1 //===-- R600ExpandSpecialInstrs.cpp - Expand special instructions ---------===// 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 /// \file 11 /// Vector, Reduction, and Cube instructions need to fill the entire instruction 12 /// group to work correctly. This pass expands these individual instructions 13 /// into several instructions that will completely fill the instruction group. 14 // 15 //===----------------------------------------------------------------------===// 16 17 #include "AMDGPU.h" 18 #include "R600Defines.h" 19 #include "R600InstrInfo.h" 20 #include "R600MachineFunctionInfo.h" 21 #include "R600RegisterInfo.h" 22 #include "llvm/CodeGen/MachineFunctionPass.h" 23 #include "llvm/CodeGen/MachineInstrBuilder.h" 24 #include "llvm/CodeGen/MachineRegisterInfo.h" 25 26 using namespace llvm; 27 28 namespace { 29 30 class R600ExpandSpecialInstrsPass : public MachineFunctionPass { 31 32 private: 33 static char ID; 34 const R600InstrInfo *TII; 35 36 bool ExpandInputPerspective(MachineInstr& MI); 37 bool ExpandInputConstant(MachineInstr& MI); 38 39 public: 40 R600ExpandSpecialInstrsPass(TargetMachine &tm) : MachineFunctionPass(ID), 41 TII (static_cast<const R600InstrInfo *>(tm.getInstrInfo())) { } 42 43 virtual bool runOnMachineFunction(MachineFunction &MF); 44 45 const char *getPassName() const { 46 return "R600 Expand special instructions pass"; 47 } 48 }; 49 50 } // End anonymous namespace 51 52 char R600ExpandSpecialInstrsPass::ID = 0; 53 54 FunctionPass *llvm::createR600ExpandSpecialInstrsPass(TargetMachine &TM) { 55 return new R600ExpandSpecialInstrsPass(TM); 56 } 57 58 bool R600ExpandSpecialInstrsPass::runOnMachineFunction(MachineFunction &MF) { 59 60 const R600RegisterInfo &TRI = TII->getRegisterInfo(); 61 62 for (MachineFunction::iterator BB = MF.begin(), BB_E = MF.end(); 63 BB != BB_E; ++BB) { 64 MachineBasicBlock &MBB = *BB; 65 MachineBasicBlock::iterator I = MBB.begin(); 66 while (I != MBB.end()) { 67 MachineInstr &MI = *I; 68 I = llvm::next(I); 69 70 switch (MI.getOpcode()) { 71 default: break; 72 // Expand PRED_X to one of the PRED_SET instructions. 73 case AMDGPU::PRED_X: { 74 uint64_t Flags = MI.getOperand(3).getImm(); 75 // The native opcode used by PRED_X is stored as an immediate in the 76 // third operand. 77 MachineInstr *PredSet = TII->buildDefaultInstruction(MBB, I, 78 MI.getOperand(2).getImm(), // opcode 79 MI.getOperand(0).getReg(), // dst 80 MI.getOperand(1).getReg(), // src0 81 AMDGPU::ZERO); // src1 82 TII->addFlag(PredSet, 0, MO_FLAG_MASK); 83 if (Flags & MO_FLAG_PUSH) { 84 TII->setImmOperand(PredSet, R600Operands::UPDATE_EXEC_MASK, 1); 85 } else { 86 TII->setImmOperand(PredSet, R600Operands::UPDATE_PREDICATE, 1); 87 } 88 MI.eraseFromParent(); 89 continue; 90 } 91 case AMDGPU::BREAK: { 92 MachineInstr *PredSet = TII->buildDefaultInstruction(MBB, I, 93 AMDGPU::PRED_SETE_INT, 94 AMDGPU::PREDICATE_BIT, 95 AMDGPU::ZERO, 96 AMDGPU::ZERO); 97 TII->addFlag(PredSet, 0, MO_FLAG_MASK); 98 TII->setImmOperand(PredSet, R600Operands::UPDATE_EXEC_MASK, 1); 99 100 BuildMI(MBB, I, MBB.findDebugLoc(I), 101 TII->get(AMDGPU::PREDICATED_BREAK)) 102 .addReg(AMDGPU::PREDICATE_BIT); 103 MI.eraseFromParent(); 104 continue; 105 } 106 107 case AMDGPU::INTERP_PAIR_XY: { 108 MachineInstr *BMI; 109 unsigned PReg = AMDGPU::R600_ArrayBaseRegClass.getRegister( 110 MI.getOperand(2).getImm()); 111 112 for (unsigned Chan = 0; Chan < 4; ++Chan) { 113 unsigned DstReg; 114 115 if (Chan < 2) 116 DstReg = MI.getOperand(Chan).getReg(); 117 else 118 DstReg = Chan == 2 ? AMDGPU::T0_Z : AMDGPU::T0_W; 119 120 BMI = TII->buildDefaultInstruction(MBB, I, AMDGPU::INTERP_XY, 121 DstReg, MI.getOperand(3 + (Chan % 2)).getReg(), PReg); 122 123 if (Chan > 0) { 124 BMI->bundleWithPred(); 125 } 126 if (Chan >= 2) 127 TII->addFlag(BMI, 0, MO_FLAG_MASK); 128 if (Chan != 3) 129 TII->addFlag(BMI, 0, MO_FLAG_NOT_LAST); 130 } 131 132 MI.eraseFromParent(); 133 continue; 134 } 135 136 case AMDGPU::INTERP_PAIR_ZW: { 137 MachineInstr *BMI; 138 unsigned PReg = AMDGPU::R600_ArrayBaseRegClass.getRegister( 139 MI.getOperand(2).getImm()); 140 141 for (unsigned Chan = 0; Chan < 4; ++Chan) { 142 unsigned DstReg; 143 144 if (Chan < 2) 145 DstReg = Chan == 0 ? AMDGPU::T0_X : AMDGPU::T0_Y; 146 else 147 DstReg = MI.getOperand(Chan-2).getReg(); 148 149 BMI = TII->buildDefaultInstruction(MBB, I, AMDGPU::INTERP_ZW, 150 DstReg, MI.getOperand(3 + (Chan % 2)).getReg(), PReg); 151 152 if (Chan > 0) { 153 BMI->bundleWithPred(); 154 } 155 if (Chan < 2) 156 TII->addFlag(BMI, 0, MO_FLAG_MASK); 157 if (Chan != 3) 158 TII->addFlag(BMI, 0, MO_FLAG_NOT_LAST); 159 } 160 161 MI.eraseFromParent(); 162 continue; 163 } 164 165 case AMDGPU::INTERP_VEC_LOAD: { 166 const R600RegisterInfo &TRI = TII->getRegisterInfo(); 167 MachineInstr *BMI; 168 unsigned PReg = AMDGPU::R600_ArrayBaseRegClass.getRegister( 169 MI.getOperand(1).getImm()); 170 unsigned DstReg = MI.getOperand(0).getReg(); 171 172 for (unsigned Chan = 0; Chan < 4; ++Chan) { 173 BMI = TII->buildDefaultInstruction(MBB, I, AMDGPU::INTERP_LOAD_P0, 174 TRI.getSubReg(DstReg, TRI.getSubRegFromChannel(Chan)), PReg); 175 if (Chan > 0) { 176 BMI->bundleWithPred(); 177 } 178 if (Chan != 3) 179 TII->addFlag(BMI, 0, MO_FLAG_NOT_LAST); 180 } 181 182 MI.eraseFromParent(); 183 continue; 184 } 185 } 186 187 bool IsReduction = TII->isReductionOp(MI.getOpcode()); 188 bool IsVector = TII->isVector(MI); 189 bool IsCube = TII->isCubeOp(MI.getOpcode()); 190 if (!IsReduction && !IsVector && !IsCube) { 191 continue; 192 } 193 194 // Expand the instruction 195 // 196 // Reduction instructions: 197 // T0_X = DP4 T1_XYZW, T2_XYZW 198 // becomes: 199 // TO_X = DP4 T1_X, T2_X 200 // TO_Y (write masked) = DP4 T1_Y, T2_Y 201 // TO_Z (write masked) = DP4 T1_Z, T2_Z 202 // TO_W (write masked) = DP4 T1_W, T2_W 203 // 204 // Vector instructions: 205 // T0_X = MULLO_INT T1_X, T2_X 206 // becomes: 207 // T0_X = MULLO_INT T1_X, T2_X 208 // T0_Y (write masked) = MULLO_INT T1_X, T2_X 209 // T0_Z (write masked) = MULLO_INT T1_X, T2_X 210 // T0_W (write masked) = MULLO_INT T1_X, T2_X 211 // 212 // Cube instructions: 213 // T0_XYZW = CUBE T1_XYZW 214 // becomes: 215 // TO_X = CUBE T1_Z, T1_Y 216 // T0_Y = CUBE T1_Z, T1_X 217 // T0_Z = CUBE T1_X, T1_Z 218 // T0_W = CUBE T1_Y, T1_Z 219 for (unsigned Chan = 0; Chan < 4; Chan++) { 220 unsigned DstReg = MI.getOperand( 221 TII->getOperandIdx(MI, R600Operands::DST)).getReg(); 222 unsigned Src0 = MI.getOperand( 223 TII->getOperandIdx(MI, R600Operands::SRC0)).getReg(); 224 unsigned Src1 = 0; 225 226 // Determine the correct source registers 227 if (!IsCube) { 228 int Src1Idx = TII->getOperandIdx(MI, R600Operands::SRC1); 229 if (Src1Idx != -1) { 230 Src1 = MI.getOperand(Src1Idx).getReg(); 231 } 232 } 233 if (IsReduction) { 234 unsigned SubRegIndex = TRI.getSubRegFromChannel(Chan); 235 Src0 = TRI.getSubReg(Src0, SubRegIndex); 236 Src1 = TRI.getSubReg(Src1, SubRegIndex); 237 } else if (IsCube) { 238 static const int CubeSrcSwz[] = {2, 2, 0, 1}; 239 unsigned SubRegIndex0 = TRI.getSubRegFromChannel(CubeSrcSwz[Chan]); 240 unsigned SubRegIndex1 = TRI.getSubRegFromChannel(CubeSrcSwz[3 - Chan]); 241 Src1 = TRI.getSubReg(Src0, SubRegIndex1); 242 Src0 = TRI.getSubReg(Src0, SubRegIndex0); 243 } 244 245 // Determine the correct destination registers; 246 bool Mask = false; 247 bool NotLast = true; 248 if (IsCube) { 249 unsigned SubRegIndex = TRI.getSubRegFromChannel(Chan); 250 DstReg = TRI.getSubReg(DstReg, SubRegIndex); 251 } else { 252 // Mask the write if the original instruction does not write to 253 // the current Channel. 254 Mask = (Chan != TRI.getHWRegChan(DstReg)); 255 unsigned DstBase = TRI.getEncodingValue(DstReg) & HW_REG_MASK; 256 DstReg = AMDGPU::R600_TReg32RegClass.getRegister((DstBase * 4) + Chan); 257 } 258 259 // Set the IsLast bit 260 NotLast = (Chan != 3 ); 261 262 // Add the new instruction 263 unsigned Opcode = MI.getOpcode(); 264 switch (Opcode) { 265 case AMDGPU::CUBE_r600_pseudo: 266 Opcode = AMDGPU::CUBE_r600_real; 267 break; 268 case AMDGPU::CUBE_eg_pseudo: 269 Opcode = AMDGPU::CUBE_eg_real; 270 break; 271 case AMDGPU::DOT4_r600_pseudo: 272 Opcode = AMDGPU::DOT4_r600_real; 273 break; 274 case AMDGPU::DOT4_eg_pseudo: 275 Opcode = AMDGPU::DOT4_eg_real; 276 break; 277 default: 278 break; 279 } 280 281 MachineInstr *NewMI = 282 TII->buildDefaultInstruction(MBB, I, Opcode, DstReg, Src0, Src1); 283 284 if (Chan != 0) 285 NewMI->bundleWithPred(); 286 if (Mask) { 287 TII->addFlag(NewMI, 0, MO_FLAG_MASK); 288 } 289 if (NotLast) { 290 TII->addFlag(NewMI, 0, MO_FLAG_NOT_LAST); 291 } 292 } 293 MI.eraseFromParent(); 294 } 295 } 296 return false; 297 } 298