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