1 //===------ RegAllocPBQP.cpp ---- PBQP Register Allocator -------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
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.
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:
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.
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,
30 //===----------------------------------------------------------------------===//
32 #define DEBUG_TYPE "regalloc"
35 #include "VirtRegMap.h"
36 #include "RegisterCoalescer.h"
37 #include "llvm/Module.h"
38 #include "llvm/Analysis/AliasAnalysis.h"
39 #include "llvm/CodeGen/CalcSpillWeights.h"
40 #include "llvm/CodeGen/LiveIntervalAnalysis.h"
41 #include "llvm/CodeGen/LiveRangeEdit.h"
42 #include "llvm/CodeGen/LiveStackAnalysis.h"
43 #include "llvm/CodeGen/RegAllocPBQP.h"
44 #include "llvm/CodeGen/MachineDominators.h"
45 #include "llvm/CodeGen/MachineFunctionPass.h"
46 #include "llvm/CodeGen/MachineLoopInfo.h"
47 #include "llvm/CodeGen/MachineRegisterInfo.h"
48 #include "llvm/CodeGen/PBQP/HeuristicSolver.h"
49 #include "llvm/CodeGen/PBQP/Graph.h"
50 #include "llvm/CodeGen/PBQP/Heuristics/Briggs.h"
51 #include "llvm/CodeGen/RegAllocRegistry.h"
52 #include "llvm/Support/Debug.h"
53 #include "llvm/Support/raw_ostream.h"
54 #include "llvm/Target/TargetInstrInfo.h"
55 #include "llvm/Target/TargetMachine.h"
64 static RegisterRegAlloc
65 registerPBQPRepAlloc("pbqp", "PBQP register allocator",
66 createDefaultPBQPRegisterAllocator);
69 pbqpCoalescing("pbqp-coalescing",
70 cl::desc("Attempt coalescing during PBQP register allocation."),
71 cl::init(false), cl::Hidden);
75 pbqpDumpGraphs("pbqp-dump-graphs",
76 cl::desc("Dump graphs for each function/round in the compilation unit."),
77 cl::init(false), cl::Hidden);
83 /// PBQP based allocators solve the register allocation problem by mapping
84 /// register allocation problems to Partitioned Boolean Quadratic
85 /// Programming problems.
86 class RegAllocPBQP : public MachineFunctionPass {
91 /// Construct a PBQP register allocator.
92 RegAllocPBQP(std::auto_ptr<PBQPBuilder> b, char *cPassID=0)
93 : MachineFunctionPass(ID), builder(b), customPassID(cPassID) {
94 initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
95 initializeLiveIntervalsPass(*PassRegistry::getPassRegistry());
96 initializeCalculateSpillWeightsPass(*PassRegistry::getPassRegistry());
97 initializeLiveStacksPass(*PassRegistry::getPassRegistry());
98 initializeMachineLoopInfoPass(*PassRegistry::getPassRegistry());
99 initializeVirtRegMapPass(*PassRegistry::getPassRegistry());
102 /// Return the pass name.
103 virtual const char* getPassName() const {
104 return "PBQP Register Allocator";
107 /// PBQP analysis usage.
108 virtual void getAnalysisUsage(AnalysisUsage &au) const;
110 /// Perform register allocation
111 virtual bool runOnMachineFunction(MachineFunction &MF);
115 typedef std::map<const LiveInterval*, unsigned> LI2NodeMap;
116 typedef std::vector<const LiveInterval*> Node2LIMap;
117 typedef std::vector<unsigned> AllowedSet;
118 typedef std::vector<AllowedSet> AllowedSetMap;
119 typedef std::pair<unsigned, unsigned> RegPair;
120 typedef std::map<RegPair, PBQP::PBQPNum> CoalesceMap;
121 typedef std::vector<PBQP::Graph::NodeItr> NodeVector;
122 typedef std::set<unsigned> RegSet;
125 std::auto_ptr<PBQPBuilder> builder;
130 const TargetMachine *tm;
131 const TargetRegisterInfo *tri;
132 const TargetInstrInfo *tii;
133 const MachineLoopInfo *loopInfo;
134 MachineRegisterInfo *mri;
136 std::auto_ptr<Spiller> spiller;
141 RegSet vregsToAlloc, emptyIntervalVRegs;
143 /// \brief Finds the initial set of vreg intervals to allocate.
144 void findVRegIntervalsToAlloc();
146 /// \brief Given a solved PBQP problem maps this solution back to a register
148 bool mapPBQPToRegAlloc(const PBQPRAProblem &problem,
149 const PBQP::Solution &solution);
151 /// \brief Postprocessing before final spilling. Sets basic block "live in"
153 void finalizeAlloc() const;
157 char RegAllocPBQP::ID = 0;
159 } // End anonymous namespace.
161 unsigned PBQPRAProblem::getVRegForNode(PBQP::Graph::ConstNodeItr node) const {
162 Node2VReg::const_iterator vregItr = node2VReg.find(node);
163 assert(vregItr != node2VReg.end() && "No vreg for node.");
164 return vregItr->second;
167 PBQP::Graph::NodeItr PBQPRAProblem::getNodeForVReg(unsigned vreg) const {
168 VReg2Node::const_iterator nodeItr = vreg2Node.find(vreg);
169 assert(nodeItr != vreg2Node.end() && "No node for vreg.");
170 return nodeItr->second;
174 const PBQPRAProblem::AllowedSet&
175 PBQPRAProblem::getAllowedSet(unsigned vreg) const {
176 AllowedSetMap::const_iterator allowedSetItr = allowedSets.find(vreg);
177 assert(allowedSetItr != allowedSets.end() && "No pregs for vreg.");
178 const AllowedSet &allowedSet = allowedSetItr->second;
182 unsigned PBQPRAProblem::getPRegForOption(unsigned vreg, unsigned option) const {
183 assert(isPRegOption(vreg, option) && "Not a preg option.");
185 const AllowedSet& allowedSet = getAllowedSet(vreg);
186 assert(option <= allowedSet.size() && "Option outside allowed set.");
187 return allowedSet[option - 1];
190 std::auto_ptr<PBQPRAProblem> PBQPBuilder::build(MachineFunction *mf,
191 const LiveIntervals *lis,
192 const MachineLoopInfo *loopInfo,
193 const RegSet &vregs) {
195 LiveIntervals *LIS = const_cast<LiveIntervals*>(lis);
196 MachineRegisterInfo *mri = &mf->getRegInfo();
197 const TargetRegisterInfo *tri = mf->getTarget().getRegisterInfo();
199 std::auto_ptr<PBQPRAProblem> p(new PBQPRAProblem());
200 PBQP::Graph &g = p->getGraph();
203 // Collect the set of preg intervals, record that they're used in the MF.
204 for (unsigned Reg = 1, e = tri->getNumRegs(); Reg != e; ++Reg) {
205 if (mri->def_empty(Reg))
208 mri->setPhysRegUsed(Reg);
211 // Iterate over vregs.
212 for (RegSet::const_iterator vregItr = vregs.begin(), vregEnd = vregs.end();
213 vregItr != vregEnd; ++vregItr) {
214 unsigned vreg = *vregItr;
215 const TargetRegisterClass *trc = mri->getRegClass(vreg);
216 LiveInterval *vregLI = &LIS->getInterval(vreg);
218 // Record any overlaps with regmask operands.
219 BitVector regMaskOverlaps;
220 LIS->checkRegMaskInterference(*vregLI, regMaskOverlaps);
222 // Compute an initial allowed set for the current vreg.
223 typedef std::vector<unsigned> VRAllowed;
225 ArrayRef<uint16_t> rawOrder = trc->getRawAllocationOrder(*mf);
226 for (unsigned i = 0; i != rawOrder.size(); ++i) {
227 unsigned preg = rawOrder[i];
228 if (mri->isReserved(preg))
231 // vregLI crosses a regmask operand that clobbers preg.
232 if (!regMaskOverlaps.empty() && !regMaskOverlaps.test(preg))
235 // vregLI overlaps fixed regunit interference.
236 bool Interference = false;
237 for (MCRegUnitIterator Units(preg, tri); Units.isValid(); ++Units) {
238 if (vregLI->overlaps(LIS->getRegUnit(*Units))) {
246 // preg is usable for this virtual register.
247 vrAllowed.push_back(preg);
250 // Construct the node.
251 PBQP::Graph::NodeItr node =
252 g.addNode(PBQP::Vector(vrAllowed.size() + 1, 0));
254 // Record the mapping and allowed set in the problem.
255 p->recordVReg(vreg, node, vrAllowed.begin(), vrAllowed.end());
257 PBQP::PBQPNum spillCost = (vregLI->weight != 0.0) ?
258 vregLI->weight : std::numeric_limits<PBQP::PBQPNum>::min();
260 addSpillCosts(g.getNodeCosts(node), spillCost);
263 for (RegSet::const_iterator vr1Itr = vregs.begin(), vrEnd = vregs.end();
264 vr1Itr != vrEnd; ++vr1Itr) {
265 unsigned vr1 = *vr1Itr;
266 const LiveInterval &l1 = lis->getInterval(vr1);
267 const PBQPRAProblem::AllowedSet &vr1Allowed = p->getAllowedSet(vr1);
269 for (RegSet::const_iterator vr2Itr = llvm::next(vr1Itr);
270 vr2Itr != vrEnd; ++vr2Itr) {
271 unsigned vr2 = *vr2Itr;
272 const LiveInterval &l2 = lis->getInterval(vr2);
273 const PBQPRAProblem::AllowedSet &vr2Allowed = p->getAllowedSet(vr2);
275 assert(!l2.empty() && "Empty interval in vreg set?");
276 if (l1.overlaps(l2)) {
277 PBQP::Graph::EdgeItr edge =
278 g.addEdge(p->getNodeForVReg(vr1), p->getNodeForVReg(vr2),
279 PBQP::Matrix(vr1Allowed.size()+1, vr2Allowed.size()+1, 0));
281 addInterferenceCosts(g.getEdgeCosts(edge), vr1Allowed, vr2Allowed, tri);
289 void PBQPBuilder::addSpillCosts(PBQP::Vector &costVec,
290 PBQP::PBQPNum spillCost) {
291 costVec[0] = spillCost;
294 void PBQPBuilder::addInterferenceCosts(
295 PBQP::Matrix &costMat,
296 const PBQPRAProblem::AllowedSet &vr1Allowed,
297 const PBQPRAProblem::AllowedSet &vr2Allowed,
298 const TargetRegisterInfo *tri) {
299 assert(costMat.getRows() == vr1Allowed.size() + 1 && "Matrix height mismatch.");
300 assert(costMat.getCols() == vr2Allowed.size() + 1 && "Matrix width mismatch.");
302 for (unsigned i = 0; i != vr1Allowed.size(); ++i) {
303 unsigned preg1 = vr1Allowed[i];
305 for (unsigned j = 0; j != vr2Allowed.size(); ++j) {
306 unsigned preg2 = vr2Allowed[j];
308 if (tri->regsOverlap(preg1, preg2)) {
309 costMat[i + 1][j + 1] = std::numeric_limits<PBQP::PBQPNum>::infinity();
315 std::auto_ptr<PBQPRAProblem> PBQPBuilderWithCoalescing::build(
317 const LiveIntervals *lis,
318 const MachineLoopInfo *loopInfo,
319 const RegSet &vregs) {
321 std::auto_ptr<PBQPRAProblem> p = PBQPBuilder::build(mf, lis, loopInfo, vregs);
322 PBQP::Graph &g = p->getGraph();
324 const TargetMachine &tm = mf->getTarget();
325 CoalescerPair cp(*tm.getRegisterInfo());
327 // Scan the machine function and add a coalescing cost whenever CoalescerPair
329 for (MachineFunction::const_iterator mbbItr = mf->begin(),
331 mbbItr != mbbEnd; ++mbbItr) {
332 const MachineBasicBlock *mbb = &*mbbItr;
334 for (MachineBasicBlock::const_iterator miItr = mbb->begin(),
336 miItr != miEnd; ++miItr) {
337 const MachineInstr *mi = &*miItr;
339 if (!cp.setRegisters(mi)) {
340 continue; // Not coalescable.
343 if (cp.getSrcReg() == cp.getDstReg()) {
344 continue; // Already coalesced.
347 unsigned dst = cp.getDstReg(),
348 src = cp.getSrcReg();
350 const float copyFactor = 0.5; // Cost of copy relative to load. Current
351 // value plucked randomly out of the air.
353 PBQP::PBQPNum cBenefit =
354 copyFactor * LiveIntervals::getSpillWeight(false, true,
355 loopInfo->getLoopDepth(mbb));
358 if (!mf->getRegInfo().isAllocatable(dst)) {
362 const PBQPRAProblem::AllowedSet &allowed = p->getAllowedSet(src);
363 unsigned pregOpt = 0;
364 while (pregOpt < allowed.size() && allowed[pregOpt] != dst) {
367 if (pregOpt < allowed.size()) {
368 ++pregOpt; // +1 to account for spill option.
369 PBQP::Graph::NodeItr node = p->getNodeForVReg(src);
370 addPhysRegCoalesce(g.getNodeCosts(node), pregOpt, cBenefit);
373 const PBQPRAProblem::AllowedSet *allowed1 = &p->getAllowedSet(dst);
374 const PBQPRAProblem::AllowedSet *allowed2 = &p->getAllowedSet(src);
375 PBQP::Graph::NodeItr node1 = p->getNodeForVReg(dst);
376 PBQP::Graph::NodeItr node2 = p->getNodeForVReg(src);
377 PBQP::Graph::EdgeItr edge = g.findEdge(node1, node2);
378 if (edge == g.edgesEnd()) {
379 edge = g.addEdge(node1, node2, PBQP::Matrix(allowed1->size() + 1,
380 allowed2->size() + 1,
383 if (g.getEdgeNode1(edge) == node2) {
384 std::swap(node1, node2);
385 std::swap(allowed1, allowed2);
389 addVirtRegCoalesce(g.getEdgeCosts(edge), *allowed1, *allowed2,
398 void PBQPBuilderWithCoalescing::addPhysRegCoalesce(PBQP::Vector &costVec,
400 PBQP::PBQPNum benefit) {
401 costVec[pregOption] += -benefit;
404 void PBQPBuilderWithCoalescing::addVirtRegCoalesce(
405 PBQP::Matrix &costMat,
406 const PBQPRAProblem::AllowedSet &vr1Allowed,
407 const PBQPRAProblem::AllowedSet &vr2Allowed,
408 PBQP::PBQPNum benefit) {
410 assert(costMat.getRows() == vr1Allowed.size() + 1 && "Size mismatch.");
411 assert(costMat.getCols() == vr2Allowed.size() + 1 && "Size mismatch.");
413 for (unsigned i = 0; i != vr1Allowed.size(); ++i) {
414 unsigned preg1 = vr1Allowed[i];
415 for (unsigned j = 0; j != vr2Allowed.size(); ++j) {
416 unsigned preg2 = vr2Allowed[j];
418 if (preg1 == preg2) {
419 costMat[i + 1][j + 1] += -benefit;
426 void RegAllocPBQP::getAnalysisUsage(AnalysisUsage &au) const {
427 au.setPreservesCFG();
428 au.addRequired<AliasAnalysis>();
429 au.addPreserved<AliasAnalysis>();
430 au.addRequired<SlotIndexes>();
431 au.addPreserved<SlotIndexes>();
432 au.addRequired<LiveIntervals>();
433 au.addPreserved<LiveIntervals>();
434 //au.addRequiredID(SplitCriticalEdgesID);
436 au.addRequiredID(*customPassID);
437 au.addRequired<CalculateSpillWeights>();
438 au.addRequired<LiveStacks>();
439 au.addPreserved<LiveStacks>();
440 au.addRequired<MachineDominatorTree>();
441 au.addPreserved<MachineDominatorTree>();
442 au.addRequired<MachineLoopInfo>();
443 au.addPreserved<MachineLoopInfo>();
444 au.addRequired<VirtRegMap>();
445 au.addPreserved<VirtRegMap>();
446 MachineFunctionPass::getAnalysisUsage(au);
449 void RegAllocPBQP::findVRegIntervalsToAlloc() {
451 // Iterate over all live ranges.
452 for (unsigned i = 0, e = mri->getNumVirtRegs(); i != e; ++i) {
453 unsigned Reg = TargetRegisterInfo::index2VirtReg(i);
454 if (mri->reg_nodbg_empty(Reg))
456 LiveInterval *li = &lis->getInterval(Reg);
458 // If this live interval is non-empty we will use pbqp to allocate it.
459 // Empty intervals we allocate in a simple post-processing stage in
462 vregsToAlloc.insert(li->reg);
464 emptyIntervalVRegs.insert(li->reg);
469 bool RegAllocPBQP::mapPBQPToRegAlloc(const PBQPRAProblem &problem,
470 const PBQP::Solution &solution) {
471 // Set to true if we have any spills
472 bool anotherRoundNeeded = false;
474 // Clear the existing allocation.
477 const PBQP::Graph &g = problem.getGraph();
478 // Iterate over the nodes mapping the PBQP solution to a register
480 for (PBQP::Graph::ConstNodeItr node = g.nodesBegin(),
481 nodeEnd = g.nodesEnd();
482 node != nodeEnd; ++node) {
483 unsigned vreg = problem.getVRegForNode(node);
484 unsigned alloc = solution.getSelection(node);
486 if (problem.isPRegOption(vreg, alloc)) {
487 unsigned preg = problem.getPRegForOption(vreg, alloc);
488 DEBUG(dbgs() << "VREG " << PrintReg(vreg, tri) << " -> "
489 << tri->getName(preg) << "\n");
490 assert(preg != 0 && "Invalid preg selected.");
491 vrm->assignVirt2Phys(vreg, preg);
492 } else if (problem.isSpillOption(vreg, alloc)) {
493 vregsToAlloc.erase(vreg);
494 SmallVector<LiveInterval*, 8> newSpills;
495 LiveRangeEdit LRE(&lis->getInterval(vreg), newSpills, *mf, *lis, vrm);
498 DEBUG(dbgs() << "VREG " << PrintReg(vreg, tri) << " -> SPILLED (Cost: "
499 << LRE.getParent().weight << ", New vregs: ");
501 // Copy any newly inserted live intervals into the list of regs to
503 for (LiveRangeEdit::iterator itr = LRE.begin(), end = LRE.end();
505 assert(!(*itr)->empty() && "Empty spill range.");
506 DEBUG(dbgs() << PrintReg((*itr)->reg, tri) << " ");
507 vregsToAlloc.insert((*itr)->reg);
510 DEBUG(dbgs() << ")\n");
512 // We need another round if spill intervals were added.
513 anotherRoundNeeded |= !LRE.empty();
515 llvm_unreachable("Unknown allocation option.");
519 return !anotherRoundNeeded;
523 void RegAllocPBQP::finalizeAlloc() const {
524 // First allocate registers for the empty intervals.
525 for (RegSet::const_iterator
526 itr = emptyIntervalVRegs.begin(), end = emptyIntervalVRegs.end();
528 LiveInterval *li = &lis->getInterval(*itr);
530 unsigned physReg = vrm->getRegAllocPref(li->reg);
533 const TargetRegisterClass *liRC = mri->getRegClass(li->reg);
534 physReg = liRC->getRawAllocationOrder(*mf).front();
537 vrm->assignVirt2Phys(li->reg, physReg);
541 bool RegAllocPBQP::runOnMachineFunction(MachineFunction &MF) {
544 tm = &mf->getTarget();
545 tri = tm->getRegisterInfo();
546 tii = tm->getInstrInfo();
547 mri = &mf->getRegInfo();
549 lis = &getAnalysis<LiveIntervals>();
550 lss = &getAnalysis<LiveStacks>();
551 loopInfo = &getAnalysis<MachineLoopInfo>();
553 vrm = &getAnalysis<VirtRegMap>();
554 spiller.reset(createInlineSpiller(*this, MF, *vrm));
556 mri->freezeReservedRegs(MF);
558 DEBUG(dbgs() << "PBQP Register Allocating for " << mf->getName() << "\n");
560 // Allocator main loop:
562 // * Map current regalloc problem to a PBQP problem
563 // * Solve the PBQP problem
564 // * Map the solution back to a register allocation
565 // * Spill if necessary
567 // This process is continued till no more spills are generated.
569 // Find the vreg intervals in need of allocation.
570 findVRegIntervalsToAlloc();
573 const Function* func = mf->getFunction();
575 func->getParent()->getModuleIdentifier() + "." +
576 func->getName().str();
579 // If there are non-empty intervals allocate them using pbqp.
580 if (!vregsToAlloc.empty()) {
582 bool pbqpAllocComplete = false;
585 while (!pbqpAllocComplete) {
586 DEBUG(dbgs() << " PBQP Regalloc round " << round << ":\n");
588 std::auto_ptr<PBQPRAProblem> problem =
589 builder->build(mf, lis, loopInfo, vregsToAlloc);
592 if (pbqpDumpGraphs) {
593 std::ostringstream rs;
595 std::string graphFileName(fqn + "." + rs.str() + ".pbqpgraph");
597 raw_fd_ostream os(graphFileName.c_str(), tmp);
598 DEBUG(dbgs() << "Dumping graph for round " << round << " to \""
599 << graphFileName << "\"\n");
600 problem->getGraph().dump(os);
604 PBQP::Solution solution =
605 PBQP::HeuristicSolver<PBQP::Heuristics::Briggs>::solve(
606 problem->getGraph());
608 pbqpAllocComplete = mapPBQPToRegAlloc(*problem, solution);
614 // Finalise allocation, allocate empty ranges.
616 vregsToAlloc.clear();
617 emptyIntervalVRegs.clear();
619 DEBUG(dbgs() << "Post alloc VirtRegMap:\n" << *vrm << "\n");
624 FunctionPass* llvm::createPBQPRegisterAllocator(
625 std::auto_ptr<PBQPBuilder> builder,
626 char *customPassID) {
627 return new RegAllocPBQP(builder, customPassID);
630 FunctionPass* llvm::createDefaultPBQPRegisterAllocator() {
631 if (pbqpCoalescing) {
632 return createPBQPRegisterAllocator(
633 std::auto_ptr<PBQPBuilder>(new PBQPBuilderWithCoalescing()));
635 return createPBQPRegisterAllocator(
636 std::auto_ptr<PBQPBuilder>(new PBQPBuilder()));