diff options
Diffstat (limited to 'lib/Target')
68 files changed, 4389 insertions, 333 deletions
diff --git a/lib/Target/ARM/ARM.h b/lib/Target/ARM/ARM.h index 9a8cab8ecc..0ac92f1ee8 100644 --- a/lib/Target/ARM/ARM.h +++ b/lib/Target/ARM/ARM.h @@ -40,6 +40,7 @@ FunctionPass *createARMJITCodeEmitterPass(ARMBaseTargetMachine &TM, FunctionPass *createARMLoadStoreOptimizationPass(bool PreAlloc = false); FunctionPass *createARMExpandPseudoPass(); +FunctionPass *createARMGlobalBaseRegPass(); FunctionPass *createARMGlobalMergePass(const TargetLowering* tli); FunctionPass *createARMConstantIslandPass(); FunctionPass *createMLxExpansionPass(); diff --git a/lib/Target/ARM/ARM.td b/lib/Target/ARM/ARM.td index 38509a3400..00bf1b85ec 100644 --- a/lib/Target/ARM/ARM.td +++ b/lib/Target/ARM/ARM.td @@ -32,9 +32,6 @@ def FeatureVFP2 : SubtargetFeature<"vfp2", "HasVFPv2", "true", def FeatureVFP3 : SubtargetFeature<"vfp3", "HasVFPv3", "true", "Enable VFP3 instructions", [FeatureVFP2]>; -def FeatureVFP4 : SubtargetFeature<"vfp4", "HasVFPv4", "true", - "Enable VFP4 instructions", - [FeatureVFP3]>; def FeatureNEON : SubtargetFeature<"neon", "HasNEON", "true", "Enable NEON instructions", [FeatureVFP3]>; @@ -44,10 +41,16 @@ def FeatureNoARM : SubtargetFeature<"noarm", "NoARM", "true", "Does not support ARM mode execution">; def FeatureFP16 : SubtargetFeature<"fp16", "HasFP16", "true", "Enable half-precision floating point">; +def FeatureVFP4 : SubtargetFeature<"vfp4", "HasVFPv4", "true", + "Enable VFP4 instructions", + [FeatureVFP3, FeatureFP16]>; def FeatureD16 : SubtargetFeature<"d16", "HasD16", "true", "Restrict VFP3 to 16 double registers">; def FeatureHWDiv : SubtargetFeature<"hwdiv", "HasHardwareDivide", "true", "Enable divide instructions">; +def FeatureHWDivARM : SubtargetFeature<"hwdiv-arm", + "HasHardwareDivideInARM", "true", + "Enable divide instructions in ARM mode">; def FeatureT2XtPk : SubtargetFeature<"t2xtpk", "HasT2ExtractPack", "true", "Enable Thumb2 extract and pack instructions">; def FeatureDB : SubtargetFeature<"db", "HasDataBarrier", "true", @@ -139,6 +142,13 @@ def ProcA9 : SubtargetFeature<"a9", "ARMProcFamily", "CortexA9", [FeatureVMLxForwarding, FeatureT2XtPk, FeatureFP16, FeatureAvoidPartialCPSR]>; +def ProcSwift : SubtargetFeature<"swift", "ARMProcFamily", "Swift", + "Swift ARM processors", + [FeatureNEONForFP, FeatureT2XtPk, + FeatureVFP4, FeatureMP, FeatureHWDiv, + FeatureHWDivARM, FeatureAvoidPartialCPSR, + FeatureHasSlowFPVMLx]>; + // FIXME: It has not been determined if A15 has these features. def ProcA15 : SubtargetFeature<"a15", "ARMProcFamily", "CortexA15", "Cortex-A15 ARM processors", @@ -240,6 +250,12 @@ def : ProcNoItin<"cortex-m4", [HasV7Ops, FeatureT2XtPk, FeatureVFP4, FeatureVFPOnlySP, FeatureMClass]>; +// Swift uArch Processors. +def : ProcessorModel<"swift", SwiftModel, + [ProcSwift, HasV7Ops, FeatureNEON, + FeatureDB, FeatureDSPThumb2, + FeatureHasRAS]>; + //===----------------------------------------------------------------------===// // Register File Description //===----------------------------------------------------------------------===// diff --git a/lib/Target/ARM/ARMBaseInstrInfo.cpp b/lib/Target/ARM/ARMBaseInstrInfo.cpp index c08294918e..42b6bc3cdc 100644 --- a/lib/Target/ARM/ARMBaseInstrInfo.cpp +++ b/lib/Target/ARM/ARMBaseInstrInfo.cpp @@ -49,6 +49,11 @@ static cl::opt<bool> WidenVMOVS("widen-vmovs", cl::Hidden, cl::init(true), cl::desc("Widen ARM vmovs to vmovd when possible")); +static cl::opt<unsigned> +SwiftPartialUpdateClearance("swift-partial-update-clearance", + cl::Hidden, cl::init(12), + cl::desc("Clearance before partial register updates")); + /// ARM_MLxEntry - Record information about MLA / MLS instructions. struct ARM_MLxEntry { uint16_t MLxOpc; // MLA / MLS opcode @@ -1389,7 +1394,6 @@ bool ARMBaseInstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2, case ARM::VLDRD: case ARM::VLDRS: case ARM::t2LDRi8: - case ARM::t2LDRDi8: case ARM::t2LDRSHi8: case ARM::t2LDRi12: case ARM::t2LDRSHi12: @@ -1528,6 +1532,14 @@ isProfitableToIfCvt(MachineBasicBlock &TMBB, return (TCycles + FCycles + TExtra + FExtra) <= UnpredCost; } +bool +ARMBaseInstrInfo::isProfitableToUnpredicate(MachineBasicBlock &TMBB, + MachineBasicBlock &FMBB) const { + // Reduce false anti-dependencies to let Swift's out-of-order execution + // engine do its thing. + return Subtarget.isSwift(); +} + /// getInstrPredicate - If instruction is predicated, returns its predicate /// condition, otherwise returns AL. It also returns the condition code /// register by reference. @@ -2344,6 +2356,229 @@ bool ARMBaseInstrInfo::FoldImmediate(MachineInstr *UseMI, return true; } +static unsigned getNumMicroOpsSwiftLdSt(const InstrItineraryData *ItinData, + const MachineInstr *MI) { + switch (MI->getOpcode()) { + default: { + const MCInstrDesc &Desc = MI->getDesc(); + int UOps = ItinData->getNumMicroOps(Desc.getSchedClass()); + assert(UOps >= 0 && "bad # UOps"); + return UOps; + } + + case ARM::LDRrs: + case ARM::LDRBrs: + case ARM::STRrs: + case ARM::STRBrs: { + unsigned ShOpVal = MI->getOperand(3).getImm(); + bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub; + unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal); + if (!isSub && + (ShImm == 0 || + ((ShImm == 1 || ShImm == 2 || ShImm == 3) && + ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))) + return 1; + return 2; + } + + case ARM::LDRH: + case ARM::STRH: { + if (!MI->getOperand(2).getReg()) + return 1; + + unsigned ShOpVal = MI->getOperand(3).getImm(); + bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub; + unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal); + if (!isSub && + (ShImm == 0 || + ((ShImm == 1 || ShImm == 2 || ShImm == 3) && + ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))) + return 1; + return 2; + } + + case ARM::LDRSB: + case ARM::LDRSH: + return (ARM_AM::getAM3Op(MI->getOperand(3).getImm()) == ARM_AM::sub) ? 3:2; + + case ARM::LDRSB_POST: + case ARM::LDRSH_POST: { + unsigned Rt = MI->getOperand(0).getReg(); + unsigned Rm = MI->getOperand(3).getReg(); + return (Rt == Rm) ? 4 : 3; + } + + case ARM::LDR_PRE_REG: + case ARM::LDRB_PRE_REG: { + unsigned Rt = MI->getOperand(0).getReg(); + unsigned Rm = MI->getOperand(3).getReg(); + if (Rt == Rm) + return 3; + unsigned ShOpVal = MI->getOperand(4).getImm(); + bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub; + unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal); + if (!isSub && + (ShImm == 0 || + ((ShImm == 1 || ShImm == 2 || ShImm == 3) && + ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))) + return 2; + return 3; + } + + case ARM::STR_PRE_REG: + case ARM::STRB_PRE_REG: { + unsigned ShOpVal = MI->getOperand(4).getImm(); + bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub; + unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal); + if (!isSub && + (ShImm == 0 || + ((ShImm == 1 || ShImm == 2 || ShImm == 3) && + ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))) + return 2; + return 3; + } + + case ARM::LDRH_PRE: + case ARM::STRH_PRE: { + unsigned Rt = MI->getOperand(0).getReg(); + unsigned Rm = MI->getOperand(3).getReg(); + if (!Rm) + return 2; + if (Rt == Rm) + return 3; + return (ARM_AM::getAM3Op(MI->getOperand(4).getImm()) == ARM_AM::sub) + ? 3 : 2; + } + + case ARM::LDR_POST_REG: + case ARM::LDRB_POST_REG: + case ARM::LDRH_POST: { + unsigned Rt = MI->getOperand(0).getReg(); + unsigned Rm = MI->getOperand(3).getReg(); + return (Rt == Rm) ? 3 : 2; + } + + case ARM::LDR_PRE_IMM: + case ARM::LDRB_PRE_IMM: + case ARM::LDR_POST_IMM: + case ARM::LDRB_POST_IMM: + case ARM::STRB_POST_IMM: + case ARM::STRB_POST_REG: + case ARM::STRB_PRE_IMM: + case ARM::STRH_POST: + case ARM::STR_POST_IMM: + case ARM::STR_POST_REG: + case ARM::STR_PRE_IMM: + return 2; + + case ARM::LDRSB_PRE: + case ARM::LDRSH_PRE: { + unsigned Rm = MI->getOperand(3).getReg(); + if (Rm == 0) + return 3; + unsigned Rt = MI->getOperand(0).getReg(); + if (Rt == Rm) + return 4; + unsigned ShOpVal = MI->getOperand(4).getImm(); + bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub; + unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal); + if (!isSub && + (ShImm == 0 || + ((ShImm == 1 || ShImm == 2 || ShImm == 3) && + ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))) + return 3; + return 4; + } + + case ARM::LDRD: { + unsigned Rt = MI->getOperand(0).getReg(); + unsigned Rn = MI->getOperand(2).getReg(); + unsigned Rm = MI->getOperand(3).getReg(); + if (Rm) + return (ARM_AM::getAM3Op(MI->getOperand(4).getImm()) == ARM_AM::sub) ?4:3; + return (Rt == Rn) ? 3 : 2; + } + + case ARM::STRD: { + unsigned Rm = MI->getOperand(3).getReg(); + if (Rm) + return (ARM_AM::getAM3Op(MI->getOperand(4).getImm()) == ARM_AM::sub) ?4:3; + return 2; + } + + case ARM::LDRD_POST: + case ARM::t2LDRD_POST: + return 3; + + case ARM::STRD_POST: + case ARM::t2STRD_POST: + return 4; + + case ARM::LDRD_PRE: { + unsigned Rt = MI->getOperand(0).getReg(); + unsigned Rn = MI->getOperand(3).getReg(); + unsigned Rm = MI->getOperand(4).getReg(); + if (Rm) + return (ARM_AM::getAM3Op(MI->getOperand(5).getImm()) == ARM_AM::sub) ?5:4; + return (Rt == Rn) ? 4 : 3; + } + + case ARM::t2LDRD_PRE: { + unsigned Rt = MI->getOperand(0).getReg(); + unsigned Rn = MI->getOperand(3).getReg(); + return (Rt == Rn) ? 4 : 3; + } + + case ARM::STRD_PRE: { + unsigned Rm = MI->getOperand(4).getReg(); + if (Rm) + return (ARM_AM::getAM3Op(MI->getOperand(5).getImm()) == ARM_AM::sub) ?5:4; + return 3; + } + + case ARM::t2STRD_PRE: + return 3; + + case ARM::t2LDR_POST: + case ARM::t2LDRB_POST: + case ARM::t2LDRB_PRE: + case ARM::t2LDRSBi12: + case ARM::t2LDRSBi8: + case ARM::t2LDRSBpci: + case ARM::t2LDRSBs: + case ARM::t2LDRH_POST: + case ARM::t2LDRH_PRE: + case ARM::t2LDRSBT: + case ARM::t2LDRSB_POST: + case ARM::t2LDRSB_PRE: + case ARM::t2LDRSH_POST: + case ARM::t2LDRSH_PRE: + case ARM::t2LDRSHi12: + case ARM::t2LDRSHi8: + case ARM::t2LDRSHpci: + case ARM::t2LDRSHs: + return 2; + + case ARM::t2LDRDi8: { + unsigned Rt = MI->getOperand(0).getReg(); + unsigned Rn = MI->getOperand(2).getReg(); + return (Rt == Rn) ? 3 : 2; + } + + case ARM::t2STRB_POST: + case ARM::t2STRB_PRE: + case ARM::t2STRBs: + case ARM::t2STRDi8: + case ARM::t2STRH_POST: + case ARM::t2STRH_PRE: + case ARM::t2STRHs: + case ARM::t2STR_POST: + case ARM::t2STR_PRE: + case ARM::t2STRs: + return 2; + } +} + // Return the number of 32-bit words loaded by LDM or stored by STM. If this // can't be easily determined return 0 (missing MachineMemOperand). // @@ -2384,8 +2619,12 @@ ARMBaseInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData, const MCInstrDesc &Desc = MI->getDesc(); unsigned Class = Desc.getSchedClass(); int ItinUOps = ItinData->getNumMicroOps(Class); - if (ItinUOps >= 0) + if (ItinUOps >= 0) { + if (Subtarget.isSwift() && (Desc.mayLoad() || Desc.mayStore())) + return getNumMicroOpsSwiftLdSt(ItinData, MI); + return ItinUOps; + } unsigned Opc = MI->getOpcode(); switch (Opc) { @@ -2454,7 +2693,43 @@ ARMBaseInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData, case ARM::t2STMIA_UPD: case ARM::t2STMDB_UPD: { unsigned NumRegs = MI->getNumOperands() - Desc.getNumOperands() + 1; - if (Subtarget.isCortexA8()) { + if (Subtarget.isSwift()) { + // rdar://8402126 + int UOps = 1 + NumRegs; // One for address computation, one for each ld / st. + switch (Opc) { + default: break; + case ARM::VLDMDIA_UPD: + case ARM::VLDMDDB_UPD: + case ARM::VLDMSIA_UPD: + case ARM::VLDMSDB_UPD: + case ARM::VSTMDIA_UPD: + case ARM::VSTMDDB_UPD: + case ARM::VSTMSIA_UPD: + case ARM::VSTMSDB_UPD: + case ARM::LDMIA_UPD: + case ARM::LDMDA_UPD: + case ARM::LDMDB_UPD: + case ARM::LDMIB_UPD: + case ARM::STMIA_UPD: + case ARM::STMDA_UPD: + case ARM::STMDB_UPD: + case ARM::STMIB_UPD: + case ARM::tLDMIA_UPD: + case ARM::tSTMIA_UPD: + case ARM::t2LDMIA_UPD: + case ARM::t2LDMDB_UPD: + case ARM::t2STMIA_UPD: + case ARM::t2STMDB_UPD: + ++UOps; // One for base register writeback. + break; + case ARM::LDMIA_RET: + case ARM::tPOP_RET: + case ARM::t2LDMIA_RET: + UOps += 2; // One for base reg wb, one for write to pc. + break; + } + return UOps; + } else if (Subtarget.isCortexA8()) { if (NumRegs < 4) return 2; // 4 registers would be issued: 2, 2. @@ -2463,7 +2738,7 @@ ARMBaseInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData, if (NumRegs % 2) ++A8UOps; return A8UOps; - } else if (Subtarget.isLikeA9()) { + } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) { int A9UOps = (NumRegs / 2); // If there are odd number of registers or if it's not 64-bit aligned, // then it takes an extra AGU (Address Generation Unit) cycle. @@ -2496,7 +2771,7 @@ ARMBaseInstrInfo::getVLDMDefCycle(const InstrItineraryData *ItinData, DefCycle = RegNo / 2 + 1; if (RegNo % 2) ++DefCycle; - } else if (Subtarget.isLikeA9()) { + } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) { DefCycle = RegNo; bool isSLoad = false; @@ -2540,7 +2815,7 @@ ARMBaseInstrInfo::getLDMDefCycle(const InstrItineraryData *ItinData, DefCycle = 1; // Result latency is issue cycle + 2: E2. DefCycle += 2; - } else if (Subtarget.isLikeA9()) { + } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) { DefCycle = (RegNo / 2); // If there are odd number of registers or if it's not 64-bit aligned, // then it takes an extra AGU (Address Generation Unit) cycle. @@ -2571,7 +2846,7 @@ ARMBaseInstrInfo::getVSTMUseCycle(const InstrItineraryData *ItinData, UseCycle = RegNo / 2 + 1; if (RegNo % 2) ++UseCycle; - } else if (Subtarget.isLikeA9()) { + } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) { UseCycle = RegNo; bool isSStore = false; @@ -2612,7 +2887,7 @@ ARMBaseInstrInfo::getSTMUseCycle(const InstrItineraryData *ItinData, UseCycle = 2; // Read in E3. UseCycle += 2; - } else if (Subtarget.isLikeA9()) { + } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) { UseCycle = (RegNo / 2); // If there are odd number of registers or if it's not 64-bit aligned, // then it takes an extra AGU (Address Generation Unit) cycle. @@ -2822,6 +3097,37 @@ static int adjustDefLatency(const ARMSubtarget &Subtarget, break; } } + } else if (Subtarget.isSwift()) { + // FIXME: Properly handle all of the latency adjustments for address + // writeback. + switch (DefMCID->getOpcode()) { + default: break; + case ARM::LDRrs: + case ARM::LDRBrs: { + unsigned ShOpVal = DefMI->getOperand(3).getImm(); + bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub; + unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal); + if (!isSub && + (ShImm == 0 || + ((ShImm == 1 || ShImm == 2 || ShImm == 3) && + ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))) + Adjust -= 2; + else if (!isSub && + ShImm == 1 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsr) + --Adjust; + break; + } + case ARM::t2LDRs: + case ARM::t2LDRBs: + case ARM::t2LDRHs: + case ARM::t2LDRSHs: { + // Thumb2 mode: lsl only. + unsigned ShAmt = DefMI->getOperand(3).getImm(); + if (ShAmt == 0 || ShAmt == 1 || ShAmt == 2 || ShAmt == 3) + Adjust -= 2; + break; + } + } } if (DefAlign < 8 && Subtarget.isLikeA9()) { @@ -2998,7 +3304,7 @@ ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData, // instructions). if (Latency > 0 && Subtarget.isThumb2()) { const MachineFunction *MF = DefMI->getParent()->getParent(); - if (MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize)) + if (MF->getFunction()->getFnAttributes().hasOptimizeForSizeAttr()) --Latency; } return Latency; @@ -3048,7 +3354,7 @@ ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData, if (!UseNode->isMachineOpcode()) { int Latency = ItinData->getOperandCycle(DefMCID.getSchedClass(), DefIdx); - if (Subtarget.isLikeA9()) + if (Subtarget.isLikeA9() || Subtarget.isSwift()) return Latency <= 2 ? 1 : Latency - 1; else return Latency <= 3 ? 1 : Latency - 2; @@ -3092,6 +3398,33 @@ ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData, break; } } + } else if (DefIdx == 0 && Latency > 2 && Subtarget.isSwift()) { + // FIXME: Properly handle all of the latency adjustments for address + // writeback. + switch (DefMCID.getOpcode()) { + default: break; + case ARM::LDRrs: + case ARM::LDRBrs: { + unsigned ShOpVal = + cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue(); + unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal); |