xref: /freebsd/contrib/llvm-project/llvm/lib/Target/X86/GISel/X86CallLowering.cpp (revision 7fdf597e96a02165cfe22ff357b857d5fa15ed8a)
1 //===- llvm/lib/Target/X86/X86CallLowering.cpp - Call lowering ------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 /// \file
10 /// This file implements the lowering of LLVM calls to machine code calls for
11 /// GlobalISel.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "X86CallLowering.h"
16 #include "X86CallingConv.h"
17 #include "X86ISelLowering.h"
18 #include "X86InstrInfo.h"
19 #include "X86MachineFunctionInfo.h"
20 #include "X86RegisterInfo.h"
21 #include "X86Subtarget.h"
22 #include "llvm/ADT/ArrayRef.h"
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/CodeGen/Analysis.h"
25 #include "llvm/CodeGen/CallingConvLower.h"
26 #include "llvm/CodeGen/FunctionLoweringInfo.h"
27 #include "llvm/CodeGen/GlobalISel/MachineIRBuilder.h"
28 #include "llvm/CodeGen/GlobalISel/Utils.h"
29 #include "llvm/CodeGen/LowLevelTypeUtils.h"
30 #include "llvm/CodeGen/MachineBasicBlock.h"
31 #include "llvm/CodeGen/MachineFrameInfo.h"
32 #include "llvm/CodeGen/MachineFunction.h"
33 #include "llvm/CodeGen/MachineInstrBuilder.h"
34 #include "llvm/CodeGen/MachineMemOperand.h"
35 #include "llvm/CodeGen/MachineOperand.h"
36 #include "llvm/CodeGen/MachineRegisterInfo.h"
37 #include "llvm/CodeGen/TargetInstrInfo.h"
38 #include "llvm/CodeGen/TargetSubtargetInfo.h"
39 #include "llvm/CodeGen/ValueTypes.h"
40 #include "llvm/CodeGenTypes/LowLevelType.h"
41 #include "llvm/CodeGenTypes/MachineValueType.h"
42 #include "llvm/IR/Attributes.h"
43 #include "llvm/IR/DataLayout.h"
44 #include "llvm/IR/Function.h"
45 #include "llvm/IR/Value.h"
46 #include "llvm/MC/MCRegisterInfo.h"
47 #include <cassert>
48 #include <cstdint>
49 
50 using namespace llvm;
51 
52 X86CallLowering::X86CallLowering(const X86TargetLowering &TLI)
53     : CallLowering(&TLI) {}
54 
55 namespace {
56 
57 struct X86OutgoingValueAssigner : public CallLowering::OutgoingValueAssigner {
58 private:
59   uint64_t StackSize = 0;
60   unsigned NumXMMRegs = 0;
61 
62 public:
63   uint64_t getStackSize() { return StackSize; }
64   unsigned getNumXmmRegs() { return NumXMMRegs; }
65 
66   X86OutgoingValueAssigner(CCAssignFn *AssignFn_)
67       : CallLowering::OutgoingValueAssigner(AssignFn_) {}
68 
69   bool assignArg(unsigned ValNo, EVT OrigVT, MVT ValVT, MVT LocVT,
70                  CCValAssign::LocInfo LocInfo,
71                  const CallLowering::ArgInfo &Info, ISD::ArgFlagsTy Flags,
72                  CCState &State) override {
73     bool Res = AssignFn(ValNo, ValVT, LocVT, LocInfo, Flags, State);
74     StackSize = State.getStackSize();
75 
76     static const MCPhysReg XMMArgRegs[] = {X86::XMM0, X86::XMM1, X86::XMM2,
77                                            X86::XMM3, X86::XMM4, X86::XMM5,
78                                            X86::XMM6, X86::XMM7};
79     if (!Info.IsFixed)
80       NumXMMRegs = State.getFirstUnallocated(XMMArgRegs);
81 
82     return Res;
83   }
84 };
85 
86 struct X86OutgoingValueHandler : public CallLowering::OutgoingValueHandler {
87   X86OutgoingValueHandler(MachineIRBuilder &MIRBuilder,
88                           MachineRegisterInfo &MRI, MachineInstrBuilder &MIB)
89       : OutgoingValueHandler(MIRBuilder, MRI), MIB(MIB),
90         DL(MIRBuilder.getMF().getDataLayout()),
91         STI(MIRBuilder.getMF().getSubtarget<X86Subtarget>()) {}
92 
93   Register getStackAddress(uint64_t Size, int64_t Offset,
94                            MachinePointerInfo &MPO,
95                            ISD::ArgFlagsTy Flags) override {
96     LLT p0 = LLT::pointer(0, DL.getPointerSizeInBits(0));
97     LLT SType = LLT::scalar(DL.getPointerSizeInBits(0));
98     auto SPReg =
99         MIRBuilder.buildCopy(p0, STI.getRegisterInfo()->getStackRegister());
100 
101     auto OffsetReg = MIRBuilder.buildConstant(SType, Offset);
102 
103     auto AddrReg = MIRBuilder.buildPtrAdd(p0, SPReg, OffsetReg);
104 
105     MPO = MachinePointerInfo::getStack(MIRBuilder.getMF(), Offset);
106     return AddrReg.getReg(0);
107   }
108 
109   void assignValueToReg(Register ValVReg, Register PhysReg,
110                         const CCValAssign &VA) override {
111     MIB.addUse(PhysReg, RegState::Implicit);
112     Register ExtReg = extendRegister(ValVReg, VA);
113     MIRBuilder.buildCopy(PhysReg, ExtReg);
114   }
115 
116   void assignValueToAddress(Register ValVReg, Register Addr, LLT MemTy,
117                             const MachinePointerInfo &MPO,
118                             const CCValAssign &VA) override {
119     MachineFunction &MF = MIRBuilder.getMF();
120     Register ExtReg = extendRegister(ValVReg, VA);
121 
122     auto *MMO = MF.getMachineMemOperand(MPO, MachineMemOperand::MOStore, MemTy,
123                                         inferAlignFromPtrInfo(MF, MPO));
124     MIRBuilder.buildStore(ExtReg, Addr, *MMO);
125   }
126 
127 protected:
128   MachineInstrBuilder &MIB;
129   const DataLayout &DL;
130   const X86Subtarget &STI;
131 };
132 
133 } // end anonymous namespace
134 
135 bool X86CallLowering::canLowerReturn(
136     MachineFunction &MF, CallingConv::ID CallConv,
137     SmallVectorImpl<CallLowering::BaseArgInfo> &Outs, bool IsVarArg) const {
138   LLVMContext &Context = MF.getFunction().getContext();
139   SmallVector<CCValAssign, 16> RVLocs;
140   CCState CCInfo(CallConv, IsVarArg, MF, RVLocs, Context);
141   return checkReturn(CCInfo, Outs, RetCC_X86);
142 }
143 
144 bool X86CallLowering::lowerReturn(MachineIRBuilder &MIRBuilder,
145                                   const Value *Val, ArrayRef<Register> VRegs,
146                                   FunctionLoweringInfo &FLI) const {
147   assert(((Val && !VRegs.empty()) || (!Val && VRegs.empty())) &&
148          "Return value without a vreg");
149   MachineFunction &MF = MIRBuilder.getMF();
150   auto MIB = MIRBuilder.buildInstrNoInsert(X86::RET).addImm(0);
151   auto FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
152   const auto &STI = MF.getSubtarget<X86Subtarget>();
153   Register RetReg = STI.is64Bit() ? X86::RAX : X86::EAX;
154 
155   if (!FLI.CanLowerReturn) {
156     insertSRetStores(MIRBuilder, Val->getType(), VRegs, FLI.DemoteRegister);
157     MIRBuilder.buildCopy(RetReg, FLI.DemoteRegister);
158     MIB.addReg(RetReg);
159   } else if (Register Reg = FuncInfo->getSRetReturnReg()) {
160     MIRBuilder.buildCopy(RetReg, Reg);
161     MIB.addReg(RetReg);
162   } else if (!VRegs.empty()) {
163     const Function &F = MF.getFunction();
164     MachineRegisterInfo &MRI = MF.getRegInfo();
165     const DataLayout &DL = MF.getDataLayout();
166 
167     ArgInfo OrigRetInfo(VRegs, Val->getType(), 0);
168     setArgFlags(OrigRetInfo, AttributeList::ReturnIndex, DL, F);
169 
170     SmallVector<ArgInfo, 4> SplitRetInfos;
171     splitToValueTypes(OrigRetInfo, SplitRetInfos, DL, F.getCallingConv());
172 
173     X86OutgoingValueAssigner Assigner(RetCC_X86);
174     X86OutgoingValueHandler Handler(MIRBuilder, MRI, MIB);
175     if (!determineAndHandleAssignments(Handler, Assigner, SplitRetInfos,
176                                        MIRBuilder, F.getCallingConv(),
177                                        F.isVarArg()))
178       return false;
179   }
180 
181   MIRBuilder.insertInstr(MIB);
182   return true;
183 }
184 
185 namespace {
186 
187 struct X86IncomingValueHandler : public CallLowering::IncomingValueHandler {
188   X86IncomingValueHandler(MachineIRBuilder &MIRBuilder,
189                           MachineRegisterInfo &MRI)
190       : IncomingValueHandler(MIRBuilder, MRI),
191         DL(MIRBuilder.getMF().getDataLayout()) {}
192 
193   Register getStackAddress(uint64_t Size, int64_t Offset,
194                            MachinePointerInfo &MPO,
195                            ISD::ArgFlagsTy Flags) override {
196     auto &MFI = MIRBuilder.getMF().getFrameInfo();
197 
198     // Byval is assumed to be writable memory, but other stack passed arguments
199     // are not.
200     const bool IsImmutable = !Flags.isByVal();
201 
202     int FI = MFI.CreateFixedObject(Size, Offset, IsImmutable);
203     MPO = MachinePointerInfo::getFixedStack(MIRBuilder.getMF(), FI);
204 
205     return MIRBuilder
206         .buildFrameIndex(LLT::pointer(0, DL.getPointerSizeInBits(0)), FI)
207         .getReg(0);
208   }
209 
210   void assignValueToAddress(Register ValVReg, Register Addr, LLT MemTy,
211                             const MachinePointerInfo &MPO,
212                             const CCValAssign &VA) override {
213     MachineFunction &MF = MIRBuilder.getMF();
214     auto *MMO = MF.getMachineMemOperand(
215         MPO, MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant, MemTy,
216         inferAlignFromPtrInfo(MF, MPO));
217     MIRBuilder.buildLoad(ValVReg, Addr, *MMO);
218   }
219 
220   void assignValueToReg(Register ValVReg, Register PhysReg,
221                         const CCValAssign &VA) override {
222     markPhysRegUsed(PhysReg);
223     IncomingValueHandler::assignValueToReg(ValVReg, PhysReg, VA);
224   }
225 
226   /// How the physical register gets marked varies between formal
227   /// parameters (it's a basic-block live-in), and a call instruction
228   /// (it's an implicit-def of the BL).
229   virtual void markPhysRegUsed(unsigned PhysReg) = 0;
230 
231 protected:
232   const DataLayout &DL;
233 };
234 
235 struct FormalArgHandler : public X86IncomingValueHandler {
236   FormalArgHandler(MachineIRBuilder &MIRBuilder, MachineRegisterInfo &MRI)
237       : X86IncomingValueHandler(MIRBuilder, MRI) {}
238 
239   void markPhysRegUsed(unsigned PhysReg) override {
240     MIRBuilder.getMRI()->addLiveIn(PhysReg);
241     MIRBuilder.getMBB().addLiveIn(PhysReg);
242   }
243 };
244 
245 struct CallReturnHandler : public X86IncomingValueHandler {
246   CallReturnHandler(MachineIRBuilder &MIRBuilder, MachineRegisterInfo &MRI,
247                     MachineInstrBuilder &MIB)
248       : X86IncomingValueHandler(MIRBuilder, MRI), MIB(MIB) {}
249 
250   void markPhysRegUsed(unsigned PhysReg) override {
251     MIB.addDef(PhysReg, RegState::Implicit);
252   }
253 
254 protected:
255   MachineInstrBuilder &MIB;
256 };
257 
258 } // end anonymous namespace
259 
260 bool X86CallLowering::lowerFormalArguments(MachineIRBuilder &MIRBuilder,
261                                            const Function &F,
262                                            ArrayRef<ArrayRef<Register>> VRegs,
263                                            FunctionLoweringInfo &FLI) const {
264   MachineFunction &MF = MIRBuilder.getMF();
265   MachineRegisterInfo &MRI = MF.getRegInfo();
266   auto DL = MF.getDataLayout();
267   auto FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
268 
269   SmallVector<ArgInfo, 8> SplitArgs;
270 
271   if (!FLI.CanLowerReturn)
272     insertSRetIncomingArgument(F, SplitArgs, FLI.DemoteRegister, MRI, DL);
273 
274   // TODO: handle variadic function
275   if (F.isVarArg())
276     return false;
277 
278   unsigned Idx = 0;
279   for (const auto &Arg : F.args()) {
280     // TODO: handle not simple cases.
281     if (Arg.hasAttribute(Attribute::ByVal) ||
282         Arg.hasAttribute(Attribute::InReg) ||
283         Arg.hasAttribute(Attribute::SwiftSelf) ||
284         Arg.hasAttribute(Attribute::SwiftError) ||
285         Arg.hasAttribute(Attribute::Nest) || VRegs[Idx].size() > 1)
286       return false;
287 
288     if (Arg.hasAttribute(Attribute::StructRet)) {
289       assert(VRegs[Idx].size() == 1 &&
290              "Unexpected amount of registers for sret argument.");
291       FuncInfo->setSRetReturnReg(VRegs[Idx][0]);
292     }
293 
294     ArgInfo OrigArg(VRegs[Idx], Arg.getType(), Idx);
295     setArgFlags(OrigArg, Idx + AttributeList::FirstArgIndex, DL, F);
296     splitToValueTypes(OrigArg, SplitArgs, DL, F.getCallingConv());
297     Idx++;
298   }
299 
300   if (SplitArgs.empty())
301     return true;
302 
303   MachineBasicBlock &MBB = MIRBuilder.getMBB();
304   if (!MBB.empty())
305     MIRBuilder.setInstr(*MBB.begin());
306 
307   X86OutgoingValueAssigner Assigner(CC_X86);
308   FormalArgHandler Handler(MIRBuilder, MRI);
309   if (!determineAndHandleAssignments(Handler, Assigner, SplitArgs, MIRBuilder,
310                                      F.getCallingConv(), F.isVarArg()))
311     return false;
312 
313   // Move back to the end of the basic block.
314   MIRBuilder.setMBB(MBB);
315 
316   return true;
317 }
318 
319 bool X86CallLowering::lowerCall(MachineIRBuilder &MIRBuilder,
320                                 CallLoweringInfo &Info) const {
321   MachineFunction &MF = MIRBuilder.getMF();
322   const Function &F = MF.getFunction();
323   MachineRegisterInfo &MRI = MF.getRegInfo();
324   const DataLayout &DL = F.getDataLayout();
325   const X86Subtarget &STI = MF.getSubtarget<X86Subtarget>();
326   const TargetInstrInfo &TII = *STI.getInstrInfo();
327   const X86RegisterInfo *TRI = STI.getRegisterInfo();
328 
329   // Handle only Linux C, X86_64_SysV calling conventions for now.
330   if (!STI.isTargetLinux() || !(Info.CallConv == CallingConv::C ||
331                                 Info.CallConv == CallingConv::X86_64_SysV))
332     return false;
333 
334   unsigned AdjStackDown = TII.getCallFrameSetupOpcode();
335   auto CallSeqStart = MIRBuilder.buildInstr(AdjStackDown);
336 
337   // Create a temporarily-floating call instruction so we can add the implicit
338   // uses of arg registers.
339   bool Is64Bit = STI.is64Bit();
340   unsigned CallOpc = Info.Callee.isReg()
341                          ? (Is64Bit ? X86::CALL64r : X86::CALL32r)
342                          : (Is64Bit ? X86::CALL64pcrel32 : X86::CALLpcrel32);
343 
344   auto MIB = MIRBuilder.buildInstrNoInsert(CallOpc)
345                  .add(Info.Callee)
346                  .addRegMask(TRI->getCallPreservedMask(MF, Info.CallConv));
347 
348   SmallVector<ArgInfo, 8> SplitArgs;
349   for (const auto &OrigArg : Info.OrigArgs) {
350 
351     // TODO: handle not simple cases.
352     if (OrigArg.Flags[0].isByVal())
353       return false;
354 
355     if (OrigArg.Regs.size() > 1)
356       return false;
357 
358     splitToValueTypes(OrigArg, SplitArgs, DL, Info.CallConv);
359   }
360   // Do the actual argument marshalling.
361   X86OutgoingValueAssigner Assigner(CC_X86);
362   X86OutgoingValueHandler Handler(MIRBuilder, MRI, MIB);
363   if (!determineAndHandleAssignments(Handler, Assigner, SplitArgs, MIRBuilder,
364                                      Info.CallConv, Info.IsVarArg))
365     return false;
366 
367   bool IsFixed = Info.OrigArgs.empty() ? true : Info.OrigArgs.back().IsFixed;
368   if (STI.is64Bit() && !IsFixed && !STI.isCallingConvWin64(Info.CallConv)) {
369     // From AMD64 ABI document:
370     // For calls that may call functions that use varargs or stdargs
371     // (prototype-less calls or calls to functions containing ellipsis (...) in
372     // the declaration) %al is used as hidden argument to specify the number
373     // of SSE registers used. The contents of %al do not need to match exactly
374     // the number of registers, but must be an ubound on the number of SSE
375     // registers used and is in the range 0 - 8 inclusive.
376 
377     MIRBuilder.buildInstr(X86::MOV8ri)
378         .addDef(X86::AL)
379         .addImm(Assigner.getNumXmmRegs());
380     MIB.addUse(X86::AL, RegState::Implicit);
381   }
382 
383   // Now we can add the actual call instruction to the correct basic block.
384   MIRBuilder.insertInstr(MIB);
385 
386   // If Callee is a reg, since it is used by a target specific
387   // instruction, it must have a register class matching the
388   // constraint of that instruction.
389   if (Info.Callee.isReg())
390     MIB->getOperand(0).setReg(constrainOperandRegClass(
391         MF, *TRI, MRI, *MF.getSubtarget().getInstrInfo(),
392         *MF.getSubtarget().getRegBankInfo(), *MIB, MIB->getDesc(), Info.Callee,
393         0));
394 
395   // Finally we can copy the returned value back into its virtual-register. In
396   // symmetry with the arguments, the physical register must be an
397   // implicit-define of the call instruction.
398 
399   if (Info.CanLowerReturn && !Info.OrigRet.Ty->isVoidTy()) {
400     if (Info.OrigRet.Regs.size() > 1)
401       return false;
402 
403     SplitArgs.clear();
404     SmallVector<Register, 8> NewRegs;
405 
406     splitToValueTypes(Info.OrigRet, SplitArgs, DL, Info.CallConv);
407 
408     X86OutgoingValueAssigner Assigner(RetCC_X86);
409     CallReturnHandler Handler(MIRBuilder, MRI, MIB);
410     if (!determineAndHandleAssignments(Handler, Assigner, SplitArgs, MIRBuilder,
411                                        Info.CallConv, Info.IsVarArg))
412       return false;
413 
414     if (!NewRegs.empty())
415       MIRBuilder.buildMergeLikeInstr(Info.OrigRet.Regs[0], NewRegs);
416   }
417 
418   CallSeqStart.addImm(Assigner.getStackSize())
419       .addImm(0 /* see getFrameTotalSize */)
420       .addImm(0 /* see getFrameAdjustment */);
421 
422   unsigned AdjStackUp = TII.getCallFrameDestroyOpcode();
423   MIRBuilder.buildInstr(AdjStackUp)
424       .addImm(Assigner.getStackSize())
425       .addImm(0 /* NumBytesForCalleeToPop */);
426 
427   if (!Info.CanLowerReturn)
428     insertSRetLoads(MIRBuilder, Info.OrigRet.Ty, Info.OrigRet.Regs,
429                     Info.DemoteRegister, Info.DemoteStackIndex);
430 
431   return true;
432 }
433