blob: 2c58d7dda20855025f8696c44587aa24ce958b3b [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.
208 for (LiveIntervals::const_iterator itr = lis->begin(), end = lis->end();
209 itr != end; ++itr) {
210 if (TargetRegisterInfo::isPhysicalRegister(itr->first)) {
211 pregs.insert(itr->first);
212 mri->setPhysRegUsed(itr->first);
Evan Chengb1290a62008-10-02 18:29:27 +0000213 }
Lang Hameseb6c8f52010-09-18 09:07:10 +0000214 }
Evan Chengb1290a62008-10-02 18:29:27 +0000215
Lang Hameseb6c8f52010-09-18 09:07:10 +0000216 BitVector reservedRegs = tri->getReservedRegs(*mf);
Evan Chengb1290a62008-10-02 18:29:27 +0000217
Andrew Trick16f72dd2012-02-10 04:10:26 +0000218 // Iterate over vregs.
Lang Hameseb6c8f52010-09-18 09:07:10 +0000219 for (RegSet::const_iterator vregItr = vregs.begin(), vregEnd = vregs.end();
220 vregItr != vregEnd; ++vregItr) {
221 unsigned vreg = *vregItr;
222 const TargetRegisterClass *trc = mri->getRegClass(vreg);
223 const LiveInterval *vregLI = &lis->getInterval(vreg);
Evan Chengb1290a62008-10-02 18:29:27 +0000224
Lang Hameseb6c8f52010-09-18 09:07:10 +0000225 // Compute an initial allowed set for the current vreg.
226 typedef std::vector<unsigned> VRAllowed;
227 VRAllowed vrAllowed;
Craig Topperb6632ba2012-03-04 10:16:38 +0000228 ArrayRef<uint16_t> rawOrder = trc->getRawAllocationOrder(*mf);
Jakob Stoklund Olesen714c0eb2011-06-16 20:37:45 +0000229 for (unsigned i = 0; i != rawOrder.size(); ++i) {
230 unsigned preg = rawOrder[i];
Lang Hameseb6c8f52010-09-18 09:07:10 +0000231 if (!reservedRegs.test(preg)) {
232 vrAllowed.push_back(preg);
Lang Hamesd0f6f012010-07-17 06:31:41 +0000233 }
Lang Hameseb6c8f52010-09-18 09:07:10 +0000234 }
Lang Hamesd0f6f012010-07-17 06:31:41 +0000235
Lang Hamesf1113ef2012-03-23 17:33:42 +0000236 RegSet overlappingPRegs;
237
238 // Record physical registers whose ranges overlap.
Lang Hameseb6c8f52010-09-18 09:07:10 +0000239 for (RegSet::const_iterator pregItr = pregs.begin(),
240 pregEnd = pregs.end();
241 pregItr != pregEnd; ++pregItr) {
242 unsigned preg = *pregItr;
243 const LiveInterval *pregLI = &lis->getInterval(preg);
Lang Hames27601ef2008-11-16 12:12:54 +0000244
Lang Hames5e77f4b2010-11-12 05:47:21 +0000245 if (pregLI->empty()) {
Lang Hameseb6c8f52010-09-18 09:07:10 +0000246 continue;
Lang Hames5e77f4b2010-11-12 05:47:21 +0000247 }
Evan Chengb1290a62008-10-02 18:29:27 +0000248
Lang Hamesf1113ef2012-03-23 17:33:42 +0000249 if (vregLI->overlaps(*pregLI))
250 overlappingPRegs.insert(preg);
251 }
252
253 // Record any overlaps with regmask operands.
254 BitVector regMaskOverlaps(tri->getNumRegs());
255 for (ArrayRef<SlotIndex>::iterator rmItr = regMaskSlots.begin(),
256 rmEnd = regMaskSlots.end();
257 rmItr != rmEnd; ++rmItr) {
258 SlotIndex rmIdx = *rmItr;
259 if (vregLI->liveAt(rmIdx)) {
260 MachineInstr *rmMI = lis->getInstructionFromIndex(rmIdx);
261 const uint32_t* regMask = 0;
262 for (MachineInstr::mop_iterator mopItr = rmMI->operands_begin(),
263 mopEnd = rmMI->operands_end();
264 mopItr != mopEnd; ++mopItr) {
265 if (mopItr->isRegMask()) {
266 regMask = mopItr->getRegMask();
267 break;
268 }
269 }
270 assert(regMask != 0 && "Couldn't find register mask.");
271 regMaskOverlaps.setBitsNotInMask(regMask);
Lang Hames5e77f4b2010-11-12 05:47:21 +0000272 }
Lang Hamesf1113ef2012-03-23 17:33:42 +0000273 }
274
275 for (unsigned preg = 0; preg < tri->getNumRegs(); ++preg) {
276 if (regMaskOverlaps.test(preg))
277 overlappingPRegs.insert(preg);
278 }
279
280 for (RegSet::const_iterator pregItr = overlappingPRegs.begin(),
281 pregEnd = overlappingPRegs.end();
282 pregItr != pregEnd; ++pregItr) {
283 unsigned preg = *pregItr;
Lang Hames030c4bf2010-01-26 04:49:58 +0000284
Lang Hameseb6c8f52010-09-18 09:07:10 +0000285 // Remove the register from the allowed set.
286 VRAllowed::iterator eraseItr =
287 std::find(vrAllowed.begin(), vrAllowed.end(), preg);
Evan Chengb1290a62008-10-02 18:29:27 +0000288
Lang Hameseb6c8f52010-09-18 09:07:10 +0000289 if (eraseItr != vrAllowed.end()) {
290 vrAllowed.erase(eraseItr);
291 }
Evan Chengb1290a62008-10-02 18:29:27 +0000292
Lang Hameseb6c8f52010-09-18 09:07:10 +0000293 // Also remove any aliases.
Jakob Stoklund Olesen396618b2012-06-01 23:28:30 +0000294 for (MCRegAliasIterator AI(preg, tri, false); AI.isValid(); ++AI) {
295 VRAllowed::iterator eraseItr =
296 std::find(vrAllowed.begin(), vrAllowed.end(), *AI);
297 if (eraseItr != vrAllowed.end()) {
298 vrAllowed.erase(eraseItr);
Lang Hameseb6c8f52010-09-18 09:07:10 +0000299 }
300 }
301 }
Misha Brukman2a835f92009-01-08 15:50:22 +0000302
Lang Hameseb6c8f52010-09-18 09:07:10 +0000303 // Construct the node.
Andrew Trick16f72dd2012-02-10 04:10:26 +0000304 PBQP::Graph::NodeItr node =
Lang Hameseb6c8f52010-09-18 09:07:10 +0000305 g.addNode(PBQP::Vector(vrAllowed.size() + 1, 0));
Evan Chengb1290a62008-10-02 18:29:27 +0000306
Lang Hameseb6c8f52010-09-18 09:07:10 +0000307 // Record the mapping and allowed set in the problem.
308 p->recordVReg(vreg, node, vrAllowed.begin(), vrAllowed.end());
Evan Chengb1290a62008-10-02 18:29:27 +0000309
Lang Hameseb6c8f52010-09-18 09:07:10 +0000310 PBQP::PBQPNum spillCost = (vregLI->weight != 0.0) ?
311 vregLI->weight : std::numeric_limits<PBQP::PBQPNum>::min();
Evan Chengb1290a62008-10-02 18:29:27 +0000312
Lang Hameseb6c8f52010-09-18 09:07:10 +0000313 addSpillCosts(g.getNodeCosts(node), spillCost);
314 }
Evan Chengb1290a62008-10-02 18:29:27 +0000315
Lang Hames481630d2010-09-18 09:49:08 +0000316 for (RegSet::const_iterator vr1Itr = vregs.begin(), vrEnd = vregs.end();
Lang Hameseb6c8f52010-09-18 09:07:10 +0000317 vr1Itr != vrEnd; ++vr1Itr) {
318 unsigned vr1 = *vr1Itr;
319 const LiveInterval &l1 = lis->getInterval(vr1);
320 const PBQPRAProblem::AllowedSet &vr1Allowed = p->getAllowedSet(vr1);
Evan Chengb1290a62008-10-02 18:29:27 +0000321
Benjamin Kramer9e8d1f92010-09-18 14:41:26 +0000322 for (RegSet::const_iterator vr2Itr = llvm::next(vr1Itr);
Lang Hameseb6c8f52010-09-18 09:07:10 +0000323 vr2Itr != vrEnd; ++vr2Itr) {
324 unsigned vr2 = *vr2Itr;
325 const LiveInterval &l2 = lis->getInterval(vr2);
326 const PBQPRAProblem::AllowedSet &vr2Allowed = p->getAllowedSet(vr2);
Evan Chengb1290a62008-10-02 18:29:27 +0000327
Lang Hameseb6c8f52010-09-18 09:07:10 +0000328 assert(!l2.empty() && "Empty interval in vreg set?");
329 if (l1.overlaps(l2)) {
330 PBQP::Graph::EdgeItr edge =
331 g.addEdge(p->getNodeForVReg(vr1), p->getNodeForVReg(vr2),
332 PBQP::Matrix(vr1Allowed.size()+1, vr2Allowed.size()+1, 0));
Lang Hames27601ef2008-11-16 12:12:54 +0000333
Lang Hameseb6c8f52010-09-18 09:07:10 +0000334 addInterferenceCosts(g.getEdgeCosts(edge), vr1Allowed, vr2Allowed, tri);
335 }
336 }
337 }
Evan Chengb1290a62008-10-02 18:29:27 +0000338
Lang Hameseb6c8f52010-09-18 09:07:10 +0000339 return p;
340}
Lang Hames27601ef2008-11-16 12:12:54 +0000341
Lang Hameseb6c8f52010-09-18 09:07:10 +0000342void PBQPBuilder::addSpillCosts(PBQP::Vector &costVec,
343 PBQP::PBQPNum spillCost) {
344 costVec[0] = spillCost;
345}
Evan Chengb1290a62008-10-02 18:29:27 +0000346
Lang Hamese9c93562010-09-21 13:19:36 +0000347void PBQPBuilder::addInterferenceCosts(
348 PBQP::Matrix &costMat,
349 const PBQPRAProblem::AllowedSet &vr1Allowed,
350 const PBQPRAProblem::AllowedSet &vr2Allowed,
351 const TargetRegisterInfo *tri) {
Lang Hameseb6c8f52010-09-18 09:07:10 +0000352 assert(costMat.getRows() == vr1Allowed.size() + 1 && "Matrix height mismatch.");
353 assert(costMat.getCols() == vr2Allowed.size() + 1 && "Matrix width mismatch.");
354
Lang Hames5e77f4b2010-11-12 05:47:21 +0000355 for (unsigned i = 0; i != vr1Allowed.size(); ++i) {
Lang Hameseb6c8f52010-09-18 09:07:10 +0000356 unsigned preg1 = vr1Allowed[i];
357
Lang Hames5e77f4b2010-11-12 05:47:21 +0000358 for (unsigned j = 0; j != vr2Allowed.size(); ++j) {
Lang Hameseb6c8f52010-09-18 09:07:10 +0000359 unsigned preg2 = vr2Allowed[j];
360
361 if (tri->regsOverlap(preg1, preg2)) {
362 costMat[i + 1][j + 1] = std::numeric_limits<PBQP::PBQPNum>::infinity();
363 }
364 }
365 }
Evan Chengb1290a62008-10-02 18:29:27 +0000366}
367
Lang Hamese9c93562010-09-21 13:19:36 +0000368std::auto_ptr<PBQPRAProblem> PBQPBuilderWithCoalescing::build(
369 MachineFunction *mf,
370 const LiveIntervals *lis,
371 const MachineLoopInfo *loopInfo,
372 const RegSet &vregs) {
373
374 std::auto_ptr<PBQPRAProblem> p = PBQPBuilder::build(mf, lis, loopInfo, vregs);
375 PBQP::Graph &g = p->getGraph();
376
377 const TargetMachine &tm = mf->getTarget();
Benjamin Kramera7542d52012-06-06 18:25:08 +0000378 CoalescerPair cp(*tm.getRegisterInfo());
Lang Hamese9c93562010-09-21 13:19:36 +0000379
380 // Scan the machine function and add a coalescing cost whenever CoalescerPair
381 // gives the Ok.
382 for (MachineFunction::const_iterator mbbItr = mf->begin(),
383 mbbEnd = mf->end();
384 mbbItr != mbbEnd; ++mbbItr) {
385 const MachineBasicBlock *mbb = &*mbbItr;
386
387 for (MachineBasicBlock::const_iterator miItr = mbb->begin(),
388 miEnd = mbb->end();
389 miItr != miEnd; ++miItr) {
390 const MachineInstr *mi = &*miItr;
391
Lang Hames5e77f4b2010-11-12 05:47:21 +0000392 if (!cp.setRegisters(mi)) {
Lang Hamese9c93562010-09-21 13:19:36 +0000393 continue; // Not coalescable.
Lang Hames5e77f4b2010-11-12 05:47:21 +0000394 }
Lang Hamese9c93562010-09-21 13:19:36 +0000395
Lang Hames5e77f4b2010-11-12 05:47:21 +0000396 if (cp.getSrcReg() == cp.getDstReg()) {
Lang Hamese9c93562010-09-21 13:19:36 +0000397 continue; // Already coalesced.
Lang Hames5e77f4b2010-11-12 05:47:21 +0000398 }
Lang Hamese9c93562010-09-21 13:19:36 +0000399
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000400 unsigned dst = cp.getDstReg(),
401 src = cp.getSrcReg();
Lang Hamese9c93562010-09-21 13:19:36 +0000402
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000403 const float copyFactor = 0.5; // Cost of copy relative to load. Current
404 // value plucked randomly out of the air.
Andrew Trick16f72dd2012-02-10 04:10:26 +0000405
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000406 PBQP::PBQPNum cBenefit =
407 copyFactor * LiveIntervals::getSpillWeight(false, true,
408 loopInfo->getLoopDepth(mbb));
Lang Hamese9c93562010-09-21 13:19:36 +0000409
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000410 if (cp.isPhys()) {
Lang Hames5e77f4b2010-11-12 05:47:21 +0000411 if (!lis->isAllocatable(dst)) {
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000412 continue;
Lang Hames5e77f4b2010-11-12 05:47:21 +0000413 }
Lang Hamese9c93562010-09-21 13:19:36 +0000414
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000415 const PBQPRAProblem::AllowedSet &allowed = p->getAllowedSet(src);
Andrew Trick16f72dd2012-02-10 04:10:26 +0000416 unsigned pregOpt = 0;
Lang Hames5e77f4b2010-11-12 05:47:21 +0000417 while (pregOpt < allowed.size() && allowed[pregOpt] != dst) {
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000418 ++pregOpt;
Lang Hames5e77f4b2010-11-12 05:47:21 +0000419 }
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000420 if (pregOpt < allowed.size()) {
421 ++pregOpt; // +1 to account for spill option.
422 PBQP::Graph::NodeItr node = p->getNodeForVReg(src);
423 addPhysRegCoalesce(g.getNodeCosts(node), pregOpt, cBenefit);
Lang Hamese9c93562010-09-21 13:19:36 +0000424 }
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000425 } else {
426 const PBQPRAProblem::AllowedSet *allowed1 = &p->getAllowedSet(dst);
427 const PBQPRAProblem::AllowedSet *allowed2 = &p->getAllowedSet(src);
428 PBQP::Graph::NodeItr node1 = p->getNodeForVReg(dst);
429 PBQP::Graph::NodeItr node2 = p->getNodeForVReg(src);
430 PBQP::Graph::EdgeItr edge = g.findEdge(node1, node2);
431 if (edge == g.edgesEnd()) {
432 edge = g.addEdge(node1, node2, PBQP::Matrix(allowed1->size() + 1,
433 allowed2->size() + 1,
434 0));
435 } else {
436 if (g.getEdgeNode1(edge) == node2) {
437 std::swap(node1, node2);
438 std::swap(allowed1, allowed2);
439 }
440 }
Andrew Trick16f72dd2012-02-10 04:10:26 +0000441
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000442 addVirtRegCoalesce(g.getEdgeCosts(edge), *allowed1, *allowed2,
443 cBenefit);
Lang Hamese9c93562010-09-21 13:19:36 +0000444 }
445 }
446 }
447
448 return p;
449}
450
Lang Hamese9c93562010-09-21 13:19:36 +0000451void PBQPBuilderWithCoalescing::addPhysRegCoalesce(PBQP::Vector &costVec,
452 unsigned pregOption,
453 PBQP::PBQPNum benefit) {
454 costVec[pregOption] += -benefit;
455}
456
457void PBQPBuilderWithCoalescing::addVirtRegCoalesce(
458 PBQP::Matrix &costMat,
459 const PBQPRAProblem::AllowedSet &vr1Allowed,
460 const PBQPRAProblem::AllowedSet &vr2Allowed,
461 PBQP::PBQPNum benefit) {
462
463 assert(costMat.getRows() == vr1Allowed.size() + 1 && "Size mismatch.");
464 assert(costMat.getCols() == vr2Allowed.size() + 1 && "Size mismatch.");
465
Lang Hames5e77f4b2010-11-12 05:47:21 +0000466 for (unsigned i = 0; i != vr1Allowed.size(); ++i) {
Lang Hamese9c93562010-09-21 13:19:36 +0000467 unsigned preg1 = vr1Allowed[i];
Lang Hames5e77f4b2010-11-12 05:47:21 +0000468 for (unsigned j = 0; j != vr2Allowed.size(); ++j) {
Lang Hamese9c93562010-09-21 13:19:36 +0000469 unsigned preg2 = vr2Allowed[j];
470
471 if (preg1 == preg2) {
472 costMat[i + 1][j + 1] += -benefit;
Andrew Trick16f72dd2012-02-10 04:10:26 +0000473 }
Lang Hamese9c93562010-09-21 13:19:36 +0000474 }
475 }
476}
Evan Chengb1290a62008-10-02 18:29:27 +0000477
Lang Hameseb6c8f52010-09-18 09:07:10 +0000478
479void RegAllocPBQP::getAnalysisUsage(AnalysisUsage &au) const {
Lang Hames9ad7e072011-12-06 01:45:57 +0000480 au.setPreservesCFG();
481 au.addRequired<AliasAnalysis>();
482 au.addPreserved<AliasAnalysis>();
Lang Hameseb6c8f52010-09-18 09:07:10 +0000483 au.addRequired<SlotIndexes>();
484 au.addPreserved<SlotIndexes>();
485 au.addRequired<LiveIntervals>();
486 //au.addRequiredID(SplitCriticalEdgesID);
Lang Hames8d857662011-06-17 07:09:01 +0000487 if (customPassID)
488 au.addRequiredID(*customPassID);
Lang Hameseb6c8f52010-09-18 09:07:10 +0000489 au.addRequired<CalculateSpillWeights>();
490 au.addRequired<LiveStacks>();
491 au.addPreserved<LiveStacks>();
Lang Hames9ad7e072011-12-06 01:45:57 +0000492 au.addRequired<MachineDominatorTree>();
493 au.addPreserved<MachineDominatorTree>();
Lang Hameseb6c8f52010-09-18 09:07:10 +0000494 au.addRequired<MachineLoopInfo>();
495 au.addPreserved<MachineLoopInfo>();
Lang Hameseb6c8f52010-09-18 09:07:10 +0000496 au.addRequired<VirtRegMap>();
497 au.addRequired<RenderMachineFunction>();
498 MachineFunctionPass::getAnalysisUsage(au);
499}
500
Lang Hameseb6c8f52010-09-18 09:07:10 +0000501void RegAllocPBQP::findVRegIntervalsToAlloc() {
Lang Hames27601ef2008-11-16 12:12:54 +0000502
503 // Iterate over all live ranges.
504 for (LiveIntervals::iterator itr = lis->begin(), end = lis->end();
505 itr != end; ++itr) {
506
507 // Ignore physical ones.
508 if (TargetRegisterInfo::isPhysicalRegister(itr->first))
509 continue;
510
511 LiveInterval *li = itr->second;
512
513 // If this live interval is non-empty we will use pbqp to allocate it.
514 // Empty intervals we allocate in a simple post-processing stage in
515 // finalizeAlloc.
516 if (!li->empty()) {
Lang Hameseb6c8f52010-09-18 09:07:10 +0000517 vregsToAlloc.insert(li->reg);
Lang Hames5e77f4b2010-11-12 05:47:21 +0000518 } else {
Lang Hameseb6c8f52010-09-18 09:07:10 +0000519 emptyIntervalVRegs.insert(li->reg);
Lang Hames27601ef2008-11-16 12:12:54 +0000520 }
521 }
Evan Chengb1290a62008-10-02 18:29:27 +0000522}
523
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000524bool RegAllocPBQP::mapPBQPToRegAlloc(const PBQPRAProblem &problem,
525 const PBQP::Solution &solution) {
Lang Hameseb6c8f52010-09-18 09:07:10 +0000526 // Set to true if we have any spills
527 bool anotherRoundNeeded = false;
528
529 // Clear the existing allocation.
530 vrm->clearAllVirt();
531
532 const PBQP::Graph &g = problem.getGraph();
533 // Iterate over the nodes mapping the PBQP solution to a register
534 // assignment.
535 for (PBQP::Graph::ConstNodeItr node = g.nodesBegin(),
536 nodeEnd = g.nodesEnd();
537 node != nodeEnd; ++node) {
538 unsigned vreg = problem.getVRegForNode(node);
539 unsigned alloc = solution.getSelection(node);
540
541 if (problem.isPRegOption(vreg, alloc)) {
Andrew Trick16f72dd2012-02-10 04:10:26 +0000542 unsigned preg = problem.getPRegForOption(vreg, alloc);
Patrik Hägglundd7693872012-05-23 12:12:58 +0000543 DEBUG(dbgs() << "VREG " << PrintReg(vreg, tri) << " -> "
544 << tri->getName(preg) << "\n");
Lang Hameseb6c8f52010-09-18 09:07:10 +0000545 assert(preg != 0 && "Invalid preg selected.");
Andrew Trick16f72dd2012-02-10 04:10:26 +0000546 vrm->assignVirt2Phys(vreg, preg);
Lang Hameseb6c8f52010-09-18 09:07:10 +0000547 } else if (problem.isSpillOption(vreg, alloc)) {
548 vregsToAlloc.erase(vreg);
Jakob Stoklund Olesencfa81012011-11-12 23:17:52 +0000549 SmallVector<LiveInterval*, 8> newSpills;
Jakob Stoklund Olesen20942dc2012-05-19 05:25:46 +0000550 LiveRangeEdit LRE(&lis->getInterval(vreg), newSpills, *mf, *lis, vrm);
Jakob Stoklund Olesencfa81012011-11-12 23:17:52 +0000551 spiller->spill(LRE);
Lang Hameseb6c8f52010-09-18 09:07:10 +0000552
Patrik Hägglundd7693872012-05-23 12:12:58 +0000553 DEBUG(dbgs() << "VREG " << PrintReg(vreg, tri) << " -> SPILLED (Cost: "
Jakob Stoklund Olesencfa81012011-11-12 23:17:52 +0000554 << LRE.getParent().weight << ", New vregs: ");
Lang Hameseb6c8f52010-09-18 09:07:10 +0000555
556 // Copy any newly inserted live intervals into the list of regs to
557 // allocate.
Jakob Stoklund Olesencfa81012011-11-12 23:17:52 +0000558 for (LiveRangeEdit::iterator itr = LRE.begin(), end = LRE.end();
Lang Hameseb6c8f52010-09-18 09:07:10 +0000559 itr != end; ++itr) {
560 assert(!(*itr)->empty() && "Empty spill range.");
Patrik Hägglundd7693872012-05-23 12:12:58 +0000561 DEBUG(dbgs() << PrintReg((*itr)->reg, tri) << " ");
Lang Hameseb6c8f52010-09-18 09:07:10 +0000562 vregsToAlloc.insert((*itr)->reg);
563 }
564
565 DEBUG(dbgs() << ")\n");
566
567 // We need another round if spill intervals were added.
Jakob Stoklund Olesencfa81012011-11-12 23:17:52 +0000568 anotherRoundNeeded |= !LRE.empty();
Lang Hameseb6c8f52010-09-18 09:07:10 +0000569 } else {
Craig Topper5e25ee82012-02-05 08:31:47 +0000570 llvm_unreachable("Unknown allocation option.");
Lang Hameseb6c8f52010-09-18 09:07:10 +0000571 }
572 }
573
574 return !anotherRoundNeeded;
575}
576
577
578void RegAllocPBQP::finalizeAlloc() const {
Lang Hames27601ef2008-11-16 12:12:54 +0000579 typedef LiveIntervals::iterator LIIterator;
580 typedef LiveInterval::Ranges::const_iterator LRIterator;
581
582 // First allocate registers for the empty intervals.
Lang Hameseb6c8f52010-09-18 09:07:10 +0000583 for (RegSet::const_iterator
584 itr = emptyIntervalVRegs.begin(), end = emptyIntervalVRegs.end();
Lang Hames27601ef2008-11-16 12:12:54 +0000585 itr != end; ++itr) {
Lang Hameseb6c8f52010-09-18 09:07:10 +0000586 LiveInterval *li = &lis->getInterval(*itr);
Lang Hames27601ef2008-11-16 12:12:54 +0000587
Evan Cheng90f95f82009-06-14 20:22:55 +0000588 unsigned physReg = vrm->getRegAllocPref(li->reg);
Lang Hames6699fb22009-08-06 23:32:48 +0000589
Lang Hames27601ef2008-11-16 12:12:54 +0000590 if (physReg == 0) {
591 const TargetRegisterClass *liRC = mri->getRegClass(li->reg);
Jakob Stoklund Olesen714c0eb2011-06-16 20:37:45 +0000592 physReg = liRC->getRawAllocationOrder(*mf).front();
Lang Hames27601ef2008-11-16 12:12:54 +0000593 }
Misha Brukman2a835f92009-01-08 15:50:22 +0000594
595 vrm->assignVirt2Phys(li->reg, physReg);
Lang Hames27601ef2008-11-16 12:12:54 +0000596 }
Lang Hames27601ef2008-11-16 12:12:54 +0000597}
598
Lang Hameseb6c8f52010-09-18 09:07:10 +0000599bool RegAllocPBQP::runOnMachineFunction(MachineFunction &MF) {
Lang Hames27601ef2008-11-16 12:12:54 +0000600
Evan Chengb1290a62008-10-02 18:29:27 +0000601 mf = &MF;
602 tm = &mf->getTarget();
603 tri = tm->getRegisterInfo();
Lang Hames27601ef2008-11-16 12:12:54 +0000604 tii = tm->getInstrInfo();
Andrew Trick16f72dd2012-02-10 04:10:26 +0000605 mri = &mf->getRegInfo();
Evan Chengb1290a62008-10-02 18:29:27 +0000606
Lang Hames27601ef2008-11-16 12:12:54 +0000607 lis = &getAnalysis<LiveIntervals>();
608 lss = &getAnalysis<LiveStacks>();
Evan Chengb1290a62008-10-02 18:29:27 +0000609 loopInfo = &getAnalysis<MachineLoopInfo>();
Lang Hames33198392010-09-02 08:27:00 +0000610 rmf = &getAnalysis<RenderMachineFunction>();
Evan Chengb1290a62008-10-02 18:29:27 +0000611
Owen Anderson49c8aa02009-03-13 05:55:11 +0000612 vrm = &getAnalysis<VirtRegMap>();
Jakob Stoklund Olesencfa81012011-11-12 23:17:52 +0000613 spiller.reset(createInlineSpiller(*this, MF, *vrm));
Evan Chengb1290a62008-10-02 18:29:27 +0000614
Jakob Stoklund Olesend9e5c762012-01-05 00:26:49 +0000615 mri->freezeReservedRegs(MF);
Lang Hames54cc2ef2010-07-19 15:22:28 +0000616
Lang Hames030c4bf2010-01-26 04:49:58 +0000617 DEBUG(dbgs() << "PBQP Register Allocating for " << mf->getFunction()->getName() << "\n");
Lang Hames27601ef2008-11-16 12:12:54 +0000618
Evan Chengb1290a62008-10-02 18:29:27 +0000619 // Allocator main loop:
Misha Brukman2a835f92009-01-08 15:50:22 +0000620 //
Evan Chengb1290a62008-10-02 18:29:27 +0000621 // * Map current regalloc problem to a PBQP problem
622 // * Solve the PBQP problem
623 // * Map the solution back to a register allocation
624 // * Spill if necessary
Misha Brukman2a835f92009-01-08 15:50:22 +0000625 //
Evan Chengb1290a62008-10-02 18:29:27 +0000626 // This process is continued till no more spills are generated.
627
Lang Hames27601ef2008-11-16 12:12:54 +0000628 // Find the vreg intervals in need of allocation.
629 findVRegIntervalsToAlloc();
Misha Brukman2a835f92009-01-08 15:50:22 +0000630
Lang Hames20df03c2012-03-26 23:07:23 +0000631 const Function* func = mf->getFunction();
632 std::string fqn =
633 func->getParent()->getModuleIdentifier() + "." +
634 func->getName().str();
635 (void)fqn;
636
Lang Hames27601ef2008-11-16 12:12:54 +0000637 // If there are non-empty intervals allocate them using pbqp.
Lang Hameseb6c8f52010-09-18 09:07:10 +0000638 if (!vregsToAlloc.empty()) {
Evan Chengb1290a62008-10-02 18:29:27 +0000639
Lang Hames27601ef2008-11-16 12:12:54 +0000640 bool pbqpAllocComplete = false;
641 unsigned round = 0;
642
Lang Hamesab62b7e2010-10-04 12:13:07 +0000643 while (!pbqpAllocComplete) {
644 DEBUG(dbgs() << " PBQP Regalloc round " << round << ":\n");
Lang Hames27601ef2008-11-16 12:12:54 +0000645
Lang Hamesab62b7e2010-10-04 12:13:07 +0000646 std::auto_ptr<PBQPRAProblem> problem =
647 builder->build(mf, lis, loopInfo, vregsToAlloc);
Lang Hames20df03c2012-03-26 23:07:23 +0000648
649#ifndef NDEBUG
650 if (pbqpDumpGraphs) {
651 std::ostringstream rs;
652 rs << round;
653 std::string graphFileName(fqn + "." + rs.str() + ".pbqpgraph");
654 std::string tmp;
655 raw_fd_ostream os(graphFileName.c_str(), tmp);
656 DEBUG(dbgs() << "Dumping graph for round " << round << " to \""
657 << graphFileName << "\"\n");
658 problem->getGraph().dump(os);
659 }
660#endif
661
Lang Hamesab62b7e2010-10-04 12:13:07 +0000662 PBQP::Solution solution =
663 PBQP::HeuristicSolver<PBQP::Heuristics::Briggs>::solve(
664 problem->getGraph());
Lang Hames233fd9c2009-08-18 23:34:50 +0000665
Lang Hamesab62b7e2010-10-04 12:13:07 +0000666 pbqpAllocComplete = mapPBQPToRegAlloc(*problem, solution);
Lang Hames27601ef2008-11-16 12:12:54 +0000667
Lang Hamesab62b7e2010-10-04 12:13:07 +0000668 ++round;
Lang Hames27601ef2008-11-16 12:12:54 +0000669 }
Evan Chengb1290a62008-10-02 18:29:27 +0000670 }
671
Lang Hames27601ef2008-11-16 12:12:54 +0000672 // Finalise allocation, allocate empty ranges.
673 finalizeAlloc();
Evan Chengb1290a62008-10-02 18:29:27 +0000674
Lang Hamesc4bcc772010-07-20 07:41:44 +0000675 rmf->renderMachineFunction("After PBQP register allocation.", vrm);
676
Lang Hameseb6c8f52010-09-18 09:07:10 +0000677 vregsToAlloc.clear();
678 emptyIntervalVRegs.clear();
Lang Hames27601ef2008-11-16 12:12:54 +0000679
David Greene30931542010-01-05 01:25:43 +0000680 DEBUG(dbgs() << "Post alloc VirtRegMap:\n" << *vrm << "\n");
Lang Hames27601ef2008-11-16 12:12:54 +0000681
Misha Brukman2a835f92009-01-08 15:50:22 +0000682 return true;
Evan Chengb1290a62008-10-02 18:29:27 +0000683}
684
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000685FunctionPass* llvm::createPBQPRegisterAllocator(
Lang Hames8d857662011-06-17 07:09:01 +0000686 std::auto_ptr<PBQPBuilder> builder,
687 char *customPassID) {
688 return new RegAllocPBQP(builder, customPassID);
Evan Chengb1290a62008-10-02 18:29:27 +0000689}
690
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000691FunctionPass* llvm::createDefaultPBQPRegisterAllocator() {
692 if (pbqpCoalescing) {
693 return createPBQPRegisterAllocator(
694 std::auto_ptr<PBQPBuilder>(new PBQPBuilderWithCoalescing()));
695 } // else
696 return createPBQPRegisterAllocator(
697 std::auto_ptr<PBQPBuilder>(new PBQPBuilder()));
Lang Hameseb6c8f52010-09-18 09:07:10 +0000698}
Evan Chengb1290a62008-10-02 18:29:27 +0000699
700#undef DEBUG_TYPE