//
//===----------------------------------------------------------------------===//
-#define DEBUG_TYPE "isel"
-#include "llvm/Function.h"
-#include "llvm/GlobalVariable.h"
-#include "llvm/Instructions.h"
-#include "llvm/IntrinsicInst.h"
-#include "llvm/Operator.h"
-#include "llvm/CodeGen/Analysis.h"
#include "llvm/CodeGen/FastISel.h"
+#include "llvm/ADT/Optional.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/Analysis/BranchProbabilityInfo.h"
+#include "llvm/Analysis/Loads.h"
+#include "llvm/CodeGen/Analysis.h"
#include "llvm/CodeGen/FunctionLoweringInfo.h"
+#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineModuleInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
-#include "llvm/Analysis/DebugInfo.h"
-#include "llvm/Analysis/Loads.h"
-#include "llvm/Target/TargetData.h"
+#include "llvm/CodeGen/StackMaps.h"
+#include "llvm/IR/DataLayout.h"
+#include "llvm/IR/DebugInfo.h"
+#include "llvm/IR/Function.h"
+#include "llvm/IR/GlobalVariable.h"
+#include "llvm/IR/Instructions.h"
+#include "llvm/IR/IntrinsicInst.h"
+#include "llvm/IR/Operator.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/ErrorHandling.h"
#include "llvm/Target/TargetInstrInfo.h"
+#include "llvm/Target/TargetLibraryInfo.h"
#include "llvm/Target/TargetLowering.h"
#include "llvm/Target/TargetMachine.h"
-#include "llvm/Support/ErrorHandling.h"
-#include "llvm/Support/Debug.h"
-#include "llvm/ADT/Statistic.h"
using namespace llvm;
+#define DEBUG_TYPE "isel"
+
STATISTIC(NumFastIselSuccessIndependent, "Number of insts selected by "
"target-independent selector");
STATISTIC(NumFastIselSuccessTarget, "Number of insts selected by "
void FastISel::startNewBlock() {
LocalValueMap.clear();
- EmitStartPt = 0;
+ // Instructions are appended to FuncInfo.MBB. If the basic block already
+ // contains labels or copies, use the last instruction as the last local
+ // value.
+ EmitStartPt = nullptr;
+ if (!FuncInfo.MBB->empty())
+ EmitStartPt = &FuncInfo.MBB->back();
+ LastLocalValue = EmitStartPt;
+}
- // Advance the emit start point past any EH_LABEL instructions.
- MachineBasicBlock::iterator
- I = FuncInfo.MBB->begin(), E = FuncInfo.MBB->end();
- while (I != E && I->getOpcode() == TargetOpcode::EH_LABEL) {
- EmitStartPt = I;
- ++I;
+bool FastISel::LowerArguments() {
+ if (!FuncInfo.CanLowerReturn)
+ // Fallback to SDISel argument lowering code to deal with sret pointer
+ // parameter.
+ return false;
+
+ if (!FastLowerArguments())
+ return false;
+
+ // Enter arguments into ValueMap for uses in non-entry BBs.
+ for (Function::const_arg_iterator I = FuncInfo.Fn->arg_begin(),
+ E = FuncInfo.Fn->arg_end(); I != E; ++I) {
+ DenseMap<const Value *, unsigned>::iterator VI = LocalValueMap.find(I);
+ assert(VI != LocalValueMap.end() && "Missed an argument?");
+ FuncInfo.ValueMap[I] = VI->second;
}
- LastLocalValue = EmitStartPt;
+ return true;
}
void FastISel::flushLocalValueMap() {
// No-op casts are trivially coalesced by fast-isel.
if (const CastInst *Cast = dyn_cast<CastInst>(I))
- if (Cast->isNoopCast(TD.getIntPtrType(Cast->getContext())) &&
+ if (Cast->isNoopCast(DL.getIntPtrType(Cast->getContext())) &&
!hasTrivialKill(Cast->getOperand(0)))
return false;
!(I->getOpcode() == Instruction::BitCast ||
I->getOpcode() == Instruction::PtrToInt ||
I->getOpcode() == Instruction::IntToPtr) &&
- cast<Instruction>(*I->use_begin())->getParent() == I->getParent();
+ cast<Instruction>(*I->user_begin())->getParent() == I->getParent();
}
unsigned FastISel::getRegForValue(const Value *V) {
return 0;
}
- // Look up the value to see if we already have a register for it. We
- // cache values defined by Instructions across blocks, and other values
- // only locally. This is because Instructions already have the SSA
- // def-dominates-use requirement enforced.
- DenseMap<const Value *, unsigned>::iterator I = FuncInfo.ValueMap.find(V);
- if (I != FuncInfo.ValueMap.end())
- return I->second;
-
- unsigned Reg = LocalValueMap[V];
+ // Look up the value to see if we already have a register for it.
+ unsigned Reg = lookUpRegForValue(V);
if (Reg != 0)
return Reg;
// Translate this as an integer zero so that it can be
// local-CSE'd with actual integer zeros.
Reg =
- getRegForValue(Constant::getNullValue(TD.getIntPtrType(V->getContext())));
+ getRegForValue(Constant::getNullValue(DL.getIntPtrType(V->getContext())));
} else if (const ConstantFP *CF = dyn_cast<ConstantFP>(V)) {
if (CF->isNullValue()) {
Reg = TargetMaterializeFloatZero(CF);
uint32_t IntBitWidth = IntVT.getSizeInBits();
bool isExact;
(void) Flt.convertToInteger(x, IntBitWidth, /*isSigned=*/true,
- APFloat::rmTowardZero, &isExact);
+ APFloat::rmTowardZero, &isExact);
if (isExact) {
APInt IntVal(IntBitWidth, x);
Reg = lookUpRegForValue(Op);
} else if (isa<UndefValue>(V)) {
Reg = createResultReg(TLI.getRegClassFor(VT));
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
TII.get(TargetOpcode::IMPLICIT_DEF), Reg);
}
FastISel::SavePoint FastISel::enterLocalValueArea() {
MachineBasicBlock::iterator OldInsertPt = FuncInfo.InsertPt;
- DebugLoc OldDL = DL;
+ DebugLoc OldDL = DbgLoc;
recomputeInsertPt();
- DL = DebugLoc();
+ DbgLoc = DebugLoc();
SavePoint SP = { OldInsertPt, OldDL };
return SP;
}
void FastISel::leaveLocalValueArea(SavePoint OldInsertPt) {
if (FuncInfo.InsertPt != FuncInfo.MBB->begin())
- LastLocalValue = llvm::prior(FuncInfo.InsertPt);
+ LastLocalValue = std::prev(FuncInfo.InsertPt);
// Restore the previous insert position.
FuncInfo.InsertPt = OldInsertPt.InsertPt;
- DL = OldInsertPt.DL;
+ DbgLoc = OldInsertPt.DL;
}
/// SelectBinaryOp - Select and emit code for a binary operator instruction,
ISDOpcode = ISD::SRA;
}
+ // Transform "urem x, pow2" -> "and x, pow2-1".
+ if (ISDOpcode == ISD::UREM && isa<BinaryOperator>(I) &&
+ isPowerOf2_64(Imm)) {
+ --Imm;
+ ISDOpcode = ISD::AND;
+ }
+
unsigned ResultReg = FastEmit_ri_(VT.getSimpleVT(), ISDOpcode, Op0,
Op0IsKill, Imm, VT.getSimpleVT());
if (ResultReg == 0) return false;
unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
if (Field) {
// N = N + Offset
- TotalOffs += TD.getStructLayout(StTy)->getElementOffset(Field);
+ TotalOffs += DL.getStructLayout(StTy)->getElementOffset(Field);
if (TotalOffs >= MaxOffs) {
N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, TotalOffs, VT);
if (N == 0)
if (const ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
if (CI->isZero()) continue;
// N = N + Offset
- TotalOffs +=
- TD.getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
+ TotalOffs +=
+ DL.getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
if (TotalOffs >= MaxOffs) {
N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, TotalOffs, VT);
if (N == 0)
}
// N = N + Idx * ElementSize;
- uint64_t ElementSize = TD.getTypeAllocSize(Ty);
+ uint64_t ElementSize = DL.getTypeAllocSize(Ty);
std::pair<unsigned, bool> Pair = getRegForGEPIndex(Idx);
unsigned IdxN = Pair.first;
bool IdxNIsKill = Pair.second;
return true;
}
+/// \brief Add a stackmap or patchpoint intrinsic call's live variable operands
+/// to a stackmap or patchpoint machine instruction.
+bool FastISel::addStackMapLiveVars(SmallVectorImpl<MachineOperand> &Ops,
+ const CallInst *CI, unsigned StartIdx) {
+ for (unsigned i = StartIdx, e = CI->getNumArgOperands(); i != e; ++i) {
+ Value *Val = CI->getArgOperand(i);
+ // Check for constants and encode them with a StackMaps::ConstantOp prefix.
+ if (auto *C = dyn_cast<ConstantInt>(Val)) {
+ Ops.push_back(MachineOperand::CreateImm(StackMaps::ConstantOp));
+ Ops.push_back(MachineOperand::CreateImm(C->getSExtValue()));
+ } else if (isa<ConstantPointerNull>(Val)) {
+ Ops.push_back(MachineOperand::CreateImm(StackMaps::ConstantOp));
+ Ops.push_back(MachineOperand::CreateImm(0));
+ } else if (auto *AI = dyn_cast<AllocaInst>(Val)) {
+ // Values coming from a stack location also require a sepcial encoding,
+ // but that is added later on by the target specific frame index
+ // elimination implementation.
+ auto SI = FuncInfo.StaticAllocaMap.find(AI);
+ if (SI != FuncInfo.StaticAllocaMap.end())
+ Ops.push_back(MachineOperand::CreateFI(SI->second));
+ else
+ return false;
+ } else {
+ unsigned Reg = getRegForValue(Val);
+ if (Reg == 0)
+ return false;
+ Ops.push_back(MachineOperand::CreateReg(Reg, /*IsDef=*/false));
+ }
+ }
+
+ return true;
+}
+
+bool FastISel::SelectStackmap(const CallInst *I) {
+ // void @llvm.experimental.stackmap(i64 <id>, i32 <numShadowBytes>,
+ // [live variables...])
+ assert(I->getCalledFunction()->getReturnType()->isVoidTy() &&
+ "Stackmap cannot return a value.");
+
+ // The stackmap intrinsic only records the live variables (the arguments
+ // passed to it) and emits NOPS (if requested). Unlike the patchpoint
+ // intrinsic, this won't be lowered to a function call. This means we don't
+ // have to worry about calling conventions and target-specific lowering code.
+ // Instead we perform the call lowering right here.
+ //
+ // CALLSEQ_START(0)
+ // STACKMAP(id, nbytes, ...)
+ // CALLSEQ_END(0, 0)
+ //
+ SmallVector<MachineOperand, 32> Ops;
+
+ // Add the <id> and <numBytes> constants.
+ assert(isa<ConstantInt>(I->getOperand(PatchPointOpers::IDPos)) &&
+ "Expected a constant integer.");
+ const auto *ID = cast<ConstantInt>(I->getOperand(PatchPointOpers::IDPos));
+ Ops.push_back(MachineOperand::CreateImm(ID->getZExtValue()));
+
+ assert(isa<ConstantInt>(I->getOperand(PatchPointOpers::NBytesPos)) &&
+ "Expected a constant integer.");
+ const auto *NumBytes =
+ cast<ConstantInt>(I->getOperand(PatchPointOpers::NBytesPos));
+ Ops.push_back(MachineOperand::CreateImm(NumBytes->getZExtValue()));
+
+ // Push live variables for the stack map (skipping the first two arguments
+ // <id> and <numBytes>).
+ if (!addStackMapLiveVars(Ops, I, 2))
+ return false;
+
+ // We are not adding any register mask info here, because the stackmap doesn't
+ // clobber anything.
+
+ // Add scratch registers as implicit def and early clobber.
+ CallingConv::ID CC = I->getCallingConv();
+ const MCPhysReg *ScratchRegs = TLI.getScratchRegisters(CC);
+ for (unsigned i = 0; ScratchRegs[i]; ++i)
+ Ops.push_back(MachineOperand::CreateReg(
+ ScratchRegs[i], /*IsDef=*/true, /*IsImp=*/true, /*IsKill=*/false,
+ /*IsDead=*/false, /*IsUndef=*/false, /*IsEarlyClobber=*/true));
+
+ // Issue CALLSEQ_START
+ unsigned AdjStackDown = TII.getCallFrameSetupOpcode();
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackDown))
+ .addImm(0);
+
+ // Issue STACKMAP.
+ MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
+ TII.get(TargetOpcode::STACKMAP));
+ for (auto const &MO : Ops)
+ MIB.addOperand(MO);
+
+ // Issue CALLSEQ_END
+ unsigned AdjStackUp = TII.getCallFrameDestroyOpcode();
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackUp))
+ .addImm(0).addImm(0);
+
+ // Inform the Frame Information that we have a stackmap in this function.
+ FuncInfo.MF->getFrameInfo()->setHasStackMap();
+
+ return true;
+}
+
bool FastISel::SelectCall(const User *I) {
const CallInst *Call = cast<CallInst>(I);
if (IA->isAlignStack())
ExtraInfo |= InlineAsm::Extra_IsAlignStack;
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
TII.get(TargetOpcode::INLINEASM))
.addExternalSymbol(IA->getAsmString().c_str())
.addImm(ExtraInfo);
return true;
}
+ MachineModuleInfo &MMI = FuncInfo.MF->getMMI();
+ ComputeUsesVAFloatArgument(*Call, &MMI);
+
const Function *F = Call->getCalledFunction();
if (!F) return false;
// Handle selected intrinsic function calls.
switch (F->getIntrinsicID()) {
default: break;
+ // At -O0 we don't care about the lifetime intrinsics.
+ case Intrinsic::lifetime_start:
+ case Intrinsic::lifetime_end:
+ // The donothing intrinsic does, well, nothing.
+ case Intrinsic::donothing:
+ return true;
+
case Intrinsic::dbg_declare: {
const DbgDeclareInst *DI = cast<DbgDeclareInst>(Call);
- if (!DIVariable(DI->getVariable()).Verify() ||
- !FuncInfo.MF->getMMI().hasDebugInfo())
+ DIVariable DIVar(DI->getVariable());
+ assert((!DIVar || DIVar.isVariable()) &&
+ "Variable in DbgDeclareInst should be either null or a DIVariable.");
+ if (!DIVar ||
+ !FuncInfo.MF->getMMI().hasDebugInfo()) {
+ DEBUG(dbgs() << "Dropping debug info for " << *DI << "\n");
return true;
+ }
const Value *Address = DI->getAddress();
- if (!Address || isa<UndefValue>(Address) || isa<AllocaInst>(Address))
+ if (!Address || isa<UndefValue>(Address)) {
+ DEBUG(dbgs() << "Dropping debug info for " << *DI << "\n");
return true;
+ }
- unsigned Reg = 0;
unsigned Offset = 0;
- if (const Argument *Arg = dyn_cast<Argument>(Address)) {
+ Optional<MachineOperand> Op;
+ if (const Argument *Arg = dyn_cast<Argument>(Address))
// Some arguments' frame index is recorded during argument lowering.
Offset = FuncInfo.getArgumentFrameIndex(Arg);
- if (Offset)
- Reg = TRI.getFrameRegister(*FuncInfo.MF);
+ if (Offset)
+ Op = MachineOperand::CreateFI(Offset);
+ if (!Op)
+ if (unsigned Reg = lookUpRegForValue(Address))
+ Op = MachineOperand::CreateReg(Reg, false);
+
+ // If we have a VLA that has a "use" in a metadata node that's then used
+ // here but it has no other uses, then we have a problem. E.g.,
+ //
+ // int foo (const int *x) {
+ // char a[*x];
+ // return 0;
+ // }
+ //
+ // If we assign 'a' a vreg and fast isel later on has to use the selection
+ // DAG isel, it will want to copy the value to the vreg. However, there are
+ // no uses, which goes counter to what selection DAG isel expects.
+ if (!Op && !Address->use_empty() && isa<Instruction>(Address) &&
+ (!isa<AllocaInst>(Address) ||
+ !FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(Address))))
+ Op = MachineOperand::CreateReg(FuncInfo.InitializeRegForValue(Address),
+ false);
+
+ if (Op) {
+ if (Op->isReg()) {
+ Op->setIsDebug(true);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
+ TII.get(TargetOpcode::DBG_VALUE), false, Op->getReg(), 0,
+ DI->getVariable());
+ } else
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
+ TII.get(TargetOpcode::DBG_VALUE))
+ .addOperand(*Op)
+ .addImm(0)
+ .addMetadata(DI->getVariable());
+ } else {
+ // We can't yet handle anything else here because it would require
+ // generating code, thus altering codegen because of debug info.
+ DEBUG(dbgs() << "Dropping debug info for " << *DI << "\n");
}
- if (!Reg)
- Reg = getRegForValue(Address);
-
- if (Reg)
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
- TII.get(TargetOpcode::DBG_VALUE))
- .addReg(Reg, RegState::Debug).addImm(Offset)
- .addMetadata(DI->getVariable());
return true;
}
case Intrinsic::dbg_value: {
if (!V) {
// Currently the optimizer can produce this; insert an undef to
// help debugging. Probably the optimizer should not do this.
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
.addReg(0U).addImm(DI->getOffset())
.addMetadata(DI->getVariable());
} else if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
if (CI->getBitWidth() > 64)
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
.addCImm(CI).addImm(DI->getOffset())
.addMetadata(DI->getVariable());
- else
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
+ else
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
.addImm(CI->getZExtValue()).addImm(DI->getOffset())
.addMetadata(DI->getVariable());
} else if (const ConstantFP *CF = dyn_cast<ConstantFP>(V)) {
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
.addFPImm(CF).addImm(DI->getOffset())
.addMetadata(DI->getVariable());
} else if (unsigned Reg = lookUpRegForValue(V)) {
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
- .addReg(Reg, RegState::Debug).addImm(DI->getOffset())
- .addMetadata(DI->getVariable());
+ // FIXME: This does not handle register-indirect values at offset 0.
+ bool IsIndirect = DI->getOffset() != 0;
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, IsIndirect,
+ Reg, DI->getOffset(), DI->getVariable());
} else {
// We can't yet handle anything else here because it would require
// generating code, thus altering codegen because of debug info.
- DEBUG(dbgs() << "Dropping debug info for " << DI);
- }
- return true;
- }
- case Intrinsic::eh_exception: {
- EVT VT = TLI.getValueType(Call->getType());
- if (TLI.getOperationAction(ISD::EXCEPTIONADDR, VT)!=TargetLowering::Expand)
- break;
-
- assert(FuncInfo.MBB->isLandingPad() &&
- "Call to eh.exception not in landing pad!");
- unsigned Reg = TLI.getExceptionAddressRegister();
- const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
- unsigned ResultReg = createResultReg(RC);
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
- ResultReg).addReg(Reg);
- UpdateValueMap(Call, ResultReg);
- return true;
- }
- case Intrinsic::eh_selector: {
- EVT VT = TLI.getValueType(Call->getType());
- if (TLI.getOperationAction(ISD::EHSELECTION, VT) != TargetLowering::Expand)
- break;
- if (FuncInfo.MBB->isLandingPad())
- AddCatchInfo(*Call, &FuncInfo.MF->getMMI(), FuncInfo.MBB);
- else {
-#ifndef NDEBUG
- FuncInfo.CatchInfoLost.insert(Call);
-#endif
- // FIXME: Mark exception selector register as live in. Hack for PR1508.
- unsigned Reg = TLI.getExceptionSelectorRegister();
- if (Reg) FuncInfo.MBB->addLiveIn(Reg);
+ DEBUG(dbgs() << "Dropping debug info for " << *DI << "\n");
}
-
- unsigned Reg = TLI.getExceptionSelectorRegister();
- EVT SrcVT = TLI.getPointerTy();
- const TargetRegisterClass *RC = TLI.getRegClassFor(SrcVT);
- unsigned ResultReg = createResultReg(RC);
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
- ResultReg).addReg(Reg);
-
- bool ResultRegIsKill = hasTrivialKill(Call);
-
- // Cast the register to the type of the selector.
- if (SrcVT.bitsGT(MVT::i32))
- ResultReg = FastEmit_r(SrcVT.getSimpleVT(), MVT::i32, ISD::TRUNCATE,
- ResultReg, ResultRegIsKill);
- else if (SrcVT.bitsLT(MVT::i32))
- ResultReg = FastEmit_r(SrcVT.getSimpleVT(), MVT::i32,
- ISD::SIGN_EXTEND, ResultReg, ResultRegIsKill);
- if (ResultReg == 0)
- // Unhandled operand. Halt "fast" selection and bail.
- return false;
-
- UpdateValueMap(Call, ResultReg);
-
return true;
}
case Intrinsic::objectsize: {
UpdateValueMap(Call, ResultReg);
return true;
}
+ case Intrinsic::expect: {
+ unsigned ResultReg = getRegForValue(Call->getArgOperand(0));
+ if (ResultReg == 0)
+ return false;
+ UpdateValueMap(Call, ResultReg);
+ return true;
+ }
+ case Intrinsic::experimental_stackmap:
+ return SelectStackmap(Call);
}
// Usually, it does not make sense to initialize a value,
// all the values which have already been materialized,
// appear after the call. It also makes sense to skip intrinsics
// since they tend to be inlined.
- if (!isa<IntrinsicInst>(F))
+ if (!isa<IntrinsicInst>(Call))
flushLocalValueMap();
// An arbitrary call. Bail.
}
// Bitcasts of other values become reg-reg copies or BITCAST operators.
- EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
- EVT DstVT = TLI.getValueType(I->getType());
-
- if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
- DstVT == MVT::Other || !DstVT.isSimple() ||
- !TLI.isTypeLegal(SrcVT) || !TLI.isTypeLegal(DstVT))
+ EVT SrcEVT = TLI.getValueType(I->getOperand(0)->getType());
+ EVT DstEVT = TLI.getValueType(I->getType());
+ if (SrcEVT == MVT::Other || DstEVT == MVT::Other ||
+ !TLI.isTypeLegal(SrcEVT) || !TLI.isTypeLegal(DstEVT))
// Unhandled type. Halt "fast" selection and bail.
return false;
+ MVT SrcVT = SrcEVT.getSimpleVT();
+ MVT DstVT = DstEVT.getSimpleVT();
unsigned Op0 = getRegForValue(I->getOperand(0));
if (Op0 == 0)
// Unhandled operand. Halt "fast" selection and bail.
// First, try to perform the bitcast by inserting a reg-reg copy.
unsigned ResultReg = 0;
- if (SrcVT.getSimpleVT() == DstVT.getSimpleVT()) {
- TargetRegisterClass* SrcClass = TLI.getRegClassFor(SrcVT);
- TargetRegisterClass* DstClass = TLI.getRegClassFor(DstVT);
+ if (SrcVT == DstVT) {
+ const TargetRegisterClass* SrcClass = TLI.getRegClassFor(SrcVT);
+ const TargetRegisterClass* DstClass = TLI.getRegClassFor(DstVT);
// Don't attempt a cross-class copy. It will likely fail.
if (SrcClass == DstClass) {
ResultReg = createResultReg(DstClass);
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
- ResultReg).addReg(Op0);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
+ TII.get(TargetOpcode::COPY), ResultReg).addReg(Op0);
}
}
// If the reg-reg copy failed, select a BITCAST opcode.
if (!ResultReg)
- ResultReg = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(),
- ISD::BITCAST, Op0, Op0IsKill);
+ ResultReg = FastEmit_r(SrcVT, DstVT, ISD::BITCAST, Op0, Op0IsKill);
if (!ResultReg)
return false;
if (!HandlePHINodesInSuccessorBlocks(I->getParent()))
return false;
- DL = I->getDebugLoc();
+ DbgLoc = I->getDebugLoc();
MachineBasicBlock::iterator SavedInsertPt = FuncInfo.InsertPt;
+ if (const CallInst *Call = dyn_cast<CallInst>(I)) {
+ const Function *F = Call->getCalledFunction();
+ LibFunc::Func Func;
+
+ // As a special case, don't handle calls to builtin library functions that
+ // may be translated directly to target instructions.
+ if (F && !F->hasLocalLinkage() && F->hasName() &&
+ LibInfo->getLibFunc(F->getName(), Func) &&
+ LibInfo->hasOptimizedCodeGen(Func))
+ return false;
+
+ // Don't handle Intrinsic::trap if a trap funciton is specified.
+ if (F && F->getIntrinsicID() == Intrinsic::trap &&
+ !TM.Options.getTrapFunctionName().empty())
+ return false;
+ }
+
// First, try doing target-independent selection.
if (SelectOperator(I, I->getOpcode())) {
++NumFastIselSuccessIndependent;
- DL = DebugLoc();
+ DbgLoc = DebugLoc();
return true;
}
- // Remove dead code. However, ignore call instructions since we've flushed
+ // Remove dead code. However, ignore call instructions since we've flushed
// the local value map and recomputed the insert point.
if (!isa<CallInst>(I)) {
recomputeInsertPt();
SavedInsertPt = FuncInfo.InsertPt;
if (TargetSelectInstruction(I)) {
++NumFastIselSuccessTarget;
- DL = DebugLoc();
+ DbgLoc = DebugLoc();
return true;
}
// Check for dead code and remove as necessary.
if (SavedInsertPt != FuncInfo.InsertPt)
removeDeadCode(FuncInfo.InsertPt, SavedInsertPt);
- DL = DebugLoc();
+ DbgLoc = DebugLoc();
return false;
}
/// unless it is the immediate (fall-through) successor, and update
/// the CFG.
void
-FastISel::FastEmitBranch(MachineBasicBlock *MSucc, DebugLoc DL) {
- if (FuncInfo.MBB->isLayoutSuccessor(MSucc)) {
- // The unconditional fall-through case, which needs no instructions.
+FastISel::FastEmitBranch(MachineBasicBlock *MSucc, DebugLoc DbgLoc) {
+ if (FuncInfo.MBB->getBasicBlock()->size() > 1 &&
+ FuncInfo.MBB->isLayoutSuccessor(MSucc)) {
+ // For more accurate line information if this is the only instruction
+ // in the block then emit it, otherwise we have the unconditional
+ // fall-through case, which needs no instructions.
} else {
// The unconditional branch case.
- TII.InsertBranch(*FuncInfo.MBB, MSucc, NULL,
- SmallVector<MachineOperand, 0>(), DL);
+ TII.InsertBranch(*FuncInfo.MBB, MSucc, nullptr,
+ SmallVector<MachineOperand, 0>(), DbgLoc);
}
- FuncInfo.MBB->addSuccessor(MSucc);
+ uint32_t BranchWeight = 0;
+ if (FuncInfo.BPI)
+ BranchWeight = FuncInfo.BPI->getEdgeWeight(FuncInfo.MBB->getBasicBlock(),
+ MSucc->getBasicBlock());
+ FuncInfo.MBB->addSuccessor(MSucc, BranchWeight);
}
/// SelectFNeg - Emit an FNeg operation.
return true;
}
-bool
-FastISel::SelectInsertValue(const User *U) {
- const InsertValueInst *IVI = dyn_cast<InsertValueInst>(U);
- if (!IVI)
- return false;
-
- // Only try to handle inserts of legal types. But also allow i16/i8/i1 because
- // they're easy.
- const Value *Val = IVI->getOperand(1);
- Type *ValTy = Val->getType();
- EVT ValVT = TLI.getValueType(ValTy, /*AllowUnknown=*/true);
- if (!ValVT.isSimple())
- return false;
- MVT VT = ValVT.getSimpleVT();
- if (!TLI.isTypeLegal(VT) && VT != MVT::i16 && VT != MVT::i8 && VT != MVT::i1)
- return false;
-
- // Get the Val register.
- unsigned ValReg = getRegForValue(Val);
- if (ValReg == 0) return false;
-
- const Value *Agg = IVI->getOperand(0);
- Type *AggTy = Agg->getType();
-
- // TODO: Is there a better way to do this? For each insertvalue we allocate
- // a new set of virtual registers, which results in a large number of
- // loads/stores from/to the stack that copies the aggregate all over the place
- // and results in lots of spill code. I believe this is necessary to preserve
- // SSA form, but maybe there's something we could do to improve this.
-
- // Get the Aggregate base register.
- unsigned AggBaseReg;
- DenseMap<const Value *, unsigned>::iterator I = FuncInfo.ValueMap.find(Agg);
- if (I != FuncInfo.ValueMap.end())
- AggBaseReg = I->second;
- else if (isa<Instruction>(Agg))
- AggBaseReg = FuncInfo.InitializeRegForValue(Agg);
- else if (isa<UndefValue>(Agg))
- // In this case we don't need to allocate a new set of register that will
- // never be defined. Just copy Val into the proper result registers.
- AggBaseReg = 0;
- else
- return false; // fast-isel can't handle aggregate constants at the moment
-
- // Create result register(s).
- unsigned ResultBaseReg = FuncInfo.CreateRegs(AggTy);
-
- // Get the actual result register, which is an offset from the base register.
- unsigned LinearIndex = ComputeLinearIndex(Agg->getType(), IVI->getIndices());
-
- SmallVector<EVT, 4> AggValueVTs;
- ComputeValueVTs(TLI, AggTy, AggValueVTs);
-
- // Copy the beginning value(s) from the original aggregate.
- unsigned SrcReg;
- unsigned DestReg;
- unsigned BaseRegOff = 0;
- unsigned i = 0;
- for (; i != LinearIndex; ++i) {
- unsigned NRE = TLI.getNumRegisters(FuncInfo.Fn->getContext(),
- AggValueVTs[i]);
- for (unsigned NRI = 0; NRI != NRE; NRI++) {
- if (AggBaseReg) {
- SrcReg = AggBaseReg + BaseRegOff + NRI;
- DestReg = ResultBaseReg + BaseRegOff + NRI;
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
- DestReg).addReg(SrcReg);
- }
- }
- BaseRegOff += NRE;
- }
-
- // FIXME: Handle aggregate inserts. Haven't seen these in practice, but..
- // Copy value(s) from the inserted value(s).
- DestReg = ResultBaseReg + BaseRegOff;
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
- DestReg).addReg(ValReg);
- ++BaseRegOff;
- ++i;
-
- // Copy remaining value(s) from the original aggregate.
- if (AggBaseReg) {
- for (unsigned NumAggValues = AggValueVTs.size(); i != NumAggValues; ++i) {
- unsigned NRE = TLI.getNumRegisters(FuncInfo.Fn->getContext(),
- AggValueVTs[i]);
- for (unsigned NRI = 0; NRI != NRE; NRI++) {
- SrcReg = AggBaseReg + BaseRegOff + NRI;
- DestReg = ResultBaseReg + BaseRegOff + NRI;
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
- DestReg).addReg(SrcReg);
-
- }
- BaseRegOff += NRE;
- }
- }
- UpdateValueMap(IVI, ResultBaseReg);
- return true;
-}
-
bool
FastISel::SelectOperator(const User *I, unsigned Opcode) {
switch (Opcode) {
}
case Instruction::Unreachable:
- // Nothing to emit.
- return true;
+ if (TM.Options.TrapUnreachable)
+ return FastEmit_(MVT::Other, MVT::Other, ISD::TRAP) != 0;
+ else
+ return true;
case Instruction::Alloca:
// FunctionLowering has the static-sized case covered.
case Instruction::ExtractValue:
return SelectExtractValue(I);
- case Instruction::InsertValue:
- return SelectInsertValue(I);
-
case Instruction::PHI:
llvm_unreachable("FastISel shouldn't visit PHI nodes!");
}
}
-FastISel::FastISel(FunctionLoweringInfo &funcInfo)
+FastISel::FastISel(FunctionLoweringInfo &funcInfo,
+ const TargetLibraryInfo *libInfo)
: FuncInfo(funcInfo),
+ MF(funcInfo.MF),
MRI(FuncInfo.MF->getRegInfo()),
MFI(*FuncInfo.MF->getFrameInfo()),
MCP(*FuncInfo.MF->getConstantPool()),
TM(FuncInfo.MF->getTarget()),
- TD(*TM.getTargetData()),
+ DL(*TM.getDataLayout()),
TII(*TM.getInstrInfo()),
TLI(*TM.getTargetLowering()),
- TRI(*TM.getRegisterInfo()) {
+ TRI(*TM.getRegisterInfo()),
+ LibInfo(libInfo) {
}
FastISel::~FastISel() {}
+bool FastISel::FastLowerArguments() {
+ return false;
+}
+
unsigned FastISel::FastEmit_(MVT, MVT,
unsigned) {
return 0;
IntegerType *ITy = IntegerType::get(FuncInfo.Fn->getContext(),
VT.getSizeInBits());
MaterialReg = getRegForValue(ConstantInt::get(ITy, Imm));
+ assert (MaterialReg != 0 && "Unable to materialize imm.");
+ if (MaterialReg == 0) return 0;
}
return FastEmit_rr(VT, VT, Opcode,
Op0, Op0IsKill,
return MRI.createVirtualRegister(RC);
}
+unsigned FastISel::constrainOperandRegClass(const MCInstrDesc &II,
+ unsigned Op, unsigned OpNum) {
+ if (TargetRegisterInfo::isVirtualRegister(Op)) {
+ const TargetRegisterClass *RegClass =
+ TII.getRegClass(II, OpNum, &TRI, *FuncInfo.MF);
+ if (!MRI.constrainRegClass(Op, RegClass)) {
+ // If it's not legal to COPY between the register classes, something
+ // has gone very wrong before we got here.
+ unsigned NewOp = createResultReg(RegClass);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
+ TII.get(TargetOpcode::COPY), NewOp).addReg(Op);
+ return NewOp;
+ }
+ }
+ return Op;
+}
+
unsigned FastISel::FastEmitInst_(unsigned MachineInstOpcode,
const TargetRegisterClass* RC) {
unsigned ResultReg = createResultReg(RC);
const MCInstrDesc &II = TII.get(MachineInstOpcode);
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg);
return ResultReg;
}
unsigned FastISel::FastEmitInst_r(unsigned MachineInstOpcode,
const TargetRegisterClass *RC,
unsigned Op0, bool Op0IsKill) {
- unsigned ResultReg = createResultReg(RC);
const MCInstrDesc &II = TII.get(MachineInstOpcode);
+ unsigned ResultReg = createResultReg(RC);
+ Op0 = constrainOperandRegClass(II, Op0, II.getNumDefs());
+
if (II.getNumDefs() >= 1)
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
.addReg(Op0, Op0IsKill * RegState::Kill);
else {
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
.addReg(Op0, Op0IsKill * RegState::Kill);
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
- ResultReg).addReg(II.ImplicitDefs[0]);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
+ TII.get(TargetOpcode::COPY), ResultReg).addReg(II.ImplicitDefs[0]);
}
return ResultReg;
const TargetRegisterClass *RC,
unsigned Op0, bool Op0IsKill,
unsigned Op1, bool Op1IsKill) {
- unsigned ResultReg = createResultReg(RC);
const MCInstrDesc &II = TII.get(MachineInstOpcode);
+ unsigned ResultReg = createResultReg(RC);
+ Op0 = constrainOperandRegClass(II, Op0, II.getNumDefs());
+ Op1 = constrainOperandRegClass(II, Op1, II.getNumDefs() + 1);
+
if (II.getNumDefs() >= 1)
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
.addReg(Op0, Op0IsKill * RegState::Kill)
.addReg(Op1, Op1IsKill * RegState::Kill);
else {
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
.addReg(Op0, Op0IsKill * RegState::Kill)
.addReg(Op1, Op1IsKill * RegState::Kill);
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
- ResultReg).addReg(II.ImplicitDefs[0]);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
+ TII.get(TargetOpcode::COPY), ResultReg).addReg(II.ImplicitDefs[0]);
}
return ResultReg;
}
unsigned Op0, bool Op0IsKill,
unsigned Op1, bool Op1IsKill,
unsigned Op2, bool Op2IsKill) {
- unsigned ResultReg = createResultReg(RC);
const MCInstrDesc &II = TII.get(MachineInstOpcode);
+ unsigned ResultReg = createResultReg(RC);
+ Op0 = constrainOperandRegClass(II, Op0, II.getNumDefs());
+ Op1 = constrainOperandRegClass(II, Op1, II.getNumDefs() + 1);
+ Op2 = constrainOperandRegClass(II, Op2, II.getNumDefs() + 2);
+
if (II.getNumDefs() >= 1)
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
.addReg(Op0, Op0IsKill * RegState::Kill)
.addReg(Op1, Op1IsKill * RegState::Kill)
.addReg(Op2, Op2IsKill * RegState::Kill);
else {
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
.addReg(Op0, Op0IsKill * RegState::Kill)
.addReg(Op1, Op1IsKill * RegState::Kill)
.addReg(Op2, Op2IsKill * RegState::Kill);
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
- ResultReg).addReg(II.ImplicitDefs[0]);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
+ TII.get(TargetOpcode::COPY), ResultReg).addReg(II.ImplicitDefs[0]);
}
return ResultReg;
}
const TargetRegisterClass *RC,
unsigned Op0, bool Op0IsKill,
uint64_t Imm) {
- unsigned ResultReg = createResultReg(RC);
const MCInstrDesc &II = TII.get(MachineInstOpcode);
+ unsigned ResultReg = createResultReg(RC);
+ RC = TII.getRegClass(II, II.getNumDefs(), &TRI, *FuncInfo.MF);
+ MRI.constrainRegClass(Op0, RC);
+
if (II.getNumDefs() >= 1)
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
.addReg(Op0, Op0IsKill * RegState::Kill)
.addImm(Imm);
else {
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
.addReg(Op0, Op0IsKill * RegState::Kill)
.addImm(Imm);
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
- ResultReg).addReg(II.ImplicitDefs[0]);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
+ TII.get(TargetOpcode::COPY), ResultReg).addReg(II.ImplicitDefs[0]);
}
return ResultReg;
}
const TargetRegisterClass *RC,
unsigned Op0, bool Op0IsKill,
uint64_t Imm1, uint64_t Imm2) {
- unsigned ResultReg = createResultReg(RC);
const MCInstrDesc &II = TII.get(MachineInstOpcode);
+ unsigned ResultReg = createResultReg(RC);
+ Op0 = constrainOperandRegClass(II, Op0, II.getNumDefs());
+
if (II.getNumDefs() >= 1)
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
.addReg(Op0, Op0IsKill * RegState::Kill)
.addImm(Imm1)
.addImm(Imm2);
else {
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
.addReg(Op0, Op0IsKill * RegState::Kill)
.addImm(Imm1)
.addImm(Imm2);
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
- ResultReg).addReg(II.ImplicitDefs[0]);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
+ TII.get(TargetOpcode::COPY), ResultReg).addReg(II.ImplicitDefs[0]);
}
return ResultReg;
}
const TargetRegisterClass *RC,
unsigned Op0, bool Op0IsKill,
const ConstantFP *FPImm) {
- unsigned ResultReg = createResultReg(RC);
const MCInstrDesc &II = TII.get(MachineInstOpcode);
+ unsigned ResultReg = createResultReg(RC);
+ Op0 = constrainOperandRegClass(II, Op0, II.getNumDefs());
+
if (II.getNumDefs() >= 1)
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
.addReg(Op0, Op0IsKill * RegState::Kill)
.addFPImm(FPImm);
else {
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
.addReg(Op0, Op0IsKill * RegState::Kill)
.addFPImm(FPImm);
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
- ResultReg).addReg(II.ImplicitDefs[0]);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
+ TII.get(TargetOpcode::COPY), ResultReg).addReg(II.ImplicitDefs[0]);
}
return ResultReg;
}
unsigned Op0, bool Op0IsKill,
unsigned Op1, bool Op1IsKill,
uint64_t Imm) {
- unsigned ResultReg = createResultReg(RC);
const MCInstrDesc &II = TII.get(MachineInstOpcode);
+ unsigned ResultReg = createResultReg(RC);
+ Op0 = constrainOperandRegClass(II, Op0, II.getNumDefs());
+ Op1 = constrainOperandRegClass(II, Op1, II.getNumDefs() + 1);
+
if (II.getNumDefs() >= 1)
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
.addReg(Op0, Op0IsKill * RegState::Kill)
.addReg(Op1, Op1IsKill * RegState::Kill)
.addImm(Imm);
else {
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
.addReg(Op0, Op0IsKill * RegState::Kill)
.addReg(Op1, Op1IsKill * RegState::Kill)
.addImm(Imm);
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
- ResultReg).addReg(II.ImplicitDefs[0]);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
+ TII.get(TargetOpcode::COPY), ResultReg).addReg(II.ImplicitDefs[0]);
+ }
+ return ResultReg;
+}
+
+unsigned FastISel::FastEmitInst_rrii(unsigned MachineInstOpcode,
+ const TargetRegisterClass *RC,
+ unsigned Op0, bool Op0IsKill,
+ unsigned Op1, bool Op1IsKill,
+ uint64_t Imm1, uint64_t Imm2) {
+ const MCInstrDesc &II = TII.get(MachineInstOpcode);
+
+ unsigned ResultReg = createResultReg(RC);
+ Op0 = constrainOperandRegClass(II, Op0, II.getNumDefs());
+ Op1 = constrainOperandRegClass(II, Op1, II.getNumDefs() + 1);
+
+ if (II.getNumDefs() >= 1)
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
+ .addReg(Op0, Op0IsKill * RegState::Kill)
+ .addReg(Op1, Op1IsKill * RegState::Kill)
+ .addImm(Imm1).addImm(Imm2);
+ else {
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
+ .addReg(Op0, Op0IsKill * RegState::Kill)
+ .addReg(Op1, Op1IsKill * RegState::Kill)
+ .addImm(Imm1).addImm(Imm2);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
+ TII.get(TargetOpcode::COPY), ResultReg).addReg(II.ImplicitDefs[0]);
}
return ResultReg;
}
const MCInstrDesc &II = TII.get(MachineInstOpcode);
if (II.getNumDefs() >= 1)
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg).addImm(Imm);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg).addImm(Imm);
else {
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II).addImm(Imm);
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
- ResultReg).addReg(II.ImplicitDefs[0]);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II).addImm(Imm);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
+ TII.get(TargetOpcode::COPY), ResultReg).addReg(II.ImplicitDefs[0]);
}
return ResultReg;
}
const MCInstrDesc &II = TII.get(MachineInstOpcode);
if (II.getNumDefs() >= 1)
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
.addImm(Imm1).addImm(Imm2);
else {
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II).addImm(Imm1).addImm(Imm2);
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
- ResultReg).addReg(II.ImplicitDefs[0]);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II).addImm(Imm1).addImm(Imm2);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
+ TII.get(TargetOpcode::COPY), ResultReg).addReg(II.ImplicitDefs[0]);
}
return ResultReg;
}
unsigned ResultReg = createResultReg(TLI.getRegClassFor(RetVT));
assert(TargetRegisterInfo::isVirtualRegister(Op0) &&
"Cannot yet extract from physregs");
+ const TargetRegisterClass *RC = MRI.getRegClass(Op0);
+ MRI.constrainRegClass(Op0, TRI.getSubClassWithSubReg(RC, Idx));
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
- DL, TII.get(TargetOpcode::COPY), ResultReg)
+ DbgLoc, TII.get(TargetOpcode::COPY), ResultReg)
.addReg(Op0, getKillRegState(Op0IsKill), Idx);
return ResultReg;
}
// exactly one register for each non-void instruction.
EVT VT = TLI.getValueType(PN->getType(), /*AllowUnknown=*/true);
if (VT == MVT::Other || !TLI.isTypeLegal(VT)) {
- // Promote MVT::i1.
- if (VT == MVT::i1)
+ // Handle integer promotions, though, because they're common and easy.
+ if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16)
VT = TLI.getTypeToTransformTo(LLVMBB->getContext(), VT);
else {
FuncInfo.PHINodesToUpdate.resize(OrigNumPHINodesToUpdate);
// Set the DebugLoc for the copy. Prefer the location of the operand
// if there is one; use the location of the PHI otherwise.
- DL = PN->getDebugLoc();
+ DbgLoc = PN->getDebugLoc();
if (const Instruction *Inst = dyn_cast<Instruction>(PHIOp))
- DL = Inst->getDebugLoc();
+ DbgLoc = Inst->getDebugLoc();
unsigned Reg = getRegForValue(PHIOp);
if (Reg == 0) {
return false;
}
FuncInfo.PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg));
- DL = DebugLoc();
+ DbgLoc = DebugLoc();
}
}
return true;
}
+
+bool FastISel::tryToFoldLoad(const LoadInst *LI, const Instruction *FoldInst) {
+ assert(LI->hasOneUse() &&
+ "tryToFoldLoad expected a LoadInst with a single use");
+ // We know that the load has a single use, but don't know what it is. If it
+ // isn't one of the folded instructions, then we can't succeed here. Handle
+ // this by scanning the single-use users of the load until we get to FoldInst.
+ unsigned MaxUsers = 6; // Don't scan down huge single-use chains of instrs.
+
+ const Instruction *TheUser = LI->user_back();
+ while (TheUser != FoldInst && // Scan up until we find FoldInst.
+ // Stay in the right block.
+ TheUser->getParent() == FoldInst->getParent() &&
+ --MaxUsers) { // Don't scan too far.
+ // If there are multiple or no uses of this instruction, then bail out.
+ if (!TheUser->hasOneUse())
+ return false;
+
+ TheUser = TheUser->user_back();
+ }
+
+ // If we didn't find the fold instruction, then we failed to collapse the
+ // sequence.
+ if (TheUser != FoldInst)
+ return false;
+
+ // Don't try to fold volatile loads. Target has to deal with alignment
+ // constraints.
+ if (LI->isVolatile())
+ return false;
+
+ // Figure out which vreg this is going into. If there is no assigned vreg yet
+ // then there actually was no reference to it. Perhaps the load is referenced
+ // by a dead instruction.
+ unsigned LoadReg = getRegForValue(LI);
+ if (LoadReg == 0)
+ return false;
+
+ // We can't fold if this vreg has no uses or more than one use. Multiple uses
+ // may mean that the instruction got lowered to multiple MIs, or the use of
+ // the loaded value ended up being multiple operands of the result.
+ if (!MRI.hasOneUse(LoadReg))
+ return false;
+
+ MachineRegisterInfo::reg_iterator RI = MRI.reg_begin(LoadReg);
+ MachineInstr *User = RI->getParent();
+
+ // Set the insertion point properly. Folding the load can cause generation of
+ // other random instructions (like sign extends) for addressing modes; make
+ // sure they get inserted in a logical place before the new instruction.
+ FuncInfo.InsertPt = User;
+ FuncInfo.MBB = User->getParent();
+
+ // Ask the target to try folding the load.
+ return tryToFoldLoadIntoMI(User, RI.getOperandNo(), LI);
+}
+
+bool FastISel::canFoldAddIntoGEP(const User *GEP, const Value *Add) {
+ // Must be an add.
+ if (!isa<AddOperator>(Add))
+ return false;
+ // Type size needs to match.
+ if (DL.getTypeSizeInBits(GEP->getType()) !=
+ DL.getTypeSizeInBits(Add->getType()))
+ return false;
+ // Must be in the same basic block.
+ if (isa<Instruction>(Add) &&
+ FuncInfo.MBBMap[cast<Instruction>(Add)->getParent()] != FuncInfo.MBB)
+ return false;
+ // Must have a constant operand.
+ return isa<ConstantInt>(cast<AddOperator>(Add)->getOperand(1));
+}
+
+MachineMemOperand *
+FastISel::createMachineMemOperandFor(const Instruction *I) const {
+ const Value *Ptr;
+ Type *ValTy;
+ unsigned Alignment;
+ unsigned Flags;
+ bool IsVolatile;
+
+ if (const auto *LI = dyn_cast<LoadInst>(I)) {
+ Alignment = LI->getAlignment();
+ IsVolatile = LI->isVolatile();
+ Flags = MachineMemOperand::MOLoad;
+ Ptr = LI->getPointerOperand();
+ ValTy = LI->getType();
+ } else if (const auto *SI = dyn_cast<StoreInst>(I)) {
+ Alignment = SI->getAlignment();
+ IsVolatile = SI->isVolatile();
+ Flags = MachineMemOperand::MOStore;
+ Ptr = SI->getPointerOperand();
+ ValTy = SI->getValueOperand()->getType();
+ } else {
+ return nullptr;
+ }
+
+ bool IsNonTemporal = I->getMetadata("nontemporal") != nullptr;
+ bool IsInvariant = I->getMetadata("invariant.load") != nullptr;
+ const MDNode *TBAAInfo = I->getMetadata(LLVMContext::MD_tbaa);
+ const MDNode *Ranges = I->getMetadata(LLVMContext::MD_range);
+
+ if (Alignment == 0) // Ensure that codegen never sees alignment 0.
+ Alignment = DL.getABITypeAlignment(ValTy);
+
+ unsigned Size = TM.getDataLayout()->getTypeStoreSize(ValTy);
+
+ if (IsVolatile)
+ Flags |= MachineMemOperand::MOVolatile;
+ if (IsNonTemporal)
+ Flags |= MachineMemOperand::MONonTemporal;
+ if (IsInvariant)
+ Flags |= MachineMemOperand::MOInvariant;
+
+ return FuncInfo.MF->getMachineMemOperand(MachinePointerInfo(Ptr), Flags, Size,
+ Alignment, TBAAInfo, Ranges);
+}
projects / oota-llvm.git / blobdiff