[PowerPC] Fix for PR35688 - handle out-of-range values for r+r to r+i conversion
[llvm-core.git] / lib / CodeGen / DFAPacketizer.cpp
blob848db444270d523a0ddeab9d157874ce6052217f
1 //=- llvm/CodeGen/DFAPacketizer.cpp - DFA Packetizer for VLIW -*- C++ -*-=====//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 // This class implements a deterministic finite automaton (DFA) based
10 // packetizing mechanism for VLIW architectures. It provides APIs to
11 // determine whether there exists a legal mapping of instructions to
12 // functional unit assignments in a packet. The DFA is auto-generated from
13 // the target's Schedule.td file.
15 // A DFA consists of 3 major elements: states, inputs, and transitions. For
16 // the packetizing mechanism, the input is the set of instruction classes for
17 // a target. The state models all possible combinations of functional unit
18 // consumption for a given set of instructions in a packet. A transition
19 // models the addition of an instruction to a packet. In the DFA constructed
20 // by this class, if an instruction can be added to a packet, then a valid
21 // transition exists from the corresponding state. Invalid transitions
22 // indicate that the instruction cannot be added to the current packet.
24 //===----------------------------------------------------------------------===//
26 #include "llvm/CodeGen/DFAPacketizer.h"
27 #include "llvm/CodeGen/MachineFunction.h"
28 #include "llvm/CodeGen/MachineInstr.h"
29 #include "llvm/CodeGen/MachineInstrBundle.h"
30 #include "llvm/CodeGen/ScheduleDAG.h"
31 #include "llvm/CodeGen/ScheduleDAGInstrs.h"
32 #include "llvm/CodeGen/TargetInstrInfo.h"
33 #include "llvm/CodeGen/TargetSubtargetInfo.h"
34 #include "llvm/MC/MCInstrDesc.h"
35 #include "llvm/MC/MCInstrItineraries.h"
36 #include "llvm/Support/CommandLine.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/raw_ostream.h"
39 #include <algorithm>
40 #include <cassert>
41 #include <iterator>
42 #include <memory>
43 #include <vector>
45 using namespace llvm;
47 #define DEBUG_TYPE "packets"
49 static cl::opt<unsigned> InstrLimit("dfa-instr-limit", cl::Hidden,
50 cl::init(0), cl::desc("If present, stops packetizing after N instructions"));
52 static unsigned InstrCount = 0;
54 // --------------------------------------------------------------------
55 // Definitions shared between DFAPacketizer.cpp and DFAPacketizerEmitter.cpp
57 static DFAInput addDFAFuncUnits(DFAInput Inp, unsigned FuncUnits) {
58 return (Inp << DFA_MAX_RESOURCES) | FuncUnits;
61 /// Return the DFAInput for an instruction class input vector.
62 /// This function is used in both DFAPacketizer.cpp and in
63 /// DFAPacketizerEmitter.cpp.
64 static DFAInput getDFAInsnInput(const std::vector<unsigned> &InsnClass) {
65 DFAInput InsnInput = 0;
66 assert((InsnClass.size() <= DFA_MAX_RESTERMS) &&
67 "Exceeded maximum number of DFA terms");
68 for (auto U : InsnClass)
69 InsnInput = addDFAFuncUnits(InsnInput, U);
70 return InsnInput;
73 // --------------------------------------------------------------------
75 DFAPacketizer::DFAPacketizer(const InstrItineraryData *I,
76 const DFAStateInput (*SIT)[2],
77 const unsigned *SET):
78 InstrItins(I), DFAStateInputTable(SIT), DFAStateEntryTable(SET) {
79 // Make sure DFA types are large enough for the number of terms & resources.
80 static_assert((DFA_MAX_RESTERMS * DFA_MAX_RESOURCES) <=
81 (8 * sizeof(DFAInput)),
82 "(DFA_MAX_RESTERMS * DFA_MAX_RESOURCES) too big for DFAInput");
83 static_assert(
84 (DFA_MAX_RESTERMS * DFA_MAX_RESOURCES) <= (8 * sizeof(DFAStateInput)),
85 "(DFA_MAX_RESTERMS * DFA_MAX_RESOURCES) too big for DFAStateInput");
88 // Read the DFA transition table and update CachedTable.
90 // Format of the transition tables:
91 // DFAStateInputTable[][2] = pairs of <Input, Transition> for all valid
92 // transitions
93 // DFAStateEntryTable[i] = Index of the first entry in DFAStateInputTable
94 // for the ith state
96 void DFAPacketizer::ReadTable(unsigned int state) {
97 unsigned ThisState = DFAStateEntryTable[state];
98 unsigned NextStateInTable = DFAStateEntryTable[state+1];
99 // Early exit in case CachedTable has already contains this
100 // state's transitions.
101 if (CachedTable.count(UnsignPair(state, DFAStateInputTable[ThisState][0])))
102 return;
104 for (unsigned i = ThisState; i < NextStateInTable; i++)
105 CachedTable[UnsignPair(state, DFAStateInputTable[i][0])] =
106 DFAStateInputTable[i][1];
109 // Return the DFAInput for an instruction class.
110 DFAInput DFAPacketizer::getInsnInput(unsigned InsnClass) {
111 // Note: this logic must match that in DFAPacketizerDefs.h for input vectors.
112 DFAInput InsnInput = 0;
113 unsigned i = 0;
114 (void)i;
115 for (const InstrStage *IS = InstrItins->beginStage(InsnClass),
116 *IE = InstrItins->endStage(InsnClass); IS != IE; ++IS) {
117 InsnInput = addDFAFuncUnits(InsnInput, IS->getUnits());
118 assert((i++ < DFA_MAX_RESTERMS) && "Exceeded maximum number of DFA inputs");
120 return InsnInput;
123 // Return the DFAInput for an instruction class input vector.
124 DFAInput DFAPacketizer::getInsnInput(const std::vector<unsigned> &InsnClass) {
125 return getDFAInsnInput(InsnClass);
128 // Check if the resources occupied by a MCInstrDesc are available in the
129 // current state.
130 bool DFAPacketizer::canReserveResources(const MCInstrDesc *MID) {
131 unsigned InsnClass = MID->getSchedClass();
132 DFAInput InsnInput = getInsnInput(InsnClass);
133 UnsignPair StateTrans = UnsignPair(CurrentState, InsnInput);
134 ReadTable(CurrentState);
135 return CachedTable.count(StateTrans) != 0;
138 // Reserve the resources occupied by a MCInstrDesc and change the current
139 // state to reflect that change.
140 void DFAPacketizer::reserveResources(const MCInstrDesc *MID) {
141 unsigned InsnClass = MID->getSchedClass();
142 DFAInput InsnInput = getInsnInput(InsnClass);
143 UnsignPair StateTrans = UnsignPair(CurrentState, InsnInput);
144 ReadTable(CurrentState);
145 assert(CachedTable.count(StateTrans) != 0);
146 CurrentState = CachedTable[StateTrans];
149 // Check if the resources occupied by a machine instruction are available
150 // in the current state.
151 bool DFAPacketizer::canReserveResources(MachineInstr &MI) {
152 const MCInstrDesc &MID = MI.getDesc();
153 return canReserveResources(&MID);
156 // Reserve the resources occupied by a machine instruction and change the
157 // current state to reflect that change.
158 void DFAPacketizer::reserveResources(MachineInstr &MI) {
159 const MCInstrDesc &MID = MI.getDesc();
160 reserveResources(&MID);
163 namespace llvm {
165 // This class extends ScheduleDAGInstrs and overrides the schedule method
166 // to build the dependence graph.
167 class DefaultVLIWScheduler : public ScheduleDAGInstrs {
168 private:
169 AliasAnalysis *AA;
170 /// Ordered list of DAG postprocessing steps.
171 std::vector<std::unique_ptr<ScheduleDAGMutation>> Mutations;
173 public:
174 DefaultVLIWScheduler(MachineFunction &MF, MachineLoopInfo &MLI,
175 AliasAnalysis *AA);
177 // Actual scheduling work.
178 void schedule() override;
180 /// DefaultVLIWScheduler takes ownership of the Mutation object.
181 void addMutation(std::unique_ptr<ScheduleDAGMutation> Mutation) {
182 Mutations.push_back(std::move(Mutation));
185 protected:
186 void postprocessDAG();
189 } // end namespace llvm
191 DefaultVLIWScheduler::DefaultVLIWScheduler(MachineFunction &MF,
192 MachineLoopInfo &MLI,
193 AliasAnalysis *AA)
194 : ScheduleDAGInstrs(MF, &MLI), AA(AA) {
195 CanHandleTerminators = true;
198 /// Apply each ScheduleDAGMutation step in order.
199 void DefaultVLIWScheduler::postprocessDAG() {
200 for (auto &M : Mutations)
201 M->apply(this);
204 void DefaultVLIWScheduler::schedule() {
205 // Build the scheduling graph.
206 buildSchedGraph(AA);
207 postprocessDAG();
210 VLIWPacketizerList::VLIWPacketizerList(MachineFunction &mf,
211 MachineLoopInfo &mli, AliasAnalysis *aa)
212 : MF(mf), TII(mf.getSubtarget().getInstrInfo()), AA(aa) {
213 ResourceTracker = TII->CreateTargetScheduleState(MF.getSubtarget());
214 VLIWScheduler = new DefaultVLIWScheduler(MF, mli, AA);
217 VLIWPacketizerList::~VLIWPacketizerList() {
218 delete VLIWScheduler;
219 delete ResourceTracker;
222 // End the current packet, bundle packet instructions and reset DFA state.
223 void VLIWPacketizerList::endPacket(MachineBasicBlock *MBB,
224 MachineBasicBlock::iterator MI) {
225 DEBUG({
226 if (!CurrentPacketMIs.empty()) {
227 dbgs() << "Finalizing packet:\n";
228 for (MachineInstr *MI : CurrentPacketMIs)
229 dbgs() << " * " << *MI;
232 if (CurrentPacketMIs.size() > 1) {
233 MachineInstr &MIFirst = *CurrentPacketMIs.front();
234 finalizeBundle(*MBB, MIFirst.getIterator(), MI.getInstrIterator());
236 CurrentPacketMIs.clear();
237 ResourceTracker->clearResources();
238 DEBUG(dbgs() << "End packet\n");
241 // Bundle machine instructions into packets.
242 void VLIWPacketizerList::PacketizeMIs(MachineBasicBlock *MBB,
243 MachineBasicBlock::iterator BeginItr,
244 MachineBasicBlock::iterator EndItr) {
245 assert(VLIWScheduler && "VLIW Scheduler is not initialized!");
246 VLIWScheduler->startBlock(MBB);
247 VLIWScheduler->enterRegion(MBB, BeginItr, EndItr,
248 std::distance(BeginItr, EndItr));
249 VLIWScheduler->schedule();
251 DEBUG({
252 dbgs() << "Scheduling DAG of the packetize region\n";
253 for (SUnit &SU : VLIWScheduler->SUnits)
254 SU.dumpAll(VLIWScheduler);
257 // Generate MI -> SU map.
258 MIToSUnit.clear();
259 for (SUnit &SU : VLIWScheduler->SUnits)
260 MIToSUnit[SU.getInstr()] = &SU;
262 bool LimitPresent = InstrLimit.getPosition();
264 // The main packetizer loop.
265 for (; BeginItr != EndItr; ++BeginItr) {
266 if (LimitPresent) {
267 if (InstrCount >= InstrLimit) {
268 EndItr = BeginItr;
269 break;
271 InstrCount++;
273 MachineInstr &MI = *BeginItr;
274 initPacketizerState();
276 // End the current packet if needed.
277 if (isSoloInstruction(MI)) {
278 endPacket(MBB, MI);
279 continue;
282 // Ignore pseudo instructions.
283 if (ignorePseudoInstruction(MI, MBB))
284 continue;
286 SUnit *SUI = MIToSUnit[&MI];
287 assert(SUI && "Missing SUnit Info!");
289 // Ask DFA if machine resource is available for MI.
290 DEBUG(dbgs() << "Checking resources for adding MI to packet " << MI);
292 bool ResourceAvail = ResourceTracker->canReserveResources(MI);
293 DEBUG({
294 if (ResourceAvail)
295 dbgs() << " Resources are available for adding MI to packet\n";
296 else
297 dbgs() << " Resources NOT available\n";
299 if (ResourceAvail && shouldAddToPacket(MI)) {
300 // Dependency check for MI with instructions in CurrentPacketMIs.
301 for (auto MJ : CurrentPacketMIs) {
302 SUnit *SUJ = MIToSUnit[MJ];
303 assert(SUJ && "Missing SUnit Info!");
305 DEBUG(dbgs() << " Checking against MJ " << *MJ);
306 // Is it legal to packetize SUI and SUJ together.
307 if (!isLegalToPacketizeTogether(SUI, SUJ)) {
308 DEBUG(dbgs() << " Not legal to add MI, try to prune\n");
309 // Allow packetization if dependency can be pruned.
310 if (!isLegalToPruneDependencies(SUI, SUJ)) {
311 // End the packet if dependency cannot be pruned.
312 DEBUG(dbgs() << " Could not prune dependencies for adding MI\n");
313 endPacket(MBB, MI);
314 break;
316 DEBUG(dbgs() << " Pruned dependence for adding MI\n");
319 } else {
320 DEBUG(if (ResourceAvail)
321 dbgs() << "Resources are available, but instruction should not be "
322 "added to packet\n " << MI);
323 // End the packet if resource is not available, or if the instruction
324 // shoud not be added to the current packet.
325 endPacket(MBB, MI);
328 // Add MI to the current packet.
329 DEBUG(dbgs() << "* Adding MI to packet " << MI << '\n');
330 BeginItr = addToPacket(MI);
331 } // For all instructions in the packetization range.
333 // End any packet left behind.
334 endPacket(MBB, EndItr);
335 VLIWScheduler->exitRegion();
336 VLIWScheduler->finishBlock();
339 bool VLIWPacketizerList::alias(const MachineMemOperand &Op1,
340 const MachineMemOperand &Op2,
341 bool UseTBAA) const {
342 if (!Op1.getValue() || !Op2.getValue())
343 return true;
345 int64_t MinOffset = std::min(Op1.getOffset(), Op2.getOffset());
346 int64_t Overlapa = Op1.getSize() + Op1.getOffset() - MinOffset;
347 int64_t Overlapb = Op2.getSize() + Op2.getOffset() - MinOffset;
349 AliasResult AAResult =
350 AA->alias(MemoryLocation(Op1.getValue(), Overlapa,
351 UseTBAA ? Op1.getAAInfo() : AAMDNodes()),
352 MemoryLocation(Op2.getValue(), Overlapb,
353 UseTBAA ? Op2.getAAInfo() : AAMDNodes()));
355 return AAResult != NoAlias;
358 bool VLIWPacketizerList::alias(const MachineInstr &MI1,
359 const MachineInstr &MI2,
360 bool UseTBAA) const {
361 if (MI1.memoperands_empty() || MI2.memoperands_empty())
362 return true;
364 for (const MachineMemOperand *Op1 : MI1.memoperands())
365 for (const MachineMemOperand *Op2 : MI2.memoperands())
366 if (alias(*Op1, *Op2, UseTBAA))
367 return true;
368 return false;
371 // Add a DAG mutation object to the ordered list.
372 void VLIWPacketizerList::addMutation(
373 std::unique_ptr<ScheduleDAGMutation> Mutation) {
374 VLIWScheduler->addMutation(std::move(Mutation));