#include "llvm/CodeGen/MachineMemOperand.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/SelectionDAGNodes.h"
+#include "llvm/CodeGen/TargetSchedule.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/GlobalValue.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
+#include "llvm/Support/raw_ostream.h"
using namespace llvm;
unsigned ARMBaseInstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
- MachineBasicBlock::iterator I = MBB.end();
- if (I == MBB.begin()) return 0;
- --I;
- while (I->isDebugValue()) {
- if (I == MBB.begin())
- return 0;
- --I;
- }
+ MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr();
+ if (I == MBB.end())
+ return 0;
+
if (!isUncondBranchOpcode(I->getOpcode()) &&
!isCondBranchOpcode(I->getOpcode()))
return 0;
unsigned
ARMBaseInstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
MachineBasicBlock *FBB,
- const SmallVectorImpl<MachineOperand> &Cond,
+ ArrayRef<MachineOperand> Cond,
DebugLoc DL) const {
ARMFunctionInfo *AFI = MBB.getParent()->getInfo<ARMFunctionInfo>();
int BOpc = !AFI->isThumbFunction()
assert((Cond.size() == 2 || Cond.size() == 0) &&
"ARM branch conditions have two components!");
+ // For conditional branches, we use addOperand to preserve CPSR flags.
+
if (!FBB) {
if (Cond.empty()) { // Unconditional branch?
if (isThumb)
BuildMI(&MBB, DL, get(BOpc)).addMBB(TBB);
} else
BuildMI(&MBB, DL, get(BccOpc)).addMBB(TBB)
- .addImm(Cond[0].getImm()).addReg(Cond[1].getReg());
+ .addImm(Cond[0].getImm()).addOperand(Cond[1]);
return 1;
}
// Two-way conditional branch.
BuildMI(&MBB, DL, get(BccOpc)).addMBB(TBB)
- .addImm(Cond[0].getImm()).addReg(Cond[1].getReg());
+ .addImm(Cond[0].getImm()).addOperand(Cond[1]);
if (isThumb)
BuildMI(&MBB, DL, get(BOpc)).addMBB(FBB).addImm(ARMCC::AL).addReg(0);
else
}
bool ARMBaseInstrInfo::
-PredicateInstruction(MachineInstr *MI,
- const SmallVectorImpl<MachineOperand> &Pred) const {
+PredicateInstruction(MachineInstr *MI, ArrayRef<MachineOperand> Pred) const {
unsigned Opc = MI->getOpcode();
if (isUncondBranchOpcode(Opc)) {
MI->setDesc(get(getMatchingCondBranchOpcode(Opc)));
return false;
}
-bool ARMBaseInstrInfo::
-SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1,
- const SmallVectorImpl<MachineOperand> &Pred2) const {
+bool ARMBaseInstrInfo::SubsumesPredicate(ArrayRef<MachineOperand> Pred1,
+ ArrayRef<MachineOperand> Pred2) const {
if (Pred1.size() > 2 || Pred2.size() > 2)
return false;
case ARM::t2MOVi32imm:
return 8;
case ARM::CONSTPOOL_ENTRY:
+ case ARM::JUMPTABLE_INSTS:
+ case ARM::JUMPTABLE_ADDRS:
+ case ARM::JUMPTABLE_TBB:
+ case ARM::JUMPTABLE_TBH:
// If this machine instr is a constant pool entry, its size is recorded as
// operand #2.
return MI->getOperand(2).getImm();
case ARM::t2Int_eh_sjlj_setjmp:
case ARM::t2Int_eh_sjlj_setjmp_nofp:
return 12;
- case ARM::BR_JTr:
- case ARM::BR_JTm:
- case ARM::BR_JTadd:
- case ARM::tBR_JTr:
- case ARM::t2BR_JT:
- case ARM::t2TBB_JT:
- case ARM::t2TBH_JT: {
- // These are jumptable branches, i.e. a branch followed by an inlined
- // jumptable. The size is 4 + 4 * number of entries. For TBB, each
- // entry is one byte; TBH two byte each.
- unsigned EntrySize = (Opc == ARM::t2TBB_JT)
- ? 1 : ((Opc == ARM::t2TBH_JT) ? 2 : 4);
- unsigned NumOps = MCID.getNumOperands();
- MachineOperand JTOP =
- MI->getOperand(NumOps - (MI->isPredicable() ? 3 : 2));
- unsigned JTI = JTOP.getIndex();
- const MachineJumpTableInfo *MJTI = MF->getJumpTableInfo();
- assert(MJTI != nullptr);
- const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables();
- assert(JTI < JT.size());
- // Thumb instructions are 2 byte aligned, but JT entries are 4 byte
- // 4 aligned. The assembler / linker may add 2 byte padding just before
- // the JT entries. The size does not include this padding; the
- // constant islands pass does separate bookkeeping for it.
- // FIXME: If we know the size of the function is less than (1 << 16) *2
- // bytes, we can use 16-bit entries instead. Then there won't be an
- // alignment issue.
- unsigned InstSize = (Opc == ARM::tBR_JTr || Opc == ARM::t2BR_JT) ? 2 : 4;
- unsigned NumEntries = JT[JTI].MBBs.size();
- if (Opc == ARM::t2TBB_JT && (NumEntries & 1))
- // Make sure the instruction that follows TBB is 2-byte aligned.
- // FIXME: Constant island pass should insert an "ALIGN" instruction
- // instead.
- ++NumEntries;
- return NumEntries * EntrySize + InstSize;
- }
case ARM::SPACE:
return MI->getOperand(1).getImm();
}
Reloc::Model RM = MF.getTarget().getRelocationModel();
if (MI->getOpcode() == TargetOpcode::LOAD_STACK_GUARD) {
- assert(getSubtarget().getTargetTriple().getObjectFormat() ==
- Triple::MachO &&
+ assert(getSubtarget().getTargetTriple().isOSBinFormatMachO() &&
"LOAD_STACK_GUARD currently supported only for MachO.");
expandLoadStackGuard(MI, RM);
MI->getParent()->erase(MI);
bool ARMBaseInstrInfo::produceSameValue(const MachineInstr *MI0,
const MachineInstr *MI1,
const MachineRegisterInfo *MRI) const {
- int Opcode = MI0->getOpcode();
+ unsigned Opcode = MI0->getOpcode();
if (Opcode == ARM::t2LDRpci ||
Opcode == ARM::t2LDRpci_pic ||
Opcode == ARM::tLDRpci ||
if (!NumCycles)
return false;
+ // If we are optimizing for size, see if the branch in the predecessor can be
+ // lowered to cbn?z by the constant island lowering pass, and return false if
+ // so. This results in a shorter instruction sequence.
+ const Function *F = MBB.getParent()->getFunction();
+ if (F->hasFnAttribute(Attribute::OptimizeForSize) ||
+ F->hasFnAttribute(Attribute::MinSize)) {
+ MachineBasicBlock *Pred = *MBB.pred_begin();
+ if (!Pred->empty()) {
+ MachineInstr *LastMI = &*Pred->rbegin();
+ if (LastMI->getOpcode() == ARM::t2Bcc) {
+ MachineBasicBlock::iterator CmpMI = LastMI;
+ if (CmpMI != Pred->begin()) {
+ --CmpMI;
+ if (CmpMI->getOpcode() == ARM::tCMPi8 ||
+ CmpMI->getOpcode() == ARM::t2CMPri) {
+ unsigned Reg = CmpMI->getOperand(0).getReg();
+ unsigned PredReg = 0;
+ ARMCC::CondCodes P = getInstrPredicate(CmpMI, PredReg);
+ if (P == ARMCC::AL && CmpMI->getOperand(1).getImm() == 0 &&
+ isARMLowRegister(Reg))
+ return false;
+ }
+ }
+ }
+ }
+ }
+
// Attempt to estimate the relative costs of predication versus branching.
unsigned UnpredCost = Probability.getNumerator() * NumCycles;
UnpredCost /= Probability.getDenominator();
}
-int llvm::getMatchingCondBranchOpcode(int Opc) {
+unsigned llvm::getMatchingCondBranchOpcode(unsigned Opc) {
if (Opc == ARM::B)
return ARM::Bcc;
if (Opc == ARM::tB)
return nullptr;
}
bool DontMoveAcrossStores = true;
- if (!MI->isSafeToMove(TII, /* AliasAnalysis = */ nullptr,
- DontMoveAcrossStores))
+ if (!MI->isSafeToMove(/* AliasAnalysis = */ nullptr, DontMoveAcrossStores))
return nullptr;
return MI;
}
return false;
}
-MachineInstr *ARMBaseInstrInfo::optimizeSelect(MachineInstr *MI,
- bool PreferFalse) const {
+MachineInstr *
+ARMBaseInstrInfo::optimizeSelect(MachineInstr *MI,
+ SmallPtrSetImpl<MachineInstr *> &SeenMIs,
+ bool PreferFalse) const {
assert((MI->getOpcode() == ARM::MOVCCr || MI->getOpcode() == ARM::t2MOVCCr) &&
"Unknown select instruction");
MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
NewMI.addOperand(FalseReg);
NewMI->tieOperands(0, NewMI->getNumOperands() - 1);
+ // Update SeenMIs set: register newly created MI and erase removed DefMI.
+ SeenMIs.insert(NewMI);
+ SeenMIs.erase(DefMI);
+
+ // If MI is inside a loop, and DefMI is outside the loop, then kill flags on
+ // DefMI would be invalid when tranferred inside the loop. Checking for a
+ // loop is expensive, but at least remove kill flags if they are in different
+ // BBs.
+ if (DefMI->getParent() != MI->getParent())
+ NewMI->clearKillInfo();
+
// The caller will erase MI, but not DefMI.
DefMI->eraseFromParent();
return NewMI;
unsigned NumBytes) {
// This optimisation potentially adds lots of load and store
// micro-operations, it's only really a great benefit to code-size.
- if (!MF.getFunction()->getAttributes().hasAttribute(
- AttributeSet::FunctionIndex, Attribute::MinSize))
+ if (!MF.getFunction()->hasFnAttribute(Attribute::MinSize))
return false;
// If only one register is pushed/popped, LLVM can use an LDR/STR
if (SrcReg == MI->getOperand(CommonUse ? 1 : 0).getReg())
return true;
break;
- case ARM::COPY: {
- // Walk down one instruction which is potentially an 'and'.
- const MachineInstr &Copy = *MI;
- MachineBasicBlock::iterator AND(
- std::next(MachineBasicBlock::iterator(MI)));
- if (AND == MI->getParent()->end()) return false;
- MI = AND;
- return isSuitableForMask(MI, Copy.getOperand(0).getReg(),
- CmpMask, true);
- }
}
return false;
else if (MI->getParent() != CmpInstr->getParent() || CmpValue != 0) {
// Conservatively refuse to convert an instruction which isn't in the same
// BB as the comparison.
- // For CMPri, we need to check Sub, thus we can't return here.
+ // For CMPri w/ CmpValue != 0, a Sub may still be a candidate.
+ // Thus we cannot return here.
if (CmpInstr->getOpcode() == ARM::CMPri ||
CmpInstr->getOpcode() == ARM::t2CMPri)
MI = nullptr;
case ARM::t2EORrr:
case ARM::t2EORri: {
// Scan forward for the use of CPSR
- // When checking against MI: if it's a conditional code requires
- // checking of V bit, then this is not safe to do.
+ // When checking against MI: if it's a conditional code that requires
+ // checking of the V bit or C bit, then this is not safe to do.
// It is safe to remove CmpInstr if CPSR is redefined or killed.
// If we are done with the basic block, we need to check whether CPSR is
// live-out.
OperandsToUpdate.push_back(
std::make_pair(&((*I).getOperand(IO - 1)), NewCC));
}
- } else
+ } else {
+ // No Sub, so this is x = <op> y, z; cmp x, 0.
switch (CC) {
- default:
+ case ARMCC::EQ: // Z
+ case ARMCC::NE: // Z
+ case ARMCC::MI: // N
+ case ARMCC::PL: // N
+ case ARMCC::AL: // none
// CPSR can be used multiple times, we should continue.
break;
- case ARMCC::VS:
- case ARMCC::VC:
- case ARMCC::GE:
- case ARMCC::LT:
- case ARMCC::GT:
- case ARMCC::LE:
+ case ARMCC::HS: // C
+ case ARMCC::LO: // C
+ case ARMCC::VS: // V
+ case ARMCC::VC: // V
+ case ARMCC::HI: // C Z
+ case ARMCC::LS: // C Z
+ case ARMCC::GE: // N V
+ case ARMCC::LT: // N V
+ case ARMCC::GT: // Z N V
+ case ARMCC::LE: // Z N V
+ // The instruction uses the V bit or C bit which is not safe.
return false;
}
+ }
}
}
// instructions).
if (Latency > 0 && Subtarget.isThumb2()) {
const MachineFunction *MF = DefMI->getParent()->getParent();
- if (MF->getFunction()->getAttributes().
- hasAttribute(AttributeSet::FunctionIndex,
- Attribute::OptimizeForSize))
+ if (MF->getFunction()->hasFnAttribute(Attribute::OptimizeForSize))
--Latency;
}
return Latency;
}
bool ARMBaseInstrInfo::
-hasHighOperandLatency(const InstrItineraryData *ItinData,
+hasHighOperandLatency(const TargetSchedModel &SchedModel,
const MachineRegisterInfo *MRI,
const MachineInstr *DefMI, unsigned DefIdx,
const MachineInstr *UseMI, unsigned UseIdx) const {
return true;
// Hoist VFP / NEON instructions with 4 or higher latency.
- int Latency = computeOperandLatency(ItinData, DefMI, DefIdx, UseMI, UseIdx);
- if (Latency < 0)
- Latency = getInstrLatency(ItinData, DefMI);
+ unsigned Latency
+ = SchedModel.computeOperandLatency(DefMI, DefIdx, UseMI, UseIdx);
if (Latency <= 3)
return false;
return DDomain == ARMII::DomainVFP || DDomain == ARMII::DomainNEON ||
}
bool ARMBaseInstrInfo::
-hasLowDefLatency(const InstrItineraryData *ItinData,
+hasLowDefLatency(const TargetSchedModel &SchedModel,
const MachineInstr *DefMI, unsigned DefIdx) const {
+ const InstrItineraryData *ItinData = SchedModel.getInstrItineraries();
if (!ItinData || ItinData->isEmpty())
return false;
//
std::pair<uint16_t, uint16_t>
ARMBaseInstrInfo::getExecutionDomain(const MachineInstr *MI) const {
- // VMOVD, VMOVRS and VMOVSR are VFP instructions, but can be changed to NEON
- // if they are not predicated.
- if (MI->getOpcode() == ARM::VMOVD && !isPredicated(MI))
- return std::make_pair(ExeVFP, (1<<ExeVFP) | (1<<ExeNEON));
-
- // CortexA9 is particularly picky about mixing the two and wants these
- // converted.
- if (Subtarget.isCortexA9() && !isPredicated(MI) &&
- (MI->getOpcode() == ARM::VMOVRS ||
- MI->getOpcode() == ARM::VMOVSR ||
- MI->getOpcode() == ARM::VMOVS))
- return std::make_pair(ExeVFP, (1<<ExeVFP) | (1<<ExeNEON));
-
+ // If we don't have access to NEON instructions then we won't be able
+ // to swizzle anything to the NEON domain. Check to make sure.
+ if (Subtarget.hasNEON()) {
+ // VMOVD, VMOVRS and VMOVSR are VFP instructions, but can be changed to NEON
+ // if they are not predicated.
+ if (MI->getOpcode() == ARM::VMOVD && !isPredicated(MI))
+ return std::make_pair(ExeVFP, (1 << ExeVFP) | (1 << ExeNEON));
+
+ // CortexA9 is particularly picky about mixing the two and wants these
+ // converted.
+ if (Subtarget.isCortexA9() && !isPredicated(MI) &&
+ (MI->getOpcode() == ARM::VMOVRS || MI->getOpcode() == ARM::VMOVSR ||
+ MI->getOpcode() == ARM::VMOVS))
+ return std::make_pair(ExeVFP, (1 << ExeVFP) | (1 << ExeNEON));
+ }
// No other instructions can be swizzled, so just determine their domain.
unsigned Domain = MI->getDesc().TSFlags & ARMII::DomainMask;
// Zap the predicate operands.
assert(!isPredicated(MI) && "Cannot predicate a VORRd");
+ // Make sure we've got NEON instructions.
+ assert(Subtarget.hasNEON() && "VORRd requires NEON");
+
// Source instruction is %DDst = VMOVD %DSrc, 14, %noreg (; implicits)
DstReg = MI->getOperand(0).getReg();
SrcReg = MI->getOperand(1).getReg();
}
bool ARMBaseInstrInfo::hasNOP() const {
- return (Subtarget.getFeatureBits() & ARM::HasV6T2Ops) != 0;
+ return Subtarget.getFeatureBits()[ARM::HasV6KOps];
}
bool ARMBaseInstrInfo::isSwiftFastImmShift(const MachineInstr *MI) const {