blob: dde6fba5294e1b06ed908e11dbf006a4792cc49d [file] [log] [blame]
Evan Chengb1290a62008-10-02 18:29:27 +00001//===------ 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//===----------------------------------------------------------------------===//
Misha Brukman2a835f92009-01-08 15:50:22 +00009//
Evan Chengb1290a62008-10-02 18:29:27 +000010// 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
Misha Brukman2a835f92009-01-08 15:50:22 +000015// code is inserted and the process repeated.
Evan Chengb1290a62008-10-02 18:29:27 +000016//
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
Misha Brukmance07e992009-01-08 16:40:25 +000019// allocation, see the following papers:
Evan Chengb1290a62008-10-02 18:29:27 +000020//
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.
Misha Brukman2a835f92009-01-08 15:50:22 +000029//
Evan Chengb1290a62008-10-02 18:29:27 +000030//===----------------------------------------------------------------------===//
31
Evan Chengb1290a62008-10-02 18:29:27 +000032#define DEBUG_TYPE "regalloc"
33
Lang Hames54cc2ef2010-07-19 15:22:28 +000034#include "RenderMachineFunction.h"
Jakob Stoklund Olesencfa81012011-11-12 23:17:52 +000035#include "Spiller.h"
Evan Chengb1290a62008-10-02 18:29:27 +000036#include "VirtRegMap.h"
Rafael Espindolafdf16ca2011-06-26 21:41:06 +000037#include "RegisterCoalescer.h"
Lang Hames20df03c2012-03-26 23:07:23 +000038#include "llvm/Module.h"
Lang Hames9ad7e072011-12-06 01:45:57 +000039#include "llvm/Analysis/AliasAnalysis.h"
Lang Hamesa937f222009-12-14 06:49:42 +000040#include "llvm/CodeGen/CalcSpillWeights.h"
Evan Chengb1290a62008-10-02 18:29:27 +000041#include "llvm/CodeGen/LiveIntervalAnalysis.h"
Pete Cooper789d5d82012-04-02 22:44:18 +000042#include "llvm/CodeGen/LiveRangeEdit.h"
Lang Hames27601ef2008-11-16 12:12:54 +000043#include "llvm/CodeGen/LiveStackAnalysis.h"
Lang Hameseb6c8f52010-09-18 09:07:10 +000044#include "llvm/CodeGen/RegAllocPBQP.h"
Lang Hames9ad7e072011-12-06 01:45:57 +000045#include "llvm/CodeGen/MachineDominators.h"
Misha Brukman2a835f92009-01-08 15:50:22 +000046#include "llvm/CodeGen/MachineFunctionPass.h"
Evan Chengb1290a62008-10-02 18:29:27 +000047#include "llvm/CodeGen/MachineLoopInfo.h"
Misha Brukman2a835f92009-01-08 15:50:22 +000048#include "llvm/CodeGen/MachineRegisterInfo.h"
Lang Hameseb6c8f52010-09-18 09:07:10 +000049#include "llvm/CodeGen/PBQP/HeuristicSolver.h"
50#include "llvm/CodeGen/PBQP/Graph.h"
51#include "llvm/CodeGen/PBQP/Heuristics/Briggs.h"
Misha Brukman2a835f92009-01-08 15:50:22 +000052#include "llvm/CodeGen/RegAllocRegistry.h"
Evan Chengb1290a62008-10-02 18:29:27 +000053#include "llvm/Support/Debug.h"
Daniel Dunbarce63ffb2009-07-25 00:23:56 +000054#include "llvm/Support/raw_ostream.h"
Misha Brukman2a835f92009-01-08 15:50:22 +000055#include "llvm/Target/TargetInstrInfo.h"
56#include "llvm/Target/TargetMachine.h"
57#include <limits>
Misha Brukman2a835f92009-01-08 15:50:22 +000058#include <memory>
Evan Chengb1290a62008-10-02 18:29:27 +000059#include <set>
Lang Hames20df03c2012-03-26 23:07:23 +000060#include <sstream>
Evan Chengb1290a62008-10-02 18:29:27 +000061#include <vector>
Evan Chengb1290a62008-10-02 18:29:27 +000062
Lang Hamesf70e7cc2010-09-23 04:28:54 +000063using namespace llvm;
Lang Hameseb6c8f52010-09-18 09:07:10 +000064
Evan Chengb1290a62008-10-02 18:29:27 +000065static RegisterRegAlloc
Duncan Sands1aecd152010-02-18 14:10:41 +000066registerPBQPRepAlloc("pbqp", "PBQP register allocator",
Lang Hamesf70e7cc2010-09-23 04:28:54 +000067 createDefaultPBQPRegisterAllocator);
Evan Chengb1290a62008-10-02 18:29:27 +000068
Lang Hames8481e3b2009-08-19 01:36:14 +000069static cl::opt<bool>
70pbqpCoalescing("pbqp-coalescing",
Lang Hames030c4bf2010-01-26 04:49:58 +000071 cl::desc("Attempt coalescing during PBQP register allocation."),
72 cl::init(false), cl::Hidden);
Lang Hames8481e3b2009-08-19 01:36:14 +000073
Lang Hames20df03c2012-03-26 23:07:23 +000074#ifndef NDEBUG
75static cl::opt<bool>
76pbqpDumpGraphs("pbqp-dump-graphs",
77 cl::desc("Dump graphs for each function/round in the compilation unit."),
78 cl::init(false), cl::Hidden);
79#endif
80
Lang Hamesf70e7cc2010-09-23 04:28:54 +000081namespace {
82
83///
84/// PBQP based allocators solve the register allocation problem by mapping
85/// register allocation problems to Partitioned Boolean Quadratic
86/// Programming problems.
87class RegAllocPBQP : public MachineFunctionPass {
88public:
89
90 static char ID;
91
92 /// Construct a PBQP register allocator.
Lang Hames8d857662011-06-17 07:09:01 +000093 RegAllocPBQP(std::auto_ptr<PBQPBuilder> b, char *cPassID=0)
94 : MachineFunctionPass(ID), builder(b), customPassID(cPassID) {
Owen Anderson081c34b2010-10-19 17:21:58 +000095 initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
96 initializeLiveIntervalsPass(*PassRegistry::getPassRegistry());
Owen Anderson081c34b2010-10-19 17:21:58 +000097 initializeCalculateSpillWeightsPass(*PassRegistry::getPassRegistry());
98 initializeLiveStacksPass(*PassRegistry::getPassRegistry());
99 initializeMachineLoopInfoPass(*PassRegistry::getPassRegistry());
Owen Anderson081c34b2010-10-19 17:21:58 +0000100 initializeVirtRegMapPass(*PassRegistry::getPassRegistry());
101 initializeRenderMachineFunctionPass(*PassRegistry::getPassRegistry());
102 }
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000103
104 /// Return the pass name.
105 virtual const char* getPassName() const {
106 return "PBQP Register Allocator";
107 }
108
109 /// PBQP analysis usage.
110 virtual void getAnalysisUsage(AnalysisUsage &au) const;
111
112 /// Perform register allocation
113 virtual bool runOnMachineFunction(MachineFunction &MF);
114
115private:
116
117 typedef std::map<const LiveInterval*, unsigned> LI2NodeMap;
118 typedef std::vector<const LiveInterval*> Node2LIMap;
119 typedef std::vector<unsigned> AllowedSet;
120 typedef std::vector<AllowedSet> AllowedSetMap;
121 typedef std::pair<unsigned, unsigned> RegPair;
122 typedef std::map<RegPair, PBQP::PBQPNum> CoalesceMap;
123 typedef std::vector<PBQP::Graph::NodeItr> NodeVector;
124 typedef std::set<unsigned> RegSet;
125
126
127 std::auto_ptr<PBQPBuilder> builder;
128
Lang Hames8d857662011-06-17 07:09:01 +0000129 char *customPassID;
130
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000131 MachineFunction *mf;
132 const TargetMachine *tm;
133 const TargetRegisterInfo *tri;
134 const TargetInstrInfo *tii;
135 const MachineLoopInfo *loopInfo;
136 MachineRegisterInfo *mri;
137 RenderMachineFunction *rmf;
138
Jakob Stoklund Olesencfa81012011-11-12 23:17:52 +0000139 std::auto_ptr<Spiller> spiller;
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000140 LiveIntervals *lis;
141 LiveStacks *lss;
142 VirtRegMap *vrm;
143
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000144 RegSet vregsToAlloc, emptyIntervalVRegs;
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000145
146 /// \brief Finds the initial set of vreg intervals to allocate.
147 void findVRegIntervalsToAlloc();
148
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000149 /// \brief Given a solved PBQP problem maps this solution back to a register
150 /// assignment.
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000151 bool mapPBQPToRegAlloc(const PBQPRAProblem &problem,
152 const PBQP::Solution &solution);
153
154 /// \brief Postprocessing before final spilling. Sets basic block "live in"
155 /// variables.
156 void finalizeAlloc() const;
157
158};
159
Lang Hameseb6c8f52010-09-18 09:07:10 +0000160char RegAllocPBQP::ID = 0;
Evan Chengb1290a62008-10-02 18:29:27 +0000161
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000162} // End anonymous namespace.
163
Lang Hameseb6c8f52010-09-18 09:07:10 +0000164unsigned PBQPRAProblem::getVRegForNode(PBQP::Graph::ConstNodeItr node) const {
165 Node2VReg::const_iterator vregItr = node2VReg.find(node);
166 assert(vregItr != node2VReg.end() && "No vreg for node.");
167 return vregItr->second;
168}
Evan Chengb1290a62008-10-02 18:29:27 +0000169
Lang Hameseb6c8f52010-09-18 09:07:10 +0000170PBQP::Graph::NodeItr PBQPRAProblem::getNodeForVReg(unsigned vreg) const {
171 VReg2Node::const_iterator nodeItr = vreg2Node.find(vreg);
172 assert(nodeItr != vreg2Node.end() && "No node for vreg.");
173 return nodeItr->second;
Andrew Trick16f72dd2012-02-10 04:10:26 +0000174
Lang Hameseb6c8f52010-09-18 09:07:10 +0000175}
Daniel Dunbara279bc32009-09-20 02:20:51 +0000176
Lang Hameseb6c8f52010-09-18 09:07:10 +0000177const PBQPRAProblem::AllowedSet&
178 PBQPRAProblem::getAllowedSet(unsigned vreg) const {
179 AllowedSetMap::const_iterator allowedSetItr = allowedSets.find(vreg);
180 assert(allowedSetItr != allowedSets.end() && "No pregs for vreg.");
181 const AllowedSet &allowedSet = allowedSetItr->second;
182 return allowedSet;
183}
Evan Chengb1290a62008-10-02 18:29:27 +0000184
Lang Hameseb6c8f52010-09-18 09:07:10 +0000185unsigned PBQPRAProblem::getPRegForOption(unsigned vreg, unsigned option) const {
186 assert(isPRegOption(vreg, option) && "Not a preg option.");
187
188 const AllowedSet& allowedSet = getAllowedSet(vreg);
189 assert(option <= allowedSet.size() && "Option outside allowed set.");
190 return allowedSet[option - 1];
191}
192
Lang Hamese9c93562010-09-21 13:19:36 +0000193std::auto_ptr<PBQPRAProblem> PBQPBuilder::build(MachineFunction *mf,
194 const LiveIntervals *lis,
195 const MachineLoopInfo *loopInfo,
196 const RegSet &vregs) {
Lang Hameseb6c8f52010-09-18 09:07:10 +0000197
198 typedef std::vector<const LiveInterval*> LIVector;
Lang Hamesf1113ef2012-03-23 17:33:42 +0000199 ArrayRef<SlotIndex> regMaskSlots = lis->getRegMaskSlots();
Lang Hameseb6c8f52010-09-18 09:07:10 +0000200 MachineRegisterInfo *mri = &mf->getRegInfo();
Andrew Trick16f72dd2012-02-10 04:10:26 +0000201 const TargetRegisterInfo *tri = mf->getTarget().getRegisterInfo();
Lang Hameseb6c8f52010-09-18 09:07:10 +0000202
203 std::auto_ptr<PBQPRAProblem> p(new PBQPRAProblem());
204 PBQP::Graph &g = p->getGraph();
205 RegSet pregs;
206
207 // Collect the set of preg intervals, record that they're used in the MF.
Jakob Stoklund Olesend67582e2012-06-20 21:25:05 +0000208 for (unsigned Reg = 1, e = tri->getNumRegs(); Reg != e; ++Reg) {
209 if (!lis->hasInterval(Reg))
210 continue;
211 pregs.insert(Reg);
212 mri->setPhysRegUsed(Reg);
Lang Hameseb6c8f52010-09-18 09:07:10 +0000213 }
Evan Chengb1290a62008-10-02 18:29:27 +0000214
Lang Hameseb6c8f52010-09-18 09:07:10 +0000215 BitVector reservedRegs = tri->getReservedRegs(*mf);
Evan Chengb1290a62008-10-02 18:29:27 +0000216
Andrew Trick16f72dd2012-02-10 04:10:26 +0000217 // Iterate over vregs.
Lang Hameseb6c8f52010-09-18 09:07:10 +0000218 for (RegSet::const_iterator vregItr = vregs.begin(), vregEnd = vregs.end();
219 vregItr != vregEnd; ++vregItr) {
220 unsigned vreg = *vregItr;
221 const TargetRegisterClass *trc = mri->getRegClass(vreg);
222 const LiveInterval *vregLI = &lis->getInterval(vreg);
Evan Chengb1290a62008-10-02 18:29:27 +0000223
Lang Hameseb6c8f52010-09-18 09:07:10 +0000224 // Compute an initial allowed set for the current vreg.
225 typedef std::vector<unsigned> VRAllowed;
226 VRAllowed vrAllowed;
Craig Topperb6632ba2012-03-04 10:16:38 +0000227 ArrayRef<uint16_t> rawOrder = trc->getRawAllocationOrder(*mf);
Jakob Stoklund Olesen714c0eb2011-06-16 20:37:45 +0000228 for (unsigned i = 0; i != rawOrder.size(); ++i) {
229 unsigned preg = rawOrder[i];
Lang Hameseb6c8f52010-09-18 09:07:10 +0000230 if (!reservedRegs.test(preg)) {
231 vrAllowed.push_back(preg);
Lang Hamesd0f6f012010-07-17 06:31:41 +0000232 }
Lang Hameseb6c8f52010-09-18 09:07:10 +0000233 }
Lang Hamesd0f6f012010-07-17 06:31:41 +0000234
Lang Hamesf1113ef2012-03-23 17:33:42 +0000235 RegSet overlappingPRegs;
236
237 // Record physical registers whose ranges overlap.
Lang Hameseb6c8f52010-09-18 09:07:10 +0000238 for (RegSet::const_iterator pregItr = pregs.begin(),
239 pregEnd = pregs.end();
240 pregItr != pregEnd; ++pregItr) {
241 unsigned preg = *pregItr;
242 const LiveInterval *pregLI = &lis->getInterval(preg);
Lang Hames27601ef2008-11-16 12:12:54 +0000243
Lang Hames5e77f4b2010-11-12 05:47:21 +0000244 if (pregLI->empty()) {
Lang Hameseb6c8f52010-09-18 09:07:10 +0000245 continue;
Lang Hames5e77f4b2010-11-12 05:47:21 +0000246 }
Evan Chengb1290a62008-10-02 18:29:27 +0000247
Lang Hamesf1113ef2012-03-23 17:33:42 +0000248 if (vregLI->overlaps(*pregLI))
249 overlappingPRegs.insert(preg);
250 }
251
252 // Record any overlaps with regmask operands.
253 BitVector regMaskOverlaps(tri->getNumRegs());
254 for (ArrayRef<SlotIndex>::iterator rmItr = regMaskSlots.begin(),
255 rmEnd = regMaskSlots.end();
256 rmItr != rmEnd; ++rmItr) {
257 SlotIndex rmIdx = *rmItr;
258 if (vregLI->liveAt(rmIdx)) {
259 MachineInstr *rmMI = lis->getInstructionFromIndex(rmIdx);
260 const uint32_t* regMask = 0;
261 for (MachineInstr::mop_iterator mopItr = rmMI->operands_begin(),
262 mopEnd = rmMI->operands_end();
263 mopItr != mopEnd; ++mopItr) {
264 if (mopItr->isRegMask()) {
265 regMask = mopItr->getRegMask();
266 break;
267 }
268 }
269 assert(regMask != 0 && "Couldn't find register mask.");
270 regMaskOverlaps.setBitsNotInMask(regMask);
Lang Hames5e77f4b2010-11-12 05:47:21 +0000271 }
Lang Hamesf1113ef2012-03-23 17:33:42 +0000272 }
273
274 for (unsigned preg = 0; preg < tri->getNumRegs(); ++preg) {
275 if (regMaskOverlaps.test(preg))
276 overlappingPRegs.insert(preg);
277 }
278
279 for (RegSet::const_iterator pregItr = overlappingPRegs.begin(),
280 pregEnd = overlappingPRegs.end();
281 pregItr != pregEnd; ++pregItr) {
282 unsigned preg = *pregItr;
Lang Hames030c4bf2010-01-26 04:49:58 +0000283
Lang Hameseb6c8f52010-09-18 09:07:10 +0000284 // Remove the register from the allowed set.
285 VRAllowed::iterator eraseItr =
286 std::find(vrAllowed.begin(), vrAllowed.end(), preg);
Evan Chengb1290a62008-10-02 18:29:27 +0000287
Lang Hameseb6c8f52010-09-18 09:07:10 +0000288 if (eraseItr != vrAllowed.end()) {
289 vrAllowed.erase(eraseItr);
290 }
Evan Chengb1290a62008-10-02 18:29:27 +0000291
Lang Hameseb6c8f52010-09-18 09:07:10 +0000292 // Also remove any aliases.
Jakob Stoklund Olesen396618b2012-06-01 23:28:30 +0000293 for (MCRegAliasIterator AI(preg, tri, false); AI.isValid(); ++AI) {
294 VRAllowed::iterator eraseItr =
295 std::find(vrAllowed.begin(), vrAllowed.end(), *AI);
296 if (eraseItr != vrAllowed.end()) {
297 vrAllowed.erase(eraseItr);
Lang Hameseb6c8f52010-09-18 09:07:10 +0000298 }
299 }
300 }
Misha Brukman2a835f92009-01-08 15:50:22 +0000301
Lang Hameseb6c8f52010-09-18 09:07:10 +0000302 // Construct the node.
Andrew Trick16f72dd2012-02-10 04:10:26 +0000303 PBQP::Graph::NodeItr node =
Lang Hameseb6c8f52010-09-18 09:07:10 +0000304 g.addNode(PBQP::Vector(vrAllowed.size() + 1, 0));
Evan Chengb1290a62008-10-02 18:29:27 +0000305
Lang Hameseb6c8f52010-09-18 09:07:10 +0000306 // Record the mapping and allowed set in the problem.
307 p->recordVReg(vreg, node, vrAllowed.begin(), vrAllowed.end());
Evan Chengb1290a62008-10-02 18:29:27 +0000308
Lang Hameseb6c8f52010-09-18 09:07:10 +0000309 PBQP::PBQPNum spillCost = (vregLI->weight != 0.0) ?
310 vregLI->weight : std::numeric_limits<PBQP::PBQPNum>::min();
Evan Chengb1290a62008-10-02 18:29:27 +0000311
Lang Hameseb6c8f52010-09-18 09:07:10 +0000312 addSpillCosts(g.getNodeCosts(node), spillCost);
313 }
Evan Chengb1290a62008-10-02 18:29:27 +0000314
Lang Hames481630d2010-09-18 09:49:08 +0000315 for (RegSet::const_iterator vr1Itr = vregs.begin(), vrEnd = vregs.end();
Lang Hameseb6c8f52010-09-18 09:07:10 +0000316 vr1Itr != vrEnd; ++vr1Itr) {
317 unsigned vr1 = *vr1Itr;
318 const LiveInterval &l1 = lis->getInterval(vr1);
319 const PBQPRAProblem::AllowedSet &vr1Allowed = p->getAllowedSet(vr1);
Evan Chengb1290a62008-10-02 18:29:27 +0000320
Benjamin Kramer9e8d1f92010-09-18 14:41:26 +0000321 for (RegSet::const_iterator vr2Itr = llvm::next(vr1Itr);
Lang Hameseb6c8f52010-09-18 09:07:10 +0000322 vr2Itr != vrEnd; ++vr2Itr) {
323 unsigned vr2 = *vr2Itr;
324 const LiveInterval &l2 = lis->getInterval(vr2);
325 const PBQPRAProblem::AllowedSet &vr2Allowed = p->getAllowedSet(vr2);
Evan Chengb1290a62008-10-02 18:29:27 +0000326
Lang Hameseb6c8f52010-09-18 09:07:10 +0000327 assert(!l2.empty() && "Empty interval in vreg set?");
328 if (l1.overlaps(l2)) {
329 PBQP::Graph::EdgeItr edge =
330 g.addEdge(p->getNodeForVReg(vr1), p->getNodeForVReg(vr2),
331 PBQP::Matrix(vr1Allowed.size()+1, vr2Allowed.size()+1, 0));
Lang Hames27601ef2008-11-16 12:12:54 +0000332
Lang Hameseb6c8f52010-09-18 09:07:10 +0000333 addInterferenceCosts(g.getEdgeCosts(edge), vr1Allowed, vr2Allowed, tri);
334 }
335 }
336 }
Evan Chengb1290a62008-10-02 18:29:27 +0000337
Lang Hameseb6c8f52010-09-18 09:07:10 +0000338 return p;
339}
Lang Hames27601ef2008-11-16 12:12:54 +0000340
Lang Hameseb6c8f52010-09-18 09:07:10 +0000341void PBQPBuilder::addSpillCosts(PBQP::Vector &costVec,
342 PBQP::PBQPNum spillCost) {
343 costVec[0] = spillCost;
344}
Evan Chengb1290a62008-10-02 18:29:27 +0000345
Lang Hamese9c93562010-09-21 13:19:36 +0000346void PBQPBuilder::addInterferenceCosts(
347 PBQP::Matrix &costMat,
348 const PBQPRAProblem::AllowedSet &vr1Allowed,
349 const PBQPRAProblem::AllowedSet &vr2Allowed,
350 const TargetRegisterInfo *tri) {
Lang Hameseb6c8f52010-09-18 09:07:10 +0000351 assert(costMat.getRows() == vr1Allowed.size() + 1 && "Matrix height mismatch.");
352 assert(costMat.getCols() == vr2Allowed.size() + 1 && "Matrix width mismatch.");
353
Lang Hames5e77f4b2010-11-12 05:47:21 +0000354 for (unsigned i = 0; i != vr1Allowed.size(); ++i) {
Lang Hameseb6c8f52010-09-18 09:07:10 +0000355 unsigned preg1 = vr1Allowed[i];
356
Lang Hames5e77f4b2010-11-12 05:47:21 +0000357 for (unsigned j = 0; j != vr2Allowed.size(); ++j) {
Lang Hameseb6c8f52010-09-18 09:07:10 +0000358 unsigned preg2 = vr2Allowed[j];
359
360 if (tri->regsOverlap(preg1, preg2)) {
361 costMat[i + 1][j + 1] = std::numeric_limits<PBQP::PBQPNum>::infinity();
362 }
363 }
364 }
Evan Chengb1290a62008-10-02 18:29:27 +0000365}
366
Lang Hamese9c93562010-09-21 13:19:36 +0000367std::auto_ptr<PBQPRAProblem> PBQPBuilderWithCoalescing::build(
368 MachineFunction *mf,
369 const LiveIntervals *lis,
370 const MachineLoopInfo *loopInfo,
371 const RegSet &vregs) {
372
373 std::auto_ptr<PBQPRAProblem> p = PBQPBuilder::build(mf, lis, loopInfo, vregs);
374 PBQP::Graph &g = p->getGraph();
375
376 const TargetMachine &tm = mf->getTarget();
Benjamin Kramera7542d52012-06-06 18:25:08 +0000377 CoalescerPair cp(*tm.getRegisterInfo());
Lang Hamese9c93562010-09-21 13:19:36 +0000378
379 // Scan the machine function and add a coalescing cost whenever CoalescerPair
380 // gives the Ok.
381 for (MachineFunction::const_iterator mbbItr = mf->begin(),
382 mbbEnd = mf->end();
383 mbbItr != mbbEnd; ++mbbItr) {
384 const MachineBasicBlock *mbb = &*mbbItr;
385
386 for (MachineBasicBlock::const_iterator miItr = mbb->begin(),
387 miEnd = mbb->end();
388 miItr != miEnd; ++miItr) {
389 const MachineInstr *mi = &*miItr;
390
Lang Hames5e77f4b2010-11-12 05:47:21 +0000391 if (!cp.setRegisters(mi)) {
Lang Hamese9c93562010-09-21 13:19:36 +0000392 continue; // Not coalescable.
Lang Hames5e77f4b2010-11-12 05:47:21 +0000393 }
Lang Hamese9c93562010-09-21 13:19:36 +0000394
Lang Hames5e77f4b2010-11-12 05:47:21 +0000395 if (cp.getSrcReg() == cp.getDstReg()) {
Lang Hamese9c93562010-09-21 13:19:36 +0000396 continue; // Already coalesced.
Lang Hames5e77f4b2010-11-12 05:47:21 +0000397 }
Lang Hamese9c93562010-09-21 13:19:36 +0000398
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000399 unsigned dst = cp.getDstReg(),
400 src = cp.getSrcReg();
Lang Hamese9c93562010-09-21 13:19:36 +0000401
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000402 const float copyFactor = 0.5; // Cost of copy relative to load. Current
403 // value plucked randomly out of the air.
Andrew Trick16f72dd2012-02-10 04:10:26 +0000404
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000405 PBQP::PBQPNum cBenefit =
406 copyFactor * LiveIntervals::getSpillWeight(false, true,
407 loopInfo->getLoopDepth(mbb));
Lang Hamese9c93562010-09-21 13:19:36 +0000408
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000409 if (cp.isPhys()) {
Lang Hames5e77f4b2010-11-12 05:47:21 +0000410 if (!lis->isAllocatable(dst)) {
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000411 continue;
Lang Hames5e77f4b2010-11-12 05:47:21 +0000412 }
Lang Hamese9c93562010-09-21 13:19:36 +0000413
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000414 const PBQPRAProblem::AllowedSet &allowed = p->getAllowedSet(src);
Andrew Trick16f72dd2012-02-10 04:10:26 +0000415 unsigned pregOpt = 0;
Lang Hames5e77f4b2010-11-12 05:47:21 +0000416 while (pregOpt < allowed.size() && allowed[pregOpt] != dst) {
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000417 ++pregOpt;
Lang Hames5e77f4b2010-11-12 05:47:21 +0000418 }
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000419 if (pregOpt < allowed.size()) {
420 ++pregOpt; // +1 to account for spill option.
421 PBQP::Graph::NodeItr node = p->getNodeForVReg(src);
422 addPhysRegCoalesce(g.getNodeCosts(node), pregOpt, cBenefit);
Lang Hamese9c93562010-09-21 13:19:36 +0000423 }
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000424 } else {
425 const PBQPRAProblem::AllowedSet *allowed1 = &p->getAllowedSet(dst);
426 const PBQPRAProblem::AllowedSet *allowed2 = &p->getAllowedSet(src);
427 PBQP::Graph::NodeItr node1 = p->getNodeForVReg(dst);
428 PBQP::Graph::NodeItr node2 = p->getNodeForVReg(src);
429 PBQP::Graph::EdgeItr edge = g.findEdge(node1, node2);
430 if (edge == g.edgesEnd()) {
431 edge = g.addEdge(node1, node2, PBQP::Matrix(allowed1->size() + 1,
432 allowed2->size() + 1,
433 0));
434 } else {
435 if (g.getEdgeNode1(edge) == node2) {
436 std::swap(node1, node2);
437 std::swap(allowed1, allowed2);
438 }
439 }
Andrew Trick16f72dd2012-02-10 04:10:26 +0000440
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000441 addVirtRegCoalesce(g.getEdgeCosts(edge), *allowed1, *allowed2,
442 cBenefit);
Lang Hamese9c93562010-09-21 13:19:36 +0000443 }
444 }
445 }
446
447 return p;
448}
449
Lang Hamese9c93562010-09-21 13:19:36 +0000450void PBQPBuilderWithCoalescing::addPhysRegCoalesce(PBQP::Vector &costVec,
451 unsigned pregOption,
452 PBQP::PBQPNum benefit) {
453 costVec[pregOption] += -benefit;
454}
455
456void PBQPBuilderWithCoalescing::addVirtRegCoalesce(
457 PBQP::Matrix &costMat,
458 const PBQPRAProblem::AllowedSet &vr1Allowed,
459 const PBQPRAProblem::AllowedSet &vr2Allowed,
460 PBQP::PBQPNum benefit) {
461
462 assert(costMat.getRows() == vr1Allowed.size() + 1 && "Size mismatch.");
463 assert(costMat.getCols() == vr2Allowed.size() + 1 && "Size mismatch.");
464
Lang Hames5e77f4b2010-11-12 05:47:21 +0000465 for (unsigned i = 0; i != vr1Allowed.size(); ++i) {
Lang Hamese9c93562010-09-21 13:19:36 +0000466 unsigned preg1 = vr1Allowed[i];
Lang Hames5e77f4b2010-11-12 05:47:21 +0000467 for (unsigned j = 0; j != vr2Allowed.size(); ++j) {
Lang Hamese9c93562010-09-21 13:19:36 +0000468 unsigned preg2 = vr2Allowed[j];
469
470 if (preg1 == preg2) {
471 costMat[i + 1][j + 1] += -benefit;
Andrew Trick16f72dd2012-02-10 04:10:26 +0000472 }
Lang Hamese9c93562010-09-21 13:19:36 +0000473 }
474 }
475}
Evan Chengb1290a62008-10-02 18:29:27 +0000476
Lang Hameseb6c8f52010-09-18 09:07:10 +0000477
478void RegAllocPBQP::getAnalysisUsage(AnalysisUsage &au) const {
Lang Hames9ad7e072011-12-06 01:45:57 +0000479 au.setPreservesCFG();
480 au.addRequired<AliasAnalysis>();
481 au.addPreserved<AliasAnalysis>();
Lang Hameseb6c8f52010-09-18 09:07:10 +0000482 au.addRequired<SlotIndexes>();
483 au.addPreserved<SlotIndexes>();
484 au.addRequired<LiveIntervals>();
485 //au.addRequiredID(SplitCriticalEdgesID);
Lang Hames8d857662011-06-17 07:09:01 +0000486 if (customPassID)
487 au.addRequiredID(*customPassID);
Lang Hameseb6c8f52010-09-18 09:07:10 +0000488 au.addRequired<CalculateSpillWeights>();
489 au.addRequired<LiveStacks>();
490 au.addPreserved<LiveStacks>();
Lang Hames9ad7e072011-12-06 01:45:57 +0000491 au.addRequired<MachineDominatorTree>();
492 au.addPreserved<MachineDominatorTree>();
Lang Hameseb6c8f52010-09-18 09:07:10 +0000493 au.addRequired<MachineLoopInfo>();
494 au.addPreserved<MachineLoopInfo>();
Lang Hameseb6c8f52010-09-18 09:07:10 +0000495 au.addRequired<VirtRegMap>();
496 au.addRequired<RenderMachineFunction>();
497 MachineFunctionPass::getAnalysisUsage(au);
498}
499
Lang Hameseb6c8f52010-09-18 09:07:10 +0000500void RegAllocPBQP::findVRegIntervalsToAlloc() {
Lang Hames27601ef2008-11-16 12:12:54 +0000501
502 // Iterate over all live ranges.
Jakob Stoklund Olesend67582e2012-06-20 21:25:05 +0000503 for (unsigned i = 0, e = mri->getNumVirtRegs(); i != e; ++i) {
504 unsigned Reg = TargetRegisterInfo::index2VirtReg(i);
505 if (mri->reg_nodbg_empty(Reg))
Lang Hames27601ef2008-11-16 12:12:54 +0000506 continue;
Jakob Stoklund Olesend67582e2012-06-20 21:25:05 +0000507 LiveInterval *li = &lis->getInterval(Reg);
Lang Hames27601ef2008-11-16 12:12:54 +0000508
509 // If this live interval is non-empty we will use pbqp to allocate it.
510 // Empty intervals we allocate in a simple post-processing stage in
511 // finalizeAlloc.
512 if (!li->empty()) {
Lang Hameseb6c8f52010-09-18 09:07:10 +0000513 vregsToAlloc.insert(li->reg);
Lang Hames5e77f4b2010-11-12 05:47:21 +0000514 } else {
Lang Hameseb6c8f52010-09-18 09:07:10 +0000515 emptyIntervalVRegs.insert(li->reg);
Lang Hames27601ef2008-11-16 12:12:54 +0000516 }
517 }
Evan Chengb1290a62008-10-02 18:29:27 +0000518}
519
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000520bool RegAllocPBQP::mapPBQPToRegAlloc(const PBQPRAProblem &problem,
521 const PBQP::Solution &solution) {
Lang Hameseb6c8f52010-09-18 09:07:10 +0000522 // Set to true if we have any spills
523 bool anotherRoundNeeded = false;
524
525 // Clear the existing allocation.
526 vrm->clearAllVirt();
527
528 const PBQP::Graph &g = problem.getGraph();
529 // Iterate over the nodes mapping the PBQP solution to a register
530 // assignment.
531 for (PBQP::Graph::ConstNodeItr node = g.nodesBegin(),
532 nodeEnd = g.nodesEnd();
533 node != nodeEnd; ++node) {
534 unsigned vreg = problem.getVRegForNode(node);
535 unsigned alloc = solution.getSelection(node);
536
537 if (problem.isPRegOption(vreg, alloc)) {
Andrew Trick16f72dd2012-02-10 04:10:26 +0000538 unsigned preg = problem.getPRegForOption(vreg, alloc);
Patrik Hägglundd7693872012-05-23 12:12:58 +0000539 DEBUG(dbgs() << "VREG " << PrintReg(vreg, tri) << " -> "
540 << tri->getName(preg) << "\n");
Lang Hameseb6c8f52010-09-18 09:07:10 +0000541 assert(preg != 0 && "Invalid preg selected.");
Andrew Trick16f72dd2012-02-10 04:10:26 +0000542 vrm->assignVirt2Phys(vreg, preg);
Lang Hameseb6c8f52010-09-18 09:07:10 +0000543 } else if (problem.isSpillOption(vreg, alloc)) {
544 vregsToAlloc.erase(vreg);
Jakob Stoklund Olesencfa81012011-11-12 23:17:52 +0000545 SmallVector<LiveInterval*, 8> newSpills;
Jakob Stoklund Olesen20942dc2012-05-19 05:25:46 +0000546 LiveRangeEdit LRE(&lis->getInterval(vreg), newSpills, *mf, *lis, vrm);
Jakob Stoklund Olesencfa81012011-11-12 23:17:52 +0000547 spiller->spill(LRE);
Lang Hameseb6c8f52010-09-18 09:07:10 +0000548
Patrik Hägglundd7693872012-05-23 12:12:58 +0000549 DEBUG(dbgs() << "VREG " << PrintReg(vreg, tri) << " -> SPILLED (Cost: "
Jakob Stoklund Olesencfa81012011-11-12 23:17:52 +0000550 << LRE.getParent().weight << ", New vregs: ");
Lang Hameseb6c8f52010-09-18 09:07:10 +0000551
552 // Copy any newly inserted live intervals into the list of regs to
553 // allocate.
Jakob Stoklund Olesencfa81012011-11-12 23:17:52 +0000554 for (LiveRangeEdit::iterator itr = LRE.begin(), end = LRE.end();
Lang Hameseb6c8f52010-09-18 09:07:10 +0000555 itr != end; ++itr) {
556 assert(!(*itr)->empty() && "Empty spill range.");
Patrik Hägglundd7693872012-05-23 12:12:58 +0000557 DEBUG(dbgs() << PrintReg((*itr)->reg, tri) << " ");
Lang Hameseb6c8f52010-09-18 09:07:10 +0000558 vregsToAlloc.insert((*itr)->reg);
559 }
560
561 DEBUG(dbgs() << ")\n");
562
563 // We need another round if spill intervals were added.
Jakob Stoklund Olesencfa81012011-11-12 23:17:52 +0000564 anotherRoundNeeded |= !LRE.empty();
Lang Hameseb6c8f52010-09-18 09:07:10 +0000565 } else {
Craig Topper5e25ee82012-02-05 08:31:47 +0000566 llvm_unreachable("Unknown allocation option.");
Lang Hameseb6c8f52010-09-18 09:07:10 +0000567 }
568 }
569
570 return !anotherRoundNeeded;
571}
572
573
574void RegAllocPBQP::finalizeAlloc() const {
Lang Hames27601ef2008-11-16 12:12:54 +0000575 typedef LiveIntervals::iterator LIIterator;
576 typedef LiveInterval::Ranges::const_iterator LRIterator;
577
578 // First allocate registers for the empty intervals.
Lang Hameseb6c8f52010-09-18 09:07:10 +0000579 for (RegSet::const_iterator
580 itr = emptyIntervalVRegs.begin(), end = emptyIntervalVRegs.end();
Lang Hames27601ef2008-11-16 12:12:54 +0000581 itr != end; ++itr) {
Lang Hameseb6c8f52010-09-18 09:07:10 +0000582 LiveInterval *li = &lis->getInterval(*itr);
Lang Hames27601ef2008-11-16 12:12:54 +0000583
Evan Cheng90f95f82009-06-14 20:22:55 +0000584 unsigned physReg = vrm->getRegAllocPref(li->reg);
Lang Hames6699fb22009-08-06 23:32:48 +0000585
Lang Hames27601ef2008-11-16 12:12:54 +0000586 if (physReg == 0) {
587 const TargetRegisterClass *liRC = mri->getRegClass(li->reg);
Jakob Stoklund Olesen714c0eb2011-06-16 20:37:45 +0000588 physReg = liRC->getRawAllocationOrder(*mf).front();
Lang Hames27601ef2008-11-16 12:12:54 +0000589 }
Misha Brukman2a835f92009-01-08 15:50:22 +0000590
591 vrm->assignVirt2Phys(li->reg, physReg);
Lang Hames27601ef2008-11-16 12:12:54 +0000592 }
Lang Hames27601ef2008-11-16 12:12:54 +0000593}
594
Lang Hameseb6c8f52010-09-18 09:07:10 +0000595bool RegAllocPBQP::runOnMachineFunction(MachineFunction &MF) {
Lang Hames27601ef2008-11-16 12:12:54 +0000596
Evan Chengb1290a62008-10-02 18:29:27 +0000597 mf = &MF;
598 tm = &mf->getTarget();
599 tri = tm->getRegisterInfo();
Lang Hames27601ef2008-11-16 12:12:54 +0000600 tii = tm->getInstrInfo();
Andrew Trick16f72dd2012-02-10 04:10:26 +0000601 mri = &mf->getRegInfo();
Evan Chengb1290a62008-10-02 18:29:27 +0000602
Lang Hames27601ef2008-11-16 12:12:54 +0000603 lis = &getAnalysis<LiveIntervals>();
604 lss = &getAnalysis<LiveStacks>();
Evan Chengb1290a62008-10-02 18:29:27 +0000605 loopInfo = &getAnalysis<MachineLoopInfo>();
Lang Hames33198392010-09-02 08:27:00 +0000606 rmf = &getAnalysis<RenderMachineFunction>();
Evan Chengb1290a62008-10-02 18:29:27 +0000607
Owen Anderson49c8aa02009-03-13 05:55:11 +0000608 vrm = &getAnalysis<VirtRegMap>();
Jakob Stoklund Olesencfa81012011-11-12 23:17:52 +0000609 spiller.reset(createInlineSpiller(*this, MF, *vrm));
Evan Chengb1290a62008-10-02 18:29:27 +0000610
Jakob Stoklund Olesend9e5c762012-01-05 00:26:49 +0000611 mri->freezeReservedRegs(MF);
Lang Hames54cc2ef2010-07-19 15:22:28 +0000612
Lang Hames030c4bf2010-01-26 04:49:58 +0000613 DEBUG(dbgs() << "PBQP Register Allocating for " << mf->getFunction()->getName() << "\n");
Lang Hames27601ef2008-11-16 12:12:54 +0000614
Evan Chengb1290a62008-10-02 18:29:27 +0000615 // Allocator main loop:
Misha Brukman2a835f92009-01-08 15:50:22 +0000616 //
Evan Chengb1290a62008-10-02 18:29:27 +0000617 // * Map current regalloc problem to a PBQP problem
618 // * Solve the PBQP problem
619 // * Map the solution back to a register allocation
620 // * Spill if necessary
Misha Brukman2a835f92009-01-08 15:50:22 +0000621 //
Evan Chengb1290a62008-10-02 18:29:27 +0000622 // This process is continued till no more spills are generated.
623
Lang Hames27601ef2008-11-16 12:12:54 +0000624 // Find the vreg intervals in need of allocation.
625 findVRegIntervalsToAlloc();
Misha Brukman2a835f92009-01-08 15:50:22 +0000626
Lang Hames20df03c2012-03-26 23:07:23 +0000627 const Function* func = mf->getFunction();
628 std::string fqn =
629 func->getParent()->getModuleIdentifier() + "." +
630 func->getName().str();
631 (void)fqn;
632
Lang Hames27601ef2008-11-16 12:12:54 +0000633 // If there are non-empty intervals allocate them using pbqp.
Lang Hameseb6c8f52010-09-18 09:07:10 +0000634 if (!vregsToAlloc.empty()) {
Evan Chengb1290a62008-10-02 18:29:27 +0000635
Lang Hames27601ef2008-11-16 12:12:54 +0000636 bool pbqpAllocComplete = false;
637 unsigned round = 0;
638
Lang Hamesab62b7e2010-10-04 12:13:07 +0000639 while (!pbqpAllocComplete) {
640 DEBUG(dbgs() << " PBQP Regalloc round " << round << ":\n");
Lang Hames27601ef2008-11-16 12:12:54 +0000641
Lang Hamesab62b7e2010-10-04 12:13:07 +0000642 std::auto_ptr<PBQPRAProblem> problem =
643 builder->build(mf, lis, loopInfo, vregsToAlloc);
Lang Hames20df03c2012-03-26 23:07:23 +0000644
645#ifndef NDEBUG
646 if (pbqpDumpGraphs) {
647 std::ostringstream rs;
648 rs << round;
649 std::string graphFileName(fqn + "." + rs.str() + ".pbqpgraph");
650 std::string tmp;
651 raw_fd_ostream os(graphFileName.c_str(), tmp);
652 DEBUG(dbgs() << "Dumping graph for round " << round << " to \""
653 << graphFileName << "\"\n");
654 problem->getGraph().dump(os);
655 }
656#endif
657
Lang Hamesab62b7e2010-10-04 12:13:07 +0000658 PBQP::Solution solution =
659 PBQP::HeuristicSolver<PBQP::Heuristics::Briggs>::solve(
660 problem->getGraph());
Lang Hames233fd9c2009-08-18 23:34:50 +0000661
Lang Hamesab62b7e2010-10-04 12:13:07 +0000662 pbqpAllocComplete = mapPBQPToRegAlloc(*problem, solution);
Lang Hames27601ef2008-11-16 12:12:54 +0000663
Lang Hamesab62b7e2010-10-04 12:13:07 +0000664 ++round;
Lang Hames27601ef2008-11-16 12:12:54 +0000665 }
Evan Chengb1290a62008-10-02 18:29:27 +0000666 }
667
Lang Hames27601ef2008-11-16 12:12:54 +0000668 // Finalise allocation, allocate empty ranges.
669 finalizeAlloc();
Evan Chengb1290a62008-10-02 18:29:27 +0000670
Lang Hamesc4bcc772010-07-20 07:41:44 +0000671 rmf->renderMachineFunction("After PBQP register allocation.", vrm);
672
Lang Hameseb6c8f52010-09-18 09:07:10 +0000673 vregsToAlloc.clear();
674 emptyIntervalVRegs.clear();
Lang Hames27601ef2008-11-16 12:12:54 +0000675
David Greene30931542010-01-05 01:25:43 +0000676 DEBUG(dbgs() << "Post alloc VirtRegMap:\n" << *vrm << "\n");
Lang Hames27601ef2008-11-16 12:12:54 +0000677
Misha Brukman2a835f92009-01-08 15:50:22 +0000678 return true;
Evan Chengb1290a62008-10-02 18:29:27 +0000679}
680
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000681FunctionPass* llvm::createPBQPRegisterAllocator(
Lang Hames8d857662011-06-17 07:09:01 +0000682 std::auto_ptr<PBQPBuilder> builder,
683 char *customPassID) {
684 return new RegAllocPBQP(builder, customPassID);
Evan Chengb1290a62008-10-02 18:29:27 +0000685}
686
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000687FunctionPass* llvm::createDefaultPBQPRegisterAllocator() {
688 if (pbqpCoalescing) {
689 return createPBQPRegisterAllocator(
690 std::auto_ptr<PBQPBuilder>(new PBQPBuilderWithCoalescing()));
691 } // else
692 return createPBQPRegisterAllocator(
693 std::auto_ptr<PBQPBuilder>(new PBQPBuilder()));
Lang Hameseb6c8f52010-09-18 09:07:10 +0000694}
Evan Chengb1290a62008-10-02 18:29:27 +0000695
696#undef DEBUG_TYPE