lib/MCA/InstrBuilder.cpp

   1 //===--------------------- InstrBuilder.cpp ---------------------*- C++ -*-===//
   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 /// \file
   9 ///
  10 /// This file implements the InstrBuilder interface.
  11 ///
  12 //===----------------------------------------------------------------------===//
  13
  14 #include "llvm/MCA/InstrBuilder.h"
  15 #include "llvm/ADT/APInt.h"
  16 #include "llvm/ADT/DenseMap.h"
  17 #include "llvm/MC/MCInst.h"
  18 #include "llvm/Support/Debug.h"
  19 #include "llvm/Support/WithColor.h"
  20 #include "llvm/Support/raw_ostream.h"
  21
  22 #define DEBUG_TYPE "llvm-mca"
  23
  24 namespace llvm {
  25 namespace mca {
  26
  27 InstrBuilder::InstrBuilder(const llvm::MCSubtargetInfo &sti,
  28                            const llvm::MCInstrInfo &mcii,
  29                            const llvm::MCRegisterInfo &mri,
  30                            const llvm::MCInstrAnalysis *mcia)
  31     : STI(sti), MCII(mcii), MRI(mri), MCIA(mcia), FirstCallInst(true),
  32       FirstReturnInst(true) {
  33   const MCSchedModel &SM = STI.getSchedModel();
  34   ProcResourceMasks.resize(SM.getNumProcResourceKinds());
  35   computeProcResourceMasks(STI.getSchedModel(), ProcResourceMasks);
  36 }
  37
  38 static void initializeUsedResources(InstrDesc &ID,
  39                                     const MCSchedClassDesc &SCDesc,
  40                                     const MCSubtargetInfo &STI,
  41                                     ArrayRef<uint64_t> ProcResourceMasks) {
  42   const MCSchedModel &SM = STI.getSchedModel();
  43
  44   // Populate resources consumed.
  45   using ResourcePlusCycles = std::pair<uint64_t, ResourceUsage>;
  46   std::vector<ResourcePlusCycles> Worklist;
  47
  48   // Track cycles contributed by resources that are in a "Super" relationship.
  49   // This is required if we want to correctly match the behavior of method
  50   // SubtargetEmitter::ExpandProcResource() in Tablegen. When computing the set
  51   // of "consumed" processor resources and resource cycles, the logic in
  52   // ExpandProcResource() doesn't update the number of resource cycles
  53   // contributed by a "Super" resource to a group.
  54   // We need to take this into account when we find that a processor resource is
  55   // part of a group, and it is also used as the "Super" of other resources.
  56   // This map stores the number of cycles contributed by sub-resources that are
  57   // part of a "Super" resource. The key value is the "Super" resource mask ID.
  58   DenseMap<uint64_t, unsigned> SuperResources;
  59
  60   unsigned NumProcResources = SM.getNumProcResourceKinds();
  61   APInt Buffers(NumProcResources, 0);
  62
  63   bool AllInOrderResources = true;
  64   bool AnyDispatchHazards = false;
  65   for (unsigned I = 0, E = SCDesc.NumWriteProcResEntries; I < E; ++I) {
  66     const MCWriteProcResEntry *PRE = STI.getWriteProcResBegin(&SCDesc) + I;
  67     const MCProcResourceDesc &PR = *SM.getProcResource(PRE->ProcResourceIdx);
  68     if (!PRE->Cycles) {
  69 #ifndef NDEBUG
  70       WithColor::warning()
  71           << "Ignoring invalid write of zero cycles on processor resource "
  72           << PR.Name << "\n";
  73       WithColor::note() << "found in scheduling class " << SCDesc.Name
  74                         << " (write index #" << I << ")\n";
  75 #endif
  76       continue;
  77     }
  78
  79     uint64_t Mask = ProcResourceMasks[PRE->ProcResourceIdx];
  80     if (PR.BufferSize < 0) {
  81       AllInOrderResources = false;
  82     } else {
  83       Buffers.setBit(getResourceStateIndex(Mask));
  84       AnyDispatchHazards |= (PR.BufferSize == 0);
  85       AllInOrderResources &= (PR.BufferSize <= 1);
  86     }
  87
  88     CycleSegment RCy(0, PRE->Cycles, false);
  89     Worklist.emplace_back(ResourcePlusCycles(Mask, ResourceUsage(RCy)));
  90     if (PR.SuperIdx) {
  91       uint64_t Super = ProcResourceMasks[PR.SuperIdx];
  92       SuperResources[Super] += PRE->Cycles;
  93     }
  94   }
  95
  96   ID.MustIssueImmediately = AllInOrderResources && AnyDispatchHazards;
  97
  98   // Sort elements by mask popcount, so that we prioritize resource units over
  99   // resource groups, and smaller groups over larger groups.
 100   sort(Worklist, [](const ResourcePlusCycles &A, const ResourcePlusCycles &B) {
 101     unsigned popcntA = countPopulation(A.first);
 102     unsigned popcntB = countPopulation(B.first);
 103     if (popcntA < popcntB)
 104       return true;
 105     if (popcntA > popcntB)
 106       return false;
 107     return A.first < B.first;
 108   });
 109
 110   uint64_t UsedResourceUnits = 0;
 111   uint64_t UsedResourceGroups = 0;
 112
 113   // Remove cycles contributed by smaller resources.
 114   for (unsigned I = 0, E = Worklist.size(); I < E; ++I) {
 115     ResourcePlusCycles &A = Worklist[I];
 116     if (!A.second.size()) {
 117       assert(countPopulation(A.first) > 1 && "Expected a group!");
 118       UsedResourceGroups |= PowerOf2Floor(A.first);
 119       continue;
 120     }
 121
 122     ID.Resources.emplace_back(A);
 123     uint64_t NormalizedMask = A.first;
 124     if (countPopulation(A.first) == 1) {
 125       UsedResourceUnits |= A.first;
 126     } else {
 127       // Remove the leading 1 from the resource group mask.
 128       NormalizedMask ^= PowerOf2Floor(NormalizedMask);
 129       UsedResourceGroups |= (A.first ^ NormalizedMask);
 130     }
 131
 132     for (unsigned J = I + 1; J < E; ++J) {
 133       ResourcePlusCycles &B = Worklist[J];
 134       if ((NormalizedMask & B.first) == NormalizedMask) {
 135         B.second.CS.subtract(A.second.size() - SuperResources[A.first]);
 136         if (countPopulation(B.first) > 1)
 137           B.second.NumUnits++;
 138       }
 139     }
 140   }
 141
 142   // A SchedWrite may specify a number of cycles in which a resource group
 143   // is reserved. For example (on target x86; cpu Haswell):
 144   //
 145   //  SchedWriteRes<[HWPort0, HWPort1, HWPort01]> {
 146   //    let ResourceCycles = [2, 2, 3];
 147   //  }
 148   //
 149   // This means:
 150   // Resource units HWPort0 and HWPort1 are both used for 2cy.
 151   // Resource group HWPort01 is the union of HWPort0 and HWPort1.
 152   // Since this write touches both HWPort0 and HWPort1 for 2cy, HWPort01
 153   // will not be usable for 2 entire cycles from instruction issue.
 154   //
 155   // On top of those 2cy, SchedWriteRes explicitly specifies an extra latency
 156   // of 3 cycles for HWPort01. This tool assumes that the 3cy latency is an
 157   // extra delay on top of the 2 cycles latency.
 158   // During those extra cycles, HWPort01 is not usable by other instructions.
 159   for (ResourcePlusCycles &RPC : ID.Resources) {
 160     if (countPopulation(RPC.first) > 1 && !RPC.second.isReserved()) {
 161       // Remove the leading 1 from the resource group mask.
 162       uint64_t Mask = RPC.first ^ PowerOf2Floor(RPC.first);
 163       if ((Mask & UsedResourceUnits) == Mask)
 164         RPC.second.setReserved();
 165     }
 166   }
 167
 168   // Identify extra buffers that are consumed through super resources.
 169   for (const std::pair<uint64_t, unsigned> &SR : SuperResources) {
 170     for (unsigned I = 1, E = NumProcResources; I < E; ++I) {
 171       const MCProcResourceDesc &PR = *SM.getProcResource(I);
 172       if (PR.BufferSize == -1)
 173         continue;
 174
 175       uint64_t Mask = ProcResourceMasks[I];
 176       if (Mask != SR.first && ((Mask & SR.first) == SR.first))
 177         Buffers.setBit(getResourceStateIndex(Mask));
 178     }
 179   }
 180
 181   ID.UsedBuffers = Buffers.getZExtValue();
 182   ID.UsedProcResUnits = UsedResourceUnits;
 183   ID.UsedProcResGroups = UsedResourceGroups;
 184
 185   LLVM_DEBUG({
 186     for (const std::pair<uint64_t, ResourceUsage> &R : ID.Resources)
 187       dbgs() << "\t\tResource Mask=" << format_hex(R.first, 16) << ", "
 188              << "Reserved=" << R.second.isReserved() << ", "
 189              << "#Units=" << R.second.NumUnits << ", "
 190              << "cy=" << R.second.size() << '\n';
 191     uint64_t BufferIDs = ID.UsedBuffers;
 192     while (BufferIDs) {
 193       uint64_t Current = BufferIDs & (-BufferIDs);
 194       dbgs() << "\t\tBuffer Mask=" << format_hex(Current, 16) << '\n';
 195       BufferIDs ^= Current;
 196     }
 197     dbgs() << "\t\t Used Units=" << format_hex(ID.UsedProcResUnits, 16) << '\n';
 198     dbgs() << "\t\tUsed Groups=" << format_hex(ID.UsedProcResGroups, 16)
 199            << '\n';
 200   });
 201 }
 202
 203 static void computeMaxLatency(InstrDesc &ID, const MCInstrDesc &MCDesc,
 204                               const MCSchedClassDesc &SCDesc,
 205                               const MCSubtargetInfo &STI) {
 206   if (MCDesc.isCall()) {
 207     // We cannot estimate how long this call will take.
 208     // Artificially set an arbitrarily high latency (100cy).
 209     ID.MaxLatency = 100U;
 210     return;
 211   }
 212
 213   int Latency = MCSchedModel::computeInstrLatency(STI, SCDesc);
 214   // If latency is unknown, then conservatively assume a MaxLatency of 100cy.
 215   ID.MaxLatency = Latency < 0 ? 100U : static_cast<unsigned>(Latency);
 216 }
 217
 218 static Error verifyOperands(const MCInstrDesc &MCDesc, const MCInst &MCI) {
 219   // Count register definitions, and skip non register operands in the process.
 220   unsigned I, E;
 221   unsigned NumExplicitDefs = MCDesc.getNumDefs();
 222   for (I = 0, E = MCI.getNumOperands(); NumExplicitDefs && I < E; ++I) {
 223     const MCOperand &Op = MCI.getOperand(I);
 224     if (Op.isReg())
 225       --NumExplicitDefs;
 226   }
 227
 228   if (NumExplicitDefs) {
 229     return make_error<InstructionError<MCInst>>(
 230         "Expected more register operand definitions.", MCI);
 231   }
 232
 233   if (MCDesc.hasOptionalDef()) {
 234     // Always assume that the optional definition is the last operand.
 235     const MCOperand &Op = MCI.getOperand(MCDesc.getNumOperands() - 1);
 236     if (I == MCI.getNumOperands() || !Op.isReg()) {
 237       std::string Message =
 238           "expected a register operand for an optional definition. Instruction "
 239           "has not been correctly analyzed.";
 240       return make_error<InstructionError<MCInst>>(Message, MCI);
 241     }
 242   }
 243
 244   return ErrorSuccess();
 245 }
 246
 247 void InstrBuilder::populateWrites(InstrDesc &ID, const MCInst &MCI,
 248                                   unsigned SchedClassID) {
 249   const MCInstrDesc &MCDesc = MCII.get(MCI.getOpcode());
 250   const MCSchedModel &SM = STI.getSchedModel();
 251   const MCSchedClassDesc &SCDesc = *SM.getSchedClassDesc(SchedClassID);
 252
 253   // Assumptions made by this algorithm:
 254   //  1. The number of explicit and implicit register definitions in a MCInst
 255   //     matches the number of explicit and implicit definitions according to
 256   //     the opcode descriptor (MCInstrDesc).
 257   //  2. Uses start at index #(MCDesc.getNumDefs()).
 258   //  3. There can only be a single optional register definition, an it is
 259   //     always the last operand of the sequence (excluding extra operands
 260   //     contributed by variadic opcodes).
 261   //
 262   // These assumptions work quite well for most out-of-order in-tree targets
 263   // like x86. This is mainly because the vast majority of instructions is
 264   // expanded to MCInst using a straightforward lowering logic that preserves
 265   // the ordering of the operands.
 266   //
 267   // About assumption 1.
 268   // The algorithm allows non-register operands between register operand
 269   // definitions. This helps to handle some special ARM instructions with
 270   // implicit operand increment (-mtriple=armv7):
 271   //
 272   // vld1.32  {d18, d19}, [r1]!  @ <MCInst #1463 VLD1q32wb_fixed
 273   //                             @  <MCOperand Reg:59>
 274   //                             @  <MCOperand Imm:0>     (!!)
 275   //                             @  <MCOperand Reg:67>
 276   //                             @  <MCOperand Imm:0>
 277   //                             @  <MCOperand Imm:14>
 278   //                             @  <MCOperand Reg:0>>
 279   //
 280   // MCDesc reports:
 281   //  6 explicit operands.
 282   //  1 optional definition
 283   //  2 explicit definitions (!!)
 284   //
 285   // The presence of an 'Imm' operand between the two register definitions
 286   // breaks the assumption that "register definitions are always at the
 287   // beginning of the operand sequence".
 288   //
 289   // To workaround this issue, this algorithm ignores (i.e. skips) any
 290   // non-register operands between register definitions.  The optional
 291   // definition is still at index #(NumOperands-1).
 292   //
 293   // According to assumption 2. register reads start at #(NumExplicitDefs-1).
 294   // That means, register R1 from the example is both read and written.
 295   unsigned NumExplicitDefs = MCDesc.getNumDefs();
 296   unsigned NumImplicitDefs = MCDesc.getNumImplicitDefs();
 297   unsigned NumWriteLatencyEntries = SCDesc.NumWriteLatencyEntries;
 298   unsigned TotalDefs = NumExplicitDefs + NumImplicitDefs;
 299   if (MCDesc.hasOptionalDef())
 300     TotalDefs++;
 301
 302   unsigned NumVariadicOps = MCI.getNumOperands() - MCDesc.getNumOperands();
 303   ID.Writes.resize(TotalDefs + NumVariadicOps);
 304   // Iterate over the operands list, and skip non-register operands.
 305   // The first NumExplictDefs register operands are expected to be register
 306   // definitions.
 307   unsigned CurrentDef = 0;
 308   unsigned i = 0;
 309   for (; i < MCI.getNumOperands() && CurrentDef < NumExplicitDefs; ++i) {
 310     const MCOperand &Op = MCI.getOperand(i);
 311     if (!Op.isReg())
 312       continue;
 313
 314     WriteDescriptor &Write = ID.Writes[CurrentDef];
 315     Write.OpIndex = i;
 316     if (CurrentDef < NumWriteLatencyEntries) {
 317       const MCWriteLatencyEntry &WLE =
 318           *STI.getWriteLatencyEntry(&SCDesc, CurrentDef);
 319       // Conservatively default to MaxLatency.
 320       Write.Latency =
 321           WLE.Cycles < 0 ? ID.MaxLatency : static_cast<unsigned>(WLE.Cycles);
 322       Write.SClassOrWriteResourceID = WLE.WriteResourceID;
 323     } else {
 324       // Assign a default latency for this write.
 325       Write.Latency = ID.MaxLatency;
 326       Write.SClassOrWriteResourceID = 0;
 327     }
 328     Write.IsOptionalDef = false;
 329     LLVM_DEBUG({
 330       dbgs() << "\t\t[Def]    OpIdx=" << Write.OpIndex
 331              << ", Latency=" << Write.Latency
 332              << ", WriteResourceID=" << Write.SClassOrWriteResourceID << '\n';
 333     });
 334     CurrentDef++;
 335   }
 336
 337   assert(CurrentDef == NumExplicitDefs &&
 338          "Expected more register operand definitions.");
 339   for (CurrentDef = 0; CurrentDef < NumImplicitDefs; ++CurrentDef) {
 340     unsigned Index = NumExplicitDefs + CurrentDef;
 341     WriteDescriptor &Write = ID.Writes[Index];
 342     Write.OpIndex = ~CurrentDef;
 343     Write.RegisterID = MCDesc.getImplicitDefs()[CurrentDef];
 344     if (Index < NumWriteLatencyEntries) {
 345       const MCWriteLatencyEntry &WLE =
 346           *STI.getWriteLatencyEntry(&SCDesc, Index);
 347       // Conservatively default to MaxLatency.
 348       Write.Latency =
 349           WLE.Cycles < 0 ? ID.MaxLatency : static_cast<unsigned>(WLE.Cycles);
 350       Write.SClassOrWriteResourceID = WLE.WriteResourceID;
 351     } else {
 352       // Assign a default latency for this write.
 353       Write.Latency = ID.MaxLatency;
 354       Write.SClassOrWriteResourceID = 0;
 355     }
 356
 357     Write.IsOptionalDef = false;
 358     assert(Write.RegisterID != 0 && "Expected a valid phys register!");
 359     LLVM_DEBUG({
 360       dbgs() << "\t\t[Def][I] OpIdx=" << ~Write.OpIndex
 361              << ", PhysReg=" << MRI.getName(Write.RegisterID)
 362              << ", Latency=" << Write.Latency
 363              << ", WriteResourceID=" << Write.SClassOrWriteResourceID << '\n';
 364     });
 365   }
 366
 367   if (MCDesc.hasOptionalDef()) {
 368     WriteDescriptor &Write = ID.Writes[NumExplicitDefs + NumImplicitDefs];
 369     Write.OpIndex = MCDesc.getNumOperands() - 1;
 370     // Assign a default latency for this write.
 371     Write.Latency = ID.MaxLatency;
 372     Write.SClassOrWriteResourceID = 0;
 373     Write.IsOptionalDef = true;
 374     LLVM_DEBUG({
 375       dbgs() << "\t\t[Def][O] OpIdx=" << Write.OpIndex
 376              << ", Latency=" << Write.Latency
 377              << ", WriteResourceID=" << Write.SClassOrWriteResourceID << '\n';
 378     });
 379   }
 380
 381   if (!NumVariadicOps)
 382     return;
 383
 384   // FIXME: if an instruction opcode is flagged 'mayStore', and it has no
 385   // "unmodeledSideEffects', then this logic optimistically assumes that any
 386   // extra register operands in the variadic sequence is not a register
 387   // definition.
 388   //
 389   // Otherwise, we conservatively assume that any register operand from the
 390   // variadic sequence is both a register read and a register write.
 391   bool AssumeUsesOnly = MCDesc.mayStore() && !MCDesc.mayLoad() &&
 392                         !MCDesc.hasUnmodeledSideEffects();
 393   CurrentDef = NumExplicitDefs + NumImplicitDefs + MCDesc.hasOptionalDef();
 394   for (unsigned I = 0, OpIndex = MCDesc.getNumOperands();
 395        I < NumVariadicOps && !AssumeUsesOnly; ++I, ++OpIndex) {
 396     const MCOperand &Op = MCI.getOperand(OpIndex);
 397     if (!Op.isReg())
 398       continue;
 399
 400     WriteDescriptor &Write = ID.Writes[CurrentDef];
 401     Write.OpIndex = OpIndex;
 402     // Assign a default latency for this write.
 403     Write.Latency = ID.MaxLatency;
 404     Write.SClassOrWriteResourceID = 0;
 405     Write.IsOptionalDef = false;
 406     ++CurrentDef;
 407     LLVM_DEBUG({
 408       dbgs() << "\t\t[Def][V] OpIdx=" << Write.OpIndex
 409              << ", Latency=" << Write.Latency
 410              << ", WriteResourceID=" << Write.SClassOrWriteResourceID << '\n';
 411     });
 412   }
 413
 414   ID.Writes.resize(CurrentDef);
 415 }
 416
 417 void InstrBuilder::populateReads(InstrDesc &ID, const MCInst &MCI,
 418                                  unsigned SchedClassID) {
 419   const MCInstrDesc &MCDesc = MCII.get(MCI.getOpcode());
 420   unsigned NumExplicitUses = MCDesc.getNumOperands() - MCDesc.getNumDefs();
 421   unsigned NumImplicitUses = MCDesc.getNumImplicitUses();
 422   // Remove the optional definition.
 423   if (MCDesc.hasOptionalDef())
 424     --NumExplicitUses;
 425   unsigned NumVariadicOps = MCI.getNumOperands() - MCDesc.getNumOperands();
 426   unsigned TotalUses = NumExplicitUses + NumImplicitUses + NumVariadicOps;
 427   ID.Reads.resize(TotalUses);
 428   unsigned CurrentUse = 0;
 429   for (unsigned I = 0, OpIndex = MCDesc.getNumDefs(); I < NumExplicitUses;
 430        ++I, ++OpIndex) {
 431     const MCOperand &Op = MCI.getOperand(OpIndex);
 432     if (!Op.isReg())
 433       continue;
 434
 435     ReadDescriptor &Read = ID.Reads[CurrentUse];
 436     Read.OpIndex = OpIndex;
 437     Read.UseIndex = I;
 438     Read.SchedClassID = SchedClassID;
 439     ++CurrentUse;
 440     LLVM_DEBUG(dbgs() << "\t\t[Use]    OpIdx=" << Read.OpIndex
 441                       << ", UseIndex=" << Read.UseIndex << '\n');
 442   }
 443
 444   // For the purpose of ReadAdvance, implicit uses come directly after explicit
 445   // uses. The "UseIndex" must be updated according to that implicit layout.
 446   for (unsigned I = 0; I < NumImplicitUses; ++I) {
 447     ReadDescriptor &Read = ID.Reads[CurrentUse + I];
 448     Read.OpIndex = ~I;
 449     Read.UseIndex = NumExplicitUses + I;
 450     Read.RegisterID = MCDesc.getImplicitUses()[I];
 451     Read.SchedClassID = SchedClassID;
 452     LLVM_DEBUG(dbgs() << "\t\t[Use][I] OpIdx=" << ~Read.OpIndex
 453                       << ", UseIndex=" << Read.UseIndex << ", RegisterID="
 454                       << MRI.getName(Read.RegisterID) << '\n');
 455   }
 456
 457   CurrentUse += NumImplicitUses;
 458
 459   // FIXME: If an instruction opcode is marked as 'mayLoad', and it has no
 460   // "unmodeledSideEffects", then this logic optimistically assumes that any
 461   // extra register operands in the variadic sequence are not register
 462   // definition.
 463
 464   bool AssumeDefsOnly = !MCDesc.mayStore() && MCDesc.mayLoad() &&
 465                         !MCDesc.hasUnmodeledSideEffects();
 466   for (unsigned I = 0, OpIndex = MCDesc.getNumOperands();
 467        I < NumVariadicOps && !AssumeDefsOnly; ++I, ++OpIndex) {
 468     const MCOperand &Op = MCI.getOperand(OpIndex);
 469     if (!Op.isReg())
 470       continue;
 471
 472     ReadDescriptor &Read = ID.Reads[CurrentUse];
 473     Read.OpIndex = OpIndex;
 474     Read.UseIndex = NumExplicitUses + NumImplicitUses + I;
 475     Read.SchedClassID = SchedClassID;
 476     ++CurrentUse;
 477     LLVM_DEBUG(dbgs() << "\t\t[Use][V] OpIdx=" << Read.OpIndex
 478                       << ", UseIndex=" << Read.UseIndex << '\n');
 479   }
 480
 481   ID.Reads.resize(CurrentUse);
 482 }
 483
 484 Error InstrBuilder::verifyInstrDesc(const InstrDesc &ID,
 485                                     const MCInst &MCI) const {
 486   if (ID.NumMicroOps != 0)
 487     return ErrorSuccess();
 488
 489   bool UsesMemory = ID.MayLoad || ID.MayStore;
 490   bool UsesBuffers = ID.UsedBuffers;
 491   bool UsesResources = !ID.Resources.empty();
 492   if (!UsesMemory && !UsesBuffers && !UsesResources)
 493     return ErrorSuccess();
 494
 495   StringRef Message;
 496   if (UsesMemory) {
 497     Message = "found an inconsistent instruction that decodes "
 498               "into zero opcodes and that consumes load/store "
 499               "unit resources.";
 500   } else {
 501     Message = "found an inconsistent instruction that decodes "
 502               "to zero opcodes and that consumes scheduler "
 503               "resources.";
 504   }
 505
 506   return make_error<InstructionError<MCInst>>(Message, MCI);
 507 }
 508
 509 Expected<const InstrDesc &>
 510 InstrBuilder::createInstrDescImpl(const MCInst &MCI) {
 511   assert(STI.getSchedModel().hasInstrSchedModel() &&
 512          "Itineraries are not yet supported!");
 513
 514   // Obtain the instruction descriptor from the opcode.
 515   unsigned short Opcode = MCI.getOpcode();
 516   const MCInstrDesc &MCDesc = MCII.get(Opcode);
 517   const MCSchedModel &SM = STI.getSchedModel();
 518
 519   // Then obtain the scheduling class information from the instruction.
 520   unsigned SchedClassID = MCDesc.getSchedClass();
 521   bool IsVariant = SM.getSchedClassDesc(SchedClassID)->isVariant();
 522
 523   // Try to solve variant scheduling classes.
 524   if (IsVariant) {
 525     unsigned CPUID = SM.getProcessorID();
 526     while (SchedClassID && SM.getSchedClassDesc(SchedClassID)->isVariant())
 527       SchedClassID = STI.resolveVariantSchedClass(SchedClassID, &MCI, CPUID);
 528
 529     if (!SchedClassID) {
 530       return make_error<InstructionError<MCInst>>(
 531           "unable to resolve scheduling class for write variant.", MCI);
 532     }
 533   }
 534
 535   // Check if this instruction is supported. Otherwise, report an error.
 536   const MCSchedClassDesc &SCDesc = *SM.getSchedClassDesc(SchedClassID);
 537   if (SCDesc.NumMicroOps == MCSchedClassDesc::InvalidNumMicroOps) {
 538     return make_error<InstructionError<MCInst>>(
 539         "found an unsupported instruction in the input assembly sequence.",
 540         MCI);
 541   }
 542
 543   LLVM_DEBUG(dbgs() << "\n\t\tOpcode Name= " << MCII.getName(Opcode) << '\n');
 544   LLVM_DEBUG(dbgs() << "\t\tSchedClassID=" << SchedClassID << '\n');
 545
 546   // Create a new empty descriptor.
 547   std::unique_ptr<InstrDesc> ID = std::make_unique<InstrDesc>();
 548   ID->NumMicroOps = SCDesc.NumMicroOps;
 549   ID->SchedClassID = SchedClassID;
 550
 551   if (MCDesc.isCall() && FirstCallInst) {
 552     // We don't correctly model calls.
 553     WithColor::warning() << "found a call in the input assembly sequence.\n";
 554     WithColor::note() << "call instructions are not correctly modeled. "
 555                       << "Assume a latency of 100cy.\n";
 556     FirstCallInst = false;
 557   }
 558
 559   if (MCDesc.isReturn() && FirstReturnInst) {
 560     WithColor::warning() << "found a return instruction in the input"
 561                          << " assembly sequence.\n";
 562     WithColor::note() << "program counter updates are ignored.\n";
 563     FirstReturnInst = false;
 564   }
 565
 566   ID->MayLoad = MCDesc.mayLoad();
 567   ID->MayStore = MCDesc.mayStore();
 568   ID->HasSideEffects = MCDesc.hasUnmodeledSideEffects();
 569   ID->BeginGroup = SCDesc.BeginGroup;
 570   ID->EndGroup = SCDesc.EndGroup;
 571
 572   initializeUsedResources(*ID, SCDesc, STI, ProcResourceMasks);
 573   computeMaxLatency(*ID, MCDesc, SCDesc, STI);
 574
 575   if (Error Err = verifyOperands(MCDesc, MCI))
 576     return std::move(Err);
 577
 578   populateWrites(*ID, MCI, SchedClassID);
 579   populateReads(*ID, MCI, SchedClassID);
 580
 581   LLVM_DEBUG(dbgs() << "\t\tMaxLatency=" << ID->MaxLatency << '\n');
 582   LLVM_DEBUG(dbgs() << "\t\tNumMicroOps=" << ID->NumMicroOps << '\n');
 583
 584   // Sanity check on the instruction descriptor.
 585   if (Error Err = verifyInstrDesc(*ID, MCI))
 586     return std::move(Err);
 587
 588   // Now add the new descriptor.
 589   bool IsVariadic = MCDesc.isVariadic();
 590   if (!IsVariadic && !IsVariant) {
 591     Descriptors[MCI.getOpcode()] = std::move(ID);
 592     return *Descriptors[MCI.getOpcode()];
 593   }
 594
 595   VariantDescriptors[&MCI] = std::move(ID);
 596   return *VariantDescriptors[&MCI];
 597 }
 598
 599 Expected<const InstrDesc &>
 600 InstrBuilder::getOrCreateInstrDesc(const MCInst &MCI) {
 601   if (Descriptors.find_as(MCI.getOpcode()) != Descriptors.end())
 602     return *Descriptors[MCI.getOpcode()];
 603
 604   if (VariantDescriptors.find(&MCI) != VariantDescriptors.end())
 605     return *VariantDescriptors[&MCI];
 606
 607   return createInstrDescImpl(MCI);
 608 }
 609
 610 Expected<std::unique_ptr<Instruction>>
 611 InstrBuilder::createInstruction(const MCInst &MCI) {
 612   Expected<const InstrDesc &> DescOrErr = getOrCreateInstrDesc(MCI);
 613   if (!DescOrErr)
 614     return DescOrErr.takeError();
 615   const InstrDesc &D = *DescOrErr;
 616   std::unique_ptr<Instruction> NewIS = std::make_unique<Instruction>(D);
 617
 618   // Check if this is a dependency breaking instruction.
 619   APInt Mask;
 620
 621   bool IsZeroIdiom = false;
 622   bool IsDepBreaking = false;
 623   if (MCIA) {
 624     unsigned ProcID = STI.getSchedModel().getProcessorID();
 625     IsZeroIdiom = MCIA->isZeroIdiom(MCI, Mask, ProcID);
 626     IsDepBreaking =
 627         IsZeroIdiom || MCIA->isDependencyBreaking(MCI, Mask, ProcID);
 628     if (MCIA->isOptimizableRegisterMove(MCI, ProcID))
 629       NewIS->setOptimizableMove();
 630   }
 631
 632   // Initialize Reads first.
 633   for (const ReadDescriptor &RD : D.Reads) {
 634     int RegID = -1;
 635     if (!RD.isImplicitRead()) {
 636       // explicit read.
 637       const MCOperand &Op = MCI.getOperand(RD.OpIndex);
 638       // Skip non-register operands.
 639       if (!Op.isReg())
 640         continue;
 641       RegID = Op.getReg();
 642     } else {
 643       // Implicit read.
 644       RegID = RD.RegisterID;
 645     }
 646
 647     // Skip invalid register operands.
 648     if (!RegID)
 649       continue;
 650
 651     // Okay, this is a register operand. Create a ReadState for it.
 652     assert(RegID > 0 && "Invalid register ID found!");
 653     NewIS->getUses().emplace_back(RD, RegID);
 654     ReadState &RS = NewIS->getUses().back();
 655
 656     if (IsDepBreaking) {
 657       // A mask of all zeroes means: explicit input operands are not
 658       // independent.
 659       if (Mask.isNullValue()) {
 660         if (!RD.isImplicitRead())
 661           RS.setIndependentFromDef();
 662       } else {
 663         // Check if this register operand is independent according to `Mask`.
 664         // Note that Mask may not have enough bits to describe all explicit and
 665         // implicit input operands. If this register operand doesn't have a
 666         // corresponding bit in Mask, then conservatively assume that it is
 667         // dependent.
 668         if (Mask.getBitWidth() > RD.UseIndex) {
 669           // Okay. This map describe register use `RD.UseIndex`.
 670           if (Mask[RD.UseIndex])
 671             RS.setIndependentFromDef();
 672         }
 673       }
 674     }
 675   }
 676
 677   // Early exit if there are no writes.
 678   if (D.Writes.empty())
 679     return std::move(NewIS);
 680
 681   // Track register writes that implicitly clear the upper portion of the
 682   // underlying super-registers using an APInt.
 683   APInt WriteMask(D.Writes.size(), 0);
 684
 685   // Now query the MCInstrAnalysis object to obtain information about which
 686   // register writes implicitly clear the upper portion of a super-register.
 687   if (MCIA)
 688     MCIA->clearsSuperRegisters(MRI, MCI, WriteMask);
 689
 690   // Initialize writes.
 691   unsigned WriteIndex = 0;
 692   for (const WriteDescriptor &WD : D.Writes) {
 693     unsigned RegID = WD.isImplicitWrite() ? WD.RegisterID
 694                                           : MCI.getOperand(WD.OpIndex).getReg();
 695     // Check if this is a optional definition that references NoReg.
 696     if (WD.IsOptionalDef && !RegID) {
 697       ++WriteIndex;
 698       continue;
 699     }
 700
 701     assert(RegID && "Expected a valid register ID!");
 702     NewIS->getDefs().emplace_back(WD, RegID,
 703                                   /* ClearsSuperRegs */ WriteMask[WriteIndex],
 704                                   /* WritesZero */ IsZeroIdiom);
 705     ++WriteIndex;
 706   }
 707
 708   return std::move(NewIS);
 709 }
 710 } // namespace mca
 711 } // namespace llvm