llvm-6502/lib/Target/AArch64/AArch64FastISel.cpp
Tim Northover 46b3076cd0 AArch64: teach FastISel how to handle offset FrameIndices
Previously we were abandonning the attempt, leading to some combination of
extra work (when selection of a load/store fails completely) and inferior code
(when this leads to a real memcpy call instead of inlining).

rdar://problem/17187463

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@210520 91177308-0d34-0410-b5e6-96231b3b80d8
2014-06-10 09:52:44 +00:00

1983 lines
61 KiB
C++

//===-- AArch6464FastISel.cpp - AArch64 FastISel implementation -----------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file defines the AArch64-specific support for the FastISel class. Some
// of the target-specific code is generated by tablegen in the file
// AArch64GenFastISel.inc, which is #included here.
//
//===----------------------------------------------------------------------===//
#include "AArch64.h"
#include "AArch64TargetMachine.h"
#include "AArch64Subtarget.h"
#include "MCTargetDesc/AArch64AddressingModes.h"
#include "llvm/CodeGen/CallingConvLower.h"
#include "llvm/CodeGen/FastISel.h"
#include "llvm/CodeGen/FunctionLoweringInfo.h"
#include "llvm/CodeGen/MachineConstantPool.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/IR/CallingConv.h"
#include "llvm/IR/DataLayout.h"
#include "llvm/IR/DerivedTypes.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/GetElementPtrTypeIterator.h"
#include "llvm/IR/GlobalAlias.h"
#include "llvm/IR/GlobalVariable.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/Operator.h"
#include "llvm/Support/CommandLine.h"
using namespace llvm;
namespace {
class AArch64FastISel : public FastISel {
class Address {
public:
typedef enum {
RegBase,
FrameIndexBase
} BaseKind;
private:
BaseKind Kind;
union {
unsigned Reg;
int FI;
} Base;
int64_t Offset;
public:
Address() : Kind(RegBase), Offset(0) { Base.Reg = 0; }
void setKind(BaseKind K) { Kind = K; }
BaseKind getKind() const { return Kind; }
bool isRegBase() const { return Kind == RegBase; }
bool isFIBase() const { return Kind == FrameIndexBase; }
void setReg(unsigned Reg) {
assert(isRegBase() && "Invalid base register access!");
Base.Reg = Reg;
}
unsigned getReg() const {
assert(isRegBase() && "Invalid base register access!");
return Base.Reg;
}
void setFI(unsigned FI) {
assert(isFIBase() && "Invalid base frame index access!");
Base.FI = FI;
}
unsigned getFI() const {
assert(isFIBase() && "Invalid base frame index access!");
return Base.FI;
}
void setOffset(int64_t O) { Offset = O; }
int64_t getOffset() { return Offset; }
bool isValid() { return isFIBase() || (isRegBase() && getReg() != 0); }
};
/// Subtarget - Keep a pointer to the AArch64Subtarget around so that we can
/// make the right decision when generating code for different targets.
const AArch64Subtarget *Subtarget;
LLVMContext *Context;
private:
// Selection routines.
bool SelectLoad(const Instruction *I);
bool SelectStore(const Instruction *I);
bool SelectBranch(const Instruction *I);
bool SelectIndirectBr(const Instruction *I);
bool SelectCmp(const Instruction *I);
bool SelectSelect(const Instruction *I);
bool SelectFPExt(const Instruction *I);
bool SelectFPTrunc(const Instruction *I);
bool SelectFPToInt(const Instruction *I, bool Signed);
bool SelectIntToFP(const Instruction *I, bool Signed);
bool SelectRem(const Instruction *I, unsigned ISDOpcode);
bool SelectCall(const Instruction *I, const char *IntrMemName);
bool SelectIntrinsicCall(const IntrinsicInst &I);
bool SelectRet(const Instruction *I);
bool SelectTrunc(const Instruction *I);
bool SelectIntExt(const Instruction *I);
bool SelectMul(const Instruction *I);
// Utility helper routines.
bool isTypeLegal(Type *Ty, MVT &VT);
bool isLoadStoreTypeLegal(Type *Ty, MVT &VT);
bool ComputeAddress(const Value *Obj, Address &Addr);
bool SimplifyAddress(Address &Addr, MVT VT, int64_t ScaleFactor,
bool UseUnscaled);
void AddLoadStoreOperands(Address &Addr, const MachineInstrBuilder &MIB,
unsigned Flags, bool UseUnscaled);
bool IsMemCpySmall(uint64_t Len, unsigned Alignment);
bool TryEmitSmallMemCpy(Address Dest, Address Src, uint64_t Len,
unsigned Alignment);
// Emit functions.
bool EmitCmp(Value *Src1Value, Value *Src2Value, bool isZExt);
bool EmitLoad(MVT VT, unsigned &ResultReg, Address Addr,
bool UseUnscaled = false);
bool EmitStore(MVT VT, unsigned SrcReg, Address Addr,
bool UseUnscaled = false);
unsigned EmitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, bool isZExt);
unsigned Emiti1Ext(unsigned SrcReg, MVT DestVT, bool isZExt);
unsigned AArch64MaterializeFP(const ConstantFP *CFP, MVT VT);
unsigned AArch64MaterializeGV(const GlobalValue *GV);
// Call handling routines.
private:
CCAssignFn *CCAssignFnForCall(CallingConv::ID CC) const;
bool ProcessCallArgs(SmallVectorImpl<Value *> &Args,
SmallVectorImpl<unsigned> &ArgRegs,
SmallVectorImpl<MVT> &ArgVTs,
SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
SmallVectorImpl<unsigned> &RegArgs, CallingConv::ID CC,
unsigned &NumBytes);
bool FinishCall(MVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
const Instruction *I, CallingConv::ID CC, unsigned &NumBytes);
public:
// Backend specific FastISel code.
unsigned TargetMaterializeAlloca(const AllocaInst *AI) override;
unsigned TargetMaterializeConstant(const Constant *C) override;
explicit AArch64FastISel(FunctionLoweringInfo &funcInfo,
const TargetLibraryInfo *libInfo)
: FastISel(funcInfo, libInfo) {
Subtarget = &TM.getSubtarget<AArch64Subtarget>();
Context = &funcInfo.Fn->getContext();
}
bool TargetSelectInstruction(const Instruction *I) override;
#include "AArch64GenFastISel.inc"
};
} // end anonymous namespace
#include "AArch64GenCallingConv.inc"
CCAssignFn *AArch64FastISel::CCAssignFnForCall(CallingConv::ID CC) const {
if (CC == CallingConv::WebKit_JS)
return CC_AArch64_WebKit_JS;
return Subtarget->isTargetDarwin() ? CC_AArch64_DarwinPCS : CC_AArch64_AAPCS;
}
unsigned AArch64FastISel::TargetMaterializeAlloca(const AllocaInst *AI) {
assert(TLI.getValueType(AI->getType(), true) == MVT::i64 &&
"Alloca should always return a pointer.");
// Don't handle dynamic allocas.
if (!FuncInfo.StaticAllocaMap.count(AI))
return 0;
DenseMap<const AllocaInst *, int>::iterator SI =
FuncInfo.StaticAllocaMap.find(AI);
if (SI != FuncInfo.StaticAllocaMap.end()) {
unsigned ResultReg = createResultReg(&AArch64::GPR64RegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri),
ResultReg)
.addFrameIndex(SI->second)
.addImm(0)
.addImm(0);
return ResultReg;
}
return 0;
}
unsigned AArch64FastISel::AArch64MaterializeFP(const ConstantFP *CFP, MVT VT) {
if (VT != MVT::f32 && VT != MVT::f64)
return 0;
const APFloat Val = CFP->getValueAPF();
bool is64bit = (VT == MVT::f64);
// This checks to see if we can use FMOV instructions to materialize
// a constant, otherwise we have to materialize via the constant pool.
if (TLI.isFPImmLegal(Val, VT)) {
int Imm;
unsigned Opc;
if (is64bit) {
Imm = AArch64_AM::getFP64Imm(Val);
Opc = AArch64::FMOVDi;
} else {
Imm = AArch64_AM::getFP32Imm(Val);
Opc = AArch64::FMOVSi;
}
unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
.addImm(Imm);
return ResultReg;
}
// Materialize via constant pool. MachineConstantPool wants an explicit
// alignment.
unsigned Align = DL.getPrefTypeAlignment(CFP->getType());
if (Align == 0)
Align = DL.getTypeAllocSize(CFP->getType());
unsigned Idx = MCP.getConstantPoolIndex(cast<Constant>(CFP), Align);
unsigned ADRPReg = createResultReg(&AArch64::GPR64commonRegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
ADRPReg).addConstantPoolIndex(Idx, 0, AArch64II::MO_PAGE);
unsigned Opc = is64bit ? AArch64::LDRDui : AArch64::LDRSui;
unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
.addReg(ADRPReg)
.addConstantPoolIndex(Idx, 0, AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
return ResultReg;
}
unsigned AArch64FastISel::AArch64MaterializeGV(const GlobalValue *GV) {
// We can't handle thread-local variables quickly yet.
if (GV->isThreadLocal())
return 0;
// MachO still uses GOT for large code-model accesses, but ELF requires
// movz/movk sequences, which FastISel doesn't handle yet.
if (TM.getCodeModel() != CodeModel::Small && !Subtarget->isTargetMachO())
return 0;
unsigned char OpFlags = Subtarget->ClassifyGlobalReference(GV, TM);
EVT DestEVT = TLI.getValueType(GV->getType(), true);
if (!DestEVT.isSimple())
return 0;
unsigned ADRPReg = createResultReg(&AArch64::GPR64commonRegClass);
unsigned ResultReg;
if (OpFlags & AArch64II::MO_GOT) {
// ADRP + LDRX
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
ADRPReg)
.addGlobalAddress(GV, 0, AArch64II::MO_GOT | AArch64II::MO_PAGE);
ResultReg = createResultReg(&AArch64::GPR64RegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::LDRXui),
ResultReg)
.addReg(ADRPReg)
.addGlobalAddress(GV, 0, AArch64II::MO_GOT | AArch64II::MO_PAGEOFF |
AArch64II::MO_NC);
} else {
// ADRP + ADDX
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
ADRPReg).addGlobalAddress(GV, 0, AArch64II::MO_PAGE);
ResultReg = createResultReg(&AArch64::GPR64spRegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri),
ResultReg)
.addReg(ADRPReg)
.addGlobalAddress(GV, 0, AArch64II::MO_PAGEOFF | AArch64II::MO_NC)
.addImm(0);
}
return ResultReg;
}
unsigned AArch64FastISel::TargetMaterializeConstant(const Constant *C) {
EVT CEVT = TLI.getValueType(C->getType(), true);
// Only handle simple types.
if (!CEVT.isSimple())
return 0;
MVT VT = CEVT.getSimpleVT();
// FIXME: Handle ConstantInt.
if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
return AArch64MaterializeFP(CFP, VT);
else if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
return AArch64MaterializeGV(GV);
return 0;
}
// Computes the address to get to an object.
bool AArch64FastISel::ComputeAddress(const Value *Obj, Address &Addr) {
const User *U = nullptr;
unsigned Opcode = Instruction::UserOp1;
if (const Instruction *I = dyn_cast<Instruction>(Obj)) {
// Don't walk into other basic blocks unless the object is an alloca from
// another block, otherwise it may not have a virtual register assigned.
if (FuncInfo.StaticAllocaMap.count(static_cast<const AllocaInst *>(Obj)) ||
FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
Opcode = I->getOpcode();
U = I;
}
} else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Obj)) {
Opcode = C->getOpcode();
U = C;
}
if (const PointerType *Ty = dyn_cast<PointerType>(Obj->getType()))
if (Ty->getAddressSpace() > 255)
// Fast instruction selection doesn't support the special
// address spaces.
return false;
switch (Opcode) {
default:
break;
case Instruction::BitCast: {
// Look through bitcasts.
return ComputeAddress(U->getOperand(0), Addr);
}
case Instruction::IntToPtr: {
// Look past no-op inttoptrs.
if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
return ComputeAddress(U->getOperand(0), Addr);
break;
}
case Instruction::PtrToInt: {
// Look past no-op ptrtoints.
if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
return ComputeAddress(U->getOperand(0), Addr);
break;
}
case Instruction::GetElementPtr: {
Address SavedAddr = Addr;
uint64_t TmpOffset = Addr.getOffset();
// Iterate through the GEP folding the constants into offsets where
// we can.
gep_type_iterator GTI = gep_type_begin(U);
for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e;
++i, ++GTI) {
const Value *Op = *i;
if (StructType *STy = dyn_cast<StructType>(*GTI)) {
const StructLayout *SL = DL.getStructLayout(STy);
unsigned Idx = cast<ConstantInt>(Op)->getZExtValue();
TmpOffset += SL->getElementOffset(Idx);
} else {
uint64_t S = DL.getTypeAllocSize(GTI.getIndexedType());
for (;;) {
if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
// Constant-offset addressing.
TmpOffset += CI->getSExtValue() * S;
break;
}
if (canFoldAddIntoGEP(U, Op)) {
// A compatible add with a constant operand. Fold the constant.
ConstantInt *CI =
cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1));
TmpOffset += CI->getSExtValue() * S;
// Iterate on the other operand.
Op = cast<AddOperator>(Op)->getOperand(0);
continue;
}
// Unsupported
goto unsupported_gep;
}
}
}
// Try to grab the base operand now.
Addr.setOffset(TmpOffset);
if (ComputeAddress(U->getOperand(0), Addr))
return true;
// We failed, restore everything and try the other options.
Addr = SavedAddr;
unsupported_gep:
break;
}
case Instruction::Alloca: {
const AllocaInst *AI = cast<AllocaInst>(Obj);
DenseMap<const AllocaInst *, int>::iterator SI =
FuncInfo.StaticAllocaMap.find(AI);
if (SI != FuncInfo.StaticAllocaMap.end()) {
Addr.setKind(Address::FrameIndexBase);
Addr.setFI(SI->second);
return true;
}
break;
}
}
// Try to get this in a register if nothing else has worked.
if (!Addr.isValid())
Addr.setReg(getRegForValue(Obj));
return Addr.isValid();
}
bool AArch64FastISel::isTypeLegal(Type *Ty, MVT &VT) {
EVT evt = TLI.getValueType(Ty, true);
// Only handle simple types.
if (evt == MVT::Other || !evt.isSimple())
return false;
VT = evt.getSimpleVT();
// This is a legal type, but it's not something we handle in fast-isel.
if (VT == MVT::f128)
return false;
// Handle all other legal types, i.e. a register that will directly hold this
// value.
return TLI.isTypeLegal(VT);
}
bool AArch64FastISel::isLoadStoreTypeLegal(Type *Ty, MVT &VT) {
if (isTypeLegal(Ty, VT))
return true;
// If this is a type than can be sign or zero-extended to a basic operation
// go ahead and accept it now. For stores, this reflects truncation.
if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16)
return true;
return false;
}
bool AArch64FastISel::SimplifyAddress(Address &Addr, MVT VT,
int64_t ScaleFactor, bool UseUnscaled) {
bool needsLowering = false;
int64_t Offset = Addr.getOffset();
switch (VT.SimpleTy) {
default:
return false;
case MVT::i1:
case MVT::i8:
case MVT::i16:
case MVT::i32:
case MVT::i64:
case MVT::f32:
case MVT::f64:
if (!UseUnscaled)
// Using scaled, 12-bit, unsigned immediate offsets.
needsLowering = ((Offset & 0xfff) != Offset);
else
// Using unscaled, 9-bit, signed immediate offsets.
needsLowering = (Offset > 256 || Offset < -256);
break;
}
//If this is a stack pointer and the offset needs to be simplified then put
// the alloca address into a register, set the base type back to register and
// continue. This should almost never happen.
if (needsLowering && Addr.getKind() == Address::FrameIndexBase) {
unsigned ResultReg = createResultReg(&AArch64::GPR64RegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri),
ResultReg)
.addFrameIndex(Addr.getFI())
.addImm(0)
.addImm(0);
Addr.setKind(Address::RegBase);
Addr.setReg(ResultReg);
}
// Since the offset is too large for the load/store instruction get the
// reg+offset into a register.
if (needsLowering) {
uint64_t UnscaledOffset = Addr.getOffset() * ScaleFactor;
unsigned ResultReg = FastEmit_ri_(MVT::i64, ISD::ADD, Addr.getReg(), false,
UnscaledOffset, MVT::i64);
if (ResultReg == 0)
return false;
Addr.setReg(ResultReg);
Addr.setOffset(0);
}
return true;
}
void AArch64FastISel::AddLoadStoreOperands(Address &Addr,
const MachineInstrBuilder &MIB,
unsigned Flags, bool UseUnscaled) {
int64_t Offset = Addr.getOffset();
// Frame base works a bit differently. Handle it separately.
if (Addr.getKind() == Address::FrameIndexBase) {
int FI = Addr.getFI();
// FIXME: We shouldn't be using getObjectSize/getObjectAlignment. The size
// and alignment should be based on the VT.
MachineMemOperand *MMO = FuncInfo.MF->getMachineMemOperand(
MachinePointerInfo::getFixedStack(FI, Offset), Flags,
MFI.getObjectSize(FI), MFI.getObjectAlignment(FI));
// Now add the rest of the operands.
MIB.addFrameIndex(FI).addImm(Offset).addMemOperand(MMO);
} else {
// Now add the rest of the operands.
MIB.addReg(Addr.getReg());
MIB.addImm(Offset);
}
}
bool AArch64FastISel::EmitLoad(MVT VT, unsigned &ResultReg, Address Addr,
bool UseUnscaled) {
// Negative offsets require unscaled, 9-bit, signed immediate offsets.
// Otherwise, we try using scaled, 12-bit, unsigned immediate offsets.
if (!UseUnscaled && Addr.getOffset() < 0)
UseUnscaled = true;
unsigned Opc;
const TargetRegisterClass *RC;
bool VTIsi1 = false;
int64_t ScaleFactor = 0;
switch (VT.SimpleTy) {
default:
return false;
case MVT::i1:
VTIsi1 = true;
// Intentional fall-through.
case MVT::i8:
Opc = UseUnscaled ? AArch64::LDURBBi : AArch64::LDRBBui;
RC = &AArch64::GPR32RegClass;
ScaleFactor = 1;
break;
case MVT::i16:
Opc = UseUnscaled ? AArch64::LDURHHi : AArch64::LDRHHui;
RC = &AArch64::GPR32RegClass;
ScaleFactor = 2;
break;
case MVT::i32:
Opc = UseUnscaled ? AArch64::LDURWi : AArch64::LDRWui;
RC = &AArch64::GPR32RegClass;
ScaleFactor = 4;
break;
case MVT::i64:
Opc = UseUnscaled ? AArch64::LDURXi : AArch64::LDRXui;
RC = &AArch64::GPR64RegClass;
ScaleFactor = 8;
break;
case MVT::f32:
Opc = UseUnscaled ? AArch64::LDURSi : AArch64::LDRSui;
RC = TLI.getRegClassFor(VT);
ScaleFactor = 4;
break;
case MVT::f64:
Opc = UseUnscaled ? AArch64::LDURDi : AArch64::LDRDui;
RC = TLI.getRegClassFor(VT);
ScaleFactor = 8;
break;
}
// Scale the offset.
if (!UseUnscaled) {
int64_t Offset = Addr.getOffset();
if (Offset & (ScaleFactor - 1))
// Retry using an unscaled, 9-bit, signed immediate offset.
return EmitLoad(VT, ResultReg, Addr, /*UseUnscaled*/ true);
Addr.setOffset(Offset / ScaleFactor);
}
// Simplify this down to something we can handle.
if (!SimplifyAddress(Addr, VT, UseUnscaled ? 1 : ScaleFactor, UseUnscaled))
return false;
// Create the base instruction, then add the operands.
ResultReg = createResultReg(RC);
MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
TII.get(Opc), ResultReg);
AddLoadStoreOperands(Addr, MIB, MachineMemOperand::MOLoad, UseUnscaled);
// Loading an i1 requires special handling.
if (VTIsi1) {
MRI.constrainRegClass(ResultReg, &AArch64::GPR32RegClass);
unsigned ANDReg = createResultReg(&AArch64::GPR32spRegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ANDWri),
ANDReg)
.addReg(ResultReg)
.addImm(AArch64_AM::encodeLogicalImmediate(1, 32));
ResultReg = ANDReg;
}
return true;
}
bool AArch64FastISel::SelectLoad(const Instruction *I) {
MVT VT;
// Verify we have a legal type before going any further. Currently, we handle
// simple types that will directly fit in a register (i32/f32/i64/f64) or
// those that can be sign or zero-extended to a basic operation (i1/i8/i16).
if (!isLoadStoreTypeLegal(I->getType(), VT) || cast<LoadInst>(I)->isAtomic())
return false;
// See if we can handle this address.
Address Addr;
if (!ComputeAddress(I->getOperand(0), Addr))
return false;
unsigned ResultReg;
if (!EmitLoad(VT, ResultReg, Addr))
return false;
UpdateValueMap(I, ResultReg);
return true;
}
bool AArch64FastISel::EmitStore(MVT VT, unsigned SrcReg, Address Addr,
bool UseUnscaled) {
// Negative offsets require unscaled, 9-bit, signed immediate offsets.
// Otherwise, we try using scaled, 12-bit, unsigned immediate offsets.
if (!UseUnscaled && Addr.getOffset() < 0)
UseUnscaled = true;
unsigned StrOpc;
bool VTIsi1 = false;
int64_t ScaleFactor = 0;
// Using scaled, 12-bit, unsigned immediate offsets.
switch (VT.SimpleTy) {
default:
return false;
case MVT::i1:
VTIsi1 = true;
case MVT::i8:
StrOpc = UseUnscaled ? AArch64::STURBBi : AArch64::STRBBui;
ScaleFactor = 1;
break;
case MVT::i16:
StrOpc = UseUnscaled ? AArch64::STURHHi : AArch64::STRHHui;
ScaleFactor = 2;
break;
case MVT::i32:
StrOpc = UseUnscaled ? AArch64::STURWi : AArch64::STRWui;
ScaleFactor = 4;
break;
case MVT::i64:
StrOpc = UseUnscaled ? AArch64::STURXi : AArch64::STRXui;
ScaleFactor = 8;
break;
case MVT::f32:
StrOpc = UseUnscaled ? AArch64::STURSi : AArch64::STRSui;
ScaleFactor = 4;
break;
case MVT::f64:
StrOpc = UseUnscaled ? AArch64::STURDi : AArch64::STRDui;
ScaleFactor = 8;
break;
}
// Scale the offset.
if (!UseUnscaled) {
int64_t Offset = Addr.getOffset();
if (Offset & (ScaleFactor - 1))
// Retry using an unscaled, 9-bit, signed immediate offset.
return EmitStore(VT, SrcReg, Addr, /*UseUnscaled*/ true);
Addr.setOffset(Offset / ScaleFactor);
}
// Simplify this down to something we can handle.
if (!SimplifyAddress(Addr, VT, UseUnscaled ? 1 : ScaleFactor, UseUnscaled))
return false;
// Storing an i1 requires special handling.
if (VTIsi1) {
MRI.constrainRegClass(SrcReg, &AArch64::GPR32RegClass);
unsigned ANDReg = createResultReg(&AArch64::GPR32spRegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ANDWri),
ANDReg)
.addReg(SrcReg)
.addImm(AArch64_AM::encodeLogicalImmediate(1, 32));
SrcReg = ANDReg;
}
// Create the base instruction, then add the operands.
MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
TII.get(StrOpc)).addReg(SrcReg);
AddLoadStoreOperands(Addr, MIB, MachineMemOperand::MOStore, UseUnscaled);
return true;
}
bool AArch64FastISel::SelectStore(const Instruction *I) {
MVT VT;
Value *Op0 = I->getOperand(0);
// Verify we have a legal type before going any further. Currently, we handle
// simple types that will directly fit in a register (i32/f32/i64/f64) or
// those that can be sign or zero-extended to a basic operation (i1/i8/i16).
if (!isLoadStoreTypeLegal(Op0->getType(), VT) ||
cast<StoreInst>(I)->isAtomic())
return false;
// Get the value to be stored into a register.
unsigned SrcReg = getRegForValue(Op0);
if (SrcReg == 0)
return false;
// See if we can handle this address.
Address Addr;
if (!ComputeAddress(I->getOperand(1), Addr))
return false;
if (!EmitStore(VT, SrcReg, Addr))
return false;
return true;
}
static AArch64CC::CondCode getCompareCC(CmpInst::Predicate Pred) {
switch (Pred) {
case CmpInst::FCMP_ONE:
case CmpInst::FCMP_UEQ:
default:
// AL is our "false" for now. The other two need more compares.
return AArch64CC::AL;
case CmpInst::ICMP_EQ:
case CmpInst::FCMP_OEQ:
return AArch64CC::EQ;
case CmpInst::ICMP_SGT:
case CmpInst::FCMP_OGT:
return AArch64CC::GT;
case CmpInst::ICMP_SGE:
case CmpInst::FCMP_OGE:
return AArch64CC::GE;
case CmpInst::ICMP_UGT:
case CmpInst::FCMP_UGT:
return AArch64CC::HI;
case CmpInst::FCMP_OLT:
return AArch64CC::MI;
case CmpInst::ICMP_ULE:
case CmpInst::FCMP_OLE:
return AArch64CC::LS;
case CmpInst::FCMP_ORD:
return AArch64CC::VC;
case CmpInst::FCMP_UNO:
return AArch64CC::VS;
case CmpInst::FCMP_UGE:
return AArch64CC::PL;
case CmpInst::ICMP_SLT:
case CmpInst::FCMP_ULT:
return AArch64CC::LT;
case CmpInst::ICMP_SLE:
case CmpInst::FCMP_ULE:
return AArch64CC::LE;
case CmpInst::FCMP_UNE:
case CmpInst::ICMP_NE:
return AArch64CC::NE;
case CmpInst::ICMP_UGE:
return AArch64CC::HS;
case CmpInst::ICMP_ULT:
return AArch64CC::LO;
}
}
bool AArch64FastISel::SelectBranch(const Instruction *I) {
const BranchInst *BI = cast<BranchInst>(I);
MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
if (CI->hasOneUse() && (CI->getParent() == I->getParent())) {
// We may not handle every CC for now.
AArch64CC::CondCode CC = getCompareCC(CI->getPredicate());
if (CC == AArch64CC::AL)
return false;
// Emit the cmp.
if (!EmitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
return false;
// Emit the branch.
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
.addImm(CC)
.addMBB(TBB);
FuncInfo.MBB->addSuccessor(TBB);
FastEmitBranch(FBB, DbgLoc);
return true;
}
} else if (TruncInst *TI = dyn_cast<TruncInst>(BI->getCondition())) {
MVT SrcVT;
if (TI->hasOneUse() && TI->getParent() == I->getParent() &&
(isLoadStoreTypeLegal(TI->getOperand(0)->getType(), SrcVT))) {
unsigned CondReg = getRegForValue(TI->getOperand(0));
if (CondReg == 0)
return false;
// Issue an extract_subreg to get the lower 32-bits.
if (SrcVT == MVT::i64)
CondReg = FastEmitInst_extractsubreg(MVT::i32, CondReg, /*Kill=*/true,
AArch64::sub_32);
MRI.constrainRegClass(CondReg, &AArch64::GPR32RegClass);
unsigned ANDReg = createResultReg(&AArch64::GPR32spRegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
TII.get(AArch64::ANDWri), ANDReg)
.addReg(CondReg)
.addImm(AArch64_AM::encodeLogicalImmediate(1, 32));
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
TII.get(AArch64::SUBSWri))
.addReg(ANDReg)
.addReg(ANDReg)
.addImm(0)
.addImm(0);
unsigned CC = AArch64CC::NE;
if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
std::swap(TBB, FBB);
CC = AArch64CC::EQ;
}
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
.addImm(CC)
.addMBB(TBB);
FuncInfo.MBB->addSuccessor(TBB);
FastEmitBranch(FBB, DbgLoc);
return true;
}
} else if (const ConstantInt *CI =
dyn_cast<ConstantInt>(BI->getCondition())) {
uint64_t Imm = CI->getZExtValue();
MachineBasicBlock *Target = (Imm == 0) ? FBB : TBB;
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::B))
.addMBB(Target);
FuncInfo.MBB->addSuccessor(Target);
return true;
}
unsigned CondReg = getRegForValue(BI->getCondition());
if (CondReg == 0)
return false;
// We've been divorced from our compare! Our block was split, and
// now our compare lives in a predecessor block. We musn't
// re-compare here, as the children of the compare aren't guaranteed
// live across the block boundary (we *could* check for this).
// Regardless, the compare has been done in the predecessor block,
// and it left a value for us in a virtual register. Ergo, we test
// the one-bit value left in the virtual register.
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::SUBSWri),
AArch64::WZR)
.addReg(CondReg)
.addImm(0)
.addImm(0);
unsigned CC = AArch64CC::NE;
if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
std::swap(TBB, FBB);
CC = AArch64CC::EQ;
}
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
.addImm(CC)
.addMBB(TBB);
FuncInfo.MBB->addSuccessor(TBB);
FastEmitBranch(FBB, DbgLoc);
return true;
}
bool AArch64FastISel::SelectIndirectBr(const Instruction *I) {
const IndirectBrInst *BI = cast<IndirectBrInst>(I);
unsigned AddrReg = getRegForValue(BI->getOperand(0));
if (AddrReg == 0)
return false;
// Emit the indirect branch.
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::BR))
.addReg(AddrReg);
// Make sure the CFG is up-to-date.
for (unsigned i = 0, e = BI->getNumSuccessors(); i != e; ++i)
FuncInfo.MBB->addSuccessor(FuncInfo.MBBMap[BI->getSuccessor(i)]);
return true;
}
bool AArch64FastISel::EmitCmp(Value *Src1Value, Value *Src2Value, bool isZExt) {
Type *Ty = Src1Value->getType();
EVT SrcEVT = TLI.getValueType(Ty, true);
if (!SrcEVT.isSimple())
return false;
MVT SrcVT = SrcEVT.getSimpleVT();
// Check to see if the 2nd operand is a constant that we can encode directly
// in the compare.
uint64_t Imm;
bool UseImm = false;
bool isNegativeImm = false;
if (const ConstantInt *ConstInt = dyn_cast<ConstantInt>(Src2Value)) {
if (SrcVT == MVT::i64 || SrcVT == MVT::i32 || SrcVT == MVT::i16 ||
SrcVT == MVT::i8 || SrcVT == MVT::i1) {
const APInt &CIVal = ConstInt->getValue();
Imm = (isZExt) ? CIVal.getZExtValue() : CIVal.getSExtValue();
if (CIVal.isNegative()) {
isNegativeImm = true;
Imm = -Imm;
}
// FIXME: We can handle more immediates using shifts.
UseImm = ((Imm & 0xfff) == Imm);
}
} else if (const ConstantFP *ConstFP = dyn_cast<ConstantFP>(Src2Value)) {
if (SrcVT == MVT::f32 || SrcVT == MVT::f64)
if (ConstFP->isZero() && !ConstFP->isNegative())
UseImm = true;
}
unsigned ZReg;
unsigned CmpOpc;
bool isICmp = true;
bool needsExt = false;
switch (SrcVT.SimpleTy) {
default:
return false;
case MVT::i1:
case MVT::i8:
case MVT::i16:
needsExt = true;
// Intentional fall-through.
case MVT::i32:
ZReg = AArch64::WZR;
if (UseImm)
CmpOpc = isNegativeImm ? AArch64::ADDSWri : AArch64::SUBSWri;
else
CmpOpc = AArch64::SUBSWrr;
break;
case MVT::i64:
ZReg = AArch64::XZR;
if (UseImm)
CmpOpc = isNegativeImm ? AArch64::ADDSXri : AArch64::SUBSXri;
else
CmpOpc = AArch64::SUBSXrr;
break;
case MVT::f32:
isICmp = false;
CmpOpc = UseImm ? AArch64::FCMPSri : AArch64::FCMPSrr;
break;
case MVT::f64:
isICmp = false;
CmpOpc = UseImm ? AArch64::FCMPDri : AArch64::FCMPDrr;
break;
}
unsigned SrcReg1 = getRegForValue(Src1Value);
if (SrcReg1 == 0)
return false;
unsigned SrcReg2;
if (!UseImm) {
SrcReg2 = getRegForValue(Src2Value);
if (SrcReg2 == 0)
return false;
}
// We have i1, i8, or i16, we need to either zero extend or sign extend.
if (needsExt) {
SrcReg1 = EmitIntExt(SrcVT, SrcReg1, MVT::i32, isZExt);
if (SrcReg1 == 0)
return false;
if (!UseImm) {
SrcReg2 = EmitIntExt(SrcVT, SrcReg2, MVT::i32, isZExt);
if (SrcReg2 == 0)
return false;
}
}
if (isICmp) {
if (UseImm)
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(CmpOpc))
.addReg(ZReg)
.addReg(SrcReg1)
.addImm(Imm)
.addImm(0);
else
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(CmpOpc))
.addReg(ZReg)
.addReg(SrcReg1)
.addReg(SrcReg2);
} else {
if (UseImm)
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(CmpOpc))
.addReg(SrcReg1);
else
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(CmpOpc))
.addReg(SrcReg1)
.addReg(SrcReg2);
}
return true;
}
bool AArch64FastISel::SelectCmp(const Instruction *I) {
const CmpInst *CI = cast<CmpInst>(I);
// We may not handle every CC for now.
AArch64CC::CondCode CC = getCompareCC(CI->getPredicate());
if (CC == AArch64CC::AL)
return false;
// Emit the cmp.
if (!EmitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
return false;
// Now set a register based on the comparison.
AArch64CC::CondCode invertedCC = getInvertedCondCode(CC);
unsigned ResultReg = createResultReg(&AArch64::GPR32RegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::CSINCWr),
ResultReg)
.addReg(AArch64::WZR)
.addReg(AArch64::WZR)
.addImm(invertedCC);
UpdateValueMap(I, ResultReg);
return true;
}
bool AArch64FastISel::SelectSelect(const Instruction *I) {
const SelectInst *SI = cast<SelectInst>(I);
EVT DestEVT = TLI.getValueType(SI->getType(), true);
if (!DestEVT.isSimple())
return false;
MVT DestVT = DestEVT.getSimpleVT();
if (DestVT != MVT::i32 && DestVT != MVT::i64 && DestVT != MVT::f32 &&
DestVT != MVT::f64)
return false;
unsigned CondReg = getRegForValue(SI->getCondition());
if (CondReg == 0)
return false;
unsigned TrueReg = getRegForValue(SI->getTrueValue());
if (TrueReg == 0)
return false;
unsigned FalseReg = getRegForValue(SI->getFalseValue());
if (FalseReg == 0)
return false;
MRI.constrainRegClass(CondReg, &AArch64::GPR32RegClass);
unsigned ANDReg = createResultReg(&AArch64::GPR32spRegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ANDWri),
ANDReg)
.addReg(CondReg)
.addImm(AArch64_AM::encodeLogicalImmediate(1, 32));
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::SUBSWri))
.addReg(ANDReg)
.addReg(ANDReg)
.addImm(0)
.addImm(0);
unsigned SelectOpc;
switch (DestVT.SimpleTy) {
default:
return false;
case MVT::i32:
SelectOpc = AArch64::CSELWr;
break;
case MVT::i64:
SelectOpc = AArch64::CSELXr;
break;
case MVT::f32:
SelectOpc = AArch64::FCSELSrrr;
break;
case MVT::f64:
SelectOpc = AArch64::FCSELDrrr;
break;
}
unsigned ResultReg = createResultReg(TLI.getRegClassFor(DestVT));
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(SelectOpc),
ResultReg)
.addReg(TrueReg)
.addReg(FalseReg)
.addImm(AArch64CC::NE);
UpdateValueMap(I, ResultReg);
return true;
}
bool AArch64FastISel::SelectFPExt(const Instruction *I) {
Value *V = I->getOperand(0);
if (!I->getType()->isDoubleTy() || !V->getType()->isFloatTy())
return false;
unsigned Op = getRegForValue(V);
if (Op == 0)
return false;
unsigned ResultReg = createResultReg(&AArch64::FPR64RegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::FCVTDSr),
ResultReg).addReg(Op);
UpdateValueMap(I, ResultReg);
return true;
}
bool AArch64FastISel::SelectFPTrunc(const Instruction *I) {
Value *V = I->getOperand(0);
if (!I->getType()->isFloatTy() || !V->getType()->isDoubleTy())
return false;
unsigned Op = getRegForValue(V);
if (Op == 0)
return false;
unsigned ResultReg = createResultReg(&AArch64::FPR32RegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::FCVTSDr),
ResultReg).addReg(Op);
UpdateValueMap(I, ResultReg);
return true;
}
// FPToUI and FPToSI
bool AArch64FastISel::SelectFPToInt(const Instruction *I, bool Signed) {
MVT DestVT;
if (!isTypeLegal(I->getType(), DestVT) || DestVT.isVector())
return false;
unsigned SrcReg = getRegForValue(I->getOperand(0));
if (SrcReg == 0)
return false;
EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType(), true);
if (SrcVT == MVT::f128)
return false;
unsigned Opc;
if (SrcVT == MVT::f64) {
if (Signed)
Opc = (DestVT == MVT::i32) ? AArch64::FCVTZSUWDr : AArch64::FCVTZSUXDr;
else
Opc = (DestVT == MVT::i32) ? AArch64::FCVTZUUWDr : AArch64::FCVTZUUXDr;
} else {
if (Signed)
Opc = (DestVT == MVT::i32) ? AArch64::FCVTZSUWSr : AArch64::FCVTZSUXSr;
else
Opc = (DestVT == MVT::i32) ? AArch64::FCVTZUUWSr : AArch64::FCVTZUUXSr;
}
unsigned ResultReg = createResultReg(
DestVT == MVT::i32 ? &AArch64::GPR32RegClass : &AArch64::GPR64RegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
.addReg(SrcReg);
UpdateValueMap(I, ResultReg);
return true;
}
bool AArch64FastISel::SelectIntToFP(const Instruction *I, bool Signed) {
MVT DestVT;
if (!isTypeLegal(I->getType(), DestVT) || DestVT.isVector())
return false;
assert ((DestVT == MVT::f32 || DestVT == MVT::f64) &&
"Unexpected value type.");
unsigned SrcReg = getRegForValue(I->getOperand(0));
if (SrcReg == 0)
return false;
EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType(), true);
// Handle sign-extension.
if (SrcVT == MVT::i16 || SrcVT == MVT::i8 || SrcVT == MVT::i1) {
SrcReg =
EmitIntExt(SrcVT.getSimpleVT(), SrcReg, MVT::i32, /*isZExt*/ !Signed);
if (SrcReg == 0)
return false;
}
MRI.constrainRegClass(SrcReg, SrcVT == MVT::i64 ? &AArch64::GPR64RegClass
: &AArch64::GPR32RegClass);
unsigned Opc;
if (SrcVT == MVT::i64) {
if (Signed)
Opc = (DestVT == MVT::f32) ? AArch64::SCVTFUXSri : AArch64::SCVTFUXDri;
else
Opc = (DestVT == MVT::f32) ? AArch64::UCVTFUXSri : AArch64::UCVTFUXDri;
} else {
if (Signed)
Opc = (DestVT == MVT::f32) ? AArch64::SCVTFUWSri : AArch64::SCVTFUWDri;
else
Opc = (DestVT == MVT::f32) ? AArch64::UCVTFUWSri : AArch64::UCVTFUWDri;
}
unsigned ResultReg = createResultReg(TLI.getRegClassFor(DestVT));
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
.addReg(SrcReg);
UpdateValueMap(I, ResultReg);
return true;
}
bool AArch64FastISel::ProcessCallArgs(
SmallVectorImpl<Value *> &Args, SmallVectorImpl<unsigned> &ArgRegs,
SmallVectorImpl<MVT> &ArgVTs, SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
SmallVectorImpl<unsigned> &RegArgs, CallingConv::ID CC,
unsigned &NumBytes) {
SmallVector<CCValAssign, 16> ArgLocs;
CCState CCInfo(CC, false, *FuncInfo.MF, TM, ArgLocs, *Context);
CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CCAssignFnForCall(CC));
// Get a count of how many bytes are to be pushed on the stack.
NumBytes = CCInfo.getNextStackOffset();
// Issue CALLSEQ_START
unsigned AdjStackDown = TII.getCallFrameSetupOpcode();
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackDown))
.addImm(NumBytes);
// Process the args.
for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
CCValAssign &VA = ArgLocs[i];
unsigned Arg = ArgRegs[VA.getValNo()];
MVT ArgVT = ArgVTs[VA.getValNo()];
// Handle arg promotion: SExt, ZExt, AExt.
switch (VA.getLocInfo()) {
case CCValAssign::Full:
break;
case CCValAssign::SExt: {
MVT DestVT = VA.getLocVT();
MVT SrcVT = ArgVT;
Arg = EmitIntExt(SrcVT, Arg, DestVT, /*isZExt*/ false);
if (Arg == 0)
return false;
break;
}
case CCValAssign::AExt:
// Intentional fall-through.
case CCValAssign::ZExt: {
MVT DestVT = VA.getLocVT();
MVT SrcVT = ArgVT;
Arg = EmitIntExt(SrcVT, Arg, DestVT, /*isZExt*/ true);
if (Arg == 0)
return false;
break;
}
default:
llvm_unreachable("Unknown arg promotion!");
}
// Now copy/store arg to correct locations.
if (VA.isRegLoc() && !VA.needsCustom()) {
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
TII.get(TargetOpcode::COPY), VA.getLocReg()).addReg(Arg);
RegArgs.push_back(VA.getLocReg());
} else if (VA.needsCustom()) {
// FIXME: Handle custom args.
return false;
} else {
assert(VA.isMemLoc() && "Assuming store on stack.");
// Need to store on the stack.
unsigned ArgSize = (ArgVT.getSizeInBits() + 7) / 8;
unsigned BEAlign = 0;
if (ArgSize < 8 && !Subtarget->isLittleEndian())
BEAlign = 8 - ArgSize;
Address Addr;
Addr.setKind(Address::RegBase);
Addr.setReg(AArch64::SP);
Addr.setOffset(VA.getLocMemOffset() + BEAlign);
if (!EmitStore(ArgVT, Arg, Addr))
return false;
}
}
return true;
}
bool AArch64FastISel::FinishCall(MVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
const Instruction *I, CallingConv::ID CC,
unsigned &NumBytes) {
// Issue CALLSEQ_END
unsigned AdjStackUp = TII.getCallFrameDestroyOpcode();
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackUp))
.addImm(NumBytes)
.addImm(0);
// Now the return value.
if (RetVT != MVT::isVoid) {
SmallVector<CCValAssign, 16> RVLocs;
CCState CCInfo(CC, false, *FuncInfo.MF, TM, RVLocs, *Context);
CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC));
// Only handle a single return value.
if (RVLocs.size() != 1)
return false;
// Copy all of the result registers out of their specified physreg.
MVT CopyVT = RVLocs[0].getValVT();
unsigned ResultReg = createResultReg(TLI.getRegClassFor(CopyVT));
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
TII.get(TargetOpcode::COPY),
ResultReg).addReg(RVLocs[0].getLocReg());
UsedRegs.push_back(RVLocs[0].getLocReg());
// Finally update the result.
UpdateValueMap(I, ResultReg);
}
return true;
}
bool AArch64FastISel::SelectCall(const Instruction *I,
const char *IntrMemName = nullptr) {
const CallInst *CI = cast<CallInst>(I);
const Value *Callee = CI->getCalledValue();
// Don't handle inline asm or intrinsics.
if (isa<InlineAsm>(Callee))
return false;
// Only handle global variable Callees.
const GlobalValue *GV = dyn_cast<GlobalValue>(Callee);
if (!GV)
return false;
// Check the calling convention.
ImmutableCallSite CS(CI);
CallingConv::ID CC = CS.getCallingConv();
// Let SDISel handle vararg functions.
PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
FunctionType *FTy = cast<FunctionType>(PT->getElementType());
if (FTy->isVarArg())
return false;
// Handle *simple* calls for now.
MVT RetVT;
Type *RetTy = I->getType();
if (RetTy->isVoidTy())
RetVT = MVT::isVoid;
else if (!isTypeLegal(RetTy, RetVT))
return false;
// Set up the argument vectors.
SmallVector<Value *, 8> Args;
SmallVector<unsigned, 8> ArgRegs;
SmallVector<MVT, 8> ArgVTs;
SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
Args.reserve(CS.arg_size());
ArgRegs.reserve(CS.arg_size());
ArgVTs.reserve(CS.arg_size());
ArgFlags.reserve(CS.arg_size());
for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
i != e; ++i) {
// If we're lowering a memory intrinsic instead of a regular call, skip the
// last two arguments, which shouldn't be passed to the underlying function.
if (IntrMemName && e - i <= 2)
break;
unsigned Arg = getRegForValue(*i);
if (Arg == 0)
return false;
ISD::ArgFlagsTy Flags;
unsigned AttrInd = i - CS.arg_begin() + 1;
if (CS.paramHasAttr(AttrInd, Attribute::SExt))
Flags.setSExt();
if (CS.paramHasAttr(AttrInd, Attribute::ZExt))
Flags.setZExt();
// FIXME: Only handle *easy* calls for now.
if (CS.paramHasAttr(AttrInd, Attribute::InReg) ||
CS.paramHasAttr(AttrInd, Attribute::StructRet) ||
CS.paramHasAttr(AttrInd, Attribute::Nest) ||
CS.paramHasAttr(AttrInd, Attribute::ByVal))
return false;
MVT ArgVT;
Type *ArgTy = (*i)->getType();
if (!isTypeLegal(ArgTy, ArgVT) &&
!(ArgVT == MVT::i1 || ArgVT == MVT::i8 || ArgVT == MVT::i16))
return false;
// We don't handle vector parameters yet.
if (ArgVT.isVector() || ArgVT.getSizeInBits() > 64)
return false;
unsigned OriginalAlignment = DL.getABITypeAlignment(ArgTy);
Flags.setOrigAlign(OriginalAlignment);
Args.push_back(*i);
ArgRegs.push_back(Arg);
ArgVTs.push_back(ArgVT);
ArgFlags.push_back(Flags);
}
// Handle the arguments now that we've gotten them.
SmallVector<unsigned, 4> RegArgs;
unsigned NumBytes;
if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
return false;
// Issue the call.
MachineInstrBuilder MIB;
MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::BL));
if (!IntrMemName)
MIB.addGlobalAddress(GV, 0, 0);
else
MIB.addExternalSymbol(IntrMemName, 0);
// Add implicit physical register uses to the call.
for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
MIB.addReg(RegArgs[i], RegState::Implicit);
// Add a register mask with the call-preserved registers.
// Proper defs for return values will be added by setPhysRegsDeadExcept().
MIB.addRegMask(TRI.getCallPreservedMask(CS.getCallingConv()));
// Finish off the call including any return values.
SmallVector<unsigned, 4> UsedRegs;
if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes))
return false;
// Set all unused physreg defs as dead.
static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
return true;
}
bool AArch64FastISel::IsMemCpySmall(uint64_t Len, unsigned Alignment) {
if (Alignment)
return Len / Alignment <= 4;
else
return Len < 32;
}
bool AArch64FastISel::TryEmitSmallMemCpy(Address Dest, Address Src,
uint64_t Len, unsigned Alignment) {
// Make sure we don't bloat code by inlining very large memcpy's.
if (!IsMemCpySmall(Len, Alignment))
return false;
int64_t UnscaledOffset = 0;
Address OrigDest = Dest;
Address OrigSrc = Src;
while (Len) {
MVT VT;
if (!Alignment || Alignment >= 8) {
if (Len >= 8)
VT = MVT::i64;
else if (Len >= 4)
VT = MVT::i32;
else if (Len >= 2)
VT = MVT::i16;
else {
VT = MVT::i8;
}
} else {
// Bound based on alignment.
if (Len >= 4 && Alignment == 4)
VT = MVT::i32;
else if (Len >= 2 && Alignment == 2)
VT = MVT::i16;
else {
VT = MVT::i8;
}
}
bool RV;
unsigned ResultReg;
RV = EmitLoad(VT, ResultReg, Src);
if (!RV)
return false;
RV = EmitStore(VT, ResultReg, Dest);
if (!RV)
return false;
int64_t Size = VT.getSizeInBits() / 8;
Len -= Size;
UnscaledOffset += Size;
// We need to recompute the unscaled offset for each iteration.
Dest.setOffset(OrigDest.getOffset() + UnscaledOffset);
Src.setOffset(OrigSrc.getOffset() + UnscaledOffset);
}
return true;
}
bool AArch64FastISel::SelectIntrinsicCall(const IntrinsicInst &I) {
// FIXME: Handle more intrinsics.
switch (I.getIntrinsicID()) {
default:
return false;
case Intrinsic::memcpy:
case Intrinsic::memmove: {
const MemTransferInst &MTI = cast<MemTransferInst>(I);
// Don't handle volatile.
if (MTI.isVolatile())
return false;
// Disable inlining for memmove before calls to ComputeAddress. Otherwise,
// we would emit dead code because we don't currently handle memmoves.
bool isMemCpy = (I.getIntrinsicID() == Intrinsic::memcpy);
if (isa<ConstantInt>(MTI.getLength()) && isMemCpy) {
// Small memcpy's are common enough that we want to do them without a call
// if possible.
uint64_t Len = cast<ConstantInt>(MTI.getLength())->getZExtValue();
unsigned Alignment = MTI.getAlignment();
if (IsMemCpySmall(Len, Alignment)) {
Address Dest, Src;
if (!ComputeAddress(MTI.getRawDest(), Dest) ||
!ComputeAddress(MTI.getRawSource(), Src))
return false;
if (TryEmitSmallMemCpy(Dest, Src, Len, Alignment))
return true;
}
}
if (!MTI.getLength()->getType()->isIntegerTy(64))
return false;
if (MTI.getSourceAddressSpace() > 255 || MTI.getDestAddressSpace() > 255)
// Fast instruction selection doesn't support the special
// address spaces.
return false;
const char *IntrMemName = isa<MemCpyInst>(I) ? "memcpy" : "memmove";
return SelectCall(&I, IntrMemName);
}
case Intrinsic::memset: {
const MemSetInst &MSI = cast<MemSetInst>(I);
// Don't handle volatile.
if (MSI.isVolatile())
return false;
if (!MSI.getLength()->getType()->isIntegerTy(64))
return false;
if (MSI.getDestAddressSpace() > 255)
// Fast instruction selection doesn't support the special
// address spaces.
return false;
return SelectCall(&I, "memset");
}
case Intrinsic::trap: {
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::BRK))
.addImm(1);
return true;
}
}
return false;
}
bool AArch64FastISel::SelectRet(const Instruction *I) {
const ReturnInst *Ret = cast<ReturnInst>(I);
const Function &F = *I->getParent()->getParent();
if (!FuncInfo.CanLowerReturn)
return false;
if (F.isVarArg())
return false;
// Build a list of return value registers.
SmallVector<unsigned, 4> RetRegs;
if (Ret->getNumOperands() > 0) {
CallingConv::ID CC = F.getCallingConv();
SmallVector<ISD::OutputArg, 4> Outs;
GetReturnInfo(F.getReturnType(), F.getAttributes(), Outs, TLI);
// Analyze operands of the call, assigning locations to each operand.
SmallVector<CCValAssign, 16> ValLocs;
CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, TM, ValLocs,
I->getContext());
CCAssignFn *RetCC = CC == CallingConv::WebKit_JS ? RetCC_AArch64_WebKit_JS
: RetCC_AArch64_AAPCS;
CCInfo.AnalyzeReturn(Outs, RetCC);
// Only handle a single return value for now.
if (ValLocs.size() != 1)
return false;
CCValAssign &VA = ValLocs[0];
const Value *RV = Ret->getOperand(0);
// Don't bother handling odd stuff for now.
if (VA.getLocInfo() != CCValAssign::Full)
return false;
// Only handle register returns for now.
if (!VA.isRegLoc())
return false;
unsigned Reg = getRegForValue(RV);
if (Reg == 0)
return false;
unsigned SrcReg = Reg + VA.getValNo();
unsigned DestReg = VA.getLocReg();
// Avoid a cross-class copy. This is very unlikely.
if (!MRI.getRegClass(SrcReg)->contains(DestReg))
return false;
EVT RVEVT = TLI.getValueType(RV->getType());
if (!RVEVT.isSimple())
return false;
// Vectors (of > 1 lane) in big endian need tricky handling.
if (RVEVT.isVector() && RVEVT.getVectorNumElements() > 1)
return false;
MVT RVVT = RVEVT.getSimpleVT();
if (RVVT == MVT::f128)
return false;
MVT DestVT = VA.getValVT();
// Special handling for extended integers.
if (RVVT != DestVT) {
if (RVVT != MVT::i1 && RVVT != MVT::i8 && RVVT != MVT::i16)
return false;
if (!Outs[0].Flags.isZExt() && !Outs[0].Flags.isSExt())
return false;
bool isZExt = Outs[0].Flags.isZExt();
SrcReg = EmitIntExt(RVVT, SrcReg, DestVT, isZExt);
if (SrcReg == 0)
return false;
}
// Make the copy.
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
TII.get(TargetOpcode::COPY), DestReg).addReg(SrcReg);
// Add register to return instruction.
RetRegs.push_back(VA.getLocReg());
}
MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
TII.get(AArch64::RET_ReallyLR));
for (unsigned i = 0, e = RetRegs.size(); i != e; ++i)
MIB.addReg(RetRegs[i], RegState::Implicit);
return true;
}
bool AArch64FastISel::SelectTrunc(const Instruction *I) {
Type *DestTy = I->getType();
Value *Op = I->getOperand(0);
Type *SrcTy = Op->getType();
EVT SrcEVT = TLI.getValueType(SrcTy, true);
EVT DestEVT = TLI.getValueType(DestTy, true);
if (!SrcEVT.isSimple())
return false;
if (!DestEVT.isSimple())
return false;
MVT SrcVT = SrcEVT.getSimpleVT();
MVT DestVT = DestEVT.getSimpleVT();
if (SrcVT != MVT::i64 && SrcVT != MVT::i32 && SrcVT != MVT::i16 &&
SrcVT != MVT::i8)
return false;
if (DestVT != MVT::i32 && DestVT != MVT::i16 && DestVT != MVT::i8 &&
DestVT != MVT::i1)
return false;
unsigned SrcReg = getRegForValue(Op);
if (!SrcReg)
return false;
// If we're truncating from i64 to a smaller non-legal type then generate an
// AND. Otherwise, we know the high bits are undefined and a truncate doesn't
// generate any code.
if (SrcVT == MVT::i64) {
uint64_t Mask = 0;
switch (DestVT.SimpleTy) {
default:
// Trunc i64 to i32 is handled by the target-independent fast-isel.
return false;
case MVT::i1:
Mask = 0x1;
break;
case MVT::i8:
Mask = 0xff;
break;
case MVT::i16:
Mask = 0xffff;
break;
}
// Issue an extract_subreg to get the lower 32-bits.
unsigned Reg32 = FastEmitInst_extractsubreg(MVT::i32, SrcReg, /*Kill=*/true,
AArch64::sub_32);
MRI.constrainRegClass(Reg32, &AArch64::GPR32RegClass);
// Create the AND instruction which performs the actual truncation.
unsigned ANDReg = createResultReg(&AArch64::GPR32spRegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ANDWri),
ANDReg)
.addReg(Reg32)
.addImm(AArch64_AM::encodeLogicalImmediate(Mask, 32));
SrcReg = ANDReg;
}
UpdateValueMap(I, SrcReg);
return true;
}
unsigned AArch64FastISel::Emiti1Ext(unsigned SrcReg, MVT DestVT, bool isZExt) {
assert((DestVT == MVT::i8 || DestVT == MVT::i16 || DestVT == MVT::i32 ||
DestVT == MVT::i64) &&
"Unexpected value type.");
// Handle i8 and i16 as i32.
if (DestVT == MVT::i8 || DestVT == MVT::i16)
DestVT = MVT::i32;
if (isZExt) {
MRI.constrainRegClass(SrcReg, &AArch64::GPR32RegClass);
unsigned ResultReg = createResultReg(&AArch64::GPR32spRegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ANDWri),
ResultReg)
.addReg(SrcReg)
.addImm(AArch64_AM::encodeLogicalImmediate(1, 32));
if (DestVT == MVT::i64) {
// We're ZExt i1 to i64. The ANDWri Wd, Ws, #1 implicitly clears the
// upper 32 bits. Emit a SUBREG_TO_REG to extend from Wd to Xd.
unsigned Reg64 = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
TII.get(AArch64::SUBREG_TO_REG), Reg64)
.addImm(0)
.addReg(ResultReg)
.addImm(AArch64::sub_32);
ResultReg = Reg64;
}
return ResultReg;
} else {
if (DestVT == MVT::i64) {
// FIXME: We're SExt i1 to i64.
return 0;
}
unsigned ResultReg = createResultReg(&AArch64::GPR32RegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::SBFMWri),
ResultReg)
.addReg(SrcReg)
.addImm(0)
.addImm(0);
return ResultReg;
}
}
unsigned AArch64FastISel::EmitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT,
bool isZExt) {
assert(DestVT != MVT::i1 && "ZeroExt/SignExt an i1?");
unsigned Opc;
unsigned Imm = 0;
switch (SrcVT.SimpleTy) {
default:
return 0;
case MVT::i1:
return Emiti1Ext(SrcReg, DestVT, isZExt);
case MVT::i8:
if (DestVT == MVT::i64)
Opc = isZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
else
Opc = isZExt ? AArch64::UBFMWri : AArch64::SBFMWri;
Imm = 7;
break;
case MVT::i16:
if (DestVT == MVT::i64)
Opc = isZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
else
Opc = isZExt ? AArch64::UBFMWri : AArch64::SBFMWri;
Imm = 15;
break;
case MVT::i32:
assert(DestVT == MVT::i64 && "IntExt i32 to i32?!?");
Opc = isZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
Imm = 31;
break;
}
// Handle i8 and i16 as i32.
if (DestVT == MVT::i8 || DestVT == MVT::i16)
DestVT = MVT::i32;
else if (DestVT == MVT::i64) {
unsigned Src64 = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
TII.get(AArch64::SUBREG_TO_REG), Src64)
.addImm(0)
.addReg(SrcReg)
.addImm(AArch64::sub_32);
SrcReg = Src64;
}
unsigned ResultReg = createResultReg(TLI.getRegClassFor(DestVT));
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
.addReg(SrcReg)
.addImm(0)
.addImm(Imm);
return ResultReg;
}
bool AArch64FastISel::SelectIntExt(const Instruction *I) {
// On ARM, in general, integer casts don't involve legal types; this code
// handles promotable integers. The high bits for a type smaller than
// the register size are assumed to be undefined.
Type *DestTy = I->getType();
Value *Src = I->getOperand(0);
Type *SrcTy = Src->getType();
bool isZExt = isa<ZExtInst>(I);
unsigned SrcReg = getRegForValue(Src);
if (!SrcReg)
return false;
EVT SrcEVT = TLI.getValueType(SrcTy, true);
EVT DestEVT = TLI.getValueType(DestTy, true);
if (!SrcEVT.isSimple())
return false;
if (!DestEVT.isSimple())
return false;
MVT SrcVT = SrcEVT.getSimpleVT();
MVT DestVT = DestEVT.getSimpleVT();
unsigned ResultReg = EmitIntExt(SrcVT, SrcReg, DestVT, isZExt);
if (ResultReg == 0)
return false;
UpdateValueMap(I, ResultReg);
return true;
}
bool AArch64FastISel::SelectRem(const Instruction *I, unsigned ISDOpcode) {
EVT DestEVT = TLI.getValueType(I->getType(), true);
if (!DestEVT.isSimple())
return false;
MVT DestVT = DestEVT.getSimpleVT();
if (DestVT != MVT::i64 && DestVT != MVT::i32)
return false;
unsigned DivOpc;
bool is64bit = (DestVT == MVT::i64);
switch (ISDOpcode) {
default:
return false;
case ISD::SREM:
DivOpc = is64bit ? AArch64::SDIVXr : AArch64::SDIVWr;
break;
case ISD::UREM:
DivOpc = is64bit ? AArch64::UDIVXr : AArch64::UDIVWr;
break;
}
unsigned MSubOpc = is64bit ? AArch64::MSUBXrrr : AArch64::MSUBWrrr;
unsigned Src0Reg = getRegForValue(I->getOperand(0));
if (!Src0Reg)
return false;
unsigned Src1Reg = getRegForValue(I->getOperand(1));
if (!Src1Reg)
return false;
unsigned QuotReg = createResultReg(TLI.getRegClassFor(DestVT));
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(DivOpc), QuotReg)
.addReg(Src0Reg)
.addReg(Src1Reg);
// The remainder is computed as numerator - (quotient * denominator) using the
// MSUB instruction.
unsigned ResultReg = createResultReg(TLI.getRegClassFor(DestVT));
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(MSubOpc), ResultReg)
.addReg(QuotReg)
.addReg(Src1Reg)
.addReg(Src0Reg);
UpdateValueMap(I, ResultReg);
return true;
}
bool AArch64FastISel::SelectMul(const Instruction *I) {
EVT SrcEVT = TLI.getValueType(I->getOperand(0)->getType(), true);
if (!SrcEVT.isSimple())
return false;
MVT SrcVT = SrcEVT.getSimpleVT();
// Must be simple value type. Don't handle vectors.
if (SrcVT != MVT::i64 && SrcVT != MVT::i32 && SrcVT != MVT::i16 &&
SrcVT != MVT::i8)
return false;
unsigned Opc;
unsigned ZReg;
switch (SrcVT.SimpleTy) {
default:
return false;
case MVT::i8:
case MVT::i16:
case MVT::i32:
ZReg = AArch64::WZR;
Opc = AArch64::MADDWrrr;
break;
case MVT::i64:
ZReg = AArch64::XZR;
Opc = AArch64::MADDXrrr;
break;
}
unsigned Src0Reg = getRegForValue(I->getOperand(0));
if (!Src0Reg)
return false;
unsigned Src1Reg = getRegForValue(I->getOperand(1));
if (!Src1Reg)
return false;
// Create the base instruction, then add the operands.
unsigned ResultReg = createResultReg(TLI.getRegClassFor(SrcVT));
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
.addReg(Src0Reg)
.addReg(Src1Reg)
.addReg(ZReg);
UpdateValueMap(I, ResultReg);
return true;
}
bool AArch64FastISel::TargetSelectInstruction(const Instruction *I) {
switch (I->getOpcode()) {
default:
break;
case Instruction::Load:
return SelectLoad(I);
case Instruction::Store:
return SelectStore(I);
case Instruction::Br:
return SelectBranch(I);
case Instruction::IndirectBr:
return SelectIndirectBr(I);
case Instruction::FCmp:
case Instruction::ICmp:
return SelectCmp(I);
case Instruction::Select:
return SelectSelect(I);
case Instruction::FPExt:
return SelectFPExt(I);
case Instruction::FPTrunc:
return SelectFPTrunc(I);
case Instruction::FPToSI:
return SelectFPToInt(I, /*Signed=*/true);
case Instruction::FPToUI:
return SelectFPToInt(I, /*Signed=*/false);
case Instruction::SIToFP:
return SelectIntToFP(I, /*Signed=*/true);
case Instruction::UIToFP:
return SelectIntToFP(I, /*Signed=*/false);
case Instruction::SRem:
return SelectRem(I, ISD::SREM);
case Instruction::URem:
return SelectRem(I, ISD::UREM);
case Instruction::Call:
if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
return SelectIntrinsicCall(*II);
return SelectCall(I);
case Instruction::Ret:
return SelectRet(I);
case Instruction::Trunc:
return SelectTrunc(I);
case Instruction::ZExt:
case Instruction::SExt:
return SelectIntExt(I);
case Instruction::Mul:
// FIXME: This really should be handled by the target-independent selector.
return SelectMul(I);
}
return false;
// Silence warnings.
(void)&CC_AArch64_DarwinPCS_VarArg;
}
namespace llvm {
llvm::FastISel *AArch64::createFastISel(FunctionLoweringInfo &funcInfo,
const TargetLibraryInfo *libInfo) {
return new AArch64FastISel(funcInfo, libInfo);
}
}