xref: /freebsd/contrib/llvm-project/llvm/lib/CodeGen/RegAllocPBQP.cpp (revision 3cbb4cc200f8a0ad7ed08233425ea54524a21f1c)
1 //===- RegAllocPBQP.cpp ---- PBQP Register Allocator ----------------------===//
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 contains a Partitioned Boolean Quadratic Programming (PBQP) based
10 // register allocator for LLVM. This allocator works by constructing a PBQP
11 // problem representing the register allocation problem under consideration,
12 // solving this using a PBQP solver, and mapping the solution back to a
13 // register assignment. If any variables are selected for spilling then spill
14 // code is inserted and the process repeated.
15 //
16 // The PBQP solver (pbqp.c) provided for this allocator uses a heuristic tuned
17 // for register allocation. For more information on PBQP for register
18 // allocation, see the following papers:
19 //
20 //   (1) Hames, L. and Scholz, B. 2006. Nearly optimal register allocation with
21 //   PBQP. In Proceedings of the 7th Joint Modular Languages Conference
22 //   (JMLC'06). LNCS, vol. 4228. Springer, New York, NY, USA. 346-361.
23 //
24 //   (2) Scholz, B., Eckstein, E. 2002. Register allocation for irregular
25 //   architectures. In Proceedings of the Joint Conference on Languages,
26 //   Compilers and Tools for Embedded Systems (LCTES'02), ACM Press, New York,
27 //   NY, USA, 139-148.
28 //
29 //===----------------------------------------------------------------------===//
30 
31 #include "llvm/CodeGen/RegAllocPBQP.h"
32 #include "RegisterCoalescer.h"
33 #include "llvm/ADT/ArrayRef.h"
34 #include "llvm/ADT/BitVector.h"
35 #include "llvm/ADT/DenseMap.h"
36 #include "llvm/ADT/DenseSet.h"
37 #include "llvm/ADT/STLExtras.h"
38 #include "llvm/ADT/SmallPtrSet.h"
39 #include "llvm/ADT/SmallVector.h"
40 #include "llvm/ADT/StringRef.h"
41 #include "llvm/Analysis/AliasAnalysis.h"
42 #include "llvm/CodeGen/CalcSpillWeights.h"
43 #include "llvm/CodeGen/LiveInterval.h"
44 #include "llvm/CodeGen/LiveIntervals.h"
45 #include "llvm/CodeGen/LiveRangeEdit.h"
46 #include "llvm/CodeGen/LiveStacks.h"
47 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
48 #include "llvm/CodeGen/MachineDominators.h"
49 #include "llvm/CodeGen/MachineFunction.h"
50 #include "llvm/CodeGen/MachineFunctionPass.h"
51 #include "llvm/CodeGen/MachineInstr.h"
52 #include "llvm/CodeGen/MachineLoopInfo.h"
53 #include "llvm/CodeGen/MachineRegisterInfo.h"
54 #include "llvm/CodeGen/PBQP/Graph.h"
55 #include "llvm/CodeGen/PBQP/Math.h"
56 #include "llvm/CodeGen/PBQP/Solution.h"
57 #include "llvm/CodeGen/PBQPRAConstraint.h"
58 #include "llvm/CodeGen/RegAllocRegistry.h"
59 #include "llvm/CodeGen/SlotIndexes.h"
60 #include "llvm/CodeGen/Spiller.h"
61 #include "llvm/CodeGen/TargetRegisterInfo.h"
62 #include "llvm/CodeGen/TargetSubtargetInfo.h"
63 #include "llvm/CodeGen/VirtRegMap.h"
64 #include "llvm/Config/llvm-config.h"
65 #include "llvm/IR/Function.h"
66 #include "llvm/IR/Module.h"
67 #include "llvm/MC/MCRegisterInfo.h"
68 #include "llvm/Pass.h"
69 #include "llvm/Support/CommandLine.h"
70 #include "llvm/Support/Compiler.h"
71 #include "llvm/Support/Debug.h"
72 #include "llvm/Support/FileSystem.h"
73 #include "llvm/Support/Printable.h"
74 #include "llvm/Support/raw_ostream.h"
75 #include <algorithm>
76 #include <cassert>
77 #include <cstddef>
78 #include <limits>
79 #include <map>
80 #include <memory>
81 #include <queue>
82 #include <set>
83 #include <sstream>
84 #include <string>
85 #include <system_error>
86 #include <tuple>
87 #include <utility>
88 #include <vector>
89 
90 using namespace llvm;
91 
92 #define DEBUG_TYPE "regalloc"
93 
94 static RegisterRegAlloc
95 RegisterPBQPRepAlloc("pbqp", "PBQP register allocator",
96                        createDefaultPBQPRegisterAllocator);
97 
98 static cl::opt<bool>
99 PBQPCoalescing("pbqp-coalescing",
100                 cl::desc("Attempt coalescing during PBQP register allocation."),
101                 cl::init(false), cl::Hidden);
102 
103 #ifndef NDEBUG
104 static cl::opt<bool>
105 PBQPDumpGraphs("pbqp-dump-graphs",
106                cl::desc("Dump graphs for each function/round in the compilation unit."),
107                cl::init(false), cl::Hidden);
108 #endif
109 
110 namespace {
111 
112 ///
113 /// PBQP based allocators solve the register allocation problem by mapping
114 /// register allocation problems to Partitioned Boolean Quadratic
115 /// Programming problems.
116 class RegAllocPBQP : public MachineFunctionPass {
117 public:
118   static char ID;
119 
120   /// Construct a PBQP register allocator.
121   RegAllocPBQP(char *cPassID = nullptr)
122       : MachineFunctionPass(ID), customPassID(cPassID) {
123     initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
124     initializeLiveIntervalsPass(*PassRegistry::getPassRegistry());
125     initializeLiveStacksPass(*PassRegistry::getPassRegistry());
126     initializeVirtRegMapPass(*PassRegistry::getPassRegistry());
127   }
128 
129   /// Return the pass name.
130   StringRef getPassName() const override { return "PBQP Register Allocator"; }
131 
132   /// PBQP analysis usage.
133   void getAnalysisUsage(AnalysisUsage &au) const override;
134 
135   /// Perform register allocation
136   bool runOnMachineFunction(MachineFunction &MF) override;
137 
138   MachineFunctionProperties getRequiredProperties() const override {
139     return MachineFunctionProperties().set(
140         MachineFunctionProperties::Property::NoPHIs);
141   }
142 
143 private:
144   using LI2NodeMap = std::map<const LiveInterval *, unsigned>;
145   using Node2LIMap = std::vector<const LiveInterval *>;
146   using AllowedSet = std::vector<unsigned>;
147   using AllowedSetMap = std::vector<AllowedSet>;
148   using RegPair = std::pair<unsigned, unsigned>;
149   using CoalesceMap = std::map<RegPair, PBQP::PBQPNum>;
150   using RegSet = std::set<unsigned>;
151 
152   char *customPassID;
153 
154   RegSet VRegsToAlloc, EmptyIntervalVRegs;
155 
156   /// Inst which is a def of an original reg and whose defs are already all
157   /// dead after remat is saved in DeadRemats. The deletion of such inst is
158   /// postponed till all the allocations are done, so its remat expr is
159   /// always available for the remat of all the siblings of the original reg.
160   SmallPtrSet<MachineInstr *, 32> DeadRemats;
161 
162   /// Finds the initial set of vreg intervals to allocate.
163   void findVRegIntervalsToAlloc(const MachineFunction &MF, LiveIntervals &LIS);
164 
165   /// Constructs an initial graph.
166   void initializeGraph(PBQPRAGraph &G, VirtRegMap &VRM, Spiller &VRegSpiller);
167 
168   /// Spill the given VReg.
169   void spillVReg(Register VReg, SmallVectorImpl<Register> &NewIntervals,
170                  MachineFunction &MF, LiveIntervals &LIS, VirtRegMap &VRM,
171                  Spiller &VRegSpiller);
172 
173   /// Given a solved PBQP problem maps this solution back to a register
174   /// assignment.
175   bool mapPBQPToRegAlloc(const PBQPRAGraph &G,
176                          const PBQP::Solution &Solution,
177                          VirtRegMap &VRM,
178                          Spiller &VRegSpiller);
179 
180   /// Postprocessing before final spilling. Sets basic block "live in"
181   /// variables.
182   void finalizeAlloc(MachineFunction &MF, LiveIntervals &LIS,
183                      VirtRegMap &VRM) const;
184 
185   void postOptimization(Spiller &VRegSpiller, LiveIntervals &LIS);
186 };
187 
188 char RegAllocPBQP::ID = 0;
189 
190 /// Set spill costs for each node in the PBQP reg-alloc graph.
191 class SpillCosts : public PBQPRAConstraint {
192 public:
193   void apply(PBQPRAGraph &G) override {
194     LiveIntervals &LIS = G.getMetadata().LIS;
195 
196     // A minimum spill costs, so that register constraints can can be set
197     // without normalization in the [0.0:MinSpillCost( interval.
198     const PBQP::PBQPNum MinSpillCost = 10.0;
199 
200     for (auto NId : G.nodeIds()) {
201       PBQP::PBQPNum SpillCost =
202         LIS.getInterval(G.getNodeMetadata(NId).getVReg()).weight;
203       if (SpillCost == 0.0)
204         SpillCost = std::numeric_limits<PBQP::PBQPNum>::min();
205       else
206         SpillCost += MinSpillCost;
207       PBQPRAGraph::RawVector NodeCosts(G.getNodeCosts(NId));
208       NodeCosts[PBQP::RegAlloc::getSpillOptionIdx()] = SpillCost;
209       G.setNodeCosts(NId, std::move(NodeCosts));
210     }
211   }
212 };
213 
214 /// Add interference edges between overlapping vregs.
215 class Interference : public PBQPRAConstraint {
216 private:
217   using AllowedRegVecPtr = const PBQP::RegAlloc::AllowedRegVector *;
218   using IKey = std::pair<AllowedRegVecPtr, AllowedRegVecPtr>;
219   using IMatrixCache = DenseMap<IKey, PBQPRAGraph::MatrixPtr>;
220   using DisjointAllowedRegsCache = DenseSet<IKey>;
221   using IEdgeKey = std::pair<PBQP::GraphBase::NodeId, PBQP::GraphBase::NodeId>;
222   using IEdgeCache = DenseSet<IEdgeKey>;
223 
224   bool haveDisjointAllowedRegs(const PBQPRAGraph &G, PBQPRAGraph::NodeId NId,
225                                PBQPRAGraph::NodeId MId,
226                                const DisjointAllowedRegsCache &D) const {
227     const auto *NRegs = &G.getNodeMetadata(NId).getAllowedRegs();
228     const auto *MRegs = &G.getNodeMetadata(MId).getAllowedRegs();
229 
230     if (NRegs == MRegs)
231       return false;
232 
233     if (NRegs < MRegs)
234       return D.count(IKey(NRegs, MRegs)) > 0;
235 
236     return D.count(IKey(MRegs, NRegs)) > 0;
237   }
238 
239   void setDisjointAllowedRegs(const PBQPRAGraph &G, PBQPRAGraph::NodeId NId,
240                               PBQPRAGraph::NodeId MId,
241                               DisjointAllowedRegsCache &D) {
242     const auto *NRegs = &G.getNodeMetadata(NId).getAllowedRegs();
243     const auto *MRegs = &G.getNodeMetadata(MId).getAllowedRegs();
244 
245     assert(NRegs != MRegs && "AllowedRegs can not be disjoint with itself");
246 
247     if (NRegs < MRegs)
248       D.insert(IKey(NRegs, MRegs));
249     else
250       D.insert(IKey(MRegs, NRegs));
251   }
252 
253   // Holds (Interval, CurrentSegmentID, and NodeId). The first two are required
254   // for the fast interference graph construction algorithm. The last is there
255   // to save us from looking up node ids via the VRegToNode map in the graph
256   // metadata.
257   using IntervalInfo =
258       std::tuple<LiveInterval*, size_t, PBQP::GraphBase::NodeId>;
259 
260   static SlotIndex getStartPoint(const IntervalInfo &I) {
261     return std::get<0>(I)->segments[std::get<1>(I)].start;
262   }
263 
264   static SlotIndex getEndPoint(const IntervalInfo &I) {
265     return std::get<0>(I)->segments[std::get<1>(I)].end;
266   }
267 
268   static PBQP::GraphBase::NodeId getNodeId(const IntervalInfo &I) {
269     return std::get<2>(I);
270   }
271 
272   static bool lowestStartPoint(const IntervalInfo &I1,
273                                const IntervalInfo &I2) {
274     // Condition reversed because priority queue has the *highest* element at
275     // the front, rather than the lowest.
276     return getStartPoint(I1) > getStartPoint(I2);
277   }
278 
279   static bool lowestEndPoint(const IntervalInfo &I1,
280                              const IntervalInfo &I2) {
281     SlotIndex E1 = getEndPoint(I1);
282     SlotIndex E2 = getEndPoint(I2);
283 
284     if (E1 < E2)
285       return true;
286 
287     if (E1 > E2)
288       return false;
289 
290     // If two intervals end at the same point, we need a way to break the tie or
291     // the set will assume they're actually equal and refuse to insert a
292     // "duplicate". Just compare the vregs - fast and guaranteed unique.
293     return std::get<0>(I1)->reg < std::get<0>(I2)->reg;
294   }
295 
296   static bool isAtLastSegment(const IntervalInfo &I) {
297     return std::get<1>(I) == std::get<0>(I)->size() - 1;
298   }
299 
300   static IntervalInfo nextSegment(const IntervalInfo &I) {
301     return std::make_tuple(std::get<0>(I), std::get<1>(I) + 1, std::get<2>(I));
302   }
303 
304 public:
305   void apply(PBQPRAGraph &G) override {
306     // The following is loosely based on the linear scan algorithm introduced in
307     // "Linear Scan Register Allocation" by Poletto and Sarkar. This version
308     // isn't linear, because the size of the active set isn't bound by the
309     // number of registers, but rather the size of the largest clique in the
310     // graph. Still, we expect this to be better than N^2.
311     LiveIntervals &LIS = G.getMetadata().LIS;
312 
313     // Interferenc matrices are incredibly regular - they're only a function of
314     // the allowed sets, so we cache them to avoid the overhead of constructing
315     // and uniquing them.
316     IMatrixCache C;
317 
318     // Finding an edge is expensive in the worst case (O(max_clique(G))). So
319     // cache locally edges we have already seen.
320     IEdgeCache EC;
321 
322     // Cache known disjoint allowed registers pairs
323     DisjointAllowedRegsCache D;
324 
325     using IntervalSet = std::set<IntervalInfo, decltype(&lowestEndPoint)>;
326     using IntervalQueue =
327         std::priority_queue<IntervalInfo, std::vector<IntervalInfo>,
328                             decltype(&lowestStartPoint)>;
329     IntervalSet Active(lowestEndPoint);
330     IntervalQueue Inactive(lowestStartPoint);
331 
332     // Start by building the inactive set.
333     for (auto NId : G.nodeIds()) {
334       unsigned VReg = G.getNodeMetadata(NId).getVReg();
335       LiveInterval &LI = LIS.getInterval(VReg);
336       assert(!LI.empty() && "PBQP graph contains node for empty interval");
337       Inactive.push(std::make_tuple(&LI, 0, NId));
338     }
339 
340     while (!Inactive.empty()) {
341       // Tentatively grab the "next" interval - this choice may be overriden
342       // below.
343       IntervalInfo Cur = Inactive.top();
344 
345       // Retire any active intervals that end before Cur starts.
346       IntervalSet::iterator RetireItr = Active.begin();
347       while (RetireItr != Active.end() &&
348              (getEndPoint(*RetireItr) <= getStartPoint(Cur))) {
349         // If this interval has subsequent segments, add the next one to the
350         // inactive list.
351         if (!isAtLastSegment(*RetireItr))
352           Inactive.push(nextSegment(*RetireItr));
353 
354         ++RetireItr;
355       }
356       Active.erase(Active.begin(), RetireItr);
357 
358       // One of the newly retired segments may actually start before the
359       // Cur segment, so re-grab the front of the inactive list.
360       Cur = Inactive.top();
361       Inactive.pop();
362 
363       // At this point we know that Cur overlaps all active intervals. Add the
364       // interference edges.
365       PBQP::GraphBase::NodeId NId = getNodeId(Cur);
366       for (const auto &A : Active) {
367         PBQP::GraphBase::NodeId MId = getNodeId(A);
368 
369         // Do not add an edge when the nodes' allowed registers do not
370         // intersect: there is obviously no interference.
371         if (haveDisjointAllowedRegs(G, NId, MId, D))
372           continue;
373 
374         // Check that we haven't already added this edge
375         IEdgeKey EK(std::min(NId, MId), std::max(NId, MId));
376         if (EC.count(EK))
377           continue;
378 
379         // This is a new edge - add it to the graph.
380         if (!createInterferenceEdge(G, NId, MId, C))
381           setDisjointAllowedRegs(G, NId, MId, D);
382         else
383           EC.insert(EK);
384       }
385 
386       // Finally, add Cur to the Active set.
387       Active.insert(Cur);
388     }
389   }
390 
391 private:
392   // Create an Interference edge and add it to the graph, unless it is
393   // a null matrix, meaning the nodes' allowed registers do not have any
394   // interference. This case occurs frequently between integer and floating
395   // point registers for example.
396   // return true iff both nodes interferes.
397   bool createInterferenceEdge(PBQPRAGraph &G,
398                               PBQPRAGraph::NodeId NId, PBQPRAGraph::NodeId MId,
399                               IMatrixCache &C) {
400     const TargetRegisterInfo &TRI =
401         *G.getMetadata().MF.getSubtarget().getRegisterInfo();
402     const auto &NRegs = G.getNodeMetadata(NId).getAllowedRegs();
403     const auto &MRegs = G.getNodeMetadata(MId).getAllowedRegs();
404 
405     // Try looking the edge costs up in the IMatrixCache first.
406     IKey K(&NRegs, &MRegs);
407     IMatrixCache::iterator I = C.find(K);
408     if (I != C.end()) {
409       G.addEdgeBypassingCostAllocator(NId, MId, I->second);
410       return true;
411     }
412 
413     PBQPRAGraph::RawMatrix M(NRegs.size() + 1, MRegs.size() + 1, 0);
414     bool NodesInterfere = false;
415     for (unsigned I = 0; I != NRegs.size(); ++I) {
416       unsigned PRegN = NRegs[I];
417       for (unsigned J = 0; J != MRegs.size(); ++J) {
418         unsigned PRegM = MRegs[J];
419         if (TRI.regsOverlap(PRegN, PRegM)) {
420           M[I + 1][J + 1] = std::numeric_limits<PBQP::PBQPNum>::infinity();
421           NodesInterfere = true;
422         }
423       }
424     }
425 
426     if (!NodesInterfere)
427       return false;
428 
429     PBQPRAGraph::EdgeId EId = G.addEdge(NId, MId, std::move(M));
430     C[K] = G.getEdgeCostsPtr(EId);
431 
432     return true;
433   }
434 };
435 
436 class Coalescing : public PBQPRAConstraint {
437 public:
438   void apply(PBQPRAGraph &G) override {
439     MachineFunction &MF = G.getMetadata().MF;
440     MachineBlockFrequencyInfo &MBFI = G.getMetadata().MBFI;
441     CoalescerPair CP(*MF.getSubtarget().getRegisterInfo());
442 
443     // Scan the machine function and add a coalescing cost whenever CoalescerPair
444     // gives the Ok.
445     for (const auto &MBB : MF) {
446       for (const auto &MI : MBB) {
447         // Skip not-coalescable or already coalesced copies.
448         if (!CP.setRegisters(&MI) || CP.getSrcReg() == CP.getDstReg())
449           continue;
450 
451         unsigned DstReg = CP.getDstReg();
452         unsigned SrcReg = CP.getSrcReg();
453 
454         const float Scale = 1.0f / MBFI.getEntryFreq();
455         PBQP::PBQPNum CBenefit = MBFI.getBlockFreq(&MBB).getFrequency() * Scale;
456 
457         if (CP.isPhys()) {
458           if (!MF.getRegInfo().isAllocatable(DstReg))
459             continue;
460 
461           PBQPRAGraph::NodeId NId = G.getMetadata().getNodeIdForVReg(SrcReg);
462 
463           const PBQPRAGraph::NodeMetadata::AllowedRegVector &Allowed =
464             G.getNodeMetadata(NId).getAllowedRegs();
465 
466           unsigned PRegOpt = 0;
467           while (PRegOpt < Allowed.size() && Allowed[PRegOpt] != DstReg)
468             ++PRegOpt;
469 
470           if (PRegOpt < Allowed.size()) {
471             PBQPRAGraph::RawVector NewCosts(G.getNodeCosts(NId));
472             NewCosts[PRegOpt + 1] -= CBenefit;
473             G.setNodeCosts(NId, std::move(NewCosts));
474           }
475         } else {
476           PBQPRAGraph::NodeId N1Id = G.getMetadata().getNodeIdForVReg(DstReg);
477           PBQPRAGraph::NodeId N2Id = G.getMetadata().getNodeIdForVReg(SrcReg);
478           const PBQPRAGraph::NodeMetadata::AllowedRegVector *Allowed1 =
479             &G.getNodeMetadata(N1Id).getAllowedRegs();
480           const PBQPRAGraph::NodeMetadata::AllowedRegVector *Allowed2 =
481             &G.getNodeMetadata(N2Id).getAllowedRegs();
482 
483           PBQPRAGraph::EdgeId EId = G.findEdge(N1Id, N2Id);
484           if (EId == G.invalidEdgeId()) {
485             PBQPRAGraph::RawMatrix Costs(Allowed1->size() + 1,
486                                          Allowed2->size() + 1, 0);
487             addVirtRegCoalesce(Costs, *Allowed1, *Allowed2, CBenefit);
488             G.addEdge(N1Id, N2Id, std::move(Costs));
489           } else {
490             if (G.getEdgeNode1Id(EId) == N2Id) {
491               std::swap(N1Id, N2Id);
492               std::swap(Allowed1, Allowed2);
493             }
494             PBQPRAGraph::RawMatrix Costs(G.getEdgeCosts(EId));
495             addVirtRegCoalesce(Costs, *Allowed1, *Allowed2, CBenefit);
496             G.updateEdgeCosts(EId, std::move(Costs));
497           }
498         }
499       }
500     }
501   }
502 
503 private:
504   void addVirtRegCoalesce(
505                     PBQPRAGraph::RawMatrix &CostMat,
506                     const PBQPRAGraph::NodeMetadata::AllowedRegVector &Allowed1,
507                     const PBQPRAGraph::NodeMetadata::AllowedRegVector &Allowed2,
508                     PBQP::PBQPNum Benefit) {
509     assert(CostMat.getRows() == Allowed1.size() + 1 && "Size mismatch.");
510     assert(CostMat.getCols() == Allowed2.size() + 1 && "Size mismatch.");
511     for (unsigned I = 0; I != Allowed1.size(); ++I) {
512       unsigned PReg1 = Allowed1[I];
513       for (unsigned J = 0; J != Allowed2.size(); ++J) {
514         unsigned PReg2 = Allowed2[J];
515         if (PReg1 == PReg2)
516           CostMat[I + 1][J + 1] -= Benefit;
517       }
518     }
519   }
520 };
521 
522 } // end anonymous namespace
523 
524 // Out-of-line destructor/anchor for PBQPRAConstraint.
525 PBQPRAConstraint::~PBQPRAConstraint() = default;
526 
527 void PBQPRAConstraint::anchor() {}
528 
529 void PBQPRAConstraintList::anchor() {}
530 
531 void RegAllocPBQP::getAnalysisUsage(AnalysisUsage &au) const {
532   au.setPreservesCFG();
533   au.addRequired<AAResultsWrapperPass>();
534   au.addPreserved<AAResultsWrapperPass>();
535   au.addRequired<SlotIndexes>();
536   au.addPreserved<SlotIndexes>();
537   au.addRequired<LiveIntervals>();
538   au.addPreserved<LiveIntervals>();
539   //au.addRequiredID(SplitCriticalEdgesID);
540   if (customPassID)
541     au.addRequiredID(*customPassID);
542   au.addRequired<LiveStacks>();
543   au.addPreserved<LiveStacks>();
544   au.addRequired<MachineBlockFrequencyInfo>();
545   au.addPreserved<MachineBlockFrequencyInfo>();
546   au.addRequired<MachineLoopInfo>();
547   au.addPreserved<MachineLoopInfo>();
548   au.addRequired<MachineDominatorTree>();
549   au.addPreserved<MachineDominatorTree>();
550   au.addRequired<VirtRegMap>();
551   au.addPreserved<VirtRegMap>();
552   MachineFunctionPass::getAnalysisUsage(au);
553 }
554 
555 void RegAllocPBQP::findVRegIntervalsToAlloc(const MachineFunction &MF,
556                                             LiveIntervals &LIS) {
557   const MachineRegisterInfo &MRI = MF.getRegInfo();
558 
559   // Iterate over all live ranges.
560   for (unsigned I = 0, E = MRI.getNumVirtRegs(); I != E; ++I) {
561     unsigned Reg = Register::index2VirtReg(I);
562     if (MRI.reg_nodbg_empty(Reg))
563       continue;
564     VRegsToAlloc.insert(Reg);
565   }
566 }
567 
568 static bool isACalleeSavedRegister(unsigned reg, const TargetRegisterInfo &TRI,
569                                    const MachineFunction &MF) {
570   const MCPhysReg *CSR = MF.getRegInfo().getCalleeSavedRegs();
571   for (unsigned i = 0; CSR[i] != 0; ++i)
572     if (TRI.regsOverlap(reg, CSR[i]))
573       return true;
574   return false;
575 }
576 
577 void RegAllocPBQP::initializeGraph(PBQPRAGraph &G, VirtRegMap &VRM,
578                                    Spiller &VRegSpiller) {
579   MachineFunction &MF = G.getMetadata().MF;
580 
581   LiveIntervals &LIS = G.getMetadata().LIS;
582   const MachineRegisterInfo &MRI = G.getMetadata().MF.getRegInfo();
583   const TargetRegisterInfo &TRI =
584       *G.getMetadata().MF.getSubtarget().getRegisterInfo();
585 
586   std::vector<unsigned> Worklist(VRegsToAlloc.begin(), VRegsToAlloc.end());
587 
588   std::map<unsigned, std::vector<unsigned>> VRegAllowedMap;
589 
590   while (!Worklist.empty()) {
591     unsigned VReg = Worklist.back();
592     Worklist.pop_back();
593 
594     LiveInterval &VRegLI = LIS.getInterval(VReg);
595 
596     // If this is an empty interval move it to the EmptyIntervalVRegs set then
597     // continue.
598     if (VRegLI.empty()) {
599       EmptyIntervalVRegs.insert(VRegLI.reg);
600       VRegsToAlloc.erase(VRegLI.reg);
601       continue;
602     }
603 
604     const TargetRegisterClass *TRC = MRI.getRegClass(VReg);
605 
606     // Record any overlaps with regmask operands.
607     BitVector RegMaskOverlaps;
608     LIS.checkRegMaskInterference(VRegLI, RegMaskOverlaps);
609 
610     // Compute an initial allowed set for the current vreg.
611     std::vector<unsigned> VRegAllowed;
612     ArrayRef<MCPhysReg> RawPRegOrder = TRC->getRawAllocationOrder(MF);
613     for (unsigned I = 0; I != RawPRegOrder.size(); ++I) {
614       unsigned PReg = RawPRegOrder[I];
615       if (MRI.isReserved(PReg))
616         continue;
617 
618       // vregLI crosses a regmask operand that clobbers preg.
619       if (!RegMaskOverlaps.empty() && !RegMaskOverlaps.test(PReg))
620         continue;
621 
622       // vregLI overlaps fixed regunit interference.
623       bool Interference = false;
624       for (MCRegUnitIterator Units(PReg, &TRI); Units.isValid(); ++Units) {
625         if (VRegLI.overlaps(LIS.getRegUnit(*Units))) {
626           Interference = true;
627           break;
628         }
629       }
630       if (Interference)
631         continue;
632 
633       // preg is usable for this virtual register.
634       VRegAllowed.push_back(PReg);
635     }
636 
637     // Check for vregs that have no allowed registers. These should be
638     // pre-spilled and the new vregs added to the worklist.
639     if (VRegAllowed.empty()) {
640       SmallVector<Register, 8> NewVRegs;
641       spillVReg(VReg, NewVRegs, MF, LIS, VRM, VRegSpiller);
642       Worklist.insert(Worklist.end(), NewVRegs.begin(), NewVRegs.end());
643       continue;
644     } else
645       VRegAllowedMap[VReg] = std::move(VRegAllowed);
646   }
647 
648   for (auto &KV : VRegAllowedMap) {
649     auto VReg = KV.first;
650 
651     // Move empty intervals to the EmptyIntervalVReg set.
652     if (LIS.getInterval(VReg).empty()) {
653       EmptyIntervalVRegs.insert(VReg);
654       VRegsToAlloc.erase(VReg);
655       continue;
656     }
657 
658     auto &VRegAllowed = KV.second;
659 
660     PBQPRAGraph::RawVector NodeCosts(VRegAllowed.size() + 1, 0);
661 
662     // Tweak cost of callee saved registers, as using then force spilling and
663     // restoring them. This would only happen in the prologue / epilogue though.
664     for (unsigned i = 0; i != VRegAllowed.size(); ++i)
665       if (isACalleeSavedRegister(VRegAllowed[i], TRI, MF))
666         NodeCosts[1 + i] += 1.0;
667 
668     PBQPRAGraph::NodeId NId = G.addNode(std::move(NodeCosts));
669     G.getNodeMetadata(NId).setVReg(VReg);
670     G.getNodeMetadata(NId).setAllowedRegs(
671       G.getMetadata().getAllowedRegs(std::move(VRegAllowed)));
672     G.getMetadata().setNodeIdForVReg(VReg, NId);
673   }
674 }
675 
676 void RegAllocPBQP::spillVReg(Register VReg,
677                              SmallVectorImpl<Register> &NewIntervals,
678                              MachineFunction &MF, LiveIntervals &LIS,
679                              VirtRegMap &VRM, Spiller &VRegSpiller) {
680   VRegsToAlloc.erase(VReg);
681   LiveRangeEdit LRE(&LIS.getInterval(VReg), NewIntervals, MF, LIS, &VRM,
682                     nullptr, &DeadRemats);
683   VRegSpiller.spill(LRE);
684 
685   const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
686   (void)TRI;
687   LLVM_DEBUG(dbgs() << "VREG " << printReg(VReg, &TRI) << " -> SPILLED (Cost: "
688                     << LRE.getParent().weight << ", New vregs: ");
689 
690   // Copy any newly inserted live intervals into the list of regs to
691   // allocate.
692   for (LiveRangeEdit::iterator I = LRE.begin(), E = LRE.end();
693        I != E; ++I) {
694     const LiveInterval &LI = LIS.getInterval(*I);
695     assert(!LI.empty() && "Empty spill range.");
696     LLVM_DEBUG(dbgs() << printReg(LI.reg, &TRI) << " ");
697     VRegsToAlloc.insert(LI.reg);
698   }
699 
700   LLVM_DEBUG(dbgs() << ")\n");
701 }
702 
703 bool RegAllocPBQP::mapPBQPToRegAlloc(const PBQPRAGraph &G,
704                                      const PBQP::Solution &Solution,
705                                      VirtRegMap &VRM,
706                                      Spiller &VRegSpiller) {
707   MachineFunction &MF = G.getMetadata().MF;
708   LiveIntervals &LIS = G.getMetadata().LIS;
709   const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
710   (void)TRI;
711 
712   // Set to true if we have any spills
713   bool AnotherRoundNeeded = false;
714 
715   // Clear the existing allocation.
716   VRM.clearAllVirt();
717 
718   // Iterate over the nodes mapping the PBQP solution to a register
719   // assignment.
720   for (auto NId : G.nodeIds()) {
721     unsigned VReg = G.getNodeMetadata(NId).getVReg();
722     unsigned AllocOption = Solution.getSelection(NId);
723 
724     if (AllocOption != PBQP::RegAlloc::getSpillOptionIdx()) {
725       unsigned PReg = G.getNodeMetadata(NId).getAllowedRegs()[AllocOption - 1];
726       LLVM_DEBUG(dbgs() << "VREG " << printReg(VReg, &TRI) << " -> "
727                         << TRI.getName(PReg) << "\n");
728       assert(PReg != 0 && "Invalid preg selected.");
729       VRM.assignVirt2Phys(VReg, PReg);
730     } else {
731       // Spill VReg. If this introduces new intervals we'll need another round
732       // of allocation.
733       SmallVector<Register, 8> NewVRegs;
734       spillVReg(VReg, NewVRegs, MF, LIS, VRM, VRegSpiller);
735       AnotherRoundNeeded |= !NewVRegs.empty();
736     }
737   }
738 
739   return !AnotherRoundNeeded;
740 }
741 
742 void RegAllocPBQP::finalizeAlloc(MachineFunction &MF,
743                                  LiveIntervals &LIS,
744                                  VirtRegMap &VRM) const {
745   MachineRegisterInfo &MRI = MF.getRegInfo();
746 
747   // First allocate registers for the empty intervals.
748   for (RegSet::const_iterator
749          I = EmptyIntervalVRegs.begin(), E = EmptyIntervalVRegs.end();
750          I != E; ++I) {
751     LiveInterval &LI = LIS.getInterval(*I);
752 
753     unsigned PReg = MRI.getSimpleHint(LI.reg);
754 
755     if (PReg == 0) {
756       const TargetRegisterClass &RC = *MRI.getRegClass(LI.reg);
757       const ArrayRef<MCPhysReg> RawPRegOrder = RC.getRawAllocationOrder(MF);
758       for (unsigned CandidateReg : RawPRegOrder) {
759         if (!VRM.getRegInfo().isReserved(CandidateReg)) {
760           PReg = CandidateReg;
761           break;
762         }
763       }
764       assert(PReg &&
765              "No un-reserved physical registers in this register class");
766     }
767 
768     VRM.assignVirt2Phys(LI.reg, PReg);
769   }
770 }
771 
772 void RegAllocPBQP::postOptimization(Spiller &VRegSpiller, LiveIntervals &LIS) {
773   VRegSpiller.postOptimization();
774   /// Remove dead defs because of rematerialization.
775   for (auto DeadInst : DeadRemats) {
776     LIS.RemoveMachineInstrFromMaps(*DeadInst);
777     DeadInst->eraseFromParent();
778   }
779   DeadRemats.clear();
780 }
781 
782 static inline float normalizePBQPSpillWeight(float UseDefFreq, unsigned Size,
783                                          unsigned NumInstr) {
784   // All intervals have a spill weight that is mostly proportional to the number
785   // of uses, with uses in loops having a bigger weight.
786   return NumInstr * normalizeSpillWeight(UseDefFreq, Size, 1);
787 }
788 
789 bool RegAllocPBQP::runOnMachineFunction(MachineFunction &MF) {
790   LiveIntervals &LIS = getAnalysis<LiveIntervals>();
791   MachineBlockFrequencyInfo &MBFI =
792     getAnalysis<MachineBlockFrequencyInfo>();
793 
794   VirtRegMap &VRM = getAnalysis<VirtRegMap>();
795 
796   calculateSpillWeightsAndHints(LIS, MF, &VRM, getAnalysis<MachineLoopInfo>(),
797                                 MBFI, normalizePBQPSpillWeight);
798 
799   std::unique_ptr<Spiller> VRegSpiller(createInlineSpiller(*this, MF, VRM));
800 
801   MF.getRegInfo().freezeReservedRegs(MF);
802 
803   LLVM_DEBUG(dbgs() << "PBQP Register Allocating for " << MF.getName() << "\n");
804 
805   // Allocator main loop:
806   //
807   // * Map current regalloc problem to a PBQP problem
808   // * Solve the PBQP problem
809   // * Map the solution back to a register allocation
810   // * Spill if necessary
811   //
812   // This process is continued till no more spills are generated.
813 
814   // Find the vreg intervals in need of allocation.
815   findVRegIntervalsToAlloc(MF, LIS);
816 
817 #ifndef NDEBUG
818   const Function &F = MF.getFunction();
819   std::string FullyQualifiedName =
820     F.getParent()->getModuleIdentifier() + "." + F.getName().str();
821 #endif
822 
823   // If there are non-empty intervals allocate them using pbqp.
824   if (!VRegsToAlloc.empty()) {
825     const TargetSubtargetInfo &Subtarget = MF.getSubtarget();
826     std::unique_ptr<PBQPRAConstraintList> ConstraintsRoot =
827       std::make_unique<PBQPRAConstraintList>();
828     ConstraintsRoot->addConstraint(std::make_unique<SpillCosts>());
829     ConstraintsRoot->addConstraint(std::make_unique<Interference>());
830     if (PBQPCoalescing)
831       ConstraintsRoot->addConstraint(std::make_unique<Coalescing>());
832     ConstraintsRoot->addConstraint(Subtarget.getCustomPBQPConstraints());
833 
834     bool PBQPAllocComplete = false;
835     unsigned Round = 0;
836 
837     while (!PBQPAllocComplete) {
838       LLVM_DEBUG(dbgs() << "  PBQP Regalloc round " << Round << ":\n");
839 
840       PBQPRAGraph G(PBQPRAGraph::GraphMetadata(MF, LIS, MBFI));
841       initializeGraph(G, VRM, *VRegSpiller);
842       ConstraintsRoot->apply(G);
843 
844 #ifndef NDEBUG
845       if (PBQPDumpGraphs) {
846         std::ostringstream RS;
847         RS << Round;
848         std::string GraphFileName = FullyQualifiedName + "." + RS.str() +
849                                     ".pbqpgraph";
850         std::error_code EC;
851         raw_fd_ostream OS(GraphFileName, EC, sys::fs::OF_Text);
852         LLVM_DEBUG(dbgs() << "Dumping graph for round " << Round << " to \""
853                           << GraphFileName << "\"\n");
854         G.dump(OS);
855       }
856 #endif
857 
858       PBQP::Solution Solution = PBQP::RegAlloc::solve(G);
859       PBQPAllocComplete = mapPBQPToRegAlloc(G, Solution, VRM, *VRegSpiller);
860       ++Round;
861     }
862   }
863 
864   // Finalise allocation, allocate empty ranges.
865   finalizeAlloc(MF, LIS, VRM);
866   postOptimization(*VRegSpiller, LIS);
867   VRegsToAlloc.clear();
868   EmptyIntervalVRegs.clear();
869 
870   LLVM_DEBUG(dbgs() << "Post alloc VirtRegMap:\n" << VRM << "\n");
871 
872   return true;
873 }
874 
875 /// Create Printable object for node and register info.
876 static Printable PrintNodeInfo(PBQP::RegAlloc::PBQPRAGraph::NodeId NId,
877                                const PBQP::RegAlloc::PBQPRAGraph &G) {
878   return Printable([NId, &G](raw_ostream &OS) {
879     const MachineRegisterInfo &MRI = G.getMetadata().MF.getRegInfo();
880     const TargetRegisterInfo *TRI = MRI.getTargetRegisterInfo();
881     unsigned VReg = G.getNodeMetadata(NId).getVReg();
882     const char *RegClassName = TRI->getRegClassName(MRI.getRegClass(VReg));
883     OS << NId << " (" << RegClassName << ':' << printReg(VReg, TRI) << ')';
884   });
885 }
886 
887 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
888 LLVM_DUMP_METHOD void PBQP::RegAlloc::PBQPRAGraph::dump(raw_ostream &OS) const {
889   for (auto NId : nodeIds()) {
890     const Vector &Costs = getNodeCosts(NId);
891     assert(Costs.getLength() != 0 && "Empty vector in graph.");
892     OS << PrintNodeInfo(NId, *this) << ": " << Costs << '\n';
893   }
894   OS << '\n';
895 
896   for (auto EId : edgeIds()) {
897     NodeId N1Id = getEdgeNode1Id(EId);
898     NodeId N2Id = getEdgeNode2Id(EId);
899     assert(N1Id != N2Id && "PBQP graphs should not have self-edges.");
900     const Matrix &M = getEdgeCosts(EId);
901     assert(M.getRows() != 0 && "No rows in matrix.");
902     assert(M.getCols() != 0 && "No cols in matrix.");
903     OS << PrintNodeInfo(N1Id, *this) << ' ' << M.getRows() << " rows / ";
904     OS << PrintNodeInfo(N2Id, *this) << ' ' << M.getCols() << " cols:\n";
905     OS << M << '\n';
906   }
907 }
908 
909 LLVM_DUMP_METHOD void PBQP::RegAlloc::PBQPRAGraph::dump() const {
910   dump(dbgs());
911 }
912 #endif
913 
914 void PBQP::RegAlloc::PBQPRAGraph::printDot(raw_ostream &OS) const {
915   OS << "graph {\n";
916   for (auto NId : nodeIds()) {
917     OS << "  node" << NId << " [ label=\""
918        << PrintNodeInfo(NId, *this) << "\\n"
919        << getNodeCosts(NId) << "\" ]\n";
920   }
921 
922   OS << "  edge [ len=" << nodeIds().size() << " ]\n";
923   for (auto EId : edgeIds()) {
924     OS << "  node" << getEdgeNode1Id(EId)
925        << " -- node" << getEdgeNode2Id(EId)
926        << " [ label=\"";
927     const Matrix &EdgeCosts = getEdgeCosts(EId);
928     for (unsigned i = 0; i < EdgeCosts.getRows(); ++i) {
929       OS << EdgeCosts.getRowAsVector(i) << "\\n";
930     }
931     OS << "\" ]\n";
932   }
933   OS << "}\n";
934 }
935 
936 FunctionPass *llvm::createPBQPRegisterAllocator(char *customPassID) {
937   return new RegAllocPBQP(customPassID);
938 }
939 
940 FunctionPass* llvm::createDefaultPBQPRegisterAllocator() {
941   return createPBQPRegisterAllocator();
942 }
943