//
//===----------------------------------------------------------------------===//
-#define DEBUG_TYPE "isel"
-#include "llvm/DebugInfo.h"
-#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/Loads.h"
-#include "llvm/DataLayout.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"
+#include "llvm/Target/TargetSubtargetInfo.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 "
"target-specific selector");
STATISTIC(NumFastIselDead, "Number of dead insts removed on failure");
+/// \brief Set CallLoweringInfo attribute flags based on a call instruction
+/// and called function attributes.
+void FastISel::ArgListEntry::setAttributes(ImmutableCallSite *CS,
+ unsigned AttrIdx) {
+ isSExt = CS->paramHasAttr(AttrIdx, Attribute::SExt);
+ isZExt = CS->paramHasAttr(AttrIdx, Attribute::ZExt);
+ isInReg = CS->paramHasAttr(AttrIdx, Attribute::InReg);
+ isSRet = CS->paramHasAttr(AttrIdx, Attribute::StructRet);
+ isNest = CS->paramHasAttr(AttrIdx, Attribute::Nest);
+ isByVal = CS->paramHasAttr(AttrIdx, Attribute::ByVal);
+ isInAlloca = CS->paramHasAttr(AttrIdx, Attribute::InAlloca);
+ isReturned = CS->paramHasAttr(AttrIdx, Attribute::Returned);
+ Alignment = CS->getParamAlignment(AttrIdx);
+}
+
/// startNewBlock - Set the current block to which generated machine
/// instructions will be appended, and clear the local CSE map.
///
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() {
recomputeInsertPt();
}
-bool FastISel::hasTrivialKill(const Value *V) const {
+bool FastISel::hasTrivialKill(const Value *V) {
// Don't consider constants or arguments to have trivial kills.
const Instruction *I = dyn_cast<Instruction>(V);
if (!I)
// 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;
+ // Even the value might have only one use in the LLVM IR, it is possible that
+ // FastISel might fold the use into another instruction and now there is more
+ // than one use at the Machine Instruction level.
+ unsigned Reg = lookUpRegForValue(V);
+ if (Reg && !MRI.use_empty(Reg))
+ return false;
+
// GEPs with all zero indices are trivially coalesced by fast-isel.
if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I))
if (GEP->hasAllZeroIndices() && !hasTrivialKill(GEP->getOperand(0)))
!(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 Reg;
}
-/// materializeRegForValue - Helper for getRegForValue. This function is
-/// called when the value isn't already available in a register and must
-/// be materialized with new instructions.
-unsigned FastISel::materializeRegForValue(const Value *V, MVT VT) {
+unsigned FastISel::MaterializeConstant(const Value *V, MVT VT) {
unsigned Reg = 0;
-
if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
if (CI->getValue().getActiveBits() <= 64)
Reg = FastEmit_i(VT, VT, ISD::Constant, CI->getZExtValue());
- } else if (isa<AllocaInst>(V)) {
+ } else if (isa<AllocaInst>(V))
Reg = TargetMaterializeAlloca(cast<AllocaInst>(V));
- } else if (isa<ConstantPointerNull>(V)) {
+ else if (isa<ConstantPointerNull>(V))
// 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())));
- } else if (const ConstantFP *CF = dyn_cast<ConstantFP>(V)) {
- if (CF->isNullValue()) {
+ getRegForValue(Constant::getNullValue(DL.getIntPtrType(V->getContext())));
+ else if (const ConstantFP *CF = dyn_cast<ConstantFP>(V)) {
+ if (CF->isNullValue())
Reg = TargetMaterializeFloatZero(CF);
- } else {
+ else
// Try to emit the constant directly.
Reg = FastEmit_f(VT, VT, ISD::ConstantFP, CF);
- }
if (!Reg) {
// Try to emit the constant by using an integer constant with a cast.
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);
}
+ return Reg;
+}
- // If target-independent code couldn't handle the value, give target-specific
- // code a try.
- if (!Reg && isa<Constant>(V))
+/// materializeRegForValue - Helper for getRegForValue. This function is
+/// called when the value isn't already available in a register and must
+/// be materialized with new instructions.
+unsigned FastISel::materializeRegForValue(const Value *V, MVT VT) {
+ unsigned Reg = 0;
+ // Give the target-specific code a try first.
+ if (isa<Constant>(V))
Reg = TargetMaterializeConstant(cast<Constant>(V));
+ // If target-specific code couldn't or didn't want to handle the value, then
+ // give target-independent code a try.
+ if (!Reg)
+ Reg = MaterializeConstant(V, VT);
+
// Don't cache constant materializations in the general ValueMap.
// To do so would require tracking what uses they dominate.
- if (Reg != 0) {
+ if (Reg) {
LocalValueMap[V] = Reg;
LastLocalValue = MRI.getVRegDef(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,
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 (CI->isZero()) continue;
// N = N + Offset
TotalOffs +=
- TD.getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
+ 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;
+}
+
+/// \brief Lower an argument list according to the target calling convention.
+///
+/// This is a helper for lowering intrinsics that follow a target calling
+/// convention or require stack pointer adjustment. Only a subset of the
+/// intrinsic's operands need to participate in the calling convention.
+bool FastISel::lowerCallOperands(const CallInst *CI, unsigned ArgIdx,
+ unsigned NumArgs, const Value *Callee,
+ bool ForceRetVoidTy, CallLoweringInfo &CLI) {
+ ArgListTy Args;
+ Args.reserve(NumArgs);
+
+ // Populate the argument list.
+ // Attributes for args start at offset 1, after the return attribute.
+ ImmutableCallSite CS(CI);
+ for (unsigned ArgI = ArgIdx, ArgE = ArgIdx + NumArgs, AttrI = ArgIdx + 1;
+ ArgI != ArgE; ++ArgI) {
+ Value *V = CI->getOperand(ArgI);
+
+ assert(!V->getType()->isEmptyTy() && "Empty type passed to intrinsic.");
+
+ ArgListEntry Entry;
+ Entry.Val = V;
+ Entry.Ty = V->getType();
+ Entry.setAttributes(&CS, AttrI);
+ Args.push_back(Entry);
+ }
+
+ Type *RetTy = ForceRetVoidTy ? Type::getVoidTy(CI->getType()->getContext())
+ : CI->getType();
+ CLI.setCallee(CI->getCallingConv(), RetTy, Callee, std::move(Args), NumArgs);
+
+ return LowerCallTo(CLI);
+}
+
+bool FastISel::SelectPatchpoint(const CallInst *I) {
+ // void|i64 @llvm.experimental.patchpoint.void|i64(i64 <id>,
+ // i32 <numBytes>,
+ // i8* <target>,
+ // i32 <numArgs>,
+ // [Args...],
+ // [live variables...])
+ CallingConv::ID CC = I->getCallingConv();
+ bool IsAnyRegCC = CC == CallingConv::AnyReg;
+ bool HasDef = !I->getType()->isVoidTy();
+ Value *Callee = I->getOperand(PatchPointOpers::TargetPos);
+
+ // Get the real number of arguments participating in the call <numArgs>
+ assert(isa<ConstantInt>(I->getOperand(PatchPointOpers::NArgPos)) &&
+ "Expected a constant integer.");
+ const auto *NumArgsVal =
+ cast<ConstantInt>(I->getOperand(PatchPointOpers::NArgPos));
+ unsigned NumArgs = NumArgsVal->getZExtValue();
+
+ // Skip the four meta args: <id>, <numNopBytes>, <target>, <numArgs>
+ // This includes all meta-operands up to but not including CC.
+ unsigned NumMetaOpers = PatchPointOpers::CCPos;
+ assert(I->getNumArgOperands() >= NumMetaOpers + NumArgs &&
+ "Not enough arguments provided to the patchpoint intrinsic");
+
+ // For AnyRegCC the arguments are lowered later on manually.
+ unsigned NumCallArgs = IsAnyRegCC ? 0 : NumArgs;
+ CallLoweringInfo CLI;
+ if (!lowerCallOperands(I, NumMetaOpers, NumCallArgs, Callee, IsAnyRegCC, CLI))
+ return false;
+
+ assert(CLI.Call && "No call instruction specified.");
+
+ SmallVector<MachineOperand, 32> Ops;
+
+ // Add an explicit result reg if we use the anyreg calling convention.
+ if (IsAnyRegCC && HasDef) {
+ assert(CLI.NumResultRegs == 0 && "Unexpected result register.");
+ CLI.ResultReg = createResultReg(TLI.getRegClassFor(MVT::i64));
+ CLI.NumResultRegs = 1;
+ Ops.push_back(MachineOperand::CreateReg(CLI.ResultReg, /*IsDef=*/true));
+ }
+
+ // 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()));
+
+ // Assume that the callee is a constant address or null pointer.
+ // FIXME: handle function symbols in the future.
+ uint64_t CalleeAddr;
+ if (const auto *C = dyn_cast<IntToPtrInst>(Callee))
+ CalleeAddr = cast<ConstantInt>(C->getOperand(0))->getZExtValue();
+ else if (const auto *C = dyn_cast<ConstantExpr>(Callee)) {
+ if (C->getOpcode() == Instruction::IntToPtr)
+ CalleeAddr = cast<ConstantInt>(C->getOperand(0))->getZExtValue();
+ else
+ llvm_unreachable("Unsupported ConstantExpr.");
+ } else if (isa<ConstantPointerNull>(Callee))
+ CalleeAddr = 0;
+ else
+ llvm_unreachable("Unsupported callee address.");
+
+ Ops.push_back(MachineOperand::CreateImm(CalleeAddr));
+
+ // Adjust <numArgs> to account for any arguments that have been passed on
+ // the stack instead.
+ unsigned NumCallRegArgs = IsAnyRegCC ? NumArgs : CLI.OutRegs.size();
+ Ops.push_back(MachineOperand::CreateImm(NumCallRegArgs));
+
+ // Add the calling convention
+ Ops.push_back(MachineOperand::CreateImm((unsigned)CC));
+
+ // Add the arguments we omitted previously. The register allocator should
+ // place these in any free register.
+ if (IsAnyRegCC) {
+ for (unsigned i = NumMetaOpers, e = NumMetaOpers + NumArgs; i != e; ++i) {
+ unsigned Reg = getRegForValue(I->getArgOperand(i));
+ if (!Reg)
+ return false;
+ Ops.push_back(MachineOperand::CreateReg(Reg, /*IsDef=*/false));
+ }
+ }
+
+ // Push the arguments from the call instruction.
+ for (auto Reg : CLI.OutRegs)
+ Ops.push_back(MachineOperand::CreateReg(Reg, /*IsDef=*/false));
+
+ // Push live variables for the stack map.
+ if (!addStackMapLiveVars(Ops, I, NumMetaOpers + NumArgs))
+ return false;
+
+ // Push the register mask info.
+ Ops.push_back(MachineOperand::CreateRegMask(TRI.getCallPreservedMask(CC)));
+
+ // Add scratch registers as implicit def and early clobber.
+ 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));
+
+ // Add implicit defs (return values).
+ for (auto Reg : CLI.InRegs)
+ Ops.push_back(MachineOperand::CreateReg(Reg, /*IsDef=*/true,
+ /*IsImpl=*/true));
+
+ // Insert the patchpoint instruction before the call generated by the target.
+ MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, CLI.Call, DbgLoc,
+ TII.get(TargetOpcode::PATCHPOINT));
+
+ for (auto &MO : Ops)
+ MIB.addOperand(MO);
+
+ MIB->setPhysRegsDeadExcept(CLI.InRegs, TRI);
+
+ // Delete the original call instruction.
+ CLI.Call->eraseFromParent();
+
+ // Inform the Frame Information that we have a patchpoint in this function.
+ FuncInfo.MF->getFrameInfo()->setHasPatchPoint();
+
+ if (CLI.NumResultRegs)
+ UpdateValueMap(I, CLI.ResultReg, CLI.NumResultRegs);
+ return true;
+}
+
+/// Returns an AttributeSet representing the attributes applied to the return
+/// value of the given call.
+static AttributeSet getReturnAttrs(FastISel::CallLoweringInfo &CLI) {
+ SmallVector<Attribute::AttrKind, 2> Attrs;
+ if (CLI.RetSExt)
+ Attrs.push_back(Attribute::SExt);
+ if (CLI.RetZExt)
+ Attrs.push_back(Attribute::ZExt);
+ if (CLI.IsInReg)
+ Attrs.push_back(Attribute::InReg);
+
+ return AttributeSet::get(CLI.RetTy->getContext(), AttributeSet::ReturnIndex,
+ Attrs);
+}
+
+bool FastISel::LowerCallTo(const CallInst *CI, const char *SymName,
+ unsigned NumArgs) {
+ ImmutableCallSite CS(CI);
+
+ PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
+ FunctionType *FTy = cast<FunctionType>(PT->getElementType());
+ Type *RetTy = FTy->getReturnType();
+
+ ArgListTy Args;
+ Args.reserve(NumArgs);
+
+ // Populate the argument list.
+ // Attributes for args start at offset 1, after the return attribute.
+ for (unsigned ArgI = 0; ArgI != NumArgs; ++ArgI) {
+ Value *V = CI->getOperand(ArgI);
+
+ assert(!V->getType()->isEmptyTy() && "Empty type passed to intrinsic.");
+
+ ArgListEntry Entry;
+ Entry.Val = V;
+ Entry.Ty = V->getType();
+ Entry.setAttributes(&CS, ArgI + 1);
+ Args.push_back(Entry);
+ }
+
+ CallLoweringInfo CLI;
+ CLI.setCallee(RetTy, FTy, SymName, std::move(Args), CS, NumArgs);
+
+ return LowerCallTo(CLI);
+}
+
+bool FastISel::LowerCallTo(CallLoweringInfo &CLI) {
+ // Handle the incoming return values from the call.
+ CLI.clearIns();
+ SmallVector<EVT, 4> RetTys;
+ ComputeValueVTs(TLI, CLI.RetTy, RetTys);
+
+ SmallVector<ISD::OutputArg, 4> Outs;
+ GetReturnInfo(CLI.RetTy, getReturnAttrs(CLI), Outs, TLI);
+
+ bool CanLowerReturn = TLI.CanLowerReturn(CLI.CallConv, *FuncInfo.MF,
+ CLI.IsVarArg, Outs,
+ CLI.RetTy->getContext());
+
+ // FIXME: sret demotion isn't supported yet - bail out.
+ if (!CanLowerReturn)
+ return false;
+
+ for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
+ EVT VT = RetTys[I];
+ MVT RegisterVT = TLI.getRegisterType(CLI.RetTy->getContext(), VT);
+ unsigned NumRegs = TLI.getNumRegisters(CLI.RetTy->getContext(), VT);
+ for (unsigned i = 0; i != NumRegs; ++i) {
+ ISD::InputArg MyFlags;
+ MyFlags.VT = RegisterVT;
+ MyFlags.ArgVT = VT;
+ MyFlags.Used = CLI.IsReturnValueUsed;
+ if (CLI.RetSExt)
+ MyFlags.Flags.setSExt();
+ if (CLI.RetZExt)
+ MyFlags.Flags.setZExt();
+ if (CLI.IsInReg)
+ MyFlags.Flags.setInReg();
+ CLI.Ins.push_back(MyFlags);
+ }
+ }
+
+ // Handle all of the outgoing arguments.
+ CLI.clearOuts();
+ for (auto &Arg : CLI.getArgs()) {
+ Type *FinalType = Arg.Ty;
+ if (Arg.isByVal)
+ FinalType = cast<PointerType>(Arg.Ty)->getElementType();
+ bool NeedsRegBlock = TLI.functionArgumentNeedsConsecutiveRegisters(
+ FinalType, CLI.CallConv, CLI.IsVarArg);
+
+ ISD::ArgFlagsTy Flags;
+ if (Arg.isZExt)
+ Flags.setZExt();
+ if (Arg.isSExt)
+ Flags.setSExt();
+ if (Arg.isInReg)
+ Flags.setInReg();
+ if (Arg.isSRet)
+ Flags.setSRet();
+ if (Arg.isByVal)
+ Flags.setByVal();
+ if (Arg.isInAlloca) {
+ Flags.setInAlloca();
+ // Set the byval flag for CCAssignFn callbacks that don't know about
+ // inalloca. This way we can know how many bytes we should've allocated
+ // and how many bytes a callee cleanup function will pop. If we port
+ // inalloca to more targets, we'll have to add custom inalloca handling in
+ // the various CC lowering callbacks.
+ Flags.setByVal();
+ }
+ if (Arg.isByVal || Arg.isInAlloca) {
+ PointerType *Ty = cast<PointerType>(Arg.Ty);
+ Type *ElementTy = Ty->getElementType();
+ unsigned FrameSize = DL.getTypeAllocSize(ElementTy);
+ // For ByVal, alignment should come from FE. BE will guess if this info is
+ // not there, but there are cases it cannot get right.
+ unsigned FrameAlign = Arg.Alignment;
+ if (!FrameAlign)
+ FrameAlign = TLI.getByValTypeAlignment(ElementTy);
+ Flags.setByValSize(FrameSize);
+ Flags.setByValAlign(FrameAlign);
+ }
+ if (Arg.isNest)
+ Flags.setNest();
+ if (NeedsRegBlock)
+ Flags.setInConsecutiveRegs();
+ unsigned OriginalAlignment = DL.getABITypeAlignment(Arg.Ty);
+ Flags.setOrigAlign(OriginalAlignment);
+
+ CLI.OutVals.push_back(Arg.Val);
+ CLI.OutFlags.push_back(Flags);
+ }
+
+ if (!FastLowerCall(CLI))
+ return false;
+
+ // Set all unused physreg defs as dead.
+ assert(CLI.Call && "No call instruction specified.");
+ CLI.Call->setPhysRegsDeadExcept(CLI.InRegs, TRI);
+
+ if (CLI.NumResultRegs && CLI.CS)
+ UpdateValueMap(CLI.CS->getInstruction(), CLI.ResultReg, CLI.NumResultRegs);
+
+ return true;
+}
+
+bool FastISel::LowerCall(const CallInst *CI) {
+ ImmutableCallSite CS(CI);
+
+ PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
+ FunctionType *FuncTy = cast<FunctionType>(PT->getElementType());
+ Type *RetTy = FuncTy->getReturnType();
+
+ ArgListTy Args;
+ ArgListEntry Entry;
+ Args.reserve(CS.arg_size());
+
+ for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
+ i != e; ++i) {
+ Value *V = *i;
+
+ // Skip empty types
+ if (V->getType()->isEmptyTy())
+ continue;
+
+ Entry.Val = V;
+ Entry.Ty = V->getType();
+
+ // Skip the first return-type Attribute to get to params.
+ Entry.setAttributes(&CS, i - CS.arg_begin() + 1);
+ Args.push_back(Entry);
+ }
+
+ // Check if target-independent constraints permit a tail call here.
+ // Target-dependent constraints are checked within FastLowerCall.
+ bool IsTailCall = CI->isTailCall();
+ if (IsTailCall && !isInTailCallPosition(CS, TM))
+ IsTailCall = false;
+
+ CallLoweringInfo CLI;
+ CLI.setCallee(RetTy, FuncTy, CI->getCalledValue(), std::move(Args), CS)
+ .setTailCall(IsTailCall);
+
+ return LowerCallTo(CLI);
+}
+
bool FastISel::SelectCall(const User *I) {
const CallInst *Call = cast<CallInst>(I);
// Handle simple inline asms.
if (const InlineAsm *IA = dyn_cast<InlineAsm>(Call->getCalledValue())) {
+ // If the inline asm has side effects, then make sure that no local value
+ // lives across by flushing the local value map.
+ if (IA->hasSideEffects())
+ flushLocalValueMap();
+
// Don't attempt to handle constraints.
if (!IA->getConstraintString().empty())
return false;
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);
MachineModuleInfo &MMI = FuncInfo.MF->getMMI();
ComputeUsesVAFloatArgument(*Call, &MMI);
- const Function *F = Call->getCalledFunction();
- if (!F) return false;
+ // Handle intrinsic function calls.
+ if (const auto *II = dyn_cast<IntrinsicInst>(Call))
+ return SelectIntrinsicCall(II);
- // Handle selected intrinsic function calls.
- switch (F->getIntrinsicID()) {
+ // Usually, it does not make sense to initialize a value,
+ // make an unrelated function call and use the value, because
+ // it tends to be spilled on the stack. So, we move the pointer
+ // to the last local value to the beginning of the block, so that
+ // 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.
+ flushLocalValueMap();
+
+ return LowerCall(Call);
+}
+
+bool FastISel::SelectIntrinsicCall(const IntrinsicInst *II) {
+ switch (II->getIntrinsicID()) {
default: break;
- // At -O0 we don't care about the lifetime intrinsics.
+ // At -O0 we don't care about the lifetime intrinsics.
case Intrinsic::lifetime_start:
case Intrinsic::lifetime_end:
- // The donothing intrinsic does, well, nothing.
+ // 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()) {
+ const DbgDeclareInst *DI = cast<DbgDeclareInst>(II);
+ 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;
}
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 (!Reg)
- Reg = lookUpRegForValue(Address);
+ 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.,
// 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) &&
+ if (!Op && !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
+ 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);
+ DEBUG(dbgs() << "Dropping debug info for " << *DI << "\n");
+ }
return true;
}
case Intrinsic::dbg_value: {
// This form of DBG_VALUE is target-independent.
- const DbgValueInst *DI = cast<DbgValueInst>(Call);
+ const DbgValueInst *DI = cast<DbgValueInst>(II);
const MCInstrDesc &II = TII.get(TargetOpcode::DBG_VALUE);
const Value *V = DI->getValue();
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)
+ 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);
+ DEBUG(dbgs() << "Dropping debug info for " << *DI << "\n");
}
return true;
}
case Intrinsic::objectsize: {
- ConstantInt *CI = cast<ConstantInt>(Call->getArgOperand(1));
+ ConstantInt *CI = cast<ConstantInt>(II->getArgOperand(1));
unsigned long long Res = CI->isZero() ? -1ULL : 0;
- Constant *ResCI = ConstantInt::get(Call->getType(), Res);
+ Constant *ResCI = ConstantInt::get(II->getType(), Res);
unsigned ResultReg = getRegForValue(ResCI);
if (ResultReg == 0)
return false;
- UpdateValueMap(Call, ResultReg);
+ UpdateValueMap(II, ResultReg);
return true;
}
+ case Intrinsic::expect: {
+ unsigned ResultReg = getRegForValue(II->getArgOperand(0));
+ if (ResultReg == 0)
+ return false;
+ UpdateValueMap(II, ResultReg);
+ return true;
+ }
+ case Intrinsic::experimental_stackmap:
+ return SelectStackmap(II);
+ case Intrinsic::experimental_patchpoint_void:
+ case Intrinsic::experimental_patchpoint_i64:
+ return SelectPatchpoint(II);
}
- // Usually, it does not make sense to initialize a value,
- // make an unrelated function call and use the value, because
- // it tends to be spilled on the stack. So, we move the pointer
- // to the last local value to the beginning of the block, so that
- // 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))
- flushLocalValueMap();
-
- // An arbitrary call. Bail.
- return false;
+ return FastLowerIntrinsicCall(II);
}
bool FastISel::SelectCast(const User *I, unsigned Opcode) {
}
// 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()) {
+ 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;
- // 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;
+
+ // 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
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();
+ // Undo phi node updates, because they will be added again by SelectionDAG.
+ if (isa<TerminatorInst>(I))
+ FuncInfo.PHINodesToUpdate.resize(FuncInfo.OrigNumPHINodesToUpdate);
return false;
}
/// unless it is the immediate (fall-through) successor, and update
/// the CFG.
void
-FastISel::FastEmitBranch(MachineBasicBlock *MSucc, DebugLoc DL) {
-
- if (FuncInfo.MBB->getBasicBlock()->size() > 1 && FuncInfo.MBB->isLayoutSuccessor(MSucc)) {
+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.
}
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.
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.getDataLayout()),
- TII(*TM.getInstrInfo()),
- TLI(*TM.getTargetLowering()),
- TRI(*TM.getRegisterInfo()),
- LibInfo(libInfo) {
-}
+ : FuncInfo(funcInfo), MF(funcInfo.MF), MRI(FuncInfo.MF->getRegInfo()),
+ MFI(*FuncInfo.MF->getFrameInfo()), MCP(*FuncInfo.MF->getConstantPool()),
+ TM(FuncInfo.MF->getTarget()), DL(*TM.getSubtargetImpl()->getDataLayout()),
+ TII(*TM.getSubtargetImpl()->getInstrInfo()),
+ TLI(*TM.getSubtargetImpl()->getTargetLowering()),
+ TRI(*TM.getSubtargetImpl()->getRegisterInfo()), LibInfo(libInfo) {}
FastISel::~FastISel() {}
+bool FastISel::FastLowerArguments() {
+ return false;
+}
+
+bool FastISel::FastLowerCall(CallLoweringInfo &/*CLI*/) {
+ return false;
+}
+
+bool FastISel::FastLowerIntrinsicCall(const IntrinsicInst * /*II*/) {
+ 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));
+ 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);
+ 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(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 Op0, bool Op0IsKill,
unsigned Op1, bool Op1IsKill,
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());
+ 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(Imm1).addImm(Imm2);
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(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 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;
}
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;
}
const TerminatorInst *TI = LLVMBB->getTerminator();
SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
- unsigned OrigNumPHINodesToUpdate = FuncInfo.PHINodesToUpdate.size();
+ FuncInfo.OrigNumPHINodesToUpdate = FuncInfo.PHINodesToUpdate.size();
// Check successor nodes' PHI nodes that expect a constant to be available
// from this block.
if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16)
VT = TLI.getTypeToTransformTo(LLVMBB->getContext(), VT);
else {
- FuncInfo.PHINodesToUpdate.resize(OrigNumPHINodesToUpdate);
+ FuncInfo.PHINodesToUpdate.resize(FuncInfo.OrigNumPHINodesToUpdate);
return false;
}
}
// 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) {
- FuncInfo.PHINodesToUpdate.resize(OrigNumPHINodesToUpdate);
+ if (!Reg) {
+ FuncInfo.PHINodesToUpdate.resize(FuncInfo.OrigNumPHINodesToUpdate);
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 *Ranges = I->getMetadata(LLVMContext::MD_range);
+
+ AAMDNodes AAInfo;
+ I->getAAMetadata(AAInfo);
+
+ if (Alignment == 0) // Ensure that codegen never sees alignment 0.
+ Alignment = DL.getABITypeAlignment(ValTy);
+
+ unsigned Size =
+ TM.getSubtargetImpl()->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, AAInfo, Ranges);
+}
projects / oota-llvm.git / blobdiff