1 //===--- AArch64CallLowering.cpp - Call lowering --------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
11 /// This file implements the lowering of LLVM calls to machine code calls for
14 //===----------------------------------------------------------------------===//
16 #include "AArch64CallLowering.h"
17 #include "AArch64ISelLowering.h"
18 #include "AArch64MachineFunctionInfo.h"
19 #include "AArch64Subtarget.h"
20 #include "llvm/ADT/ArrayRef.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/CodeGen/Analysis.h"
23 #include "llvm/CodeGen/CallingConvLower.h"
24 #include "llvm/CodeGen/GlobalISel/MachineIRBuilder.h"
25 #include "llvm/CodeGen/GlobalISel/Utils.h"
26 #include "llvm/CodeGen/LowLevelType.h"
27 #include "llvm/CodeGen/MachineBasicBlock.h"
28 #include "llvm/CodeGen/MachineFrameInfo.h"
29 #include "llvm/CodeGen/MachineFunction.h"
30 #include "llvm/CodeGen/MachineInstrBuilder.h"
31 #include "llvm/CodeGen/MachineMemOperand.h"
32 #include "llvm/CodeGen/MachineOperand.h"
33 #include "llvm/CodeGen/MachineRegisterInfo.h"
34 #include "llvm/CodeGen/TargetRegisterInfo.h"
35 #include "llvm/CodeGen/TargetSubtargetInfo.h"
36 #include "llvm/CodeGen/ValueTypes.h"
37 #include "llvm/IR/Argument.h"
38 #include "llvm/IR/Attributes.h"
39 #include "llvm/IR/Function.h"
40 #include "llvm/IR/Type.h"
41 #include "llvm/IR/Value.h"
42 #include "llvm/Support/MachineValueType.h"
50 AArch64CallLowering::AArch64CallLowering(const AArch64TargetLowering &TLI)
51 : CallLowering(&TLI) {}
54 struct IncomingArgHandler : public CallLowering::ValueHandler {
55 IncomingArgHandler(MachineIRBuilder &MIRBuilder, MachineRegisterInfo &MRI,
57 : ValueHandler(MIRBuilder, MRI, AssignFn), StackUsed(0) {}
59 unsigned getStackAddress(uint64_t Size, int64_t Offset,
60 MachinePointerInfo &MPO) override {
61 auto &MFI = MIRBuilder.getMF().getFrameInfo();
62 int FI = MFI.CreateFixedObject(Size, Offset, true);
63 MPO = MachinePointerInfo::getFixedStack(MIRBuilder.getMF(), FI);
64 unsigned AddrReg = MRI.createGenericVirtualRegister(LLT::pointer(0, 64));
65 MIRBuilder.buildFrameIndex(AddrReg, FI);
66 StackUsed = std::max(StackUsed, Size + Offset);
70 void assignValueToReg(unsigned ValVReg, unsigned PhysReg,
71 CCValAssign &VA) override {
72 markPhysRegUsed(PhysReg);
73 switch (VA.getLocInfo()) {
75 MIRBuilder.buildCopy(ValVReg, PhysReg);
77 case CCValAssign::LocInfo::SExt:
78 case CCValAssign::LocInfo::ZExt:
79 case CCValAssign::LocInfo::AExt: {
80 auto Copy = MIRBuilder.buildCopy(LLT{VA.getLocVT()}, PhysReg);
81 MIRBuilder.buildTrunc(ValVReg, Copy);
87 void assignValueToAddress(unsigned ValVReg, unsigned Addr, uint64_t Size,
88 MachinePointerInfo &MPO, CCValAssign &VA) override {
89 auto MMO = MIRBuilder.getMF().getMachineMemOperand(
90 MPO, MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant, Size,
92 MIRBuilder.buildLoad(ValVReg, Addr, *MMO);
95 /// How the physical register gets marked varies between formal
96 /// parameters (it's a basic-block live-in), and a call instruction
97 /// (it's an implicit-def of the BL).
98 virtual void markPhysRegUsed(unsigned PhysReg) = 0;
103 struct FormalArgHandler : public IncomingArgHandler {
104 FormalArgHandler(MachineIRBuilder &MIRBuilder, MachineRegisterInfo &MRI,
105 CCAssignFn *AssignFn)
106 : IncomingArgHandler(MIRBuilder, MRI, AssignFn) {}
108 void markPhysRegUsed(unsigned PhysReg) override {
109 MIRBuilder.getMBB().addLiveIn(PhysReg);
113 struct CallReturnHandler : public IncomingArgHandler {
114 CallReturnHandler(MachineIRBuilder &MIRBuilder, MachineRegisterInfo &MRI,
115 MachineInstrBuilder MIB, CCAssignFn *AssignFn)
116 : IncomingArgHandler(MIRBuilder, MRI, AssignFn), MIB(MIB) {}
118 void markPhysRegUsed(unsigned PhysReg) override {
119 MIB.addDef(PhysReg, RegState::Implicit);
122 MachineInstrBuilder MIB;
125 struct OutgoingArgHandler : public CallLowering::ValueHandler {
126 OutgoingArgHandler(MachineIRBuilder &MIRBuilder, MachineRegisterInfo &MRI,
127 MachineInstrBuilder MIB, CCAssignFn *AssignFn,
128 CCAssignFn *AssignFnVarArg)
129 : ValueHandler(MIRBuilder, MRI, AssignFn), MIB(MIB),
130 AssignFnVarArg(AssignFnVarArg), StackSize(0) {}
132 unsigned getStackAddress(uint64_t Size, int64_t Offset,
133 MachinePointerInfo &MPO) override {
134 LLT p0 = LLT::pointer(0, 64);
135 LLT s64 = LLT::scalar(64);
136 unsigned SPReg = MRI.createGenericVirtualRegister(p0);
137 MIRBuilder.buildCopy(SPReg, AArch64::SP);
139 unsigned OffsetReg = MRI.createGenericVirtualRegister(s64);
140 MIRBuilder.buildConstant(OffsetReg, Offset);
142 unsigned AddrReg = MRI.createGenericVirtualRegister(p0);
143 MIRBuilder.buildGEP(AddrReg, SPReg, OffsetReg);
145 MPO = MachinePointerInfo::getStack(MIRBuilder.getMF(), Offset);
149 void assignValueToReg(unsigned ValVReg, unsigned PhysReg,
150 CCValAssign &VA) override {
151 MIB.addUse(PhysReg, RegState::Implicit);
152 unsigned ExtReg = extendRegister(ValVReg, VA);
153 MIRBuilder.buildCopy(PhysReg, ExtReg);
156 void assignValueToAddress(unsigned ValVReg, unsigned Addr, uint64_t Size,
157 MachinePointerInfo &MPO, CCValAssign &VA) override {
158 if (VA.getLocInfo() == CCValAssign::LocInfo::AExt) {
159 Size = VA.getLocVT().getSizeInBits() / 8;
160 ValVReg = MIRBuilder.buildAnyExt(LLT::scalar(Size * 8), ValVReg)
164 auto MMO = MIRBuilder.getMF().getMachineMemOperand(
165 MPO, MachineMemOperand::MOStore, Size, 0);
166 MIRBuilder.buildStore(ValVReg, Addr, *MMO);
169 bool assignArg(unsigned ValNo, MVT ValVT, MVT LocVT,
170 CCValAssign::LocInfo LocInfo,
171 const CallLowering::ArgInfo &Info,
172 CCState &State) override {
175 Res = AssignFn(ValNo, ValVT, LocVT, LocInfo, Info.Flags, State);
177 Res = AssignFnVarArg(ValNo, ValVT, LocVT, LocInfo, Info.Flags, State);
179 StackSize = State.getNextStackOffset();
183 MachineInstrBuilder MIB;
184 CCAssignFn *AssignFnVarArg;
189 void AArch64CallLowering::splitToValueTypes(
190 const ArgInfo &OrigArg, SmallVectorImpl<ArgInfo> &SplitArgs,
191 const DataLayout &DL, MachineRegisterInfo &MRI, CallingConv::ID CallConv,
192 const SplitArgTy &PerformArgSplit) const {
193 const AArch64TargetLowering &TLI = *getTLI<AArch64TargetLowering>();
194 LLVMContext &Ctx = OrigArg.Ty->getContext();
196 if (OrigArg.Ty->isVoidTy())
199 SmallVector<EVT, 4> SplitVTs;
200 SmallVector<uint64_t, 4> Offsets;
201 ComputeValueVTs(TLI, DL, OrigArg.Ty, SplitVTs, &Offsets, 0);
203 if (SplitVTs.size() == 1) {
204 // No splitting to do, but we want to replace the original type (e.g. [1 x
205 // double] -> double).
206 SplitArgs.emplace_back(OrigArg.Reg, SplitVTs[0].getTypeForEVT(Ctx),
207 OrigArg.Flags, OrigArg.IsFixed);
211 unsigned FirstRegIdx = SplitArgs.size();
212 bool NeedsRegBlock = TLI.functionArgumentNeedsConsecutiveRegisters(
213 OrigArg.Ty, CallConv, false);
214 for (auto SplitVT : SplitVTs) {
215 Type *SplitTy = SplitVT.getTypeForEVT(Ctx);
217 ArgInfo{MRI.createGenericVirtualRegister(getLLTForType(*SplitTy, DL)),
218 SplitTy, OrigArg.Flags, OrigArg.IsFixed});
220 SplitArgs.back().Flags.setInConsecutiveRegs();
223 SplitArgs.back().Flags.setInConsecutiveRegsLast();
225 for (unsigned i = 0; i < Offsets.size(); ++i)
226 PerformArgSplit(SplitArgs[FirstRegIdx + i].Reg, Offsets[i] * 8);
229 bool AArch64CallLowering::lowerReturn(MachineIRBuilder &MIRBuilder,
230 const Value *Val, unsigned VReg) const {
231 MachineFunction &MF = MIRBuilder.getMF();
232 const Function &F = MF.getFunction();
234 auto MIB = MIRBuilder.buildInstrNoInsert(AArch64::RET_ReallyLR);
235 assert(((Val && VReg) || (!Val && !VReg)) && "Return value without a vreg");
238 MachineRegisterInfo &MRI = MF.getRegInfo();
240 // We zero-extend i1s to i8.
241 if (MRI.getType(VReg).getSizeInBits() == 1)
242 VReg = MIRBuilder.buildZExt(LLT::scalar(8), VReg)->getOperand(0).getReg();
244 const AArch64TargetLowering &TLI = *getTLI<AArch64TargetLowering>();
245 CCAssignFn *AssignFn = TLI.CCAssignFnForReturn(F.getCallingConv());
246 auto &DL = F.getParent()->getDataLayout();
248 ArgInfo OrigArg{VReg, Val->getType()};
249 setArgFlags(OrigArg, AttributeList::ReturnIndex, DL, F);
251 SmallVector<ArgInfo, 8> SplitArgs;
252 splitToValueTypes(OrigArg, SplitArgs, DL, MRI, F.getCallingConv(),
253 [&](unsigned Reg, uint64_t Offset) {
254 MIRBuilder.buildExtract(Reg, VReg, Offset);
257 OutgoingArgHandler Handler(MIRBuilder, MRI, MIB, AssignFn, AssignFn);
258 Success = handleAssignments(MIRBuilder, SplitArgs, Handler);
261 MIRBuilder.insertInstr(MIB);
265 bool AArch64CallLowering::lowerFormalArguments(MachineIRBuilder &MIRBuilder,
267 ArrayRef<unsigned> VRegs) const {
268 MachineFunction &MF = MIRBuilder.getMF();
269 MachineBasicBlock &MBB = MIRBuilder.getMBB();
270 MachineRegisterInfo &MRI = MF.getRegInfo();
271 auto &DL = F.getParent()->getDataLayout();
273 SmallVector<ArgInfo, 8> SplitArgs;
275 for (auto &Arg : F.args()) {
276 if (DL.getTypeStoreSize(Arg.getType()) == 0)
278 ArgInfo OrigArg{VRegs[i], Arg.getType()};
279 setArgFlags(OrigArg, i + AttributeList::FirstArgIndex, DL, F);
281 LLT Ty = MRI.getType(VRegs[i]);
282 unsigned Dst = VRegs[i];
284 splitToValueTypes(OrigArg, SplitArgs, DL, MRI, F.getCallingConv(),
285 [&](unsigned Reg, uint64_t Offset) {
288 Dst = MRI.createGenericVirtualRegister(Ty);
289 MIRBuilder.buildUndef(Dst);
291 unsigned Tmp = MRI.createGenericVirtualRegister(Ty);
292 MIRBuilder.buildInsert(Tmp, Dst, Reg, Offset);
297 MIRBuilder.buildCopy(VRegs[i], Dst);
302 MIRBuilder.setInstr(*MBB.begin());
304 const AArch64TargetLowering &TLI = *getTLI<AArch64TargetLowering>();
305 CCAssignFn *AssignFn =
306 TLI.CCAssignFnForCall(F.getCallingConv(), /*IsVarArg=*/false);
308 FormalArgHandler Handler(MIRBuilder, MRI, AssignFn);
309 if (!handleAssignments(MIRBuilder, SplitArgs, Handler))
313 if (!MF.getSubtarget<AArch64Subtarget>().isTargetDarwin()) {
314 // FIXME: we need to reimplement saveVarArgsRegisters from
315 // AArch64ISelLowering.
319 // We currently pass all varargs at 8-byte alignment.
320 uint64_t StackOffset = alignTo(Handler.StackUsed, 8);
322 auto &MFI = MIRBuilder.getMF().getFrameInfo();
323 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
324 FuncInfo->setVarArgsStackIndex(MFI.CreateFixedObject(4, StackOffset, true));
327 // Move back to the end of the basic block.
328 MIRBuilder.setMBB(MBB);
333 bool AArch64CallLowering::lowerCall(MachineIRBuilder &MIRBuilder,
334 CallingConv::ID CallConv,
335 const MachineOperand &Callee,
336 const ArgInfo &OrigRet,
337 ArrayRef<ArgInfo> OrigArgs) const {
338 MachineFunction &MF = MIRBuilder.getMF();
339 const Function &F = MF.getFunction();
340 MachineRegisterInfo &MRI = MF.getRegInfo();
341 auto &DL = F.getParent()->getDataLayout();
343 SmallVector<ArgInfo, 8> SplitArgs;
344 for (auto &OrigArg : OrigArgs) {
345 splitToValueTypes(OrigArg, SplitArgs, DL, MRI, CallConv,
346 [&](unsigned Reg, uint64_t Offset) {
347 MIRBuilder.buildExtract(Reg, OrigArg.Reg, Offset);
351 // Find out which ABI gets to decide where things go.
352 const AArch64TargetLowering &TLI = *getTLI<AArch64TargetLowering>();
353 CCAssignFn *AssignFnFixed =
354 TLI.CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
355 CCAssignFn *AssignFnVarArg =
356 TLI.CCAssignFnForCall(CallConv, /*IsVarArg=*/true);
358 auto CallSeqStart = MIRBuilder.buildInstr(AArch64::ADJCALLSTACKDOWN);
360 // Create a temporarily-floating call instruction so we can add the implicit
361 // uses of arg registers.
362 auto MIB = MIRBuilder.buildInstrNoInsert(Callee.isReg() ? AArch64::BLR
366 // Tell the call which registers are clobbered.
367 auto TRI = MF.getSubtarget().getRegisterInfo();
368 MIB.addRegMask(TRI->getCallPreservedMask(MF, F.getCallingConv()));
370 // Do the actual argument marshalling.
371 SmallVector<unsigned, 8> PhysRegs;
372 OutgoingArgHandler Handler(MIRBuilder, MRI, MIB, AssignFnFixed,
374 if (!handleAssignments(MIRBuilder, SplitArgs, Handler))
377 // Now we can add the actual call instruction to the correct basic block.
378 MIRBuilder.insertInstr(MIB);
380 // If Callee is a reg, since it is used by a target specific
381 // instruction, it must have a register class matching the
382 // constraint of that instruction.
384 MIB->getOperand(0).setReg(constrainOperandRegClass(
385 MF, *TRI, MRI, *MF.getSubtarget().getInstrInfo(),
386 *MF.getSubtarget().getRegBankInfo(), *MIB, MIB->getDesc(), Callee, 0));
388 // Finally we can copy the returned value back into its virtual-register. In
389 // symmetry with the arugments, the physical register must be an
390 // implicit-define of the call instruction.
391 CCAssignFn *RetAssignFn = TLI.CCAssignFnForReturn(F.getCallingConv());
395 SmallVector<uint64_t, 8> RegOffsets;
396 SmallVector<unsigned, 8> SplitRegs;
397 splitToValueTypes(OrigRet, SplitArgs, DL, MRI, F.getCallingConv(),
398 [&](unsigned Reg, uint64_t Offset) {
399 RegOffsets.push_back(Offset);
400 SplitRegs.push_back(Reg);
403 CallReturnHandler Handler(MIRBuilder, MRI, MIB, RetAssignFn);
404 if (!handleAssignments(MIRBuilder, SplitArgs, Handler))
407 if (!RegOffsets.empty())
408 MIRBuilder.buildSequence(OrigRet.Reg, SplitRegs, RegOffsets);
411 CallSeqStart.addImm(Handler.StackSize).addImm(0);
412 MIRBuilder.buildInstr(AArch64::ADJCALLSTACKUP)
413 .addImm(Handler.StackSize)