1 //===- SIPreAllocateWWMRegs.cpp - WWM Register Pre-allocation -------------===// 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 /// Pass to pre-allocated WWM registers 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "AMDGPU.h" 15 #include "GCNSubtarget.h" 16 #include "MCTargetDesc/AMDGPUMCTargetDesc.h" 17 #include "SIMachineFunctionInfo.h" 18 #include "llvm/ADT/PostOrderIterator.h" 19 #include "llvm/CodeGen/LiveIntervals.h" 20 #include "llvm/CodeGen/LiveRegMatrix.h" 21 #include "llvm/CodeGen/MachineFunctionPass.h" 22 #include "llvm/InitializePasses.h" 23 24 using namespace llvm; 25 26 #define DEBUG_TYPE "si-pre-allocate-wwm-regs" 27 28 namespace { 29 30 class SIPreAllocateWWMRegs : public MachineFunctionPass { 31 private: 32 const SIInstrInfo *TII; 33 const SIRegisterInfo *TRI; 34 MachineRegisterInfo *MRI; 35 LiveIntervals *LIS; 36 LiveRegMatrix *Matrix; 37 VirtRegMap *VRM; 38 RegisterClassInfo RegClassInfo; 39 40 std::vector<unsigned> RegsToRewrite; 41 #ifndef NDEBUG 42 void printWWMInfo(const MachineInstr &MI); 43 #endif 44 45 public: 46 static char ID; 47 48 SIPreAllocateWWMRegs() : MachineFunctionPass(ID) { 49 initializeSIPreAllocateWWMRegsPass(*PassRegistry::getPassRegistry()); 50 } 51 52 bool runOnMachineFunction(MachineFunction &MF) override; 53 54 void getAnalysisUsage(AnalysisUsage &AU) const override { 55 AU.addRequired<LiveIntervals>(); 56 AU.addPreserved<LiveIntervals>(); 57 AU.addRequired<VirtRegMap>(); 58 AU.addRequired<LiveRegMatrix>(); 59 AU.addPreserved<SlotIndexes>(); 60 AU.setPreservesCFG(); 61 MachineFunctionPass::getAnalysisUsage(AU); 62 } 63 64 private: 65 bool processDef(MachineOperand &MO); 66 void rewriteRegs(MachineFunction &MF); 67 }; 68 69 } // End anonymous namespace. 70 71 INITIALIZE_PASS_BEGIN(SIPreAllocateWWMRegs, DEBUG_TYPE, 72 "SI Pre-allocate WWM Registers", false, false) 73 INITIALIZE_PASS_DEPENDENCY(LiveIntervals) 74 INITIALIZE_PASS_DEPENDENCY(VirtRegMap) 75 INITIALIZE_PASS_DEPENDENCY(LiveRegMatrix) 76 INITIALIZE_PASS_END(SIPreAllocateWWMRegs, DEBUG_TYPE, 77 "SI Pre-allocate WWM Registers", false, false) 78 79 char SIPreAllocateWWMRegs::ID = 0; 80 81 char &llvm::SIPreAllocateWWMRegsID = SIPreAllocateWWMRegs::ID; 82 83 FunctionPass *llvm::createSIPreAllocateWWMRegsPass() { 84 return new SIPreAllocateWWMRegs(); 85 } 86 87 bool SIPreAllocateWWMRegs::processDef(MachineOperand &MO) { 88 if (!MO.isReg()) 89 return false; 90 91 Register Reg = MO.getReg(); 92 if (Reg.isPhysical()) 93 return false; 94 95 if (!TRI->isVGPR(*MRI, Reg)) 96 return false; 97 98 if (VRM->hasPhys(Reg)) 99 return false; 100 101 LiveInterval &LI = LIS->getInterval(Reg); 102 103 for (MCRegister PhysReg : RegClassInfo.getOrder(MRI->getRegClass(Reg))) { 104 if (!MRI->isPhysRegUsed(PhysReg) && 105 Matrix->checkInterference(LI, PhysReg) == LiveRegMatrix::IK_Free) { 106 Matrix->assign(LI, PhysReg); 107 assert(PhysReg != 0); 108 RegsToRewrite.push_back(Reg); 109 return true; 110 } 111 } 112 113 llvm_unreachable("physreg not found for WWM expression"); 114 return false; 115 } 116 117 void SIPreAllocateWWMRegs::rewriteRegs(MachineFunction &MF) { 118 for (MachineBasicBlock &MBB : MF) { 119 for (MachineInstr &MI : MBB) { 120 for (MachineOperand &MO : MI.operands()) { 121 if (!MO.isReg()) 122 continue; 123 124 const Register VirtReg = MO.getReg(); 125 if (VirtReg.isPhysical()) 126 continue; 127 128 if (!VRM->hasPhys(VirtReg)) 129 continue; 130 131 Register PhysReg = VRM->getPhys(VirtReg); 132 const unsigned SubReg = MO.getSubReg(); 133 if (SubReg != 0) { 134 PhysReg = TRI->getSubReg(PhysReg, SubReg); 135 MO.setSubReg(0); 136 } 137 138 MO.setReg(PhysReg); 139 MO.setIsRenamable(false); 140 } 141 } 142 } 143 144 SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); 145 146 for (unsigned Reg : RegsToRewrite) { 147 LIS->removeInterval(Reg); 148 149 const Register PhysReg = VRM->getPhys(Reg); 150 assert(PhysReg != 0); 151 MFI->ReserveWWMRegister(PhysReg); 152 } 153 154 RegsToRewrite.clear(); 155 156 // Update the set of reserved registers to include WWM ones. 157 MRI->freezeReservedRegs(MF); 158 } 159 160 #ifndef NDEBUG 161 LLVM_DUMP_METHOD void 162 SIPreAllocateWWMRegs::printWWMInfo(const MachineInstr &MI) { 163 164 unsigned Opc = MI.getOpcode(); 165 166 if (Opc == AMDGPU::ENTER_STRICT_WWM || Opc == AMDGPU::ENTER_STRICT_WQM) { 167 dbgs() << "Entering "; 168 } else { 169 assert(Opc == AMDGPU::EXIT_STRICT_WWM || Opc == AMDGPU::EXIT_STRICT_WQM); 170 dbgs() << "Exiting "; 171 } 172 173 if (Opc == AMDGPU::ENTER_STRICT_WWM || Opc == AMDGPU::EXIT_STRICT_WWM) { 174 dbgs() << "Strict WWM "; 175 } else { 176 assert(Opc == AMDGPU::ENTER_STRICT_WQM || Opc == AMDGPU::EXIT_STRICT_WQM); 177 dbgs() << "Strict WQM "; 178 } 179 180 dbgs() << "region: " << MI; 181 } 182 183 #endif 184 185 bool SIPreAllocateWWMRegs::runOnMachineFunction(MachineFunction &MF) { 186 LLVM_DEBUG(dbgs() << "SIPreAllocateWWMRegs: function " << MF.getName() << "\n"); 187 188 const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); 189 190 TII = ST.getInstrInfo(); 191 TRI = &TII->getRegisterInfo(); 192 MRI = &MF.getRegInfo(); 193 194 LIS = &getAnalysis<LiveIntervals>(); 195 Matrix = &getAnalysis<LiveRegMatrix>(); 196 VRM = &getAnalysis<VirtRegMap>(); 197 198 RegClassInfo.runOnMachineFunction(MF); 199 200 bool RegsAssigned = false; 201 202 // We use a reverse post-order traversal of the control-flow graph to 203 // guarantee that we visit definitions in dominance order. Since WWM 204 // expressions are guaranteed to never involve phi nodes, and we can only 205 // escape WWM through the special WWM instruction, this means that this is a 206 // perfect elimination order, so we can never do any better. 207 ReversePostOrderTraversal<MachineFunction*> RPOT(&MF); 208 209 for (MachineBasicBlock *MBB : RPOT) { 210 bool InWWM = false; 211 for (MachineInstr &MI : *MBB) { 212 if (MI.getOpcode() == AMDGPU::V_SET_INACTIVE_B32 || 213 MI.getOpcode() == AMDGPU::V_SET_INACTIVE_B64) 214 RegsAssigned |= processDef(MI.getOperand(0)); 215 216 if (MI.getOpcode() == AMDGPU::ENTER_STRICT_WWM || 217 MI.getOpcode() == AMDGPU::ENTER_STRICT_WQM) { 218 LLVM_DEBUG(printWWMInfo(MI)); 219 InWWM = true; 220 continue; 221 } 222 223 if (MI.getOpcode() == AMDGPU::EXIT_STRICT_WWM || 224 MI.getOpcode() == AMDGPU::EXIT_STRICT_WQM) { 225 LLVM_DEBUG(printWWMInfo(MI)); 226 InWWM = false; 227 } 228 229 if (!InWWM) 230 continue; 231 232 LLVM_DEBUG(dbgs() << "Processing " << MI); 233 234 for (MachineOperand &DefOpnd : MI.defs()) { 235 RegsAssigned |= processDef(DefOpnd); 236 } 237 } 238 } 239 240 if (!RegsAssigned) 241 return false; 242 243 rewriteRegs(MF); 244 return true; 245 } 246