Hexagon/HexagonInstrInfoV5.td b/lib/Target/Hexagon/HexagonInstrInfoV5.td index 92d098c..9da6074 100644 --- a/lib/Target/Hexagon/HexagonInstrInfoV5.td +++ b/lib/Target/Hexagon/HexagonInstrInfoV5.td @@ -26,22 +26,29 @@ def CONST32_Float_Real : LDInst<(outs IntRegs:$dst), (ins f32imm:$src1), // Only works with single precision fp value. // For double precision, use CONST64_float_real, as 64bit transfer // can only hold 40-bit values - 32 from const ext + 8 bit immediate. -let isMoveImm = 1, isReMaterializable = 1, isPredicable = 1 in -def TFRI_f : ALU32_ri<(outs IntRegs:$dst), (ins f32imm:$src1), - "$dst = ##$src1", +// Make sure that complexity is more than the CONST32 pattern in +// HexagonInstrInfo.td patterns. +let isExtended = 1, opExtendable = 1, isMoveImm = 1, isReMaterializable = 1, +isPredicable = 1, AddedComplexity = 30, validSubTargets = HasV5SubT, +isCodeGenOnly = 1 in +def TFRI_f : ALU32_ri<(outs IntRegs:$dst), (ins f32Ext:$src1), + "$dst = #$src1", [(set IntRegs:$dst, fpimm:$src1)]>, Requires<[HasV5T]>; +let isExtended = 1, opExtendable = 2, isPredicated = 1, +neverHasSideEffects = 1, validSubTargets = HasV5SubT in def TFRI_cPt_f : ALU32_ri<(outs IntRegs:$dst), - (ins PredRegs:$src1, f32imm:$src2), - "if ($src1) $dst = ##$src2", + (ins PredRegs:$src1, f32Ext:$src2), + "if ($src1) $dst = #$src2", []>, Requires<[HasV5T]>; -let isPredicated = 1 in +let isExtended = 1, opExtendable = 2, isPredicated = 1, isPredicatedFalse = 1, +neverHasSideEffects = 1, validSubTargets = HasV5SubT in def TFRI_cNotPt_f : ALU32_ri<(outs IntRegs:$dst), - (ins PredRegs:$src1, f32imm:$src2), - "if (!$src1) $dst = ##$src2", + (ins PredRegs:$src1, f32Ext:$src2), + "if (!$src1) $dst =#$src2", []>, Requires<[HasV5T]>; diff --git a/lib/Target/Hexagon/HexagonMCInst.h b/lib/Target/Hexagon/HexagonMCInst.h deleted file mode 100644 index e16636e..0000000 --- a/lib/Target/Hexagon/HexagonMCInst.h +++ /dev/null @@ -1,41 +0,0 @@ -//===- HexagonMCInst.h - Hexagon sub-class of MCInst ----------------------===// -// -// The LLVM Compiler Infrastructure -// -// This file is distributed under the University of Illinois Open Source -// License. See LICENSE.TXT for details. -// -//===----------------------------------------------------------------------===// -// -// This class extends MCInst to allow some VLIW annotation. -// -//===----------------------------------------------------------------------===// - -#ifndef HEXAGONMCINST_H -#define HEXAGONMCINST_H - -#include "llvm/CodeGen/MachineInstr.h" -#include "llvm/MC/MCInst.h" - -namespace llvm { - class HexagonMCInst: public MCInst { - // Packet start and end markers - unsigned startPacket: 1, endPacket: 1; - const MachineInstr *MachineI; - public: - explicit HexagonMCInst(): MCInst(), - startPacket(0), endPacket(0) {} - - const MachineInstr* getMI() const { return MachineI; } - - void setMI(const MachineInstr *MI) { MachineI = MI; } - - bool isStartPacket() const { return (startPacket); } - bool isEndPacket() const { return (endPacket); } - - void setStartPacket(bool yes) { startPacket = yes; } - void setEndPacket(bool yes) { endPacket = yes; } - }; -} - -#endif diff --git a/lib/Target/Hexagon/HexagonMachineFunctionInfo.h b/lib/Target/Hexagon/HexagonMachineFunctionInfo.h index 0318c51..bd7b26a 100644 --- a/lib/Target/Hexagon/HexagonMachineFunctionInfo.h +++ b/lib/Target/Hexagon/HexagonMachineFunctionInfo.h @@ -29,15 +29,18 @@ class HexagonMachineFunctionInfo : public MachineFunctionInfo { std::vector<MachineInstr*> AllocaAdjustInsts; int VarArgsFrameIndex; bool HasClobberLR; + bool HasEHReturn; std::map<const MachineInstr*, unsigned> PacketInfo; public: - HexagonMachineFunctionInfo() : SRetReturnReg(0), HasClobberLR(0) {} + HexagonMachineFunctionInfo() : SRetReturnReg(0), HasClobberLR(0), + HasEHReturn(false) {} HexagonMachineFunctionInfo(MachineFunction &MF) : SRetReturnReg(0), - HasClobberLR(0) {} + HasClobberLR(0), + HasEHReturn(false) {} unsigned getSRetReturnReg() const { return SRetReturnReg; } void setSRetReturnReg(unsigned Reg) { SRetReturnReg = Reg; } @@ -69,6 +72,8 @@ public: void setHasClobberLR(bool v) { HasClobberLR = v; } bool hasClobberLR() const { return HasClobberLR; } + bool hasEHReturn() const { return HasEHReturn; }; + void setHasEHReturn(bool H = true) { HasEHReturn = H; }; }; } // End llvm namespace diff --git a/lib/Target/Hexagon/HexagonMachineScheduler.cpp b/lib/Target/Hexagon/HexagonMachineScheduler.cpp index 1388ad4..10bb3e9 100644 --- a/lib/Target/Hexagon/HexagonMachineScheduler.cpp +++ b/lib/Target/Hexagon/HexagonMachineScheduler.cpp @@ -195,7 +195,6 @@ void VLIWMachineScheduler::schedule() { void ConvergingVLIWScheduler::initialize(ScheduleDAGMI *dag) { DAG = static_cast<VLIWMachineScheduler*>(dag); SchedModel = DAG->getSchedModel(); - TRI = DAG->TRI; Top.init(DAG, SchedModel); Bot.init(DAG, SchedModel); @@ -209,6 +208,8 @@ void ConvergingVLIWScheduler::initialize(ScheduleDAGMI *dag) { Top.HazardRec = TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG); Bot.HazardRec = TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG); + delete Top.ResourceModel; + delete Bot.ResourceModel; Top.ResourceModel = new VLIWResourceModel(TM, DAG->getSchedModel()); Bot.ResourceModel = new VLIWResourceModel(TM, DAG->getSchedModel()); @@ -223,7 +224,7 @@ void ConvergingVLIWScheduler::releaseTopNode(SUnit *SU) { for (SUnit::succ_iterator I = SU->Preds.begin(), E = SU->Preds.end(); I != E; ++I) { unsigned PredReadyCycle = I->getSUnit()->TopReadyCycle; - unsigned MinLatency = I->getMinLatency(); + unsigned MinLatency = I->getLatency(); #ifndef NDEBUG Top.MaxMinLatency = std::max(MinLatency, Top.MaxMinLatency); #endif @@ -242,7 +243,7 @@ void ConvergingVLIWScheduler::releaseBottomNode(SUnit *SU) { for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end(); I != E; ++I) { unsigned SuccReadyCycle = I->getSUnit()->BotReadyCycle; - unsigned MinLatency = I->getMinLatency(); + unsigned MinLatency = I->getLatency(); #ifndef NDEBUG Bot.MaxMinLatency = std::max(MinLatency, Bot.MaxMinLatency); #endif @@ -409,7 +410,7 @@ void ConvergingVLIWScheduler::traceCandidate(const char *Label, SUnit *SU, PressureElement P) { dbgs() << Label << " " << Q.getName() << " "; if (P.isValid()) - dbgs() << TRI->getRegPressureSetName(P.PSetID) << ":" << P.UnitIncrease + dbgs() << DAG->TRI->getRegPressureSetName(P.PSetID) << ":" << P.UnitIncrease << " "; else dbgs() << " "; diff --git a/lib/Target/Hexagon/HexagonMachineScheduler.h b/lib/Target/Hexagon/HexagonMachineScheduler.h index f68dadf..171193e 100644 --- a/lib/Target/Hexagon/HexagonMachineScheduler.h +++ b/lib/Target/Hexagon/HexagonMachineScheduler.h @@ -190,7 +190,6 @@ class ConvergingVLIWScheduler : public MachineSchedStrategy { VLIWMachineScheduler *DAG; const TargetSchedModel *SchedModel; - const TargetRegisterInfo *TRI; // State of the top and bottom scheduled instruction boundaries. SchedBoundary Top; @@ -205,7 +204,7 @@ public: }; ConvergingVLIWScheduler(): - DAG(0), SchedModel(0), TRI(0), Top(TopQID, "TopQ"), Bot(BotQID, "BotQ") {} + DAG(0), SchedModel(0), Top(TopQID, "TopQ"), Bot(BotQID, "BotQ") {} virtual void initialize(ScheduleDAGMI *dag); diff --git a/lib/Target/Hexagon/HexagonNewValueJump.cpp b/lib/Target/Hexagon/HexagonNewValueJump.cpp index cd3d289..f7c4513 100644 --- a/lib/Target/Hexagon/HexagonNewValueJump.cpp +++ b/lib/Target/Hexagon/HexagonNewValueJump.cpp @@ -22,29 +22,31 @@ // //===----------------------------------------------------------------------===// #define DEBUG_TYPE "hexagon-nvj" -#include "Hexagon.h" -#include "HexagonInstrInfo.h" -#include "HexagonMachineFunctionInfo.h" -#include "HexagonRegisterInfo.h" -#include "HexagonSubtarget.h" -#include "HexagonTargetMachine.h" +#include "llvm/PassSupport.h" +#include "llvm/Support/Compiler.h" +#include "llvm/Support/Debug.h" #include "llvm/ADT/DenseMap.h" #include "llvm/ADT/Statistic.h" -#include "llvm/CodeGen/LiveVariables.h" -#include "llvm/CodeGen/MachineFunctionAnalysis.h" -#include "llvm/CodeGen/MachineFunctionPass.h" -#include "llvm/CodeGen/MachineInstrBuilder.h" -#include "llvm/CodeGen/MachineRegisterInfo.h" #include "llvm/CodeGen/Passes.h" #include "llvm/CodeGen/ScheduleDAGInstrs.h" -#include "llvm/PassSupport.h" -#include "llvm/Support/CommandLine.h" -#include "llvm/Support/Compiler.h" -#include "llvm/Support/Debug.h" -#include "llvm/Target/TargetInstrInfo.h" +#include "llvm/CodeGen/MachineInstrBuilder.h" +#include "llvm/CodeGen/MachineFunctionPass.h" +#include "llvm/CodeGen/LiveVariables.h" +#include "llvm/CodeGen/MachineRegisterInfo.h" +#include "llvm/CodeGen/MachineFunctionAnalysis.h" #include "llvm/Target/TargetMachine.h" +#include "llvm/Target/TargetInstrInfo.h" #include "llvm/Target/TargetRegisterInfo.h" +#include "Hexagon.h" +#include "HexagonTargetMachine.h" +#include "HexagonRegisterInfo.h" +#include "HexagonSubtarget.h" +#include "HexagonInstrInfo.h" +#include "HexagonMachineFunctionInfo.h" + #include <map> + +#include "llvm/Support/CommandLine.h" using namespace llvm; STATISTIC(NumNVJGenerated, "Number of New Value Jump Instructions created"); @@ -57,6 +59,11 @@ static cl::opt<bool> DisableNewValueJumps("disable-nvjump", cl::Hidden, cl::ZeroOrMore, cl::init(false), cl::desc("Disable New Value Jumps")); +namespace llvm { + void initializeHexagonNewValueJumpPass(PassRegistry&); +} + + namespace { struct HexagonNewValueJump : public MachineFunctionPass { const HexagonInstrInfo *QII; @@ -65,9 +72,12 @@ namespace { public: static char ID; - HexagonNewValueJump() : MachineFunctionPass(ID) { } + HexagonNewValueJump() : MachineFunctionPass(ID) { + initializeHexagonNewValueJumpPass(*PassRegistry::getPassRegistry()); + } virtual void getAnalysisUsage(AnalysisUsage &AU) const { + AU.addRequired<MachineBranchProbabilityInfo>(); MachineFunctionPass::getAnalysisUsage(AU); } @@ -78,6 +88,8 @@ namespace { virtual bool runOnMachineFunction(MachineFunction &Fn); private: + /// \brief A handle to the branch probability pass. + const MachineBranchProbabilityInfo *MBPI; }; @@ -85,6 +97,13 @@ namespace { char HexagonNewValueJump::ID = 0; +INITIALIZE_PASS_BEGIN(HexagonNewValueJump, "hexagon-nvj", + "Hexagon NewValueJump", false, false) +INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo) +INITIALIZE_PASS_END(HexagonNewValueJump, "hexagon-nvj", + "Hexagon NewValueJump", false, false) + + // We have identified this II could be feeder to NVJ, // verify that it can be. static bool canBeFeederToNewValueJump(const HexagonInstrInfo *QII, @@ -208,14 +227,10 @@ static bool canCompareBeNewValueJump(const HexagonInstrInfo *QII, // range specified by the arch. if (!secondReg) { int64_t v = MI->getOperand(2).getImm(); - if (MI->getOpcode() == Hexagon::CMPGEri || - (MI->getOpcode() == Hexagon::CMPGEUri && v > 0)) - --v; if (!(isUInt<5>(v) || ((MI->getOpcode() == Hexagon::CMPEQri || - MI->getOpcode() == Hexagon::CMPGTri || - MI->getOpcode() == Hexagon::CMPGEri) && + MI->getOpcode() == Hexagon::CMPGTri) && (v == -1)))) return false; } @@ -271,58 +286,63 @@ static bool canCompareBeNewValueJump(const HexagonInstrInfo *QII, // Given a compare operator, return a matching New Value Jump // compare operator. Make sure that MI here is included in // HexagonInstrInfo.cpp::isNewValueJumpCandidate -static unsigned getNewValueJumpOpcode(const MachineInstr *MI, int reg, - bool secondRegNewified) { +static unsigned getNewValueJumpOpcode(MachineInstr *MI, int reg, + bool secondRegNewified, + MachineBasicBlock *jmpTarget, + const MachineBranchProbabilityInfo + *MBPI) { + bool taken = false; + MachineBasicBlock *Src = MI->getParent(); + const BranchProbability Prediction = + MBPI->getEdgeProbability(Src, jmpTarget); + + if (Prediction >= BranchProbability(1,2)) + taken = true; + switch (MI->getOpcode()) { case Hexagon::CMPEQrr: - return Hexagon::JMP_EQrrPt_nv_V4; + return taken ? Hexagon::CMPEQrr_t_Jumpnv_t_V4 + : Hexagon::CMPEQrr_t_Jumpnv_nt_V4; case Hexagon::CMPEQri: { if (reg >= 0) - return Hexagon::JMP_EQriPt_nv_V4; + return taken ? Hexagon::CMPEQri_t_Jumpnv_t_V4 + : Hexagon::CMPEQri_t_Jumpnv_nt_V4; else - return Hexagon::JMP_EQriPtneg_nv_V4; + return taken ? Hexagon::CMPEQn1_t_Jumpnv_t_V4 + : Hexagon::CMPEQn1_t_Jumpnv_nt_V4; } - case Hexagon::CMPLTrr: case Hexagon::CMPGTrr: { if (secondRegNewified) - return Hexagon::JMP_GTrrdnPt_nv_V4; + return taken ? Hexagon::CMPLTrr_t_Jumpnv_t_V4 + : Hexagon::CMPLTrr_t_Jumpnv_nt_V4; else - return Hexagon::JMP_GTrrPt_nv_V4; - } - - case Hexagon::CMPGEri: { - if (reg >= 1) - return Hexagon::JMP_GTriPt_nv_V4; - else - return Hexagon::JMP_GTriPtneg_nv_V4; + return taken ? Hexagon::CMPGTrr_t_Jumpnv_t_V4 + : Hexagon::CMPGTrr_t_Jumpnv_nt_V4; } case Hexagon::CMPGTri: { if (reg >= 0) - return Hexagon::JMP_GTriPt_nv_V4; + return taken ? Hexagon::CMPGTri_t_Jumpnv_t_V4 + : Hexagon::CMPGTri_t_Jumpnv_nt_V4; else - return Hexagon::JMP_GTriPtneg_nv_V4; + return taken ? Hexagon::CMPGTn1_t_Jumpnv_t_V4 + : Hexagon::CMPGTn1_t_Jumpnv_nt_V4; } - case Hexagon::CMPLTUrr: case Hexagon::CMPGTUrr: { if (secondRegNewified) - return Hexagon::JMP_GTUrrdnPt_nv_V4; + return taken ? Hexagon::CMPLTUrr_t_Jumpnv_t_V4 + : Hexagon::CMPLTUrr_t_Jumpnv_nt_V4; else - return Hexagon::JMP_GTUrrPt_nv_V4; + return taken ? Hexagon::CMPGTUrr_t_Jumpnv_t_V4 + : Hexagon::CMPGTUrr_t_Jumpnv_nt_V4; } case Hexagon::CMPGTUri: - return Hexagon::JMP_GTUriPt_nv_V4; - - case Hexagon::CMPGEUri: { - if (reg == 0) - return Hexagon::JMP_EQrrPt_nv_V4; - else - return Hexagon::JMP_GTUriPt_nv_V4; - } + return taken ? Hexagon::CMPGTUri_t_Jumpnv_t_V4 + : Hexagon::CMPGTUri_t_Jumpnv_nt_V4; default: llvm_unreachable("Could not find matching New Value Jump instruction."); @@ -346,6 +366,7 @@ bool HexagonNewValueJump::runOnMachineFunction(MachineFunction &MF) { QII = static_cast<const HexagonInstrInfo *>(MF.getTarget().getInstrInfo()); QRI = static_cast<const HexagonRegisterInfo *>(MF.getTarget().getRegisterInfo()); + MBPI = &getAnalysis<MachineBranchProbabilityInfo>(); if (!QRI->Subtarget.hasV4TOps() || DisableNewValueJumps) { @@ -393,12 +414,12 @@ bool HexagonNewValueJump::runOnMachineFunction(MachineFunction &MF) { DEBUG(dbgs() << "Instr: "; MI->dump(); dbgs() << "\n"); if (!foundJump && - (MI->getOpcode() == Hexagon::JMP_c || - MI->getOpcode() == Hexagon::JMP_cNot || - MI->getOpcode() == Hexagon::JMP_cdnPt || - MI->getOpcode() == Hexagon::JMP_cdnPnt || - MI->getOpcode() == Hexagon::JMP_cdnNotPt || - MI->getOpcode() == Hexagon::JMP_cdnNotPnt)) { + (MI->getOpcode() == Hexagon::JMP_t || + MI->getOpcode() == Hexagon::JMP_f || + MI->getOpcode() == Hexagon::JMP_tnew_t || + MI->getOpcode() == Hexagon::JMP_tnew_nt || + MI->getOpcode() == Hexagon::JMP_fnew_t || + MI->getOpcode() == Hexagon::JMP_fnew_nt)) { // This is where you would insert your compare and // instr that feeds compare jmpPos = MII; @@ -434,9 +455,9 @@ bool HexagonNewValueJump::runOnMachineFunction(MachineFunction &MF) { jmpTarget = MI->getOperand(1).getMBB(); foundJump = true; - if (MI->getOpcode() == Hexagon::JMP_cNot || - MI->getOpcode() == Hexagon::JMP_cdnNotPt || - MI->getOpcode() == Hexagon::JMP_cdnNotPnt) { + if (MI->getOpcode() == Hexagon::JMP_f || + MI->getOpcode() == Hexagon::JMP_fnew_t || + MI->getOpcode() == Hexagon::JMP_fnew_nt) { invertPredicate = true; } continue; @@ -525,10 +546,8 @@ bool HexagonNewValueJump::runOnMachineFunction(MachineFunction &MF) { if (isSecondOpReg) { // In case of CMPLT, or CMPLTU, or EQ with the second register // to newify, swap the operands. - if (cmpInstr->getOpcode() == Hexagon::CMPLTrr || - cmpInstr->getOpcode() == Hexagon::CMPLTUrr || - (cmpInstr->getOpcode() == Hexagon::CMPEQrr && - feederReg == (unsigned) cmpOp2)) { + if (cmpInstr->getOpcode() == Hexagon::CMPEQrr && + feederReg == (unsigned) cmpOp2) { unsigned tmp = cmpReg1; bool tmpIsKill = MO1IsKill; cmpReg1 = cmpOp2; @@ -582,44 +601,37 @@ bool HexagonNewValueJump::runOnMachineFunction(MachineFunction &MF) { assert((QII->isNewValueJumpCandidate(cmpInstr)) && "This compare is not a New Value Jump candidate."); unsigned opc = getNewValueJumpOpcode(cmpInstr, cmpOp2, - isSecondOpNewified); + isSecondOpNewified, + jmpTarget, MBPI); if (invertPredicate) opc = QII->getInvertedPredicatedOpcode(opc); - // Manage the conversions from CMPGEUri to either CMPEQrr - // or CMPGTUri properly. See Arch spec for CMPGEUri instructions. - // This has to be after the getNewValueJumpOpcode function call as - // second operand of the compare could be modified in this logic. - if (cmpInstr->getOpcode() == Hexagon::CMPGEUri) { - if (cmpOp2 == 0) { - cmpOp2 = cmpReg1; - MO2IsKill = MO1IsKill; - isSecondOpReg = true; - } else - --cmpOp2; - } - - // Manage the conversions from CMPGEri to CMPGTUri properly. - // See Arch spec for CMPGEri instructions. - if (cmpInstr->getOpcode() == Hexagon::CMPGEri) - --cmpOp2; - - if (isSecondOpReg) { + if (isSecondOpReg) NewMI = BuildMI(*MBB, jmpPos, dl, QII->get(opc)) .addReg(cmpReg1, getKillRegState(MO1IsKill)) .addReg(cmpOp2, getKillRegState(MO2IsKill)) .addMBB(jmpTarget); - } - else { + + else if ((cmpInstr->getOpcode() == Hexagon::CMPEQri || + cmpInstr->getOpcode() == Hexagon::CMPGTri) && + cmpOp2 == -1 ) + // Corresponding new-value compare jump instructions don't have the + // operand for -1 immediate value. + NewMI = BuildMI(*MBB, jmpPos, dl, + QII->get(opc)) + .addReg(cmpReg1, getKillRegState(MO1IsKill)) + .addMBB(jmpTarget); + + else NewMI = BuildMI(*MBB, jmpPos, dl, QII->get(opc)) .addReg(cmpReg1, getKillRegState(MO1IsKill)) .addImm(cmpOp2) .addMBB(jmpTarget); - } assert(NewMI && "New Value Jump Instruction Not created!"); + (void)NewMI; if (cmpInstr->getOperand(0).isReg() && cmpInstr->getOperand(0).isKill()) cmpInstr->getOperand(0).setIsKill(false); diff --git a/lib/Target/Hexagon/HexagonPeephole.cpp b/lib/Target/Hexagon/HexagonPeephole.cpp index 576f1d7..89e3406 100644 --- a/lib/Target/Hexagon/HexagonPeephole.cpp +++ b/lib/Target/Hexagon/HexagonPeephole.cpp @@ -61,10 +61,6 @@ static cl::opt<bool> DisableHexagonPeephole("disable-hexagon-peephole", cl::Hidden, cl::ZeroOrMore, cl::init(false), cl::desc("Disable Peephole Optimization")); -static cl::opt<int> -DbgPNPCount("pnp-count", cl::init(-1), cl::Hidden, - cl::desc("Maximum number of P=NOT(P) to be optimized")); - static cl::opt<bool> DisablePNotP("disable-hexagon-pnotp", cl::Hidden, cl::ZeroOrMore, cl::init(false), cl::desc("Disable Optimization of PNotP")); @@ -73,6 +69,14 @@ static cl::opt<bool> DisableOptSZExt("disable-hexagon-optszext", cl::Hidden, cl::ZeroOrMore, cl::init(false), cl::desc("Disable Optimization of Sign/Zero Extends")); +static cl::opt<bool> DisableOptExtTo64("disable-hexagon-opt-ext-to-64", + cl::Hidden, cl::ZeroOrMore, cl::init(false), + cl::desc("Disable Optimization of extensions to i64.")); + +namespace llvm { + void initializeHexagonPeepholePass(PassRegistry&); +} + namespace { struct HexagonPeephole : public MachineFunctionPass { const HexagonInstrInfo *QII; @@ -81,7 +85,9 @@ namespace { public: static char ID; - HexagonPeephole() : MachineFunctionPass(ID) { } + HexagonPeephole() : MachineFunctionPass(ID) { + initializeHexagonPeepholePass(*PassRegistry::getPassRegistry()); + } bool runOnMachineFunction(MachineFunction &MF); @@ -100,8 +106,10 @@ namespace { char HexagonPeephole::ID = 0; -bool HexagonPeephole::runOnMachineFunction(MachineFunction &MF) { +INITIALIZE_PASS(HexagonPeephole, "hexagon-peephole", "Hexagon Peephole", + false, false) +bool HexagonPeephole::runOnMachineFunction(MachineFunction &MF) { QII = static_cast<const HexagonInstrInfo *>(MF.getTarget(). getInstrInfo()); QRI = static_cast<const HexagonRegisterInfo *>(MF.getTarget(). @@ -142,6 +150,21 @@ bool HexagonPeephole::runOnMachineFunction(MachineFunction &MF) { } } + // Look for %vreg170<def> = COMBINE_ir_V4 (0, %vreg169) + // %vreg170:DoublRegs, %vreg169:IntRegs + if (!DisableOptExtTo64 && + MI->getOpcode () == Hexagon::COMBINE_Ir_V4) { + assert (MI->getNumOperands() == 3); + MachineOperand &Dst = MI->getOperand(0); + MachineOperand &Src1 = MI->getOperand(1); + MachineOperand &Src2 = MI->getOperand(2); + if (Src1.getImm() != 0) + continue; + unsigned DstReg = Dst.getReg(); + unsigned SrcReg = Src2.getReg(); + PeepholeMap[DstReg] = SrcReg; + } + // Look for this sequence below // %vregDoubleReg1 = LSRd_ri %vregDoubleReg0, 32 // %vregIntReg = COPY %vregDoubleReg1:subreg_loreg. diff --git a/lib/Target/Hexagon/HexagonRegisterInfo.cpp b/lib/Target/Hexagon/HexagonRegisterInfo.cpp index f947dfc..d5ca4d7 100644 --- a/lib/Target/Hexagon/HexagonRegisterInfo.cpp +++ b/lib/Target/Hexagon/HexagonRegisterInfo.cpp @@ -14,34 +14,33 @@ #include "HexagonRegisterInfo.h" #include "Hexagon.h" -#include "HexagonMachineFunctionInfo.h" #include "HexagonSubtarget.h" #include "HexagonTargetMachine.h" +#include "HexagonMachineFunctionInfo.h" #include "llvm/ADT/BitVector.h" #include "llvm/ADT/STLExtras.h" -#include "llvm/CodeGen/MachineFrameInfo.h" +#include "llvm/CodeGen/MachineInstrBuilder.h" #include "llvm/CodeGen/MachineFunction.h" #include "llvm/CodeGen/MachineFunctionPass.h" -#include "llvm/CodeGen/MachineInstrBuilder.h" +#include "llvm/CodeGen/MachineFrameInfo.h" #include "llvm/CodeGen/MachineRegisterInfo.h" +#include "llvm/CodeGen/PseudoSourceValue.h" #include "llvm/CodeGen/RegisterScavenging.h" #include "llvm/IR/Function.h" #include "llvm/IR/Type.h" #include "llvm/MC/MachineLocation.h" -#include "llvm/Support/CommandLine.h" -#include "llvm/Support/ErrorHandling.h" #include "llvm/Target/TargetInstrInfo.h" #include "llvm/Target/TargetMachine.h" #include "llvm/Target/TargetOptions.h" +#include "llvm/Support/CommandLine.h" +#include "llvm/Support/ErrorHandling.h" using namespace llvm; -HexagonRegisterInfo::HexagonRegisterInfo(HexagonSubtarget &st, - const HexagonInstrInfo &tii) +HexagonRegisterInfo::HexagonRegisterInfo(HexagonSubtarget &st) : HexagonGenRegisterInfo(Hexagon::R31), - Subtarget(st), - TII(tii) { + Subtarget(st) { } const uint16_t* HexagonRegisterInfo::getCalleeSavedRegs(const MachineFunction @@ -129,6 +128,8 @@ void HexagonRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II, // Addressable stack objects are accessed using neg. offsets from %fp. MachineFunction &MF = *MI.getParent()->getParent(); + const HexagonInstrInfo &TII = + *static_cast<const HexagonInstrInfo*>(MF.getTarget().getInstrInfo()); int Offset = MF.getFrameInfo()->getObjectOffset(FrameIndex); MachineFrameInfo &MFI = *MF.getFrameInfo(); @@ -215,28 +216,41 @@ void HexagonRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II, MI.getOperand(FIOperandNum).ChangeToRegister(resReg, false, false,true); MI.getOperand(FIOperandNum+1).ChangeToImmediate(0); } else if (TII.isMemOp(&MI)) { - unsigned resReg = HEXAGON_RESERVED_REG_1; - if (!MFI.hasVarSizedObjects() && - TII.isValidOffset(MI.getOpcode(), (FrameSize+Offset))) { - MI.getOperand(FIOperandNum).ChangeToRegister(getStackRegister(), - false, false, true); - MI.getOperand(FIOperandNum+1).ChangeToImmediate(FrameSize+Offset); - } else if (!TII.isValidOffset(Hexagon::ADD_ri, Offset)) { - BuildMI(*MI.getParent(), II, MI.getDebugLoc(), - TII.get(Hexagon::CONST32_Int_Real), resReg).addImm(Offset); - BuildMI(*MI.getParent(), II, MI.getDebugLoc(), - TII.get(Hexagon::ADD_rr), - resReg).addReg(FrameReg).addReg(resReg); - MI.getOperand(FIOperandNum).ChangeToRegister(resReg, false, false, - true); - MI.getOperand(FIOperandNum+1).ChangeToImmediate(0); + // use the constant extender if the instruction provides it + // and we are V4TOps. + if (Subtarget.hasV4TOps()) { + if (TII.isConstExtended(&MI)) { + MI.getOperand(FIOperandNum).ChangeToRegister(FrameReg, false); + MI.getOperand(FIOperandNum+1).ChangeToImmediate(Offset); + TII.immediateExtend(&MI); + } else { + llvm_unreachable("Need to implement for memops"); + } } else { - BuildMI(*MI.getParent(), II, MI.getDebugLoc(), - TII.get(Hexagon::ADD_ri), - resReg).addReg(FrameReg).addImm(Offset); - MI.getOperand(FIOperandNum).ChangeToRegister(resReg, false, false, - true); - MI.getOperand(FIOperandNum+1).ChangeToImmediate(0); + // Only V3 and older instructions here. + unsigned ResReg = HEXAGON_RESERVED_REG_1; + if (!MFI.hasVarSizedObjects() && + TII.isValidOffset(MI.getOpcode(), (FrameSize+Offset))) { + MI.getOperand(FIOperandNum).ChangeToRegister(getStackRegister(), + false, false, false); + MI.getOperand(FIOperandNum+1).ChangeToImmediate(FrameSize+Offset); + } else if (!TII.isValidOffset(Hexagon::ADD_ri, Offset)) { + BuildMI(*MI.getParent(), II, MI.getDebugLoc(), + TII.get(Hexagon::CONST32_Int_Real), ResReg).addImm(Offset); + BuildMI(*MI.getParent(), II, MI.getDebugLoc(), + TII.get(Hexagon::ADD_rr), ResReg).addReg(FrameReg). + addReg(ResReg); + MI.getOperand(FIOperandNum).ChangeToRegister(ResReg, false, false, + true); + MI.getOperand(FIOperandNum+1).ChangeToImmediate(0); + } else { + BuildMI(*MI.getParent(), II, MI.getDebugLoc(), + TII.get(Hexagon::ADD_ri), ResReg).addReg(FrameReg). + addImm(Offset); + MI.getOperand(FIOperandNum).ChangeToRegister(ResReg, false, false, + true); + MI.getOperand(FIOperandNum+1).ChangeToImmediate(0); + } } } else { unsigned dstReg = MI.getOperand(0).getReg(); @@ -281,16 +295,6 @@ unsigned HexagonRegisterInfo::getStackRegister() const { return Hexagon::R29; } -void HexagonRegisterInfo::getInitialFrameState(std::vector<MachineMove> - &Moves) const -{ - // VirtualFP = (R30 + #0). - unsigned FPReg = getFrameRegister(); - MachineLocation Dst(MachineLocation::VirtualFP); - MachineLocation Src(FPReg, 0); - Moves.push_back(MachineMove(0, Dst, Src)); -} - unsigned HexagonRegisterInfo::getEHExceptionRegister() const { llvm_unreachable("What is the exception register"); } diff --git a/lib/Target/Hexagon/HexagonRegisterInfo.h b/lib/Target/Hexagon/HexagonRegisterInfo.h index 8a3f94a..fec86df 100644 --- a/lib/Target/Hexagon/HexagonRegisterInfo.h +++ b/lib/Target/Hexagon/HexagonRegisterInfo.h @@ -44,9 +44,8 @@ class Type; struct HexagonRegisterInfo : public HexagonGenRegisterInfo { HexagonSubtarget &Subtarget; - const HexagonInstrInfo &TII; - HexagonRegisterInfo(HexagonSubtarget &st, const HexagonInstrInfo &tii); + HexagonRegisterInfo(HexagonSubtarget &st); /// Code Generation virtual methods... const uint16_t *getCalleeSavedRegs(const MachineFunction *MF = 0) const; @@ -78,7 +77,6 @@ struct HexagonRegisterInfo : public HexagonGenRegisterInfo { unsigned getRARegister() const; unsigned getFrameRegister(const MachineFunction &MF) const; unsigned getFrameRegister() const; - void getInitialFrameState(std::vector<MachineMove> &Moves) const; unsigned getStackRegister() const; // Exception handling queries. diff --git a/lib/Target/Hexagon/HexagonRegisterInfo.td b/lib/Target/Hexagon/HexagonRegisterInfo.td index fe41fc3..8ea1b7e 100644 --- a/lib/Target/Hexagon/HexagonRegisterInfo.td +++ b/lib/Target/Hexagon/HexagonRegisterInfo.td @@ -57,8 +57,8 @@ let Namespace = "Hexagon" in { let Aliases = [R]; } - def subreg_loreg : SubRegIndex; - def subreg_hireg : SubRegIndex; + def subreg_loreg : SubRegIndex<32>; + def subreg_hireg : SubRegIndex<32, 32>; // Integer registers. def R0 : Ri< 0, "r0">, DwarfRegNum<[0]>; diff --git a/lib/Target/Hexagon/HexagonRemoveSZExtArgs.cpp b/lib/Target/Hexagon/HexagonRemoveSZExtArgs.cpp index 34bf4ea..44234e8 100644 --- a/lib/Target/Hexagon/HexagonRemoveSZExtArgs.cpp +++ b/lib/Target/Hexagon/HexagonRemoveSZExtArgs.cpp @@ -21,11 +21,18 @@ #include "llvm/Transforms/Scalar.h" using namespace llvm; + +namespace llvm { + void initializeHexagonRemoveExtendArgsPass(PassRegistry&); +} + namespace { struct HexagonRemoveExtendArgs : public FunctionPass { public: static char ID; - HexagonRemoveExtendArgs() : FunctionPass(ID) {} + HexagonRemoveExtendArgs() : FunctionPass(ID) { + initializeHexagonRemoveExtendArgsPass(*PassRegistry::getPassRegistry()); + } virtual bool runOnFunction(Function &F); const char *getPassName() const { @@ -41,11 +48,9 @@ namespace { } char HexagonRemoveExtendArgs::ID = 0; -RegisterPass<HexagonRemoveExtendArgs> X("reargs", - "Remove Sign and Zero Extends for Args" - ); - +INITIALIZE_PASS(HexagonRemoveExtendArgs, "reargs", + "Remove Sign and Zero Extends for Args", false, false) bool HexagonRemoveExtendArgs::runOnFunction(Function &F) { unsigned Idx = 1; @@ -78,6 +83,7 @@ bool HexagonRemoveExtendArgs::runOnFunction(Function &F) { -FunctionPass *llvm::createHexagonRemoveExtendOps(HexagonTargetMachine &TM) { +FunctionPass* +llvm::createHexagonRemoveExtendArgs(const HexagonTargetMachine &TM) { return new HexagonRemoveExtendArgs(); } diff --git a/lib/Target/Hexagon/HexagonSelectionDAGInfo.cpp b/lib/Target/Hexagon/HexagonSelectionDAGInfo.cpp index a52c604..c37bf9f 100644 --- a/lib/Target/Hexagon/HexagonSelectionDAGInfo.cpp +++ b/lib/Target/Hexagon/HexagonSelectionDAGInfo.cpp @@ -27,7 +27,7 @@ HexagonSelectionDAGInfo::~HexagonSelectionDAGInfo() { SDValue HexagonSelectionDAGInfo:: -EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl, SDValue Chain, +EmitTargetCodeForMemcpy(SelectionDAG &DAG, SDLoc dl, SDValue Chain, SDValue Dst, SDValue Src, SDValue Size, unsigned Align, bool isVolatile, bool AlwaysInline, MachinePointerInfo DstPtrInfo, diff --git a/lib/Target/Hexagon/HexagonSelectionDAGInfo.h b/lib/Target/Hexagon/HexagonSelectionDAGInfo.h index 0673e4d..31f278a 100644 --- a/lib/Target/Hexagon/HexagonSelectionDAGInfo.h +++ b/lib/Target/Hexagon/HexagonSelectionDAGInfo.h @@ -26,7 +26,7 @@ public: ~HexagonSelectionDAGInfo(); virtual - SDValue EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl, + SDValue EmitTargetCodeForMemcpy(SelectionDAG &DAG, SDLoc dl, SDValue Chain, SDValue Dst, SDValue Src, SDValue Size, unsigned Align, diff --git a/lib/Target/Hexagon/HexagonSplitConst32AndConst64.cpp b/lib/Target/Hexagon/HexagonSplitConst32AndConst64.cpp new file mode 100644 index 0000000..3bf2f20 --- /dev/null +++ b/lib/Target/Hexagon/HexagonSplitConst32AndConst64.cpp @@ -0,0 +1,176 @@ +//=== HexagonSplitConst32AndConst64.cpp - split CONST32/Const64 into HI/LO ===// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +// +// When the compiler is invoked with no small data, for instance, with the -G0 +// command line option, then all CONST32_* opcodes should be broken down into +// appropriate LO and HI instructions. This splitting is done by this pass. +// The only reason this is not done in the DAG lowering itself is that there +// is no simple way of getting the register allocator to allot the same hard +// register to the result of LO and HI instructions. This pass is always +// scheduled after register allocation. +// +//===----------------------------------------------------------------------===// +#define DEBUG_TYPE "xfer" +#include "llvm/CodeGen/Passes.h" +#include "llvm/CodeGen/ScheduleDAGInstrs.h" +#include "llvm/CodeGen/LatencyPriorityQueue.h" +#include "llvm/CodeGen/SchedulerRegistry.h" +#include "llvm/CodeGen/MachineDominators.h" +#include "llvm/CodeGen/MachineFunctionPass.h" +#include "llvm/CodeGen/MachineLoopInfo.h" +#include "llvm/CodeGen/MachineRegisterInfo.h" +#include "llvm/CodeGen/ScheduleHazardRecognizer.h" +#include "llvm/Target/TargetMachine.h" +#include "llvm/Target/TargetInstrInfo.h" +#include "llvm/Target/TargetRegisterInfo.h" +#include "llvm/Support/Compiler.h" +#include "llvm/Support/Debug.h" +#include "llvm/ADT/Statistic.h" +#include "llvm/Support/MathExtras.h" +#include "llvm/CodeGen/MachineInstrBuilder.h" +#include "HexagonTargetMachine.h" +#include "HexagonSubtarget.h" +#include "HexagonMachineFunctionInfo.h" +#include <map> +#include <iostream> + +#include "llvm/Support/CommandLine.h" +#define DEBUG_TYPE "xfer" + + +using namespace llvm; + +namespace { + +class HexagonSplitConst32AndConst64 : public MachineFunctionPass { + const HexagonTargetMachine& QTM; + const HexagonSubtarget &QST; + + public: + static char ID; + HexagonSplitConst32AndConst64(const HexagonTargetMachine& TM) + : MachineFunctionPass(ID), QTM(TM), QST(*TM.getSubtargetImpl()) {} + + const char *getPassName() const { + return "Hexagon Split Const32s and Const64s"; + } + bool runOnMachineFunction(MachineFunction &Fn); +}; + + +char HexagonSplitConst32AndConst64::ID = 0; + + +bool HexagonSplitConst32AndConst64::runOnMachineFunction(MachineFunction &Fn) { + + const TargetInstrInfo *TII = QTM.getInstrInfo(); + + // Loop over all of the basic blocks + for (MachineFunction::iterator MBBb = Fn.begin(), MBBe = Fn.end(); + MBBb != MBBe; ++MBBb) { + MachineBasicBlock* MBB = MBBb; + // Traverse the basic block + MachineBasicBlock::iterator MII = MBB->begin(); + MachineBasicBlock::iterator MIE = MBB->end (); + while (MII != MIE) { + MachineInstr *MI = MII; + int Opc = MI->getOpcode(); + if (Opc == Hexagon::CONST32_set) { + int DestReg = MI->getOperand(0).getReg(); + MachineOperand &Symbol = MI->getOperand (1); + + BuildMI (*MBB, MII, MI->getDebugLoc(), + TII->get(Hexagon::LO), DestReg).addOperand(Symbol); + BuildMI (*MBB, MII, MI->getDebugLoc(), + TII->get(Hexagon::HI), DestReg).addOperand(Symbol); + // MBB->erase returns the iterator to the next instruction, which is the + // one we want to process next + MII = MBB->erase (MI); + continue; + } + else if (Opc == Hexagon::CONST32_set_jt) { + int DestReg = MI->getOperand(0).getReg(); + MachineOperand &Symbol = MI->getOperand (1); + + BuildMI (*MBB, MII, MI->getDebugLoc(), + TII->get(Hexagon::LO_jt), DestReg).addOperand(Symbol); + BuildMI (*MBB, MII, MI->getDebugLoc(), + TII->get(Hexagon::HI_jt), DestReg).addOperand(Symbol); + // MBB->erase returns the iterator to the next instruction, which is the + // one we want to process next + MII = MBB->erase (MI); + continue; + } + else if (Opc == Hexagon::CONST32_Label) { + int DestReg = MI->getOperand(0).getReg(); + MachineOperand &Symbol = MI->getOperand (1); + + BuildMI (*MBB, MII, MI->getDebugLoc(), + TII->get(Hexagon::LO_label), DestReg).addOperand(Symbol); + BuildMI (*MBB, MII, MI->getDebugLoc(), + TII->get(Hexagon::HI_label), DestReg).addOperand(Symbol); + // MBB->erase returns the iterator to the next instruction, which is the + // one we want to process next + MII = MBB->erase (MI); + continue; + } + else if (Opc == Hexagon::CONST32_Int_Real) { + int DestReg = MI->getOperand(0).getReg(); + int64_t ImmValue = MI->getOperand(1).getImm (); + + BuildMI (*MBB, MII, MI->getDebugLoc(), + TII->get(Hexagon::LOi), DestReg).addImm(ImmValue); + BuildMI (*MBB, MII, MI->getDebugLoc(), + TII->get(Hexagon::HIi), DestReg).addImm(ImmValue); + MII = MBB->erase (MI); + continue; + } + else if (Opc == Hexagon::CONST64_Int_Real) { + int DestReg = MI->getOperand(0).getReg(); + int64_t ImmValue = MI->getOperand(1).getImm (); + unsigned DestLo = + QTM.getRegisterInfo()->getSubReg (DestReg, Hexagon::subreg_loreg); + unsigned DestHi = + QTM.getRegisterInfo()->getSubReg (DestReg, Hexagon::subreg_hireg); + + int32_t LowWord = (ImmValue & 0xFFFFFFFF); + int32_t HighWord = (ImmValue >> 32) & 0xFFFFFFFF; + + // Lower Registers Lower Half + BuildMI (*MBB, MII, MI->getDebugLoc(), + TII->get(Hexagon::LOi), DestLo).addImm(LowWord); + // Lower Registers Higher Half + BuildMI (*MBB, MII, MI->getDebugLoc(), + TII->get(Hexagon::HIi), DestLo).addImm(LowWord); + // Higher Registers Lower Half + BuildMI (*MBB, MII, MI->getDebugLoc(), + TII->get(Hexagon::LOi), DestHi).addImm(HighWord); + // Higher Registers Higher Half. + BuildMI (*MBB, MII, MI->getDebugLoc(), + TII->get(Hexagon::HIi), DestHi).addImm(HighWord); + MII = MBB->erase (MI); + continue; + } + ++MII; + } + } + + return true; +} + +} + +//===----------------------------------------------------------------------===// +// Public Constructor Functions +//===----------------------------------------------------------------------===// + +FunctionPass * +llvm::createHexagonSplitConst32AndConst64(const HexagonTargetMachine &TM) { + return new HexagonSplitConst32AndConst64(TM); +} diff --git a/lib/Target/Hexagon/HexagonSplitTFRCondSets.cpp b/lib/Target/Hexagon/HexagonSplitTFRCondSets.cpp index 814249f..8608e08 100644 --- a/lib/Target/Hexagon/HexagonSplitTFRCondSets.cpp +++ b/lib/Target/Hexagon/HexagonSplitTFRCondSets.cpp @@ -49,16 +49,23 @@ using namespace llvm; +namespace llvm { + void initializeHexagonSplitTFRCondSetsPass(PassRegistry&); +} + + namespace { class HexagonSplitTFRCondSets : public MachineFunctionPass { - HexagonTargetMachine& QTM; + const HexagonTargetMachine &QTM; const HexagonSubtarget &QST; public: static char ID; - HexagonSplitTFRCondSets(HexagonTargetMachine& TM) : - MachineFunctionPass(ID), QTM(TM), QST(*TM.getSubtargetImpl()) {} + HexagonSplitTFRCondSets(const HexagonTargetMachine& TM) : + MachineFunctionPass(ID), QTM(TM), QST(*TM.getSubtargetImpl()) { + initializeHexagonSplitTFRCondSetsPass(*PassRegistry::getPassRegistry()); + } const char *getPassName() const { return "Hexagon Split TFRCondSets"; @@ -211,6 +218,18 @@ bool HexagonSplitTFRCondSets::runOnMachineFunction(MachineFunction &Fn) { // Public Constructor Functions //===----------------------------------------------------------------------===// -FunctionPass *llvm::createHexagonSplitTFRCondSets(HexagonTargetMachine &TM) { +static void initializePassOnce(PassRegistry &Registry) { + const char *Name = "Hexagon Split TFRCondSets"; + PassInfo *PI = new PassInfo(Name, "hexagon-split-tfr", + &HexagonSplitTFRCondSets::ID, 0, false, false); + Registry.registerPass(*PI, true); +} + +void llvm::initializeHexagonSplitTFRCondSetsPass(PassRegistry &Registry) { + CALL_ONCE_INITIALIZATION(initializePassOnce) +} + +FunctionPass* +llvm::createHexagonSplitTFRCondSets(const HexagonTargetMachine &TM) { return new HexagonSplitTFRCondSets(TM); } diff --git a/lib/Target/Hexagon/HexagonSubtarget.cpp b/lib/Target/Hexagon/HexagonSubtarget.cpp index 4bacb8f..07d5ce1 100644 --- a/lib/Target/Hexagon/HexagonSubtarget.cpp +++ b/lib/Target/Hexagon/HexagonSubtarget.cpp @@ -29,8 +29,16 @@ EnableV3("enable-hexagon-v3", cl::Hidden, static cl::opt<bool> EnableMemOps( "enable-hexagon-memops", - cl::Hidden, cl::ZeroOrMore, cl::ValueDisallowed, - cl::desc("Generate V4 memop instructions.")); + cl::Hidden, cl::ZeroOrMore, cl::ValueDisallowed, cl::init(true), + cl::desc( + "Generate V4 MEMOP in code generation for Hexagon target")); + +static cl::opt<bool> +DisableMemOps( + "disable-hexagon-memops", + cl::Hidden, cl::ZeroOrMore, cl::ValueDisallowed, cl::init(false), + cl::desc( + "Do not generate V4 MEMOP in code generation for Hexagon target")); static cl::opt<bool> EnableIEEERndNear( @@ -64,7 +72,10 @@ HexagonSubtarget::HexagonSubtarget(StringRef TT, StringRef CPU, StringRef FS): // Initialize scheduling itinerary for the specified CPU. InstrItins = getInstrItineraryForCPU(CPUString); - if (EnableMemOps) + // UseMemOps on by default unless disabled explicitly + if (DisableMemOps) + UseMemOps = false; + else if (EnableMemOps) UseMemOps = true; else UseMemOps = false; diff --git a/lib/Target/Hexagon/HexagonTargetMachine.cpp b/lib/Target/Hexagon/HexagonTargetMachine.cpp index d9fef3e..cd96b58 100644 --- a/lib/Target/Hexagon/HexagonTargetMachine.cpp +++ b/lib/Target/Hexagon/HexagonTargetMachine.cpp @@ -15,6 +15,7 @@ #include "Hexagon.h" #include "HexagonISelLowering.h" #include "HexagonMachineScheduler.h" +#include "HexagonTargetObjectFile.h" #include "llvm/CodeGen/Passes.h" #include "llvm/IR/Module.h" #include "llvm/PassManager.h" @@ -25,15 +26,17 @@ using namespace llvm; -static cl:: -opt<bool> DisableHardwareLoops( - "disable-hexagon-hwloops", cl::Hidden, - cl::desc("Disable Hardware Loops for Hexagon target")); +static cl:: opt<bool> DisableHardwareLoops("disable-hexagon-hwloops", + cl::Hidden, cl::desc("Disable Hardware Loops for Hexagon target")); + +static cl::opt<bool> DisableHexagonMISched("disable-hexagon-misched", + cl::Hidden, cl::ZeroOrMore, cl::init(false), + cl::desc("Disable Hexagon MI Scheduling")); + +static cl::opt<bool> DisableHexagonCFGOpt("disable-hexagon-cfgopt", + cl::Hidden, cl::ZeroOrMore, cl::init(false), + cl::desc("Disable Hexagon CFG Optimization")); -static cl:: -opt<bool> DisableHexagonMISched("disable-hexagon-misched", - cl::Hidden, cl::ZeroOrMore, cl::init(false), - cl::desc("Disable Hexagon MI Scheduling")); /// HexagonTargetMachineModule - Note that this is used on hosts that /// cannot link in a library unless there are references into the @@ -75,19 +78,21 @@ HexagonTargetMachine::HexagonTargetMachine(const Target &T, StringRef TT, TSInfo(*this), FrameLowering(Subtarget), InstrItins(&Subtarget.getInstrItineraryData()) { - setMCUseCFI(false); + setMCUseCFI(false); + initAsmInfo(); } // addPassesForOptimizations - Allow the backend (target) to add Target // Independent Optimization passes to the Pass Manager. bool HexagonTargetMachine::addPassesForOptimizations(PassManagerBase &PM) { - - PM.add(createConstantPropagationPass()); - PM.add(createLoopSimplifyPass()); - PM.add(createDeadCodeEliminationPass()); - PM.add(createConstantPropagationPass()); - PM.add(createLoopUnrollPass()); - PM.add(createLoopStrengthReducePass()); + if (getOptLevel() != CodeGenOpt::None) { + PM.add(createConstantPropagationPass()); + PM.add(createLoopSimplifyPass()); + PM.add(createDeadCodeEliminationPass()); + PM.add(createConstantPropagationPass()); + PM.add(createLoopUnrollPass()); + PM.add(createLoopStrengthReducePass()); + } return true; } @@ -121,47 +126,71 @@ TargetPassConfig *HexagonTargetMachine::createPassConfig(PassManagerBase &PM) { } bool HexagonPassConfig::addInstSelector() { - addPass(createHexagonRemoveExtendOps(getHexagonTargetMachine())); - addPass(createHexagonISelDag(getHexagonTargetMachine(), getOptLevel())); - addPass(createHexagonPeephole()); + HexagonTargetMachine &TM = getHexagonTargetMachine(); + bool NoOpt = (getOptLevel() == CodeGenOpt::None); + + if (!NoOpt) + addPass(createHexagonRemoveExtendArgs(TM)); + + addPass(createHexagonISelDag(TM, getOptLevel())); + + if (!NoOpt) { + addPass(createHexagonPeephole()); + printAndVerify("After hexagon peephole pass"); + } + return false; } - bool HexagonPassConfig::addPreRegAlloc() { - if (!DisableHardwareLoops) { - addPass(createHexagonHardwareLoops()); - } + if (getOptLevel() != CodeGenOpt::None) + if (!DisableHardwareLoops) + addPass(createHexagonHardwareLoops()); return false; } bool HexagonPassConfig::addPostRegAlloc() { - addPass(createHexagonCFGOptimizer(getHexagonTargetMachine())); - return true; + const HexagonTargetMachine &TM = getHexagonTargetMachine(); + if (getOptLevel() != CodeGenOpt::None) + if (!DisableHexagonCFGOpt) + addPass(createHexagonCFGOptimizer(TM)); + return false; } - bool HexagonPassConfig::addPreSched2() { - addPass(&IfConverterID); + const HexagonTargetMachine &TM = getHexagonTargetMachine(); + const HexagonTargetObjectFile &TLOF = + (const HexagonTargetObjectFile &)getTargetLowering()->getObjFileLowering(); + + addPass(createHexagonCopyToCombine()); + if (getOptLevel() != CodeGenOpt::None) + addPass(&IfConverterID); + if (!TLOF.IsSmallDataEnabled()) { + addPass(createHexagonSplitConst32AndConst64(TM)); + printAndVerify("After hexagon split const32/64 pass"); + } return true; } bool HexagonPassConfig::addPreEmitPass() { + const HexagonTargetMachine &TM = getHexagonTargetMachine(); + bool NoOpt = (getOptLevel() == CodeGenOpt::None); - if (!DisableHardwareLoops) { - addPass(createHexagonFixupHwLoops()); - } - - addPass(createHexagonNewValueJump()); + if (!NoOpt) + addPass(createHexagonNewValueJump()); // Expand Spill code for predicate registers. - addPass(createHexagonExpandPredSpillCode(getHexagonTargetMachine())); + addPass(createHexagonExpandPredSpillCode(TM)); // Split up TFRcondsets into conditional transfers. - addPass(createHexagonSplitTFRCondSets(getHexagonTargetMachine())); + addPass(createHexagonSplitTFRCondSets(TM)); // Create Packets. - addPass(createHexagonPacketizer()); + if (!NoOpt) { + if (!DisableHardwareLoops) + addPass(createHexagonFixupHwLoops()); + addPass(createHexagonPacketizer()); + } return false; } diff --git a/lib/Target/Hexagon/HexagonTargetObjectFile.cpp b/lib/Target/Hexagon/HexagonTargetObjectFile.cpp index 993fcfa..7773cff 100644 --- a/lib/Target/Hexagon/HexagonTargetObjectFile.cpp +++ b/lib/Target/Hexagon/HexagonTargetObjectFile.cpp @@ -25,7 +25,8 @@ using namespace llvm; static cl::opt<int> SmallDataThreshold("hexagon-small-data-threshold", - cl::init(8), cl::Hidden); + cl::init(8), cl::Hidden, + cl::desc("The maximum size of an object in the sdata section")); void HexagonTargetObjectFile::Initialize(MCContext &Ctx, const TargetMachine &TM) { @@ -46,6 +47,11 @@ void HexagonTargetObjectFile::Initialize(MCContext &Ctx, static bool IsInSmallSection(uint64_t Size) { return Size > 0 && Size <= (uint64_t)SmallDataThreshold; } + +bool HexagonTargetObjectFile::IsSmallDataEnabled () const { + return SmallDataThreshold > 0; +} + /// IsGlobalInSmallSection - Return true if this global value should be /// placed into small data/bss section. bool HexagonTargetObjectFile::IsGlobalInSmallSection(const GlobalValue *GV, diff --git a/lib/Target/Hexagon/HexagonTargetObjectFile.h b/lib/Target/Hexagon/HexagonTargetObjectFile.h index 6933450..41f6792 100644 --- a/lib/Target/Hexagon/HexagonTargetObjectFile.h +++ b/lib/Target/Hexagon/HexagonTargetObjectFile.h @@ -29,6 +29,7 @@ namespace llvm { bool IsGlobalInSmallSection(const GlobalValue *GV, const TargetMachine &TM) const; + bool IsSmallDataEnabled () const; const MCSection* SelectSectionForGlobal(const GlobalValue *GV, SectionKind Kind, Mangler *Mang, diff --git a/lib/Target/Hexagon/HexagonVLIWPacketizer.cpp b/lib/Target/Hexagon/HexagonVLIWPacketizer.cpp index 866beb1..41e382d 100644 --- a/lib/Target/Hexagon/HexagonVLIWPacketizer.cpp +++ b/lib/Target/Hexagon/HexagonVLIWPacketizer.cpp @@ -48,19 +48,32 @@ #include "HexagonMachineFunctionInfo.h" #include <map> +#include <vector> using namespace llvm; +static cl::opt<bool> PacketizeVolatiles("hexagon-packetize-volatiles", + cl::ZeroOrMore, cl::Hidden, cl::init(true), + cl::desc("Allow non-solo packetization of volatile memory references")); + +namespace llvm { + void initializeHexagonPacketizerPass(PassRegistry&); +} + + namespace { class HexagonPacketizer : public MachineFunctionPass { public: static char ID; - HexagonPacketizer() : MachineFunctionPass(ID) {} + HexagonPacketizer() : MachineFunctionPass(ID) { + initializeHexagonPacketizerPass(*PassRegistry::getPassRegistry()); + } void getAnalysisUsage(AnalysisUsage &AU) const { AU.setPreservesCFG(); AU.addRequired<MachineDominatorTree>(); + AU.addRequired<MachineBranchProbabilityInfo>(); AU.addPreserved<MachineDominatorTree>(); AU.addRequired<MachineLoopInfo>(); AU.addPreserved<MachineLoopInfo>(); @@ -96,10 +109,17 @@ namespace { // schedule this instruction. bool FoundSequentialDependence; + /// \brief A handle to the branch probability pass. + const MachineBranchProbabilityInfo *MBPI; + + // Track MIs with ignored dependece. + std::vector<MachineInstr*> IgnoreDepMIs; + public: // Ctor. HexagonPacketizerList(MachineFunction &MF, MachineLoopInfo &MLI, - MachineDominatorTree &MDT); + MachineDominatorTree &MDT, + const MachineBranchProbabilityInfo *MBPI); // initPacketizerState - initialize some internal flags. void initPacketizerState(); @@ -123,20 +143,20 @@ namespace { private: bool IsCallDependent(MachineInstr* MI, SDep::Kind DepType, unsigned DepReg); bool PromoteToDotNew(MachineInstr* MI, SDep::Kind DepType, - MachineBasicBlock::iterator &MII, - const TargetRegisterClass* RC); + MachineBasicBlock::iterator &MII, + const TargetRegisterClass* RC); bool CanPromoteToDotNew(MachineInstr* MI, SUnit* PacketSU, - unsigned DepReg, - std::map <MachineInstr*, SUnit*> MIToSUnit, - MachineBasicBlock::iterator &MII, - const TargetRegisterClass* RC); + unsigned DepReg, + std::map <MachineInstr*, SUnit*> MIToSUnit, + MachineBasicBlock::iterator &MII, + const TargetRegisterClass* RC); bool CanPromoteToNewValue(MachineInstr* MI, SUnit* PacketSU, - unsigned DepReg, - std::map <MachineInstr*, SUnit*> MIToSUnit, - MachineBasicBlock::iterator &MII); + unsigned DepReg, + std::map <MachineInstr*, SUnit*> MIToSUnit, + MachineBasicBlock::iterator &MII); bool CanPromoteToNewValueStore(MachineInstr* MI, MachineInstr* PacketMI, - unsigned DepReg, - std::map <MachineInstr*, SUnit*> MIToSUnit); + unsigned DepReg, + std::map <MachineInstr*, SUnit*> MIToSUnit); bool DemoteToDotOld(MachineInstr* MI); bool ArePredicatesComplements(MachineInstr* MI1, MachineInstr* MI2, std::map <MachineInstr*, SUnit*> MIToSUnit); @@ -144,28 +164,39 @@ namespace { unsigned, std::map <MachineInstr*, SUnit*>); bool isNewifiable(MachineInstr* MI); bool isCondInst(MachineInstr* MI); - bool IsNewifyStore (MachineInstr* MI); bool tryAllocateResourcesForConstExt(MachineInstr* MI); bool canReserveResourcesForConstExt(MachineInstr *MI); void reserveResourcesForConstExt(MachineInstr* MI); bool isNewValueInst(MachineInstr* MI); - bool isDotNewInst(MachineInstr* MI); }; } +INITIALIZE_PASS_BEGIN(HexagonPacketizer, "packets", "Hexagon Packetizer", + false, false) +INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree) +INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo) +INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo) +INITIALIZE_AG_DEPENDENCY(AliasAnalysis) +INITIALIZE_PASS_END(HexagonPacketizer, "packets", "Hexagon Packetizer", + false, false) + + // HexagonPacketizerList Ctor. HexagonPacketizerList::HexagonPacketizerList( - MachineFunction &MF, MachineLoopInfo &MLI,MachineDominatorTree &MDT) + MachineFunction &MF, MachineLoopInfo &MLI,MachineDominatorTree &MDT, + const MachineBranchProbabilityInfo *MBPI) : VLIWPacketizerList(MF, MLI, MDT, true){ + this->MBPI = MBPI; } bool HexagonPacketizer::runOnMachineFunction(MachineFunction &Fn) { const TargetInstrInfo *TII = Fn.getTarget().getInstrInfo(); MachineLoopInfo &MLI = getAnalysis<MachineLoopInfo>(); MachineDominatorTree &MDT = getAnalysis<MachineDominatorTree>(); - + const MachineBranchProbabilityInfo *MBPI = + &getAnalysis<MachineBranchProbabilityInfo>(); // Instantiate the packetizer. - HexagonPacketizerList Packetizer(Fn, MLI, MDT); + HexagonPacketizerList Packetizer(Fn, MLI, MDT, MBPI); // DFA state table should not be empty. assert(Packetizer.getResourceTracker() && "Empty DFA table!"); @@ -351,104 +382,6 @@ static bool IsControlFlow(MachineInstr* MI) { return (MI->getDesc().isTerminator() || MI->getDesc().isCall()); } -// Function returns true if an instruction can be promoted to the new-value -// store. It will always return false for v2 and v3. -// It lists all the conditional and unconditional stores that can be promoted -// to the new-value stores. - -bool HexagonPacketizerList::IsNewifyStore (MachineInstr* MI) { - const HexagonRegisterInfo* QRI = - (const HexagonRegisterInfo *) TM.getRegisterInfo(); - switch (MI->getOpcode()) - { - // store byte - case Hexagon::STrib: - case Hexagon::STrib_indexed: - case Hexagon::STrib_indexed_shl_V4: - case Hexagon::STrib_shl_V4: - case Hexagon::STb_GP_V4: - case Hexagon::POST_STbri: - case Hexagon::STrib_cPt: - case Hexagon::STrib_cdnPt_V4: - case Hexagon::STrib_cNotPt: - case Hexagon::STrib_cdnNotPt_V4: - case Hexagon::STrib_indexed_cPt: - case Hexagon::STrib_indexed_cdnPt_V4: - case Hexagon::STrib_indexed_cNotPt: - case Hexagon::STrib_indexed_cdnNotPt_V4: - case Hexagon::STrib_indexed_shl_cPt_V4: - case Hexagon::STrib_indexed_shl_cdnPt_V4: - case Hexagon::STrib_indexed_shl_cNotPt_V4: - case Hexagon::STrib_indexed_shl_cdnNotPt_V4: - case Hexagon::POST_STbri_cPt: - case Hexagon::POST_STbri_cdnPt_V4: - case Hexagon::POST_STbri_cNotPt: - case Hexagon::POST_STbri_cdnNotPt_V4: - case Hexagon::STb_GP_cPt_V4: - case Hexagon::STb_GP_cNotPt_V4: - case Hexagon::STb_GP_cdnPt_V4: - case Hexagon::STb_GP_cdnNotPt_V4: - - // store halfword - case Hexagon::STrih: - case Hexagon::STrih_indexed: - case Hexagon::STrih_indexed_shl_V4: - case Hexagon::STrih_shl_V4: - case Hexagon::STh_GP_V4: - case Hexagon::POST_SThri: - case Hexagon::STrih_cPt: - case Hexagon::STrih_cdnPt_V4: - case Hexagon::STrih_cNotPt: - case Hexagon::STrih_cdnNotPt_V4: - case Hexagon::STrih_indexed_cPt: - case Hexagon::STrih_indexed_cdnPt_V4: - case Hexagon::STrih_indexed_cNotPt: - case Hexagon::STrih_indexed_cdnNotPt_V4: - case Hexagon::STrih_indexed_shl_cPt_V4: - case Hexagon::STrih_indexed_shl_cdnPt_V4: - case Hexagon::STrih_indexed_shl_cNotPt_V4: - case Hexagon::STrih_indexed_shl_cdnNotPt_V4: - case Hexagon::POST_SThri_cPt: - case Hexagon::POST_SThri_cdnPt_V4: - case Hexagon::POST_SThri_cNotPt: - case Hexagon::POST_SThri_cdnNotPt_V4: - case Hexagon::STh_GP_cPt_V4: - case Hexagon::STh_GP_cNotPt_V4: - case Hexagon::STh_GP_cdnPt_V4: - case Hexagon::STh_GP_cdnNotPt_V4: - - // store word - case Hexagon::STriw: - case Hexagon::STriw_indexed: - case Hexagon::STriw_indexed_shl_V4: - case Hexagon::STriw_shl_V4: - case Hexagon::STw_GP_V4: - case Hexagon::POST_STwri: - case Hexagon::STriw_cPt: - case Hexagon::STriw_cdnPt_V4: - case Hexagon::STriw_cNotPt: - case Hexagon::STriw_cdnNotPt_V4: - case Hexagon::STriw_indexed_cPt: - case Hexagon::STriw_indexed_cdnPt_V4: - case Hexagon::STriw_indexed_cNotPt: - case Hexagon::STriw_indexed_cdnNotPt_V4: - case Hexagon::STriw_indexed_shl_cPt_V4: - case Hexagon::STriw_indexed_shl_cdnPt_V4: - case Hexagon::STriw_indexed_shl_cNotPt_V4: - case Hexagon::STriw_indexed_shl_cdnNotPt_V4: - case Hexagon::POST_STwri_cPt: - case Hexagon::POST_STwri_cdnPt_V4: - case Hexagon::POST_STwri_cNotPt: - case Hexagon::POST_STwri_cdnNotPt_V4: - case Hexagon::STw_GP_cPt_V4: - case Hexagon::STw_GP_cNotPt_V4: - case Hexagon::STw_GP_cdnPt_V4: - case Hexagon::STw_GP_cdnNotPt_V4: - return QRI->Subtarget.hasV4TOps(); - } - return false; -} - static bool IsLoopN(MachineInstr *MI) { return (MI->getOpcode() == Hexagon::LOOP0_i || MI->getOpcode() == Hexagon::LOOP0_r); @@ -466,769 +399,11 @@ static bool DoesModifyCalleeSavedReg(MachineInstr *MI, return false; } -// Return the new value instruction for a given store. -static int GetDotNewOp(const int opc) { - switch (opc) { - default: llvm_unreachable("Unknown .new type"); - // store new value byte - case Hexagon::STrib: - return Hexagon::STrib_nv_V4; - - case Hexagon::STrib_indexed: - return Hexagon::STrib_indexed_nv_V4; - - case Hexagon::STrib_indexed_shl_V4: - return Hexagon::STrib_indexed_shl_nv_V4; - - case Hexagon::STrib_shl_V4: - return Hexagon::STrib_shl_nv_V4; - - case Hexagon::STb_GP_V4: - return Hexagon::STb_GP_nv_V4; - - case Hexagon::POST_STbri: - return Hexagon::POST_STbri_nv_V4; - - case Hexagon::STrib_cPt: - return Hexagon::STrib_cPt_nv_V4; - - case Hexagon::STrib_cdnPt_V4: - return Hexagon::STrib_cdnPt_nv_V4; - - case Hexagon::STrib_cNotPt: - return Hexagon::STrib_cNotPt_nv_V4; - - case Hexagon::STrib_cdnNotPt_V4: - return Hexagon::STrib_cdnNotPt_nv_V4; - - case Hexagon::STrib_indexed_cPt: - return Hexagon::STrib_indexed_cPt_nv_V4; - - case Hexagon::STrib_indexed_cdnPt_V4: - return Hexagon::STrib_indexed_cdnPt_nv_V4; - - case Hexagon::STrib_indexed_cNotPt: - return Hexagon::STrib_indexed_cNotPt_nv_V4; - - case Hexagon::STrib_indexed_cdnNotPt_V4: - return Hexagon::STrib_indexed_cdnNotPt_nv_V4; - - case Hexagon::STrib_indexed_shl_cPt_V4: - return Hexagon::STrib_indexed_shl_cPt_nv_V4; - - case Hexagon::STrib_indexed_shl_cdnPt_V4: - return Hexagon::STrib_indexed_shl_cdnPt_nv_V4; - - case Hexagon::STrib_indexed_shl_cNotPt_V4: - return Hexagon::STrib_indexed_shl_cNotPt_nv_V4; - - case Hexagon::STrib_indexed_shl_cdnNotPt_V4: - return Hexagon::STrib_indexed_shl_cdnNotPt_nv_V4; - - case Hexagon::POST_STbri_cPt: - return Hexagon::POST_STbri_cPt_nv_V4; - - case Hexagon::POST_STbri_cdnPt_V4: - return Hexagon::POST_STbri_cdnPt_nv_V4; - - case Hexagon::POST_STbri_cNotPt: - return Hexagon::POST_STbri_cNotPt_nv_V4; - - case Hexagon::POST_STbri_cdnNotPt_V4: - return Hexagon::POST_STbri_cdnNotPt_nv_V4; - - case Hexagon::STb_GP_cPt_V4: - return Hexagon::STb_GP_cPt_nv_V4; - - case Hexagon::STb_GP_cNotPt_V4: - return Hexagon::STb_GP_cNotPt_nv_V4; - - case Hexagon::STb_GP_cdnPt_V4: - return Hexagon::STb_GP_cdnPt_nv_V4; - - case Hexagon::STb_GP_cdnNotPt_V4: - return Hexagon::STb_GP_cdnNotPt_nv_V4; - - // store new value halfword - case Hexagon::STrih: - return Hexagon::STrih_nv_V4; - - case Hexagon::STrih_indexed: - return Hexagon::STrih_indexed_nv_V4; - - case Hexagon::STrih_indexed_shl_V4: - return Hexagon::STrih_indexed_shl_nv_V4; - - case Hexagon::STrih_shl_V4: - return Hexagon::STrih_shl_nv_V4; - - case Hexagon::STh_GP_V4: - return Hexagon::STh_GP_nv_V4; - - case Hexagon::POST_SThri: - return Hexagon::POST_SThri_nv_V4; - - case Hexagon::STrih_cPt: - return Hexagon::STrih_cPt_nv_V4; - - case Hexagon::STrih_cdnPt_V4: - return Hexagon::STrih_cdnPt_nv_V4; - - case Hexagon::STrih_cNotPt: - return Hexagon::STrih_cNotPt_nv_V4; - - case Hexagon::STrih_cdnNotPt_V4: - return Hexagon::STrih_cdnNotPt_nv_V4; - - case Hexagon::STrih_indexed_cPt: - return Hexagon::STrih_indexed_cPt_nv_V4; - - case Hexagon::STrih_indexed_cdnPt_V4: - return Hexagon::STrih_indexed_cdnPt_nv_V4; - - case Hexagon::STrih_indexed_cNotPt: - return Hexagon::STrih_indexed_cNotPt_nv_V4; - - case Hexagon::STrih_indexed_cdnNotPt_V4: - return Hexagon::STrih_indexed_cdnNotPt_nv_V4; - - case Hexagon::STrih_indexed_shl_cPt_V4: - return Hexagon::STrih_indexed_shl_cPt_nv_V4; - - case Hexagon::STrih_indexed_shl_cdnPt_V4: - return Hexagon::STrih_indexed_shl_cdnPt_nv_V4; - - case Hexagon::STrih_indexed_shl_cNotPt_V4: - return Hexagon::STrih_indexed_shl_cNotPt_nv_V4; - - case Hexagon::STrih_indexed_shl_cdnNotPt_V4: - return Hexagon::STrih_indexed_shl_cdnNotPt_nv_V4; - - case Hexagon::POST_SThri_cPt: - return Hexagon::POST_SThri_cPt_nv_V4; - - case Hexagon::POST_SThri_cdnPt_V4: - return Hexagon::POST_SThri_cdnPt_nv_V4; - - case Hexagon::POST_SThri_cNotPt: - return Hexagon::POST_SThri_cNotPt_nv_V4; - - case Hexagon::POST_SThri_cdnNotPt_V4: - return Hexagon::POST_SThri_cdnNotPt_nv_V4; - - case Hexagon::STh_GP_cPt_V4: - return Hexagon::STh_GP_cPt_nv_V4; - - case Hexagon::STh_GP_cNotPt_V4: - return Hexagon::STh_GP_cNotPt_nv_V4; - - case Hexagon::STh_GP_cdnPt_V4: - return Hexagon::STh_GP_cdnPt_nv_V4; - - case Hexagon::STh_GP_cdnNotPt_V4: - return Hexagon::STh_GP_cdnNotPt_nv_V4; - - // store new value word - case Hexagon::STriw: - return Hexagon::STriw_nv_V4; - - case Hexagon::STriw_indexed: - return Hexagon::STriw_indexed_nv_V4; - - case Hexagon::STriw_indexed_shl_V4: - return Hexagon::STriw_indexed_shl_nv_V4; - - case Hexagon::STriw_shl_V4: - return Hexagon::STriw_shl_nv_V4; - - case Hexagon::STw_GP_V4: - return Hexagon::STw_GP_nv_V4; - - case Hexagon::POST_STwri: - return Hexagon::POST_STwri_nv_V4; - - case Hexagon::STriw_cPt: - return Hexagon::STriw_cPt_nv_V4; - - case Hexagon::STriw_cdnPt_V4: - return Hexagon::STriw_cdnPt_nv_V4; - - case Hexagon::STriw_cNotPt: - return Hexagon::STriw_cNotPt_nv_V4; - - case Hexagon::STriw_cdnNotPt_V4: - return Hexagon::STriw_cdnNotPt_nv_V4; - - case Hexagon::STriw_indexed_cPt: - return Hexagon::STriw_indexed_cPt_nv_V4; - - case Hexagon::STriw_indexed_cdnPt_V4: - return Hexagon::STriw_indexed_cdnPt_nv_V4; - - case Hexagon::STriw_indexed_cNotPt: - return Hexagon::STriw_indexed_cNotPt_nv_V4; - - case Hexagon::STriw_indexed_cdnNotPt_V4: - return Hexagon::STriw_indexed_cdnNotPt_nv_V4; - - case Hexagon::STriw_indexed_shl_cPt_V4: - return Hexagon::STriw_indexed_shl_cPt_nv_V4; - - case Hexagon::STriw_indexed_shl_cdnPt_V4: - return Hexagon::STriw_indexed_shl_cdnPt_nv_V4; - - case Hexagon::STriw_indexed_shl_cNotPt_V4: - return Hexagon::STriw_indexed_shl_cNotPt_nv_V4; - - case Hexagon::STriw_indexed_shl_cdnNotPt_V4: - return Hexagon::STriw_indexed_shl_cdnNotPt_nv_V4; - - case Hexagon::POST_STwri_cPt: - return Hexagon::POST_STwri_cPt_nv_V4; - - case Hexagon::POST_STwri_cdnPt_V4: - return Hexagon::POST_STwri_cdnPt_nv_V4; - - case Hexagon::POST_STwri_cNotPt: - return Hexagon::POST_STwri_cNotPt_nv_V4; - - case Hexagon::POST_STwri_cdnNotPt_V4: - return Hexagon::POST_STwri_cdnNotPt_nv_V4; - - case Hexagon::STw_GP_cPt_V4: - return Hexagon::STw_GP_cPt_nv_V4; - - case Hexagon::STw_GP_cNotPt_V4: - return Hexagon::STw_GP_cNotPt_nv_V4; - - case Hexagon::STw_GP_cdnPt_V4: - return Hexagon::STw_GP_cdnPt_nv_V4; - - case Hexagon::STw_GP_cdnNotPt_V4: - return Hexagon::STw_GP_cdnNotPt_nv_V4; - - } -} - -// Return .new predicate version for an instruction -static int GetDotNewPredOp(const int opc) { - switch (opc) { - default: llvm_unreachable("Unknown .new type"); - // Conditional stores - // Store byte conditionally - case Hexagon::STrib_cPt : - return Hexagon::STrib_cdnPt_V4; - - case Hexagon::STrib_cNotPt : - return Hexagon::STrib_cdnNotPt_V4; - - case Hexagon::STrib_indexed_cPt : - return Hexagon::STrib_indexed_cdnPt_V4; - - case Hexagon::STrib_indexed_cNotPt : - return Hexagon::STrib_indexed_cdnNotPt_V4; - - case Hexagon::STrib_imm_cPt_V4 : - return Hexagon::STrib_imm_cdnPt_V4; - - case Hexagon::STrib_imm_cNotPt_V4 : - return Hexagon::STrib_imm_cdnNotPt_V4; - - case Hexagon::POST_STbri_cPt : - return Hexagon::POST_STbri_cdnPt_V4; - - case Hexagon::POST_STbri_cNotPt : - return Hexagon::POST_STbri_cdnNotPt_V4; - - case Hexagon::STrib_indexed_shl_cPt_V4 : - return Hexagon::STrib_indexed_shl_cdnPt_V4; - - case Hexagon::STrib_indexed_shl_cNotPt_V4 : - return Hexagon::STrib_indexed_shl_cdnNotPt_V4; - - case Hexagon::STb_GP_cPt_V4 : - return Hexagon::STb_GP_cdnPt_V4; - - case Hexagon::STb_GP_cNotPt_V4 : - return Hexagon::STb_GP_cdnNotPt_V4; - - // Store doubleword conditionally - case Hexagon::STrid_cPt : - return Hexagon::STrid_cdnPt_V4; - - case Hexagon::STrid_cNotPt : - return Hexagon::STrid_cdnNotPt_V4; - - case Hexagon::STrid_indexed_cPt : - return Hexagon::STrid_indexed_cdnPt_V4; - - case Hexagon::STrid_indexed_cNotPt : - return Hexagon::STrid_indexed_cdnNotPt_V4; - - case Hexagon::STrid_indexed_shl_cPt_V4 : - return Hexagon::STrid_indexed_shl_cdnPt_V4; - - case Hexagon::STrid_indexed_shl_cNotPt_V4 : - return Hexagon::STrid_indexed_shl_cdnNotPt_V4; - - case Hexagon::POST_STdri_cPt : - return Hexagon::POST_STdri_cdnPt_V4; - - case Hexagon::POST_STdri_cNotPt : - return Hexagon::POST_STdri_cdnNotPt_V4; - - case Hexagon::STd_GP_cPt_V4 : - return Hexagon::STd_GP_cdnPt_V4; - - case Hexagon::STd_GP_cNotPt_V4 : - return Hexagon::STd_GP_cdnNotPt_V4; - - // Store halfword conditionally - case Hexagon::STrih_cPt : - return Hexagon::STrih_cdnPt_V4; - - case Hexagon::STrih_cNotPt : - return Hexagon::STrih_cdnNotPt_V4; - - case Hexagon::STrih_indexed_cPt : - return Hexagon::STrih_indexed_cdnPt_V4; - - case Hexagon::STrih_indexed_cNotPt : - return Hexagon::STrih_indexed_cdnNotPt_V4; - - case Hexagon::STrih_imm_cPt_V4 : - return Hexagon::STrih_imm_cdnPt_V4; - - case Hexagon::STrih_imm_cNotPt_V4 : - return Hexagon::STrih_imm_cdnNotPt_V4; - - case Hexagon::STrih_indexed_shl_cPt_V4 : - return Hexagon::STrih_indexed_shl_cdnPt_V4; - - case Hexagon::STrih_indexed_shl_cNotPt_V4 : - return Hexagon::STrih_indexed_shl_cdnNotPt_V4; - - case Hexagon::POST_SThri_cPt : - return Hexagon::POST_SThri_cdnPt_V4; - - case Hexagon::POST_SThri_cNotPt : - return Hexagon::POST_SThri_cdnNotPt_V4; - - case Hexagon::STh_GP_cPt_V4 : - return Hexagon::STh_GP_cdnPt_V4; - - case Hexagon::STh_GP_cNotPt_V4 : - return Hexagon::STh_GP_cdnNotPt_V4; - - // Store word conditionally - case Hexagon::STriw_cPt : - return Hexagon::STriw_cdnPt_V4; - - case Hexagon::STriw_cNotPt : - return Hexagon::STriw_cdnNotPt_V4; - - case Hexagon::STriw_indexed_cPt : - return Hexagon::STriw_indexed_cdnPt_V4; - - case Hexagon::STriw_indexed_cNotPt : - return Hexagon::STriw_indexed_cdnNotPt_V4; - - case Hexagon::STriw_imm_cPt_V4 : - return Hexagon::STriw_imm_cdnPt_V4; - - case Hexagon::STriw_imm_cNotPt_V4 : - return Hexagon::STriw_imm_cdnNotPt_V4; - - case Hexagon::STriw_indexed_shl_cPt_V4 : - return Hexagon::STriw_indexed_shl_cdnPt_V4; - - case Hexagon::STriw_indexed_shl_cNotPt_V4 : - return Hexagon::STriw_indexed_shl_cdnNotPt_V4; - - case Hexagon::POST_STwri_cPt : - return Hexagon::POST_STwri_cdnPt_V4; - - case Hexagon::POST_STwri_cNotPt : - return Hexagon::POST_STwri_cdnNotPt_V4; - - case Hexagon::STw_GP_cPt_V4 : - return Hexagon::STw_GP_cdnPt_V4; - - case Hexagon::STw_GP_cNotPt_V4 : - return Hexagon::STw_GP_cdnNotPt_V4; - - // Condtional Jumps - case Hexagon::JMP_c: - return Hexagon::JMP_cdnPt; - - case Hexagon::JMP_cNot: - return Hexagon::JMP_cdnNotPt; - - case Hexagon::JMPR_cPt: - return Hexagon::JMPR_cdnPt_V3; - - case Hexagon::JMPR_cNotPt: - return Hexagon::JMPR_cdnNotPt_V3; - - // Conditional Transfers - case Hexagon::TFR_cPt: - return Hexagon::TFR_cdnPt; - - case Hexagon::TFR_cNotPt: - return Hexagon::TFR_cdnNotPt; - - case Hexagon::TFRI_cPt: - return Hexagon::TFRI_cdnPt; - - case Hexagon::TFRI_cNotPt: - return Hexagon::TFRI_cdnNotPt; - - // Load double word - case Hexagon::LDrid_cPt : - return Hexagon::LDrid_cdnPt; - - case Hexagon::LDrid_cNotPt : - return Hexagon::LDrid_cdnNotPt; - - case Hexagon::LDrid_indexed_cPt : - return Hexagon::LDrid_indexed_cdnPt; - - case Hexagon::LDrid_indexed_cNotPt : - return Hexagon::LDrid_indexed_cdnNotPt; - - case Hexagon::POST_LDrid_cPt : - return Hexagon::POST_LDrid_cdnPt_V4; - - case Hexagon::POST_LDrid_cNotPt : - return Hexagon::POST_LDrid_cdnNotPt_V4; - - // Load word - case Hexagon::LDriw_cPt : - return Hexagon::LDriw_cdnPt; - - case Hexagon::LDriw_cNotPt : - return Hexagon::LDriw_cdnNotPt; - - case Hexagon::LDriw_indexed_cPt : - return Hexagon::LDriw_indexed_cdnPt; - - case Hexagon::LDriw_indexed_cNotPt : - return Hexagon::LDriw_indexed_cdnNotPt; - - case Hexagon::POST_LDriw_cPt : - return Hexagon::POST_LDriw_cdnPt_V4; - - case Hexagon::POST_LDriw_cNotPt : - return Hexagon::POST_LDriw_cdnNotPt_V4; - - // Load halfword - case Hexagon::LDrih_cPt : - return Hexagon::LDrih_cdnPt; - - case Hexagon::LDrih_cNotPt : - return Hexagon::LDrih_cdnNotPt; - - case Hexagon::LDrih_indexed_cPt : - return Hexagon::LDrih_indexed_cdnPt; - - case Hexagon::LDrih_indexed_cNotPt : - return Hexagon::LDrih_indexed_cdnNotPt; - - case Hexagon::POST_LDrih_cPt : - return Hexagon::POST_LDrih_cdnPt_V4; - - case Hexagon::POST_LDrih_cNotPt : - return Hexagon::POST_LDrih_cdnNotPt_V4; - - // Load byte - case Hexagon::LDrib_cPt : - return Hexagon::LDrib_cdnPt; - - case Hexagon::LDrib_cNotPt : - return Hexagon::LDrib_cdnNotPt; - - case Hexagon::LDrib_indexed_cPt : - return Hexagon::LDrib_indexed_cdnPt; - - case Hexagon::LDrib_indexed_cNotPt : - return Hexagon::LDrib_indexed_cdnNotPt; - - case Hexagon::POST_LDrib_cPt : - return Hexagon::POST_LDrib_cdnPt_V4; - - case Hexagon::POST_LDrib_cNotPt : - return Hexagon::POST_LDrib_cdnNotPt_V4; - - // Load unsigned halfword - case Hexagon::LDriuh_cPt : - return Hexagon::LDriuh_cdnPt; - - case Hexagon::LDriuh_cNotPt : - return Hexagon::LDriuh_cdnNotPt; - - case Hexagon::LDriuh_indexed_cPt : - return Hexagon::LDriuh_indexed_cdnPt; - - case Hexagon::LDriuh_indexed_cNotPt : - return Hexagon::LDriuh_indexed_cdnNotPt; - - case Hexagon::POST_LDriuh_cPt : - return Hexagon::POST_LDriuh_cdnPt_V4; - - case Hexagon::POST_LDriuh_cNotPt : - return Hexagon::POST_LDriuh_cdnNotPt_V4; - - // Load unsigned byte - case Hexagon::LDriub_cPt : - return Hexagon::LDriub_cdnPt; - - case Hexagon::LDriub_cNotPt : - return Hexagon::LDriub_cdnNotPt; - - case Hexagon::LDriub_indexed_cPt : - return Hexagon::LDriub_indexed_cdnPt; - - case Hexagon::LDriub_indexed_cNotPt : - return Hexagon::LDriub_indexed_cdnNotPt; - - case Hexagon::POST_LDriub_cPt : - return Hexagon::POST_LDriub_cdnPt_V4; - - case Hexagon::POST_LDriub_cNotPt : - return Hexagon::POST_LDriub_cdnNotPt_V4; - - // V4 indexed+scaled load - - case Hexagon::LDrid_indexed_shl_cPt_V4 : - return Hexagon::LDrid_indexed_shl_cdnPt_V4; - - case Hexagon::LDrid_indexed_shl_cNotPt_V4 : - return Hexagon::LDrid_indexed_shl_cdnNotPt_V4; - - case Hexagon::LDrib_indexed_shl_cPt_V4 : - return Hexagon::LDrib_indexed_shl_cdnPt_V4; - - case Hexagon::LDrib_indexed_shl_cNotPt_V4 : - return Hexagon::LDrib_indexed_shl_cdnNotPt_V4; - - case Hexagon::LDriub_indexed_shl_cPt_V4 : - return Hexagon::LDriub_indexed_shl_cdnPt_V4; - - case Hexagon::LDriub_indexed_shl_cNotPt_V4 : - return Hexagon::LDriub_indexed_shl_cdnNotPt_V4; - - case Hexagon::LDrih_indexed_shl_cPt_V4 : - return Hexagon::LDrih_indexed_shl_cdnPt_V4; - - case Hexagon::LDrih_indexed_shl_cNotPt_V4 : - return Hexagon::LDrih_indexed_shl_cdnNotPt_V4; - - case Hexagon::LDriuh_indexed_shl_cPt_V4 : - return Hexagon::LDriuh_indexed_shl_cdnPt_V4; - - case Hexagon::LDriuh_indexed_shl_cNotPt_V4 : - return Hexagon::LDriuh_indexed_shl_cdnNotPt_V4; - - case Hexagon::LDriw_indexed_shl_cPt_V4 : - return Hexagon::LDriw_indexed_shl_cdnPt_V4; - - case Hexagon::LDriw_indexed_shl_cNotPt_V4 : - return Hexagon::LDriw_indexed_shl_cdnNotPt_V4; - - // V4 global address load - - case Hexagon::LDd_GP_cPt_V4: - return Hexagon::LDd_GP_cdnPt_V4; - - case Hexagon::LDd_GP_cNotPt_V4: - return Hexagon::LDd_GP_cdnNotPt_V4; - - case Hexagon::LDb_GP_cPt_V4: - return Hexagon::LDb_GP_cdnPt_V4; - - case Hexagon::LDb_GP_cNotPt_V4: - return Hexagon::LDb_GP_cdnNotPt_V4; - - case Hexagon::LDub_GP_cPt_V4: - return Hexagon::LDub_GP_cdnPt_V4; - - case Hexagon::LDub_GP_cNotPt_V4: - return Hexagon::LDub_GP_cdnNotPt_V4; - - case Hexagon::LDh_GP_cPt_V4: - return Hexagon::LDh_GP_cdnPt_V4; - - case Hexagon::LDh_GP_cNotPt_V4: - return Hexagon::LDh_GP_cdnNotPt_V4; - - case Hexagon::LDuh_GP_cPt_V4: - return Hexagon::LDuh_GP_cdnPt_V4; - - case Hexagon::LDuh_GP_cNotPt_V4: - return Hexagon::LDuh_GP_cdnNotPt_V4; - - case Hexagon::LDw_GP_cPt_V4: - return Hexagon::LDw_GP_cdnPt_V4; - - case Hexagon::LDw_GP_cNotPt_V4: - return Hexagon::LDw_GP_cdnNotPt_V4; - - // Conditional store new-value byte - case Hexagon::STrib_cPt_nv_V4 : - return Hexagon::STrib_cdnPt_nv_V4; - case Hexagon::STrib_cNotPt_nv_V4 : - return Hexagon::STrib_cdnNotPt_nv_V4; - - case Hexagon::STrib_indexed_cPt_nv_V4 : - return Hexagon::STrib_indexed_cdnPt_nv_V4; - case Hexagon::STrib_indexed_cNotPt_nv_V4 : - return Hexagon::STrib_indexed_cdnNotPt_nv_V4; - - case Hexagon::STrib_indexed_shl_cPt_nv_V4 : - return Hexagon::STrib_indexed_shl_cdnPt_nv_V4; - case Hexagon::STrib_indexed_shl_cNotPt_nv_V4 : - return Hexagon::STrib_indexed_shl_cdnNotPt_nv_V4; - - case Hexagon::POST_STbri_cPt_nv_V4 : - return Hexagon::POST_STbri_cdnPt_nv_V4; - case Hexagon::POST_STbri_cNotPt_nv_V4 : - return Hexagon::POST_STbri_cdnNotPt_nv_V4; - - case Hexagon::STb_GP_cPt_nv_V4 : - return Hexagon::STb_GP_cdnPt_nv_V4; - - case Hexagon::STb_GP_cNotPt_nv_V4 : - return Hexagon::STb_GP_cdnNotPt_nv_V4; - - // Conditional store new-value halfword - case Hexagon::STrih_cPt_nv_V4 : - return Hexagon::STrih_cdnPt_nv_V4; - case Hexagon::STrih_cNotPt_nv_V4 : - return Hexagon::STrih_cdnNotPt_nv_V4; - - case Hexagon::STrih_indexed_cPt_nv_V4 : - return Hexagon::STrih_indexed_cdnPt_nv_V4; - case Hexagon::STrih_indexed_cNotPt_nv_V4 : - return Hexagon::STrih_indexed_cdnNotPt_nv_V4; - - case Hexagon::STrih_indexed_shl_cPt_nv_V4 : - return Hexagon::STrih_indexed_shl_cdnPt_nv_V4; - case Hexagon::STrih_indexed_shl_cNotPt_nv_V4 : - return Hexagon::STrih_indexed_shl_cdnNotPt_nv_V4; - - case Hexagon::POST_SThri_cPt_nv_V4 : - return Hexagon::POST_SThri_cdnPt_nv_V4; - case Hexagon::POST_SThri_cNotPt_nv_V4 : - return Hexagon::POST_SThri_cdnNotPt_nv_V4; - - case Hexagon::STh_GP_cPt_nv_V4 : - return Hexagon::STh_GP_cdnPt_nv_V4; - - case Hexagon::STh_GP_cNotPt_nv_V4 : - return Hexagon::STh_GP_cdnNotPt_nv_V4; - - // Conditional store new-value word - case Hexagon::STriw_cPt_nv_V4 : - return Hexagon::STriw_cdnPt_nv_V4; - case Hexagon::STriw_cNotPt_nv_V4 : - return Hexagon::STriw_cdnNotPt_nv_V4; - - case Hexagon::STriw_indexed_cPt_nv_V4 : - return Hexagon::STriw_indexed_cdnPt_nv_V4; - case Hexagon::STriw_indexed_cNotPt_nv_V4 : - return Hexagon::STriw_indexed_cdnNotPt_nv_V4; - - case Hexagon::STriw_indexed_shl_cPt_nv_V4 : - return Hexagon::STriw_indexed_shl_cdnPt_nv_V4; - case Hexagon::STriw_indexed_shl_cNotPt_nv_V4 : - return Hexagon::STriw_indexed_shl_cdnNotPt_nv_V4; - - case Hexagon::POST_STwri_cPt_nv_V4 : - return Hexagon::POST_STwri_cdnPt_nv_V4; - case Hexagon::POST_STwri_cNotPt_nv_V4: - return Hexagon::POST_STwri_cdnNotPt_nv_V4; - - case Hexagon::STw_GP_cPt_nv_V4 : - return Hexagon::STw_GP_cdnPt_nv_V4; - - case Hexagon::STw_GP_cNotPt_nv_V4 : - return Hexagon::STw_GP_cdnNotPt_nv_V4; - - // Conditional add - case Hexagon::ADD_ri_cPt : - return Hexagon::ADD_ri_cdnPt; - case Hexagon::ADD_ri_cNotPt : - return Hexagon::ADD_ri_cdnNotPt; - - case Hexagon::ADD_rr_cPt : - return Hexagon::ADD_rr_cdnPt; - case Hexagon::ADD_rr_cNotPt : - return Hexagon::ADD_rr_cdnNotPt; - - // Conditional logical Operations - case Hexagon::XOR_rr_cPt : - return Hexagon::XOR_rr_cdnPt; - case Hexagon::XOR_rr_cNotPt : - return Hexagon::XOR_rr_cdnNotPt; - - case Hexagon::AND_rr_cPt : - return Hexagon::AND_rr_cdnPt; - case Hexagon::AND_rr_cNotPt : - return Hexagon::AND_rr_cdnNotPt; - - case Hexagon::OR_rr_cPt : - return Hexagon::OR_rr_cdnPt; - case Hexagon::OR_rr_cNotPt : - return Hexagon::OR_rr_cdnNotPt; - - // Conditional Subtract - case Hexagon::SUB_rr_cPt : - return Hexagon::SUB_rr_cdnPt; - case Hexagon::SUB_rr_cNotPt : - return Hexagon::SUB_rr_cdnNotPt; - - // Conditional combine - case Hexagon::COMBINE_rr_cPt : - return Hexagon::COMBINE_rr_cdnPt; - case Hexagon::COMBINE_rr_cNotPt : - return Hexagon::COMBINE_rr_cdnNotPt; - - case Hexagon::ASLH_cPt_V4 : - return Hexagon::ASLH_cdnPt_V4; - case Hexagon::ASLH_cNotPt_V4 : - return Hexagon::ASLH_cdnNotPt_V4; - - case Hexagon::ASRH_cPt_V4 : - return Hexagon::ASRH_cdnPt_V4; - case Hexagon::ASRH_cNotPt_V4 : - return Hexagon::ASRH_cdnNotPt_V4; - - case Hexagon::SXTB_cPt_V4 : - return Hexagon::SXTB_cdnPt_V4; - case Hexagon::SXTB_cNotPt_V4 : - return Hexagon::SXTB_cdnNotPt_V4; - - case Hexagon::SXTH_cPt_V4 : - return Hexagon::SXTH_cdnPt_V4; - case Hexagon::SXTH_cNotPt_V4 : - return Hexagon::SXTH_cdnNotPt_V4; - - case Hexagon::ZXTB_cPt_V4 : - return Hexagon::ZXTB_cdnPt_V4; - case Hexagon::ZXTB_cNotPt_V4 : - return Hexagon::ZXTB_cdnNotPt_V4; - - case Hexagon::ZXTH_cPt_V4 : - return Hexagon::ZXTH_cdnPt_V4; - case Hexagon::ZXTH_cNotPt_V4 : - return Hexagon::ZXTH_cdnNotPt_V4; - } -} - // Returns true if an instruction can be promoted to .new predicate // or new-value store. bool HexagonPacketizerList::isNewifiable(MachineInstr* MI) { - if ( isCondInst(MI) || IsNewifyStore(MI)) + const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII; + if ( isCondInst(MI) || QII->mayBeNewStore(MI)) return true; else return false; @@ -1262,1062 +437,38 @@ bool HexagonPacketizerList::PromoteToDotNew(MachineInstr* MI, int NewOpcode; if (RC == &Hexagon::PredRegsRegClass) - NewOpcode = GetDotNewPredOp(MI->getOpcode()); + NewOpcode = QII->GetDotNewPredOp(MI, MBPI); else - NewOpcode = GetDotNewOp(MI->getOpcode()); + NewOpcode = QII->GetDotNewOp(MI); MI->setDesc(QII->get(NewOpcode)); return true; } -// Returns the most basic instruction for the .new predicated instructions and -// new-value stores. -// For example, all of the following instructions will be converted back to the -// same instruction: -// 1) if (p0.new) memw(R0+#0) = R1.new ---> -// 2) if (p0) memw(R0+#0)= R1.new -------> if (p0) memw(R0+#0) = R1 -// 3) if (p0.new) memw(R0+#0) = R1 ---> -// -// To understand the translation of instruction 1 to its original form, consider -// a packet with 3 instructions. -// { p0 = cmp.eq(R0,R1) -// if (p0.new) R2 = add(R3, R4) -// R5 = add (R3, R1) -// } -// if (p0) memw(R5+#0) = R2 <--- trying to include it in the previous packet -// -// This instruction can be part of the previous packet only if both p0 and R2 -// are promoted to .new values. This promotion happens in steps, first -// predicate register is promoted to .new and in the next iteration R2 is -// promoted. Therefore, in case of dependence check failure (due to R5) during -// next iteration, it should be converted back to its most basic form. - -static int GetDotOldOp(const int opc) { - switch (opc) { - default: llvm_unreachable("Unknown .old type"); - case Hexagon::TFR_cdnPt: - return Hexagon::TFR_cPt; - - case Hexagon::TFR_cdnNotPt: - return Hexagon::TFR_cNotPt; - - case Hexagon::TFRI_cdnPt: - return Hexagon::TFRI_cPt; - - case Hexagon::TFRI_cdnNotPt: - return Hexagon::TFRI_cNotPt; - - case Hexagon::JMP_cdnPt: - return Hexagon::JMP_c; - - case Hexagon::JMP_cdnNotPt: - return Hexagon::JMP_cNot; - - case Hexagon::JMPR_cdnPt_V3: - return Hexagon::JMPR_cPt; - - case Hexagon::JMPR_cdnNotPt_V3: - return Hexagon::JMPR_cNotPt; - - // Load double word - - case Hexagon::LDrid_cdnPt : - return Hexagon::LDrid_cPt; - - case Hexagon::LDrid_cdnNotPt : - return Hexagon::LDrid_cNotPt; - - case Hexagon::LDrid_indexed_cdnPt : - return Hexagon::LDrid_indexed_cPt; - - case Hexagon::LDrid_indexed_cdnNotPt : - return Hexagon::LDrid_indexed_cNotPt; - - case Hexagon::POST_LDrid_cdnPt_V4 : - return Hexagon::POST_LDrid_cPt; - - case Hexagon::POST_LDrid_cdnNotPt_V4 : - return Hexagon::POST_LDrid_cNotPt; - - // Load word - - case Hexagon::LDriw_cdnPt : - return Hexagon::LDriw_cPt; - - case Hexagon::LDriw_cdnNotPt : - return Hexagon::LDriw_cNotPt; - - case Hexagon::LDriw_indexed_cdnPt : - return Hexagon::LDriw_indexed_cPt; - - case Hexagon::LDriw_indexed_cdnNotPt : - return Hexagon::LDriw_indexed_cNotPt; - - case Hexagon::POST_LDriw_cdnPt_V4 : - return Hexagon::POST_LDriw_cPt; - - case Hexagon::POST_LDriw_cdnNotPt_V4 : - return Hexagon::POST_LDriw_cNotPt; - - // Load half - - case Hexagon::LDrih_cdnPt : - return Hexagon::LDrih_cPt; - - case Hexagon::LDrih_cdnNotPt : - return Hexagon::LDrih_cNotPt; - - case Hexagon::LDrih_indexed_cdnPt : - return Hexagon::LDrih_indexed_cPt; - - case Hexagon::LDrih_indexed_cdnNotPt : - return Hexagon::LDrih_indexed_cNotPt; - - case Hexagon::POST_LDrih_cdnPt_V4 : - return Hexagon::POST_LDrih_cPt; - - case Hexagon::POST_LDrih_cdnNotPt_V4 : - return Hexagon::POST_LDrih_cNotPt; - - // Load byte - - case Hexagon::LDrib_cdnPt : - return Hexagon::LDrib_cPt; - - case Hexagon::LDrib_cdnNotPt : - return Hexagon::LDrib_cNotPt; - - case Hexagon::LDrib_indexed_cdnPt : - return Hexagon::LDrib_indexed_cPt; - - case Hexagon::LDrib_indexed_cdnNotPt : - return Hexagon::LDrib_indexed_cNotPt; - - case Hexagon::POST_LDrib_cdnPt_V4 : - return Hexagon::POST_LDrib_cPt; - - case Hexagon::POST_LDrib_cdnNotPt_V4 : - return Hexagon::POST_LDrib_cNotPt; - - // Load unsigned half - - case Hexagon::LDriuh_cdnPt : - return Hexagon::LDriuh_cPt; - - case Hexagon::LDriuh_cdnNotPt : - return Hexagon::LDriuh_cNotPt; - - case Hexagon::LDriuh_indexed_cdnPt : - return Hexagon::LDriuh_indexed_cPt; - - case Hexagon::LDriuh_indexed_cdnNotPt : - return Hexagon::LDriuh_indexed_cNotPt; - - case Hexagon::POST_LDriuh_cdnPt_V4 : - return Hexagon::POST_LDriuh_cPt; - - case Hexagon::POST_LDriuh_cdnNotPt_V4 : - return Hexagon::POST_LDriuh_cNotPt; - - // Load unsigned byte - case Hexagon::LDriub_cdnPt : - return Hexagon::LDriub_cPt; - - case Hexagon::LDriub_cdnNotPt : - return Hexagon::LDriub_cNotPt; - - case Hexagon::LDriub_indexed_cdnPt : - return Hexagon::LDriub_indexed_cPt; - - case Hexagon::LDriub_indexed_cdnNotPt : - return Hexagon::LDriub_indexed_cNotPt; - - case Hexagon::POST_LDriub_cdnPt_V4 : - return Hexagon::POST_LDriub_cPt; - - case Hexagon::POST_LDriub_cdnNotPt_V4 : - return Hexagon::POST_LDriub_cNotPt; - - // V4 indexed+scaled Load - - case Hexagon::LDrid_indexed_shl_cdnPt_V4 : - return Hexagon::LDrid_indexed_shl_cPt_V4; - - case Hexagon::LDrid_indexed_shl_cdnNotPt_V4 : - return Hexagon::LDrid_indexed_shl_cNotPt_V4; - - case Hexagon::LDrib_indexed_shl_cdnPt_V4 : - return Hexagon::LDrib_indexed_shl_cPt_V4; - - case Hexagon::LDrib_indexed_shl_cdnNotPt_V4 : - return Hexagon::LDrib_indexed_shl_cNotPt_V4; - - case Hexagon::LDriub_indexed_shl_cdnPt_V4 : - return Hexagon::LDriub_indexed_shl_cPt_V4; - - case Hexagon::LDriub_indexed_shl_cdnNotPt_V4 : - return Hexagon::LDriub_indexed_shl_cNotPt_V4; - - case Hexagon::LDrih_indexed_shl_cdnPt_V4 : - return Hexagon::LDrih_indexed_shl_cPt_V4; - - case Hexagon::LDrih_indexed_shl_cdnNotPt_V4 : - return Hexagon::LDrih_indexed_shl_cNotPt_V4; - - case Hexagon::LDriuh_indexed_shl_cdnPt_V4 : - return Hexagon::LDriuh_indexed_shl_cPt_V4; - - case Hexagon::LDriuh_indexed_shl_cdnNotPt_V4 : - return Hexagon::LDriuh_indexed_shl_cNotPt_V4; - - case Hexagon::LDriw_indexed_shl_cdnPt_V4 : - return Hexagon::LDriw_indexed_shl_cPt_V4; - - case Hexagon::LDriw_indexed_shl_cdnNotPt_V4 : - return Hexagon::LDriw_indexed_shl_cNotPt_V4; - - // V4 global address load - - case Hexagon::LDd_GP_cdnPt_V4: - return Hexagon::LDd_GP_cPt_V4; - - case Hexagon::LDd_GP_cdnNotPt_V4: - return Hexagon::LDd_GP_cNotPt_V4; - - case Hexagon::LDb_GP_cdnPt_V4: - return Hexagon::LDb_GP_cPt_V4; - - case Hexagon::LDb_GP_cdnNotPt_V4: - return Hexagon::LDb_GP_cNotPt_V4; - - case Hexagon::LDub_GP_cdnPt_V4: - return Hexagon::LDub_GP_cPt_V4; - - case Hexagon::LDub_GP_cdnNotPt_V4: - return Hexagon::LDub_GP_cNotPt_V4; - - case Hexagon::LDh_GP_cdnPt_V4: - return Hexagon::LDh_GP_cPt_V4; - - case Hexagon::LDh_GP_cdnNotPt_V4: - return Hexagon::LDh_GP_cNotPt_V4; - - case Hexagon::LDuh_GP_cdnPt_V4: - return Hexagon::LDuh_GP_cPt_V4; - - case Hexagon::LDuh_GP_cdnNotPt_V4: - return Hexagon::LDuh_GP_cNotPt_V4; - - case Hexagon::LDw_GP_cdnPt_V4: - return Hexagon::LDw_GP_cPt_V4; - - case Hexagon::LDw_GP_cdnNotPt_V4: - return Hexagon::LDw_GP_cNotPt_V4; - - // Conditional add - - case Hexagon::ADD_ri_cdnPt : - return Hexagon::ADD_ri_cPt; - case Hexagon::ADD_ri_cdnNotPt : - return Hexagon::ADD_ri_cNotPt; - - case Hexagon::ADD_rr_cdnPt : - return Hexagon::ADD_rr_cPt; - case Hexagon::ADD_rr_cdnNotPt: - return Hexagon::ADD_rr_cNotPt; - - // Conditional logical Operations - - case Hexagon::XOR_rr_cdnPt : - return Hexagon::XOR_rr_cPt; - case Hexagon::XOR_rr_cdnNotPt : - return Hexagon::XOR_rr_cNotPt; - - case Hexagon::AND_rr_cdnPt : - return Hexagon::AND_rr_cPt; - case Hexagon::AND_rr_cdnNotPt : - return Hexagon::AND_rr_cNotPt; - - case Hexagon::OR_rr_cdnPt : - return Hexagon::OR_rr_cPt; - case Hexagon::OR_rr_cdnNotPt : - return Hexagon::OR_rr_cNotPt; - - // Conditional Subtract - - case Hexagon::SUB_rr_cdnPt : - return Hexagon::SUB_rr_cPt; - case Hexagon::SUB_rr_cdnNotPt : - return Hexagon::SUB_rr_cNotPt; - - // Conditional combine - - case Hexagon::COMBINE_rr_cdnPt : - return Hexagon::COMBINE_rr_cPt; - case Hexagon::COMBINE_rr_cdnNotPt : - return Hexagon::COMBINE_rr_cNotPt; - -// Conditional shift operations - - case Hexagon::ASLH_cdnPt_V4 : - return Hexagon::ASLH_cPt_V4; - case Hexagon::ASLH_cdnNotPt_V4 : - return Hexagon::ASLH_cNotPt_V4; - - case Hexagon::ASRH_cdnPt_V4 : - return Hexagon::ASRH_cPt_V4; - case Hexagon::ASRH_cdnNotPt_V4 : - return Hexagon::ASRH_cNotPt_V4; - - case Hexagon::SXTB_cdnPt_V4 : - return Hexagon::SXTB_cPt_V4; - case Hexagon::SXTB_cdnNotPt_V4 : - return Hexagon::SXTB_cNotPt_V4; - - case Hexagon::SXTH_cdnPt_V4 : - return Hexagon::SXTH_cPt_V4; - case Hexagon::SXTH_cdnNotPt_V4 : - return Hexagon::SXTH_cNotPt_V4; - - case Hexagon::ZXTB_cdnPt_V4 : - return Hexagon::ZXTB_cPt_V4; - case Hexagon::ZXTB_cdnNotPt_V4 : - return Hexagon::ZXTB_cNotPt_V4; - - case Hexagon::ZXTH_cdnPt_V4 : - return Hexagon::ZXTH_cPt_V4; - case Hexagon::ZXTH_cdnNotPt_V4 : - return Hexagon::ZXTH_cNotPt_V4; - - // Store byte - - case Hexagon::STrib_imm_cdnPt_V4 : - return Hexagon::STrib_imm_cPt_V4; - - case Hexagon::STrib_imm_cdnNotPt_V4 : - return Hexagon::STrib_imm_cNotPt_V4; - - case Hexagon::STrib_cdnPt_nv_V4 : - case Hexagon::STrib_cPt_nv_V4 : - case Hexagon::STrib_cdnPt_V4 : - return Hexagon::STrib_cPt; - - case Hexagon::STrib_cdnNotPt_nv_V4 : - case Hexagon::STrib_cNotPt_nv_V4 : - case Hexagon::STrib_cdnNotPt_V4 : - return Hexagon::STrib_cNotPt; - - case Hexagon::STrib_indexed_cdnPt_V4 : - case Hexagon::STrib_indexed_cPt_nv_V4 : - case Hexagon::STrib_indexed_cdnPt_nv_V4 : - return Hexagon::STrib_indexed_cPt; - - case Hexagon::STrib_indexed_cdnNotPt_V4 : - case Hexagon::STrib_indexed_cNotPt_nv_V4 : - case Hexagon::STrib_indexed_cdnNotPt_nv_V4 : - return Hexagon::STrib_indexed_cNotPt; - - case Hexagon::STrib_indexed_shl_cdnPt_nv_V4: - case Hexagon::STrib_indexed_shl_cPt_nv_V4 : - case Hexagon::STrib_indexed_shl_cdnPt_V4 : - return Hexagon::STrib_indexed_shl_cPt_V4; - - case Hexagon::STrib_indexed_shl_cdnNotPt_nv_V4: - case Hexagon::STrib_indexed_shl_cNotPt_nv_V4 : - case Hexagon::STrib_indexed_shl_cdnNotPt_V4 : - return Hexagon::STrib_indexed_shl_cNotPt_V4; - - case Hexagon::POST_STbri_cdnPt_nv_V4 : - case Hexagon::POST_STbri_cPt_nv_V4 : - case Hexagon::POST_STbri_cdnPt_V4 : - return Hexagon::POST_STbri_cPt; - - case Hexagon::POST_STbri_cdnNotPt_nv_V4 : - case Hexagon::POST_STbri_cNotPt_nv_V4: - case Hexagon::POST_STbri_cdnNotPt_V4 : - return Hexagon::POST_STbri_cNotPt; - - case Hexagon::STb_GP_cdnPt_nv_V4: - case Hexagon::STb_GP_cdnPt_V4: - case Hexagon::STb_GP_cPt_nv_V4: - return Hexagon::STb_GP_cPt_V4; - - case Hexagon::STb_GP_cdnNotPt_nv_V4: - case Hexagon::STb_GP_cdnNotPt_V4: - case Hexagon::STb_GP_cNotPt_nv_V4: - return Hexagon::STb_GP_cNotPt_V4; - - // Store new-value byte - unconditional - case Hexagon::STrib_nv_V4: - return Hexagon::STrib; - - case Hexagon::STrib_indexed_nv_V4: - return Hexagon::STrib_indexed; - - case Hexagon::STrib_indexed_shl_nv_V4: - return Hexagon::STrib_indexed_shl_V4; - - case Hexagon::STrib_shl_nv_V4: - return Hexagon::STrib_shl_V4; - - case Hexagon::STb_GP_nv_V4: - return Hexagon::STb_GP_V4; - - case Hexagon::POST_STbri_nv_V4: - return Hexagon::POST_STbri; - - // Store halfword - case Hexagon::STrih_imm_cdnPt_V4 : - return Hexagon::STrih_imm_cPt_V4; - - case Hexagon::STrih_imm_cdnNotPt_V4 : - return Hexagon::STrih_imm_cNotPt_V4; - - case Hexagon::STrih_cdnPt_nv_V4 : - case Hexagon::STrih_cPt_nv_V4 : - case Hexagon::STrih_cdnPt_V4 : - return Hexagon::STrih_cPt; - - case Hexagon::STrih_cdnNotPt_nv_V4 : - case Hexagon::STrih_cNotPt_nv_V4 : - case Hexagon::STrih_cdnNotPt_V4 : - return Hexagon::STrih_cNotPt; - - case Hexagon::STrih_indexed_cdnPt_nv_V4: - case Hexagon::STrih_indexed_cPt_nv_V4 : - case Hexagon::STrih_indexed_cdnPt_V4 : - return Hexagon::STrih_indexed_cPt; - - case Hexagon::STrih_indexed_cdnNotPt_nv_V4: - case Hexagon::STrih_indexed_cNotPt_nv_V4 : - case Hexagon::STrih_indexed_cdnNotPt_V4 : - return Hexagon::STrih_indexed_cNotPt; - - case Hexagon::STrih_indexed_shl_cdnPt_nv_V4 : - case Hexagon::STrih_indexed_shl_cPt_nv_V4 : - case Hexagon::STrih_indexed_shl_cdnPt_V4 : - return Hexagon::STrih_indexed_shl_cPt_V4; - - case Hexagon::STrih_indexed_shl_cdnNotPt_nv_V4 : - case Hexagon::STrih_indexed_shl_cNotPt_nv_V4 : - case Hexagon::STrih_indexed_shl_cdnNotPt_V4 : - return Hexagon::STrih_indexed_shl_cNotPt_V4; - - case Hexagon::POST_SThri_cdnPt_nv_V4 : - case Hexagon::POST_SThri_cPt_nv_V4 : - case Hexagon::POST_SThri_cdnPt_V4 : - return Hexagon::POST_SThri_cPt; - - case Hexagon::POST_SThri_cdnNotPt_nv_V4 : - case Hexagon::POST_SThri_cNotPt_nv_V4 : - case Hexagon::POST_SThri_cdnNotPt_V4 : - return Hexagon::POST_SThri_cNotPt; - - case Hexagon::STh_GP_cdnPt_nv_V4: - case Hexagon::STh_GP_cdnPt_V4: - case Hexagon::STh_GP_cPt_nv_V4: - return Hexagon::STh_GP_cPt_V4; - - case Hexagon::STh_GP_cdnNotPt_nv_V4: - case Hexagon::STh_GP_cdnNotPt_V4: - case Hexagon::STh_GP_cNotPt_nv_V4: - return Hexagon::STh_GP_cNotPt_V4; - - // Store new-value halfword - unconditional - - case Hexagon::STrih_nv_V4: - return Hexagon::STrih; - - case Hexagon::STrih_indexed_nv_V4: - return Hexagon::STrih_indexed; - - case Hexagon::STrih_indexed_shl_nv_V4: - return Hexagon::STrih_indexed_shl_V4; - - case Hexagon::STrih_shl_nv_V4: - return Hexagon::STrih_shl_V4; - - case Hexagon::STh_GP_nv_V4: - return Hexagon::STh_GP_V4; - - case Hexagon::POST_SThri_nv_V4: - return Hexagon::POST_SThri; - - // Store word - - case Hexagon::STriw_imm_cdnPt_V4 : - return Hexagon::STriw_imm_cPt_V4; - - case Hexagon::STriw_imm_cdnNotPt_V4 : - return Hexagon::STriw_imm_cNotPt_V4; - - case Hexagon::STriw_cdnPt_nv_V4 : - case Hexagon::STriw_cPt_nv_V4 : - case Hexagon::STriw_cdnPt_V4 : - return Hexagon::STriw_cPt; - - case Hexagon::STriw_cdnNotPt_nv_V4 : - case Hexagon::STriw_cNotPt_nv_V4 : - case Hexagon::STriw_cdnNotPt_V4 : - return Hexagon::STriw_cNotPt; - - case Hexagon::STriw_indexed_cdnPt_nv_V4 : - case Hexagon::STriw_indexed_cPt_nv_V4 : - case Hexagon::STriw_indexed_cdnPt_V4 : - return Hexagon::STriw_indexed_cPt; - - case Hexagon::STriw_indexed_cdnNotPt_nv_V4 : - case Hexagon::STriw_indexed_cNotPt_nv_V4 : - case Hexagon::STriw_indexed_cdnNotPt_V4 : - return Hexagon::STriw_indexed_cNotPt; - - case Hexagon::STriw_indexed_shl_cdnPt_nv_V4 : - case Hexagon::STriw_indexed_shl_cPt_nv_V4 : - case Hexagon::STriw_indexed_shl_cdnPt_V4 : - return Hexagon::STriw_indexed_shl_cPt_V4; - - case Hexagon::STriw_indexed_shl_cdnNotPt_nv_V4 : - case Hexagon::STriw_indexed_shl_cNotPt_nv_V4 : - case Hexagon::STriw_indexed_shl_cdnNotPt_V4 : - return Hexagon::STriw_indexed_shl_cNotPt_V4; - - case Hexagon::POST_STwri_cdnPt_nv_V4 : - case Hexagon::POST_STwri_cPt_nv_V4 : - case Hexagon::POST_STwri_cdnPt_V4 : - return Hexagon::POST_STwri_cPt; - - case Hexagon::POST_STwri_cdnNotPt_nv_V4 : - case Hexagon::POST_STwri_cNotPt_nv_V4 : - case Hexagon::POST_STwri_cdnNotPt_V4 : - return Hexagon::POST_STwri_cNotPt; - - case Hexagon::STw_GP_cdnPt_nv_V4: - case Hexagon::STw_GP_cdnPt_V4: - case Hexagon::STw_GP_cPt_nv_V4: - return Hexagon::STw_GP_cPt_V4; - - case Hexagon::STw_GP_cdnNotPt_nv_V4: - case Hexagon::STw_GP_cdnNotPt_V4: - case Hexagon::STw_GP_cNotPt_nv_V4: - return Hexagon::STw_GP_cNotPt_V4; - - // Store new-value word - unconditional - - case Hexagon::STriw_nv_V4: - return Hexagon::STriw; - - case Hexagon::STriw_indexed_nv_V4: - return Hexagon::STriw_indexed; - - case Hexagon::STriw_indexed_shl_nv_V4: - return Hexagon::STriw_indexed_shl_V4; - - case Hexagon::STriw_shl_nv_V4: - return Hexagon::STriw_shl_V4; - - case Hexagon::STw_GP_nv_V4: - return Hexagon::STw_GP_V4; - - case Hexagon::POST_STwri_nv_V4: - return Hexagon::POST_STwri; - - // Store doubleword - - case Hexagon::STrid_cdnPt_V4 : - return Hexagon::STrid_cPt; - - case Hexagon::STrid_cdnNotPt_V4 : - return Hexagon::STrid_cNotPt; - - case Hexagon::STrid_indexed_cdnPt_V4 : - return Hexagon::STrid_indexed_cPt; - - case Hexagon::STrid_indexed_cdnNotPt_V4 : - return Hexagon::STrid_indexed_cNotPt; - - case Hexagon::STrid_indexed_shl_cdnPt_V4 : - return Hexagon::STrid_indexed_shl_cPt_V4; - - case Hexagon::STrid_indexed_shl_cdnNotPt_V4 : - return Hexagon::STrid_indexed_shl_cNotPt_V4; - - case Hexagon::POST_STdri_cdnPt_V4 : - return Hexagon::POST_STdri_cPt; - - case Hexagon::POST_STdri_cdnNotPt_V4 : - return Hexagon::POST_STdri_cNotPt; - - case Hexagon::STd_GP_cdnPt_V4 : - return Hexagon::STd_GP_cPt_V4; - - case Hexagon::STd_GP_cdnNotPt_V4 : - return Hexagon::STd_GP_cNotPt_V4; - - } -} - bool HexagonPacketizerList::DemoteToDotOld(MachineInstr* MI) { const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII; - int NewOpcode = GetDotOldOp(MI->getOpcode()); + int NewOpcode = QII->GetDotOldOp(MI->getOpcode()); MI->setDesc(QII->get(NewOpcode)); return true; } -// Returns true if an instruction is predicated on p0 and false if it's -// predicated on !p0. - -static bool GetPredicateSense(MachineInstr* MI, - const HexagonInstrInfo *QII) { - - switch (MI->getOpcode()) { - default: llvm_unreachable("Unknown predicate sense of the instruction"); - case Hexagon::TFR_cPt: - case Hexagon::TFR_cdnPt: - case Hexagon::TFRI_cPt: - case Hexagon::TFRI_cdnPt: - case Hexagon::STrib_cPt : - case Hexagon::STrib_cdnPt_V4 : - case Hexagon::STrib_indexed_cPt : - case Hexagon::STrib_indexed_cdnPt_V4 : - case Hexagon::STrib_indexed_shl_cPt_V4 : - case Hexagon::STrib_indexed_shl_cdnPt_V4 : - case Hexagon::POST_STbri_cPt : - case Hexagon::POST_STbri_cdnPt_V4 : - case Hexagon::STrih_cPt : - case Hexagon::STrih_cdnPt_V4 : - case Hexagon::STrih_indexed_cPt : - case Hexagon::STrih_indexed_cdnPt_V4 : - case Hexagon::STrih_indexed_shl_cPt_V4 : - case Hexagon::STrih_indexed_shl_cdnPt_V4 : - case Hexagon::POST_SThri_cPt : - case Hexagon::POST_SThri_cdnPt_V4 : - case Hexagon::STriw_cPt : - case Hexagon::STriw_cdnPt_V4 : - case Hexagon::STriw_indexed_cPt : - case Hexagon::STriw_indexed_cdnPt_V4 : - case Hexagon::STriw_indexed_shl_cPt_V4 : - case Hexagon::STriw_indexed_shl_cdnPt_V4 : - case Hexagon::POST_STwri_cPt : - case Hexagon::POST_STwri_cdnPt_V4 : - case Hexagon::STrib_imm_cPt_V4 : - case Hexagon::STrib_imm_cdnPt_V4 : - case Hexagon::STrid_cPt : - case Hexagon::STrid_cdnPt_V4 : - case Hexagon::STrid_indexed_cPt : - case Hexagon::STrid_indexed_cdnPt_V4 : - case Hexagon::STrid_indexed_shl_cPt_V4 : - case Hexagon::STrid_indexed_shl_cdnPt_V4 : - case Hexagon::POST_STdri_cPt : - case Hexagon::POST_STdri_cdnPt_V4 : - case Hexagon::STrih_imm_cPt_V4 : - case Hexagon::STrih_imm_cdnPt_V4 : - case Hexagon::STriw_imm_cPt_V4 : - case Hexagon::STriw_imm_cdnPt_V4 : - case Hexagon::JMP_cdnPt : - case Hexagon::LDrid_cPt : - case Hexagon::LDrid_cdnPt : - case Hexagon::LDrid_indexed_cPt : - case Hexagon::LDrid_indexed_cdnPt : - case Hexagon::POST_LDrid_cPt : - case Hexagon::POST_LDrid_cdnPt_V4 : - case Hexagon::LDriw_cPt : - case Hexagon::LDriw_cdnPt : - case Hexagon::LDriw_indexed_cPt : - case Hexagon::LDriw_indexed_cdnPt : - case Hexagon::POST_LDriw_cPt : - case Hexagon::POST_LDriw_cdnPt_V4 : - case Hexagon::LDrih_cPt : - case Hexagon::LDrih_cdnPt : - case Hexagon::LDrih_indexed_cPt : - case Hexagon::LDrih_indexed_cdnPt : - case Hexagon::POST_LDrih_cPt : - case Hexagon::POST_LDrih_cdnPt_V4 : - case Hexagon::LDrib_cPt : - case Hexagon::LDrib_cdnPt : - case Hexagon::LDrib_indexed_cPt : - case Hexagon::LDrib_indexed_cdnPt : - case Hexagon::POST_LDrib_cPt : - case Hexagon::POST_LDrib_cdnPt_V4 : - case Hexagon::LDriuh_cPt : - case Hexagon::LDriuh_cdnPt : - case Hexagon::LDriuh_indexed_cPt : - case Hexagon::LDriuh_indexed_cdnPt : - case Hexagon::POST_LDriuh_cPt : - case Hexagon::POST_LDriuh_cdnPt_V4 : - case Hexagon::LDriub_cPt : - case Hexagon::LDriub_cdnPt : - case Hexagon::LDriub_indexed_cPt : - case Hexagon::LDriub_indexed_cdnPt : - case Hexagon::POST_LDriub_cPt : - case Hexagon::POST_LDriub_cdnPt_V4 : - case Hexagon::LDrid_indexed_shl_cPt_V4 : - case Hexagon::LDrid_indexed_shl_cdnPt_V4 : - case Hexagon::LDrib_indexed_shl_cPt_V4 : - case Hexagon::LDrib_indexed_shl_cdnPt_V4 : - case Hexagon::LDriub_indexed_shl_cPt_V4 : - case Hexagon::LDriub_indexed_shl_cdnPt_V4 : - case Hexagon::LDrih_indexed_shl_cPt_V4 : - case Hexagon::LDrih_indexed_shl_cdnPt_V4 : - case Hexagon::LDriuh_indexed_shl_cPt_V4 : - case Hexagon::LDriuh_indexed_shl_cdnPt_V4 : - case Hexagon::LDriw_indexed_shl_cPt_V4 : - case Hexagon::LDriw_indexed_shl_cdnPt_V4 : - case Hexagon::ADD_ri_cPt : - case Hexagon::ADD_ri_cdnPt : - case Hexagon::ADD_rr_cPt : - case Hexagon::ADD_rr_cdnPt : - case Hexagon::XOR_rr_cPt : - case Hexagon::XOR_rr_cdnPt : - case Hexagon::AND_rr_cPt : - case Hexagon::AND_rr_cdnPt : - case Hexagon::OR_rr_cPt : - case Hexagon::OR_rr_cdnPt : - case Hexagon::SUB_rr_cPt : - case Hexagon::SUB_rr_cdnPt : - case Hexagon::COMBINE_rr_cPt : - case Hexagon::COMBINE_rr_cdnPt : - case Hexagon::ASLH_cPt_V4 : - case Hexagon::ASLH_cdnPt_V4 : - case Hexagon::ASRH_cPt_V4 : - case Hexagon::ASRH_cdnPt_V4 : - case Hexagon::SXTB_cPt_V4 : - case Hexagon::SXTB_cdnPt_V4 : - case Hexagon::SXTH_cPt_V4 : - case Hexagon::SXTH_cdnPt_V4 : - case Hexagon::ZXTB_cPt_V4 : - case Hexagon::ZXTB_cdnPt_V4 : - case Hexagon::ZXTH_cPt_V4 : - case Hexagon::ZXTH_cdnPt_V4 : - case Hexagon::LDd_GP_cPt_V4 : - case Hexagon::LDb_GP_cPt_V4 : - case Hexagon::LDub_GP_cPt_V4 : - case Hexagon::LDh_GP_cPt_V4 : - case Hexagon::LDuh_GP_cPt_V4 : - case Hexagon::LDw_GP_cPt_V4 : - case Hexagon::STd_GP_cPt_V4 : - case Hexagon::STb_GP_cPt_V4 : - case Hexagon::STh_GP_cPt_V4 : - case Hexagon::STw_GP_cPt_V4 : - case Hexagon::LDd_GP_cdnPt_V4 : - case Hexagon::LDb_GP_cdnPt_V4 : - case Hexagon::LDub_GP_cdnPt_V4 : - case Hexagon::LDh_GP_cdnPt_V4 : - case Hexagon::LDuh_GP_cdnPt_V4 : - case Hexagon::LDw_GP_cdnPt_V4 : - case Hexagon::STd_GP_cdnPt_V4 : - case Hexagon::STb_GP_cdnPt_V4 : - case Hexagon::STh_GP_cdnPt_V4 : - case Hexagon::STw_GP_cdnPt_V4 : - return true; +enum PredicateKind { + PK_False, + PK_True, + PK_Unknown +}; - case Hexagon::TFR_cNotPt: - case Hexagon::TFR_cdnNotPt: - case Hexagon::TFRI_cNotPt: - case Hexagon::TFRI_cdnNotPt: - case Hexagon::STrib_cNotPt : - case Hexagon::STrib_cdnNotPt_V4 : - case Hexagon::STrib_indexed_cNotPt : - case Hexagon::STrib_indexed_cdnNotPt_V4 : - case Hexagon::STrib_indexed_shl_cNotPt_V4 : - case Hexagon::STrib_indexed_shl_cdnNotPt_V4 : - case Hexagon::POST_STbri_cNotPt : - case Hexagon::POST_STbri_cdnNotPt_V4 : - case Hexagon::STrih_cNotPt : - case Hexagon::STrih_cdnNotPt_V4 : - case Hexagon::STrih_indexed_cNotPt : - case Hexagon::STrih_indexed_cdnNotPt_V4 : - case Hexagon::STrih_indexed_shl_cNotPt_V4 : - case Hexagon::STrih_indexed_shl_cdnNotPt_V4 : - case Hexagon::POST_SThri_cNotPt : - case Hexagon::POST_SThri_cdnNotPt_V4 : - case Hexagon::STriw_cNotPt : - case Hexagon::STriw_cdnNotPt_V4 : - case Hexagon::STriw_indexed_cNotPt : - case Hexagon::STriw_indexed_cdnNotPt_V4 : - case Hexagon::STriw_indexed_shl_cNotPt_V4 : - case Hexagon::STriw_indexed_shl_cdnNotPt_V4 : - case Hexagon::POST_STwri_cNotPt : - case Hexagon::POST_STwri_cdnNotPt_V4 : - case Hexagon::STrib_imm_cNotPt_V4 : - case Hexagon::STrib_imm_cdnNotPt_V4 : - case Hexagon::STrid_cNotPt : - case Hexagon::STrid_cdnNotPt_V4 : - case Hexagon::STrid_indexed_cdnNotPt_V4 : - case Hexagon::STrid_indexed_cNotPt : - case Hexagon::STrid_indexed_shl_cNotPt_V4 : - case Hexagon::STrid_indexed_shl_cdnNotPt_V4 : - case Hexagon::POST_STdri_cNotPt : - case Hexagon::POST_STdri_cdnNotPt_V4 : - case Hexagon::STrih_imm_cNotPt_V4 : - case Hexagon::STrih_imm_cdnNotPt_V4 : - case Hexagon::STriw_imm_cNotPt_V4 : - case Hexagon::STriw_imm_cdnNotPt_V4 : - case Hexagon::JMP_cdnNotPt : - case Hexagon::LDrid_cNotPt : - case Hexagon::LDrid_cdnNotPt : - case Hexagon::LDrid_indexed_cNotPt : - case Hexagon::LDrid_indexed_cdnNotPt : - case Hexagon::POST_LDrid_cNotPt : - case Hexagon::POST_LDrid_cdnNotPt_V4 : - case Hexagon::LDriw_cNotPt : - case Hexagon::LDriw_cdnNotPt : - case Hexagon::LDriw_indexed_cNotPt : - case Hexagon::LDriw_indexed_cdnNotPt : - case Hexagon::POST_LDriw_cNotPt : - case Hexagon::POST_LDriw_cdnNotPt_V4 : - case Hexagon::LDrih_cNotPt : - case Hexagon::LDrih_cdnNotPt : - case Hexagon::LDrih_indexed_cNotPt : - case Hexagon::LDrih_indexed_cdnNotPt : - case Hexagon::POST_LDrih_cNotPt : - case Hexagon::POST_LDrih_cdnNotPt_V4 : - case Hexagon::LDrib_cNotPt : - case Hexagon::LDrib_cdnNotPt : - case Hexagon::LDrib_indexed_cNotPt : - case Hexagon::LDrib_indexed_cdnNotPt : - case Hexagon::POST_LDrib_cNotPt : - case Hexagon::POST_LDrib_cdnNotPt_V4 : - case Hexagon::LDriuh_cNotPt : - case Hexagon::LDriuh_cdnNotPt : - case Hexagon::LDriuh_indexed_cNotPt : - case Hexagon::LDriuh_indexed_cdnNotPt : - case Hexagon::POST_LDriuh_cNotPt : - case Hexagon::POST_LDriuh_cdnNotPt_V4 : - case Hexagon::LDriub_cNotPt : - case Hexagon::LDriub_cdnNotPt : - case Hexagon::LDriub_indexed_cNotPt : - case Hexagon::LDriub_indexed_cdnNotPt : - case Hexagon::POST_LDriub_cNotPt : - case Hexagon::POST_LDriub_cdnNotPt_V4 : - case Hexagon::LDrid_indexed_shl_cNotPt_V4 : - case Hexagon::LDrid_indexed_shl_cdnNotPt_V4 : - case Hexagon::LDrib_indexed_shl_cNotPt_V4 : - case Hexagon::LDrib_indexed_shl_cdnNotPt_V4 : - case Hexagon::LDriub_indexed_shl_cNotPt_V4 : - case Hexagon::LDriub_indexed_shl_cdnNotPt_V4 : - case Hexagon::LDrih_indexed_shl_cNotPt_V4 : - case Hexagon::LDrih_indexed_shl_cdnNotPt_V4 : - case Hexagon::LDriuh_indexed_shl_cNotPt_V4 : - case Hexagon::LDriuh_indexed_shl_cdnNotPt_V4 : - case Hexagon::LDriw_indexed_shl_cNotPt_V4 : - case Hexagon::LDriw_indexed_shl_cdnNotPt_V4 : - case Hexagon::ADD_ri_cNotPt : - case Hexagon::ADD_ri_cdnNotPt : - case Hexagon::ADD_rr_cNotPt : - case Hexagon::ADD_rr_cdnNotPt : - case Hexagon::XOR_rr_cNotPt : - case Hexagon::XOR_rr_cdnNotPt : - case Hexagon::AND_rr_cNotPt : - case Hexagon::AND_rr_cdnNotPt : - case Hexagon::OR_rr_cNotPt : - case Hexagon::OR_rr_cdnNotPt : - case Hexagon::SUB_rr_cNotPt : - case Hexagon::SUB_rr_cdnNotPt : - case Hexagon::COMBINE_rr_cNotPt : - case Hexagon::COMBINE_rr_cdnNotPt : - case Hexagon::ASLH_cNotPt_V4 : - case Hexagon::ASLH_cdnNotPt_V4 : - case Hexagon::ASRH_cNotPt_V4 : - case Hexagon::ASRH_cdnNotPt_V4 : - case Hexagon::SXTB_cNotPt_V4 : - case Hexagon::SXTB_cdnNotPt_V4 : - case Hexagon::SXTH_cNotPt_V4 : - case Hexagon::SXTH_cdnNotPt_V4 : - case Hexagon::ZXTB_cNotPt_V4 : - case Hexagon::ZXTB_cdnNotPt_V4 : - case Hexagon::ZXTH_cNotPt_V4 : - case Hexagon::ZXTH_cdnNotPt_V4 : - - case Hexagon::LDd_GP_cNotPt_V4 : - case Hexagon::LDb_GP_cNotPt_V4 : - case Hexagon::LDub_GP_cNotPt_V4 : - case Hexagon::LDh_GP_cNotPt_V4 : - case Hexagon::LDuh_GP_cNotPt_V4 : - case Hexagon::LDw_GP_cNotPt_V4 : - case Hexagon::STd_GP_cNotPt_V4 : - case Hexagon::STb_GP_cNotPt_V4 : - case Hexagon::STh_GP_cNotPt_V4 : - case Hexagon::STw_GP_cNotPt_V4 : - case Hexagon::LDd_GP_cdnNotPt_V4 : - case Hexagon::LDb_GP_cdnNotPt_V4 : - case Hexagon::LDub_GP_cdnNotPt_V4 : - case Hexagon::LDh_GP_cdnNotPt_V4 : - case Hexagon::LDuh_GP_cdnNotPt_V4 : - case Hexagon::LDw_GP_cdnNotPt_V4 : - case Hexagon::STd_GP_cdnNotPt_V4 : - case Hexagon::STb_GP_cdnNotPt_V4 : - case Hexagon::STh_GP_cdnNotPt_V4 : - case Hexagon::STw_GP_cdnNotPt_V4 : - return false; - } - // return *some value* to avoid compiler warning - return false; -} +/// Returns true if an instruction is predicated on p0 and false if it's +/// predicated on !p0. +static PredicateKind getPredicateSense(MachineInstr* MI, + const HexagonInstrInfo *QII) { + if (!QII->isPredicated(MI)) + return PK_Unknown; -bool HexagonPacketizerList::isDotNewInst(MachineInstr* MI) { - const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII; - if (QII->isNewValueInst(MI)) - return true; + if (QII->isPredicatedTrue(MI)) + return PK_True; - switch (MI->getOpcode()) { - case Hexagon::TFR_cdnNotPt: - case Hexagon::TFR_cdnPt: - case Hexagon::TFRI_cdnNotPt: - case Hexagon::TFRI_cdnPt: - case Hexagon::LDrid_cdnPt : - case Hexagon::LDrid_cdnNotPt : - case Hexagon::LDrid_indexed_cdnPt : - case Hexagon::LDrid_indexed_cdnNotPt : - case Hexagon::POST_LDrid_cdnPt_V4 : - case Hexagon::POST_LDrid_cdnNotPt_V4 : - case Hexagon::LDriw_cdnPt : - case Hexagon::LDriw_cdnNotPt : - case Hexagon::LDriw_indexed_cdnPt : - case Hexagon::LDriw_indexed_cdnNotPt : - case Hexagon::POST_LDriw_cdnPt_V4 : - case Hexagon::POST_LDriw_cdnNotPt_V4 : - case Hexagon::LDrih_cdnPt : - case Hexagon::LDrih_cdnNotPt : - case Hexagon::LDrih_indexed_cdnPt : - case Hexagon::LDrih_indexed_cdnNotPt : - case Hexagon::POST_LDrih_cdnPt_V4 : - case Hexagon::POST_LDrih_cdnNotPt_V4 : - case Hexagon::LDrib_cdnPt : - case Hexagon::LDrib_cdnNotPt : - case Hexagon::LDrib_indexed_cdnPt : - case Hexagon::LDrib_indexed_cdnNotPt : - case Hexagon::POST_LDrib_cdnPt_V4 : - case Hexagon::POST_LDrib_cdnNotPt_V4 : - case Hexagon::LDriuh_cdnPt : - case Hexagon::LDriuh_cdnNotPt : - case Hexagon::LDriuh_indexed_cdnPt : - case Hexagon::LDriuh_indexed_cdnNotPt : - case Hexagon::POST_LDriuh_cdnPt_V4 : - case Hexagon::POST_LDriuh_cdnNotPt_V4 : - case Hexagon::LDriub_cdnPt : - case Hexagon::LDriub_cdnNotPt : - case Hexagon::LDriub_indexed_cdnPt : - case Hexagon::LDriub_indexed_cdnNotPt : - case Hexagon::POST_LDriub_cdnPt_V4 : - case Hexagon::POST_LDriub_cdnNotPt_V4 : - - case Hexagon::LDrid_indexed_shl_cdnPt_V4 : - case Hexagon::LDrid_indexed_shl_cdnNotPt_V4 : - case Hexagon::LDrib_indexed_shl_cdnPt_V4 : - case Hexagon::LDrib_indexed_shl_cdnNotPt_V4 : - case Hexagon::LDriub_indexed_shl_cdnPt_V4 : - case Hexagon::LDriub_indexed_shl_cdnNotPt_V4 : - case Hexagon::LDrih_indexed_shl_cdnPt_V4 : - case Hexagon::LDrih_indexed_shl_cdnNotPt_V4 : - case Hexagon::LDriuh_indexed_shl_cdnPt_V4 : - case Hexagon::LDriuh_indexed_shl_cdnNotPt_V4 : - case Hexagon::LDriw_indexed_shl_cdnPt_V4 : - case Hexagon::LDriw_indexed_shl_cdnNotPt_V4 : - -// Coditional add - case Hexagon::ADD_ri_cdnPt: - case Hexagon::ADD_ri_cdnNotPt: - case Hexagon::ADD_rr_cdnPt: - case Hexagon::ADD_rr_cdnNotPt: - - // Conditional logical operations - case Hexagon::XOR_rr_cdnPt : - case Hexagon::XOR_rr_cdnNotPt : - case Hexagon::AND_rr_cdnPt : - case Hexagon::AND_rr_cdnNotPt : - case Hexagon::OR_rr_cdnPt : - case Hexagon::OR_rr_cdnNotPt : - - // Conditonal subtract - case Hexagon::SUB_rr_cdnPt : - case Hexagon::SUB_rr_cdnNotPt : - - // Conditional combine - case Hexagon::COMBINE_rr_cdnPt : - case Hexagon::COMBINE_rr_cdnNotPt : - - // Conditional shift operations - case Hexagon::ASLH_cdnPt_V4: - case Hexagon::ASLH_cdnNotPt_V4: - case Hexagon::ASRH_cdnPt_V4: - case Hexagon::ASRH_cdnNotPt_V4: - case Hexagon::SXTB_cdnPt_V4: - case Hexagon::SXTB_cdnNotPt_V4: - case Hexagon::SXTH_cdnPt_V4: - case Hexagon::SXTH_cdnNotPt_V4: - case Hexagon::ZXTB_cdnPt_V4: - case Hexagon::ZXTB_cdnNotPt_V4: - case Hexagon::ZXTH_cdnPt_V4: - case Hexagon::ZXTH_cdnNotPt_V4: - - // Conditional stores - case Hexagon::STrib_imm_cdnPt_V4 : - case Hexagon::STrib_imm_cdnNotPt_V4 : - case Hexagon::STrib_cdnPt_V4 : - case Hexagon::STrib_cdnNotPt_V4 : - case Hexagon::STrib_indexed_cdnPt_V4 : - case Hexagon::STrib_indexed_cdnNotPt_V4 : - case Hexagon::POST_STbri_cdnPt_V4 : - case Hexagon::POST_STbri_cdnNotPt_V4 : - case Hexagon::STrib_indexed_shl_cdnPt_V4 : - case Hexagon::STrib_indexed_shl_cdnNotPt_V4 : - - // Store doubleword conditionally - case Hexagon::STrid_indexed_cdnPt_V4 : - case Hexagon::STrid_indexed_cdnNotPt_V4 : - case Hexagon::STrid_indexed_shl_cdnPt_V4 : - case Hexagon::STrid_indexed_shl_cdnNotPt_V4 : - case Hexagon::POST_STdri_cdnPt_V4 : - case Hexagon::POST_STdri_cdnNotPt_V4 : - - // Store halfword conditionally - case Hexagon::STrih_cdnPt_V4 : - case Hexagon::STrih_cdnNotPt_V4 : - case Hexagon::STrih_indexed_cdnPt_V4 : - case Hexagon::STrih_indexed_cdnNotPt_V4 : - case Hexagon::STrih_imm_cdnPt_V4 : - case Hexagon::STrih_imm_cdnNotPt_V4 : - case Hexagon::STrih_indexed_shl_cdnPt_V4 : - case Hexagon::STrih_indexed_shl_cdnNotPt_V4 : - case Hexagon::POST_SThri_cdnPt_V4 : - case Hexagon::POST_SThri_cdnNotPt_V4 : - - // Store word conditionally - case Hexagon::STriw_cdnPt_V4 : - case Hexagon::STriw_cdnNotPt_V4 : - case Hexagon::STriw_indexed_cdnPt_V4 : - case Hexagon::STriw_indexed_cdnNotPt_V4 : - case Hexagon::STriw_imm_cdnPt_V4 : - case Hexagon::STriw_imm_cdnNotPt_V4 : - case Hexagon::STriw_indexed_shl_cdnPt_V4 : - case Hexagon::STriw_indexed_shl_cdnNotPt_V4 : - case Hexagon::POST_STwri_cdnPt_V4 : - case Hexagon::POST_STwri_cdnNotPt_V4 : - - case Hexagon::LDd_GP_cdnPt_V4: - case Hexagon::LDd_GP_cdnNotPt_V4: - case Hexagon::LDb_GP_cdnPt_V4: - case Hexagon::LDb_GP_cdnNotPt_V4: - case Hexagon::LDub_GP_cdnPt_V4: - case Hexagon::LDub_GP_cdnNotPt_V4: - case Hexagon::LDh_GP_cdnPt_V4: - case Hexagon::LDh_GP_cdnNotPt_V4: - case Hexagon::LDuh_GP_cdnPt_V4: - case Hexagon::LDuh_GP_cdnNotPt_V4: - case Hexagon::LDw_GP_cdnPt_V4: - case Hexagon::LDw_GP_cdnNotPt_V4: - - case Hexagon::STd_GP_cdnPt_V4: - case Hexagon::STd_GP_cdnNotPt_V4: - case Hexagon::STb_GP_cdnPt_V4: - case Hexagon::STb_GP_cdnNotPt_V4: - case Hexagon::STh_GP_cdnPt_V4: - case Hexagon::STh_GP_cdnNotPt_V4: - case Hexagon::STw_GP_cdnPt_V4: - case Hexagon::STw_GP_cdnNotPt_V4: - return true; - } - return false; + return PK_False; } static MachineOperand& GetPostIncrementOperand(MachineInstr *MI, @@ -2386,10 +537,10 @@ static MachineOperand& GetStoreValueOperand(MachineInstr *MI) { // Arch Spec: 3.4.4.2 bool HexagonPacketizerList::CanPromoteToNewValueStore( MachineInstr *MI, MachineInstr *PacketMI, unsigned DepReg, - std::map <MachineInstr*, SUnit*> MIToSUnit) -{ - // Make sure we are looking at the store - if (!IsNewifyStore(MI)) + std::map <MachineInstr*, SUnit*> MIToSUnit) { + const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII; + // Make sure we are looking at the store, that can be promoted. + if (!QII->mayBeNewStore(MI)) return false; // Make sure there is dependency and can be new value'ed @@ -2397,12 +548,11 @@ bool HexagonPacketizerList::CanPromoteToNewValueStore( MachineInstr *MI, GetStoreValueOperand(MI).getReg() != DepReg) return false; - const HexagonRegisterInfo* QRI = + const HexagonRegisterInfo* QRI = (const HexagonRegisterInfo *) TM.getRegisterInfo(); const MCInstrDesc& MCID = PacketMI->getDesc(); // first operand is always the result - const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII; const TargetRegisterClass* PacketRC = QII->getRegClass(MCID, 0, QRI, MF); // if there is already an store in the packet, no can do new value store @@ -2445,7 +595,7 @@ bool HexagonPacketizerList::CanPromoteToNewValueStore( MachineInstr *MI, } // If the source that feeds the store is predicated, new value store must - // also be also predicated. + // also be predicated. if (QII->isPredicated(PacketMI)) { if (!QII->isPredicated(MI)) return false; @@ -2490,8 +640,8 @@ bool HexagonPacketizerList::CanPromoteToNewValueStore( MachineInstr *MI, // sense, i.e, either both should be negated or both should be none negated. if (( predRegNumDst != predRegNumSrc) || - isDotNewInst(PacketMI) != isDotNewInst(MI) || - GetPredicateSense(MI, QII) != GetPredicateSense(PacketMI, QII)) { + QII->isDotNewInst(PacketMI) != QII->isDotNewInst(MI) || + getPredicateSense(MI, QII) != getPredicateSense(PacketMI, QII)) { return false; } } @@ -2572,10 +722,11 @@ bool HexagonPacketizerList::CanPromoteToNewValue( MachineInstr *MI, MachineBasicBlock::iterator &MII) { + const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII; const HexagonRegisterInfo* QRI = (const HexagonRegisterInfo *) TM.getRegisterInfo(); if (!QRI->Subtarget.hasV4TOps() || - !IsNewifyStore(MI)) + !QII->mayBeNewStore(MI)) return false; MachineInstr *PacketMI = PacketSU->getInstr(); @@ -2600,8 +751,9 @@ bool HexagonPacketizerList::CanPromoteToDotNew( MachineInstr *MI, MachineBasicBlock::iterator &MII, const TargetRegisterClass* RC ) { - // already a dot new instruction - if (isDotNewInst(MI) && !IsNewifyStore(MI)) + const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII; + // Already a dot new instruction. + if (QII->isDotNewInst(MI) && !QII->mayBeNewStore(MI)) return false; if (!isNewifiable(MI)) @@ -2611,13 +763,12 @@ bool HexagonPacketizerList::CanPromoteToDotNew( MachineInstr *MI, if (RC == &Hexagon::PredRegsRegClass && isCondInst(MI)) return true; else if (RC != &Hexagon::PredRegsRegClass && - !IsNewifyStore(MI)) // MI is not a new-value store + !QII->mayBeNewStore(MI)) // MI is not a new-value store return false; else { // Create a dot new machine instruction to see if resources can be // allocated. If not, bail out now. - const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII; - int NewOpcode = GetDotNewOp(MI->getOpcode()); + int NewOpcode = QII->GetDotNewOp(MI); const MCInstrDesc &desc = QII->get(NewOpcode); DebugLoc dl; MachineInstr *NewMI = @@ -2686,16 +837,39 @@ bool HexagonPacketizerList::RestrictingDepExistInPacket (MachineInstr* MI, } +/// Gets the predicate register of a predicated instruction. +static unsigned getPredicatedRegister(MachineInstr *MI, + const HexagonInstrInfo *QII) { + /// We use the following rule: The first predicate register that is a use is + /// the predicate register of a predicated instruction. + + assert(QII->isPredicated(MI) && "Must be predicated instruction"); + + for (MachineInstr::mop_iterator OI = MI->operands_begin(), + OE = MI->operands_end(); OI != OE; ++OI) { + MachineOperand &Op = *OI; + if (Op.isReg() && Op.getReg() && Op.isUse() && + Hexagon::PredRegsRegClass.contains(Op.getReg())) + return Op.getReg(); + } + + llvm_unreachable("Unknown instruction operand layout"); + + return 0; +} + // Given two predicated instructions, this function detects whether // the predicates are complements bool HexagonPacketizerList::ArePredicatesComplements (MachineInstr* MI1, MachineInstr* MI2, std::map <MachineInstr*, SUnit*> MIToSUnit) { const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII; - // Currently can only reason about conditional transfers - if (!QII->isConditionalTransfer(MI1) || !QII->isConditionalTransfer(MI2)) { + + // If we don't know the predicate sense of the instructions bail out early, we + // need it later. + if (getPredicateSense(MI1, QII) == PK_Unknown || + getPredicateSense(MI2, QII) == PK_Unknown) return false; - } // Scheduling unit for candidate SUnit* SU = MIToSUnit[MI1]; @@ -2734,9 +908,9 @@ bool HexagonPacketizerList::ArePredicatesComplements (MachineInstr* MI1, // there already exist anti dep on the same pred in // the packet. if (PacketSU->Succs[i].getSUnit() == SU && + PacketSU->Succs[i].getKind() == SDep::Data && Hexagon::PredRegsRegClass.contains( PacketSU->Succs[i].getReg()) && - PacketSU->Succs[i].getKind() == SDep::Data && // Here I know that *VIN is predicate setting instruction // with true data dep to candidate on the register // we care about - c) in the above example. @@ -2757,9 +931,13 @@ bool HexagonPacketizerList::ArePredicatesComplements (MachineInstr* MI1, // that the predicate sense is different // We also need to differentiate .old vs. .new: // !p0 is not complimentary to p0.new - return ((MI1->getOperand(1).getReg() == MI2->getOperand(1).getReg()) && - (GetPredicateSense(MI1, QII) != GetPredicateSense(MI2, QII)) && - (isDotNewInst(MI1) == isDotNewInst(MI2))); + unsigned PReg1 = getPredicatedRegister(MI1, QII); + unsigned PReg2 = getPredicatedRegister(MI2, QII); + return ((PReg1 == PReg2) && + Hexagon::PredRegsRegClass.contains(PReg1) && + Hexagon::PredRegsRegClass.contains(PReg2) && + (getPredicateSense(MI1, QII) != getPredicateSense(MI2, QII)) && + (QII->isDotNewInst(MI1) == QII->isDotNewInst(MI2))); } // initPacketizerState - Initialize packetizer flags @@ -2856,24 +1034,21 @@ bool HexagonPacketizerList::isLegalToPacketizeTogether(SUnit *SUI, SUnit *SUJ) { } // A LoopN instruction cannot appear in the same packet as a jump or call. - if (IsLoopN(I) && ( IsDirectJump(J) - || MCIDJ.isCall() - || QII->isDeallocRet(J))) { + if (IsLoopN(I) && + (IsDirectJump(J) || MCIDJ.isCall() || QII->isDeallocRet(J))) { Dependence = true; return false; } - if (IsLoopN(J) && ( IsDirectJump(I) - || MCIDI.isCall() - || QII->isDeallocRet(I))) { + if (IsLoopN(J) && + (IsDirectJump(I) || MCIDI.isCall() || QII->isDeallocRet(I))) { Dependence = true; return false; } // dealloc_return cannot appear in the same packet as a conditional or // unconditional jump. - if (QII->isDeallocRet(I) && ( MCIDJ.isBranch() - || MCIDJ.isCall() - || MCIDJ.isBarrier())) { + if (QII->isDeallocRet(I) && + (MCIDJ.isBranch() || MCIDJ.isCall() || MCIDJ.isBarrier())) { Dependence = true; return false; } @@ -2898,7 +1073,7 @@ bool HexagonPacketizerList::isLegalToPacketizeTogether(SUnit *SUI, SUnit *SUJ) { } //if dealloc_return - if (MCIDJ.mayStore() && QII->isDeallocRet(I)){ + if (MCIDJ.mayStore() && QII->isDeallocRet(I)) { Dependence = true; return false; } @@ -2906,9 +1081,8 @@ bool HexagonPacketizerList::isLegalToPacketizeTogether(SUnit *SUI, SUnit *SUJ) { // If an instruction feeds new value jump, glue it. MachineBasicBlock::iterator NextMII = I; ++NextMII; - MachineInstr *NextMI = NextMII; - - if (QII->isNewValueJump(NextMI)) { + if (NextMII != I->getParent()->end() && QII->isNewValueJump(NextMII)) { + MachineInstr *NextMI = NextMII; bool secondRegMatch = false; bool maintainNewValueJump = false; diff --git a/lib/Target/Hexagon/InstPrinter/HexagonInstPrinter.cpp b/lib/Target/Hexagon/InstPrinter/HexagonInstPrinter.cpp index 36da6df..2ea0d2e 100644 --- a/lib/Target/Hexagon/InstPrinter/HexagonInstPrinter.cpp +++ b/lib/Target/Hexagon/InstPrinter/HexagonInstPrinter.cpp @@ -179,7 +179,7 @@ void HexagonInstPrinter::printBranchOperand(const MCInst *MI, unsigned OpNo, raw_ostream &O) const { // Branches can take an immediate operand. This is used by the branch // selection pass to print $+8, an eight byte displacement from the PC. - assert("Unknown branch operand."); + llvm_unreachable("Unknown branch operand."); } void HexagonInstPrinter::printCallOperand(const MCInst *MI, unsigned OpNo, @@ -196,15 +196,9 @@ void HexagonInstPrinter::printPredicateOperand(const MCInst *MI, unsigned OpNo, void HexagonInstPrinter::printSymbol(const MCInst *MI, unsigned OpNo, raw_ostream &O, bool hi) const { - const MCOperand& MO = MI->getOperand(OpNo); + assert(MI->getOperand(OpNo).isImm() && "Unknown symbol operand"); - O << '#' << (hi? "HI": "LO") << '('; - if (MO.isImm()) { - O << '#'; - printOperand(MI, OpNo, O); - } else { - assert("Unknown symbol operand"); - printOperand(MI, OpNo, O); - } + O << '#' << (hi ? "HI" : "LO") << "(#"; + printOperand(MI, OpNo, O); O << ')'; } diff --git a/lib/Target/Hexagon/MCTargetDesc/HexagonBaseInfo.h b/lib/Target/Hexagon/MCTargetDesc/HexagonBaseInfo.h index d4a93b5..e0f5a27 100644 --- a/lib/Target/Hexagon/MCTargetDesc/HexagonBaseInfo.h +++ b/lib/Target/Hexagon/MCTargetDesc/HexagonBaseInfo.h @@ -65,7 +65,8 @@ namespace HexagonII { AbsoluteSet = 2, // Absolute set addressing mode BaseImmOffset = 3, // Indirect with offset BaseLongOffset = 4, // Indirect with long offset - BaseRegOffset = 5 // Indirect with register offset + BaseRegOffset = 5, // Indirect with register offset + PostInc = 6 // Post increment addressing mode }; enum MemAccessSize { diff --git a/lib/Target/Hexagon/MCTargetDesc/HexagonMCAsmInfo.cpp b/lib/Target/Hexagon/MCTargetDesc/HexagonMCAsmInfo.cpp index 86f75d1..495dbb9 100644 --- a/lib/Target/Hexagon/MCTargetDesc/HexagonMCAsmInfo.cpp +++ b/lib/Target/Hexagon/MCTargetDesc/HexagonMCAsmInfo.cpp @@ -15,7 +15,7 @@ using namespace llvm; -HexagonMCAsmInfo::HexagonMCAsmInfo(const Target &T, StringRef TT) { +HexagonMCAsmInfo::HexagonMCAsmInfo(StringRef TT) { Data16bitsDirective = "\t.half\t"; Data32bitsDirective = "\t.word\t"; Data64bitsDirective = 0; // .xword is only supported by V9. @@ -31,6 +31,7 @@ HexagonMCAsmInfo::HexagonMCAsmInfo(const Target &T, StringRef TT) { AscizDirective = "\t.string\t"; WeakRefDirective = "\t.weak\t"; + SupportsDebugInformation = true; UsesELFSectionDirectiveForBSS = true; ExceptionsType = ExceptionHandling::DwarfCFI; } diff --git a/lib/Target/Hexagon/MCTargetDesc/HexagonMCAsmInfo.h b/lib/Target/Hexagon/MCTargetDesc/HexagonMCAsmInfo.h index d336cd5..0b94d21 100644 --- a/lib/Target/Hexagon/MCTargetDesc/HexagonMCAsmInfo.h +++ b/lib/Target/Hexagon/MCTargetDesc/HexagonMCAsmInfo.h @@ -18,11 +18,9 @@ #include "llvm/MC/MCAsmInfo.h" namespace llvm { - class Target; - class HexagonMCAsmInfo : public MCAsmInfo { public: - explicit HexagonMCAsmInfo(const Target &T, StringRef TT); + explicit HexagonMCAsmInfo(StringRef TT); }; } // namespace llvm diff --git a/lib/Target/Hexagon/MCTargetDesc/HexagonMCTargetDesc.cpp b/lib/Target/Hexagon/MCTargetDesc/HexagonMCTargetDesc.cpp index 6b1d2d1..2f93a52 100644 --- a/lib/Target/Hexagon/MCTargetDesc/HexagonMCTargetDesc.cpp +++ b/lib/Target/Hexagon/MCTargetDesc/HexagonMCTargetDesc.cpp @@ -54,13 +54,14 @@ static MCSubtargetInfo *createHexagonMCSubtargetInfo(StringRef TT, return X; } -static MCAsmInfo *createHexagonMCAsmInfo(const Target &T, StringRef TT) { - MCAsmInfo *MAI = new HexagonMCAsmInfo(T, TT); +static MCAsmInfo *createHexagonMCAsmInfo(const MCRegisterInfo &MRI, + StringRef TT) { + MCAsmInfo *MAI = new HexagonMCAsmInfo(TT); // VirtualFP = (R30 + #0). - MachineLocation Dst(MachineLocation::VirtualFP); - MachineLocation Src(Hexagon::R30, 0); - MAI->addInitialFrameState(0, Dst, Src); + MCCFIInstruction Inst = MCCFIInstruction::createDefCfa( + 0, Hexagon::R30, 0); + MAI->addInitialFrameState(Inst); return MAI; } |