17330f729Sjoerg //===- DFAPacketizerEmitter.cpp - Packetization DFA for a VLIW machine ----===//
27330f729Sjoerg //
37330f729Sjoerg // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
47330f729Sjoerg // See https://llvm.org/LICENSE.txt for license information.
57330f729Sjoerg // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
67330f729Sjoerg //
77330f729Sjoerg //===----------------------------------------------------------------------===//
87330f729Sjoerg //
97330f729Sjoerg // This class parses the Schedule.td file and produces an API that can be used
107330f729Sjoerg // to reason about whether an instruction can be added to a packet on a VLIW
117330f729Sjoerg // architecture. The class internally generates a deterministic finite
127330f729Sjoerg // automaton (DFA) that models all possible mappings of machine instructions
137330f729Sjoerg // to functional units as instructions are added to a packet.
147330f729Sjoerg //
157330f729Sjoerg //===----------------------------------------------------------------------===//
167330f729Sjoerg
177330f729Sjoerg #define DEBUG_TYPE "dfa-emitter"
187330f729Sjoerg
19*82d56013Sjoerg #include "CodeGenSchedule.h"
207330f729Sjoerg #include "CodeGenTarget.h"
217330f729Sjoerg #include "DFAEmitter.h"
227330f729Sjoerg #include "llvm/ADT/DenseSet.h"
237330f729Sjoerg #include "llvm/ADT/SmallVector.h"
247330f729Sjoerg #include "llvm/ADT/StringExtras.h"
257330f729Sjoerg #include "llvm/Support/Debug.h"
267330f729Sjoerg #include "llvm/Support/raw_ostream.h"
27*82d56013Sjoerg #include "llvm/TableGen/Record.h"
28*82d56013Sjoerg #include "llvm/TableGen/TableGenBackend.h"
297330f729Sjoerg #include <cassert>
307330f729Sjoerg #include <cstdint>
317330f729Sjoerg #include <map>
327330f729Sjoerg #include <set>
337330f729Sjoerg #include <string>
347330f729Sjoerg #include <unordered_map>
357330f729Sjoerg #include <vector>
367330f729Sjoerg
377330f729Sjoerg using namespace llvm;
387330f729Sjoerg
39*82d56013Sjoerg // We use a uint64_t to represent a resource bitmask.
40*82d56013Sjoerg #define DFA_MAX_RESOURCES 64
417330f729Sjoerg
427330f729Sjoerg namespace {
43*82d56013Sjoerg using ResourceVector = SmallVector<uint64_t, 4>;
447330f729Sjoerg
45*82d56013Sjoerg struct ScheduleClass {
46*82d56013Sjoerg /// The parent itinerary index (processor model ID).
47*82d56013Sjoerg unsigned ItineraryID;
487330f729Sjoerg
49*82d56013Sjoerg /// Index within this itinerary of the schedule class.
50*82d56013Sjoerg unsigned Idx;
517330f729Sjoerg
52*82d56013Sjoerg /// The index within the uniqued set of required resources of Resources.
53*82d56013Sjoerg unsigned ResourcesIdx;
547330f729Sjoerg
55*82d56013Sjoerg /// Conjunctive list of resource requirements:
56*82d56013Sjoerg /// {a|b, b|c} => (a OR b) AND (b or c).
57*82d56013Sjoerg /// Resources are unique across all itineraries.
58*82d56013Sjoerg ResourceVector Resources;
59*82d56013Sjoerg };
607330f729Sjoerg
61*82d56013Sjoerg // Generates and prints out the DFA for resource tracking.
627330f729Sjoerg class DFAPacketizerEmitter {
637330f729Sjoerg private:
647330f729Sjoerg std::string TargetName;
657330f729Sjoerg RecordKeeper &Records;
667330f729Sjoerg
67*82d56013Sjoerg UniqueVector<ResourceVector> UniqueResources;
68*82d56013Sjoerg std::vector<ScheduleClass> ScheduleClasses;
69*82d56013Sjoerg std::map<std::string, uint64_t> FUNameToBitsMap;
70*82d56013Sjoerg std::map<unsigned, uint64_t> ComboBitToBitsMap;
71*82d56013Sjoerg
727330f729Sjoerg public:
737330f729Sjoerg DFAPacketizerEmitter(RecordKeeper &R);
747330f729Sjoerg
75*82d56013Sjoerg // Construct a map of function unit names to bits.
76*82d56013Sjoerg int collectAllFuncUnits(
77*82d56013Sjoerg ArrayRef<const CodeGenProcModel *> ProcModels);
787330f729Sjoerg
79*82d56013Sjoerg // Construct a map from a combo function unit bit to the bits of all included
80*82d56013Sjoerg // functional units.
81*82d56013Sjoerg int collectAllComboFuncs(ArrayRef<Record *> ComboFuncList);
827330f729Sjoerg
83*82d56013Sjoerg ResourceVector getResourcesForItinerary(Record *Itinerary);
84*82d56013Sjoerg void createScheduleClasses(unsigned ItineraryIdx, const RecVec &Itineraries);
857330f729Sjoerg
867330f729Sjoerg // Emit code for a subset of itineraries.
877330f729Sjoerg void emitForItineraries(raw_ostream &OS,
88*82d56013Sjoerg std::vector<const CodeGenProcModel *> &ProcItinList,
897330f729Sjoerg std::string DFAName);
907330f729Sjoerg
917330f729Sjoerg void run(raw_ostream &OS);
927330f729Sjoerg };
937330f729Sjoerg } // end anonymous namespace
947330f729Sjoerg
DFAPacketizerEmitter(RecordKeeper & R)95*82d56013Sjoerg DFAPacketizerEmitter::DFAPacketizerEmitter(RecordKeeper &R)
96*82d56013Sjoerg : TargetName(std::string(CodeGenTarget(R).getName())), Records(R) {}
977330f729Sjoerg
collectAllFuncUnits(ArrayRef<const CodeGenProcModel * > ProcModels)987330f729Sjoerg int DFAPacketizerEmitter::collectAllFuncUnits(
99*82d56013Sjoerg ArrayRef<const CodeGenProcModel *> ProcModels) {
1007330f729Sjoerg LLVM_DEBUG(dbgs() << "-------------------------------------------------------"
1017330f729Sjoerg "----------------------\n");
1027330f729Sjoerg LLVM_DEBUG(dbgs() << "collectAllFuncUnits");
103*82d56013Sjoerg LLVM_DEBUG(dbgs() << " (" << ProcModels.size() << " itineraries)\n");
104*82d56013Sjoerg
105*82d56013Sjoerg std::set<Record *> ProcItinList;
106*82d56013Sjoerg for (const CodeGenProcModel *Model : ProcModels)
107*82d56013Sjoerg ProcItinList.insert(Model->ItinsDef);
1087330f729Sjoerg
1097330f729Sjoerg int totalFUs = 0;
1107330f729Sjoerg // Parse functional units for all the itineraries.
111*82d56013Sjoerg for (Record *Proc : ProcItinList) {
1127330f729Sjoerg std::vector<Record *> FUs = Proc->getValueAsListOfDefs("FU");
1137330f729Sjoerg
114*82d56013Sjoerg LLVM_DEBUG(dbgs() << " FU:"
115*82d56013Sjoerg << " (" << FUs.size() << " FUs) " << Proc->getName());
1167330f729Sjoerg
1177330f729Sjoerg // Convert macros to bits for each stage.
1187330f729Sjoerg unsigned numFUs = FUs.size();
1197330f729Sjoerg for (unsigned j = 0; j < numFUs; ++j) {
1207330f729Sjoerg assert((j < DFA_MAX_RESOURCES) &&
1217330f729Sjoerg "Exceeded maximum number of representable resources");
122*82d56013Sjoerg uint64_t FuncResources = 1ULL << j;
123*82d56013Sjoerg FUNameToBitsMap[std::string(FUs[j]->getName())] = FuncResources;
1247330f729Sjoerg LLVM_DEBUG(dbgs() << " " << FUs[j]->getName() << ":0x"
1257330f729Sjoerg << Twine::utohexstr(FuncResources));
1267330f729Sjoerg }
1277330f729Sjoerg totalFUs += numFUs;
1287330f729Sjoerg LLVM_DEBUG(dbgs() << "\n");
1297330f729Sjoerg }
1307330f729Sjoerg return totalFUs;
1317330f729Sjoerg }
1327330f729Sjoerg
collectAllComboFuncs(ArrayRef<Record * > ComboFuncList)133*82d56013Sjoerg int DFAPacketizerEmitter::collectAllComboFuncs(ArrayRef<Record *> ComboFuncList) {
1347330f729Sjoerg LLVM_DEBUG(dbgs() << "-------------------------------------------------------"
1357330f729Sjoerg "----------------------\n");
1367330f729Sjoerg LLVM_DEBUG(dbgs() << "collectAllComboFuncs");
1377330f729Sjoerg LLVM_DEBUG(dbgs() << " (" << ComboFuncList.size() << " sets)\n");
1387330f729Sjoerg
1397330f729Sjoerg int numCombos = 0;
1407330f729Sjoerg for (unsigned i = 0, N = ComboFuncList.size(); i < N; ++i) {
1417330f729Sjoerg Record *Func = ComboFuncList[i];
1427330f729Sjoerg std::vector<Record *> FUs = Func->getValueAsListOfDefs("CFD");
1437330f729Sjoerg
1447330f729Sjoerg LLVM_DEBUG(dbgs() << " CFD:" << i << " (" << FUs.size() << " combo FUs) "
1457330f729Sjoerg << Func->getName() << "\n");
1467330f729Sjoerg
1477330f729Sjoerg // Convert macros to bits for each stage.
1487330f729Sjoerg for (unsigned j = 0, N = FUs.size(); j < N; ++j) {
1497330f729Sjoerg assert((j < DFA_MAX_RESOURCES) &&
1507330f729Sjoerg "Exceeded maximum number of DFA resources");
1517330f729Sjoerg Record *FuncData = FUs[j];
1527330f729Sjoerg Record *ComboFunc = FuncData->getValueAsDef("TheComboFunc");
1537330f729Sjoerg const std::vector<Record *> &FuncList =
1547330f729Sjoerg FuncData->getValueAsListOfDefs("FuncList");
155*82d56013Sjoerg const std::string &ComboFuncName = std::string(ComboFunc->getName());
156*82d56013Sjoerg uint64_t ComboBit = FUNameToBitsMap[ComboFuncName];
157*82d56013Sjoerg uint64_t ComboResources = ComboBit;
1587330f729Sjoerg LLVM_DEBUG(dbgs() << " combo: " << ComboFuncName << ":0x"
1597330f729Sjoerg << Twine::utohexstr(ComboResources) << "\n");
160*82d56013Sjoerg for (auto *K : FuncList) {
161*82d56013Sjoerg std::string FuncName = std::string(K->getName());
162*82d56013Sjoerg uint64_t FuncResources = FUNameToBitsMap[FuncName];
1637330f729Sjoerg LLVM_DEBUG(dbgs() << " " << FuncName << ":0x"
1647330f729Sjoerg << Twine::utohexstr(FuncResources) << "\n");
1657330f729Sjoerg ComboResources |= FuncResources;
1667330f729Sjoerg }
1677330f729Sjoerg ComboBitToBitsMap[ComboBit] = ComboResources;
1687330f729Sjoerg numCombos++;
1697330f729Sjoerg LLVM_DEBUG(dbgs() << " => combo bits: " << ComboFuncName << ":0x"
1707330f729Sjoerg << Twine::utohexstr(ComboBit) << " = 0x"
1717330f729Sjoerg << Twine::utohexstr(ComboResources) << "\n");
1727330f729Sjoerg }
1737330f729Sjoerg }
1747330f729Sjoerg return numCombos;
1757330f729Sjoerg }
1767330f729Sjoerg
177*82d56013Sjoerg ResourceVector
getResourcesForItinerary(Record * Itinerary)178*82d56013Sjoerg DFAPacketizerEmitter::getResourcesForItinerary(Record *Itinerary) {
179*82d56013Sjoerg ResourceVector Resources;
180*82d56013Sjoerg assert(Itinerary);
181*82d56013Sjoerg for (Record *StageDef : Itinerary->getValueAsListOfDefs("Stages")) {
182*82d56013Sjoerg uint64_t StageResources = 0;
183*82d56013Sjoerg for (Record *Unit : StageDef->getValueAsListOfDefs("Units")) {
184*82d56013Sjoerg StageResources |= FUNameToBitsMap[std::string(Unit->getName())];
185*82d56013Sjoerg }
186*82d56013Sjoerg if (StageResources != 0)
187*82d56013Sjoerg Resources.push_back(StageResources);
188*82d56013Sjoerg }
189*82d56013Sjoerg return Resources;
1907330f729Sjoerg }
1917330f729Sjoerg
createScheduleClasses(unsigned ItineraryIdx,const RecVec & Itineraries)192*82d56013Sjoerg void DFAPacketizerEmitter::createScheduleClasses(unsigned ItineraryIdx,
193*82d56013Sjoerg const RecVec &Itineraries) {
194*82d56013Sjoerg unsigned Idx = 0;
195*82d56013Sjoerg for (Record *Itinerary : Itineraries) {
196*82d56013Sjoerg if (!Itinerary) {
197*82d56013Sjoerg ScheduleClasses.push_back({ItineraryIdx, Idx++, 0, ResourceVector{}});
198*82d56013Sjoerg continue;
1997330f729Sjoerg }
200*82d56013Sjoerg ResourceVector Resources = getResourcesForItinerary(Itinerary);
201*82d56013Sjoerg ScheduleClasses.push_back(
202*82d56013Sjoerg {ItineraryIdx, Idx++, UniqueResources.insert(Resources), Resources});
2037330f729Sjoerg }
2047330f729Sjoerg }
2057330f729Sjoerg
2067330f729Sjoerg //
2077330f729Sjoerg // Run the worklist algorithm to generate the DFA.
2087330f729Sjoerg //
run(raw_ostream & OS)2097330f729Sjoerg void DFAPacketizerEmitter::run(raw_ostream &OS) {
2107330f729Sjoerg OS << "\n"
2117330f729Sjoerg << "#include \"llvm/CodeGen/DFAPacketizer.h\"\n";
2127330f729Sjoerg OS << "namespace llvm {\n";
2137330f729Sjoerg
214*82d56013Sjoerg CodeGenTarget CGT(Records);
215*82d56013Sjoerg CodeGenSchedModels CGS(Records, CGT);
2167330f729Sjoerg
217*82d56013Sjoerg std::unordered_map<std::string, std::vector<const CodeGenProcModel *>>
218*82d56013Sjoerg ItinsByNamespace;
219*82d56013Sjoerg for (const CodeGenProcModel &ProcModel : CGS.procModels()) {
220*82d56013Sjoerg if (ProcModel.hasItineraries()) {
221*82d56013Sjoerg auto NS = ProcModel.ItinsDef->getValueAsString("PacketizerNamespace");
222*82d56013Sjoerg ItinsByNamespace[std::string(NS)].push_back(&ProcModel);
223*82d56013Sjoerg }
224*82d56013Sjoerg }
2257330f729Sjoerg
2267330f729Sjoerg for (auto &KV : ItinsByNamespace)
2277330f729Sjoerg emitForItineraries(OS, KV.second, KV.first);
2287330f729Sjoerg OS << "} // end namespace llvm\n";
2297330f729Sjoerg }
2307330f729Sjoerg
emitForItineraries(raw_ostream & OS,std::vector<const CodeGenProcModel * > & ProcModels,std::string DFAName)2317330f729Sjoerg void DFAPacketizerEmitter::emitForItineraries(
232*82d56013Sjoerg raw_ostream &OS, std::vector<const CodeGenProcModel *> &ProcModels,
2337330f729Sjoerg std::string DFAName) {
234*82d56013Sjoerg OS << "} // end namespace llvm\n\n";
235*82d56013Sjoerg OS << "namespace {\n";
236*82d56013Sjoerg collectAllFuncUnits(ProcModels);
237*82d56013Sjoerg collectAllComboFuncs(Records.getAllDerivedDefinitions("ComboFuncUnits"));
2387330f729Sjoerg
2397330f729Sjoerg // Collect the itineraries.
240*82d56013Sjoerg DenseMap<const CodeGenProcModel *, unsigned> ProcModelStartIdx;
241*82d56013Sjoerg for (const CodeGenProcModel *Model : ProcModels) {
242*82d56013Sjoerg assert(Model->hasItineraries());
243*82d56013Sjoerg ProcModelStartIdx[Model] = ScheduleClasses.size();
244*82d56013Sjoerg createScheduleClasses(Model->Index, Model->ItinDefList);
2457330f729Sjoerg }
2467330f729Sjoerg
247*82d56013Sjoerg // Output the mapping from ScheduleClass to ResourcesIdx.
248*82d56013Sjoerg unsigned Idx = 0;
249*82d56013Sjoerg OS << "constexpr unsigned " << TargetName << DFAName
250*82d56013Sjoerg << "ResourceIndices[] = {";
251*82d56013Sjoerg for (const ScheduleClass &SC : ScheduleClasses) {
252*82d56013Sjoerg if (Idx++ % 32 == 0)
253*82d56013Sjoerg OS << "\n ";
254*82d56013Sjoerg OS << SC.ResourcesIdx << ", ";
255*82d56013Sjoerg }
256*82d56013Sjoerg OS << "\n};\n\n";
257*82d56013Sjoerg
258*82d56013Sjoerg // And the mapping from Itinerary index into the previous table.
259*82d56013Sjoerg OS << "constexpr unsigned " << TargetName << DFAName
260*82d56013Sjoerg << "ProcResourceIndexStart[] = {\n";
261*82d56013Sjoerg OS << " 0, // NoSchedModel\n";
262*82d56013Sjoerg for (const CodeGenProcModel *Model : ProcModels) {
263*82d56013Sjoerg OS << " " << ProcModelStartIdx[Model] << ", // " << Model->ModelName
264*82d56013Sjoerg << "\n";
265*82d56013Sjoerg }
266*82d56013Sjoerg OS << " " << ScheduleClasses.size() << "\n};\n\n";
267*82d56013Sjoerg
2687330f729Sjoerg // The type of a state in the nondeterministic automaton we're defining.
269*82d56013Sjoerg using NfaStateTy = uint64_t;
2707330f729Sjoerg
2717330f729Sjoerg // Given a resource state, return all resource states by applying
2727330f729Sjoerg // InsnClass.
273*82d56013Sjoerg auto applyInsnClass = [&](const ResourceVector &InsnClass,
274*82d56013Sjoerg NfaStateTy State) -> std::deque<NfaStateTy> {
275*82d56013Sjoerg std::deque<NfaStateTy> V(1, State);
2767330f729Sjoerg // Apply every stage in the class individually.
277*82d56013Sjoerg for (NfaStateTy Stage : InsnClass) {
2787330f729Sjoerg // Apply this stage to every existing member of V in turn.
2797330f729Sjoerg size_t Sz = V.size();
2807330f729Sjoerg for (unsigned I = 0; I < Sz; ++I) {
281*82d56013Sjoerg NfaStateTy S = V.front();
2827330f729Sjoerg V.pop_front();
2837330f729Sjoerg
2847330f729Sjoerg // For this stage, state combination, try all possible resources.
2857330f729Sjoerg for (unsigned J = 0; J < DFA_MAX_RESOURCES; ++J) {
286*82d56013Sjoerg NfaStateTy ResourceMask = 1ULL << J;
2877330f729Sjoerg if ((ResourceMask & Stage) == 0)
2887330f729Sjoerg // This resource isn't required by this stage.
2897330f729Sjoerg continue;
290*82d56013Sjoerg NfaStateTy Combo = ComboBitToBitsMap[ResourceMask];
2917330f729Sjoerg if (Combo && ((~S & Combo) != Combo))
2927330f729Sjoerg // This combo units bits are not available.
2937330f729Sjoerg continue;
294*82d56013Sjoerg NfaStateTy ResultingResourceState = S | ResourceMask | Combo;
2957330f729Sjoerg if (ResultingResourceState == S)
2967330f729Sjoerg continue;
2977330f729Sjoerg V.push_back(ResultingResourceState);
2987330f729Sjoerg }
2997330f729Sjoerg }
3007330f729Sjoerg }
3017330f729Sjoerg return V;
3027330f729Sjoerg };
3037330f729Sjoerg
3047330f729Sjoerg // Given a resource state, return a quick (conservative) guess as to whether
3057330f729Sjoerg // InsnClass can be applied. This is a filter for the more heavyweight
3067330f729Sjoerg // applyInsnClass.
307*82d56013Sjoerg auto canApplyInsnClass = [](const ResourceVector &InsnClass,
3087330f729Sjoerg NfaStateTy State) -> bool {
309*82d56013Sjoerg for (NfaStateTy Resources : InsnClass) {
3107330f729Sjoerg if ((State | Resources) == State)
3117330f729Sjoerg return false;
3127330f729Sjoerg }
3137330f729Sjoerg return true;
3147330f729Sjoerg };
3157330f729Sjoerg
3167330f729Sjoerg DfaEmitter Emitter;
3177330f729Sjoerg std::deque<NfaStateTy> Worklist(1, 0);
3187330f729Sjoerg std::set<NfaStateTy> SeenStates;
3197330f729Sjoerg SeenStates.insert(Worklist.front());
3207330f729Sjoerg while (!Worklist.empty()) {
3217330f729Sjoerg NfaStateTy State = Worklist.front();
3227330f729Sjoerg Worklist.pop_front();
323*82d56013Sjoerg for (const ResourceVector &Resources : UniqueResources) {
324*82d56013Sjoerg if (!canApplyInsnClass(Resources, State))
3257330f729Sjoerg continue;
326*82d56013Sjoerg unsigned ResourcesID = UniqueResources.idFor(Resources);
327*82d56013Sjoerg for (uint64_t NewState : applyInsnClass(Resources, State)) {
3287330f729Sjoerg if (SeenStates.emplace(NewState).second)
3297330f729Sjoerg Worklist.emplace_back(NewState);
330*82d56013Sjoerg Emitter.addTransition(State, NewState, ResourcesID);
3317330f729Sjoerg }
3327330f729Sjoerg }
3337330f729Sjoerg }
3347330f729Sjoerg
3357330f729Sjoerg std::string TargetAndDFAName = TargetName + DFAName;
3367330f729Sjoerg Emitter.emit(TargetAndDFAName, OS);
3377330f729Sjoerg OS << "} // end anonymous namespace\n\n";
3387330f729Sjoerg
3397330f729Sjoerg std::string SubTargetClassName = TargetName + "GenSubtargetInfo";
3407330f729Sjoerg OS << "namespace llvm {\n";
3417330f729Sjoerg OS << "DFAPacketizer *" << SubTargetClassName << "::"
3427330f729Sjoerg << "create" << DFAName
3437330f729Sjoerg << "DFAPacketizer(const InstrItineraryData *IID) const {\n"
3447330f729Sjoerg << " static Automaton<uint64_t> A(ArrayRef<" << TargetAndDFAName
3457330f729Sjoerg << "Transition>(" << TargetAndDFAName << "Transitions), "
3467330f729Sjoerg << TargetAndDFAName << "TransitionInfo);\n"
347*82d56013Sjoerg << " unsigned ProcResIdxStart = " << TargetAndDFAName
348*82d56013Sjoerg << "ProcResourceIndexStart[IID->SchedModel.ProcID];\n"
349*82d56013Sjoerg << " unsigned ProcResIdxNum = " << TargetAndDFAName
350*82d56013Sjoerg << "ProcResourceIndexStart[IID->SchedModel.ProcID + 1] - "
351*82d56013Sjoerg "ProcResIdxStart;\n"
352*82d56013Sjoerg << " return new DFAPacketizer(IID, A, {&" << TargetAndDFAName
353*82d56013Sjoerg << "ResourceIndices[ProcResIdxStart], ProcResIdxNum});\n"
3547330f729Sjoerg << "\n}\n\n";
3557330f729Sjoerg }
3567330f729Sjoerg
3577330f729Sjoerg namespace llvm {
3587330f729Sjoerg
EmitDFAPacketizer(RecordKeeper & RK,raw_ostream & OS)3597330f729Sjoerg void EmitDFAPacketizer(RecordKeeper &RK, raw_ostream &OS) {
3607330f729Sjoerg emitSourceFileHeader("Target DFA Packetizer Tables", OS);
3617330f729Sjoerg DFAPacketizerEmitter(RK).run(OS);
3627330f729Sjoerg }
3637330f729Sjoerg
3647330f729Sjoerg } // end namespace llvm
365