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::set<unsigned> RegSet;
124 std::auto_ptr<PBQPBuilder> builder;
129 const TargetMachine *tm;
130 const TargetRegisterInfo *tri;
131 const TargetInstrInfo *tii;
132 const MachineLoopInfo *loopInfo;
133 MachineRegisterInfo *mri;
135 std::auto_ptr<Spiller> spiller;
140 RegSet vregsToAlloc, emptyIntervalVRegs;
142 /// \brief Finds the initial set of vreg intervals to allocate.
143 void findVRegIntervalsToAlloc();
145 /// \brief Given a solved PBQP problem maps this solution back to a register
147 bool mapPBQPToRegAlloc(const PBQPRAProblem &problem,
148 const PBQP::Solution &solution);
150 /// \brief Postprocessing before final spilling. Sets basic block "live in"
152 void finalizeAlloc() const;
156 char RegAllocPBQP::ID = 0;
158 } // End anonymous namespace.
160 unsigned PBQPRAProblem::getVRegForNode(PBQP::Graph::ConstNodeItr node) const {
161 Node2VReg::const_iterator vregItr = node2VReg.find(node);
162 assert(vregItr != node2VReg.end() && "No vreg for node.");
163 return vregItr->second;
166 PBQP::Graph::NodeItr PBQPRAProblem::getNodeForVReg(unsigned vreg) const {
167 VReg2Node::const_iterator nodeItr = vreg2Node.find(vreg);
168 assert(nodeItr != vreg2Node.end() && "No node for vreg.");
169 return nodeItr->second;
173 const PBQPRAProblem::AllowedSet&
174 PBQPRAProblem::getAllowedSet(unsigned vreg) const {
175 AllowedSetMap::const_iterator allowedSetItr = allowedSets.find(vreg);
176 assert(allowedSetItr != allowedSets.end() && "No pregs for vreg.");
177 const AllowedSet &allowedSet = allowedSetItr->second;
181 unsigned PBQPRAProblem::getPRegForOption(unsigned vreg, unsigned option) const {
182 assert(isPRegOption(vreg, option) && "Not a preg option.");
184 const AllowedSet& allowedSet = getAllowedSet(vreg);
185 assert(option <= allowedSet.size() && "Option outside allowed set.");
186 return allowedSet[option - 1];
189 std::auto_ptr<PBQPRAProblem> PBQPBuilder::build(MachineFunction *mf,
190 const LiveIntervals *lis,
191 const MachineLoopInfo *loopInfo,
192 const RegSet &vregs) {
194 LiveIntervals *LIS = const_cast<LiveIntervals*>(lis);
195 MachineRegisterInfo *mri = &mf->getRegInfo();
196 const TargetRegisterInfo *tri = mf->getTarget().getRegisterInfo();
198 std::auto_ptr<PBQPRAProblem> p(new PBQPRAProblem());
199 PBQP::Graph &g = p->getGraph();
202 // Collect the set of preg intervals, record that they're used in the MF.
203 for (unsigned Reg = 1, e = tri->getNumRegs(); Reg != e; ++Reg) {
204 if (mri->def_empty(Reg))
207 mri->setPhysRegUsed(Reg);
210 // Iterate over vregs.
211 for (RegSet::const_iterator vregItr = vregs.begin(), vregEnd = vregs.end();
212 vregItr != vregEnd; ++vregItr) {
213 unsigned vreg = *vregItr;
214 const TargetRegisterClass *trc = mri->getRegClass(vreg);
215 LiveInterval *vregLI = &LIS->getInterval(vreg);
217 // Record any overlaps with regmask operands.
218 BitVector regMaskOverlaps;
219 LIS->checkRegMaskInterference(*vregLI, regMaskOverlaps);
221 // Compute an initial allowed set for the current vreg.
222 typedef std::vector<unsigned> VRAllowed;
224 ArrayRef<uint16_t> rawOrder = trc->getRawAllocationOrder(*mf);
225 for (unsigned i = 0; i != rawOrder.size(); ++i) {
226 unsigned preg = rawOrder[i];
227 if (mri->isReserved(preg))
230 // vregLI crosses a regmask operand that clobbers preg.
231 if (!regMaskOverlaps.empty() && !regMaskOverlaps.test(preg))
234 // vregLI overlaps fixed regunit interference.
235 bool Interference = false;
236 for (MCRegUnitIterator Units(preg, tri); Units.isValid(); ++Units) {
237 if (vregLI->overlaps(LIS->getRegUnit(*Units))) {
245 // preg is usable for this virtual register.
246 vrAllowed.push_back(preg);
249 // Construct the node.
250 PBQP::Graph::NodeItr node =
251 g.addNode(PBQP::Vector(vrAllowed.size() + 1, 0));
253 // Record the mapping and allowed set in the problem.
254 p->recordVReg(vreg, node, vrAllowed.begin(), vrAllowed.end());
256 PBQP::PBQPNum spillCost = (vregLI->weight != 0.0) ?
257 vregLI->weight : std::numeric_limits<PBQP::PBQPNum>::min();
259 addSpillCosts(g.getNodeCosts(node), spillCost);
262 for (RegSet::const_iterator vr1Itr = vregs.begin(), vrEnd = vregs.end();
263 vr1Itr != vrEnd; ++vr1Itr) {
264 unsigned vr1 = *vr1Itr;
265 const LiveInterval &l1 = lis->getInterval(vr1);
266 const PBQPRAProblem::AllowedSet &vr1Allowed = p->getAllowedSet(vr1);
268 for (RegSet::const_iterator vr2Itr = llvm::next(vr1Itr);
269 vr2Itr != vrEnd; ++vr2Itr) {
270 unsigned vr2 = *vr2Itr;
271 const LiveInterval &l2 = lis->getInterval(vr2);
272 const PBQPRAProblem::AllowedSet &vr2Allowed = p->getAllowedSet(vr2);
274 assert(!l2.empty() && "Empty interval in vreg set?");
275 if (l1.overlaps(l2)) {
276 PBQP::Graph::EdgeItr edge =
277 g.addEdge(p->getNodeForVReg(vr1), p->getNodeForVReg(vr2),
278 PBQP::Matrix(vr1Allowed.size()+1, vr2Allowed.size()+1, 0));
280 addInterferenceCosts(g.getEdgeCosts(edge), vr1Allowed, vr2Allowed, tri);
288 void PBQPBuilder::addSpillCosts(PBQP::Vector &costVec,
289 PBQP::PBQPNum spillCost) {
290 costVec[0] = spillCost;
293 void PBQPBuilder::addInterferenceCosts(
294 PBQP::Matrix &costMat,
295 const PBQPRAProblem::AllowedSet &vr1Allowed,
296 const PBQPRAProblem::AllowedSet &vr2Allowed,
297 const TargetRegisterInfo *tri) {
298 assert(costMat.getRows() == vr1Allowed.size() + 1 && "Matrix height mismatch.");
299 assert(costMat.getCols() == vr2Allowed.size() + 1 && "Matrix width mismatch.");
301 for (unsigned i = 0; i != vr1Allowed.size(); ++i) {
302 unsigned preg1 = vr1Allowed[i];
304 for (unsigned j = 0; j != vr2Allowed.size(); ++j) {
305 unsigned preg2 = vr2Allowed[j];
307 if (tri->regsOverlap(preg1, preg2)) {
308 costMat[i + 1][j + 1] = std::numeric_limits<PBQP::PBQPNum>::infinity();
314 std::auto_ptr<PBQPRAProblem> PBQPBuilderWithCoalescing::build(
316 const LiveIntervals *lis,
317 const MachineLoopInfo *loopInfo,
318 const RegSet &vregs) {
320 std::auto_ptr<PBQPRAProblem> p = PBQPBuilder::build(mf, lis, loopInfo, vregs);
321 PBQP::Graph &g = p->getGraph();
323 const TargetMachine &tm = mf->getTarget();
324 CoalescerPair cp(*tm.getRegisterInfo());
326 // Scan the machine function and add a coalescing cost whenever CoalescerPair
328 for (MachineFunction::const_iterator mbbItr = mf->begin(),
330 mbbItr != mbbEnd; ++mbbItr) {
331 const MachineBasicBlock *mbb = &*mbbItr;
333 for (MachineBasicBlock::const_iterator miItr = mbb->begin(),
335 miItr != miEnd; ++miItr) {
336 const MachineInstr *mi = &*miItr;
338 if (!cp.setRegisters(mi)) {
339 continue; // Not coalescable.
342 if (cp.getSrcReg() == cp.getDstReg()) {
343 continue; // Already coalesced.
346 unsigned dst = cp.getDstReg(),
347 src = cp.getSrcReg();
349 const float copyFactor = 0.5; // Cost of copy relative to load. Current
350 // value plucked randomly out of the air.
352 PBQP::PBQPNum cBenefit =
353 copyFactor * LiveIntervals::getSpillWeight(false, true,
354 loopInfo->getLoopDepth(mbb));
357 if (!mf->getRegInfo().isAllocatable(dst)) {
361 const PBQPRAProblem::AllowedSet &allowed = p->getAllowedSet(src);
362 unsigned pregOpt = 0;
363 while (pregOpt < allowed.size() && allowed[pregOpt] != dst) {
366 if (pregOpt < allowed.size()) {
367 ++pregOpt; // +1 to account for spill option.
368 PBQP::Graph::NodeItr node = p->getNodeForVReg(src);
369 addPhysRegCoalesce(g.getNodeCosts(node), pregOpt, cBenefit);
372 const PBQPRAProblem::AllowedSet *allowed1 = &p->getAllowedSet(dst);
373 const PBQPRAProblem::AllowedSet *allowed2 = &p->getAllowedSet(src);
374 PBQP::Graph::NodeItr node1 = p->getNodeForVReg(dst);
375 PBQP::Graph::NodeItr node2 = p->getNodeForVReg(src);
376 PBQP::Graph::EdgeItr edge = g.findEdge(node1, node2);
377 if (edge == g.edgesEnd()) {
378 edge = g.addEdge(node1, node2, PBQP::Matrix(allowed1->size() + 1,
379 allowed2->size() + 1,
382 if (g.getEdgeNode1(edge) == node2) {
383 std::swap(node1, node2);
384 std::swap(allowed1, allowed2);
388 addVirtRegCoalesce(g.getEdgeCosts(edge), *allowed1, *allowed2,
397 void PBQPBuilderWithCoalescing::addPhysRegCoalesce(PBQP::Vector &costVec,
399 PBQP::PBQPNum benefit) {
400 costVec[pregOption] += -benefit;
403 void PBQPBuilderWithCoalescing::addVirtRegCoalesce(
404 PBQP::Matrix &costMat,
405 const PBQPRAProblem::AllowedSet &vr1Allowed,
406 const PBQPRAProblem::AllowedSet &vr2Allowed,
407 PBQP::PBQPNum benefit) {
409 assert(costMat.getRows() == vr1Allowed.size() + 1 && "Size mismatch.");
410 assert(costMat.getCols() == vr2Allowed.size() + 1 && "Size mismatch.");
412 for (unsigned i = 0; i != vr1Allowed.size(); ++i) {
413 unsigned preg1 = vr1Allowed[i];
414 for (unsigned j = 0; j != vr2Allowed.size(); ++j) {
415 unsigned preg2 = vr2Allowed[j];
417 if (preg1 == preg2) {
418 costMat[i + 1][j + 1] += -benefit;
425 void RegAllocPBQP::getAnalysisUsage(AnalysisUsage &au) const {
426 au.setPreservesCFG();
427 au.addRequired<AliasAnalysis>();
428 au.addPreserved<AliasAnalysis>();
429 au.addRequired<SlotIndexes>();
430 au.addPreserved<SlotIndexes>();
431 au.addRequired<LiveIntervals>();
432 au.addPreserved<LiveIntervals>();
433 //au.addRequiredID(SplitCriticalEdgesID);
435 au.addRequiredID(*customPassID);
436 au.addRequired<CalculateSpillWeights>();
437 au.addRequired<LiveStacks>();
438 au.addPreserved<LiveStacks>();
439 au.addRequired<MachineDominatorTree>();
440 au.addPreserved<MachineDominatorTree>();
441 au.addRequired<MachineLoopInfo>();
442 au.addPreserved<MachineLoopInfo>();
443 au.addRequired<VirtRegMap>();
444 au.addPreserved<VirtRegMap>();
445 MachineFunctionPass::getAnalysisUsage(au);
448 void RegAllocPBQP::findVRegIntervalsToAlloc() {
450 // Iterate over all live ranges.
451 for (unsigned i = 0, e = mri->getNumVirtRegs(); i != e; ++i) {
452 unsigned Reg = TargetRegisterInfo::index2VirtReg(i);
453 if (mri->reg_nodbg_empty(Reg))
455 LiveInterval *li = &lis->getInterval(Reg);
457 // If this live interval is non-empty we will use pbqp to allocate it.
458 // Empty intervals we allocate in a simple post-processing stage in
461 vregsToAlloc.insert(li->reg);
463 emptyIntervalVRegs.insert(li->reg);
468 bool RegAllocPBQP::mapPBQPToRegAlloc(const PBQPRAProblem &problem,
469 const PBQP::Solution &solution) {
470 // Set to true if we have any spills
471 bool anotherRoundNeeded = false;
473 // Clear the existing allocation.
476 const PBQP::Graph &g = problem.getGraph();
477 // Iterate over the nodes mapping the PBQP solution to a register
479 for (PBQP::Graph::ConstNodeItr node = g.nodesBegin(),
480 nodeEnd = g.nodesEnd();
481 node != nodeEnd; ++node) {
482 unsigned vreg = problem.getVRegForNode(node);
483 unsigned alloc = solution.getSelection(node);
485 if (problem.isPRegOption(vreg, alloc)) {
486 unsigned preg = problem.getPRegForOption(vreg, alloc);
487 DEBUG(dbgs() << "VREG " << PrintReg(vreg, tri) << " -> "
488 << tri->getName(preg) << "\n");
489 assert(preg != 0 && "Invalid preg selected.");
490 vrm->assignVirt2Phys(vreg, preg);
491 } else if (problem.isSpillOption(vreg, alloc)) {
492 vregsToAlloc.erase(vreg);
493 SmallVector<LiveInterval*, 8> newSpills;
494 LiveRangeEdit LRE(&lis->getInterval(vreg), newSpills, *mf, *lis, vrm);
497 DEBUG(dbgs() << "VREG " << PrintReg(vreg, tri) << " -> SPILLED (Cost: "
498 << LRE.getParent().weight << ", New vregs: ");
500 // Copy any newly inserted live intervals into the list of regs to
502 for (LiveRangeEdit::iterator itr = LRE.begin(), end = LRE.end();
504 assert(!(*itr)->empty() && "Empty spill range.");
505 DEBUG(dbgs() << PrintReg((*itr)->reg, tri) << " ");
506 vregsToAlloc.insert((*itr)->reg);
509 DEBUG(dbgs() << ")\n");
511 // We need another round if spill intervals were added.
512 anotherRoundNeeded |= !LRE.empty();
514 llvm_unreachable("Unknown allocation option.");
518 return !anotherRoundNeeded;
522 void RegAllocPBQP::finalizeAlloc() const {
523 // First allocate registers for the empty intervals.
524 for (RegSet::const_iterator
525 itr = emptyIntervalVRegs.begin(), end = emptyIntervalVRegs.end();
527 LiveInterval *li = &lis->getInterval(*itr);
529 unsigned physReg = vrm->getRegAllocPref(li->reg);
532 const TargetRegisterClass *liRC = mri->getRegClass(li->reg);
533 physReg = liRC->getRawAllocationOrder(*mf).front();
536 vrm->assignVirt2Phys(li->reg, physReg);
540 bool RegAllocPBQP::runOnMachineFunction(MachineFunction &MF) {
543 tm = &mf->getTarget();
544 tri = tm->getRegisterInfo();
545 tii = tm->getInstrInfo();
546 mri = &mf->getRegInfo();
548 lis = &getAnalysis<LiveIntervals>();
549 lss = &getAnalysis<LiveStacks>();
550 loopInfo = &getAnalysis<MachineLoopInfo>();
552 vrm = &getAnalysis<VirtRegMap>();
553 spiller.reset(createInlineSpiller(*this, MF, *vrm));
555 mri->freezeReservedRegs(MF);
557 DEBUG(dbgs() << "PBQP Register Allocating for " << mf->getName() << "\n");
559 // Allocator main loop:
561 // * Map current regalloc problem to a PBQP problem
562 // * Solve the PBQP problem
563 // * Map the solution back to a register allocation
564 // * Spill if necessary
566 // This process is continued till no more spills are generated.
568 // Find the vreg intervals in need of allocation.
569 findVRegIntervalsToAlloc();
572 const Function* func = mf->getFunction();
574 func->getParent()->getModuleIdentifier() + "." +
575 func->getName().str();
578 // If there are non-empty intervals allocate them using pbqp.
579 if (!vregsToAlloc.empty()) {
581 bool pbqpAllocComplete = false;
584 while (!pbqpAllocComplete) {
585 DEBUG(dbgs() << " PBQP Regalloc round " << round << ":\n");
587 std::auto_ptr<PBQPRAProblem> problem =
588 builder->build(mf, lis, loopInfo, vregsToAlloc);
591 if (pbqpDumpGraphs) {
592 std::ostringstream rs;
594 std::string graphFileName(fqn + "." + rs.str() + ".pbqpgraph");
596 raw_fd_ostream os(graphFileName.c_str(), tmp);
597 DEBUG(dbgs() << "Dumping graph for round " << round << " to \""
598 << graphFileName << "\"\n");
599 problem->getGraph().dump(os);
603 PBQP::Solution solution =
604 PBQP::HeuristicSolver<PBQP::Heuristics::Briggs>::solve(
605 problem->getGraph());
607 pbqpAllocComplete = mapPBQPToRegAlloc(*problem, solution);
613 // Finalise allocation, allocate empty ranges.
615 vregsToAlloc.clear();
616 emptyIntervalVRegs.clear();
618 DEBUG(dbgs() << "Post alloc VirtRegMap:\n" << *vrm << "\n");
623 FunctionPass* llvm::createPBQPRegisterAllocator(
624 std::auto_ptr<PBQPBuilder> builder,
625 char *customPassID) {
626 return new RegAllocPBQP(builder, customPassID);
629 FunctionPass* llvm::createDefaultPBQPRegisterAllocator() {
630 if (pbqpCoalescing) {
631 return createPBQPRegisterAllocator(
632 std::auto_ptr<PBQPBuilder>(new PBQPBuilderWithCoalescing()));
634 return createPBQPRegisterAllocator(
635 std::auto_ptr<PBQPBuilder>(new PBQPBuilder()));