RegAllocPBQP.cpp revision 233a60ec40b41027ff429e2f2c27fa2be762f2e9
1//===------ RegAllocPBQP.cpp ---- PBQP Register Allocator -------*- 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//
10// This file contains a Partitioned Boolean Quadratic Programming (PBQP) based
11// register allocator for LLVM. This allocator works by constructing a PBQP
12// problem representing the register allocation problem under consideration,
13// solving this using a PBQP solver, and mapping the solution back to a
14// register assignment. If any variables are selected for spilling then spill
15// code is inserted and the process repeated.
16//
17// The PBQP solver (pbqp.c) provided for this allocator uses a heuristic tuned
18// for register allocation. For more information on PBQP for register
19// allocation, see the following papers:
20//
21//   (1) Hames, L. and Scholz, B. 2006. Nearly optimal register allocation with
22//   PBQP. In Proceedings of the 7th Joint Modular Languages Conference
23//   (JMLC'06). LNCS, vol. 4228. Springer, New York, NY, USA. 346-361.
24//
25//   (2) Scholz, B., Eckstein, E. 2002. Register allocation for irregular
26//   architectures. In Proceedings of the Joint Conference on Languages,
27//   Compilers and Tools for Embedded Systems (LCTES'02), ACM Press, New York,
28//   NY, USA, 139-148.
29//
30//===----------------------------------------------------------------------===//
31
32#define DEBUG_TYPE "regalloc"
33
34#include "PBQP/HeuristicSolver.h"
35#include "PBQP/SimpleGraph.h"
36#include "PBQP/Heuristics/Briggs.h"
37#include "VirtRegMap.h"
38#include "VirtRegRewriter.h"
39#include "llvm/CodeGen/LiveIntervalAnalysis.h"
40#include "llvm/CodeGen/LiveStackAnalysis.h"
41#include "llvm/CodeGen/MachineFunctionPass.h"
42#include "llvm/CodeGen/MachineLoopInfo.h"
43#include "llvm/CodeGen/MachineRegisterInfo.h"
44#include "llvm/CodeGen/RegAllocRegistry.h"
45#include "llvm/CodeGen/RegisterCoalescer.h"
46#include "llvm/Support/Debug.h"
47#include "llvm/Support/raw_ostream.h"
48#include "llvm/Target/TargetInstrInfo.h"
49#include "llvm/Target/TargetMachine.h"
50#include <limits>
51#include <map>
52#include <memory>
53#include <set>
54#include <vector>
55
56using namespace llvm;
57
58static RegisterRegAlloc
59registerPBQPRepAlloc("pbqp", "PBQP register allocator.",
60                      llvm::createPBQPRegisterAllocator);
61
62static cl::opt<bool>
63pbqpCoalescing("pbqp-coalescing",
64               cl::desc("Attempt coalescing during PBQP register allocation."),
65               cl::init(false), cl::Hidden);
66
67namespace {
68
69  ///
70  /// PBQP based allocators solve the register allocation problem by mapping
71  /// register allocation problems to Partitioned Boolean Quadratic
72  /// Programming problems.
73  class PBQPRegAlloc : public MachineFunctionPass {
74  public:
75
76    static char ID;
77
78    /// Construct a PBQP register allocator.
79    PBQPRegAlloc() : MachineFunctionPass(&ID) {}
80
81    /// Return the pass name.
82    virtual const char* getPassName() const {
83      return "PBQP Register Allocator";
84    }
85
86    /// PBQP analysis usage.
87    virtual void getAnalysisUsage(AnalysisUsage &au) const {
88      au.addRequired<SlotIndexes>();
89      au.addPreserved<SlotIndexes>();
90      au.addRequired<LiveIntervals>();
91      //au.addRequiredID(SplitCriticalEdgesID);
92      au.addRequired<RegisterCoalescer>();
93      au.addRequired<LiveStacks>();
94      au.addPreserved<LiveStacks>();
95      au.addRequired<MachineLoopInfo>();
96      au.addPreserved<MachineLoopInfo>();
97      au.addRequired<VirtRegMap>();
98      MachineFunctionPass::getAnalysisUsage(au);
99    }
100
101    /// Perform register allocation
102    virtual bool runOnMachineFunction(MachineFunction &MF);
103
104  private:
105    typedef std::map<const LiveInterval*, unsigned> LI2NodeMap;
106    typedef std::vector<const LiveInterval*> Node2LIMap;
107    typedef std::vector<unsigned> AllowedSet;
108    typedef std::vector<AllowedSet> AllowedSetMap;
109    typedef std::set<unsigned> RegSet;
110    typedef std::pair<unsigned, unsigned> RegPair;
111    typedef std::map<RegPair, PBQP::PBQPNum> CoalesceMap;
112
113    typedef std::set<LiveInterval*> LiveIntervalSet;
114
115    MachineFunction *mf;
116    const TargetMachine *tm;
117    const TargetRegisterInfo *tri;
118    const TargetInstrInfo *tii;
119    const MachineLoopInfo *loopInfo;
120    MachineRegisterInfo *mri;
121
122    LiveIntervals *lis;
123    LiveStacks *lss;
124    VirtRegMap *vrm;
125
126    LI2NodeMap li2Node;
127    Node2LIMap node2LI;
128    AllowedSetMap allowedSets;
129    LiveIntervalSet vregIntervalsToAlloc,
130                    emptyVRegIntervals;
131
132
133    /// Builds a PBQP cost vector.
134    template <typename RegContainer>
135    PBQP::Vector buildCostVector(unsigned vReg,
136                                 const RegContainer &allowed,
137                                 const CoalesceMap &cealesces,
138                                 PBQP::PBQPNum spillCost) const;
139
140    /// \brief Builds a PBQP interference matrix.
141    ///
142    /// @return Either a pointer to a non-zero PBQP matrix representing the
143    ///         allocation option costs, or a null pointer for a zero matrix.
144    ///
145    /// Expects allowed sets for two interfering LiveIntervals. These allowed
146    /// sets should contain only allocable registers from the LiveInterval's
147    /// register class, with any interfering pre-colored registers removed.
148    template <typename RegContainer>
149    PBQP::Matrix* buildInterferenceMatrix(const RegContainer &allowed1,
150                                          const RegContainer &allowed2) const;
151
152    ///
153    /// Expects allowed sets for two potentially coalescable LiveIntervals,
154    /// and an estimated benefit due to coalescing. The allowed sets should
155    /// contain only allocable registers from the LiveInterval's register
156    /// classes, with any interfering pre-colored registers removed.
157    template <typename RegContainer>
158    PBQP::Matrix* buildCoalescingMatrix(const RegContainer &allowed1,
159                                        const RegContainer &allowed2,
160                                        PBQP::PBQPNum cBenefit) const;
161
162    /// \brief Finds coalescing opportunities and returns them as a map.
163    ///
164    /// Any entries in the map are guaranteed coalescable, even if their
165    /// corresponding live intervals overlap.
166    CoalesceMap findCoalesces();
167
168    /// \brief Finds the initial set of vreg intervals to allocate.
169    void findVRegIntervalsToAlloc();
170
171    /// \brief Constructs a PBQP problem representation of the register
172    /// allocation problem for this function.
173    ///
174    /// @return a PBQP solver object for the register allocation problem.
175    PBQP::SimpleGraph constructPBQPProblem();
176
177    /// \brief Adds a stack interval if the given live interval has been
178    /// spilled. Used to support stack slot coloring.
179    void addStackInterval(const LiveInterval *spilled,MachineRegisterInfo* mri);
180
181    /// \brief Given a solved PBQP problem maps this solution back to a register
182    /// assignment.
183    bool mapPBQPToRegAlloc(const PBQP::Solution &solution);
184
185    /// \brief Postprocessing before final spilling. Sets basic block "live in"
186    /// variables.
187    void finalizeAlloc() const;
188
189  };
190
191  char PBQPRegAlloc::ID = 0;
192}
193
194
195template <typename RegContainer>
196PBQP::Vector PBQPRegAlloc::buildCostVector(unsigned vReg,
197                                           const RegContainer &allowed,
198                                           const CoalesceMap &coalesces,
199                                           PBQP::PBQPNum spillCost) const {
200
201  typedef typename RegContainer::const_iterator AllowedItr;
202
203  // Allocate vector. Additional element (0th) used for spill option
204  PBQP::Vector v(allowed.size() + 1, 0);
205
206  v[0] = spillCost;
207
208  // Iterate over the allowed registers inserting coalesce benefits if there
209  // are any.
210  unsigned ai = 0;
211  for (AllowedItr itr = allowed.begin(), end = allowed.end();
212       itr != end; ++itr, ++ai) {
213
214    unsigned pReg = *itr;
215
216    CoalesceMap::const_iterator cmItr =
217      coalesces.find(RegPair(vReg, pReg));
218
219    // No coalesce - on to the next preg.
220    if (cmItr == coalesces.end())
221      continue;
222
223    // We have a coalesce - insert the benefit.
224    v[ai + 1] = -cmItr->second;
225  }
226
227  return v;
228}
229
230template <typename RegContainer>
231PBQP::Matrix* PBQPRegAlloc::buildInterferenceMatrix(
232      const RegContainer &allowed1, const RegContainer &allowed2) const {
233
234  typedef typename RegContainer::const_iterator RegContainerIterator;
235
236  // Construct a PBQP matrix representing the cost of allocation options. The
237  // rows and columns correspond to the allocation options for the two live
238  // intervals.  Elements will be infinite where corresponding registers alias,
239  // since we cannot allocate aliasing registers to interfering live intervals.
240  // All other elements (non-aliasing combinations) will have zero cost. Note
241  // that the spill option (element 0,0) has zero cost, since we can allocate
242  // both intervals to memory safely (the cost for each individual allocation
243  // to memory is accounted for by the cost vectors for each live interval).
244  PBQP::Matrix *m =
245    new PBQP::Matrix(allowed1.size() + 1, allowed2.size() + 1, 0);
246
247  // Assume this is a zero matrix until proven otherwise.  Zero matrices occur
248  // between interfering live ranges with non-overlapping register sets (e.g.
249  // non-overlapping reg classes, or disjoint sets of allowed regs within the
250  // same class). The term "overlapping" is used advisedly: sets which do not
251  // intersect, but contain registers which alias, will have non-zero matrices.
252  // We optimize zero matrices away to improve solver speed.
253  bool isZeroMatrix = true;
254
255
256  // Row index. Starts at 1, since the 0th row is for the spill option, which
257  // is always zero.
258  unsigned ri = 1;
259
260  // Iterate over allowed sets, insert infinities where required.
261  for (RegContainerIterator a1Itr = allowed1.begin(), a1End = allowed1.end();
262       a1Itr != a1End; ++a1Itr) {
263
264    // Column index, starts at 1 as for row index.
265    unsigned ci = 1;
266    unsigned reg1 = *a1Itr;
267
268    for (RegContainerIterator a2Itr = allowed2.begin(), a2End = allowed2.end();
269         a2Itr != a2End; ++a2Itr) {
270
271      unsigned reg2 = *a2Itr;
272
273      // If the row/column regs are identical or alias insert an infinity.
274      if (tri->regsOverlap(reg1, reg2)) {
275        (*m)[ri][ci] = std::numeric_limits<PBQP::PBQPNum>::infinity();
276        isZeroMatrix = false;
277      }
278
279      ++ci;
280    }
281
282    ++ri;
283  }
284
285  // If this turns out to be a zero matrix...
286  if (isZeroMatrix) {
287    // free it and return null.
288    delete m;
289    return 0;
290  }
291
292  // ...otherwise return the cost matrix.
293  return m;
294}
295
296template <typename RegContainer>
297PBQP::Matrix* PBQPRegAlloc::buildCoalescingMatrix(
298      const RegContainer &allowed1, const RegContainer &allowed2,
299      PBQP::PBQPNum cBenefit) const {
300
301  typedef typename RegContainer::const_iterator RegContainerIterator;
302
303  // Construct a PBQP Matrix representing the benefits of coalescing. As with
304  // interference matrices the rows and columns represent allowed registers
305  // for the LiveIntervals which are (potentially) to be coalesced. The amount
306  // -cBenefit will be placed in any element representing the same register
307  // for both intervals.
308  PBQP::Matrix *m =
309    new PBQP::Matrix(allowed1.size() + 1, allowed2.size() + 1, 0);
310
311  // Reset costs to zero.
312  m->reset(0);
313
314  // Assume the matrix is zero till proven otherwise. Zero matrices will be
315  // optimized away as in the interference case.
316  bool isZeroMatrix = true;
317
318  // Row index. Starts at 1, since the 0th row is for the spill option, which
319  // is always zero.
320  unsigned ri = 1;
321
322  // Iterate over the allowed sets, insert coalescing benefits where
323  // appropriate.
324  for (RegContainerIterator a1Itr = allowed1.begin(), a1End = allowed1.end();
325       a1Itr != a1End; ++a1Itr) {
326
327    // Column index, starts at 1 as for row index.
328    unsigned ci = 1;
329    unsigned reg1 = *a1Itr;
330
331    for (RegContainerIterator a2Itr = allowed2.begin(), a2End = allowed2.end();
332         a2Itr != a2End; ++a2Itr) {
333
334      // If the row and column represent the same register insert a beneficial
335      // cost to preference this allocation - it would allow us to eliminate a
336      // move instruction.
337      if (reg1 == *a2Itr) {
338        (*m)[ri][ci] = -cBenefit;
339        isZeroMatrix = false;
340      }
341
342      ++ci;
343    }
344
345    ++ri;
346  }
347
348  // If this turns out to be a zero matrix...
349  if (isZeroMatrix) {
350    // ...free it and return null.
351    delete m;
352    return 0;
353  }
354
355  return m;
356}
357
358PBQPRegAlloc::CoalesceMap PBQPRegAlloc::findCoalesces() {
359
360  typedef MachineFunction::const_iterator MFIterator;
361  typedef MachineBasicBlock::const_iterator MBBIterator;
362  typedef LiveInterval::const_vni_iterator VNIIterator;
363
364  CoalesceMap coalescesFound;
365
366  // To find coalesces we need to iterate over the function looking for
367  // copy instructions.
368  for (MFIterator bbItr = mf->begin(), bbEnd = mf->end();
369       bbItr != bbEnd; ++bbItr) {
370
371    const MachineBasicBlock *mbb = &*bbItr;
372
373    for (MBBIterator iItr = mbb->begin(), iEnd = mbb->end();
374         iItr != iEnd; ++iItr) {
375
376      const MachineInstr *instr = &*iItr;
377      unsigned srcReg, dstReg, srcSubReg, dstSubReg;
378
379      // If this isn't a copy then continue to the next instruction.
380      if (!tii->isMoveInstr(*instr, srcReg, dstReg, srcSubReg, dstSubReg))
381        continue;
382
383      // If the registers are already the same our job is nice and easy.
384      if (dstReg == srcReg)
385        continue;
386
387      bool srcRegIsPhysical = TargetRegisterInfo::isPhysicalRegister(srcReg),
388           dstRegIsPhysical = TargetRegisterInfo::isPhysicalRegister(dstReg);
389
390      // If both registers are physical then we can't coalesce.
391      if (srcRegIsPhysical && dstRegIsPhysical)
392        continue;
393
394      // If it's a copy that includes a virtual register but the source and
395      // destination classes differ then we can't coalesce, so continue with
396      // the next instruction.
397      const TargetRegisterClass *srcRegClass = srcRegIsPhysical ?
398          tri->getPhysicalRegisterRegClass(srcReg) : mri->getRegClass(srcReg);
399
400      const TargetRegisterClass *dstRegClass = dstRegIsPhysical ?
401          tri->getPhysicalRegisterRegClass(dstReg) : mri->getRegClass(dstReg);
402
403      if (srcRegClass != dstRegClass)
404        continue;
405
406      // We also need any physical regs to be allocable, coalescing with
407      // a non-allocable register is invalid.
408      if (srcRegIsPhysical) {
409        if (std::find(srcRegClass->allocation_order_begin(*mf),
410                      srcRegClass->allocation_order_end(*mf), srcReg) ==
411            srcRegClass->allocation_order_end(*mf))
412          continue;
413      }
414
415      if (dstRegIsPhysical) {
416        if (std::find(dstRegClass->allocation_order_begin(*mf),
417                      dstRegClass->allocation_order_end(*mf), dstReg) ==
418            dstRegClass->allocation_order_end(*mf))
419          continue;
420      }
421
422      // If we've made it here we have a copy with compatible register classes.
423      // We can probably coalesce, but we need to consider overlap.
424      const LiveInterval *srcLI = &lis->getInterval(srcReg),
425                         *dstLI = &lis->getInterval(dstReg);
426
427      if (srcLI->overlaps(*dstLI)) {
428        // Even in the case of an overlap we might still be able to coalesce,
429        // but we need to make sure that no definition of either range occurs
430        // while the other range is live.
431
432        // Otherwise start by assuming we're ok.
433        bool badDef = false;
434
435        // Test all defs of the source range.
436        for (VNIIterator
437               vniItr = srcLI->vni_begin(), vniEnd = srcLI->vni_end();
438               vniItr != vniEnd; ++vniItr) {
439
440          // If we find a def that kills the coalescing opportunity then
441          // record it and break from the loop.
442          if (dstLI->liveAt((*vniItr)->def)) {
443            badDef = true;
444            break;
445          }
446        }
447
448        // If we have a bad def give up, continue to the next instruction.
449        if (badDef)
450          continue;
451
452        // Otherwise test definitions of the destination range.
453        for (VNIIterator
454               vniItr = dstLI->vni_begin(), vniEnd = dstLI->vni_end();
455               vniItr != vniEnd; ++vniItr) {
456
457          // We want to make sure we skip the copy instruction itself.
458          if ((*vniItr)->getCopy() == instr)
459            continue;
460
461          if (srcLI->liveAt((*vniItr)->def)) {
462            badDef = true;
463            break;
464          }
465        }
466
467        // As before a bad def we give up and continue to the next instr.
468        if (badDef)
469          continue;
470      }
471
472      // If we make it to here then either the ranges didn't overlap, or they
473      // did, but none of their definitions would prevent us from coalescing.
474      // We're good to go with the coalesce.
475
476      float cBenefit = powf(10.0f, loopInfo->getLoopDepth(mbb)) / 5.0;
477
478      coalescesFound[RegPair(srcReg, dstReg)] = cBenefit;
479      coalescesFound[RegPair(dstReg, srcReg)] = cBenefit;
480    }
481
482  }
483
484  return coalescesFound;
485}
486
487void PBQPRegAlloc::findVRegIntervalsToAlloc() {
488
489  // Iterate over all live ranges.
490  for (LiveIntervals::iterator itr = lis->begin(), end = lis->end();
491       itr != end; ++itr) {
492
493    // Ignore physical ones.
494    if (TargetRegisterInfo::isPhysicalRegister(itr->first))
495      continue;
496
497    LiveInterval *li = itr->second;
498
499    // If this live interval is non-empty we will use pbqp to allocate it.
500    // Empty intervals we allocate in a simple post-processing stage in
501    // finalizeAlloc.
502    if (!li->empty()) {
503      vregIntervalsToAlloc.insert(li);
504    }
505    else {
506      emptyVRegIntervals.insert(li);
507    }
508  }
509}
510
511PBQP::SimpleGraph PBQPRegAlloc::constructPBQPProblem() {
512
513  typedef std::vector<const LiveInterval*> LIVector;
514  typedef std::vector<unsigned> RegVector;
515  typedef std::vector<PBQP::SimpleGraph::NodeIterator> NodeVector;
516
517  // This will store the physical intervals for easy reference.
518  LIVector physIntervals;
519
520  // Start by clearing the old node <-> live interval mappings & allowed sets
521  li2Node.clear();
522  node2LI.clear();
523  allowedSets.clear();
524
525  // Populate physIntervals, update preg use:
526  for (LiveIntervals::iterator itr = lis->begin(), end = lis->end();
527       itr != end; ++itr) {
528
529    if (TargetRegisterInfo::isPhysicalRegister(itr->first)) {
530      physIntervals.push_back(itr->second);
531      mri->setPhysRegUsed(itr->second->reg);
532    }
533  }
534
535  // Iterate over vreg intervals, construct live interval <-> node number
536  //  mappings.
537  for (LiveIntervalSet::const_iterator
538       itr = vregIntervalsToAlloc.begin(), end = vregIntervalsToAlloc.end();
539       itr != end; ++itr) {
540    const LiveInterval *li = *itr;
541
542    li2Node[li] = node2LI.size();
543    node2LI.push_back(li);
544  }
545
546  // Get the set of potential coalesces.
547  CoalesceMap coalesces;
548
549  if (pbqpCoalescing) {
550    coalesces = findCoalesces();
551  }
552
553  // Construct a PBQP solver for this problem
554  PBQP::SimpleGraph problem;
555  NodeVector problemNodes(vregIntervalsToAlloc.size());
556
557  // Resize allowedSets container appropriately.
558  allowedSets.resize(vregIntervalsToAlloc.size());
559
560  // Iterate over virtual register intervals to compute allowed sets...
561  for (unsigned node = 0; node < node2LI.size(); ++node) {
562
563    // Grab pointers to the interval and its register class.
564    const LiveInterval *li = node2LI[node];
565    const TargetRegisterClass *liRC = mri->getRegClass(li->reg);
566
567    // Start by assuming all allocable registers in the class are allowed...
568    RegVector liAllowed(liRC->allocation_order_begin(*mf),
569                        liRC->allocation_order_end(*mf));
570
571    // Eliminate the physical registers which overlap with this range, along
572    // with all their aliases.
573    for (LIVector::iterator pItr = physIntervals.begin(),
574       pEnd = physIntervals.end(); pItr != pEnd; ++pItr) {
575
576      if (!li->overlaps(**pItr))
577        continue;
578
579      unsigned pReg = (*pItr)->reg;
580
581      // If we get here then the live intervals overlap, but we're still ok
582      // if they're coalescable.
583      if (coalesces.find(RegPair(li->reg, pReg)) != coalesces.end())
584        continue;
585
586      // If we get here then we have a genuine exclusion.
587
588      // Remove the overlapping reg...
589      RegVector::iterator eraseItr =
590        std::find(liAllowed.begin(), liAllowed.end(), pReg);
591
592      if (eraseItr != liAllowed.end())
593        liAllowed.erase(eraseItr);
594
595      const unsigned *aliasItr = tri->getAliasSet(pReg);
596
597      if (aliasItr != 0) {
598        // ...and its aliases.
599        for (; *aliasItr != 0; ++aliasItr) {
600          RegVector::iterator eraseItr =
601            std::find(liAllowed.begin(), liAllowed.end(), *aliasItr);
602
603          if (eraseItr != liAllowed.end()) {
604            liAllowed.erase(eraseItr);
605          }
606        }
607      }
608    }
609
610    // Copy the allowed set into a member vector for use when constructing cost
611    // vectors & matrices, and mapping PBQP solutions back to assignments.
612    allowedSets[node] = AllowedSet(liAllowed.begin(), liAllowed.end());
613
614    // Set the spill cost to the interval weight, or epsilon if the
615    // interval weight is zero
616    PBQP::PBQPNum spillCost = (li->weight != 0.0) ?
617        li->weight : std::numeric_limits<PBQP::PBQPNum>::min();
618
619    // Build a cost vector for this interval.
620    problemNodes[node] =
621      problem.addNode(
622        buildCostVector(li->reg, allowedSets[node], coalesces, spillCost));
623
624  }
625
626
627  // Now add the cost matrices...
628  for (unsigned node1 = 0; node1 < node2LI.size(); ++node1) {
629    const LiveInterval *li = node2LI[node1];
630
631    // Test for live range overlaps and insert interference matrices.
632    for (unsigned node2 = node1 + 1; node2 < node2LI.size(); ++node2) {
633      const LiveInterval *li2 = node2LI[node2];
634
635      CoalesceMap::const_iterator cmItr =
636        coalesces.find(RegPair(li->reg, li2->reg));
637
638      PBQP::Matrix *m = 0;
639
640      if (cmItr != coalesces.end()) {
641        m = buildCoalescingMatrix(allowedSets[node1], allowedSets[node2],
642                                  cmItr->second);
643      }
644      else if (li->overlaps(*li2)) {
645        m = buildInterferenceMatrix(allowedSets[node1], allowedSets[node2]);
646      }
647
648      if (m != 0) {
649        problem.addEdge(problemNodes[node1],
650                        problemNodes[node2],
651                        *m);
652
653        delete m;
654      }
655    }
656  }
657
658  problem.assignNodeIDs();
659
660  assert(problem.getNumNodes() == allowedSets.size());
661  for (unsigned i = 0; i < allowedSets.size(); ++i) {
662    assert(problem.getNodeItr(i) == problemNodes[i]);
663  }
664/*
665  std::cerr << "Allocating for " << problem.getNumNodes() << " nodes, "
666            << problem.getNumEdges() << " edges.\n";
667
668  problem.printDot(std::cerr);
669*/
670  // We're done, PBQP problem constructed - return it.
671  return problem;
672}
673
674void PBQPRegAlloc::addStackInterval(const LiveInterval *spilled,
675                                    MachineRegisterInfo* mri) {
676  int stackSlot = vrm->getStackSlot(spilled->reg);
677
678  if (stackSlot == VirtRegMap::NO_STACK_SLOT)
679    return;
680
681  const TargetRegisterClass *RC = mri->getRegClass(spilled->reg);
682  LiveInterval &stackInterval = lss->getOrCreateInterval(stackSlot, RC);
683
684  VNInfo *vni;
685  if (stackInterval.getNumValNums() != 0)
686    vni = stackInterval.getValNumInfo(0);
687  else
688    vni = stackInterval.getNextValue(
689      SlotIndex(), 0, false, lss->getVNInfoAllocator());
690
691  LiveInterval &rhsInterval = lis->getInterval(spilled->reg);
692  stackInterval.MergeRangesInAsValue(rhsInterval, vni);
693}
694
695bool PBQPRegAlloc::mapPBQPToRegAlloc(const PBQP::Solution &solution) {
696  // Set to true if we have any spills
697  bool anotherRoundNeeded = false;
698
699  // Clear the existing allocation.
700  vrm->clearAllVirt();
701
702  // Iterate over the nodes mapping the PBQP solution to a register assignment.
703  for (unsigned node = 0; node < node2LI.size(); ++node) {
704    unsigned virtReg = node2LI[node]->reg,
705             allocSelection = solution.getSelection(node);
706
707
708    // If the PBQP solution is non-zero it's a physical register...
709    if (allocSelection != 0) {
710      // Get the physical reg, subtracting 1 to account for the spill option.
711      unsigned physReg = allowedSets[node][allocSelection - 1];
712
713      DEBUG(errs() << "VREG " << virtReg << " -> "
714                   << tri->getName(physReg) << "\n");
715
716      assert(physReg != 0);
717
718      // Add to the virt reg map and update the used phys regs.
719      vrm->assignVirt2Phys(virtReg, physReg);
720    }
721    // ...Otherwise it's a spill.
722    else {
723
724      // Make sure we ignore this virtual reg on the next round
725      // of allocation
726      vregIntervalsToAlloc.erase(&lis->getInterval(virtReg));
727
728      // Insert spill ranges for this live range
729      const LiveInterval *spillInterval = node2LI[node];
730      double oldSpillWeight = spillInterval->weight;
731      SmallVector<LiveInterval*, 8> spillIs;
732      std::vector<LiveInterval*> newSpills =
733        lis->addIntervalsForSpills(*spillInterval, spillIs, loopInfo, *vrm);
734      addStackInterval(spillInterval, mri);
735
736      (void) oldSpillWeight;
737      DEBUG(errs() << "VREG " << virtReg << " -> SPILLED (Cost: "
738                   << oldSpillWeight << ", New vregs: ");
739
740      // Copy any newly inserted live intervals into the list of regs to
741      // allocate.
742      for (std::vector<LiveInterval*>::const_iterator
743           itr = newSpills.begin(), end = newSpills.end();
744           itr != end; ++itr) {
745
746        assert(!(*itr)->empty() && "Empty spill range.");
747
748        DEBUG(errs() << (*itr)->reg << " ");
749
750        vregIntervalsToAlloc.insert(*itr);
751      }
752
753      DEBUG(errs() << ")\n");
754
755      // We need another round if spill intervals were added.
756      anotherRoundNeeded |= !newSpills.empty();
757    }
758  }
759
760  return !anotherRoundNeeded;
761}
762
763void PBQPRegAlloc::finalizeAlloc() const {
764  typedef LiveIntervals::iterator LIIterator;
765  typedef LiveInterval::Ranges::const_iterator LRIterator;
766
767  // First allocate registers for the empty intervals.
768  for (LiveIntervalSet::const_iterator
769         itr = emptyVRegIntervals.begin(), end = emptyVRegIntervals.end();
770         itr != end; ++itr) {
771    LiveInterval *li = *itr;
772
773    unsigned physReg = vrm->getRegAllocPref(li->reg);
774
775    if (physReg == 0) {
776      const TargetRegisterClass *liRC = mri->getRegClass(li->reg);
777      physReg = *liRC->allocation_order_begin(*mf);
778    }
779
780    vrm->assignVirt2Phys(li->reg, physReg);
781  }
782
783  // Finally iterate over the basic blocks to compute and set the live-in sets.
784  SmallVector<MachineBasicBlock*, 8> liveInMBBs;
785  MachineBasicBlock *entryMBB = &*mf->begin();
786
787  for (LIIterator liItr = lis->begin(), liEnd = lis->end();
788       liItr != liEnd; ++liItr) {
789
790    const LiveInterval *li = liItr->second;
791    unsigned reg = 0;
792
793    // Get the physical register for this interval
794    if (TargetRegisterInfo::isPhysicalRegister(li->reg)) {
795      reg = li->reg;
796    }
797    else if (vrm->isAssignedReg(li->reg)) {
798      reg = vrm->getPhys(li->reg);
799    }
800    else {
801      // Ranges which are assigned a stack slot only are ignored.
802      continue;
803    }
804
805    if (reg == 0) {
806      // Filter out zero regs - they're for intervals that were spilled.
807      continue;
808    }
809
810    // Iterate over the ranges of the current interval...
811    for (LRIterator lrItr = li->begin(), lrEnd = li->end();
812         lrItr != lrEnd; ++lrItr) {
813
814      // Find the set of basic blocks which this range is live into...
815      if (lis->findLiveInMBBs(lrItr->start, lrItr->end,  liveInMBBs)) {
816        // And add the physreg for this interval to their live-in sets.
817        for (unsigned i = 0; i < liveInMBBs.size(); ++i) {
818          if (liveInMBBs[i] != entryMBB) {
819            if (!liveInMBBs[i]->isLiveIn(reg)) {
820              liveInMBBs[i]->addLiveIn(reg);
821            }
822          }
823        }
824        liveInMBBs.clear();
825      }
826    }
827  }
828
829}
830
831bool PBQPRegAlloc::runOnMachineFunction(MachineFunction &MF) {
832
833  mf = &MF;
834  tm = &mf->getTarget();
835  tri = tm->getRegisterInfo();
836  tii = tm->getInstrInfo();
837  mri = &mf->getRegInfo();
838
839  lis = &getAnalysis<LiveIntervals>();
840  lss = &getAnalysis<LiveStacks>();
841  loopInfo = &getAnalysis<MachineLoopInfo>();
842
843  vrm = &getAnalysis<VirtRegMap>();
844
845  DEBUG(errs() << "PBQP2 Register Allocating for " << mf->getFunction()->getName() << "\n");
846
847  // Allocator main loop:
848  //
849  // * Map current regalloc problem to a PBQP problem
850  // * Solve the PBQP problem
851  // * Map the solution back to a register allocation
852  // * Spill if necessary
853  //
854  // This process is continued till no more spills are generated.
855
856  // Find the vreg intervals in need of allocation.
857  findVRegIntervalsToAlloc();
858
859  // If there aren't any then we're done here.
860  if (vregIntervalsToAlloc.empty() && emptyVRegIntervals.empty())
861    return true;
862
863  // If there are non-empty intervals allocate them using pbqp.
864  if (!vregIntervalsToAlloc.empty()) {
865
866    bool pbqpAllocComplete = false;
867    unsigned round = 0;
868
869    while (!pbqpAllocComplete) {
870      DEBUG(errs() << "  PBQP Regalloc round " << round << ":\n");
871
872      PBQP::SimpleGraph problem = constructPBQPProblem();
873      PBQP::HeuristicSolver<PBQP::Heuristics::Briggs> solver;
874      problem.assignNodeIDs();
875      PBQP::Solution solution = solver.solve(problem);
876
877      pbqpAllocComplete = mapPBQPToRegAlloc(solution);
878
879      ++round;
880    }
881  }
882
883  // Finalise allocation, allocate empty ranges.
884  finalizeAlloc();
885
886  vregIntervalsToAlloc.clear();
887  emptyVRegIntervals.clear();
888  li2Node.clear();
889  node2LI.clear();
890  allowedSets.clear();
891
892  DEBUG(errs() << "Post alloc VirtRegMap:\n" << *vrm << "\n");
893
894  // Run rewriter
895  std::auto_ptr<VirtRegRewriter> rewriter(createVirtRegRewriter());
896
897  rewriter->runOnMachineFunction(*mf, *vrm, lis);
898
899  return true;
900}
901
902FunctionPass* llvm::createPBQPRegisterAllocator() {
903  return new PBQPRegAlloc();
904}
905
906
907#undef DEBUG_TYPE
908