//===----------------------------------------------------------------------===//
#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/Statistic.h"
+#include "llvm/Analysis/Loads.h"
+#include "llvm/CodeGen/Analysis.h"
#include "llvm/CodeGen/FunctionLoweringInfo.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/DebugInfo.h"
+#include "llvm/IR/DataLayout.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;
STATISTIC(NumFastIselSuccessIndependent, "Number of insts selected by "
LastLocalValue = EmitStartPt;
}
+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 non-dead 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) {
+ if (!I->use_empty()) {
+ DenseMap<const Value *, unsigned>::iterator VI = LocalValueMap.find(I);
+ assert(VI != LocalValueMap.end() && "Missed an argument?");
+ FuncInfo.ValueMap[I] = VI->second;
+ }
+ }
+ return true;
+}
+
void FastISel::flushLocalValueMap() {
LocalValueMap.clear();
LastLocalValue = EmitStartPt;
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;
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);
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;
if (const ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
if (CI->isZero()) continue;
// N = N + Offset
- TotalOffs +=
+ TotalOffs +=
TD.getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
if (TotalOffs >= MaxOffs) {
N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, TotalOffs, VT);
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())
+ !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;
// Some arguments' frame index is recorded during argument lowering.
Offset = FuncInfo.getArgumentFrameIndex(Arg);
if (Offset)
- Reg = TRI.getFrameRegister(*FuncInfo.MF);
+ Reg = TRI.getFrameRegister(*FuncInfo.MF);
}
if (!Reg)
- Reg = getRegForValue(Address);
+ Reg = lookUpRegForValue(Address);
+
+ // 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 (!Reg && !Address->use_empty() && isa<Instruction>(Address) &&
+ (!isa<AllocaInst>(Address) ||
+ !FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(Address))))
+ Reg = FuncInfo.InitializeRegForValue(Address);
if (Reg)
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(TargetOpcode::DBG_VALUE))
.addReg(Reg, RegState::Debug).addImm(Offset)
.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);
return true;
}
case Intrinsic::dbg_value: {
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
.addCImm(CI).addImm(DI->getOffset())
.addMetadata(DI->getVariable());
- else
+ else
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
.addImm(CI->getZExtValue()).addImm(DI->getOffset())
.addMetadata(DI->getVariable());
}
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);
- }
-
- 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: {
ConstantInt *CI = cast<ConstantInt>(Call->getArgOperand(1));
unsigned long long Res = CI->isZero() ? -1ULL : 0;
UpdateValueMap(Call, ResultReg);
return true;
}
+ case Intrinsic::expect: {
+ unsigned ResultReg = getRegForValue(Call->getArgOperand(0));
+ if (ResultReg == 0)
+ return false;
+ UpdateValueMap(Call, ResultReg);
+ return true;
+ }
}
// 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);
// 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;
MachineBasicBlock::iterator SavedInsertPt = FuncInfo.InsertPt;
+ // As a special case, don't handle calls to builtin library functions that
+ // may be translated directly to target instructions.
+ if (const CallInst *Call = dyn_cast<CallInst>(I)) {
+ const Function *F = Call->getCalledFunction();
+ LibFunc::Func Func;
+ if (F && !F->hasLocalLinkage() && F->hasName() &&
+ LibInfo->getLibFunc(F->getName(), Func) &&
+ LibInfo->hasOptimizedCodeGen(Func))
+ return false;
+ }
+
// First, try doing target-independent selection.
if (SelectOperator(I, I->getOpcode())) {
++NumFastIselSuccessIndependent;
DL = 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();
/// the CFG.
void
FastISel::FastEmitBranch(MachineBasicBlock *MSucc, DebugLoc DL) {
- if (FuncInfo.MBB->isLayoutSuccessor(MSucc)) {
- // The unconditional fall-through case, which needs no instructions.
+
+ 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,
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::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),
MRI(FuncInfo.MF->getRegInfo()),
MFI(*FuncInfo.MF->getFrameInfo()),
MCP(*FuncInfo.MF->getConstantPool()),
TM(FuncInfo.MF->getTarget()),
- TD(*TM.getTargetData()),
+ TD(*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 ResultReg;
}
+unsigned FastISel::FastEmitInst_rrii(unsigned MachineInstOpcode,
+ const TargetRegisterClass *RC,
+ unsigned Op0, bool Op0IsKill,
+ unsigned Op1, bool Op1IsKill,
+ uint64_t Imm1, uint64_t Imm2) {
+ unsigned ResultReg = createResultReg(RC);
+ const MCInstrDesc &II = TII.get(MachineInstOpcode);
+
+ if (II.getNumDefs() >= 1)
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
+ .addReg(Op0, Op0IsKill * RegState::Kill)
+ .addReg(Op1, Op1IsKill * RegState::Kill)
+ .addImm(Imm1).addImm(Imm2);
+ else {
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
+ .addReg(Op0, Op0IsKill * RegState::Kill)
+ .addReg(Op1, Op1IsKill * RegState::Kill)
+ .addImm(Imm1).addImm(Imm2);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
+ ResultReg).addReg(II.ImplicitDefs[0]);
+ }
+ return ResultReg;
+}
+
unsigned FastISel::FastEmitInst_i(unsigned MachineInstOpcode,
const TargetRegisterClass *RC,
uint64_t Imm) {
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)
.addReg(Op0, getKillRegState(Op0IsKill), Idx);
// 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);
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->use_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->use_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;
+
+ // 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);
+}
+
+