1 //===- SIMachineFunctionInfo.cpp - SI Machine Function Info ---------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 10 #include "SIMachineFunctionInfo.h" 11 #include "AMDGPUArgumentUsageInfo.h" 12 #include "AMDGPUSubtarget.h" 13 #include "SIRegisterInfo.h" 14 #include "Utils/AMDGPUBaseInfo.h" 15 #include "llvm/ADT/Optional.h" 16 #include "llvm/CodeGen/MachineBasicBlock.h" 17 #include "llvm/CodeGen/MachineFrameInfo.h" 18 #include "llvm/CodeGen/MachineFunction.h" 19 #include "llvm/CodeGen/MachineRegisterInfo.h" 20 #include "llvm/IR/CallingConv.h" 21 #include "llvm/IR/Function.h" 22 #include <cassert> 23 #include <vector> 24 25 #define MAX_LANES 64 26 27 using namespace llvm; 28 29 SIMachineFunctionInfo::SIMachineFunctionInfo(const MachineFunction &MF) 30 : AMDGPUMachineFunction(MF), 31 BufferPSV(*(MF.getSubtarget().getInstrInfo())), 32 ImagePSV(*(MF.getSubtarget().getInstrInfo())), 33 PrivateSegmentBuffer(false), 34 DispatchPtr(false), 35 QueuePtr(false), 36 KernargSegmentPtr(false), 37 DispatchID(false), 38 FlatScratchInit(false), 39 GridWorkgroupCountX(false), 40 GridWorkgroupCountY(false), 41 GridWorkgroupCountZ(false), 42 WorkGroupIDX(false), 43 WorkGroupIDY(false), 44 WorkGroupIDZ(false), 45 WorkGroupInfo(false), 46 PrivateSegmentWaveByteOffset(false), 47 WorkItemIDX(false), 48 WorkItemIDY(false), 49 WorkItemIDZ(false), 50 ImplicitBufferPtr(false), 51 ImplicitArgPtr(false) { 52 const SISubtarget &ST = MF.getSubtarget<SISubtarget>(); 53 const Function *F = MF.getFunction(); 54 FlatWorkGroupSizes = ST.getFlatWorkGroupSizes(*F); 55 WavesPerEU = ST.getWavesPerEU(*F); 56 57 if (!isEntryFunction()) { 58 // Non-entry functions have no special inputs for now, other registers 59 // required for scratch access. 60 ScratchRSrcReg = AMDGPU::SGPR0_SGPR1_SGPR2_SGPR3; 61 ScratchWaveOffsetReg = AMDGPU::SGPR4; 62 FrameOffsetReg = AMDGPU::SGPR5; 63 StackPtrOffsetReg = AMDGPU::SGPR32; 64 65 ArgInfo.PrivateSegmentBuffer = 66 ArgDescriptor::createRegister(ScratchRSrcReg); 67 ArgInfo.PrivateSegmentWaveByteOffset = 68 ArgDescriptor::createRegister(ScratchWaveOffsetReg); 69 70 if (F->hasFnAttribute("amdgpu-implicitarg-ptr")) 71 ImplicitArgPtr = true; 72 } else { 73 if (F->hasFnAttribute("amdgpu-implicitarg-ptr")) 74 KernargSegmentPtr = true; 75 } 76 77 CallingConv::ID CC = F->getCallingConv(); 78 if (CC == CallingConv::AMDGPU_KERNEL || CC == CallingConv::SPIR_KERNEL) { 79 if (!F->arg_empty()) 80 KernargSegmentPtr = true; 81 WorkGroupIDX = true; 82 WorkItemIDX = true; 83 } else if (CC == CallingConv::AMDGPU_PS) { 84 PSInputAddr = AMDGPU::getInitialPSInputAddr(*F); 85 } 86 87 if (ST.debuggerEmitPrologue()) { 88 // Enable everything. 89 WorkGroupIDX = true; 90 WorkGroupIDY = true; 91 WorkGroupIDZ = true; 92 WorkItemIDX = true; 93 WorkItemIDY = true; 94 WorkItemIDZ = true; 95 } else { 96 if (F->hasFnAttribute("amdgpu-work-group-id-x")) 97 WorkGroupIDX = true; 98 99 if (F->hasFnAttribute("amdgpu-work-group-id-y")) 100 WorkGroupIDY = true; 101 102 if (F->hasFnAttribute("amdgpu-work-group-id-z")) 103 WorkGroupIDZ = true; 104 105 if (F->hasFnAttribute("amdgpu-work-item-id-x")) 106 WorkItemIDX = true; 107 108 if (F->hasFnAttribute("amdgpu-work-item-id-y")) 109 WorkItemIDY = true; 110 111 if (F->hasFnAttribute("amdgpu-work-item-id-z")) 112 WorkItemIDZ = true; 113 } 114 115 const MachineFrameInfo &FrameInfo = MF.getFrameInfo(); 116 bool MaySpill = ST.isVGPRSpillingEnabled(*F); 117 bool HasStackObjects = FrameInfo.hasStackObjects(); 118 119 if (isEntryFunction()) { 120 // X, XY, and XYZ are the only supported combinations, so make sure Y is 121 // enabled if Z is. 122 if (WorkItemIDZ) 123 WorkItemIDY = true; 124 125 if (HasStackObjects || MaySpill) { 126 PrivateSegmentWaveByteOffset = true; 127 128 // HS and GS always have the scratch wave offset in SGPR5 on GFX9. 129 if (ST.getGeneration() >= AMDGPUSubtarget::GFX9 && 130 (CC == CallingConv::AMDGPU_HS || CC == CallingConv::AMDGPU_GS)) 131 ArgInfo.PrivateSegmentWaveByteOffset 132 = ArgDescriptor::createRegister(AMDGPU::SGPR5); 133 } 134 } 135 136 bool IsCOV2 = ST.isAmdCodeObjectV2(MF); 137 if (IsCOV2) { 138 if (HasStackObjects || MaySpill) 139 PrivateSegmentBuffer = true; 140 141 if (F->hasFnAttribute("amdgpu-dispatch-ptr")) 142 DispatchPtr = true; 143 144 if (F->hasFnAttribute("amdgpu-queue-ptr")) 145 QueuePtr = true; 146 147 if (F->hasFnAttribute("amdgpu-dispatch-id")) 148 DispatchID = true; 149 } else if (ST.isMesaGfxShader(MF)) { 150 if (HasStackObjects || MaySpill) 151 ImplicitBufferPtr = true; 152 } 153 154 if (F->hasFnAttribute("amdgpu-kernarg-segment-ptr")) 155 KernargSegmentPtr = true; 156 157 if (ST.hasFlatAddressSpace() && isEntryFunction() && IsCOV2) { 158 // TODO: This could be refined a lot. The attribute is a poor way of 159 // detecting calls that may require it before argument lowering. 160 if (HasStackObjects || F->hasFnAttribute("amdgpu-flat-scratch")) 161 FlatScratchInit = true; 162 } 163 } 164 165 unsigned SIMachineFunctionInfo::addPrivateSegmentBuffer( 166 const SIRegisterInfo &TRI) { 167 ArgInfo.PrivateSegmentBuffer = 168 ArgDescriptor::createRegister(TRI.getMatchingSuperReg( 169 getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SReg_128RegClass)); 170 NumUserSGPRs += 4; 171 return ArgInfo.PrivateSegmentBuffer.getRegister(); 172 } 173 174 unsigned SIMachineFunctionInfo::addDispatchPtr(const SIRegisterInfo &TRI) { 175 ArgInfo.DispatchPtr = ArgDescriptor::createRegister(TRI.getMatchingSuperReg( 176 getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SReg_64RegClass)); 177 NumUserSGPRs += 2; 178 return ArgInfo.DispatchPtr.getRegister(); 179 } 180 181 unsigned SIMachineFunctionInfo::addQueuePtr(const SIRegisterInfo &TRI) { 182 ArgInfo.QueuePtr = ArgDescriptor::createRegister(TRI.getMatchingSuperReg( 183 getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SReg_64RegClass)); 184 NumUserSGPRs += 2; 185 return ArgInfo.QueuePtr.getRegister(); 186 } 187 188 unsigned SIMachineFunctionInfo::addKernargSegmentPtr(const SIRegisterInfo &TRI) { 189 ArgInfo.KernargSegmentPtr 190 = ArgDescriptor::createRegister(TRI.getMatchingSuperReg( 191 getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SReg_64RegClass)); 192 NumUserSGPRs += 2; 193 return ArgInfo.KernargSegmentPtr.getRegister(); 194 } 195 196 unsigned SIMachineFunctionInfo::addDispatchID(const SIRegisterInfo &TRI) { 197 ArgInfo.DispatchID = ArgDescriptor::createRegister(TRI.getMatchingSuperReg( 198 getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SReg_64RegClass)); 199 NumUserSGPRs += 2; 200 return ArgInfo.DispatchID.getRegister(); 201 } 202 203 unsigned SIMachineFunctionInfo::addFlatScratchInit(const SIRegisterInfo &TRI) { 204 ArgInfo.FlatScratchInit = ArgDescriptor::createRegister(TRI.getMatchingSuperReg( 205 getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SReg_64RegClass)); 206 NumUserSGPRs += 2; 207 return ArgInfo.FlatScratchInit.getRegister(); 208 } 209 210 unsigned SIMachineFunctionInfo::addImplicitBufferPtr(const SIRegisterInfo &TRI) { 211 ArgInfo.ImplicitBufferPtr = ArgDescriptor::createRegister(TRI.getMatchingSuperReg( 212 getNextUserSGPR(), AMDGPU::sub0, &AMDGPU::SReg_64RegClass)); 213 NumUserSGPRs += 2; 214 return ArgInfo.ImplicitBufferPtr.getRegister(); 215 } 216 217 static bool isCalleeSavedReg(const MCPhysReg *CSRegs, MCPhysReg Reg) { 218 for (unsigned I = 0; CSRegs[I]; ++I) { 219 if (CSRegs[I] == Reg) 220 return true; 221 } 222 223 return false; 224 } 225 226 /// Reserve a slice of a VGPR to support spilling for FrameIndex \p FI. 227 bool SIMachineFunctionInfo::allocateSGPRSpillToVGPR(MachineFunction &MF, 228 int FI) { 229 std::vector<SpilledReg> &SpillLanes = SGPRToVGPRSpills[FI]; 230 231 // This has already been allocated. 232 if (!SpillLanes.empty()) 233 return true; 234 235 const SISubtarget &ST = MF.getSubtarget<SISubtarget>(); 236 const SIRegisterInfo *TRI = ST.getRegisterInfo(); 237 MachineFrameInfo &FrameInfo = MF.getFrameInfo(); 238 MachineRegisterInfo &MRI = MF.getRegInfo(); 239 unsigned WaveSize = ST.getWavefrontSize(); 240 241 unsigned Size = FrameInfo.getObjectSize(FI); 242 assert(Size >= 4 && Size <= 64 && "invalid sgpr spill size"); 243 assert(TRI->spillSGPRToVGPR() && "not spilling SGPRs to VGPRs"); 244 245 int NumLanes = Size / 4; 246 247 const MCPhysReg *CSRegs = TRI->getCalleeSavedRegs(&MF); 248 249 // Make sure to handle the case where a wide SGPR spill may span between two 250 // VGPRs. 251 for (int I = 0; I < NumLanes; ++I, ++NumVGPRSpillLanes) { 252 unsigned LaneVGPR; 253 unsigned VGPRIndex = (NumVGPRSpillLanes % WaveSize); 254 255 if (VGPRIndex == 0) { 256 LaneVGPR = TRI->findUnusedRegister(MRI, &AMDGPU::VGPR_32RegClass, MF); 257 if (LaneVGPR == AMDGPU::NoRegister) { 258 // We have no VGPRs left for spilling SGPRs. Reset because we will not 259 // partially spill the SGPR to VGPRs. 260 SGPRToVGPRSpills.erase(FI); 261 NumVGPRSpillLanes -= I; 262 return false; 263 } 264 265 Optional<int> CSRSpillFI; 266 if (FrameInfo.hasCalls() && CSRegs && isCalleeSavedReg(CSRegs, LaneVGPR)) { 267 // TODO: Should this be a CreateSpillStackObject? This is technically a 268 // weird CSR spill. 269 CSRSpillFI = FrameInfo.CreateStackObject(4, 4, false); 270 } 271 272 SpillVGPRs.push_back(SGPRSpillVGPRCSR(LaneVGPR, CSRSpillFI)); 273 274 // Add this register as live-in to all blocks to avoid machine verifer 275 // complaining about use of an undefined physical register. 276 for (MachineBasicBlock &BB : MF) 277 BB.addLiveIn(LaneVGPR); 278 } else { 279 LaneVGPR = SpillVGPRs.back().VGPR; 280 } 281 282 SpillLanes.push_back(SpilledReg(LaneVGPR, VGPRIndex)); 283 } 284 285 return true; 286 } 287 288 void SIMachineFunctionInfo::removeSGPRToVGPRFrameIndices(MachineFrameInfo &MFI) { 289 for (auto &R : SGPRToVGPRSpills) 290 MFI.RemoveStackObject(R.first); 291 } 292