xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/AMDGPUCallLowering.cpp (revision 5b27928474e6a4103d65b347544705c40c9618fd)
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 "SIISelLowering.h"
20 #include "SIMachineFunctionInfo.h"
21 #include "SIRegisterInfo.h"
22 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
23 #include "llvm/CodeGen/Analysis.h"
24 #include "llvm/CodeGen/CallingConvLower.h"
25 #include "llvm/CodeGen/GlobalISel/MachineIRBuilder.h"
26 #include "llvm/CodeGen/MachineInstrBuilder.h"
27 #include "llvm/Support/LowLevelTypeImpl.h"
28 
29 using namespace llvm;
30 
31 namespace {
32 
33 struct OutgoingValueHandler : public CallLowering::ValueHandler {
34   OutgoingValueHandler(MachineIRBuilder &B, MachineRegisterInfo &MRI,
35                        MachineInstrBuilder MIB, CCAssignFn *AssignFn)
36       : ValueHandler(B, MRI, AssignFn), MIB(MIB) {}
37 
38   MachineInstrBuilder MIB;
39 
40   bool isIncomingArgumentHandler() const override { return false; }
41 
42   Register getStackAddress(uint64_t Size, int64_t Offset,
43                            MachinePointerInfo &MPO) override {
44     llvm_unreachable("not implemented");
45   }
46 
47   void assignValueToAddress(Register ValVReg, Register Addr, uint64_t Size,
48                             MachinePointerInfo &MPO, CCValAssign &VA) override {
49     llvm_unreachable("not implemented");
50   }
51 
52   void assignValueToReg(Register ValVReg, Register PhysReg,
53                         CCValAssign &VA) override {
54     Register ExtReg;
55     if (VA.getLocVT().getSizeInBits() < 32) {
56       // 16-bit types are reported as legal for 32-bit registers. We need to
57       // extend and do a 32-bit copy to avoid the verifier complaining about it.
58       ExtReg = MIRBuilder.buildAnyExt(LLT::scalar(32), ValVReg).getReg(0);
59     } else
60       ExtReg = extendRegister(ValVReg, VA);
61 
62     MIRBuilder.buildCopy(PhysReg, ExtReg);
63     MIB.addUse(PhysReg, RegState::Implicit);
64   }
65 
66   bool assignArg(unsigned ValNo, MVT ValVT, MVT LocVT,
67                  CCValAssign::LocInfo LocInfo,
68                  const CallLowering::ArgInfo &Info,
69                  ISD::ArgFlagsTy Flags,
70                  CCState &State) override {
71     return AssignFn(ValNo, ValVT, LocVT, LocInfo, Flags, State);
72   }
73 };
74 
75 struct IncomingArgHandler : public CallLowering::ValueHandler {
76   uint64_t StackUsed = 0;
77 
78   IncomingArgHandler(MachineIRBuilder &B, MachineRegisterInfo &MRI,
79                      CCAssignFn *AssignFn)
80     : ValueHandler(B, MRI, AssignFn) {}
81 
82   Register getStackAddress(uint64_t Size, int64_t Offset,
83                            MachinePointerInfo &MPO) override {
84     auto &MFI = MIRBuilder.getMF().getFrameInfo();
85     int FI = MFI.CreateFixedObject(Size, Offset, true);
86     MPO = MachinePointerInfo::getFixedStack(MIRBuilder.getMF(), FI);
87     Register AddrReg = MRI.createGenericVirtualRegister(
88       LLT::pointer(AMDGPUAS::PRIVATE_ADDRESS, 32));
89     MIRBuilder.buildFrameIndex(AddrReg, FI);
90     StackUsed = std::max(StackUsed, Size + Offset);
91     return AddrReg;
92   }
93 
94   void assignValueToReg(Register ValVReg, Register PhysReg,
95                         CCValAssign &VA) override {
96     markPhysRegUsed(PhysReg);
97 
98     if (VA.getLocVT().getSizeInBits() < 32) {
99       // 16-bit types are reported as legal for 32-bit registers. We need to do
100       // a 32-bit copy, and truncate to avoid the verifier complaining about it.
101       auto Copy = MIRBuilder.buildCopy(LLT::scalar(32), PhysReg);
102       MIRBuilder.buildTrunc(ValVReg, Copy);
103       return;
104     }
105 
106     switch (VA.getLocInfo()) {
107     case CCValAssign::LocInfo::SExt:
108     case CCValAssign::LocInfo::ZExt:
109     case CCValAssign::LocInfo::AExt: {
110       auto Copy = MIRBuilder.buildCopy(LLT{VA.getLocVT()}, PhysReg);
111       MIRBuilder.buildTrunc(ValVReg, Copy);
112       break;
113     }
114     default:
115       MIRBuilder.buildCopy(ValVReg, PhysReg);
116       break;
117     }
118   }
119 
120   void assignValueToAddress(Register ValVReg, Register Addr, uint64_t Size,
121                             MachinePointerInfo &MPO, CCValAssign &VA) override {
122     // FIXME: Get alignment
123     auto MMO = MIRBuilder.getMF().getMachineMemOperand(
124       MPO, MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant, Size, 1);
125     MIRBuilder.buildLoad(ValVReg, Addr, *MMO);
126   }
127 
128   /// How the physical register gets marked varies between formal
129   /// parameters (it's a basic-block live-in), and a call instruction
130   /// (it's an implicit-def of the BL).
131   virtual void markPhysRegUsed(unsigned PhysReg) = 0;
132 
133   // FIXME: What is the point of this being a callback?
134   bool isIncomingArgumentHandler() const override { return true; }
135 };
136 
137 struct FormalArgHandler : public IncomingArgHandler {
138   FormalArgHandler(MachineIRBuilder &B, MachineRegisterInfo &MRI,
139                    CCAssignFn *AssignFn)
140     : IncomingArgHandler(B, MRI, AssignFn) {}
141 
142   void markPhysRegUsed(unsigned PhysReg) override {
143     MIRBuilder.getMBB().addLiveIn(PhysReg);
144   }
145 };
146 
147 }
148 
149 AMDGPUCallLowering::AMDGPUCallLowering(const AMDGPUTargetLowering &TLI)
150   : CallLowering(&TLI) {
151 }
152 
153 void AMDGPUCallLowering::splitToValueTypes(
154     const ArgInfo &OrigArg, SmallVectorImpl<ArgInfo> &SplitArgs,
155     const DataLayout &DL, MachineRegisterInfo &MRI, CallingConv::ID CallConv,
156     SplitArgTy PerformArgSplit) const {
157   const SITargetLowering &TLI = *getTLI<SITargetLowering>();
158   LLVMContext &Ctx = OrigArg.Ty->getContext();
159 
160   if (OrigArg.Ty->isVoidTy())
161     return;
162 
163   SmallVector<EVT, 4> SplitVTs;
164   ComputeValueVTs(TLI, DL, OrigArg.Ty, SplitVTs);
165 
166   assert(OrigArg.Regs.size() == SplitVTs.size());
167 
168   int SplitIdx = 0;
169   for (EVT VT : SplitVTs) {
170     unsigned NumParts = TLI.getNumRegistersForCallingConv(Ctx, CallConv, VT);
171     Type *Ty = VT.getTypeForEVT(Ctx);
172 
173 
174 
175     if (NumParts == 1) {
176       // No splitting to do, but we want to replace the original type (e.g. [1 x
177       // double] -> double).
178       SplitArgs.emplace_back(OrigArg.Regs[SplitIdx], Ty,
179                              OrigArg.Flags, OrigArg.IsFixed);
180 
181       ++SplitIdx;
182       continue;
183     }
184 
185     LLT LLTy = getLLTForType(*Ty, DL);
186 
187     SmallVector<Register, 8> SplitRegs;
188 
189     EVT PartVT = TLI.getRegisterTypeForCallingConv(Ctx, CallConv, VT);
190     Type *PartTy = PartVT.getTypeForEVT(Ctx);
191     LLT PartLLT = getLLTForType(*PartTy, DL);
192 
193     // FIXME: Should we be reporting all of the part registers for a single
194     // argument, and let handleAssignments take care of the repacking?
195     for (unsigned i = 0; i < NumParts; ++i) {
196       Register PartReg = MRI.createGenericVirtualRegister(PartLLT);
197       SplitRegs.push_back(PartReg);
198       SplitArgs.emplace_back(ArrayRef<Register>(PartReg), PartTy, OrigArg.Flags);
199     }
200 
201     PerformArgSplit(SplitRegs, LLTy, PartLLT, SplitIdx);
202 
203     ++SplitIdx;
204   }
205 }
206 
207 // Get the appropriate type to make \p OrigTy \p Factor times bigger.
208 static LLT getMultipleType(LLT OrigTy, int Factor) {
209   if (OrigTy.isVector()) {
210     return LLT::vector(OrigTy.getNumElements() * Factor,
211                        OrigTy.getElementType());
212   }
213 
214   return LLT::scalar(OrigTy.getSizeInBits() * Factor);
215 }
216 
217 // TODO: Move to generic code
218 static void unpackRegsToOrigType(MachineIRBuilder &B,
219                                  ArrayRef<Register> DstRegs,
220                                  Register SrcReg,
221                                  LLT SrcTy,
222                                  LLT PartTy) {
223   assert(DstRegs.size() > 1 && "Nothing to unpack");
224 
225   MachineFunction &MF = B.getMF();
226   MachineRegisterInfo &MRI = MF.getRegInfo();
227 
228   const unsigned SrcSize = SrcTy.getSizeInBits();
229   const unsigned PartSize = PartTy.getSizeInBits();
230 
231   if (SrcTy.isVector() && !PartTy.isVector() &&
232       PartSize > SrcTy.getElementType().getSizeInBits()) {
233     // Vector was scalarized, and the elements extended.
234     auto UnmergeToEltTy = B.buildUnmerge(SrcTy.getElementType(),
235                                                   SrcReg);
236     for (int i = 0, e = DstRegs.size(); i != e; ++i)
237       B.buildAnyExt(DstRegs[i], UnmergeToEltTy.getReg(i));
238     return;
239   }
240 
241   if (SrcSize % PartSize == 0) {
242     B.buildUnmerge(DstRegs, SrcReg);
243     return;
244   }
245 
246   const int NumRoundedParts = (SrcSize + PartSize - 1) / PartSize;
247 
248   LLT BigTy = getMultipleType(PartTy, NumRoundedParts);
249   auto ImpDef = B.buildUndef(BigTy);
250 
251   Register BigReg = MRI.createGenericVirtualRegister(BigTy);
252   B.buildInsert(BigReg, ImpDef.getReg(0), SrcReg, 0).getReg(0);
253 
254   int64_t Offset = 0;
255   for (unsigned i = 0, e = DstRegs.size(); i != e; ++i, Offset += PartSize)
256     B.buildExtract(DstRegs[i], BigReg, Offset);
257 }
258 
259 /// Lower the return value for the already existing \p Ret. This assumes that
260 /// \p B's insertion point is correct.
261 bool AMDGPUCallLowering::lowerReturnVal(MachineIRBuilder &B,
262                                         const Value *Val, ArrayRef<Register> VRegs,
263                                         MachineInstrBuilder &Ret) const {
264   if (!Val)
265     return true;
266 
267   auto &MF = B.getMF();
268   const auto &F = MF.getFunction();
269   const DataLayout &DL = MF.getDataLayout();
270 
271   CallingConv::ID CC = F.getCallingConv();
272   const SITargetLowering &TLI = *getTLI<SITargetLowering>();
273   MachineRegisterInfo &MRI = MF.getRegInfo();
274 
275   ArgInfo OrigRetInfo(VRegs, Val->getType());
276   setArgFlags(OrigRetInfo, AttributeList::ReturnIndex, DL, F);
277   SmallVector<ArgInfo, 4> SplitRetInfos;
278 
279   splitToValueTypes(
280     OrigRetInfo, SplitRetInfos, DL, MRI, CC,
281     [&](ArrayRef<Register> Regs, LLT LLTy, LLT PartLLT, int VTSplitIdx) {
282       unpackRegsToOrigType(B, Regs, VRegs[VTSplitIdx], LLTy, PartLLT);
283     });
284 
285   CCAssignFn *AssignFn = TLI.CCAssignFnForReturn(CC, F.isVarArg());
286 
287   OutgoingValueHandler RetHandler(B, MF.getRegInfo(), Ret, AssignFn);
288   return handleAssignments(B, SplitRetInfos, RetHandler);
289 }
290 
291 bool AMDGPUCallLowering::lowerReturn(MachineIRBuilder &B,
292                                      const Value *Val,
293                                      ArrayRef<Register> VRegs) const {
294 
295   MachineFunction &MF = B.getMF();
296   MachineRegisterInfo &MRI = MF.getRegInfo();
297   SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
298   MFI->setIfReturnsVoid(!Val);
299 
300   assert(!Val == VRegs.empty() && "Return value without a vreg");
301 
302   CallingConv::ID CC = B.getMF().getFunction().getCallingConv();
303   const bool IsShader = AMDGPU::isShader(CC);
304   const bool IsWaveEnd = (IsShader && MFI->returnsVoid()) ||
305                          AMDGPU::isKernel(CC);
306   if (IsWaveEnd) {
307     B.buildInstr(AMDGPU::S_ENDPGM)
308       .addImm(0);
309     return true;
310   }
311 
312   auto const &ST = B.getMF().getSubtarget<GCNSubtarget>();
313 
314   unsigned ReturnOpc =
315       IsShader ? AMDGPU::SI_RETURN_TO_EPILOG : AMDGPU::S_SETPC_B64_return;
316 
317   auto Ret = B.buildInstrNoInsert(ReturnOpc);
318   Register ReturnAddrVReg;
319   if (ReturnOpc == AMDGPU::S_SETPC_B64_return) {
320     ReturnAddrVReg = MRI.createVirtualRegister(&AMDGPU::CCR_SGPR_64RegClass);
321     Ret.addUse(ReturnAddrVReg);
322   }
323 
324   if (!lowerReturnVal(B, Val, VRegs, Ret))
325     return false;
326 
327   if (ReturnOpc == AMDGPU::S_SETPC_B64_return) {
328     const SIRegisterInfo *TRI = ST.getRegisterInfo();
329     Register LiveInReturn = MF.addLiveIn(TRI->getReturnAddressReg(MF),
330                                          &AMDGPU::SGPR_64RegClass);
331     B.buildCopy(ReturnAddrVReg, LiveInReturn);
332   }
333 
334   // TODO: Handle CalleeSavedRegsViaCopy.
335 
336   B.insertInstr(Ret);
337   return true;
338 }
339 
340 Register AMDGPUCallLowering::lowerParameterPtr(MachineIRBuilder &B,
341                                                Type *ParamTy,
342                                                uint64_t Offset) const {
343 
344   MachineFunction &MF = B.getMF();
345   const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
346   MachineRegisterInfo &MRI = MF.getRegInfo();
347   const Function &F = MF.getFunction();
348   const DataLayout &DL = F.getParent()->getDataLayout();
349   PointerType *PtrTy = PointerType::get(ParamTy, AMDGPUAS::CONSTANT_ADDRESS);
350   LLT PtrType = getLLTForType(*PtrTy, DL);
351   Register DstReg = MRI.createGenericVirtualRegister(PtrType);
352   Register KernArgSegmentPtr =
353     MFI->getPreloadedReg(AMDGPUFunctionArgInfo::KERNARG_SEGMENT_PTR);
354   Register KernArgSegmentVReg = MRI.getLiveInVirtReg(KernArgSegmentPtr);
355 
356   Register OffsetReg = MRI.createGenericVirtualRegister(LLT::scalar(64));
357   B.buildConstant(OffsetReg, Offset);
358 
359   B.buildPtrAdd(DstReg, KernArgSegmentVReg, OffsetReg);
360 
361   return DstReg;
362 }
363 
364 void AMDGPUCallLowering::lowerParameter(MachineIRBuilder &B,
365                                         Type *ParamTy, uint64_t Offset,
366                                         unsigned Align,
367                                         Register DstReg) const {
368   MachineFunction &MF = B.getMF();
369   const Function &F = MF.getFunction();
370   const DataLayout &DL = F.getParent()->getDataLayout();
371   MachinePointerInfo PtrInfo(AMDGPUAS::CONSTANT_ADDRESS);
372   unsigned TypeSize = DL.getTypeStoreSize(ParamTy);
373   Register PtrReg = lowerParameterPtr(B, ParamTy, Offset);
374 
375   MachineMemOperand *MMO =
376       MF.getMachineMemOperand(PtrInfo, MachineMemOperand::MOLoad |
377                                        MachineMemOperand::MODereferenceable |
378                                        MachineMemOperand::MOInvariant,
379                                        TypeSize, Align);
380 
381   B.buildLoad(DstReg, PtrReg, *MMO);
382 }
383 
384 // Allocate special inputs passed in user SGPRs.
385 static void allocateHSAUserSGPRs(CCState &CCInfo,
386                                  MachineIRBuilder &B,
387                                  MachineFunction &MF,
388                                  const SIRegisterInfo &TRI,
389                                  SIMachineFunctionInfo &Info) {
390   // FIXME: How should these inputs interact with inreg / custom SGPR inputs?
391   if (Info.hasPrivateSegmentBuffer()) {
392     unsigned PrivateSegmentBufferReg = Info.addPrivateSegmentBuffer(TRI);
393     MF.addLiveIn(PrivateSegmentBufferReg, &AMDGPU::SGPR_128RegClass);
394     CCInfo.AllocateReg(PrivateSegmentBufferReg);
395   }
396 
397   if (Info.hasDispatchPtr()) {
398     unsigned DispatchPtrReg = Info.addDispatchPtr(TRI);
399     MF.addLiveIn(DispatchPtrReg, &AMDGPU::SGPR_64RegClass);
400     CCInfo.AllocateReg(DispatchPtrReg);
401   }
402 
403   if (Info.hasQueuePtr()) {
404     unsigned QueuePtrReg = Info.addQueuePtr(TRI);
405     MF.addLiveIn(QueuePtrReg, &AMDGPU::SGPR_64RegClass);
406     CCInfo.AllocateReg(QueuePtrReg);
407   }
408 
409   if (Info.hasKernargSegmentPtr()) {
410     MachineRegisterInfo &MRI = MF.getRegInfo();
411     Register InputPtrReg = Info.addKernargSegmentPtr(TRI);
412     const LLT P4 = LLT::pointer(AMDGPUAS::CONSTANT_ADDRESS, 64);
413     Register VReg = MRI.createGenericVirtualRegister(P4);
414     MRI.addLiveIn(InputPtrReg, VReg);
415     B.getMBB().addLiveIn(InputPtrReg);
416     B.buildCopy(VReg, InputPtrReg);
417     CCInfo.AllocateReg(InputPtrReg);
418   }
419 
420   if (Info.hasDispatchID()) {
421     unsigned DispatchIDReg = Info.addDispatchID(TRI);
422     MF.addLiveIn(DispatchIDReg, &AMDGPU::SGPR_64RegClass);
423     CCInfo.AllocateReg(DispatchIDReg);
424   }
425 
426   if (Info.hasFlatScratchInit()) {
427     unsigned FlatScratchInitReg = Info.addFlatScratchInit(TRI);
428     MF.addLiveIn(FlatScratchInitReg, &AMDGPU::SGPR_64RegClass);
429     CCInfo.AllocateReg(FlatScratchInitReg);
430   }
431 
432   // TODO: Add GridWorkGroupCount user SGPRs when used. For now with HSA we read
433   // these from the dispatch pointer.
434 }
435 
436 bool AMDGPUCallLowering::lowerFormalArgumentsKernel(
437     MachineIRBuilder &B, const Function &F,
438     ArrayRef<ArrayRef<Register>> VRegs) const {
439   MachineFunction &MF = B.getMF();
440   const GCNSubtarget *Subtarget = &MF.getSubtarget<GCNSubtarget>();
441   MachineRegisterInfo &MRI = MF.getRegInfo();
442   SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
443   const SIRegisterInfo *TRI = Subtarget->getRegisterInfo();
444   const SITargetLowering &TLI = *getTLI<SITargetLowering>();
445 
446   const DataLayout &DL = F.getParent()->getDataLayout();
447 
448   SmallVector<CCValAssign, 16> ArgLocs;
449   CCState CCInfo(F.getCallingConv(), F.isVarArg(), MF, ArgLocs, F.getContext());
450 
451   allocateHSAUserSGPRs(CCInfo, B, MF, *TRI, *Info);
452 
453   unsigned i = 0;
454   const unsigned KernArgBaseAlign = 16;
455   const unsigned BaseOffset = Subtarget->getExplicitKernelArgOffset(F);
456   uint64_t ExplicitArgOffset = 0;
457 
458   // TODO: Align down to dword alignment and extract bits for extending loads.
459   for (auto &Arg : F.args()) {
460     Type *ArgTy = Arg.getType();
461     unsigned AllocSize = DL.getTypeAllocSize(ArgTy);
462     if (AllocSize == 0)
463       continue;
464 
465     unsigned ABIAlign = DL.getABITypeAlignment(ArgTy);
466 
467     uint64_t ArgOffset = alignTo(ExplicitArgOffset, ABIAlign) + BaseOffset;
468     ExplicitArgOffset = alignTo(ExplicitArgOffset, ABIAlign) + AllocSize;
469 
470     ArrayRef<Register> OrigArgRegs = VRegs[i];
471     Register ArgReg =
472       OrigArgRegs.size() == 1
473       ? OrigArgRegs[0]
474       : MRI.createGenericVirtualRegister(getLLTForType(*ArgTy, DL));
475     unsigned Align = MinAlign(KernArgBaseAlign, ArgOffset);
476     ArgOffset = alignTo(ArgOffset, DL.getABITypeAlignment(ArgTy));
477     lowerParameter(B, ArgTy, ArgOffset, Align, ArgReg);
478     if (OrigArgRegs.size() > 1)
479       unpackRegs(OrigArgRegs, ArgReg, ArgTy, B);
480     ++i;
481   }
482 
483   TLI.allocateSpecialEntryInputVGPRs(CCInfo, MF, *TRI, *Info);
484   TLI.allocateSystemSGPRs(CCInfo, MF, *Info, F.getCallingConv(), false);
485   return true;
486 }
487 
488 // TODO: Move this to generic code
489 static void packSplitRegsToOrigType(MachineIRBuilder &B,
490                                     ArrayRef<Register> OrigRegs,
491                                     ArrayRef<Register> Regs,
492                                     LLT LLTy,
493                                     LLT PartLLT) {
494   if (!LLTy.isVector() && !PartLLT.isVector()) {
495     B.buildMerge(OrigRegs[0], Regs);
496     return;
497   }
498 
499   if (LLTy.isVector() && PartLLT.isVector()) {
500     assert(LLTy.getElementType() == PartLLT.getElementType());
501 
502     int DstElts = LLTy.getNumElements();
503     int PartElts = PartLLT.getNumElements();
504     if (DstElts % PartElts == 0)
505       B.buildConcatVectors(OrigRegs[0], Regs);
506     else {
507       // Deal with v3s16 split into v2s16
508       assert(PartElts == 2 && DstElts % 2 != 0);
509       int RoundedElts = PartElts * ((DstElts + PartElts - 1) / PartElts);
510 
511       LLT RoundedDestTy = LLT::vector(RoundedElts, PartLLT.getElementType());
512       auto RoundedConcat = B.buildConcatVectors(RoundedDestTy, Regs);
513       B.buildExtract(OrigRegs[0], RoundedConcat, 0);
514     }
515 
516     return;
517   }
518 
519   MachineRegisterInfo &MRI = *B.getMRI();
520 
521   assert(LLTy.isVector() && !PartLLT.isVector());
522 
523   LLT DstEltTy = LLTy.getElementType();
524 
525   // Pointer information was discarded. We'll need to coerce some register types
526   // to avoid violating type constraints.
527   LLT RealDstEltTy = MRI.getType(OrigRegs[0]).getElementType();
528 
529   assert(DstEltTy.getSizeInBits() == RealDstEltTy.getSizeInBits());
530 
531   if (DstEltTy == PartLLT) {
532     // Vector was trivially scalarized.
533 
534     if (RealDstEltTy.isPointer()) {
535       for (Register Reg : Regs)
536         MRI.setType(Reg, RealDstEltTy);
537     }
538 
539     B.buildBuildVector(OrigRegs[0], Regs);
540   } else if (DstEltTy.getSizeInBits() > PartLLT.getSizeInBits()) {
541     // Deal with vector with 64-bit elements decomposed to 32-bit
542     // registers. Need to create intermediate 64-bit elements.
543     SmallVector<Register, 8> EltMerges;
544     int PartsPerElt = DstEltTy.getSizeInBits() / PartLLT.getSizeInBits();
545 
546     assert(DstEltTy.getSizeInBits() % PartLLT.getSizeInBits() == 0);
547 
548     for (int I = 0, NumElts = LLTy.getNumElements(); I != NumElts; ++I)  {
549       auto Merge = B.buildMerge(RealDstEltTy, Regs.take_front(PartsPerElt));
550       // Fix the type in case this is really a vector of pointers.
551       MRI.setType(Merge.getReg(0), RealDstEltTy);
552       EltMerges.push_back(Merge.getReg(0));
553       Regs = Regs.drop_front(PartsPerElt);
554     }
555 
556     B.buildBuildVector(OrigRegs[0], EltMerges);
557   } else {
558     // Vector was split, and elements promoted to a wider type.
559     LLT BVType = LLT::vector(LLTy.getNumElements(), PartLLT);
560     auto BV = B.buildBuildVector(BVType, Regs);
561     B.buildTrunc(OrigRegs[0], BV);
562   }
563 }
564 
565 bool AMDGPUCallLowering::lowerFormalArguments(
566     MachineIRBuilder &B, const Function &F,
567     ArrayRef<ArrayRef<Register>> VRegs) const {
568   CallingConv::ID CC = F.getCallingConv();
569 
570   // The infrastructure for normal calling convention lowering is essentially
571   // useless for kernels. We want to avoid any kind of legalization or argument
572   // splitting.
573   if (CC == CallingConv::AMDGPU_KERNEL)
574     return lowerFormalArgumentsKernel(B, F, VRegs);
575 
576   const bool IsShader = AMDGPU::isShader(CC);
577   const bool IsEntryFunc = AMDGPU::isEntryFunctionCC(CC);
578 
579   MachineFunction &MF = B.getMF();
580   MachineBasicBlock &MBB = B.getMBB();
581   MachineRegisterInfo &MRI = MF.getRegInfo();
582   SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
583   const GCNSubtarget &Subtarget = MF.getSubtarget<GCNSubtarget>();
584   const SIRegisterInfo *TRI = Subtarget.getRegisterInfo();
585   const DataLayout &DL = F.getParent()->getDataLayout();
586 
587 
588   SmallVector<CCValAssign, 16> ArgLocs;
589   CCState CCInfo(CC, F.isVarArg(), MF, ArgLocs, F.getContext());
590 
591   if (!IsEntryFunc) {
592     Register ReturnAddrReg = TRI->getReturnAddressReg(MF);
593     Register LiveInReturn = MF.addLiveIn(ReturnAddrReg,
594                                          &AMDGPU::SGPR_64RegClass);
595     MBB.addLiveIn(ReturnAddrReg);
596     B.buildCopy(LiveInReturn, ReturnAddrReg);
597   }
598 
599   if (Info->hasImplicitBufferPtr()) {
600     Register ImplicitBufferPtrReg = Info->addImplicitBufferPtr(*TRI);
601     MF.addLiveIn(ImplicitBufferPtrReg, &AMDGPU::SGPR_64RegClass);
602     CCInfo.AllocateReg(ImplicitBufferPtrReg);
603   }
604 
605 
606   SmallVector<ArgInfo, 32> SplitArgs;
607   unsigned Idx = 0;
608   unsigned PSInputNum = 0;
609 
610   for (auto &Arg : F.args()) {
611     if (DL.getTypeStoreSize(Arg.getType()) == 0)
612       continue;
613 
614     const bool InReg = Arg.hasAttribute(Attribute::InReg);
615 
616     // SGPR arguments to functions not implemented.
617     if (!IsShader && InReg)
618       return false;
619 
620     if (Arg.hasAttribute(Attribute::SwiftSelf) ||
621         Arg.hasAttribute(Attribute::SwiftError) ||
622         Arg.hasAttribute(Attribute::Nest))
623       return false;
624 
625     if (CC == CallingConv::AMDGPU_PS && !InReg && PSInputNum <= 15) {
626       const bool ArgUsed = !Arg.use_empty();
627       bool SkipArg = !ArgUsed && !Info->isPSInputAllocated(PSInputNum);
628 
629       if (!SkipArg) {
630         Info->markPSInputAllocated(PSInputNum);
631         if (ArgUsed)
632           Info->markPSInputEnabled(PSInputNum);
633       }
634 
635       ++PSInputNum;
636 
637       if (SkipArg) {
638         for (int I = 0, E = VRegs[Idx].size(); I != E; ++I)
639           B.buildUndef(VRegs[Idx][I]);
640 
641         ++Idx;
642         continue;
643       }
644     }
645 
646     ArgInfo OrigArg(VRegs[Idx], Arg.getType());
647     setArgFlags(OrigArg, Idx + AttributeList::FirstArgIndex, DL, F);
648 
649     splitToValueTypes(
650       OrigArg, SplitArgs, DL, MRI, CC,
651       // FIXME: We should probably be passing multiple registers to
652       // handleAssignments to do this
653       [&](ArrayRef<Register> Regs, LLT LLTy, LLT PartLLT, int VTSplitIdx) {
654         packSplitRegsToOrigType(B, VRegs[Idx][VTSplitIdx], Regs,
655                                 LLTy, PartLLT);
656       });
657 
658     ++Idx;
659   }
660 
661   // At least one interpolation mode must be enabled or else the GPU will
662   // hang.
663   //
664   // Check PSInputAddr instead of PSInputEnable. The idea is that if the user
665   // set PSInputAddr, the user wants to enable some bits after the compilation
666   // based on run-time states. Since we can't know what the final PSInputEna
667   // will look like, so we shouldn't do anything here and the user should take
668   // responsibility for the correct programming.
669   //
670   // Otherwise, the following restrictions apply:
671   // - At least one of PERSP_* (0xF) or LINEAR_* (0x70) must be enabled.
672   // - If POS_W_FLOAT (11) is enabled, at least one of PERSP_* must be
673   //   enabled too.
674   if (CC == CallingConv::AMDGPU_PS) {
675     if ((Info->getPSInputAddr() & 0x7F) == 0 ||
676         ((Info->getPSInputAddr() & 0xF) == 0 &&
677          Info->isPSInputAllocated(11))) {
678       CCInfo.AllocateReg(AMDGPU::VGPR0);
679       CCInfo.AllocateReg(AMDGPU::VGPR1);
680       Info->markPSInputAllocated(0);
681       Info->markPSInputEnabled(0);
682     }
683 
684     if (Subtarget.isAmdPalOS()) {
685       // For isAmdPalOS, the user does not enable some bits after compilation
686       // based on run-time states; the register values being generated here are
687       // the final ones set in hardware. Therefore we need to apply the
688       // workaround to PSInputAddr and PSInputEnable together.  (The case where
689       // a bit is set in PSInputAddr but not PSInputEnable is where the frontend
690       // set up an input arg for a particular interpolation mode, but nothing
691       // uses that input arg. Really we should have an earlier pass that removes
692       // such an arg.)
693       unsigned PsInputBits = Info->getPSInputAddr() & Info->getPSInputEnable();
694       if ((PsInputBits & 0x7F) == 0 ||
695           ((PsInputBits & 0xF) == 0 &&
696            (PsInputBits >> 11 & 1)))
697         Info->markPSInputEnabled(
698           countTrailingZeros(Info->getPSInputAddr(), ZB_Undefined));
699     }
700   }
701 
702   const SITargetLowering &TLI = *getTLI<SITargetLowering>();
703   CCAssignFn *AssignFn = TLI.CCAssignFnForCall(CC, F.isVarArg());
704 
705   if (!MBB.empty())
706     B.setInstr(*MBB.begin());
707 
708   FormalArgHandler Handler(B, MRI, AssignFn);
709   if (!handleAssignments(CCInfo, ArgLocs, B, SplitArgs, Handler))
710     return false;
711 
712   if (!IsEntryFunc) {
713     // Special inputs come after user arguments.
714     TLI.allocateSpecialInputVGPRs(CCInfo, MF, *TRI, *Info);
715   }
716 
717   // Start adding system SGPRs.
718   if (IsEntryFunc) {
719     TLI.allocateSystemSGPRs(CCInfo, MF, *Info, CC, IsShader);
720   } else {
721     CCInfo.AllocateReg(Info->getScratchRSrcReg());
722     CCInfo.AllocateReg(Info->getScratchWaveOffsetReg());
723     CCInfo.AllocateReg(Info->getFrameOffsetReg());
724     TLI.allocateSpecialInputSGPRs(CCInfo, MF, *TRI, *Info);
725   }
726 
727   // Move back to the end of the basic block.
728   B.setMBB(MBB);
729 
730   return true;
731 }
732