1 //===- R600ExpandSpecialInstrs.cpp - Expand special instructions ----------===// 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 /// Vector, Reduction, and Cube instructions need to fill the entire instruction 11 /// group to work correctly. This pass expands these individual instructions 12 /// into several instructions that will completely fill the instruction group. 13 // 14 //===----------------------------------------------------------------------===// 15 16 #include "AMDGPU.h" 17 #include "AMDGPUSubtarget.h" 18 #include "R600Defines.h" 19 #include "R600InstrInfo.h" 20 #include "R600RegisterInfo.h" 21 #include "llvm/CodeGen/MachineBasicBlock.h" 22 #include "llvm/CodeGen/MachineFunction.h" 23 #include "MCTargetDesc/AMDGPUMCTargetDesc.h" 24 #include "llvm/CodeGen/MachineFunctionPass.h" 25 #include "llvm/CodeGen/MachineInstr.h" 26 #include "llvm/CodeGen/MachineInstrBuilder.h" 27 #include "llvm/CodeGen/MachineOperand.h" 28 #include "llvm/Pass.h" 29 #include <cassert> 30 #include <cstdint> 31 #include <iterator> 32 33 using namespace llvm; 34 35 #define DEBUG_TYPE "r600-expand-special-instrs" 36 37 namespace { 38 39 class R600ExpandSpecialInstrsPass : public MachineFunctionPass { 40 private: 41 const R600InstrInfo *TII = nullptr; 42 43 void SetFlagInNewMI(MachineInstr *NewMI, const MachineInstr *OldMI, 44 unsigned Op); 45 46 public: 47 static char ID; 48 49 R600ExpandSpecialInstrsPass() : MachineFunctionPass(ID) {} 50 51 bool runOnMachineFunction(MachineFunction &MF) override; 52 53 StringRef getPassName() const override { 54 return "R600 Expand special instructions pass"; 55 } 56 }; 57 58 } // end anonymous namespace 59 60 INITIALIZE_PASS_BEGIN(R600ExpandSpecialInstrsPass, DEBUG_TYPE, 61 "R600 Expand Special Instrs", false, false) 62 INITIALIZE_PASS_END(R600ExpandSpecialInstrsPass, DEBUG_TYPE, 63 "R600ExpandSpecialInstrs", false, false) 64 65 char R600ExpandSpecialInstrsPass::ID = 0; 66 67 char &llvm::R600ExpandSpecialInstrsPassID = R600ExpandSpecialInstrsPass::ID; 68 69 FunctionPass *llvm::createR600ExpandSpecialInstrsPass() { 70 return new R600ExpandSpecialInstrsPass(); 71 } 72 73 void R600ExpandSpecialInstrsPass::SetFlagInNewMI(MachineInstr *NewMI, 74 const MachineInstr *OldMI, unsigned Op) { 75 int OpIdx = TII->getOperandIdx(*OldMI, Op); 76 if (OpIdx > -1) { 77 uint64_t Val = OldMI->getOperand(OpIdx).getImm(); 78 TII->setImmOperand(*NewMI, Op, Val); 79 } 80 } 81 82 bool R600ExpandSpecialInstrsPass::runOnMachineFunction(MachineFunction &MF) { 83 const R600Subtarget &ST = MF.getSubtarget<R600Subtarget>(); 84 TII = ST.getInstrInfo(); 85 86 const R600RegisterInfo &TRI = TII->getRegisterInfo(); 87 88 for (MachineFunction::iterator BB = MF.begin(), BB_E = MF.end(); 89 BB != BB_E; ++BB) { 90 MachineBasicBlock &MBB = *BB; 91 MachineBasicBlock::iterator I = MBB.begin(); 92 while (I != MBB.end()) { 93 MachineInstr &MI = *I; 94 I = std::next(I); 95 96 // Expand LDS_*_RET instructions 97 if (TII->isLDSRetInstr(MI.getOpcode())) { 98 int DstIdx = TII->getOperandIdx(MI.getOpcode(), R600::OpName::dst); 99 assert(DstIdx != -1); 100 MachineOperand &DstOp = MI.getOperand(DstIdx); 101 MachineInstr *Mov = TII->buildMovInstr(&MBB, I, 102 DstOp.getReg(), R600::OQAP); 103 DstOp.setReg(R600::OQAP); 104 int LDSPredSelIdx = TII->getOperandIdx(MI.getOpcode(), 105 R600::OpName::pred_sel); 106 int MovPredSelIdx = TII->getOperandIdx(Mov->getOpcode(), 107 R600::OpName::pred_sel); 108 // Copy the pred_sel bit 109 Mov->getOperand(MovPredSelIdx).setReg( 110 MI.getOperand(LDSPredSelIdx).getReg()); 111 } 112 113 switch (MI.getOpcode()) { 114 default: break; 115 // Expand PRED_X to one of the PRED_SET instructions. 116 case R600::PRED_X: { 117 uint64_t Flags = MI.getOperand(3).getImm(); 118 // The native opcode used by PRED_X is stored as an immediate in the 119 // third operand. 120 MachineInstr *PredSet = TII->buildDefaultInstruction(MBB, I, 121 MI.getOperand(2).getImm(), // opcode 122 MI.getOperand(0).getReg(), // dst 123 MI.getOperand(1).getReg(), // src0 124 R600::ZERO); // src1 125 TII->addFlag(*PredSet, 0, MO_FLAG_MASK); 126 if (Flags & MO_FLAG_PUSH) { 127 TII->setImmOperand(*PredSet, R600::OpName::update_exec_mask, 1); 128 } else { 129 TII->setImmOperand(*PredSet, R600::OpName::update_pred, 1); 130 } 131 MI.eraseFromParent(); 132 continue; 133 } 134 case R600::DOT_4: { 135 136 const R600RegisterInfo &TRI = TII->getRegisterInfo(); 137 138 unsigned DstReg = MI.getOperand(0).getReg(); 139 unsigned DstBase = TRI.getEncodingValue(DstReg) & HW_REG_MASK; 140 141 for (unsigned Chan = 0; Chan < 4; ++Chan) { 142 bool Mask = (Chan != TRI.getHWRegChan(DstReg)); 143 unsigned SubDstReg = 144 R600::R600_TReg32RegClass.getRegister((DstBase * 4) + Chan); 145 MachineInstr *BMI = 146 TII->buildSlotOfVectorInstruction(MBB, &MI, Chan, SubDstReg); 147 if (Chan > 0) { 148 BMI->bundleWithPred(); 149 } 150 if (Mask) { 151 TII->addFlag(*BMI, 0, MO_FLAG_MASK); 152 } 153 if (Chan != 3) 154 TII->addFlag(*BMI, 0, MO_FLAG_NOT_LAST); 155 unsigned Opcode = BMI->getOpcode(); 156 // While not strictly necessary from hw point of view, we force 157 // all src operands of a dot4 inst to belong to the same slot. 158 unsigned Src0 = BMI->getOperand( 159 TII->getOperandIdx(Opcode, R600::OpName::src0)) 160 .getReg(); 161 unsigned Src1 = BMI->getOperand( 162 TII->getOperandIdx(Opcode, R600::OpName::src1)) 163 .getReg(); 164 (void) Src0; 165 (void) Src1; 166 if ((TRI.getEncodingValue(Src0) & 0xff) < 127 && 167 (TRI.getEncodingValue(Src1) & 0xff) < 127) 168 assert(TRI.getHWRegChan(Src0) == TRI.getHWRegChan(Src1)); 169 } 170 MI.eraseFromParent(); 171 continue; 172 } 173 } 174 175 bool IsReduction = TII->isReductionOp(MI.getOpcode()); 176 bool IsVector = TII->isVector(MI); 177 bool IsCube = TII->isCubeOp(MI.getOpcode()); 178 if (!IsReduction && !IsVector && !IsCube) { 179 continue; 180 } 181 182 // Expand the instruction 183 // 184 // Reduction instructions: 185 // T0_X = DP4 T1_XYZW, T2_XYZW 186 // becomes: 187 // TO_X = DP4 T1_X, T2_X 188 // TO_Y (write masked) = DP4 T1_Y, T2_Y 189 // TO_Z (write masked) = DP4 T1_Z, T2_Z 190 // TO_W (write masked) = DP4 T1_W, T2_W 191 // 192 // Vector instructions: 193 // T0_X = MULLO_INT T1_X, T2_X 194 // becomes: 195 // T0_X = MULLO_INT T1_X, T2_X 196 // T0_Y (write masked) = MULLO_INT T1_X, T2_X 197 // T0_Z (write masked) = MULLO_INT T1_X, T2_X 198 // T0_W (write masked) = MULLO_INT T1_X, T2_X 199 // 200 // Cube instructions: 201 // T0_XYZW = CUBE T1_XYZW 202 // becomes: 203 // TO_X = CUBE T1_Z, T1_Y 204 // T0_Y = CUBE T1_Z, T1_X 205 // T0_Z = CUBE T1_X, T1_Z 206 // T0_W = CUBE T1_Y, T1_Z 207 for (unsigned Chan = 0; Chan < 4; Chan++) { 208 unsigned DstReg = MI.getOperand( 209 TII->getOperandIdx(MI, R600::OpName::dst)).getReg(); 210 unsigned Src0 = MI.getOperand( 211 TII->getOperandIdx(MI, R600::OpName::src0)).getReg(); 212 unsigned Src1 = 0; 213 214 // Determine the correct source registers 215 if (!IsCube) { 216 int Src1Idx = TII->getOperandIdx(MI, R600::OpName::src1); 217 if (Src1Idx != -1) { 218 Src1 = MI.getOperand(Src1Idx).getReg(); 219 } 220 } 221 if (IsReduction) { 222 unsigned SubRegIndex = AMDGPURegisterInfo::getSubRegFromChannel(Chan); 223 Src0 = TRI.getSubReg(Src0, SubRegIndex); 224 Src1 = TRI.getSubReg(Src1, SubRegIndex); 225 } else if (IsCube) { 226 static const int CubeSrcSwz[] = {2, 2, 0, 1}; 227 unsigned SubRegIndex0 = AMDGPURegisterInfo::getSubRegFromChannel(CubeSrcSwz[Chan]); 228 unsigned SubRegIndex1 = AMDGPURegisterInfo::getSubRegFromChannel(CubeSrcSwz[3 - Chan]); 229 Src1 = TRI.getSubReg(Src0, SubRegIndex1); 230 Src0 = TRI.getSubReg(Src0, SubRegIndex0); 231 } 232 233 // Determine the correct destination registers; 234 bool Mask = false; 235 bool NotLast = true; 236 if (IsCube) { 237 unsigned SubRegIndex = AMDGPURegisterInfo::getSubRegFromChannel(Chan); 238 DstReg = TRI.getSubReg(DstReg, SubRegIndex); 239 } else { 240 // Mask the write if the original instruction does not write to 241 // the current Channel. 242 Mask = (Chan != TRI.getHWRegChan(DstReg)); 243 unsigned DstBase = TRI.getEncodingValue(DstReg) & HW_REG_MASK; 244 DstReg = R600::R600_TReg32RegClass.getRegister((DstBase * 4) + Chan); 245 } 246 247 // Set the IsLast bit 248 NotLast = (Chan != 3 ); 249 250 // Add the new instruction 251 unsigned Opcode = MI.getOpcode(); 252 switch (Opcode) { 253 case R600::CUBE_r600_pseudo: 254 Opcode = R600::CUBE_r600_real; 255 break; 256 case R600::CUBE_eg_pseudo: 257 Opcode = R600::CUBE_eg_real; 258 break; 259 default: 260 break; 261 } 262 263 MachineInstr *NewMI = 264 TII->buildDefaultInstruction(MBB, I, Opcode, DstReg, Src0, Src1); 265 266 if (Chan != 0) 267 NewMI->bundleWithPred(); 268 if (Mask) { 269 TII->addFlag(*NewMI, 0, MO_FLAG_MASK); 270 } 271 if (NotLast) { 272 TII->addFlag(*NewMI, 0, MO_FLAG_NOT_LAST); 273 } 274 SetFlagInNewMI(NewMI, &MI, R600::OpName::clamp); 275 SetFlagInNewMI(NewMI, &MI, R600::OpName::literal); 276 SetFlagInNewMI(NewMI, &MI, R600::OpName::src0_abs); 277 SetFlagInNewMI(NewMI, &MI, R600::OpName::src1_abs); 278 SetFlagInNewMI(NewMI, &MI, R600::OpName::src0_neg); 279 SetFlagInNewMI(NewMI, &MI, R600::OpName::src1_neg); 280 } 281 MI.eraseFromParent(); 282 } 283 } 284 return false; 285 } 286