xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/AMDGPUCallLowering.cpp (revision e25152834cdf3b353892835a4f3b157e066a8ed4)
1 //===-- llvm/lib/Target/AMDGPU/AMDGPUCallLowering.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 "AMDGPUCallLowering.h"
16 #include "AMDGPU.h"
17 #include "AMDGPUISelLowering.h"
18 #include "AMDGPUSubtarget.h"
19 #include "AMDGPUTargetMachine.h"
20 #include "SIISelLowering.h"
21 #include "SIMachineFunctionInfo.h"
22 #include "SIRegisterInfo.h"
23 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
24 #include "llvm/CodeGen/Analysis.h"
25 #include "llvm/CodeGen/CallingConvLower.h"
26 #include "llvm/CodeGen/GlobalISel/MachineIRBuilder.h"
27 #include "llvm/CodeGen/MachineInstrBuilder.h"
28 #include "llvm/Support/LowLevelTypeImpl.h"
29 
30 using namespace llvm;
31 
32 namespace {
33 
34 struct OutgoingValueHandler : public CallLowering::ValueHandler {
35   OutgoingValueHandler(MachineIRBuilder &B, MachineRegisterInfo &MRI,
36                        MachineInstrBuilder MIB, CCAssignFn *AssignFn)
37       : ValueHandler(B, MRI, AssignFn), MIB(MIB) {}
38 
39   MachineInstrBuilder MIB;
40 
41   bool isIncomingArgumentHandler() const override { return false; }
42 
43   Register getStackAddress(uint64_t Size, int64_t Offset,
44                            MachinePointerInfo &MPO) override {
45     llvm_unreachable("not implemented");
46   }
47 
48   void assignValueToAddress(Register ValVReg, Register Addr, uint64_t Size,
49                             MachinePointerInfo &MPO, CCValAssign &VA) override {
50     llvm_unreachable("not implemented");
51   }
52 
53   void assignValueToReg(Register ValVReg, Register PhysReg,
54                         CCValAssign &VA) override {
55     Register ExtReg;
56     if (VA.getLocVT().getSizeInBits() < 32) {
57       // 16-bit types are reported as legal for 32-bit registers. We need to
58       // extend and do a 32-bit copy to avoid the verifier complaining about it.
59       ExtReg = MIRBuilder.buildAnyExt(LLT::scalar(32), ValVReg).getReg(0);
60     } else
61       ExtReg = extendRegister(ValVReg, VA);
62 
63     // If this is a scalar return, insert a readfirstlane just in case the value
64     // ends up in a VGPR.
65     // FIXME: Assert this is a shader return.
66     const SIRegisterInfo *TRI
67       = static_cast<const SIRegisterInfo *>(MRI.getTargetRegisterInfo());
68     if (TRI->isSGPRReg(MRI, PhysReg)) {
69       auto ToSGPR = MIRBuilder.buildIntrinsic(Intrinsic::amdgcn_readfirstlane,
70                                               {MRI.getType(ExtReg)}, false)
71         .addReg(ExtReg);
72       ExtReg = ToSGPR.getReg(0);
73     }
74 
75     MIRBuilder.buildCopy(PhysReg, ExtReg);
76     MIB.addUse(PhysReg, RegState::Implicit);
77   }
78 
79   bool assignArg(unsigned ValNo, MVT ValVT, MVT LocVT,
80                  CCValAssign::LocInfo LocInfo,
81                  const CallLowering::ArgInfo &Info,
82                  ISD::ArgFlagsTy Flags,
83                  CCState &State) override {
84     return AssignFn(ValNo, ValVT, LocVT, LocInfo, Flags, State);
85   }
86 };
87 
88 struct IncomingArgHandler : public CallLowering::ValueHandler {
89   uint64_t StackUsed = 0;
90 
91   IncomingArgHandler(MachineIRBuilder &B, MachineRegisterInfo &MRI,
92                      CCAssignFn *AssignFn)
93     : ValueHandler(B, MRI, AssignFn) {}
94 
95   Register getStackAddress(uint64_t Size, int64_t Offset,
96                            MachinePointerInfo &MPO) override {
97     auto &MFI = MIRBuilder.getMF().getFrameInfo();
98     int FI = MFI.CreateFixedObject(Size, Offset, true);
99     MPO = MachinePointerInfo::getFixedStack(MIRBuilder.getMF(), FI);
100     auto AddrReg = MIRBuilder.buildFrameIndex(
101         LLT::pointer(AMDGPUAS::PRIVATE_ADDRESS, 32), FI);
102     StackUsed = std::max(StackUsed, Size + Offset);
103     return AddrReg.getReg(0);
104   }
105 
106   void assignValueToReg(Register ValVReg, Register PhysReg,
107                         CCValAssign &VA) override {
108     markPhysRegUsed(PhysReg);
109 
110     if (VA.getLocVT().getSizeInBits() < 32) {
111       // 16-bit types are reported as legal for 32-bit registers. We need to do
112       // a 32-bit copy, and truncate to avoid the verifier complaining about it.
113       auto Copy = MIRBuilder.buildCopy(LLT::scalar(32), PhysReg);
114       MIRBuilder.buildTrunc(ValVReg, Copy);
115       return;
116     }
117 
118     switch (VA.getLocInfo()) {
119     case CCValAssign::LocInfo::SExt:
120     case CCValAssign::LocInfo::ZExt:
121     case CCValAssign::LocInfo::AExt: {
122       auto Copy = MIRBuilder.buildCopy(LLT{VA.getLocVT()}, PhysReg);
123       MIRBuilder.buildTrunc(ValVReg, Copy);
124       break;
125     }
126     default:
127       MIRBuilder.buildCopy(ValVReg, PhysReg);
128       break;
129     }
130   }
131 
132   void assignValueToAddress(Register ValVReg, Register Addr, uint64_t Size,
133                             MachinePointerInfo &MPO, CCValAssign &VA) override {
134     MachineFunction &MF = MIRBuilder.getMF();
135 
136     // FIXME: Get alignment
137     auto MMO = MF.getMachineMemOperand(
138         MPO, MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant, Size,
139         inferAlignFromPtrInfo(MF, MPO));
140     MIRBuilder.buildLoad(ValVReg, Addr, *MMO);
141   }
142 
143   /// How the physical register gets marked varies between formal
144   /// parameters (it's a basic-block live-in), and a call instruction
145   /// (it's an implicit-def of the BL).
146   virtual void markPhysRegUsed(unsigned PhysReg) = 0;
147 
148   // FIXME: What is the point of this being a callback?
149   bool isIncomingArgumentHandler() const override { return true; }
150 };
151 
152 struct FormalArgHandler : public IncomingArgHandler {
153   FormalArgHandler(MachineIRBuilder &B, MachineRegisterInfo &MRI,
154                    CCAssignFn *AssignFn)
155     : IncomingArgHandler(B, MRI, AssignFn) {}
156 
157   void markPhysRegUsed(unsigned PhysReg) override {
158     MIRBuilder.getMBB().addLiveIn(PhysReg);
159   }
160 };
161 
162 }
163 
164 AMDGPUCallLowering::AMDGPUCallLowering(const AMDGPUTargetLowering &TLI)
165   : CallLowering(&TLI) {
166 }
167 
168 // FIXME: Compatability shim
169 static ISD::NodeType extOpcodeToISDExtOpcode(unsigned MIOpc) {
170   switch (MIOpc) {
171   case TargetOpcode::G_SEXT:
172     return ISD::SIGN_EXTEND;
173   case TargetOpcode::G_ZEXT:
174     return ISD::ZERO_EXTEND;
175   case TargetOpcode::G_ANYEXT:
176     return ISD::ANY_EXTEND;
177   default:
178     llvm_unreachable("not an extend opcode");
179   }
180 }
181 
182 void AMDGPUCallLowering::splitToValueTypes(
183   MachineIRBuilder &B,
184   const ArgInfo &OrigArg, unsigned OrigArgIdx,
185   SmallVectorImpl<ArgInfo> &SplitArgs,
186   const DataLayout &DL, CallingConv::ID CallConv,
187   SplitArgTy PerformArgSplit) const {
188   const SITargetLowering &TLI = *getTLI<SITargetLowering>();
189   LLVMContext &Ctx = OrigArg.Ty->getContext();
190 
191   if (OrigArg.Ty->isVoidTy())
192     return;
193 
194   SmallVector<EVT, 4> SplitVTs;
195   ComputeValueVTs(TLI, DL, OrigArg.Ty, SplitVTs);
196 
197   assert(OrigArg.Regs.size() == SplitVTs.size());
198 
199   int SplitIdx = 0;
200   for (EVT VT : SplitVTs) {
201     Register Reg = OrigArg.Regs[SplitIdx];
202     Type *Ty = VT.getTypeForEVT(Ctx);
203     LLT LLTy = getLLTForType(*Ty, DL);
204 
205     if (OrigArgIdx == AttributeList::ReturnIndex && VT.isScalarInteger()) {
206       unsigned ExtendOp = TargetOpcode::G_ANYEXT;
207       if (OrigArg.Flags[0].isSExt()) {
208         assert(OrigArg.Regs.size() == 1 && "expect only simple return values");
209         ExtendOp = TargetOpcode::G_SEXT;
210       } else if (OrigArg.Flags[0].isZExt()) {
211         assert(OrigArg.Regs.size() == 1 && "expect only simple return values");
212         ExtendOp = TargetOpcode::G_ZEXT;
213       }
214 
215       EVT ExtVT = TLI.getTypeForExtReturn(Ctx, VT,
216                                           extOpcodeToISDExtOpcode(ExtendOp));
217       if (ExtVT != VT) {
218         VT = ExtVT;
219         Ty = ExtVT.getTypeForEVT(Ctx);
220         LLTy = getLLTForType(*Ty, DL);
221         Reg = B.buildInstr(ExtendOp, {LLTy}, {Reg}).getReg(0);
222       }
223     }
224 
225     unsigned NumParts = TLI.getNumRegistersForCallingConv(Ctx, CallConv, VT);
226     MVT RegVT = TLI.getRegisterTypeForCallingConv(Ctx, CallConv, VT);
227 
228     if (NumParts == 1) {
229       // No splitting to do, but we want to replace the original type (e.g. [1 x
230       // double] -> double).
231       SplitArgs.emplace_back(Reg, Ty, OrigArg.Flags, OrigArg.IsFixed);
232 
233       ++SplitIdx;
234       continue;
235     }
236 
237     SmallVector<Register, 8> SplitRegs;
238     Type *PartTy = EVT(RegVT).getTypeForEVT(Ctx);
239     LLT PartLLT = getLLTForType(*PartTy, DL);
240     MachineRegisterInfo &MRI = *B.getMRI();
241 
242     // FIXME: Should we be reporting all of the part registers for a single
243     // argument, and let handleAssignments take care of the repacking?
244     for (unsigned i = 0; i < NumParts; ++i) {
245       Register PartReg = MRI.createGenericVirtualRegister(PartLLT);
246       SplitRegs.push_back(PartReg);
247       SplitArgs.emplace_back(ArrayRef<Register>(PartReg), PartTy, OrigArg.Flags);
248     }
249 
250     PerformArgSplit(SplitRegs, Reg, LLTy, PartLLT, SplitIdx);
251 
252     ++SplitIdx;
253   }
254 }
255 
256 // Get the appropriate type to make \p OrigTy \p Factor times bigger.
257 static LLT getMultipleType(LLT OrigTy, int Factor) {
258   if (OrigTy.isVector()) {
259     return LLT::vector(OrigTy.getNumElements() * Factor,
260                        OrigTy.getElementType());
261   }
262 
263   return LLT::scalar(OrigTy.getSizeInBits() * Factor);
264 }
265 
266 // TODO: Move to generic code
267 static void unpackRegsToOrigType(MachineIRBuilder &B,
268                                  ArrayRef<Register> DstRegs,
269                                  Register SrcReg,
270                                  const CallLowering::ArgInfo &Info,
271                                  LLT SrcTy,
272                                  LLT PartTy) {
273   assert(DstRegs.size() > 1 && "Nothing to unpack");
274 
275   const unsigned SrcSize = SrcTy.getSizeInBits();
276   const unsigned PartSize = PartTy.getSizeInBits();
277 
278   if (SrcTy.isVector() && !PartTy.isVector() &&
279       PartSize > SrcTy.getElementType().getSizeInBits()) {
280     // Vector was scalarized, and the elements extended.
281     auto UnmergeToEltTy = B.buildUnmerge(SrcTy.getElementType(),
282                                                   SrcReg);
283     for (int i = 0, e = DstRegs.size(); i != e; ++i)
284       B.buildAnyExt(DstRegs[i], UnmergeToEltTy.getReg(i));
285     return;
286   }
287 
288   if (SrcSize % PartSize == 0) {
289     B.buildUnmerge(DstRegs, SrcReg);
290     return;
291   }
292 
293   const int NumRoundedParts = (SrcSize + PartSize - 1) / PartSize;
294 
295   LLT BigTy = getMultipleType(PartTy, NumRoundedParts);
296   auto ImpDef = B.buildUndef(BigTy);
297 
298   auto Big = B.buildInsert(BigTy, ImpDef.getReg(0), SrcReg, 0).getReg(0);
299 
300   int64_t Offset = 0;
301   for (unsigned i = 0, e = DstRegs.size(); i != e; ++i, Offset += PartSize)
302     B.buildExtract(DstRegs[i], Big, Offset);
303 }
304 
305 /// Lower the return value for the already existing \p Ret. This assumes that
306 /// \p B's insertion point is correct.
307 bool AMDGPUCallLowering::lowerReturnVal(MachineIRBuilder &B,
308                                         const Value *Val, ArrayRef<Register> VRegs,
309                                         MachineInstrBuilder &Ret) const {
310   if (!Val)
311     return true;
312 
313   auto &MF = B.getMF();
314   const auto &F = MF.getFunction();
315   const DataLayout &DL = MF.getDataLayout();
316   MachineRegisterInfo *MRI = B.getMRI();
317 
318   CallingConv::ID CC = F.getCallingConv();
319   const SITargetLowering &TLI = *getTLI<SITargetLowering>();
320 
321   ArgInfo OrigRetInfo(VRegs, Val->getType());
322   setArgFlags(OrigRetInfo, AttributeList::ReturnIndex, DL, F);
323   SmallVector<ArgInfo, 4> SplitRetInfos;
324 
325   splitToValueTypes(
326     B, OrigRetInfo, AttributeList::ReturnIndex, SplitRetInfos, DL, CC,
327     [&](ArrayRef<Register> Regs, Register SrcReg, LLT LLTy, LLT PartLLT,
328         int VTSplitIdx) {
329       unpackRegsToOrigType(B, Regs, SrcReg,
330                            SplitRetInfos[VTSplitIdx],
331                            LLTy, PartLLT);
332     });
333 
334   CCAssignFn *AssignFn = TLI.CCAssignFnForReturn(CC, F.isVarArg());
335   OutgoingValueHandler RetHandler(B, *MRI, Ret, AssignFn);
336   return handleAssignments(B, SplitRetInfos, RetHandler);
337 }
338 
339 bool AMDGPUCallLowering::lowerReturn(MachineIRBuilder &B,
340                                      const Value *Val,
341                                      ArrayRef<Register> VRegs) const {
342 
343   MachineFunction &MF = B.getMF();
344   MachineRegisterInfo &MRI = MF.getRegInfo();
345   SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
346   MFI->setIfReturnsVoid(!Val);
347 
348   assert(!Val == VRegs.empty() && "Return value without a vreg");
349 
350   CallingConv::ID CC = B.getMF().getFunction().getCallingConv();
351   const bool IsShader = AMDGPU::isShader(CC);
352   const bool IsWaveEnd = (IsShader && MFI->returnsVoid()) ||
353                          AMDGPU::isKernel(CC);
354   if (IsWaveEnd) {
355     B.buildInstr(AMDGPU::S_ENDPGM)
356       .addImm(0);
357     return true;
358   }
359 
360   auto const &ST = MF.getSubtarget<GCNSubtarget>();
361 
362   unsigned ReturnOpc =
363       IsShader ? AMDGPU::SI_RETURN_TO_EPILOG : AMDGPU::S_SETPC_B64_return;
364 
365   auto Ret = B.buildInstrNoInsert(ReturnOpc);
366   Register ReturnAddrVReg;
367   if (ReturnOpc == AMDGPU::S_SETPC_B64_return) {
368     ReturnAddrVReg = MRI.createVirtualRegister(&AMDGPU::CCR_SGPR_64RegClass);
369     Ret.addUse(ReturnAddrVReg);
370   }
371 
372   if (!lowerReturnVal(B, Val, VRegs, Ret))
373     return false;
374 
375   if (ReturnOpc == AMDGPU::S_SETPC_B64_return) {
376     const SIRegisterInfo *TRI = ST.getRegisterInfo();
377     Register LiveInReturn = MF.addLiveIn(TRI->getReturnAddressReg(MF),
378                                          &AMDGPU::SGPR_64RegClass);
379     B.buildCopy(ReturnAddrVReg, LiveInReturn);
380   }
381 
382   // TODO: Handle CalleeSavedRegsViaCopy.
383 
384   B.insertInstr(Ret);
385   return true;
386 }
387 
388 Register AMDGPUCallLowering::lowerParameterPtr(MachineIRBuilder &B,
389                                                Type *ParamTy,
390                                                uint64_t Offset) const {
391 
392   MachineFunction &MF = B.getMF();
393   const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
394   MachineRegisterInfo &MRI = MF.getRegInfo();
395   const Function &F = MF.getFunction();
396   const DataLayout &DL = F.getParent()->getDataLayout();
397   PointerType *PtrTy = PointerType::get(ParamTy, AMDGPUAS::CONSTANT_ADDRESS);
398   LLT PtrType = getLLTForType(*PtrTy, DL);
399   Register KernArgSegmentPtr =
400     MFI->getPreloadedReg(AMDGPUFunctionArgInfo::KERNARG_SEGMENT_PTR);
401   Register KernArgSegmentVReg = MRI.getLiveInVirtReg(KernArgSegmentPtr);
402 
403   auto OffsetReg = B.buildConstant(LLT::scalar(64), Offset);
404 
405   return B.buildPtrAdd(PtrType, KernArgSegmentVReg, OffsetReg).getReg(0);
406 }
407 
408 void AMDGPUCallLowering::lowerParameter(MachineIRBuilder &B, Type *ParamTy,
409                                         uint64_t Offset, Align Alignment,
410                                         Register DstReg) const {
411   MachineFunction &MF = B.getMF();
412   const Function &F = MF.getFunction();
413   const DataLayout &DL = F.getParent()->getDataLayout();
414   MachinePointerInfo PtrInfo(AMDGPUAS::CONSTANT_ADDRESS);
415   unsigned TypeSize = DL.getTypeStoreSize(ParamTy);
416   Register PtrReg = lowerParameterPtr(B, ParamTy, Offset);
417 
418   MachineMemOperand *MMO = MF.getMachineMemOperand(
419       PtrInfo,
420       MachineMemOperand::MOLoad | MachineMemOperand::MODereferenceable |
421           MachineMemOperand::MOInvariant,
422       TypeSize, Alignment);
423 
424   B.buildLoad(DstReg, PtrReg, *MMO);
425 }
426 
427 // Allocate special inputs passed in user SGPRs.
428 static void allocateHSAUserSGPRs(CCState &CCInfo,
429                                  MachineIRBuilder &B,
430                                  MachineFunction &MF,
431                                  const SIRegisterInfo &TRI,
432                                  SIMachineFunctionInfo &Info) {
433   // FIXME: How should these inputs interact with inreg / custom SGPR inputs?
434   if (Info.hasPrivateSegmentBuffer()) {
435     Register PrivateSegmentBufferReg = Info.addPrivateSegmentBuffer(TRI);
436     MF.addLiveIn(PrivateSegmentBufferReg, &AMDGPU::SGPR_128RegClass);
437     CCInfo.AllocateReg(PrivateSegmentBufferReg);
438   }
439 
440   if (Info.hasDispatchPtr()) {
441     Register DispatchPtrReg = Info.addDispatchPtr(TRI);
442     MF.addLiveIn(DispatchPtrReg, &AMDGPU::SGPR_64RegClass);
443     CCInfo.AllocateReg(DispatchPtrReg);
444   }
445 
446   if (Info.hasQueuePtr()) {
447     Register QueuePtrReg = Info.addQueuePtr(TRI);
448     MF.addLiveIn(QueuePtrReg, &AMDGPU::SGPR_64RegClass);
449     CCInfo.AllocateReg(QueuePtrReg);
450   }
451 
452   if (Info.hasKernargSegmentPtr()) {
453     MachineRegisterInfo &MRI = MF.getRegInfo();
454     Register InputPtrReg = Info.addKernargSegmentPtr(TRI);
455     const LLT P4 = LLT::pointer(AMDGPUAS::CONSTANT_ADDRESS, 64);
456     Register VReg = MRI.createGenericVirtualRegister(P4);
457     MRI.addLiveIn(InputPtrReg, VReg);
458     B.getMBB().addLiveIn(InputPtrReg);
459     B.buildCopy(VReg, InputPtrReg);
460     CCInfo.AllocateReg(InputPtrReg);
461   }
462 
463   if (Info.hasDispatchID()) {
464     Register DispatchIDReg = Info.addDispatchID(TRI);
465     MF.addLiveIn(DispatchIDReg, &AMDGPU::SGPR_64RegClass);
466     CCInfo.AllocateReg(DispatchIDReg);
467   }
468 
469   if (Info.hasFlatScratchInit()) {
470     Register FlatScratchInitReg = Info.addFlatScratchInit(TRI);
471     MF.addLiveIn(FlatScratchInitReg, &AMDGPU::SGPR_64RegClass);
472     CCInfo.AllocateReg(FlatScratchInitReg);
473   }
474 
475   // TODO: Add GridWorkGroupCount user SGPRs when used. For now with HSA we read
476   // these from the dispatch pointer.
477 }
478 
479 bool AMDGPUCallLowering::lowerFormalArgumentsKernel(
480     MachineIRBuilder &B, const Function &F,
481     ArrayRef<ArrayRef<Register>> VRegs) const {
482   MachineFunction &MF = B.getMF();
483   const GCNSubtarget *Subtarget = &MF.getSubtarget<GCNSubtarget>();
484   MachineRegisterInfo &MRI = MF.getRegInfo();
485   SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
486   const SIRegisterInfo *TRI = Subtarget->getRegisterInfo();
487   const SITargetLowering &TLI = *getTLI<SITargetLowering>();
488 
489   const DataLayout &DL = F.getParent()->getDataLayout();
490 
491   SmallVector<CCValAssign, 16> ArgLocs;
492   CCState CCInfo(F.getCallingConv(), F.isVarArg(), MF, ArgLocs, F.getContext());
493 
494   allocateHSAUserSGPRs(CCInfo, B, MF, *TRI, *Info);
495 
496   unsigned i = 0;
497   const Align KernArgBaseAlign(16);
498   const unsigned BaseOffset = Subtarget->getExplicitKernelArgOffset(F);
499   uint64_t ExplicitArgOffset = 0;
500 
501   // TODO: Align down to dword alignment and extract bits for extending loads.
502   for (auto &Arg : F.args()) {
503     Type *ArgTy = Arg.getType();
504     unsigned AllocSize = DL.getTypeAllocSize(ArgTy);
505     if (AllocSize == 0)
506       continue;
507 
508     Align ABIAlign = DL.getABITypeAlign(ArgTy);
509 
510     uint64_t ArgOffset = alignTo(ExplicitArgOffset, ABIAlign) + BaseOffset;
511     ExplicitArgOffset = alignTo(ExplicitArgOffset, ABIAlign) + AllocSize;
512 
513     if (Arg.use_empty()) {
514       ++i;
515       continue;
516     }
517 
518     ArrayRef<Register> OrigArgRegs = VRegs[i];
519     Register ArgReg =
520       OrigArgRegs.size() == 1
521       ? OrigArgRegs[0]
522       : MRI.createGenericVirtualRegister(getLLTForType(*ArgTy, DL));
523 
524     Align Alignment = commonAlignment(KernArgBaseAlign, ArgOffset);
525     lowerParameter(B, ArgTy, ArgOffset, Alignment, ArgReg);
526     if (OrigArgRegs.size() > 1)
527       unpackRegs(OrigArgRegs, ArgReg, ArgTy, B);
528     ++i;
529   }
530 
531   TLI.allocateSpecialEntryInputVGPRs(CCInfo, MF, *TRI, *Info);
532   TLI.allocateSystemSGPRs(CCInfo, MF, *Info, F.getCallingConv(), false);
533   return true;
534 }
535 
536 /// Pack values \p SrcRegs to cover the vector type result \p DstRegs.
537 static MachineInstrBuilder mergeVectorRegsToResultRegs(
538   MachineIRBuilder &B, ArrayRef<Register> DstRegs, ArrayRef<Register> SrcRegs) {
539   MachineRegisterInfo &MRI = *B.getMRI();
540   LLT LLTy = MRI.getType(DstRegs[0]);
541   LLT PartLLT = MRI.getType(SrcRegs[0]);
542 
543   // Deal with v3s16 split into v2s16
544   LLT LCMTy = getLCMType(LLTy, PartLLT);
545   if (LCMTy == LLTy) {
546     // Common case where no padding is needed.
547     assert(DstRegs.size() == 1);
548     return B.buildConcatVectors(DstRegs[0], SrcRegs);
549   }
550 
551   const int NumWide =  LCMTy.getSizeInBits() / PartLLT.getSizeInBits();
552   Register Undef = B.buildUndef(PartLLT).getReg(0);
553 
554   // Build vector of undefs.
555   SmallVector<Register, 8> WidenedSrcs(NumWide, Undef);
556 
557   // Replace the first sources with the real registers.
558   std::copy(SrcRegs.begin(), SrcRegs.end(), WidenedSrcs.begin());
559 
560   auto Widened = B.buildConcatVectors(LCMTy, WidenedSrcs);
561   int NumDst = LCMTy.getSizeInBits() / LLTy.getSizeInBits();
562 
563   SmallVector<Register, 8> PadDstRegs(NumDst);
564   std::copy(DstRegs.begin(), DstRegs.end(), PadDstRegs.begin());
565 
566   // Create the excess dead defs for the unmerge.
567   for (int I = DstRegs.size(); I != NumDst; ++I)
568     PadDstRegs[I] = MRI.createGenericVirtualRegister(LLTy);
569 
570   return B.buildUnmerge(PadDstRegs, Widened);
571 }
572 
573 // TODO: Move this to generic code
574 static void packSplitRegsToOrigType(MachineIRBuilder &B,
575                                     ArrayRef<Register> OrigRegs,
576                                     ArrayRef<Register> Regs,
577                                     LLT LLTy,
578                                     LLT PartLLT) {
579   MachineRegisterInfo &MRI = *B.getMRI();
580 
581   if (!LLTy.isVector() && !PartLLT.isVector()) {
582     assert(OrigRegs.size() == 1);
583     LLT OrigTy = MRI.getType(OrigRegs[0]);
584 
585     unsigned SrcSize = PartLLT.getSizeInBits() * Regs.size();
586     if (SrcSize == OrigTy.getSizeInBits())
587       B.buildMerge(OrigRegs[0], Regs);
588     else {
589       auto Widened = B.buildMerge(LLT::scalar(SrcSize), Regs);
590       B.buildTrunc(OrigRegs[0], Widened);
591     }
592 
593     return;
594   }
595 
596   if (LLTy.isVector() && PartLLT.isVector()) {
597     assert(OrigRegs.size() == 1);
598     assert(LLTy.getElementType() == PartLLT.getElementType());
599     mergeVectorRegsToResultRegs(B, OrigRegs, Regs);
600     return;
601   }
602 
603   assert(LLTy.isVector() && !PartLLT.isVector());
604 
605   LLT DstEltTy = LLTy.getElementType();
606 
607   // Pointer information was discarded. We'll need to coerce some register types
608   // to avoid violating type constraints.
609   LLT RealDstEltTy = MRI.getType(OrigRegs[0]).getElementType();
610 
611   assert(DstEltTy.getSizeInBits() == RealDstEltTy.getSizeInBits());
612 
613   if (DstEltTy == PartLLT) {
614     // Vector was trivially scalarized.
615 
616     if (RealDstEltTy.isPointer()) {
617       for (Register Reg : Regs)
618         MRI.setType(Reg, RealDstEltTy);
619     }
620 
621     B.buildBuildVector(OrigRegs[0], Regs);
622   } else if (DstEltTy.getSizeInBits() > PartLLT.getSizeInBits()) {
623     // Deal with vector with 64-bit elements decomposed to 32-bit
624     // registers. Need to create intermediate 64-bit elements.
625     SmallVector<Register, 8> EltMerges;
626     int PartsPerElt = DstEltTy.getSizeInBits() / PartLLT.getSizeInBits();
627 
628     assert(DstEltTy.getSizeInBits() % PartLLT.getSizeInBits() == 0);
629 
630     for (int I = 0, NumElts = LLTy.getNumElements(); I != NumElts; ++I)  {
631       auto Merge = B.buildMerge(RealDstEltTy, Regs.take_front(PartsPerElt));
632       // Fix the type in case this is really a vector of pointers.
633       MRI.setType(Merge.getReg(0), RealDstEltTy);
634       EltMerges.push_back(Merge.getReg(0));
635       Regs = Regs.drop_front(PartsPerElt);
636     }
637 
638     B.buildBuildVector(OrigRegs[0], EltMerges);
639   } else {
640     // Vector was split, and elements promoted to a wider type.
641     LLT BVType = LLT::vector(LLTy.getNumElements(), PartLLT);
642     auto BV = B.buildBuildVector(BVType, Regs);
643     B.buildTrunc(OrigRegs[0], BV);
644   }
645 }
646 
647 bool AMDGPUCallLowering::lowerFormalArguments(
648     MachineIRBuilder &B, const Function &F,
649     ArrayRef<ArrayRef<Register>> VRegs) const {
650   CallingConv::ID CC = F.getCallingConv();
651 
652   // The infrastructure for normal calling convention lowering is essentially
653   // useless for kernels. We want to avoid any kind of legalization or argument
654   // splitting.
655   if (CC == CallingConv::AMDGPU_KERNEL)
656     return lowerFormalArgumentsKernel(B, F, VRegs);
657 
658   const bool IsShader = AMDGPU::isShader(CC);
659   const bool IsEntryFunc = AMDGPU::isEntryFunctionCC(CC);
660 
661   MachineFunction &MF = B.getMF();
662   MachineBasicBlock &MBB = B.getMBB();
663   MachineRegisterInfo &MRI = MF.getRegInfo();
664   SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
665   const GCNSubtarget &Subtarget = MF.getSubtarget<GCNSubtarget>();
666   const SIRegisterInfo *TRI = Subtarget.getRegisterInfo();
667   const DataLayout &DL = F.getParent()->getDataLayout();
668 
669 
670   SmallVector<CCValAssign, 16> ArgLocs;
671   CCState CCInfo(CC, F.isVarArg(), MF, ArgLocs, F.getContext());
672 
673   if (!IsEntryFunc) {
674     Register ReturnAddrReg = TRI->getReturnAddressReg(MF);
675     Register LiveInReturn = MF.addLiveIn(ReturnAddrReg,
676                                          &AMDGPU::SGPR_64RegClass);
677     MBB.addLiveIn(ReturnAddrReg);
678     B.buildCopy(LiveInReturn, ReturnAddrReg);
679   }
680 
681   if (Info->hasImplicitBufferPtr()) {
682     Register ImplicitBufferPtrReg = Info->addImplicitBufferPtr(*TRI);
683     MF.addLiveIn(ImplicitBufferPtrReg, &AMDGPU::SGPR_64RegClass);
684     CCInfo.AllocateReg(ImplicitBufferPtrReg);
685   }
686 
687 
688   SmallVector<ArgInfo, 32> SplitArgs;
689   unsigned Idx = 0;
690   unsigned PSInputNum = 0;
691 
692   for (auto &Arg : F.args()) {
693     if (DL.getTypeStoreSize(Arg.getType()) == 0)
694       continue;
695 
696     const bool InReg = Arg.hasAttribute(Attribute::InReg);
697 
698     // SGPR arguments to functions not implemented.
699     if (!IsShader && InReg)
700       return false;
701 
702     if (Arg.hasAttribute(Attribute::SwiftSelf) ||
703         Arg.hasAttribute(Attribute::SwiftError) ||
704         Arg.hasAttribute(Attribute::Nest))
705       return false;
706 
707     if (CC == CallingConv::AMDGPU_PS && !InReg && PSInputNum <= 15) {
708       const bool ArgUsed = !Arg.use_empty();
709       bool SkipArg = !ArgUsed && !Info->isPSInputAllocated(PSInputNum);
710 
711       if (!SkipArg) {
712         Info->markPSInputAllocated(PSInputNum);
713         if (ArgUsed)
714           Info->markPSInputEnabled(PSInputNum);
715       }
716 
717       ++PSInputNum;
718 
719       if (SkipArg) {
720         for (int I = 0, E = VRegs[Idx].size(); I != E; ++I)
721           B.buildUndef(VRegs[Idx][I]);
722 
723         ++Idx;
724         continue;
725       }
726     }
727 
728     ArgInfo OrigArg(VRegs[Idx], Arg.getType());
729     const unsigned OrigArgIdx = Idx + AttributeList::FirstArgIndex;
730     setArgFlags(OrigArg, OrigArgIdx, DL, F);
731 
732     splitToValueTypes(
733       B, OrigArg, OrigArgIdx, SplitArgs, DL, CC,
734       // FIXME: We should probably be passing multiple registers to
735       // handleAssignments to do this
736       [&](ArrayRef<Register> Regs, Register DstReg,
737           LLT LLTy, LLT PartLLT, int VTSplitIdx) {
738         assert(DstReg == VRegs[Idx][VTSplitIdx]);
739         packSplitRegsToOrigType(B, VRegs[Idx][VTSplitIdx], Regs,
740                                 LLTy, PartLLT);
741       });
742 
743     ++Idx;
744   }
745 
746   // At least one interpolation mode must be enabled or else the GPU will
747   // hang.
748   //
749   // Check PSInputAddr instead of PSInputEnable. The idea is that if the user
750   // set PSInputAddr, the user wants to enable some bits after the compilation
751   // based on run-time states. Since we can't know what the final PSInputEna
752   // will look like, so we shouldn't do anything here and the user should take
753   // responsibility for the correct programming.
754   //
755   // Otherwise, the following restrictions apply:
756   // - At least one of PERSP_* (0xF) or LINEAR_* (0x70) must be enabled.
757   // - If POS_W_FLOAT (11) is enabled, at least one of PERSP_* must be
758   //   enabled too.
759   if (CC == CallingConv::AMDGPU_PS) {
760     if ((Info->getPSInputAddr() & 0x7F) == 0 ||
761         ((Info->getPSInputAddr() & 0xF) == 0 &&
762          Info->isPSInputAllocated(11))) {
763       CCInfo.AllocateReg(AMDGPU::VGPR0);
764       CCInfo.AllocateReg(AMDGPU::VGPR1);
765       Info->markPSInputAllocated(0);
766       Info->markPSInputEnabled(0);
767     }
768 
769     if (Subtarget.isAmdPalOS()) {
770       // For isAmdPalOS, the user does not enable some bits after compilation
771       // based on run-time states; the register values being generated here are
772       // the final ones set in hardware. Therefore we need to apply the
773       // workaround to PSInputAddr and PSInputEnable together.  (The case where
774       // a bit is set in PSInputAddr but not PSInputEnable is where the frontend
775       // set up an input arg for a particular interpolation mode, but nothing
776       // uses that input arg. Really we should have an earlier pass that removes
777       // such an arg.)
778       unsigned PsInputBits = Info->getPSInputAddr() & Info->getPSInputEnable();
779       if ((PsInputBits & 0x7F) == 0 ||
780           ((PsInputBits & 0xF) == 0 &&
781            (PsInputBits >> 11 & 1)))
782         Info->markPSInputEnabled(
783           countTrailingZeros(Info->getPSInputAddr(), ZB_Undefined));
784     }
785   }
786 
787   const SITargetLowering &TLI = *getTLI<SITargetLowering>();
788   CCAssignFn *AssignFn = TLI.CCAssignFnForCall(CC, F.isVarArg());
789 
790   if (!MBB.empty())
791     B.setInstr(*MBB.begin());
792 
793   if (!IsEntryFunc) {
794     // For the fixed ABI, pass workitem IDs in the last argument register.
795     if (AMDGPUTargetMachine::EnableFixedFunctionABI)
796       TLI.allocateSpecialInputVGPRsFixed(CCInfo, MF, *TRI, *Info);
797   }
798 
799   FormalArgHandler Handler(B, MRI, AssignFn);
800   if (!handleAssignments(CCInfo, ArgLocs, B, SplitArgs, Handler))
801     return false;
802 
803   if (!IsEntryFunc && !AMDGPUTargetMachine::EnableFixedFunctionABI) {
804     // Special inputs come after user arguments.
805     TLI.allocateSpecialInputVGPRs(CCInfo, MF, *TRI, *Info);
806   }
807 
808   // Start adding system SGPRs.
809   if (IsEntryFunc) {
810     TLI.allocateSystemSGPRs(CCInfo, MF, *Info, CC, IsShader);
811   } else {
812     CCInfo.AllocateReg(Info->getScratchRSrcReg());
813     TLI.allocateSpecialInputSGPRs(CCInfo, MF, *TRI, *Info);
814   }
815 
816   // Move back to the end of the basic block.
817   B.setMBB(MBB);
818 
819   return true;
820 }
821