| 1 | //===- HexagonSubtarget.cpp - Hexagon Subtarget Information ---------------===// |
| 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 | // This file implements the Hexagon specific subclass of TargetSubtarget. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "HexagonSubtarget.h" |
| 14 | #include "HexagonInstrInfo.h" |
| 15 | #include "HexagonRegisterInfo.h" |
| 16 | #include "MCTargetDesc/HexagonMCTargetDesc.h" |
| 17 | #include "llvm/ADT/STLExtras.h" |
| 18 | #include "llvm/ADT/SmallVector.h" |
| 19 | #include "llvm/ADT/StringRef.h" |
| 20 | #include "llvm/CodeGen/MachineInstr.h" |
| 21 | #include "llvm/CodeGen/MachineOperand.h" |
| 22 | #include "llvm/CodeGen/MachineScheduler.h" |
| 23 | #include "llvm/CodeGen/ScheduleDAG.h" |
| 24 | #include "llvm/CodeGen/ScheduleDAGInstrs.h" |
| 25 | #include "llvm/IR/IntrinsicsHexagon.h" |
| 26 | #include "llvm/Support/CommandLine.h" |
| 27 | #include "llvm/Support/ErrorHandling.h" |
| 28 | #include "llvm/Target/TargetMachine.h" |
| 29 | #include <algorithm> |
| 30 | #include <cassert> |
| 31 | #include <optional> |
| 32 | |
| 33 | using namespace llvm; |
| 34 | |
| 35 | #define DEBUG_TYPE "hexagon-subtarget" |
| 36 | |
| 37 | #define GET_SUBTARGETINFO_CTOR |
| 38 | #define GET_SUBTARGETINFO_TARGET_DESC |
| 39 | #include "HexagonGenSubtargetInfo.inc" |
| 40 | |
| 41 | static cl::opt<bool> EnableBSBSched("enable-bsb-sched" , cl::Hidden, |
| 42 | cl::init(Val: true)); |
| 43 | |
| 44 | static cl::opt<bool> EnableTCLatencySched("enable-tc-latency-sched" , cl::Hidden, |
| 45 | cl::init(Val: false)); |
| 46 | |
| 47 | static cl::opt<bool> |
| 48 | EnableDotCurSched("enable-cur-sched" , cl::Hidden, cl::init(Val: true), |
| 49 | cl::desc("Enable the scheduler to generate .cur" )); |
| 50 | |
| 51 | static cl::opt<bool> |
| 52 | DisableHexagonMISched("disable-hexagon-misched" , cl::Hidden, |
| 53 | cl::desc("Disable Hexagon MI Scheduling" )); |
| 54 | |
| 55 | static cl::opt<bool> OverrideLongCalls( |
| 56 | "hexagon-long-calls" , cl::Hidden, |
| 57 | cl::desc("If present, forces/disables the use of long calls" )); |
| 58 | |
| 59 | static cl::opt<bool> |
| 60 | EnablePredicatedCalls("hexagon-pred-calls" , cl::Hidden, |
| 61 | cl::desc("Consider calls to be predicable" )); |
| 62 | |
| 63 | static cl::opt<bool> SchedPredsCloser("sched-preds-closer" , cl::Hidden, |
| 64 | cl::init(Val: true)); |
| 65 | |
| 66 | static cl::opt<bool> SchedRetvalOptimization("sched-retval-optimization" , |
| 67 | cl::Hidden, cl::init(Val: true)); |
| 68 | |
| 69 | static cl::opt<bool> EnableCheckBankConflict( |
| 70 | "hexagon-check-bank-conflict" , cl::Hidden, cl::init(Val: true), |
| 71 | cl::desc("Enable checking for cache bank conflicts" )); |
| 72 | |
| 73 | HexagonSubtarget::HexagonSubtarget(const Triple &TT, StringRef CPU, |
| 74 | StringRef FS, const TargetMachine &TM) |
| 75 | : HexagonGenSubtargetInfo(TT, CPU, /*TuneCPU*/ CPU, FS), |
| 76 | OptLevel(TM.getOptLevel()), |
| 77 | CPUString(std::string(Hexagon_MC::selectHexagonCPU(CPU))), |
| 78 | TargetTriple(TT), InstrInfo(initializeSubtargetDependencies(CPU, FS)), |
| 79 | TLInfo(TM, *this), InstrItins(getInstrItineraryForCPU(CPU: CPUString)) { |
| 80 | Hexagon_MC::addArchSubtarget(STI: this, FS); |
| 81 | // Beware of the default constructor of InstrItineraryData: it will |
| 82 | // reset all members to 0. |
| 83 | assert(InstrItins.Itineraries != nullptr && "InstrItins not initialized" ); |
| 84 | } |
| 85 | |
| 86 | HexagonSubtarget & |
| 87 | HexagonSubtarget::initializeSubtargetDependencies(StringRef CPU, StringRef FS) { |
| 88 | std::optional<Hexagon::ArchEnum> ArchVer = Hexagon::getCpu(CPU: CPUString); |
| 89 | if (ArchVer) |
| 90 | HexagonArchVersion = *ArchVer; |
| 91 | else |
| 92 | llvm_unreachable("Unrecognized Hexagon processor version" ); |
| 93 | |
| 94 | UseHVX128BOps = false; |
| 95 | UseHVX64BOps = false; |
| 96 | UseAudioOps = false; |
| 97 | UseLongCalls = false; |
| 98 | |
| 99 | SubtargetFeatures Features(FS); |
| 100 | |
| 101 | // Turn on QFloat if the HVX version is v68+. |
| 102 | // The function ParseSubtargetFeatures will set feature bits and initialize |
| 103 | // subtarget's variables all in one, so there isn't a good way to preprocess |
| 104 | // the feature string, other than by tinkering with it directly. |
| 105 | auto IsQFloatFS = [](StringRef F) { |
| 106 | return F == "+hvx-qfloat" || F == "-hvx-qfloat" ; |
| 107 | }; |
| 108 | if (!llvm::count_if(Range: Features.getFeatures(), P: IsQFloatFS)) { |
| 109 | auto getHvxVersion = [&Features](StringRef FS) -> StringRef { |
| 110 | for (StringRef F : llvm::reverse(C: Features.getFeatures())) { |
| 111 | if (F.starts_with(Prefix: "+hvxv" )) |
| 112 | return F; |
| 113 | } |
| 114 | for (StringRef F : llvm::reverse(C: Features.getFeatures())) { |
| 115 | if (F == "-hvx" ) |
| 116 | return StringRef(); |
| 117 | if (F.starts_with(Prefix: "+hvx" ) || F == "-hvx" ) |
| 118 | return F.take_front(N: 4); // Return "+hvx" or "-hvx". |
| 119 | } |
| 120 | return StringRef(); |
| 121 | }; |
| 122 | |
| 123 | bool AddQFloat = false; |
| 124 | StringRef HvxVer = getHvxVersion(FS); |
| 125 | if (HvxVer.starts_with(Prefix: "+hvxv" )) { |
| 126 | int Ver = 0; |
| 127 | if (!HvxVer.drop_front(N: 5).consumeInteger(Radix: 10, Result&: Ver) && Ver >= 68) |
| 128 | AddQFloat = true; |
| 129 | } else if (HvxVer == "+hvx" ) { |
| 130 | if (hasV68Ops()) |
| 131 | AddQFloat = true; |
| 132 | } |
| 133 | |
| 134 | if (AddQFloat) |
| 135 | Features.AddFeature(String: "+hvx-qfloat" ); |
| 136 | } |
| 137 | |
| 138 | std::string FeatureString = Features.getString(); |
| 139 | ParseSubtargetFeatures(CPU: CPUString, /*TuneCPU*/ CPUString, FS: FeatureString); |
| 140 | |
| 141 | if (useHVXV68Ops()) |
| 142 | UseHVXFloatingPoint = UseHVXIEEEFPOps || UseHVXQFloatOps; |
| 143 | |
| 144 | if (UseHVXQFloatOps && UseHVXIEEEFPOps && UseHVXFloatingPoint) |
| 145 | LLVM_DEBUG( |
| 146 | dbgs() << "Behavior is undefined for simultaneous qfloat and ieee hvx codegen..." ); |
| 147 | |
| 148 | if (OverrideLongCalls.getPosition()) |
| 149 | UseLongCalls = OverrideLongCalls; |
| 150 | |
| 151 | UseBSBScheduling = hasV60Ops() && EnableBSBSched; |
| 152 | |
| 153 | if (isTinyCore()) { |
| 154 | // Tiny core has a single thread, so back-to-back scheduling is enabled by |
| 155 | // default. |
| 156 | if (!EnableBSBSched.getPosition()) |
| 157 | UseBSBScheduling = false; |
| 158 | } |
| 159 | |
| 160 | FeatureBitset FeatureBits = getFeatureBits(); |
| 161 | if (HexagonDisableDuplex) |
| 162 | setFeatureBits(FeatureBits.reset(I: Hexagon::FeatureDuplex)); |
| 163 | setFeatureBits(Hexagon_MC::completeHVXFeatures(FB: FeatureBits)); |
| 164 | |
| 165 | return *this; |
| 166 | } |
| 167 | |
| 168 | bool HexagonSubtarget::isHVXElementType(MVT Ty, bool IncludeBool) const { |
| 169 | if (!useHVXOps()) |
| 170 | return false; |
| 171 | if (Ty.isVector()) |
| 172 | Ty = Ty.getVectorElementType(); |
| 173 | if (IncludeBool && Ty == MVT::i1) |
| 174 | return true; |
| 175 | ArrayRef<MVT> ElemTypes = getHVXElementTypes(); |
| 176 | return llvm::is_contained(Range&: ElemTypes, Element: Ty); |
| 177 | } |
| 178 | |
| 179 | bool HexagonSubtarget::isHVXVectorType(EVT VecTy, bool IncludeBool) const { |
| 180 | if (!VecTy.isSimple()) |
| 181 | return false; |
| 182 | if (!VecTy.isVector() || !useHVXOps() || VecTy.isScalableVector()) |
| 183 | return false; |
| 184 | MVT ElemTy = VecTy.getSimpleVT().getVectorElementType(); |
| 185 | if (!IncludeBool && ElemTy == MVT::i1) |
| 186 | return false; |
| 187 | |
| 188 | unsigned HwLen = getVectorLength(); |
| 189 | unsigned NumElems = VecTy.getVectorNumElements(); |
| 190 | ArrayRef<MVT> ElemTypes = getHVXElementTypes(); |
| 191 | |
| 192 | if (IncludeBool && ElemTy == MVT::i1) { |
| 193 | // Boolean HVX vector types are formed from regular HVX vector types |
| 194 | // by replacing the element type with i1. |
| 195 | for (MVT T : ElemTypes) |
| 196 | if (NumElems * T.getSizeInBits() == 8 * HwLen) |
| 197 | return true; |
| 198 | return false; |
| 199 | } |
| 200 | |
| 201 | unsigned VecWidth = VecTy.getSizeInBits(); |
| 202 | if (VecWidth != 8 * HwLen && VecWidth != 16 * HwLen) |
| 203 | return false; |
| 204 | return llvm::is_contained(Range&: ElemTypes, Element: ElemTy); |
| 205 | } |
| 206 | |
| 207 | bool HexagonSubtarget::isTypeForHVX(Type *VecTy, bool IncludeBool) const { |
| 208 | if (!VecTy->isVectorTy() || isa<ScalableVectorType>(Val: VecTy)) |
| 209 | return false; |
| 210 | // Avoid types like <2 x i32*>. |
| 211 | Type *ScalTy = VecTy->getScalarType(); |
| 212 | if (!ScalTy->isIntegerTy() && |
| 213 | !(ScalTy->isFloatingPointTy() && useHVXFloatingPoint())) |
| 214 | return false; |
| 215 | // The given type may be something like <17 x i32>, which is not MVT, |
| 216 | // but can be represented as (non-simple) EVT. |
| 217 | EVT Ty = EVT::getEVT(Ty: VecTy, /*HandleUnknown*/false); |
| 218 | if (!Ty.getVectorElementType().isSimple()) |
| 219 | return false; |
| 220 | |
| 221 | auto isHvxTy = [this, IncludeBool](MVT SimpleTy) { |
| 222 | if (isHVXVectorType(VecTy: SimpleTy, IncludeBool)) |
| 223 | return true; |
| 224 | auto Action = getTargetLowering()->getPreferredVectorAction(VT: SimpleTy); |
| 225 | return Action == TargetLoweringBase::TypeWidenVector; |
| 226 | }; |
| 227 | |
| 228 | // Round up EVT to have power-of-2 elements, and keep checking if it |
| 229 | // qualifies for HVX, dividing it in half after each step. |
| 230 | MVT ElemTy = Ty.getVectorElementType().getSimpleVT(); |
| 231 | unsigned VecLen = PowerOf2Ceil(A: Ty.getVectorNumElements()); |
| 232 | while (VecLen > 1) { |
| 233 | MVT SimpleTy = MVT::getVectorVT(VT: ElemTy, NumElements: VecLen); |
| 234 | if (SimpleTy.isValid() && isHvxTy(SimpleTy)) |
| 235 | return true; |
| 236 | VecLen /= 2; |
| 237 | } |
| 238 | |
| 239 | return false; |
| 240 | } |
| 241 | |
| 242 | void HexagonSubtarget::UsrOverflowMutation::apply(ScheduleDAGInstrs *DAG) { |
| 243 | for (SUnit &SU : DAG->SUnits) { |
| 244 | if (!SU.isInstr()) |
| 245 | continue; |
| 246 | SmallVector<SDep, 4> Erase; |
| 247 | for (auto &D : SU.Preds) |
| 248 | if (D.getKind() == SDep::Output && D.getReg() == Hexagon::USR_OVF) |
| 249 | Erase.push_back(Elt: D); |
| 250 | for (auto &E : Erase) |
| 251 | SU.removePred(D: E); |
| 252 | } |
| 253 | } |
| 254 | |
| 255 | void HexagonSubtarget::HVXMemLatencyMutation::apply(ScheduleDAGInstrs *DAG) { |
| 256 | for (SUnit &SU : DAG->SUnits) { |
| 257 | // Update the latency of chain edges between v60 vector load or store |
| 258 | // instructions to be 1. These instruction cannot be scheduled in the |
| 259 | // same packet. |
| 260 | MachineInstr &MI1 = *SU.getInstr(); |
| 261 | auto *QII = static_cast<const HexagonInstrInfo*>(DAG->TII); |
| 262 | bool IsStoreMI1 = MI1.mayStore(); |
| 263 | bool IsLoadMI1 = MI1.mayLoad(); |
| 264 | if (!QII->isHVXVec(MI: MI1) || !(IsStoreMI1 || IsLoadMI1)) |
| 265 | continue; |
| 266 | for (SDep &SI : SU.Succs) { |
| 267 | if (SI.getKind() != SDep::Order || SI.getLatency() != 0) |
| 268 | continue; |
| 269 | MachineInstr &MI2 = *SI.getSUnit()->getInstr(); |
| 270 | if (!QII->isHVXVec(MI: MI2)) |
| 271 | continue; |
| 272 | if ((IsStoreMI1 && MI2.mayStore()) || (IsLoadMI1 && MI2.mayLoad())) { |
| 273 | SI.setLatency(1); |
| 274 | SU.setHeightDirty(); |
| 275 | // Change the dependence in the opposite direction too. |
| 276 | for (SDep &PI : SI.getSUnit()->Preds) { |
| 277 | if (PI.getSUnit() != &SU || PI.getKind() != SDep::Order) |
| 278 | continue; |
| 279 | PI.setLatency(1); |
| 280 | SI.getSUnit()->setDepthDirty(); |
| 281 | } |
| 282 | } |
| 283 | } |
| 284 | } |
| 285 | } |
| 286 | |
| 287 | // Check if a call and subsequent A2_tfrpi instructions should maintain |
| 288 | // scheduling affinity. We are looking for the TFRI to be consumed in |
| 289 | // the next instruction. This should help reduce the instances of |
| 290 | // double register pairs being allocated and scheduled before a call |
| 291 | // when not used until after the call. This situation is exacerbated |
| 292 | // by the fact that we allocate the pair from the callee saves list, |
| 293 | // leading to excess spills and restores. |
| 294 | bool HexagonSubtarget::CallMutation::shouldTFRICallBind( |
| 295 | const HexagonInstrInfo &HII, const SUnit &Inst1, |
| 296 | const SUnit &Inst2) const { |
| 297 | if (Inst1.getInstr()->getOpcode() != Hexagon::A2_tfrpi) |
| 298 | return false; |
| 299 | |
| 300 | // TypeXTYPE are 64 bit operations. |
| 301 | unsigned Type = HII.getType(MI: *Inst2.getInstr()); |
| 302 | return Type == HexagonII::TypeS_2op || Type == HexagonII::TypeS_3op || |
| 303 | Type == HexagonII::TypeALU64 || Type == HexagonII::TypeM; |
| 304 | } |
| 305 | |
| 306 | void HexagonSubtarget::CallMutation::apply(ScheduleDAGInstrs *DAGInstrs) { |
| 307 | ScheduleDAGMI *DAG = static_cast<ScheduleDAGMI*>(DAGInstrs); |
| 308 | SUnit* LastSequentialCall = nullptr; |
| 309 | // Map from virtual register to physical register from the copy. |
| 310 | DenseMap<unsigned, unsigned> VRegHoldingReg; |
| 311 | // Map from the physical register to the instruction that uses virtual |
| 312 | // register. This is used to create the barrier edge. |
| 313 | DenseMap<unsigned, SUnit *> LastVRegUse; |
| 314 | auto &TRI = *DAG->MF.getSubtarget().getRegisterInfo(); |
| 315 | auto &HII = *DAG->MF.getSubtarget<HexagonSubtarget>().getInstrInfo(); |
| 316 | |
| 317 | // Currently we only catch the situation when compare gets scheduled |
| 318 | // before preceding call. |
| 319 | for (unsigned su = 0, e = DAG->SUnits.size(); su != e; ++su) { |
| 320 | // Remember the call. |
| 321 | if (DAG->SUnits[su].getInstr()->isCall()) |
| 322 | LastSequentialCall = &DAG->SUnits[su]; |
| 323 | // Look for a compare that defines a predicate. |
| 324 | else if (DAG->SUnits[su].getInstr()->isCompare() && LastSequentialCall) |
| 325 | DAG->addEdge(SuccSU: &DAG->SUnits[su], PredDep: SDep(LastSequentialCall, SDep::Barrier)); |
| 326 | // Look for call and tfri* instructions. |
| 327 | else if (SchedPredsCloser && LastSequentialCall && su > 1 && su < e-1 && |
| 328 | shouldTFRICallBind(HII, Inst1: DAG->SUnits[su], Inst2: DAG->SUnits[su+1])) |
| 329 | DAG->addEdge(SuccSU: &DAG->SUnits[su], PredDep: SDep(&DAG->SUnits[su-1], SDep::Barrier)); |
| 330 | // Prevent redundant register copies due to reads and writes of physical |
| 331 | // registers. The original motivation for this was the code generated |
| 332 | // between two calls, which are caused both the return value and the |
| 333 | // argument for the next call being in %r0. |
| 334 | // Example: |
| 335 | // 1: <call1> |
| 336 | // 2: %vreg = COPY %r0 |
| 337 | // 3: <use of %vreg> |
| 338 | // 4: %r0 = ... |
| 339 | // 5: <call2> |
| 340 | // The scheduler would often swap 3 and 4, so an additional register is |
| 341 | // needed. This code inserts a Barrier dependence between 3 & 4 to prevent |
| 342 | // this. |
| 343 | // The code below checks for all the physical registers, not just R0/D0/V0. |
| 344 | else if (SchedRetvalOptimization) { |
| 345 | const MachineInstr *MI = DAG->SUnits[su].getInstr(); |
| 346 | if (MI->isCopy() && MI->getOperand(i: 1).getReg().isPhysical()) { |
| 347 | // %vregX = COPY %r0 |
| 348 | VRegHoldingReg[MI->getOperand(i: 0).getReg()] = MI->getOperand(i: 1).getReg(); |
| 349 | LastVRegUse.erase(Val: MI->getOperand(i: 1).getReg()); |
| 350 | } else { |
| 351 | for (const MachineOperand &MO : MI->operands()) { |
| 352 | if (!MO.isReg()) |
| 353 | continue; |
| 354 | if (MO.isUse() && !MI->isCopy() && |
| 355 | VRegHoldingReg.count(Val: MO.getReg())) { |
| 356 | // <use of %vregX> |
| 357 | LastVRegUse[VRegHoldingReg[MO.getReg()]] = &DAG->SUnits[su]; |
| 358 | } else if (MO.isDef() && MO.getReg().isPhysical()) { |
| 359 | for (MCRegAliasIterator AI(MO.getReg(), &TRI, true); AI.isValid(); |
| 360 | ++AI) { |
| 361 | if (auto It = LastVRegUse.find(Val: *AI); It != LastVRegUse.end()) { |
| 362 | if (It->second != &DAG->SUnits[su]) |
| 363 | // %r0 = ... |
| 364 | DAG->addEdge(SuccSU: &DAG->SUnits[su], |
| 365 | PredDep: SDep(It->second, SDep::Barrier)); |
| 366 | LastVRegUse.erase(I: It); |
| 367 | } |
| 368 | } |
| 369 | } |
| 370 | } |
| 371 | } |
| 372 | } |
| 373 | } |
| 374 | } |
| 375 | |
| 376 | void HexagonSubtarget::BankConflictMutation::apply(ScheduleDAGInstrs *DAG) { |
| 377 | if (!EnableCheckBankConflict) |
| 378 | return; |
| 379 | |
| 380 | const auto &HII = static_cast<const HexagonInstrInfo&>(*DAG->TII); |
| 381 | |
| 382 | // Create artificial edges between loads that could likely cause a bank |
| 383 | // conflict. Since such loads would normally not have any dependency |
| 384 | // between them, we cannot rely on existing edges. |
| 385 | for (unsigned i = 0, e = DAG->SUnits.size(); i != e; ++i) { |
| 386 | SUnit &S0 = DAG->SUnits[i]; |
| 387 | MachineInstr &L0 = *S0.getInstr(); |
| 388 | if (!L0.mayLoad() || L0.mayStore() || |
| 389 | HII.getAddrMode(MI: L0) != HexagonII::BaseImmOffset) |
| 390 | continue; |
| 391 | int64_t Offset0; |
| 392 | LocationSize Size0 = LocationSize::precise(Value: 0); |
| 393 | MachineOperand *BaseOp0 = HII.getBaseAndOffset(MI: L0, Offset&: Offset0, AccessSize&: Size0); |
| 394 | // Is the access size is longer than the L1 cache line, skip the check. |
| 395 | if (BaseOp0 == nullptr || !BaseOp0->isReg() || !Size0.hasValue() || |
| 396 | Size0.getValue() >= 32) |
| 397 | continue; |
| 398 | // Scan only up to 32 instructions ahead (to avoid n^2 complexity). |
| 399 | for (unsigned j = i+1, m = std::min(a: i+32, b: e); j != m; ++j) { |
| 400 | SUnit &S1 = DAG->SUnits[j]; |
| 401 | MachineInstr &L1 = *S1.getInstr(); |
| 402 | if (!L1.mayLoad() || L1.mayStore() || |
| 403 | HII.getAddrMode(MI: L1) != HexagonII::BaseImmOffset) |
| 404 | continue; |
| 405 | int64_t Offset1; |
| 406 | LocationSize Size1 = LocationSize::precise(Value: 0); |
| 407 | MachineOperand *BaseOp1 = HII.getBaseAndOffset(MI: L1, Offset&: Offset1, AccessSize&: Size1); |
| 408 | if (BaseOp1 == nullptr || !BaseOp1->isReg() || !Size0.hasValue() || |
| 409 | Size1.getValue() >= 32 || BaseOp0->getReg() != BaseOp1->getReg()) |
| 410 | continue; |
| 411 | // Check bits 3 and 4 of the offset: if they differ, a bank conflict |
| 412 | // is unlikely. |
| 413 | if (((Offset0 ^ Offset1) & 0x18) != 0) |
| 414 | continue; |
| 415 | // Bits 3 and 4 are the same, add an artificial edge and set extra |
| 416 | // latency. |
| 417 | SDep A(&S0, SDep::Artificial); |
| 418 | A.setLatency(1); |
| 419 | S1.addPred(D: A, Required: true); |
| 420 | } |
| 421 | } |
| 422 | } |
| 423 | |
| 424 | /// Enable use of alias analysis during code generation (during MI |
| 425 | /// scheduling, DAGCombine, etc.). |
| 426 | bool HexagonSubtarget::useAA() const { |
| 427 | if (OptLevel != CodeGenOptLevel::None) |
| 428 | return true; |
| 429 | return false; |
| 430 | } |
| 431 | |
| 432 | /// Perform target specific adjustments to the latency of a schedule |
| 433 | /// dependency. |
| 434 | void HexagonSubtarget::adjustSchedDependency( |
| 435 | SUnit *Src, int SrcOpIdx, SUnit *Dst, int DstOpIdx, SDep &Dep, |
| 436 | const TargetSchedModel *SchedModel) const { |
| 437 | if (!Src->isInstr() || !Dst->isInstr()) |
| 438 | return; |
| 439 | |
| 440 | MachineInstr *SrcInst = Src->getInstr(); |
| 441 | MachineInstr *DstInst = Dst->getInstr(); |
| 442 | const HexagonInstrInfo *QII = getInstrInfo(); |
| 443 | |
| 444 | // Instructions with .new operands have zero latency. |
| 445 | SmallPtrSet<SUnit *, 4> ExclSrc; |
| 446 | SmallPtrSet<SUnit *, 4> ExclDst; |
| 447 | if (QII->canExecuteInBundle(First: *SrcInst, Second: *DstInst) && |
| 448 | isBestZeroLatency(Src, Dst, TII: QII, ExclSrc, ExclDst)) { |
| 449 | Dep.setLatency(0); |
| 450 | return; |
| 451 | } |
| 452 | |
| 453 | // Set the latency for a copy to zero since we hope that is will get |
| 454 | // removed. |
| 455 | if (DstInst->isCopy()) |
| 456 | Dep.setLatency(0); |
| 457 | |
| 458 | // If it's a REG_SEQUENCE/COPY, use its destination instruction to determine |
| 459 | // the correct latency. |
| 460 | // If there are multiple uses of the def of COPY/REG_SEQUENCE, set the latency |
| 461 | // only if the latencies on all the uses are equal, otherwise set it to |
| 462 | // default. |
| 463 | if ((DstInst->isRegSequence() || DstInst->isCopy())) { |
| 464 | Register DReg = DstInst->getOperand(i: 0).getReg(); |
| 465 | std::optional<unsigned> DLatency; |
| 466 | for (const auto &DDep : Dst->Succs) { |
| 467 | MachineInstr *DDst = DDep.getSUnit()->getInstr(); |
| 468 | int UseIdx = -1; |
| 469 | for (unsigned OpNum = 0; OpNum < DDst->getNumOperands(); OpNum++) { |
| 470 | const MachineOperand &MO = DDst->getOperand(i: OpNum); |
| 471 | if (MO.isReg() && MO.getReg() && MO.isUse() && MO.getReg() == DReg) { |
| 472 | UseIdx = OpNum; |
| 473 | break; |
| 474 | } |
| 475 | } |
| 476 | |
| 477 | if (UseIdx == -1) |
| 478 | continue; |
| 479 | |
| 480 | std::optional<unsigned> Latency = |
| 481 | InstrInfo.getOperandLatency(ItinData: &InstrItins, DefMI: *SrcInst, DefIdx: 0, UseMI: *DDst, UseIdx); |
| 482 | |
| 483 | // Set DLatency for the first time. |
| 484 | if (!DLatency) |
| 485 | DLatency = Latency; |
| 486 | |
| 487 | // For multiple uses, if the Latency is different across uses, reset |
| 488 | // DLatency. |
| 489 | if (DLatency != Latency) { |
| 490 | DLatency = std::nullopt; |
| 491 | break; |
| 492 | } |
| 493 | } |
| 494 | Dep.setLatency(DLatency.value_or(u: 0)); |
| 495 | } |
| 496 | |
| 497 | // Try to schedule uses near definitions to generate .cur. |
| 498 | ExclSrc.clear(); |
| 499 | ExclDst.clear(); |
| 500 | if (EnableDotCurSched && QII->isToBeScheduledASAP(MI1: *SrcInst, MI2: *DstInst) && |
| 501 | isBestZeroLatency(Src, Dst, TII: QII, ExclSrc, ExclDst)) { |
| 502 | Dep.setLatency(0); |
| 503 | return; |
| 504 | } |
| 505 | int Latency = Dep.getLatency(); |
| 506 | bool IsArtificial = Dep.isArtificial(); |
| 507 | Latency = updateLatency(SrcInst&: *SrcInst, DstInst&: *DstInst, IsArtificial, Latency); |
| 508 | Dep.setLatency(Latency); |
| 509 | } |
| 510 | |
| 511 | void HexagonSubtarget::getPostRAMutations( |
| 512 | std::vector<std::unique_ptr<ScheduleDAGMutation>> &Mutations) const { |
| 513 | Mutations.push_back(x: std::make_unique<UsrOverflowMutation>()); |
| 514 | Mutations.push_back(x: std::make_unique<HVXMemLatencyMutation>()); |
| 515 | Mutations.push_back(x: std::make_unique<BankConflictMutation>()); |
| 516 | } |
| 517 | |
| 518 | void HexagonSubtarget::getSMSMutations( |
| 519 | std::vector<std::unique_ptr<ScheduleDAGMutation>> &Mutations) const { |
| 520 | Mutations.push_back(x: std::make_unique<UsrOverflowMutation>()); |
| 521 | Mutations.push_back(x: std::make_unique<HVXMemLatencyMutation>()); |
| 522 | } |
| 523 | |
| 524 | // Pin the vtable to this file. |
| 525 | void HexagonSubtarget::anchor() {} |
| 526 | |
| 527 | bool HexagonSubtarget::enableMachineScheduler() const { |
| 528 | if (DisableHexagonMISched.getNumOccurrences()) |
| 529 | return !DisableHexagonMISched; |
| 530 | return true; |
| 531 | } |
| 532 | |
| 533 | bool HexagonSubtarget::usePredicatedCalls() const { |
| 534 | return EnablePredicatedCalls; |
| 535 | } |
| 536 | |
| 537 | int HexagonSubtarget::updateLatency(MachineInstr &SrcInst, |
| 538 | MachineInstr &DstInst, bool IsArtificial, |
| 539 | int Latency) const { |
| 540 | if (IsArtificial) |
| 541 | return 1; |
| 542 | if (!hasV60Ops()) |
| 543 | return Latency; |
| 544 | |
| 545 | const HexagonInstrInfo &QII = *getInstrInfo(); |
| 546 | // BSB scheduling. |
| 547 | if (QII.isHVXVec(MI: SrcInst) || useBSBScheduling()) |
| 548 | Latency = (Latency + 1) >> 1; |
| 549 | return Latency; |
| 550 | } |
| 551 | |
| 552 | void HexagonSubtarget::restoreLatency(SUnit *Src, SUnit *Dst) const { |
| 553 | MachineInstr *SrcI = Src->getInstr(); |
| 554 | for (auto &I : Src->Succs) { |
| 555 | if (!I.isAssignedRegDep() || I.getSUnit() != Dst) |
| 556 | continue; |
| 557 | Register DepR = I.getReg(); |
| 558 | int DefIdx = -1; |
| 559 | for (unsigned OpNum = 0; OpNum < SrcI->getNumOperands(); OpNum++) { |
| 560 | const MachineOperand &MO = SrcI->getOperand(i: OpNum); |
| 561 | bool IsSameOrSubReg = false; |
| 562 | if (MO.isReg()) { |
| 563 | Register MOReg = MO.getReg(); |
| 564 | if (DepR.isVirtual()) { |
| 565 | IsSameOrSubReg = (MOReg == DepR); |
| 566 | } else { |
| 567 | IsSameOrSubReg = getRegisterInfo()->isSubRegisterEq(RegA: DepR, RegB: MOReg); |
| 568 | } |
| 569 | if (MO.isDef() && IsSameOrSubReg) |
| 570 | DefIdx = OpNum; |
| 571 | } |
| 572 | } |
| 573 | assert(DefIdx >= 0 && "Def Reg not found in Src MI" ); |
| 574 | MachineInstr *DstI = Dst->getInstr(); |
| 575 | SDep T = I; |
| 576 | for (unsigned OpNum = 0; OpNum < DstI->getNumOperands(); OpNum++) { |
| 577 | const MachineOperand &MO = DstI->getOperand(i: OpNum); |
| 578 | if (MO.isReg() && MO.isUse() && MO.getReg() == DepR) { |
| 579 | std::optional<unsigned> Latency = InstrInfo.getOperandLatency( |
| 580 | ItinData: &InstrItins, DefMI: *SrcI, DefIdx, UseMI: *DstI, UseIdx: OpNum); |
| 581 | |
| 582 | // For some instructions (ex: COPY), we might end up with < 0 latency |
| 583 | // as they don't have any Itinerary class associated with them. |
| 584 | if (!Latency) |
| 585 | Latency = 0; |
| 586 | bool IsArtificial = I.isArtificial(); |
| 587 | Latency = updateLatency(SrcInst&: *SrcI, DstInst&: *DstI, IsArtificial, Latency: *Latency); |
| 588 | I.setLatency(*Latency); |
| 589 | } |
| 590 | } |
| 591 | |
| 592 | // Update the latency of opposite edge too. |
| 593 | T.setSUnit(Src); |
| 594 | auto F = find(Range&: Dst->Preds, Val: T); |
| 595 | assert(F != Dst->Preds.end()); |
| 596 | F->setLatency(I.getLatency()); |
| 597 | } |
| 598 | } |
| 599 | |
| 600 | /// Change the latency between the two SUnits. |
| 601 | void HexagonSubtarget::changeLatency(SUnit *Src, SUnit *Dst, unsigned Lat) |
| 602 | const { |
| 603 | for (auto &I : Src->Succs) { |
| 604 | if (!I.isAssignedRegDep() || I.getSUnit() != Dst) |
| 605 | continue; |
| 606 | SDep T = I; |
| 607 | I.setLatency(Lat); |
| 608 | |
| 609 | // Update the latency of opposite edge too. |
| 610 | T.setSUnit(Src); |
| 611 | auto F = find(Range&: Dst->Preds, Val: T); |
| 612 | assert(F != Dst->Preds.end()); |
| 613 | F->setLatency(Lat); |
| 614 | } |
| 615 | } |
| 616 | |
| 617 | /// If the SUnit has a zero latency edge, return the other SUnit. |
| 618 | static SUnit *getZeroLatency(SUnit *N, SmallVector<SDep, 4> &Deps) { |
| 619 | for (auto &I : Deps) |
| 620 | if (I.isAssignedRegDep() && I.getLatency() == 0 && |
| 621 | !I.getSUnit()->getInstr()->isPseudo()) |
| 622 | return I.getSUnit(); |
| 623 | return nullptr; |
| 624 | } |
| 625 | |
| 626 | // Return true if these are the best two instructions to schedule |
| 627 | // together with a zero latency. Only one dependence should have a zero |
| 628 | // latency. If there are multiple choices, choose the best, and change |
| 629 | // the others, if needed. |
| 630 | bool HexagonSubtarget::isBestZeroLatency( |
| 631 | SUnit *Src, SUnit *Dst, const HexagonInstrInfo *TII, |
| 632 | SmallPtrSet<SUnit *, 4> &ExclSrc, SmallPtrSet<SUnit *, 4> &ExclDst) const { |
| 633 | MachineInstr &SrcInst = *Src->getInstr(); |
| 634 | MachineInstr &DstInst = *Dst->getInstr(); |
| 635 | |
| 636 | // Ignore Boundary SU nodes as these have null instructions. |
| 637 | if (Dst->isBoundaryNode()) |
| 638 | return false; |
| 639 | |
| 640 | if (SrcInst.isPHI() || DstInst.isPHI()) |
| 641 | return false; |
| 642 | |
| 643 | if (!TII->isToBeScheduledASAP(MI1: SrcInst, MI2: DstInst) && |
| 644 | !TII->canExecuteInBundle(First: SrcInst, Second: DstInst)) |
| 645 | return false; |
| 646 | |
| 647 | // The architecture doesn't allow three dependent instructions in the same |
| 648 | // packet. So, if the destination has a zero latency successor, then it's |
| 649 | // not a candidate for a zero latency predecessor. |
| 650 | if (getZeroLatency(N: Dst, Deps&: Dst->Succs) != nullptr) |
| 651 | return false; |
| 652 | |
| 653 | // Check if the Dst instruction is the best candidate first. |
| 654 | SUnit *Best = nullptr; |
| 655 | SUnit *DstBest = nullptr; |
| 656 | SUnit *SrcBest = getZeroLatency(N: Dst, Deps&: Dst->Preds); |
| 657 | if (SrcBest == nullptr || Src->NodeNum >= SrcBest->NodeNum) { |
| 658 | // Check that Src doesn't have a better candidate. |
| 659 | DstBest = getZeroLatency(N: Src, Deps&: Src->Succs); |
| 660 | if (DstBest == nullptr || Dst->NodeNum <= DstBest->NodeNum) |
| 661 | Best = Dst; |
| 662 | } |
| 663 | if (Best != Dst) |
| 664 | return false; |
| 665 | |
| 666 | // The caller frequently adds the same dependence twice. If so, then |
| 667 | // return true for this case too. |
| 668 | if ((Src == SrcBest && Dst == DstBest ) || |
| 669 | (SrcBest == nullptr && Dst == DstBest) || |
| 670 | (Src == SrcBest && Dst == nullptr)) |
| 671 | return true; |
| 672 | |
| 673 | // Reassign the latency for the previous bests, which requires setting |
| 674 | // the dependence edge in both directions. |
| 675 | if (SrcBest != nullptr) { |
| 676 | if (!hasV60Ops()) |
| 677 | changeLatency(Src: SrcBest, Dst, Lat: 1); |
| 678 | else |
| 679 | restoreLatency(Src: SrcBest, Dst); |
| 680 | } |
| 681 | if (DstBest != nullptr) { |
| 682 | if (!hasV60Ops()) |
| 683 | changeLatency(Src, Dst: DstBest, Lat: 1); |
| 684 | else |
| 685 | restoreLatency(Src, Dst: DstBest); |
| 686 | } |
| 687 | |
| 688 | // Attempt to find another opportunity for zero latency in a different |
| 689 | // dependence. |
| 690 | if (SrcBest && DstBest) |
| 691 | // If there is an edge from SrcBest to DstBst, then try to change that |
| 692 | // to 0 now. |
| 693 | changeLatency(Src: SrcBest, Dst: DstBest, Lat: 0); |
| 694 | else if (DstBest) { |
| 695 | // Check if the previous best destination instruction has a new zero |
| 696 | // latency dependence opportunity. |
| 697 | ExclSrc.insert(Ptr: Src); |
| 698 | for (auto &I : DstBest->Preds) |
| 699 | if (ExclSrc.count(Ptr: I.getSUnit()) == 0 && |
| 700 | isBestZeroLatency(Src: I.getSUnit(), Dst: DstBest, TII, ExclSrc, ExclDst)) |
| 701 | changeLatency(Src: I.getSUnit(), Dst: DstBest, Lat: 0); |
| 702 | } else if (SrcBest) { |
| 703 | // Check if previous best source instruction has a new zero latency |
| 704 | // dependence opportunity. |
| 705 | ExclDst.insert(Ptr: Dst); |
| 706 | for (auto &I : SrcBest->Succs) |
| 707 | if (ExclDst.count(Ptr: I.getSUnit()) == 0 && |
| 708 | isBestZeroLatency(Src: SrcBest, Dst: I.getSUnit(), TII, ExclSrc, ExclDst)) |
| 709 | changeLatency(Src: SrcBest, Dst: I.getSUnit(), Lat: 0); |
| 710 | } |
| 711 | |
| 712 | return true; |
| 713 | } |
| 714 | |
| 715 | unsigned HexagonSubtarget::getL1CacheLineSize() const { |
| 716 | return 32; |
| 717 | } |
| 718 | |
| 719 | unsigned HexagonSubtarget::getL1PrefetchDistance() const { |
| 720 | return 32; |
| 721 | } |
| 722 | |
| 723 | bool HexagonSubtarget::enableSubRegLiveness() const { return true; } |
| 724 | |
| 725 | Intrinsic::ID HexagonSubtarget::getIntrinsicId(unsigned Opc) const { |
| 726 | struct Scalar { |
| 727 | unsigned Opcode; |
| 728 | Intrinsic::ID IntId; |
| 729 | }; |
| 730 | struct Hvx { |
| 731 | unsigned Opcode; |
| 732 | Intrinsic::ID Int64Id, Int128Id; |
| 733 | }; |
| 734 | |
| 735 | static Scalar ScalarInts[] = { |
| 736 | #define GET_SCALAR_INTRINSICS |
| 737 | #include "HexagonDepInstrIntrinsics.inc" |
| 738 | #undef GET_SCALAR_INTRINSICS |
| 739 | }; |
| 740 | |
| 741 | static Hvx HvxInts[] = { |
| 742 | #define GET_HVX_INTRINSICS |
| 743 | #include "HexagonDepInstrIntrinsics.inc" |
| 744 | #undef GET_HVX_INTRINSICS |
| 745 | }; |
| 746 | |
| 747 | const auto CmpOpcode = [](auto A, auto B) { return A.Opcode < B.Opcode; }; |
| 748 | [[maybe_unused]] static bool SortedScalar = |
| 749 | (llvm::sort(C&: ScalarInts, Comp: CmpOpcode), true); |
| 750 | [[maybe_unused]] static bool SortedHvx = |
| 751 | (llvm::sort(C&: HvxInts, Comp: CmpOpcode), true); |
| 752 | |
| 753 | auto [BS, ES] = std::make_pair(x: std::begin(arr&: ScalarInts), y: std::end(arr&: ScalarInts)); |
| 754 | auto [BH, EH] = std::make_pair(x: std::begin(arr&: HvxInts), y: std::end(arr&: HvxInts)); |
| 755 | |
| 756 | auto FoundScalar = std::lower_bound(first: BS, last: ES, val: Scalar{.Opcode: Opc, .IntId: 0}, comp: CmpOpcode); |
| 757 | if (FoundScalar != ES && FoundScalar->Opcode == Opc) |
| 758 | return FoundScalar->IntId; |
| 759 | |
| 760 | auto FoundHvx = std::lower_bound(first: BH, last: EH, val: Hvx{.Opcode: Opc, .Int64Id: 0, .Int128Id: 0}, comp: CmpOpcode); |
| 761 | if (FoundHvx != EH && FoundHvx->Opcode == Opc) { |
| 762 | unsigned HwLen = getVectorLength(); |
| 763 | if (HwLen == 64) |
| 764 | return FoundHvx->Int64Id; |
| 765 | if (HwLen == 128) |
| 766 | return FoundHvx->Int128Id; |
| 767 | } |
| 768 | |
| 769 | std::string error = "Invalid opcode (" + std::to_string(val: Opc) + ")" ; |
| 770 | llvm_unreachable(error.c_str()); |
| 771 | return 0; |
| 772 | } |
| 773 | |