| 1 | //===--------------------- HexagonPostRAHandleQFP.cpp -------------------------- |
| 2 | //===// |
| 3 | // |
| 4 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 5 | // See https://llvm.org/LICENSE.txt for license information. |
| 6 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 7 | // |
| 8 | //===---------------------------------------------------------------------===// |
| 9 | // For v79 and above, we generate qf operations for HVX which includes vadd, |
| 10 | // vsub and vmpy instructions. These qf operations with qf operands are fast, |
| 11 | // maintain similar accuracy as IEEE and saves power. |
| 12 | // |
| 13 | // However, these qf operands should always be converted back to IEEE format |
| 14 | // when used in non-HVX instructions. This is because of how the qf values |
| 15 | // are stored in memory. qf operands have 4 extra bits. If used in non-HVX |
| 16 | // operations, these bits get dropped resulting in incorrect value being |
| 17 | // used. So, before use in any non-HVX operation we need to convert these |
| 18 | // qf values to IEEE format. |
| 19 | // |
| 20 | // During register allocation, when no more physical registers are available |
| 21 | // the qf operands may be spilled to memory. This instantly causes loss of |
| 22 | // accuracy. This pass prevents that by: |
| 23 | // 1. Inserting qf type to IEEE type conversion instructions before the spill. |
| 24 | // 2. Iterating over the uses of qf def (created before the spill) and |
| 25 | // changing their opcodes to handle IEEE type operands for saturating |
| 26 | // instructions. This is because, the refills will use IEEE type operands, but |
| 27 | // the instructions will still assume qf operands. For non-saturating |
| 28 | // instructions which uses qf, we incorporate a conversion to IEEE before that. |
| 29 | // 3. Iterating over the uses of qf def created by the spill and replacing |
| 30 | // them with appropiate opcode (which uses IEEE operands) for saturating |
| 31 | // instructions. For non-saturating instructions which uses qf, |
| 32 | // we incorporate a conversion to IEEE before that. |
| 33 | // 4. Iterating over the copy instructions and checking their uses, |
| 34 | // inserting conversions from qf to IEEE whenever required. The conversions |
| 35 | // are inserted after their reaching def since there can be multiple defs |
| 36 | // for use in non-SSA form. |
| 37 | // |
| 38 | // To get the use-def chains, we make use of Register DataFlow Graph (RDF), |
| 39 | // since after register allocation SSA form is lost. This can be done during |
| 40 | // spills and fills during Frame Lowering for register allocation. However, |
| 41 | // that was abandoned due to the intermediate state of the code. |
| 42 | // Liveness is preserved in this pass. |
| 43 | // |
| 44 | // NOTE: |
| 45 | // Saturating instructions: Instructions for which transformation involves |
| 46 | // only changing the opcode. Eg. vmpy(qf32, sf) saturates to vmpy(sf, sf) when |
| 47 | // we see that the first operand is now a sf type. |
| 48 | // Non-Saturating instructions: Instructions for which conversion(s) have |
| 49 | // to be inserted. Eg. Vd.f8=Vu.qf16. If the use operand is now hf type, |
| 50 | // we have to insert a conversion qf16 = hf before this instruction. |
| 51 | // |
| 52 | // FIXME tags have been added for potential errors, along with the underlying |
| 53 | // assumption. |
| 54 | // FIXME Implement v81 specific optimizations as below. At the moment, we add |
| 55 | // converts. |
| 56 | // Vd.qf16=Vu.hf |
| 57 | // Vd.qf16=Vu.qf16 |
| 58 | // Vd.qf32=Vu.qf32 |
| 59 | // Vd.qf32=Vu.sf |
| 60 | //===---------------------------------------------------------------------===// |
| 61 | |
| 62 | #include "HexagonTargetMachine.h" |
| 63 | #include "llvm/CodeGen/LiveInterval.h" |
| 64 | #include "llvm/CodeGen/LiveIntervals.h" |
| 65 | #include "llvm/CodeGen/LivePhysRegs.h" |
| 66 | #include "llvm/CodeGen/MachineDominanceFrontier.h" |
| 67 | #include "llvm/CodeGen/MachineDominators.h" |
| 68 | #include "llvm/CodeGen/MachineFunction.h" |
| 69 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
| 70 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
| 71 | #include "llvm/CodeGen/Passes.h" |
| 72 | #include "llvm/CodeGen/RDFGraph.h" |
| 73 | #include "llvm/CodeGen/RDFLiveness.h" |
| 74 | #include "llvm/CodeGen/RDFRegisters.h" |
| 75 | #include "llvm/CodeGen/TargetInstrInfo.h" |
| 76 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
| 77 | #include "llvm/InitializePasses.h" |
| 78 | #include "llvm/Support/CommandLine.h" |
| 79 | #include "llvm/Support/Debug.h" |
| 80 | #include "llvm/Support/raw_ostream.h" |
| 81 | #include "llvm/Target/TargetMachine.h" |
| 82 | |
| 83 | #define DEBUG_TYPE "handle-qfp" |
| 84 | |
| 85 | using namespace llvm; |
| 86 | using namespace rdf; |
| 87 | |
| 88 | extern cl::opt<QFloatMode> QFloatModeValue; |
| 89 | |
| 90 | cl::opt<bool> DisablePostRAHandleQFloat( |
| 91 | "disable-handle-qfp" , cl::init(Val: false), |
| 92 | cl::desc("Disable handling of Qfloat spills/refills after register " |
| 93 | "allocation." )); |
| 94 | |
| 95 | // This static function gets all reached uses of a def. |
| 96 | // When it encounters a phi node, it goes over the |
| 97 | // reached uses of the phi node too. |
| 98 | static void getAllRealUses(NodeAddr<DefNode *> DA, NodeSet &UNodeSet, |
| 99 | Liveness *L, DataFlowGraph *G, |
| 100 | bool comprehensive = false) { |
| 101 | RegisterRef DR = DA.Addr->getRegRef(G: *G); |
| 102 | NodeAddr<StmtNode *> DefStmt = DA.Addr->getOwner(G: *G); |
| 103 | MachineInstr *Instr = DefStmt.Addr->getCode(); |
| 104 | auto UseSet = L->getAllReachedUses(RefRR: DR, DefA: DA); |
| 105 | |
| 106 | for (auto UI : UseSet) { |
| 107 | NodeAddr<UseNode *> UA = G->addr<UseNode *>(N: UI); |
| 108 | |
| 109 | /*LLVM_DEBUG( |
| 110 | NodeAddr<StmtNode *> UseStmt = UA.Addr->getOwner(*G); |
| 111 | MachineInstr* UseInstr = UseStmt.Addr->getCode(); |
| 112 | if (UseInstr != nullptr) |
| 113 | {dbgs() << "\t\t[Reached Use]: "; UseInstr->dump();} |
| 114 | );*/ |
| 115 | |
| 116 | MachineFunction *MF = Instr->getMF(); |
| 117 | const auto &HRI = MF->getSubtarget<HexagonSubtarget>().getRegisterInfo(); |
| 118 | Register RR = UA.Addr->getRegRef(G: *G).Id; |
| 119 | if (HRI->isFakeReg(Reg: RR)) |
| 120 | continue; |
| 121 | |
| 122 | if (UA.Addr->getFlags() & NodeAttrs::PhiRef) { |
| 123 | NodeAddr<PhiNode *> PA = UA.Addr->getOwner(G: *G); |
| 124 | NodeId id = PA.Id; |
| 125 | const Liveness::RefMap &phiUse = L->getRealUses(P: id); |
| 126 | for (auto I : phiUse) { |
| 127 | if (!G->getPRI().alias(RA: RegisterRef(I.first), RB: DR)) |
| 128 | continue; |
| 129 | auto phiUseSet = I.second; |
| 130 | for (auto phiUI : phiUseSet) { |
| 131 | NodeAddr<UseNode *> phiUA = G->addr<UseNode *>(N: phiUI.first); |
| 132 | UNodeSet.insert(x: phiUA.Id); |
| 133 | } |
| 134 | } |
| 135 | } else { |
| 136 | // FIXME Due to bug in RDF, check if the reaching def of the use |
| 137 | // reaches this instruction |
| 138 | if (comprehensive) { |
| 139 | UNodeSet.insert(x: UA.Id); |
| 140 | continue; |
| 141 | } |
| 142 | NodeAddr<StmtNode *> UseStmt = UA.Addr->getOwner(G: *G); |
| 143 | for (NodeAddr<UseNode *> UA : UseStmt.Addr->members_if(P: G->IsUse, G: *G)) { |
| 144 | NodeId QFPDefNode = UA.Addr->getReachingDef(); |
| 145 | NodeAddr<DefNode *> RegDef = G->addr<DefNode *>(N: QFPDefNode); |
| 146 | // FIXME Reaching def computation error |
| 147 | if (QFPDefNode == 0) |
| 148 | continue; |
| 149 | NodeAddr<StmtNode *> RegStmt = RegDef.Addr->getOwner(G: *G); |
| 150 | MachineInstr *ReachDefInstr = RegStmt.Addr->getCode(); |
| 151 | if (ReachDefInstr && ReachDefInstr == Instr) |
| 152 | UNodeSet.insert(x: UA.Id); |
| 153 | } |
| 154 | } |
| 155 | } |
| 156 | } |
| 157 | |
| 158 | namespace llvm { |
| 159 | FunctionPass *createHexagonPostRAHandleQFP(); |
| 160 | void initializeHexagonPostRAHandleQFPPass(PassRegistry &); |
| 161 | } // namespace llvm |
| 162 | |
| 163 | // QF Instructions list which need to be analyzed. |
| 164 | // The value of the key denotes a pair |
| 165 | // pair.first|pair.second = True if IEEE type, false otherwise. |
| 166 | // We only need to change the opcode to handling qf/sf |
| 167 | // misuses for these, or these instructions can be 'saturated'. |
| 168 | DenseMap<unsigned short, std::pair<bool, bool>> QFPSatInstsMap{ |
| 169 | {Hexagon::V6_vadd_qf16_mix, {false, true}}, |
| 170 | {Hexagon::V6_vadd_qf16, {false, false}}, |
| 171 | {Hexagon::V6_vadd_qf32_mix, {false, true}}, |
| 172 | {Hexagon::V6_vadd_qf32, {false, false}}, |
| 173 | {Hexagon::V6_vsub_qf16_mix, {false, true}}, |
| 174 | {Hexagon::V6_vsub_hf_mix, {true, false}}, |
| 175 | {Hexagon::V6_vsub_qf16, {false, false}}, |
| 176 | {Hexagon::V6_vsub_qf32_mix, {false, true}}, |
| 177 | {Hexagon::V6_vsub_sf_mix, {true, false}}, |
| 178 | {Hexagon::V6_vsub_qf32, {false, false}}, |
| 179 | {Hexagon::V6_vmpy_qf16_mix_hf, {false, true}}, |
| 180 | {Hexagon::V6_vmpy_qf16, {false, false}}, |
| 181 | {Hexagon::V6_vmpy_qf32_mix_hf, {false, true}}, |
| 182 | {Hexagon::V6_vmpy_qf32_qf16, {false, false}}, |
| 183 | {Hexagon::V6_vmpy_qf32, {false, false}}, |
| 184 | {Hexagon::V6_vmpy_rt_qf16, {false, true}}, |
| 185 | // These opcodes take a single operand only. |
| 186 | // Second placeholder op is true always. |
| 187 | {Hexagon::V6_vabs_qf32_qf32, {false, true}}, |
| 188 | {Hexagon::V6_vabs_qf16_qf16, {false, true}}, |
| 189 | {Hexagon::V6_vneg_qf32_qf32, {false, true}}, |
| 190 | {Hexagon::V6_vneg_qf16_qf16, {false, true}}, |
| 191 | {Hexagon::V6_vilog2_qf32, {false, true}}, |
| 192 | {Hexagon::V6_vilog2_qf16, {false, true}}, |
| 193 | {Hexagon::V6_vconv_qf32_qf32, {false, true}}, |
| 194 | {Hexagon::V6_vconv_qf16_qf16, {false, true}}, |
| 195 | }; |
| 196 | |
| 197 | // This holds the instruction opcodes for which there are |
| 198 | // no 'saturating' opcodes. The only way is to insert |
| 199 | // convert instructions before them. |
| 200 | SmallVector<unsigned short, 5> QFNonSatInstr{ |
| 201 | Hexagon::V6_vconv_hf_qf16, Hexagon::V6_vconv_hf_qf32, |
| 202 | Hexagon::V6_vconv_sf_qf32, |
| 203 | // v81 instructions |
| 204 | Hexagon::V6_vconv_bf_qf32, Hexagon::V6_vconv_f8_qf16}; |
| 205 | |
| 206 | namespace { |
| 207 | class HexagonPostRAHandleQFP : public MachineFunctionPass { |
| 208 | public: |
| 209 | static char ID; |
| 210 | HexagonPostRAHandleQFP() : MachineFunctionPass(ID) { |
| 211 | PassRegistry &R = *PassRegistry::getPassRegistry(); |
| 212 | initializeHexagonPostRAHandleQFPPass(R); |
| 213 | } |
| 214 | StringRef getPassName() const override { |
| 215 | return "Hexagon handle QFloat spills and refills post RA." ; |
| 216 | } |
| 217 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
| 218 | MachineFunctionPass::getAnalysisUsage(AU); |
| 219 | AU.addRequired<MachineDominatorTreeWrapperPass>(); |
| 220 | AU.addRequired<MachineDominanceFrontierWrapperPass>(); |
| 221 | AU.setPreservesCFG(); |
| 222 | } |
| 223 | bool runOnMachineFunction(MachineFunction &MF) override; |
| 224 | |
| 225 | private: |
| 226 | // QFUses collects the instructions which uses QF operands. |
| 227 | // These have to be deleted and transformed to opcodes |
| 228 | // to denote usage of IEEE operands. |
| 229 | // It might involve changing the order of the Register operands. |
| 230 | using QFUses = std::map<MachineInstr *, std::pair<bool, bool>>; |
| 231 | QFUses QFUsesMap; |
| 232 | |
| 233 | // Holds the Register Dataglow Graph. |
| 234 | DataFlowGraph *DFG = nullptr; |
| 235 | |
| 236 | // Stores spill nodes and their reaching definition instructions |
| 237 | // which generates the qf operand to be stored. |
| 238 | std::vector<std::pair<MachineInstr *, NodeAddr<DefNode *>>> SpillMIs; |
| 239 | // Stores the refill nodes consisting of load instructions. |
| 240 | std::vector<NodeAddr<DefNode *>> RefillMIs; |
| 241 | |
| 242 | // Stores the type of op. |
| 243 | enum ConvOperand { |
| 244 | Undefined = 0x0, |
| 245 | Lo = 0x1, |
| 246 | Hi = 0x2, |
| 247 | HiLo = 0x3, |
| 248 | }; |
| 249 | // Stores the convert instructions which take qf operands. |
| 250 | MapVector<MachineInstr *, unsigned> QFNonSatMIs; |
| 251 | |
| 252 | // Stores the qf-generating vmul/vadd/etc. nodes with mutiple reaching defs |
| 253 | std::set<NodeId> PossibleMultiReachDefs; |
| 254 | // Qf generating instructions to ignore. Do not insert conversion instruction |
| 255 | // to sf/hf from qf, if the instr is present in this list; since that means |
| 256 | // a conversion has already been inserted after the instruction. |
| 257 | SmallPtrSet<MachineInstr *, 4> IgnoreInsertConvList; |
| 258 | |
| 259 | // Register type |
| 260 | enum class RegType { qf32, qf16, qf32_double, qf16_double, ieee, undefined }; |
| 261 | // Stores the copy instructions which their reaching def, along with the op |
| 262 | // type |
| 263 | std::map<std::pair<NodeId, NodeId>, RegType> QFCopys; |
| 264 | |
| 265 | // Stores the reaching defs of copies whose result has to be converted to IEEE |
| 266 | DenseMap<MachineInstr *, RegType> ReachDefOfCopies; |
| 267 | |
| 268 | // Stores copies which need to be converted back to qf. The uses of these |
| 269 | // copies feed to qf type instructions and hence can be converted back to qf |
| 270 | // type. |
| 271 | DenseMap<MachineInstr *, std::pair<NodeAddr<DefNode *>, RegType>> |
| 272 | ConvertToQfCopies; |
| 273 | |
| 274 | // Subregister kill set for a doubletype use. The pair of bool,bool |
| 275 | // represents the hi and lo subregisters of the double register. |
| 276 | DenseMap<MachineInstr *, std::pair<bool, bool>> SubRegKillSet; |
| 277 | |
| 278 | const HexagonInstrInfo *HII = nullptr; |
| 279 | const HexagonRegisterInfo *HRI = nullptr; |
| 280 | MachineRegisterInfo *MRI = nullptr; |
| 281 | Liveness *LV = nullptr; |
| 282 | const HexagonSubtarget *HST = nullptr; |
| 283 | |
| 284 | void collectQFPStackSpill(NodeAddr<StmtNode *> *); |
| 285 | void collectQFPStackRefill(NodeAddr<StmtNode *> *); |
| 286 | void collectCopies(NodeAddr<StmtNode *> *); |
| 287 | bool HandleRefills(); |
| 288 | bool HandleSpills(); |
| 289 | bool HandleCopies(); |
| 290 | bool HandleNonSatInstr(); |
| 291 | bool HandleMultiReachingDefs(); |
| 292 | bool HandleReachDefOfCopies(); |
| 293 | bool HandleConvertToQfCopies(); |
| 294 | RegType HasQfUses(NodeAddr<DefNode *>, MachineInstr *); |
| 295 | void collectConvQFInstr(NodeAddr<DefNode *> &); |
| 296 | void collectQFUses(NodeAddr<DefNode *>, MachineInstr *DefMI); |
| 297 | void conditionallyInsert(MachineInstr &, Register &); |
| 298 | |
| 299 | // Helper functions |
| 300 | unsigned short getreplacedQFOpcode(unsigned, bool, bool); |
| 301 | MCPhysReg findAllocatableReg(MachineInstr *MI) const; |
| 302 | void insertIEEEToQF(MachineInstr *, Register, MachineOperand, bool is32bit); |
| 303 | void collectLivenessForSubregs(NodeAddr<UseNode *> &); |
| 304 | void insertInstr(MachineInstr *, unsigned, unsigned, unsigned, RegState); |
| 305 | }; |
| 306 | } // namespace |
| 307 | |
| 308 | // This class handles spurious vector instrutions which do not |
| 309 | // follow the ABI. For eg, vcombine(qf,qf) takes qf operands |
| 310 | // instead of IEEE type. This diagnostic pass can be used |
| 311 | // as a final verifier for XQF implementation. Turned off by |
| 312 | // default |
| 313 | char HexagonPostRAHandleQFP::ID = 0; |
| 314 | |
| 315 | namespace llvm { |
| 316 | char &HexagonPostRAHandleQFPID = HexagonPostRAHandleQFP::ID; |
| 317 | } |
| 318 | |
| 319 | // Check whether the instruction is added already, if not add it |
| 320 | // along with the Register values and qf type. |
| 321 | // If already added, then check the register values and edit them. |
| 322 | void HexagonPostRAHandleQFP::conditionallyInsert(MachineInstr &MI, |
| 323 | Register &DefReg) { |
| 324 | LLVM_DEBUG(dbgs() << "\nCollecting instruction using QF: " ; MI.dump()); |
| 325 | // check if the key exists. |
| 326 | Register Reg1 = MI.getOperand(i: 1).getReg(); |
| 327 | |
| 328 | // If the use is a unary operation, make second register point to Defreg |
| 329 | // This ensures that secondOp is always true |
| 330 | Register Reg2 = MI.getNumOperands() == 2 ? DefReg : MI.getOperand(i: 2).getReg(); |
| 331 | |
| 332 | if (QFUsesMap.find(x: &MI) != QFUsesMap.end()) { |
| 333 | auto Entry = QFUsesMap[&MI]; |
| 334 | bool firstOp = ((Reg1 == DefReg) ? true : false) | Entry.first; |
| 335 | bool secondOp = ((Reg2 == DefReg) ? true : false) | Entry.second; |
| 336 | QFUsesMap[&MI] = std::make_pair(x&: firstOp, y&: secondOp); |
| 337 | |
| 338 | } else { // encountered first time. |
| 339 | // Get the default type of the operand: |
| 340 | // True : IEEE type |
| 341 | // False : QF type |
| 342 | auto defaultPair = QFPSatInstsMap[MI.getOpcode()]; |
| 343 | bool firstOp = (Reg1 == DefReg) ? true : defaultPair.first; |
| 344 | bool secondOp = (Reg2 == DefReg) ? true : defaultPair.second; |
| 345 | QFUsesMap[&MI] = std::make_pair(x&: firstOp, y&: secondOp); |
| 346 | } |
| 347 | } |
| 348 | |
| 349 | unsigned short HexagonPostRAHandleQFP::getreplacedQFOpcode(unsigned srcOpcode, |
| 350 | bool firstOp, |
| 351 | bool secondOp) { |
| 352 | if (firstOp && secondOp) { |
| 353 | switch (srcOpcode) { |
| 354 | case Hexagon::V6_vadd_qf32: |
| 355 | case Hexagon::V6_vadd_qf32_mix: |
| 356 | return Hexagon::V6_vadd_sf; |
| 357 | case Hexagon::V6_vadd_qf16: |
| 358 | case Hexagon::V6_vadd_qf16_mix: |
| 359 | return Hexagon::V6_vadd_hf; |
| 360 | |
| 361 | case Hexagon::V6_vsub_qf32: |
| 362 | case Hexagon::V6_vsub_qf32_mix: |
| 363 | case Hexagon::V6_vsub_sf_mix: |
| 364 | return Hexagon::V6_vsub_sf; |
| 365 | case Hexagon::V6_vsub_qf16: |
| 366 | case Hexagon::V6_vsub_qf16_mix: |
| 367 | case Hexagon::V6_vsub_hf_mix: |
| 368 | return Hexagon::V6_vsub_hf; |
| 369 | |
| 370 | case Hexagon::V6_vmpy_qf32: |
| 371 | return Hexagon::V6_vmpy_qf32_sf; |
| 372 | case Hexagon::V6_vmpy_qf16: |
| 373 | case Hexagon::V6_vmpy_qf16_mix_hf: |
| 374 | return Hexagon::V6_vmpy_qf16_hf; |
| 375 | case Hexagon::V6_vmpy_qf32_qf16: |
| 376 | case Hexagon::V6_vmpy_qf32_mix_hf: |
| 377 | return Hexagon::V6_vmpy_qf32_hf; |
| 378 | |
| 379 | case Hexagon::V6_vmpy_rt_qf16: |
| 380 | return Hexagon::V6_vmpy_rt_hf; |
| 381 | // v81 opcodes start |
| 382 | case Hexagon::V6_vabs_qf32_qf32: |
| 383 | return Hexagon::V6_vabs_qf32_sf; |
| 384 | case Hexagon::V6_vabs_qf16_qf16: |
| 385 | return Hexagon::V6_vabs_qf16_hf; |
| 386 | case Hexagon::V6_vneg_qf32_qf32: |
| 387 | return Hexagon::V6_vneg_qf32_sf; |
| 388 | case Hexagon::V6_vneg_qf16_qf16: |
| 389 | return Hexagon::V6_vneg_qf16_hf; |
| 390 | case Hexagon::V6_vilog2_qf32: |
| 391 | return Hexagon::V6_vilog2_sf; |
| 392 | case Hexagon::V6_vilog2_qf16: |
| 393 | return Hexagon::V6_vilog2_hf; |
| 394 | case Hexagon::V6_vconv_qf32_qf32: |
| 395 | return Hexagon::V6_vconv_qf32_sf; |
| 396 | case Hexagon::V6_vconv_qf16_qf16: |
| 397 | return Hexagon::V6_vconv_qf16_hf; |
| 398 | // v81 opcodes end |
| 399 | |
| 400 | default: |
| 401 | llvm_unreachable("Invalid qf opcode in this scenario!" ); |
| 402 | } |
| 403 | } else if (firstOp) { |
| 404 | switch (srcOpcode) { |
| 405 | case Hexagon::V6_vadd_qf32: |
| 406 | return Hexagon::V6_vadd_qf32_mix; // interchange reqd |
| 407 | case Hexagon::V6_vadd_qf16: |
| 408 | return Hexagon::V6_vadd_qf16_mix; // interchange reqd |
| 409 | |
| 410 | case Hexagon::V6_vsub_qf32: |
| 411 | if (HST->useHVXV81Ops()) |
| 412 | return Hexagon::V6_vsub_sf_mix; |
| 413 | else if (HST->useHVXV79Ops()) |
| 414 | return Hexagon::V6_vsub_sf; // conv reqd |
| 415 | else |
| 416 | llvm_unreachable("Invalid Hexagon Arch for this scenario!" ); |
| 417 | case Hexagon::V6_vsub_qf16: |
| 418 | if (HST->useHVXV81Ops()) |
| 419 | return Hexagon::V6_vsub_hf_mix; |
| 420 | else if (HST->useHVXV79Ops()) |
| 421 | return Hexagon::V6_vsub_hf; // conv reqd |
| 422 | else |
| 423 | llvm_unreachable("Invalid Hexagon Arch for this scenario!" ); |
| 424 | case Hexagon::V6_vsub_qf32_mix: |
| 425 | return Hexagon::V6_vsub_sf; |
| 426 | case Hexagon::V6_vsub_qf16_mix: |
| 427 | return Hexagon::V6_vsub_hf; |
| 428 | |
| 429 | // This opcode does not have a mixed type. Hence if one |
| 430 | // of op1 or op2 is IEEE type and another qf type, |
| 431 | // send the opcode which takes in both as IEEE type. |
| 432 | case Hexagon::V6_vmpy_qf32: |
| 433 | return Hexagon::V6_vmpy_qf32_sf; // conv reqd |
| 434 | case Hexagon::V6_vmpy_qf16: |
| 435 | return Hexagon::V6_vmpy_qf16_mix_hf; // interchange reqd |
| 436 | case Hexagon::V6_vmpy_qf32_qf16: |
| 437 | return Hexagon::V6_vmpy_qf32_mix_hf; // interchange reqd |
| 438 | |
| 439 | default: |
| 440 | return srcOpcode; |
| 441 | } |
| 442 | } else if (secondOp) { |
| 443 | switch (srcOpcode) { |
| 444 | case Hexagon::V6_vadd_qf32: |
| 445 | return Hexagon::V6_vadd_qf32_mix; |
| 446 | case Hexagon::V6_vadd_qf16: |
| 447 | return Hexagon::V6_vadd_qf16_mix; |
| 448 | |
| 449 | case Hexagon::V6_vsub_qf32: |
| 450 | return Hexagon::V6_vsub_qf32_mix; |
| 451 | case Hexagon::V6_vsub_qf16: |
| 452 | return Hexagon::V6_vsub_qf16_mix; |
| 453 | case Hexagon::V6_vsub_sf_mix: |
| 454 | return Hexagon::V6_vsub_sf; |
| 455 | case Hexagon::V6_vsub_hf_mix: |
| 456 | return Hexagon::V6_vsub_hf; |
| 457 | |
| 458 | case Hexagon::V6_vmpy_qf32: |
| 459 | return Hexagon::V6_vmpy_qf32_sf; // conv reqd |
| 460 | |
| 461 | case Hexagon::V6_vmpy_qf16: |
| 462 | return Hexagon::V6_vmpy_qf16_mix_hf; |
| 463 | case Hexagon::V6_vmpy_qf32_qf16: |
| 464 | return Hexagon::V6_vmpy_qf32_mix_hf; |
| 465 | |
| 466 | default: |
| 467 | return srcOpcode; |
| 468 | } |
| 469 | } else |
| 470 | return srcOpcode; |
| 471 | } |
| 472 | |
| 473 | // Insert IEEE to Qf conversion instructions |
| 474 | // is32bit: If true, SrcReg holds sf type, else a hf type |
| 475 | void HexagonPostRAHandleQFP::insertIEEEToQF(MachineInstr *MI, Register SrcReg, |
| 476 | MachineOperand SrcOp, |
| 477 | bool is32bit = false) { |
| 478 | |
| 479 | auto MBB = MI->getParent(); |
| 480 | MachineInstrBuilder MIB; |
| 481 | const DebugLoc &DL = MI->getDebugLoc(); |
| 482 | |
| 483 | if (HST->useHVXV81Ops()) { |
| 484 | auto Op = is32bit ? Hexagon::V6_vconv_qf32_sf : Hexagon::V6_vconv_qf16_hf; |
| 485 | MIB = BuildMI(BB&: *MBB, I&: *MI, MIMD: DL, MCID: HII->get(Opcode: Op), DestReg: SrcReg) |
| 486 | .addReg(RegNo: SrcReg, Flags: RegState::Renamable | RegState::Kill); |
| 487 | LLVM_DEBUG(dbgs() << "\nInserting new instruction: " ; |
| 488 | MIB.getInstr()->dump()); |
| 489 | |
| 490 | } else if (HST->useHVXV79Ops()) { |
| 491 | // Get an available register |
| 492 | auto V0_Reg = findAllocatableReg(MI); |
| 493 | |
| 494 | MIB = BuildMI(BB&: *MBB, I&: *MI, MIMD: DL, MCID: HII->get(Opcode: Hexagon::V6_vd0), DestReg: V0_Reg); |
| 495 | LLVM_DEBUG(dbgs() << "\nInserting new instruction: " ; |
| 496 | MIB.getInstr()->dump()); |
| 497 | auto Op = is32bit ? Hexagon::V6_vadd_sf : Hexagon::V6_vadd_hf; |
| 498 | MIB = BuildMI(BB&: *MBB, I&: *MI, MIMD: DL, MCID: HII->get(Opcode: Op), DestReg: SrcReg) |
| 499 | .addReg(RegNo: SrcReg, Flags: RegState::Renamable | RegState::Kill) |
| 500 | .addReg(RegNo: V0_Reg, Flags: RegState::Kill); |
| 501 | LLVM_DEBUG(dbgs() << "Inserting new instruction: " ; MIB.getInstr()->dump()); |
| 502 | } else |
| 503 | llvm_unreachable("Not possible to insert qf = hf/sf for this unknown\ |
| 504 | subtarget!" ); |
| 505 | } |
| 506 | |
| 507 | // Create a new instruction which handle sf/hf types to replace |
| 508 | // qf type handling instructions. |
| 509 | bool HexagonPostRAHandleQFP::HandleRefills() { |
| 510 | |
| 511 | bool Changed = false; |
| 512 | LLVM_DEBUG(dbgs() << "HandleRefills: " ); |
| 513 | std::vector<MachineInstr *> eraseList; |
| 514 | |
| 515 | for (auto It : QFUsesMap) { |
| 516 | |
| 517 | // Separately handle unary qf opcodes |
| 518 | MachineInstr *MI = It.first; |
| 519 | auto SrcOpcode = MI->getOpcode(); |
| 520 | auto Pair = It.second; |
| 521 | auto SrcOp1 = MI->getOperand(i: 1); |
| 522 | Register DestReg = MI->getOperand(i: 0).getReg(); |
| 523 | auto MBB = MI->getParent(); |
| 524 | MachineInstrBuilder MIB; |
| 525 | LLVM_DEBUG(dbgs() << "\nProcessing: " ; MI->dump()); |
| 526 | const DebugLoc &DL = MI->getDebugLoc(); |
| 527 | |
| 528 | // lambda to handle unary qf operations |
| 529 | // ieee: True if the 1st operand is sf/hf type, false if qf type |
| 530 | auto HandleUnaryRefill = [&](MachineInstr *MI, bool isIeee) -> bool { |
| 531 | if (isIeee) { |
| 532 | auto finalOpcode = getreplacedQFOpcode(srcOpcode: SrcOpcode, firstOp: true, secondOp: true); |
| 533 | MIB = BuildMI(BB&: *MBB, I&: *MI, MIMD: DL, MCID: HII->get(Opcode: finalOpcode), DestReg) |
| 534 | .addReg(RegNo: SrcOp1.getReg(), Flags: getRegState(RegOp: SrcOp1)); |
| 535 | Changed |= true; |
| 536 | LLVM_DEBUG(dbgs() << "\nInserting new instruction: " ; |
| 537 | MIB.getInstr()->dump()); |
| 538 | } else |
| 539 | eraseList.push_back(x: MI); |
| 540 | return Changed; |
| 541 | }; |
| 542 | |
| 543 | if (MI->getNumOperands() == 2) { |
| 544 | Changed |= HandleUnaryRefill(It.first, It.second.first); |
| 545 | continue; |
| 546 | } |
| 547 | auto SrcOp2 = MI->getOperand(i: 2); |
| 548 | |
| 549 | // lambda to handle mixed type vsub instructions for v79 |
| 550 | auto HandleSub = [&](auto srcOpcode) -> bool { |
| 551 | auto ConvOp = (srcOpcode == Hexagon::V6_vsub_qf32) |
| 552 | ? Hexagon::V6_vconv_sf_qf32 |
| 553 | : Hexagon::V6_vconv_hf_qf16; |
| 554 | auto SubOp = (ConvOp == Hexagon::V6_vconv_sf_qf32) ? Hexagon::V6_vsub_sf |
| 555 | : Hexagon::V6_vsub_hf; |
| 556 | |
| 557 | Register SrcOp2Reg = SrcOp2.getReg(); |
| 558 | MIB = BuildMI(*MBB, *MI, DL, HII->get(Opcode: ConvOp), SrcOp2Reg) |
| 559 | .addReg(SrcOp2Reg, getRegState(RegOp: SrcOp2) | RegState::Kill); |
| 560 | LLVM_DEBUG(dbgs() << "\nInserting new instruction: " ; |
| 561 | MIB.getInstr()->dump()); |
| 562 | MIB = BuildMI(*MBB, *MI, DL, HII->get(Opcode: SubOp), DestReg) |
| 563 | .addReg(SrcOp1.getReg(), getRegState(RegOp: SrcOp1)) |
| 564 | .addReg(SrcOp2Reg, getRegState(RegOp: SrcOp2)); |
| 565 | // If Op2 is not killed, it is used after this instruction. |
| 566 | // convert it back to original qf form. |
| 567 | if (!SrcOp2.isKill()) |
| 568 | insertIEEEToQF(MI: &*(++MI->getIterator()), SrcReg: SrcOp2.getReg(), SrcOp: SrcOp2); |
| 569 | return true; |
| 570 | }; |
| 571 | |
| 572 | // If both operands are sf type, we only need to replace the opcode. |
| 573 | if (Pair.first == true && Pair.second == true) { |
| 574 | auto finalOpcode = getreplacedQFOpcode(srcOpcode: SrcOpcode, firstOp: true, secondOp: true); |
| 575 | MIB = BuildMI(BB&: *MBB, I&: *MI, MIMD: DL, MCID: HII->get(Opcode: finalOpcode), DestReg) |
| 576 | .addReg(RegNo: SrcOp1.getReg(), Flags: getRegState(RegOp: SrcOp1)) |
| 577 | .addReg(RegNo: SrcOp2.getReg(), Flags: getRegState(RegOp: SrcOp2)); |
| 578 | Changed |= true; |
| 579 | LLVM_DEBUG(dbgs() << "\nInserting new instruction: " ; |
| 580 | MIB.getInstr()->dump()); |
| 581 | |
| 582 | } else if (Pair.first == true && Pair.second == false) { |
| 583 | auto finalOpcode = getreplacedQFOpcode(srcOpcode: SrcOpcode, firstOp: true, secondOp: false); |
| 584 | |
| 585 | // If 2nd op is qf, first op is sf, convert the 2nd |
| 586 | // op to sf before inserting the vmpy instruction. |
| 587 | if (SrcOpcode == Hexagon::V6_vmpy_qf32) { |
| 588 | Register SrcOp2Reg = SrcOp2.getReg(); |
| 589 | MIB = BuildMI(BB&: *MBB, I&: *MI, MIMD: DL, MCID: HII->get(Opcode: Hexagon::V6_vconv_sf_qf32), |
| 590 | DestReg: SrcOp2Reg) |
| 591 | .addReg(RegNo: SrcOp2Reg, Flags: getRegState(RegOp: SrcOp2) | RegState::Kill); |
| 592 | LLVM_DEBUG(dbgs() << "\nInserting new instruction before: " ; |
| 593 | MIB.getInstr()->dump()); |
| 594 | MIB = BuildMI(BB&: *MBB, I&: *MI, MIMD: DL, MCID: HII->get(Opcode: finalOpcode), DestReg) |
| 595 | .addReg(RegNo: SrcOp1.getReg(), Flags: getRegState(RegOp: SrcOp1)) |
| 596 | .addReg(RegNo: SrcOp2Reg, Flags: getRegState(RegOp: SrcOp2)); |
| 597 | // If Op2 is not killed convert back to qf, since there |
| 598 | // are uses for this qf op. |
| 599 | if (!SrcOp2.isKill()) |
| 600 | insertIEEEToQF(MI: &*(++MI->getIterator()), SrcReg: SrcOp2.getReg(), SrcOp: SrcOp2, |
| 601 | is32bit: true /* sf type reg */); |
| 602 | |
| 603 | // if the opcode is mixed type, we use Op2 as first operand |
| 604 | // since that takes in qf type. Op1 is taken as second op. |
| 605 | } else if (finalOpcode == Hexagon::V6_vadd_qf16_mix || |
| 606 | finalOpcode == Hexagon::V6_vadd_qf32_mix || |
| 607 | finalOpcode == Hexagon::V6_vmpy_qf16_mix_hf || |
| 608 | finalOpcode == Hexagon::V6_vmpy_qf32_mix_hf) { |
| 609 | MIB = BuildMI(BB&: *MBB, I&: *MI, MIMD: DL, MCID: HII->get(Opcode: finalOpcode), DestReg) |
| 610 | .addReg(RegNo: SrcOp2.getReg(), Flags: getRegState(RegOp: SrcOp2)) |
| 611 | .addReg(RegNo: SrcOp1.getReg(), Flags: getRegState(RegOp: SrcOp1)); |
| 612 | |
| 613 | // Subtracting is not associative, so if Op1 is sf/hf type and |
| 614 | // Op2 is qf type, we cannot interchange the operands. |
| 615 | // For v79, we convert Op2 to IEEE and use the non-mix type |
| 616 | // instruction for the subtraction. |
| 617 | // For v81, we have an appropiate opcode with vsub(sf/hf, qf) type |
| 618 | } else if ((SrcOpcode == Hexagon::V6_vsub_qf32 || |
| 619 | SrcOpcode == Hexagon::V6_vsub_qf16) && |
| 620 | HST->useHVXV79Ops()) { |
| 621 | Changed |= HandleSub(SrcOpcode); |
| 622 | |
| 623 | } else { |
| 624 | MIB = BuildMI(BB&: *MBB, I&: *MI, MIMD: DL, MCID: HII->get(Opcode: finalOpcode), DestReg) |
| 625 | .addReg(RegNo: SrcOp1.getReg(), Flags: getRegState(RegOp: SrcOp1)) |
| 626 | .addReg(RegNo: SrcOp2.getReg(), Flags: getRegState(RegOp: SrcOp2)); |
| 627 | } |
| 628 | LLVM_DEBUG(dbgs() << "\nInserting new instruction: " ; |
| 629 | MIB.getInstr()->dump()); |
| 630 | Changed |= true; |
| 631 | } else if (Pair.first == false && Pair.second == true) { |
| 632 | |
| 633 | auto finalOpcode = getreplacedQFOpcode(srcOpcode: SrcOpcode, firstOp: false, secondOp: true); |
| 634 | // If 2nd op is sf, first op is qf, convert the 1st |
| 635 | // op to sf before inserting the vmpy instruction. |
| 636 | if (SrcOpcode == Hexagon::V6_vmpy_qf32) { |
| 637 | Register SrcOp1Reg = SrcOp1.getReg(); |
| 638 | MIB = BuildMI(BB&: *MBB, I&: *MI, MIMD: DL, MCID: HII->get(Opcode: Hexagon::V6_vconv_sf_qf32), |
| 639 | DestReg: SrcOp1Reg) |
| 640 | .addReg(RegNo: SrcOp1Reg, Flags: getRegState(RegOp: SrcOp1) | RegState::Kill); |
| 641 | LLVM_DEBUG(dbgs() << "\nInserting new instruction before: " ; |
| 642 | MIB.getInstr()->dump()); |
| 643 | MIB = BuildMI(BB&: *MBB, I&: *MI, MIMD: DL, MCID: HII->get(Opcode: finalOpcode), DestReg) |
| 644 | .addReg(RegNo: SrcOp1Reg, Flags: getRegState(RegOp: SrcOp1)) |
| 645 | .addReg(RegNo: SrcOp2.getReg(), Flags: getRegState(RegOp: SrcOp2)); |
| 646 | LLVM_DEBUG(dbgs() << "\nInserting new instruction: " ; |
| 647 | MIB.getInstr()->dump()); |
| 648 | // If Op1 is not killed convert back to qf, since there |
| 649 | // are uses for this qf op. |
| 650 | if (!SrcOp1.isKill()) |
| 651 | insertIEEEToQF(MI: &*(++MI->getIterator()), SrcReg: SrcOp1.getReg(), SrcOp: SrcOp1, |
| 652 | is32bit: true /*sf type reg*/); |
| 653 | } else { |
| 654 | |
| 655 | MIB = BuildMI(BB&: *MBB, I&: *MI, MIMD: DL, MCID: HII->get(Opcode: finalOpcode), DestReg) |
| 656 | .addReg(RegNo: SrcOp1.getReg(), Flags: getRegState(RegOp: SrcOp1)) |
| 657 | .addReg(RegNo: SrcOp2.getReg(), Flags: getRegState(RegOp: SrcOp2)); |
| 658 | LLVM_DEBUG(dbgs() << "\nInserting new instruction: " ; |
| 659 | MIB.getInstr()->dump()); |
| 660 | } |
| 661 | Changed |= true; |
| 662 | } else { |
| 663 | // Both the operands of this instructions are valid, so no use of |
| 664 | // this instruction is to be modified. We need to remove this |
| 665 | // instruction from the action map QFUsesMap. |
| 666 | eraseList.push_back(x: MI); |
| 667 | } |
| 668 | } |
| 669 | |
| 670 | for (MachineInstr *delMI : eraseList) |
| 671 | QFUsesMap.erase(x: delMI); |
| 672 | |
| 673 | return Changed; |
| 674 | } |
| 675 | |
| 676 | // Insert a new instruction. |
| 677 | void HexagonPostRAHandleQFP::insertInstr(MachineInstr *MI, unsigned MIOpcode, |
| 678 | unsigned SrcReg, unsigned DstReg, |
| 679 | RegState Flags) { |
| 680 | |
| 681 | MachineInstrBuilder MIB; |
| 682 | MachineBasicBlock *MBB = MI->getParent(); |
| 683 | DebugLoc DL = MI->getDebugLoc(); |
| 684 | MachineBasicBlock::iterator MIt = MI; |
| 685 | auto MINext = ++MI->getIterator(); |
| 686 | if (++MIt == MBB->end()) |
| 687 | MIB = BuildMI(BB: MBB, MIMD: DL, MCID: HII->get(Opcode: MIOpcode), DestReg: DstReg).addReg(RegNo: SrcReg, Flags); |
| 688 | else |
| 689 | MIB = BuildMI(BB&: *MBB, I: MINext, MIMD: DL, MCID: HII->get(Opcode: MIOpcode), DestReg: DstReg) |
| 690 | .addReg(RegNo: SrcReg, Flags); |
| 691 | LLVM_DEBUG(dbgs() << "\t\tInserting after conv: " ; MIB.getInstr()->dump()); |
| 692 | } |
| 693 | |
| 694 | // Find an available vector register to store 0x0. We have reserved vector |
| 695 | // register v30 to be exempted from being used during register allocation |
| 696 | // for this purpose. |
| 697 | MCPhysReg HexagonPostRAHandleQFP::findAllocatableReg(MachineInstr *MI) const { |
| 698 | LLVM_DEBUG(dbgs() << "\tUsing V30 register to store a vector of zeroes!" ); |
| 699 | return Hexagon::V30; |
| 700 | } |
| 701 | |
| 702 | // Insert qf = sf/hf conversions before non-saturating instructions |
| 703 | bool HexagonPostRAHandleQFP::HandleNonSatInstr() { |
| 704 | |
| 705 | for (auto It : QFNonSatMIs) { |
| 706 | MachineInstr *MI = It.first; |
| 707 | auto MIOpcode = MI->getOpcode(); |
| 708 | auto Op = MI->getOperand(i: 1); |
| 709 | Register DefReg = Op.getReg(); |
| 710 | LLVM_DEBUG(dbgs() << "Analyzing convert instruction: " ; MI->dump()); |
| 711 | // Handle hf = qf16. |
| 712 | // Handle f8 = qf16 |
| 713 | if (MIOpcode == Hexagon::V6_vconv_hf_qf16 || |
| 714 | MIOpcode == Hexagon::V6_vconv_f8_qf16) { |
| 715 | |
| 716 | insertIEEEToQF(MI, SrcReg: DefReg, SrcOp: Op); |
| 717 | // TODO Check if there are any reaching def which is qf generating type. |
| 718 | // That op should be converted to sf/hf |
| 719 | if (!Op.isKill()) |
| 720 | insertInstr(MI, MIOpcode: Hexagon::V6_vconv_hf_qf16, SrcReg: DefReg, DstReg: DefReg, |
| 721 | Flags: getRegState(RegOp: Op) | RegState::Kill); |
| 722 | |
| 723 | // Handle hf = qf.qf. |
| 724 | // Handle bf = qf.qf |
| 725 | } else if (MIOpcode == Hexagon::V6_vconv_hf_qf32 || |
| 726 | MIOpcode == Hexagon::V6_vconv_bf_qf32) { |
| 727 | Register DefLo = HRI->getSubReg(Reg: DefReg, Idx: Hexagon::vsub_lo); |
| 728 | Register DefHi = HRI->getSubReg(Reg: DefReg, Idx: Hexagon::vsub_hi); |
| 729 | |
| 730 | if (It.second == ConvOperand::HiLo) { |
| 731 | insertIEEEToQF(MI, SrcReg: DefLo, SrcOp: Op, is32bit: true /* sf type */); |
| 732 | insertIEEEToQF(MI, SrcReg: DefHi, SrcOp: Op, is32bit: true /* sf type */); |
| 733 | |
| 734 | // Check which subregister is live and convert it |
| 735 | // and according insert conversion for that subreg |
| 736 | auto KillState = SubRegKillSet[MI]; |
| 737 | if (!KillState.first) |
| 738 | insertInstr(MI, MIOpcode: Hexagon::V6_vconv_sf_qf32, SrcReg: DefHi, DstReg: DefHi, |
| 739 | Flags: getRegState(RegOp: Op) | RegState::Kill); |
| 740 | |
| 741 | if (!KillState.second) |
| 742 | insertInstr(MI, MIOpcode: Hexagon::V6_vconv_sf_qf32, SrcReg: DefLo, DstReg: DefLo, |
| 743 | Flags: getRegState(RegOp: Op) | RegState::Kill); |
| 744 | |
| 745 | } else if (It.second == ConvOperand::Hi) { |
| 746 | insertIEEEToQF(MI, SrcReg: DefHi, SrcOp: Op, is32bit: true /* sf type */); |
| 747 | if (!Op.isKill()) |
| 748 | insertInstr(MI, MIOpcode: Hexagon::V6_vconv_sf_qf32, SrcReg: DefHi, DstReg: DefHi, |
| 749 | Flags: getRegState(RegOp: Op) | RegState::Kill); |
| 750 | |
| 751 | } else { // It.second == ConvOperand::Lo |
| 752 | insertIEEEToQF(MI, SrcReg: DefLo, SrcOp: Op, is32bit: true /* sf type */); |
| 753 | if (!Op.isKill()) |
| 754 | insertInstr(MI, MIOpcode: Hexagon::V6_vconv_sf_qf32, SrcReg: DefLo, DstReg: DefLo, |
| 755 | Flags: getRegState(RegOp: Op) | RegState::Kill); |
| 756 | } |
| 757 | // Handle sf = qf32. |
| 758 | } else if (MIOpcode == Hexagon::V6_vconv_sf_qf32) { |
| 759 | insertIEEEToQF(MI, SrcReg: DefReg, SrcOp: Op, is32bit: true /* sf type */); |
| 760 | if (!Op.isKill()) |
| 761 | insertInstr(MI, MIOpcode: Hexagon::V6_vconv_sf_qf32, SrcReg: DefReg, DstReg: DefReg, |
| 762 | Flags: getRegState(RegOp: Op) | RegState::Kill); |
| 763 | |
| 764 | } else { |
| 765 | llvm_unreachable("Unhandled non-saturating instruction!" ); |
| 766 | } |
| 767 | } |
| 768 | |
| 769 | if (QFNonSatMIs.empty()) |
| 770 | return false; |
| 771 | return true; |
| 772 | } |
| 773 | |
| 774 | // Calculates the liveness of subregisters (whether killed or not) |
| 775 | // when double register is used. This is necessary because RDF |
| 776 | // carries liveness of the superreg and not the subregisters individually |
| 777 | void HexagonPostRAHandleQFP::collectLivenessForSubregs( |
| 778 | NodeAddr<UseNode *> &UsedNode) { |
| 779 | RegisterRef UR = UsedNode.Addr->getRegRef(G: *DFG); |
| 780 | NodeAddr<StmtNode *> UseStmt = UsedNode.Addr->getOwner(G: *DFG); |
| 781 | MachineInstr *UseInstr = UseStmt.Addr->getCode(); |
| 782 | auto UseOp = UseInstr->getOperand(i: 1); |
| 783 | Register UseDefLo = HRI->getSubReg(Reg: UseOp.getReg(), Idx: Hexagon::vsub_lo); |
| 784 | Register UseDefHi = HRI->getSubReg(Reg: UseOp.getReg(), Idx: Hexagon::vsub_hi); |
| 785 | |
| 786 | NodeSet Visited, Defs; |
| 787 | bool isHiSubRegKilled = true, isLoSubRegKilled = true; |
| 788 | const auto &P = LV->getAllReachingDefsRec(RefRR: UR, RefA: UsedNode, Visited, Defs); |
| 789 | |
| 790 | if (!P.second) |
| 791 | return; |
| 792 | |
| 793 | for (auto RD : P.first) { |
| 794 | NodeAddr<DefNode *> RegDef = DFG->addr<DefNode *>(N: RD); |
| 795 | Register RR = RegDef.Addr->getRegRef(G: *DFG).Id; |
| 796 | if (HRI->isFakeReg(Reg: RR)) |
| 797 | continue; |
| 798 | NodeAddr<StmtNode *> RegStmt = RegDef.Addr->getOwner(G: *DFG); |
| 799 | MachineInstr *ReachDefInstr = RegStmt.Addr->getCode(); |
| 800 | if (ReachDefInstr == nullptr) |
| 801 | continue; |
| 802 | |
| 803 | // If the reaching def is WReg, then the kill flag in the use is correct |
| 804 | // since there is no subreg |
| 805 | Register DefReg = ReachDefInstr->getOperand(i: 0).getReg(); |
| 806 | if (Hexagon::HvxWRRegClass.contains(Reg: DefReg)) { |
| 807 | if (!UseOp.isKill()) |
| 808 | isHiSubRegKilled = isLoSubRegKilled = false; |
| 809 | |
| 810 | // If the reaching ref is VReg, the liveness might be different between |
| 811 | // each of the subreg. Handle them individually. |
| 812 | // Find the other uses after this use for the reaching def. If it exists, |
| 813 | // the subregister is live after the use. |
| 814 | // NOTE: Assumption: The uses are in order in RDF. |
| 815 | } else { |
| 816 | NodeSet UseSet; |
| 817 | getAllRealUses(DA: RegDef, UNodeSet&: UseSet, L: LV, G: DFG); |
| 818 | for (auto UIntr : UseSet) { |
| 819 | NodeAddr<UseNode *> UA = DFG->addr<UseNode *>(N: UIntr); |
| 820 | NodeAddr<StmtNode *> UseStmt = UA.Addr->getOwner(G: *DFG); |
| 821 | MachineInstr *UseMI = UseStmt.Addr->getCode(); |
| 822 | if (UseMI == nullptr) |
| 823 | continue; |
| 824 | // When we reach the use set a flag to see if there are other uses |
| 825 | // after this. If yes, then the register is not killed. |
| 826 | if (UseMI == UseInstr) |
| 827 | continue; |
| 828 | if (HII->isMIBefore(A: UseInstr, B: UseMI) && DefReg == UseDefLo) { |
| 829 | isLoSubRegKilled = false; |
| 830 | break; |
| 831 | } |
| 832 | if (HII->isMIBefore(A: UseInstr, B: UseMI) && DefReg == UseDefHi) { |
| 833 | isHiSubRegKilled = false; |
| 834 | break; |
| 835 | } |
| 836 | } |
| 837 | } |
| 838 | } |
| 839 | SubRegKillSet[UseInstr] = std::make_pair(x&: isHiSubRegKilled, y&: isLoSubRegKilled); |
| 840 | } |
| 841 | |
| 842 | // Store all refill instructions. |
| 843 | void HexagonPostRAHandleQFP::collectQFPStackRefill( |
| 844 | NodeAddr<StmtNode *> *StNode) { |
| 845 | NodeAddr<DefNode *> DfNode = |
| 846 | StNode->Addr->members_if(P: DFG->IsDef, G: *DFG).front(); |
| 847 | MachineInstr *MI = StNode->Addr->getCode(); |
| 848 | // Check if operand to this instruction is a frame index. |
| 849 | const MachineOperand &OpFI = MI->getOperand(i: 1); |
| 850 | if (!OpFI.isFI()) |
| 851 | return; |
| 852 | |
| 853 | // LLVM_DEBUG(dbgs() << "\n[Stack Refill]: Collecting: "; MI->dump()); |
| 854 | RefillMIs.push_back(x: DfNode); |
| 855 | } |
| 856 | |
| 857 | // Iterate over the uses of the qf generating instruction in RDG graph |
| 858 | // If we get a qf to IEEE convert instruction, add it to a list. |
| 859 | void HexagonPostRAHandleQFP::collectConvQFInstr(NodeAddr<DefNode *> &RegDef) { |
| 860 | |
| 861 | NodeSet UseSet; |
| 862 | NodeAddr<StmtNode *> DefStmt = RegDef.Addr->getOwner(G: *DFG); |
| 863 | MachineInstr *DefInstr = DefStmt.Addr->getCode(); |
| 864 | getAllRealUses(DA: RegDef, UNodeSet&: UseSet, L: LV, G: DFG); |
| 865 | for (auto UI : UseSet) { |
| 866 | NodeAddr<UseNode *> UA = DFG->addr<UseNode *>(N: UI); |
| 867 | if (UA.Addr->getFlags() & NodeAttrs::PhiRef) |
| 868 | continue; |
| 869 | NodeAddr<StmtNode *> UseStmt = UA.Addr->getOwner(G: *DFG); |
| 870 | MachineInstr *QFConvInstr = UseStmt.Addr->getCode(); |
| 871 | if (std::find(first: QFNonSatInstr.begin(), last: QFNonSatInstr.end(), |
| 872 | val: QFConvInstr->getOpcode()) != QFNonSatInstr.end()) { |
| 873 | |
| 874 | // The use is a double register type. But the def can be hi/lo or double |
| 875 | // type. So conversion needs to be inserted only for the type |
| 876 | // which is in IEEE form. |
| 877 | auto UseReg = QFConvInstr->getOperand(i: 1).getReg(); |
| 878 | auto DefReg = DefInstr->getOperand(i: 0).getReg(); |
| 879 | if (Hexagon::HvxWRRegClass.contains(Reg: UseReg)) { |
| 880 | |
| 881 | collectLivenessForSubregs(UsedNode&: UA); |
| 882 | unsigned Op = ConvOperand::Undefined; |
| 883 | if (QFNonSatMIs.contains(Key: QFConvInstr)) |
| 884 | Op = QFNonSatMIs[QFConvInstr]; |
| 885 | |
| 886 | // Def is double type |
| 887 | if (Hexagon::HvxWRRegClass.contains(Reg: DefReg)) |
| 888 | Op = ConvOperand::HiLo; |
| 889 | // Def is lo of double type |
| 890 | else if (DefReg == HRI->getSubReg(Reg: UseReg, Idx: Hexagon::vsub_lo)) |
| 891 | Op |= ConvOperand::Lo; |
| 892 | // Def is hi of double type |
| 893 | else |
| 894 | Op |= ConvOperand::Hi; |
| 895 | QFNonSatMIs[QFConvInstr] = Op; |
| 896 | } else // for other def-use, BothOp is used as default |
| 897 | QFNonSatMIs[QFConvInstr] = ConvOperand::HiLo; |
| 898 | |
| 899 | IgnoreInsertConvList.insert(Ptr: DefInstr); |
| 900 | LLVM_DEBUG(std::string OpType = "" ; switch (QFNonSatMIs[QFConvInstr]) { |
| 901 | case ConvOperand::HiLo: |
| 902 | OpType = "HiLo Op" ; |
| 903 | break; |
| 904 | case ConvOperand::Lo: |
| 905 | OpType = "Lo Op" ; |
| 906 | break; |
| 907 | case ConvOperand::Hi: |
| 908 | OpType = "Hi Op" ; |
| 909 | break; |
| 910 | default: |
| 911 | OpType = "Undefined" ; |
| 912 | } dbgs() << "Collecting convert instruction with type " |
| 913 | << OpType << " : " ; |
| 914 | QFConvInstr->dump()); |
| 915 | } |
| 916 | } |
| 917 | } |
| 918 | |
| 919 | // Check if the COPY statements use came from a def which generates |
| 920 | // a qf type. If yes, collect it in a vector. Also, collect copies |
| 921 | // with reaching def other copies (nested copies). |
| 922 | void HexagonPostRAHandleQFP::collectCopies(NodeAddr<StmtNode *> *StNode) { |
| 923 | |
| 924 | NodeAddr<DefNode *> CopyDef = |
| 925 | StNode->Addr->members_if(P: DFG->IsDef, G: *DFG).front(); |
| 926 | MachineInstr *CopyInstr = StNode->Addr->getCode(); |
| 927 | LLVM_DEBUG(dbgs() << "\nAnalyzing copy: " ; StNode->Addr->getCode()->dump()); |
| 928 | |
| 929 | for (NodeAddr<UseNode *> UA : StNode->Addr->members_if(P: DFG->IsUse, G: *DFG)) { |
| 930 | RegisterRef UR = UA.Addr->getRegRef(G: *DFG); |
| 931 | NodeSet Visited, Defs; |
| 932 | const auto &P = LV->getAllReachingDefsRec(RefRR: UR, RefA: UA, Visited, Defs); |
| 933 | if (!P.second) { |
| 934 | LLVM_DEBUG({ |
| 935 | dbgs() << "*** Unable to collect all reaching defs for use ***\n" |
| 936 | << PrintNode<UseNode *>(UA, *DFG) << '\n'; |
| 937 | }); |
| 938 | continue; |
| 939 | } |
| 940 | |
| 941 | // Note: there can be multiple reaching defs of the copy |
| 942 | for (auto RD : P.first) { |
| 943 | NodeAddr<DefNode *> RegDef = DFG->addr<DefNode *>(N: RD); |
| 944 | Register RR = RegDef.Addr->getRegRef(G: *DFG).Id; |
| 945 | if (HRI->isFakeReg(Reg: RR)) |
| 946 | continue; |
| 947 | NodeAddr<StmtNode *> RegStmt = RegDef.Addr->getOwner(G: *DFG); |
| 948 | MachineInstr *ReachDefInstr = RegStmt.Addr->getCode(); |
| 949 | if (ReachDefInstr == nullptr) |
| 950 | continue; |
| 951 | LLVM_DEBUG(dbgs() << "\t[Reaching Def]: " ; ReachDefInstr->dump()); |
| 952 | |
| 953 | // If the reaching def is a COPY,collect it with reg type ieee |
| 954 | if (ReachDefInstr->getOpcode() == TargetOpcode::COPY) { |
| 955 | auto pairKey = std::make_pair(x&: CopyDef.Id, y&: RegDef.Id); |
| 956 | QFCopys[pairKey] = RegType::ieee; |
| 957 | continue; |
| 958 | } |
| 959 | |
| 960 | // If the reaching def is a qf instr, collect the copy. |
| 961 | // reg type is selected based on the op |
| 962 | auto RegT = RegType::undefined; |
| 963 | if (HII->isQFPInstr(MI: ReachDefInstr)) { |
| 964 | if (HII->isQFP32Instr(MI: ReachDefInstr)) { |
| 965 | // check whether the copies register is hvxWR or hvxVR type |
| 966 | // NOTE: Assumption: A copy's reaching def shall not be 2, |
| 967 | // i.e., for each of the subregister. |
| 968 | if (Hexagon::HvxWRRegClass.contains( |
| 969 | Reg: ReachDefInstr->getOperand(i: 0).getReg())) |
| 970 | RegT = RegType::qf32_double; |
| 971 | else |
| 972 | RegT = RegType::qf32; |
| 973 | } else if (HII->isQFP16Instr(MI: ReachDefInstr)) { |
| 974 | // Check if qf16 instruction outputs double-wide register |
| 975 | if (Hexagon::HvxWRRegClass.contains( |
| 976 | Reg: ReachDefInstr->getOperand(i: 0).getReg())) { |
| 977 | RegT = RegType::qf16_double; |
| 978 | } else { |
| 979 | RegT = RegType::qf16; |
| 980 | } |
| 981 | } |
| 982 | } else { |
| 983 | // if the copy involves non-qf vector registers collect it too |
| 984 | Register CopyReg = CopyInstr->getOperand(i: 1).getReg(); |
| 985 | if (Hexagon::HvxWRRegClass.contains(Reg: CopyReg) || |
| 986 | Hexagon::HvxVRRegClass.contains(Reg: CopyReg)) |
| 987 | RegT = RegType::ieee; |
| 988 | else |
| 989 | continue; |
| 990 | } |
| 991 | auto pairKey = std::make_pair(x&: CopyDef.Id, y&: RegDef.Id); |
| 992 | QFCopys[pairKey] = RegT; |
| 993 | } |
| 994 | } |
| 995 | } |
| 996 | |
| 997 | // Inserts an qf instruction to a list. These instruction |
| 998 | // values are spilled to the stack. |
| 999 | void HexagonPostRAHandleQFP::collectQFPStackSpill( |
| 1000 | NodeAddr<StmtNode *> *StNode) { |
| 1001 | |
| 1002 | MachineInstr *MI = StNode->Addr->getCode(); |
| 1003 | LLVM_DEBUG(dbgs() << "\n[Stack Spill]: Analyzing: " ; MI->dump()); |
| 1004 | // Check if operand to this instruction is a frame index. |
| 1005 | const MachineOperand &OpFI = MI->getOperand(i: 0); |
| 1006 | if (!OpFI.isFI()) |
| 1007 | return; |
| 1008 | |
| 1009 | // Pre-RegAlloc |
| 1010 | //%46:hvxwr = V6_vmpy_qf32_hf %7:hvxvr, %10:hvxvr |
| 1011 | // PS_vstorerw_ai %stack.3, 0, %46:hvxwr :: (store (s2048) into %stack.3, |
| 1012 | // align 128) |
| 1013 | // Post-RegAlloc |
| 1014 | // renamable $w4 = V6_vmpy_qf32_hf killed renamable $v1, renamable $v0 |
| 1015 | // PS_vstorerw_ai %stack.3, 0, renamable $w4 :: (store (s2048) into %stack.3, |
| 1016 | // align 128) |
| 1017 | |
| 1018 | if (!MI->getOperand(i: 2).isReg()) |
| 1019 | return; |
| 1020 | |
| 1021 | // Iterate over the operands of the store instruction to get their reaching |
| 1022 | // defs |
| 1023 | NodeId QFPDefNode = 0; |
| 1024 | for (NodeAddr<UseNode *> UA : StNode->Addr->members_if(P: DFG->IsUse, G: *DFG)) { |
| 1025 | QFPDefNode = UA.Addr->getReachingDef(); |
| 1026 | |
| 1027 | // Get the defining instruction node(s) |
| 1028 | NodeAddr<DefNode *> RegDef = DFG->addr<DefNode *>(N: QFPDefNode); |
| 1029 | assert(QFPDefNode != 0 && "Reaching def computation error" ); |
| 1030 | NodeAddr<StmtNode *> RegStmt = RegDef.Addr->getOwner(G: *DFG); |
| 1031 | MachineInstr *ReachDefInstr = RegStmt.Addr->getCode(); |
| 1032 | if (ReachDefInstr == nullptr) |
| 1033 | continue; |
| 1034 | LLVM_DEBUG(dbgs() << "[Stack Spill]:\tReaching Def of operand:" ; |
| 1035 | ReachDefInstr->dump()); |
| 1036 | // Reaching Def cannot be a phi instruction. |
| 1037 | if (RegDef.Addr->getFlags() & NodeAttrs::PhiRef) |
| 1038 | continue; |
| 1039 | |
| 1040 | if (!HII->isQFPInstr(MI: ReachDefInstr)) |
| 1041 | continue; |
| 1042 | |
| 1043 | auto RR = RegDef.Addr->getRegRef(G: *DFG).Id; |
| 1044 | if (HRI->isFakeReg(Reg: RR)) |
| 1045 | continue; |
| 1046 | |
| 1047 | LLVM_DEBUG(dbgs() << "Found a QFPStackSpill via \n" ; MI->dump(); |
| 1048 | dbgs() << "The corresponding XQF instruction is:\n" ; |
| 1049 | ReachDefInstr->dump()); |
| 1050 | |
| 1051 | // Collect the spills. |
| 1052 | SpillMIs.push_back(x: std::make_pair(x&: MI, y&: RegDef)); |
| 1053 | } |
| 1054 | } |
| 1055 | |
| 1056 | // Find the uses of qf generating instructions and conditionally add them |
| 1057 | // to a list. |
| 1058 | void HexagonPostRAHandleQFP::collectQFUses(NodeAddr<DefNode *> RegDef, |
| 1059 | MachineInstr *DefMI) { |
| 1060 | |
| 1061 | NodeSet UseSet; |
| 1062 | LLVM_DEBUG(dbgs() << " Finding uses of: " ; DefMI->dump();); |
| 1063 | getAllRealUses(DA: RegDef, UNodeSet&: UseSet, L: LV, G: DFG); |
| 1064 | |
| 1065 | for (auto UI : UseSet) { |
| 1066 | NodeAddr<UseNode *> UA = DFG->addr<UseNode *>(N: UI); |
| 1067 | if (UA.Addr->getFlags() & NodeAttrs::PhiRef) |
| 1068 | continue; |
| 1069 | NodeAddr<StmtNode *> UseStmt = UA.Addr->getOwner(G: *DFG); |
| 1070 | MachineInstr *UseMI = UseStmt.Addr->getCode(); |
| 1071 | LLVM_DEBUG(dbgs() << "\t\t\t[Reached Use of QF operand]: " ; UseMI->dump()); |
| 1072 | |
| 1073 | Register UsedReg = UA.Addr->getRegRef(G: *DFG).Id; |
| 1074 | if (QFPSatInstsMap.find(Val: UseMI->getOpcode()) != QFPSatInstsMap.end()) { |
| 1075 | if (PossibleMultiReachDefs.count(x: UseStmt.Id) == 0) { |
| 1076 | PossibleMultiReachDefs.insert(x: UseStmt.Id); |
| 1077 | LLVM_DEBUG(dbgs() << "\n[Collect instr with possible multidef]:" ; |
| 1078 | UseMI->dump()); |
| 1079 | } |
| 1080 | conditionallyInsert(MI&: *UseMI, DefReg&: UsedReg); |
| 1081 | } |
| 1082 | } |
| 1083 | } |
| 1084 | |
| 1085 | // Process the list which can have multiple definitions. A possible case |
| 1086 | // can be reaching defs to be a copy and a qf-generating instr respectively. |
| 1087 | // Only handle the qf-generating instruction by inserting convert to sf/hf |
| 1088 | // after it. Additionally, then handle the reached uses of this reaching |
| 1089 | // def since the type has changed to sf/hf from qf after the conversion. |
| 1090 | bool HexagonPostRAHandleQFP::HandleMultiReachingDefs() { |
| 1091 | |
| 1092 | bool Changed = false; |
| 1093 | // Note: It may seem this loop can further add to PossibleMultiReachDefs. |
| 1094 | // But it is not expected to since if any instruction has multiple |
| 1095 | // definitions it should already be present in it. |
| 1096 | for (auto It : PossibleMultiReachDefs) { |
| 1097 | NodeAddr<StmtNode *> Stmt = DFG->addr<StmtNode *>(N: It); |
| 1098 | MachineInstr *Instr = Stmt.Addr->getCode(); |
| 1099 | // get the op type for the original instruction. |
| 1100 | // True is sf/hf, false is qf |
| 1101 | auto Pair = QFUsesMap[Instr]; |
| 1102 | |
| 1103 | unsigned short UseNo = 1; |
| 1104 | // Iterate over the operands |
| 1105 | for (NodeAddr<UseNode *> UA : Stmt.Addr->members_if(P: DFG->IsUse, G: *DFG)) { |
| 1106 | |
| 1107 | // If the type is qf for the operand, |
| 1108 | // we skip since there is no scope for mismatch |
| 1109 | if ((UseNo == 1 && Pair.first == false) || |
| 1110 | (UseNo == 2 && Pair.second == false)) { |
| 1111 | ++UseNo; |
| 1112 | continue; |
| 1113 | } |
| 1114 | |
| 1115 | RegisterRef UR = UA.Addr->getRegRef(G: *DFG); |
| 1116 | NodeSet Visited, Defs; |
| 1117 | const auto &P = LV->getAllReachingDefsRec(RefRR: UR, RefA: UA, Visited, Defs); |
| 1118 | if (!P.second) { |
| 1119 | LLVM_DEBUG({ |
| 1120 | dbgs() << "*** Unable to collect all reaching defs for use ***\n" |
| 1121 | << PrintNode<UseNode *>(UA, *DFG) << '\n'; |
| 1122 | }); |
| 1123 | continue; |
| 1124 | } |
| 1125 | |
| 1126 | // Iterate over the reaching defs and process the ones which |
| 1127 | // generate qf. Ignore the ones which have already been handled |
| 1128 | for (auto RD : P.first) { |
| 1129 | NodeAddr<DefNode *> RegDef = DFG->addr<DefNode *>(N: RD); |
| 1130 | |
| 1131 | // Ignore fake reaches |
| 1132 | auto RR = RegDef.Addr->getRegRef(G: *DFG).Id; |
| 1133 | if (HRI->isFakeReg(Reg: RR)) |
| 1134 | continue; |
| 1135 | |
| 1136 | NodeAddr<StmtNode *> RegStmt = RegDef.Addr->getOwner(G: *DFG); |
| 1137 | MachineInstr *ReachDefInstr = RegStmt.Addr->getCode(); |
| 1138 | |
| 1139 | if (ReachDefInstr == nullptr) |
| 1140 | continue; |
| 1141 | |
| 1142 | if (!HII->isQFPInstr(MI: ReachDefInstr)) |
| 1143 | continue; |
| 1144 | if (IgnoreInsertConvList.find(Ptr: ReachDefInstr) != |
| 1145 | IgnoreInsertConvList.end()) |
| 1146 | continue; |
| 1147 | LLVM_DEBUG(dbgs() << "[Multidef] Handling reaching def:" ; |
| 1148 | ReachDefInstr->dump()); |
| 1149 | |
| 1150 | auto *MBB = ReachDefInstr->getParent(); |
| 1151 | auto &dl = ReachDefInstr->getDebugLoc(); |
| 1152 | auto NextReachMI = ++ReachDefInstr->getIterator(); |
| 1153 | auto DefOp = ReachDefInstr->getOperand(i: 0); |
| 1154 | Register OpReg = DefOp.getReg(); |
| 1155 | MachineInstrBuilder MIB; |
| 1156 | |
| 1157 | // For double vector regs, two conversions are inserted. Single |
| 1158 | // conversion for qf32 type |
| 1159 | if (HII->isQFP32Instr(MI: ReachDefInstr)) { |
| 1160 | // if the reaching def is a qf double type |
| 1161 | if (Hexagon::HvxWRRegClass.contains( |
| 1162 | Reg: ReachDefInstr->getOperand(i: 0).getReg())) { |
| 1163 | Register RegLo = HRI->getSubReg(Reg: OpReg, Idx: Hexagon::vsub_lo); |
| 1164 | Register RegHi = HRI->getSubReg(Reg: OpReg, Idx: Hexagon::vsub_hi); |
| 1165 | MIB = BuildMI(BB&: *MBB, I: NextReachMI, MIMD: dl, |
| 1166 | MCID: HII->get(Opcode: Hexagon::V6_vconv_sf_qf32), DestReg: RegLo) |
| 1167 | .addReg(RegNo: RegLo, Flags: RegState::Renamable | RegState::Kill); |
| 1168 | LLVM_DEBUG(dbgs() << "[MultiDef] Inserting convert instruction: " ; |
| 1169 | MIB.getInstr()->dump()); |
| 1170 | MIB = BuildMI(BB&: *MBB, I: NextReachMI, MIMD: dl, |
| 1171 | MCID: HII->get(Opcode: Hexagon::V6_vconv_sf_qf32), DestReg: RegHi) |
| 1172 | .addReg(RegNo: RegHi, Flags: RegState::Renamable | RegState::Kill); |
| 1173 | } else { // If the reaching def is a qf type |
| 1174 | MIB = BuildMI(BB&: *MBB, I: NextReachMI, MIMD: dl, |
| 1175 | MCID: HII->get(Opcode: Hexagon::V6_vconv_sf_qf32), DestReg: OpReg) |
| 1176 | .addReg(RegNo: OpReg, Flags: RegState::Renamable | RegState::Kill); |
| 1177 | } |
| 1178 | } |
| 1179 | if (HII->isQFP16Instr(MI: ReachDefInstr)) { |
| 1180 | MIB = BuildMI(BB&: *MBB, I: NextReachMI, MIMD: dl, |
| 1181 | MCID: HII->get(Opcode: Hexagon::V6_vconv_hf_qf16), DestReg: OpReg) |
| 1182 | .addReg(RegNo: OpReg, Flags: RegState::Renamable | RegState::Kill); |
| 1183 | } |
| 1184 | LLVM_DEBUG(dbgs() << "[MultiDef] Inserting convert instruction: " ; |
| 1185 | MIB.getInstr()->dump(); dbgs() << "\tafter instruction: " ; |
| 1186 | ReachDefInstr->dump()); |
| 1187 | |
| 1188 | // find the uses of the newly transformed to sf/hf and handle |
| 1189 | // accordingly. Uses can be vmul/vadd/etc. types or converts which take |
| 1190 | // in qf types. |
| 1191 | collectQFUses(RegDef, DefMI: ReachDefInstr); |
| 1192 | collectConvQFInstr(RegDef); |
| 1193 | IgnoreInsertConvList.insert(Ptr: ReachDefInstr); |
| 1194 | Changed = true; |
| 1195 | } |
| 1196 | UseNo++; |
| 1197 | } |
| 1198 | } |
| 1199 | return Changed; |
| 1200 | } |
| 1201 | |
| 1202 | bool HexagonPostRAHandleQFP::HandleConvertToQfCopies() { |
| 1203 | if (ConvertToQfCopies.empty()) |
| 1204 | return false; |
| 1205 | |
| 1206 | LLVM_DEBUG( |
| 1207 | dbgs() << "\n*** Inserting convert to qf for selected copies ***\n" ); |
| 1208 | |
| 1209 | // Any reached use of the copy should not already be collected to be |
| 1210 | // converted to IEEE. If present, it means that the reached use has |
| 1211 | // other reaching def with type IEEE, other than this copy. |
| 1212 | auto CanTransform = [&](MachineInstr *MI, unsigned OpNo) -> bool { |
| 1213 | if (QFUsesMap.find(x: MI) != QFUsesMap.end()) { |
| 1214 | auto Entry = QFUsesMap[MI]; |
| 1215 | if (OpNo == 1 && Entry.first == true) |
| 1216 | return false; |
| 1217 | if (OpNo == 2 && Entry.second == true) |
| 1218 | return false; |
| 1219 | } |
| 1220 | return true; |
| 1221 | }; |
| 1222 | |
| 1223 | for (auto It : ConvertToQfCopies) { |
| 1224 | NodeSet UseSet; |
| 1225 | getAllRealUses(DA: It.second.first, UNodeSet&: UseSet, L: LV, G: DFG); |
| 1226 | |
| 1227 | bool transform = true; |
| 1228 | for (auto UI : UseSet) { |
| 1229 | NodeAddr<UseNode *> UA = DFG->addr<UseNode *>(N: UI); |
| 1230 | if (UA.Addr->getFlags() & NodeAttrs::PhiRef) |
| 1231 | continue; |
| 1232 | NodeAddr<StmtNode *> UseStmt = UA.Addr->getOwner(G: *DFG); |
| 1233 | MachineInstr *UseMI = UseStmt.Addr->getCode(); |
| 1234 | unsigned OpNo = UA.Addr->getOp().getOperandNo(); |
| 1235 | |
| 1236 | if (!CanTransform(UseMI, OpNo)) { |
| 1237 | transform = false; |
| 1238 | break; |
| 1239 | } |
| 1240 | } |
| 1241 | |
| 1242 | if (transform) { |
| 1243 | |
| 1244 | LLVM_DEBUG(dbgs() << "\n[HandleConvertToQfCopies]\tProcessing Copy:" ; |
| 1245 | It.first->dump()); |
| 1246 | auto CopyOp = It.first->getOperand(i: 0); |
| 1247 | auto NextMIIter = std::next(x: It.first->getIterator()); |
| 1248 | switch (It.second.second) { |
| 1249 | case RegType::qf32_double: { |
| 1250 | Register DefLo = HRI->getSubReg(Reg: CopyOp.getReg(), Idx: Hexagon::vsub_lo); |
| 1251 | Register DefHi = HRI->getSubReg(Reg: CopyOp.getReg(), Idx: Hexagon::vsub_hi); |
| 1252 | insertIEEEToQF(MI: &*NextMIIter, SrcReg: DefLo, SrcOp: CopyOp, /*is32bit=*/true); |
| 1253 | insertIEEEToQF(MI: &*NextMIIter, SrcReg: DefHi, SrcOp: CopyOp, /*is32bit=*/true); |
| 1254 | break; |
| 1255 | } |
| 1256 | case RegType::qf16_double: { |
| 1257 | Register DefLo = HRI->getSubReg(Reg: CopyOp.getReg(), Idx: Hexagon::vsub_lo); |
| 1258 | Register DefHi = HRI->getSubReg(Reg: CopyOp.getReg(), Idx: Hexagon::vsub_hi); |
| 1259 | insertIEEEToQF(MI: &*NextMIIter, SrcReg: DefLo, SrcOp: CopyOp, /*is32bit=*/false); |
| 1260 | insertIEEEToQF(MI: &*NextMIIter, SrcReg: DefHi, SrcOp: CopyOp, /*is32bit=*/false); |
| 1261 | break; |
| 1262 | } |
| 1263 | case RegType::qf16: |
| 1264 | insertIEEEToQF(MI: &*NextMIIter, SrcReg: CopyOp.getReg(), SrcOp: CopyOp, |
| 1265 | /*is32bit=*/false); |
| 1266 | break; |
| 1267 | case RegType::qf32: |
| 1268 | insertIEEEToQF(MI: &*NextMIIter, SrcReg: CopyOp.getReg(), SrcOp: CopyOp, /*is32bit=*/true); |
| 1269 | break; |
| 1270 | default: |
| 1271 | break; |
| 1272 | } |
| 1273 | } else { |
| 1274 | collectQFUses(RegDef: It.second.first, DefMI: It.first); |
| 1275 | collectConvQFInstr(RegDef&: It.second.first); |
| 1276 | } |
| 1277 | } |
| 1278 | return true; |
| 1279 | } |
| 1280 | |
| 1281 | bool HexagonPostRAHandleQFP::HandleReachDefOfCopies() { |
| 1282 | if (ReachDefOfCopies.empty()) |
| 1283 | return false; |
| 1284 | |
| 1285 | MachineInstrBuilder MIB; |
| 1286 | for (auto It : ReachDefOfCopies) { |
| 1287 | auto *MBB = It.first->getParent(); |
| 1288 | auto &dl = It.first->getDebugLoc(); |
| 1289 | auto NextMI = ++(It.first)->getIterator(); |
| 1290 | auto RegOp = It.first->getOperand(i: 0); |
| 1291 | Register OpReg = RegOp.getReg(); |
| 1292 | |
| 1293 | if (It.second == RegType::qf32) |
| 1294 | MIB = |
| 1295 | BuildMI(BB&: *MBB, I: NextMI, MIMD: dl, MCID: HII->get(Opcode: Hexagon::V6_vconv_sf_qf32), DestReg: OpReg) |
| 1296 | .addReg(RegNo: OpReg, Flags: RegState::Renamable | RegState::Kill); |
| 1297 | else if (It.second == RegType::qf16) |
| 1298 | MIB = |
| 1299 | BuildMI(BB&: *MBB, I: NextMI, MIMD: dl, MCID: HII->get(Opcode: Hexagon::V6_vconv_hf_qf16), DestReg: OpReg) |
| 1300 | .addReg(RegNo: OpReg, Flags: RegState::Renamable | RegState::Kill); |
| 1301 | else if (It.second == RegType::qf32_double) { |
| 1302 | Register RegLo = HRI->getSubReg(Reg: OpReg, Idx: Hexagon::vsub_lo); |
| 1303 | Register RegHi = HRI->getSubReg(Reg: OpReg, Idx: Hexagon::vsub_hi); |
| 1304 | MIB = |
| 1305 | BuildMI(BB&: *MBB, I: NextMI, MIMD: dl, MCID: HII->get(Opcode: Hexagon::V6_vconv_sf_qf32), DestReg: RegLo) |
| 1306 | .addReg(RegNo: RegLo, Flags: RegState::Renamable | RegState::Kill); |
| 1307 | LLVM_DEBUG(dbgs() << "Inserting convert instruction: " ; |
| 1308 | MIB.getInstr()->dump()); |
| 1309 | MIB = |
| 1310 | BuildMI(BB&: *MBB, I: NextMI, MIMD: dl, MCID: HII->get(Opcode: Hexagon::V6_vconv_sf_qf32), DestReg: RegHi) |
| 1311 | .addReg(RegNo: RegHi, Flags: RegState::Renamable | RegState::Kill); |
| 1312 | } else if (It.second == RegType::qf16_double) { |
| 1313 | Register RegLo = HRI->getSubReg(Reg: OpReg, Idx: Hexagon::vsub_lo); |
| 1314 | Register RegHi = HRI->getSubReg(Reg: OpReg, Idx: Hexagon::vsub_hi); |
| 1315 | MIB = |
| 1316 | BuildMI(BB&: *MBB, I: NextMI, MIMD: dl, MCID: HII->get(Opcode: Hexagon::V6_vconv_hf_qf16), DestReg: RegLo) |
| 1317 | .addReg(RegNo: RegLo, Flags: RegState::Renamable | RegState::Kill); |
| 1318 | LLVM_DEBUG(dbgs() << "Inserting convert instruction: " ; |
| 1319 | MIB.getInstr()->dump()); |
| 1320 | MIB = |
| 1321 | BuildMI(BB&: *MBB, I: NextMI, MIMD: dl, MCID: HII->get(Opcode: Hexagon::V6_vconv_hf_qf16), DestReg: RegHi) |
| 1322 | .addReg(RegNo: RegHi, Flags: RegState::Renamable | RegState::Kill); |
| 1323 | } |
| 1324 | LLVM_DEBUG(dbgs() << "Inserting convert instruction: " ; |
| 1325 | MIB.getInstr()->dump(); dbgs() << "\tafter instruction: " ; |
| 1326 | It.first->dump()); |
| 1327 | } |
| 1328 | return true; |
| 1329 | } |
| 1330 | |
| 1331 | HexagonPostRAHandleQFP::RegType |
| 1332 | HexagonPostRAHandleQFP::HasQfUses(NodeAddr<DefNode *> CopyDef, |
| 1333 | MachineInstr *CopyMI) { |
| 1334 | NodeSet UseSet; |
| 1335 | getAllRealUses(DA: CopyDef, UNodeSet&: UseSet, L: LV, G: DFG); |
| 1336 | |
| 1337 | if (UseSet.size() == 0) |
| 1338 | return RegType::undefined; |
| 1339 | |
| 1340 | bool hasQf16Use = false; |
| 1341 | bool hasQf32Use = false; |
| 1342 | |
| 1343 | LLVM_DEBUG(dbgs() << "[COPY]\nUses of the copy are: " ); |
| 1344 | for (auto UI : UseSet) { |
| 1345 | NodeAddr<UseNode *> UA = DFG->addr<UseNode *>(N: UI); |
| 1346 | if (UA.Addr->getFlags() & NodeAttrs::PhiRef) |
| 1347 | continue; |
| 1348 | NodeAddr<StmtNode *> UseStmt = UA.Addr->getOwner(G: *DFG); |
| 1349 | MachineInstr *UseMI = UseStmt.Addr->getCode(); |
| 1350 | unsigned OpNo = UA.Addr->getOp().getOperandNo(); |
| 1351 | |
| 1352 | LLVM_DEBUG(dbgs() << "\nCopy's use: " ; UseMI->dump()); |
| 1353 | // Any reached use should not be a non-qf instruction |
| 1354 | if (!HII->usesQFOperand(MI: UseMI, Index: OpNo)) |
| 1355 | return RegType::ieee; |
| 1356 | |
| 1357 | // Determine the qf type from the use |
| 1358 | if (HII->usesQF16Operand(MI: UseMI, Index: OpNo)) |
| 1359 | hasQf16Use = true; |
| 1360 | else if (HII->usesQF32Operand(MI: UseMI, Index: OpNo)) |
| 1361 | hasQf32Use = true; |
| 1362 | |
| 1363 | // Any reached use should not already be converted to IEEE. |
| 1364 | // If present, it means that the reached use has other reaching def |
| 1365 | // other than the copy. |
| 1366 | if (QFUsesMap.find(x: UseMI) != QFUsesMap.end()) { |
| 1367 | auto Entry = QFUsesMap[UseMI]; |
| 1368 | if (OpNo == 1 && Entry.first == true) |
| 1369 | return RegType::ieee; |
| 1370 | if (OpNo == 2 && Entry.second == true) |
| 1371 | return RegType::ieee; |
| 1372 | } |
| 1373 | } |
| 1374 | |
| 1375 | // Set the output type based on uses |
| 1376 | if (hasQf16Use) { |
| 1377 | // Check if copy destination is double-wide |
| 1378 | if (Hexagon::HvxWRRegClass.contains(Reg: CopyMI->getOperand(i: 0).getReg())) |
| 1379 | return RegType::qf16_double; |
| 1380 | else |
| 1381 | return RegType::qf16; |
| 1382 | } else if (hasQf32Use) { |
| 1383 | if (Hexagon::HvxWRRegClass.contains(Reg: CopyMI->getOperand(i: 0).getReg())) |
| 1384 | return RegType::qf32_double; |
| 1385 | else |
| 1386 | return RegType::qf32; |
| 1387 | } |
| 1388 | |
| 1389 | return RegType::undefined; |
| 1390 | } |
| 1391 | |
| 1392 | // Go through the collected copies and insert conversion to sf/hf |
| 1393 | // conditionally *after their reaching defs*. This is done because there |
| 1394 | // can be mutliple reaching defs of the copies. Also, check for the uses |
| 1395 | // of the reaching def and handle qf uses too by changing opcode or |
| 1396 | // inserting converts. |
| 1397 | // Additionally, check for the uses of the copy |
| 1398 | // and handle them via changing opcode or inserting converts. |
| 1399 | bool HexagonPostRAHandleQFP::HandleCopies() { |
| 1400 | |
| 1401 | bool Changed = false; |
| 1402 | |
| 1403 | // If a convert is inserted after a reaching def, add it to ignorelist. |
| 1404 | // This is because this reaching def can be reaching def of other copies |
| 1405 | // due to non-SSA form. |
| 1406 | for (auto It : QFCopys) { |
| 1407 | |
| 1408 | // Get details of the copy node |
| 1409 | NodeAddr<DefNode *> CopyNode = DFG->addr<DefNode *>(N: It.first.first); |
| 1410 | NodeAddr<StmtNode *> StNode = CopyNode.Addr->getOwner(G: *DFG); |
| 1411 | [[maybe_unused]] auto *CopyMI = StNode.Addr->getCode(); |
| 1412 | LLVM_DEBUG(dbgs() << "\nHandling Reaching Defs of COPY: " ; CopyMI->dump(); |
| 1413 | std::string Type; switch (It.second) { |
| 1414 | case RegType::qf32_double: |
| 1415 | Type = "qf32_double" ; |
| 1416 | break; |
| 1417 | case RegType::qf32: |
| 1418 | Type = "qf32" ; |
| 1419 | break; |
| 1420 | case RegType::qf16: |
| 1421 | Type = "qf16" ; |
| 1422 | break; |
| 1423 | case RegType::qf16_double: |
| 1424 | Type = "qf16_double" ; |
| 1425 | break; |
| 1426 | default: |
| 1427 | Type = "ieee" ; |
| 1428 | } dbgs() << "\t Type: " |
| 1429 | << Type << "\n" ); |
| 1430 | |
| 1431 | // insert convert to IEEE after the reaching def if it generates qf type |
| 1432 | RegType RTy = It.second; |
| 1433 | if (RTy != RegType::ieee) { |
| 1434 | |
| 1435 | // get details of the reaching def node |
| 1436 | NodeAddr<DefNode *> ReachDefNode = DFG->addr<DefNode *>(N: It.first.second); |
| 1437 | NodeAddr<StmtNode *> StNode = ReachDefNode.Addr->getOwner(G: *DFG); |
| 1438 | auto *ReachingDef = StNode.Addr->getCode(); |
| 1439 | |
| 1440 | if (IgnoreInsertConvList.find(Ptr: ReachingDef) != IgnoreInsertConvList.end()) |
| 1441 | continue; |
| 1442 | |
| 1443 | // Collect the reaching defs to be processed later. |
| 1444 | ReachDefOfCopies.insert(KV: std::make_pair(x&: ReachingDef, y&: RTy)); |
| 1445 | |
| 1446 | // Process the reached uses of the reaching def now for |
| 1447 | // incorrect usage, since the register type has changed |
| 1448 | // following the conversion. |
| 1449 | LLVM_DEBUG(dbgs() << "\n[COPY]\tAnalyzing uses of the reaching defs \ |
| 1450 | of the copy..." ); |
| 1451 | collectQFUses(RegDef: ReachDefNode, DefMI: ReachingDef); |
| 1452 | collectConvQFInstr(RegDef&: ReachDefNode); |
| 1453 | IgnoreInsertConvList.insert(Ptr: ReachingDef); |
| 1454 | Changed = true; |
| 1455 | } |
| 1456 | } |
| 1457 | |
| 1458 | // Loop through copies with qf uses |
| 1459 | for (auto It : QFCopys) { |
| 1460 | |
| 1461 | // Get details of the copy node |
| 1462 | NodeAddr<DefNode *> CopyNode = DFG->addr<DefNode *>(N: It.first.first); |
| 1463 | NodeAddr<StmtNode *> StNode = CopyNode.Addr->getOwner(G: *DFG); |
| 1464 | auto *CopyMI = StNode.Addr->getCode(); |
| 1465 | LLVM_DEBUG(dbgs() << "\nHandling COPY: " ; CopyMI->dump()); |
| 1466 | RegType RTy = It.second; |
| 1467 | |
| 1468 | // Process the reached uses of the copy to find any incorrect |
| 1469 | // qf uses. If the copy's uses are all qf types, we need to convert |
| 1470 | // its result back to qf |
| 1471 | // FIXME: don't include the copy if its the last instruction since |
| 1472 | // it is *probably* not possible to insert via BuildMI at the end of BB |
| 1473 | RTy = HasQfUses(CopyDef: CopyNode, CopyMI); |
| 1474 | if (RTy != RegType::ieee && RTy != RegType::undefined && |
| 1475 | (++CopyMI->getIterator() != CopyMI->getParent()->end())) { |
| 1476 | if (!ConvertToQfCopies.contains(Val: CopyMI)) { |
| 1477 | ConvertToQfCopies[CopyMI] = std::make_pair(x&: CopyNode, y&: RTy); |
| 1478 | LLVM_DEBUG(dbgs() << "\n[ConvertToQfCopies]\tAdded copy: " ; |
| 1479 | CopyMI->dump(); std::string Type; switch (RTy) { |
| 1480 | case RegType::qf32_double: |
| 1481 | Type = "qf32_double" ; |
| 1482 | break; |
| 1483 | case RegType::qf32: |
| 1484 | Type = "qf32" ; |
| 1485 | break; |
| 1486 | case RegType::qf16: |
| 1487 | Type = "qf16" ; |
| 1488 | break; |
| 1489 | case RegType::qf16_double: |
| 1490 | Type = "qf16_double" ; |
| 1491 | break; |
| 1492 | default: |
| 1493 | Type = "ieee" ; |
| 1494 | } dbgs() << "\t Type: " |
| 1495 | << Type << "\n" ); |
| 1496 | } |
| 1497 | continue; |
| 1498 | } |
| 1499 | LLVM_DEBUG(dbgs() << "\n[COPY]\tAnalyzing uses of the copy..." ); |
| 1500 | collectQFUses(RegDef: CopyNode, DefMI: CopyMI); |
| 1501 | collectConvQFInstr(RegDef&: CopyNode); |
| 1502 | } |
| 1503 | |
| 1504 | Changed |= HandleReachDefOfCopies(); |
| 1505 | Changed |= HandleMultiReachingDefs(); |
| 1506 | Changed |= HandleConvertToQfCopies(); |
| 1507 | |
| 1508 | return Changed; |
| 1509 | } |
| 1510 | |
| 1511 | // Inserts conversion instruction sf/hf = qf before spilling |
| 1512 | // Uses the same physical register for conversion. |
| 1513 | // Additinally checks for the uses of the register; and |
| 1514 | // conditionally store them to handle later. |
| 1515 | bool HexagonPostRAHandleQFP::HandleSpills() { |
| 1516 | |
| 1517 | LLVM_DEBUG(dbgs() << "\n[Handling Spill]\n" ); |
| 1518 | bool Changed = false; |
| 1519 | for (auto It : SpillMIs) { |
| 1520 | |
| 1521 | MachineInstr *MI = It.first; |
| 1522 | auto OpC = MI->getOpcode(); |
| 1523 | |
| 1524 | auto NodeDef = It.second; |
| 1525 | NodeAddr<StmtNode *> Stmt = NodeDef.Addr->getOwner(G: *DFG); |
| 1526 | MachineInstr *DefMI = Stmt.Addr->getCode(); |
| 1527 | auto RegOp = MI->getOperand(i: 2); |
| 1528 | Register DefR = RegOp.getReg(); |
| 1529 | |
| 1530 | // handles widened qf16/qf32 instructions. |
| 1531 | if (OpC == Hexagon::PS_vstorerw_ai) { |
| 1532 | if (!Hexagon::HvxWRRegClass.contains(Reg: DefR)) |
| 1533 | assert(false && " Unhandled Vector Register class passed\n" ); |
| 1534 | // Walk through the uses of DefLo and DefHi and if there is QFP |
| 1535 | // instructions, the instruction needs to be updated to use sf operands |
| 1536 | // instead of qf operands. |
| 1537 | collectQFUses(RegDef: NodeDef, DefMI); |
| 1538 | |
| 1539 | if (IgnoreInsertConvList.find(Ptr: DefMI) != IgnoreInsertConvList.end()) |
| 1540 | continue; |
| 1541 | |
| 1542 | // Collect the reached uses of ReachDefInstr |
| 1543 | // which are sf/hf = qf conversion instructions. |
| 1544 | collectConvQFInstr(RegDef&: NodeDef); |
| 1545 | Register DefLo = HRI->getSubReg(Reg: DefR, Idx: Hexagon::vsub_lo); |
| 1546 | Register DefHi = HRI->getSubReg(Reg: DefR, Idx: Hexagon::vsub_hi); |
| 1547 | |
| 1548 | // Create two copy instructions, one each for Hi and Lo conditionally. |
| 1549 | // Liveness is the same is for the store instruction for the register. |
| 1550 | // If both are double registers, two insertions are done. |
| 1551 | // If one of the subregs are reaching to the store, conversion is done |
| 1552 | // for that subreg. |
| 1553 | Register DReg = DefMI->getOperand(i: 0).getReg(); |
| 1554 | if (HII->isQFP16Instr(MI: DefMI)) { |
| 1555 | if (DefLo == DReg || Hexagon::HvxWRRegClass.contains(Reg: DReg)) |
| 1556 | insertInstr(MI: DefMI, MIOpcode: Hexagon::V6_vconv_hf_qf16, SrcReg: DefLo, DstReg: DefLo, |
| 1557 | Flags: getRegState(RegOp) | RegState::Kill); |
| 1558 | |
| 1559 | if (DefHi == DReg || Hexagon::HvxWRRegClass.contains(Reg: DReg)) |
| 1560 | insertInstr(MI: DefMI, MIOpcode: Hexagon::V6_vconv_hf_qf16, SrcReg: DefHi, DstReg: DefHi, |
| 1561 | Flags: getRegState(RegOp) | RegState::Kill); |
| 1562 | } else if (HII->isQFP32Instr(MI: DefMI)) { |
| 1563 | if (DefLo == DReg || Hexagon::HvxWRRegClass.contains(Reg: DReg)) |
| 1564 | insertInstr(MI: DefMI, MIOpcode: Hexagon::V6_vconv_sf_qf32, SrcReg: DefLo, DstReg: DefLo, |
| 1565 | Flags: getRegState(RegOp) | RegState::Kill); |
| 1566 | |
| 1567 | if (DefHi == DReg || Hexagon::HvxWRRegClass.contains(Reg: DReg)) |
| 1568 | insertInstr(MI: DefMI, MIOpcode: Hexagon::V6_vconv_sf_qf32, SrcReg: DefHi, DstReg: DefHi, |
| 1569 | Flags: getRegState(RegOp) | RegState::Kill); |
| 1570 | } |
| 1571 | IgnoreInsertConvList.insert(Ptr: DefMI); |
| 1572 | Changed = true; |
| 1573 | |
| 1574 | // Handles instructions which output qf32 type. |
| 1575 | } else if (OpC == Hexagon::PS_vstorerv_ai && HII->isQFP32Instr(MI: DefMI)) { |
| 1576 | collectQFUses(RegDef: NodeDef, DefMI); |
| 1577 | if (IgnoreInsertConvList.find(Ptr: DefMI) != IgnoreInsertConvList.end()) |
| 1578 | continue; |
| 1579 | collectConvQFInstr(RegDef&: NodeDef); |
| 1580 | |
| 1581 | insertInstr(MI: DefMI, MIOpcode: Hexagon::V6_vconv_sf_qf32, SrcReg: DefR, DstReg: DefR, |
| 1582 | Flags: getRegState(RegOp) | RegState::Kill); |
| 1583 | |
| 1584 | IgnoreInsertConvList.insert(Ptr: DefMI); |
| 1585 | Changed = true; |
| 1586 | |
| 1587 | // Handles instructions which output qf16 type. |
| 1588 | } else if (OpC == Hexagon::PS_vstorerv_ai && HII->isQFP16Instr(MI: DefMI)) { |
| 1589 | collectQFUses(RegDef: NodeDef, DefMI); |
| 1590 | if (IgnoreInsertConvList.find(Ptr: DefMI) != IgnoreInsertConvList.end()) |
| 1591 | continue; |
| 1592 | collectConvQFInstr(RegDef&: NodeDef); |
| 1593 | |
| 1594 | insertInstr(MI: DefMI, MIOpcode: Hexagon::V6_vconv_hf_qf16, SrcReg: DefR, DstReg: DefR, |
| 1595 | Flags: getRegState(RegOp) | RegState::Kill); |
| 1596 | |
| 1597 | IgnoreInsertConvList.insert(Ptr: DefMI); |
| 1598 | Changed = true; |
| 1599 | } else { |
| 1600 | LLVM_DEBUG(MI->dump()); |
| 1601 | llvm_unreachable("This case is not handled. Look above for MI\n" ); |
| 1602 | } |
| 1603 | } |
| 1604 | return Changed; |
| 1605 | } |
| 1606 | |
| 1607 | bool HexagonPostRAHandleQFP::runOnMachineFunction(MachineFunction &MF) { |
| 1608 | |
| 1609 | if (DisablePostRAHandleQFloat) |
| 1610 | return false; |
| 1611 | |
| 1612 | LLVM_DEBUG( |
| 1613 | dbgs() << "\n=== Entering Hexagon Fixup QF spills and refills pass ===\n" |
| 1614 | << "Mode: " ; |
| 1615 | switch (QFloatModeValue) { |
| 1616 | case QFloatMode::StrictIEEE: |
| 1617 | dbgs() << "Strict IEEE" ; |
| 1618 | break; |
| 1619 | case QFloatMode::IEEE: |
| 1620 | dbgs() << "IEEE" ; |
| 1621 | break; |
| 1622 | case QFloatMode::Lossy: |
| 1623 | dbgs() << "Lossy" ; |
| 1624 | break; |
| 1625 | default: |
| 1626 | dbgs() << "Legacy" ; |
| 1627 | break; |
| 1628 | }; |
| 1629 | dbgs() << "\n" ;); |
| 1630 | bool Changed = false; |
| 1631 | |
| 1632 | auto &_HST = MF.getSubtarget<HexagonSubtarget>(); |
| 1633 | if (!_HST.useHVXOps()) |
| 1634 | return false; |
| 1635 | |
| 1636 | HII = _HST.getInstrInfo(); |
| 1637 | |
| 1638 | // If the mode is legacy, the function may not contain qf instructions |
| 1639 | // check if this pass is required to run for legacy mode. |
| 1640 | if (QFloatModeValue == QFloatMode::Legacy) |
| 1641 | if (!HII->hasQFPInstrs(MF)) |
| 1642 | return false; |
| 1643 | |
| 1644 | HRI = _HST.getRegisterInfo(); |
| 1645 | MRI = &MF.getRegInfo(); |
| 1646 | const auto &MDF = getAnalysis<MachineDominanceFrontierWrapperPass>().getMDF(); |
| 1647 | MachineDominatorTree *MDT = |
| 1648 | &getAnalysis<MachineDominatorTreeWrapperPass>().getDomTree(); |
| 1649 | HST = &_HST; |
| 1650 | |
| 1651 | // We need Register Dataflow Graph(RDG) to calculate reaching definitions |
| 1652 | // since the Machine code is not in SSA. |
| 1653 | // DDG holds the graph on which we iterate for the nodes. |
| 1654 | DataFlowGraph G(MF, *HII, *HRI, *MDT, MDF); |
| 1655 | G.build(); |
| 1656 | DFG = &G; |
| 1657 | |
| 1658 | Liveness L(*MRI, *DFG); |
| 1659 | L.computePhiInfo(); |
| 1660 | LV = &L; |
| 1661 | |
| 1662 | // Find and save the list of QFP stack spills. |
| 1663 | // For refills store all refill instructions to process conditionally later. |
| 1664 | NodeAddr<FuncNode *> FA = DFG->getFunc(); |
| 1665 | LLVM_DEBUG(dbgs() << "==== [RefMap#]=====:\n " |
| 1666 | << Print<NodeAddr<FuncNode *>>(FA, *DFG) << "\n" ); |
| 1667 | for (NodeAddr<BlockNode *> BA : FA.Addr->members(G: *DFG)) { |
| 1668 | for (auto IA : BA.Addr->members(G: *DFG)) { |
| 1669 | |
| 1670 | if (!DFG->IsCode<NodeAttrs::Stmt>(BA: IA)) |
| 1671 | continue; |
| 1672 | |
| 1673 | // 'SA' holds the Statement node which contains the machine instruction. |
| 1674 | NodeAddr<StmtNode *> SA = IA; |
| 1675 | MachineInstr *I = SA.Addr->getCode(); |
| 1676 | |
| 1677 | switch (I->getOpcode()) { |
| 1678 | case Hexagon::PS_vstorerw_ai: |
| 1679 | case Hexagon::PS_vstorerv_ai: |
| 1680 | collectQFPStackSpill(StNode: &SA); |
| 1681 | break; |
| 1682 | case Hexagon::PS_vloadrw_ai: |
| 1683 | case Hexagon::PS_vloadrv_ai: |
| 1684 | collectQFPStackRefill(StNode: &SA); |
| 1685 | break; |
| 1686 | case TargetOpcode::COPY: |
| 1687 | collectCopies(StNode: &SA); |
| 1688 | break; |
| 1689 | default: |
| 1690 | break; |
| 1691 | } |
| 1692 | } |
| 1693 | } |
| 1694 | |
| 1695 | // Walk through the spills and insert converts when necessary. |
| 1696 | // Additionally, walk though the uses of the converts and |
| 1697 | // store them conditionally for later processing. |
| 1698 | LLVM_DEBUG(dbgs() << "\nHandling spills...." ); |
| 1699 | Changed |= HandleSpills(); |
| 1700 | SpillMIs.clear(); |
| 1701 | |
| 1702 | // Walk through the uses of the refill instructions. |
| 1703 | // Process them if they are used as qf operands. |
| 1704 | LLVM_DEBUG(dbgs() << "\nCollecting refills....\n" ); |
| 1705 | for (NodeAddr<DefNode *> DfNode : RefillMIs) { |
| 1706 | |
| 1707 | NodeAddr<StmtNode *> Stmt = DfNode.Addr->getOwner(G: *DFG); |
| 1708 | MachineInstr *DefMI = Stmt.Addr->getCode(); |
| 1709 | collectQFUses(RegDef: DfNode, DefMI); |
| 1710 | collectConvQFInstr(RegDef&: DfNode); |
| 1711 | } |
| 1712 | RefillMIs.clear(); |
| 1713 | |
| 1714 | LLVM_DEBUG(dbgs() << "\nHandling copies...." ); |
| 1715 | Changed |= HandleCopies(); |
| 1716 | QFCopys.clear(); |
| 1717 | PossibleMultiReachDefs.clear(); |
| 1718 | ReachDefOfCopies.clear(); |
| 1719 | ConvertToQfCopies.clear(); |
| 1720 | |
| 1721 | LLVM_DEBUG(dbgs() << "\n === QF Uses map === " ; for (auto It : QFUsesMap) { |
| 1722 | dbgs() << "\nInstruction: " ; |
| 1723 | It.first->dump(); |
| 1724 | dbgs() << "\t Property: " << It.second.first << " ," << It.second.second; |
| 1725 | }); |
| 1726 | |
| 1727 | // Insert new opcodes as applicable for the refill uses. |
| 1728 | // Delete the original instructions. |
| 1729 | Changed |= HandleRefills(); |
| 1730 | |
| 1731 | // Handle non-saturating instructions by inserting convert(s) from sf to qf. |
| 1732 | Changed |= HandleNonSatInstr(); |
| 1733 | QFNonSatMIs.clear(); |
| 1734 | // Cleanup |
| 1735 | for (auto It : QFUsesMap) |
| 1736 | It.first->eraseFromParent(); |
| 1737 | QFUsesMap.clear(); |
| 1738 | IgnoreInsertConvList.clear(); |
| 1739 | |
| 1740 | return Changed; |
| 1741 | } |
| 1742 | |
| 1743 | //===----------------------------------------------------------------------===// |
| 1744 | // Public Constructor Functions |
| 1745 | //===----------------------------------------------------------------------===// |
| 1746 | INITIALIZE_PASS_BEGIN(HexagonPostRAHandleQFP, "handle-qfp-spills-refills" , |
| 1747 | "Hexagon Post RA Handle QFloat" , false, false) |
| 1748 | INITIALIZE_PASS_DEPENDENCY(MachineDominatorTreeWrapperPass) |
| 1749 | INITIALIZE_PASS_DEPENDENCY(MachineDominanceFrontierWrapperPass) |
| 1750 | INITIALIZE_PASS_END(HexagonPostRAHandleQFP, "handle-qfp-spills-refills" , |
| 1751 | "Hexagon PostRA Handle QFloat" , false, false) |
| 1752 | |
| 1753 | FunctionPass *llvm::createHexagonPostRAHandleQFP() { |
| 1754 | return new HexagonPostRAHandleQFP(); |
| 1755 | } |
| 1756 | |