[RISCV] Fix mgather -> riscv.masked.strided.load combine not extending indices (...
[llvm-project.git] / llvm / lib / CodeGen / RegAllocPBQP.cpp
blobb8ee5dc0f8494b61b613c87f7a952b0e60326a08
1 //===- RegAllocPBQP.cpp ---- PBQP Register Allocator ----------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file contains a Partitioned Boolean Quadratic Programming (PBQP) based
10 // register allocator for LLVM. This allocator works by constructing a PBQP
11 // problem representing the register allocation problem under consideration,
12 // solving this using a PBQP solver, and mapping the solution back to a
13 // register assignment. If any variables are selected for spilling then spill
14 // code is inserted and the process repeated.
16 // The PBQP solver (pbqp.c) provided for this allocator uses a heuristic tuned
17 // for register allocation. For more information on PBQP for register
18 // allocation, see the following papers:
20 // (1) Hames, L. and Scholz, B. 2006. Nearly optimal register allocation with
21 // PBQP. In Proceedings of the 7th Joint Modular Languages Conference
22 // (JMLC'06). LNCS, vol. 4228. Springer, New York, NY, USA. 346-361.
24 // (2) Scholz, B., Eckstein, E. 2002. Register allocation for irregular
25 // architectures. In Proceedings of the Joint Conference on Languages,
26 // Compilers and Tools for Embedded Systems (LCTES'02), ACM Press, New York,
27 // NY, USA, 139-148.
29 //===----------------------------------------------------------------------===//
31 #include "llvm/CodeGen/RegAllocPBQP.h"
32 #include "RegisterCoalescer.h"
33 #include "llvm/ADT/ArrayRef.h"
34 #include "llvm/ADT/BitVector.h"
35 #include "llvm/ADT/DenseMap.h"
36 #include "llvm/ADT/DenseSet.h"
37 #include "llvm/ADT/STLExtras.h"
38 #include "llvm/ADT/SmallPtrSet.h"
39 #include "llvm/ADT/SmallVector.h"
40 #include "llvm/ADT/StringRef.h"
41 #include "llvm/Analysis/AliasAnalysis.h"
42 #include "llvm/CodeGen/CalcSpillWeights.h"
43 #include "llvm/CodeGen/LiveInterval.h"
44 #include "llvm/CodeGen/LiveIntervals.h"
45 #include "llvm/CodeGen/LiveRangeEdit.h"
46 #include "llvm/CodeGen/LiveStacks.h"
47 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
48 #include "llvm/CodeGen/MachineDominators.h"
49 #include "llvm/CodeGen/MachineFunction.h"
50 #include "llvm/CodeGen/MachineFunctionPass.h"
51 #include "llvm/CodeGen/MachineInstr.h"
52 #include "llvm/CodeGen/MachineLoopInfo.h"
53 #include "llvm/CodeGen/MachineRegisterInfo.h"
54 #include "llvm/CodeGen/PBQP/Graph.h"
55 #include "llvm/CodeGen/PBQP/Math.h"
56 #include "llvm/CodeGen/PBQP/Solution.h"
57 #include "llvm/CodeGen/PBQPRAConstraint.h"
58 #include "llvm/CodeGen/RegAllocRegistry.h"
59 #include "llvm/CodeGen/SlotIndexes.h"
60 #include "llvm/CodeGen/Spiller.h"
61 #include "llvm/CodeGen/TargetRegisterInfo.h"
62 #include "llvm/CodeGen/TargetSubtargetInfo.h"
63 #include "llvm/CodeGen/VirtRegMap.h"
64 #include "llvm/Config/llvm-config.h"
65 #include "llvm/IR/Function.h"
66 #include "llvm/IR/Module.h"
67 #include "llvm/MC/MCRegisterInfo.h"
68 #include "llvm/Pass.h"
69 #include "llvm/Support/CommandLine.h"
70 #include "llvm/Support/Compiler.h"
71 #include "llvm/Support/Debug.h"
72 #include "llvm/Support/FileSystem.h"
73 #include "llvm/Support/Printable.h"
74 #include "llvm/Support/raw_ostream.h"
75 #include <algorithm>
76 #include <cassert>
77 #include <cstddef>
78 #include <limits>
79 #include <map>
80 #include <memory>
81 #include <queue>
82 #include <set>
83 #include <sstream>
84 #include <string>
85 #include <system_error>
86 #include <tuple>
87 #include <utility>
88 #include <vector>
90 using namespace llvm;
92 #define DEBUG_TYPE "regalloc"
94 static RegisterRegAlloc
95 RegisterPBQPRepAlloc("pbqp", "PBQP register allocator",
96 createDefaultPBQPRegisterAllocator);
98 static cl::opt<bool>
99 PBQPCoalescing("pbqp-coalescing",
100 cl::desc("Attempt coalescing during PBQP register allocation."),
101 cl::init(false), cl::Hidden);
103 #ifndef NDEBUG
104 static cl::opt<bool>
105 PBQPDumpGraphs("pbqp-dump-graphs",
106 cl::desc("Dump graphs for each function/round in the compilation unit."),
107 cl::init(false), cl::Hidden);
108 #endif
110 namespace {
113 /// PBQP based allocators solve the register allocation problem by mapping
114 /// register allocation problems to Partitioned Boolean Quadratic
115 /// Programming problems.
116 class RegAllocPBQP : public MachineFunctionPass {
117 public:
118 static char ID;
120 /// Construct a PBQP register allocator.
121 RegAllocPBQP(char *cPassID = nullptr)
122 : MachineFunctionPass(ID), customPassID(cPassID) {
123 initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
124 initializeLiveIntervalsPass(*PassRegistry::getPassRegistry());
125 initializeLiveStacksPass(*PassRegistry::getPassRegistry());
126 initializeVirtRegMapPass(*PassRegistry::getPassRegistry());
129 /// Return the pass name.
130 StringRef getPassName() const override { return "PBQP Register Allocator"; }
132 /// PBQP analysis usage.
133 void getAnalysisUsage(AnalysisUsage &au) const override;
135 /// Perform register allocation
136 bool runOnMachineFunction(MachineFunction &MF) override;
138 MachineFunctionProperties getRequiredProperties() const override {
139 return MachineFunctionProperties().set(
140 MachineFunctionProperties::Property::NoPHIs);
143 MachineFunctionProperties getClearedProperties() const override {
144 return MachineFunctionProperties().set(
145 MachineFunctionProperties::Property::IsSSA);
148 private:
149 using RegSet = std::set<Register>;
151 char *customPassID;
153 RegSet VRegsToAlloc, EmptyIntervalVRegs;
155 /// Inst which is a def of an original reg and whose defs are already all
156 /// dead after remat is saved in DeadRemats. The deletion of such inst is
157 /// postponed till all the allocations are done, so its remat expr is
158 /// always available for the remat of all the siblings of the original reg.
159 SmallPtrSet<MachineInstr *, 32> DeadRemats;
161 /// Finds the initial set of vreg intervals to allocate.
162 void findVRegIntervalsToAlloc(const MachineFunction &MF, LiveIntervals &LIS);
164 /// Constructs an initial graph.
165 void initializeGraph(PBQPRAGraph &G, VirtRegMap &VRM, Spiller &VRegSpiller);
167 /// Spill the given VReg.
168 void spillVReg(Register VReg, SmallVectorImpl<Register> &NewIntervals,
169 MachineFunction &MF, LiveIntervals &LIS, VirtRegMap &VRM,
170 Spiller &VRegSpiller);
172 /// Given a solved PBQP problem maps this solution back to a register
173 /// assignment.
174 bool mapPBQPToRegAlloc(const PBQPRAGraph &G,
175 const PBQP::Solution &Solution,
176 VirtRegMap &VRM,
177 Spiller &VRegSpiller);
179 /// Postprocessing before final spilling. Sets basic block "live in"
180 /// variables.
181 void finalizeAlloc(MachineFunction &MF, LiveIntervals &LIS,
182 VirtRegMap &VRM) const;
184 void postOptimization(Spiller &VRegSpiller, LiveIntervals &LIS);
187 char RegAllocPBQP::ID = 0;
189 /// Set spill costs for each node in the PBQP reg-alloc graph.
190 class SpillCosts : public PBQPRAConstraint {
191 public:
192 void apply(PBQPRAGraph &G) override {
193 LiveIntervals &LIS = G.getMetadata().LIS;
195 // A minimum spill costs, so that register constraints can be set
196 // without normalization in the [0.0:MinSpillCost( interval.
197 const PBQP::PBQPNum MinSpillCost = 10.0;
199 for (auto NId : G.nodeIds()) {
200 PBQP::PBQPNum SpillCost =
201 LIS.getInterval(G.getNodeMetadata(NId).getVReg()).weight();
202 if (SpillCost == 0.0)
203 SpillCost = std::numeric_limits<PBQP::PBQPNum>::min();
204 else
205 SpillCost += MinSpillCost;
206 PBQPRAGraph::RawVector NodeCosts(G.getNodeCosts(NId));
207 NodeCosts[PBQP::RegAlloc::getSpillOptionIdx()] = SpillCost;
208 G.setNodeCosts(NId, std::move(NodeCosts));
213 /// Add interference edges between overlapping vregs.
214 class Interference : public PBQPRAConstraint {
215 private:
216 using AllowedRegVecPtr = const PBQP::RegAlloc::AllowedRegVector *;
217 using IKey = std::pair<AllowedRegVecPtr, AllowedRegVecPtr>;
218 using IMatrixCache = DenseMap<IKey, PBQPRAGraph::MatrixPtr>;
219 using DisjointAllowedRegsCache = DenseSet<IKey>;
220 using IEdgeKey = std::pair<PBQP::GraphBase::NodeId, PBQP::GraphBase::NodeId>;
221 using IEdgeCache = DenseSet<IEdgeKey>;
223 bool haveDisjointAllowedRegs(const PBQPRAGraph &G, PBQPRAGraph::NodeId NId,
224 PBQPRAGraph::NodeId MId,
225 const DisjointAllowedRegsCache &D) const {
226 const auto *NRegs = &G.getNodeMetadata(NId).getAllowedRegs();
227 const auto *MRegs = &G.getNodeMetadata(MId).getAllowedRegs();
229 if (NRegs == MRegs)
230 return false;
232 if (NRegs < MRegs)
233 return D.contains(IKey(NRegs, MRegs));
235 return D.contains(IKey(MRegs, NRegs));
238 void setDisjointAllowedRegs(const PBQPRAGraph &G, PBQPRAGraph::NodeId NId,
239 PBQPRAGraph::NodeId MId,
240 DisjointAllowedRegsCache &D) {
241 const auto *NRegs = &G.getNodeMetadata(NId).getAllowedRegs();
242 const auto *MRegs = &G.getNodeMetadata(MId).getAllowedRegs();
244 assert(NRegs != MRegs && "AllowedRegs can not be disjoint with itself");
246 if (NRegs < MRegs)
247 D.insert(IKey(NRegs, MRegs));
248 else
249 D.insert(IKey(MRegs, NRegs));
252 // Holds (Interval, CurrentSegmentID, and NodeId). The first two are required
253 // for the fast interference graph construction algorithm. The last is there
254 // to save us from looking up node ids via the VRegToNode map in the graph
255 // metadata.
256 using IntervalInfo =
257 std::tuple<LiveInterval*, size_t, PBQP::GraphBase::NodeId>;
259 static SlotIndex getStartPoint(const IntervalInfo &I) {
260 return std::get<0>(I)->segments[std::get<1>(I)].start;
263 static SlotIndex getEndPoint(const IntervalInfo &I) {
264 return std::get<0>(I)->segments[std::get<1>(I)].end;
267 static PBQP::GraphBase::NodeId getNodeId(const IntervalInfo &I) {
268 return std::get<2>(I);
271 static bool lowestStartPoint(const IntervalInfo &I1,
272 const IntervalInfo &I2) {
273 // Condition reversed because priority queue has the *highest* element at
274 // the front, rather than the lowest.
275 return getStartPoint(I1) > getStartPoint(I2);
278 static bool lowestEndPoint(const IntervalInfo &I1,
279 const IntervalInfo &I2) {
280 SlotIndex E1 = getEndPoint(I1);
281 SlotIndex E2 = getEndPoint(I2);
283 if (E1 < E2)
284 return true;
286 if (E1 > E2)
287 return false;
289 // If two intervals end at the same point, we need a way to break the tie or
290 // the set will assume they're actually equal and refuse to insert a
291 // "duplicate". Just compare the vregs - fast and guaranteed unique.
292 return std::get<0>(I1)->reg() < std::get<0>(I2)->reg();
295 static bool isAtLastSegment(const IntervalInfo &I) {
296 return std::get<1>(I) == std::get<0>(I)->size() - 1;
299 static IntervalInfo nextSegment(const IntervalInfo &I) {
300 return std::make_tuple(std::get<0>(I), std::get<1>(I) + 1, std::get<2>(I));
303 public:
304 void apply(PBQPRAGraph &G) override {
305 // The following is loosely based on the linear scan algorithm introduced in
306 // "Linear Scan Register Allocation" by Poletto and Sarkar. This version
307 // isn't linear, because the size of the active set isn't bound by the
308 // number of registers, but rather the size of the largest clique in the
309 // graph. Still, we expect this to be better than N^2.
310 LiveIntervals &LIS = G.getMetadata().LIS;
312 // Interferenc matrices are incredibly regular - they're only a function of
313 // the allowed sets, so we cache them to avoid the overhead of constructing
314 // and uniquing them.
315 IMatrixCache C;
317 // Finding an edge is expensive in the worst case (O(max_clique(G))). So
318 // cache locally edges we have already seen.
319 IEdgeCache EC;
321 // Cache known disjoint allowed registers pairs
322 DisjointAllowedRegsCache D;
324 using IntervalSet = std::set<IntervalInfo, decltype(&lowestEndPoint)>;
325 using IntervalQueue =
326 std::priority_queue<IntervalInfo, std::vector<IntervalInfo>,
327 decltype(&lowestStartPoint)>;
328 IntervalSet Active(lowestEndPoint);
329 IntervalQueue Inactive(lowestStartPoint);
331 // Start by building the inactive set.
332 for (auto NId : G.nodeIds()) {
333 Register VReg = G.getNodeMetadata(NId).getVReg();
334 LiveInterval &LI = LIS.getInterval(VReg);
335 assert(!LI.empty() && "PBQP graph contains node for empty interval");
336 Inactive.push(std::make_tuple(&LI, 0, NId));
339 while (!Inactive.empty()) {
340 // Tentatively grab the "next" interval - this choice may be overriden
341 // below.
342 IntervalInfo Cur = Inactive.top();
344 // Retire any active intervals that end before Cur starts.
345 IntervalSet::iterator RetireItr = Active.begin();
346 while (RetireItr != Active.end() &&
347 (getEndPoint(*RetireItr) <= getStartPoint(Cur))) {
348 // If this interval has subsequent segments, add the next one to the
349 // inactive list.
350 if (!isAtLastSegment(*RetireItr))
351 Inactive.push(nextSegment(*RetireItr));
353 ++RetireItr;
355 Active.erase(Active.begin(), RetireItr);
357 // One of the newly retired segments may actually start before the
358 // Cur segment, so re-grab the front of the inactive list.
359 Cur = Inactive.top();
360 Inactive.pop();
362 // At this point we know that Cur overlaps all active intervals. Add the
363 // interference edges.
364 PBQP::GraphBase::NodeId NId = getNodeId(Cur);
365 for (const auto &A : Active) {
366 PBQP::GraphBase::NodeId MId = getNodeId(A);
368 // Do not add an edge when the nodes' allowed registers do not
369 // intersect: there is obviously no interference.
370 if (haveDisjointAllowedRegs(G, NId, MId, D))
371 continue;
373 // Check that we haven't already added this edge
374 IEdgeKey EK(std::min(NId, MId), std::max(NId, MId));
375 if (EC.count(EK))
376 continue;
378 // This is a new edge - add it to the graph.
379 if (!createInterferenceEdge(G, NId, MId, C))
380 setDisjointAllowedRegs(G, NId, MId, D);
381 else
382 EC.insert(EK);
385 // Finally, add Cur to the Active set.
386 Active.insert(Cur);
390 private:
391 // Create an Interference edge and add it to the graph, unless it is
392 // a null matrix, meaning the nodes' allowed registers do not have any
393 // interference. This case occurs frequently between integer and floating
394 // point registers for example.
395 // return true iff both nodes interferes.
396 bool createInterferenceEdge(PBQPRAGraph &G,
397 PBQPRAGraph::NodeId NId, PBQPRAGraph::NodeId MId,
398 IMatrixCache &C) {
399 const TargetRegisterInfo &TRI =
400 *G.getMetadata().MF.getSubtarget().getRegisterInfo();
401 const auto &NRegs = G.getNodeMetadata(NId).getAllowedRegs();
402 const auto &MRegs = G.getNodeMetadata(MId).getAllowedRegs();
404 // Try looking the edge costs up in the IMatrixCache first.
405 IKey K(&NRegs, &MRegs);
406 IMatrixCache::iterator I = C.find(K);
407 if (I != C.end()) {
408 G.addEdgeBypassingCostAllocator(NId, MId, I->second);
409 return true;
412 PBQPRAGraph::RawMatrix M(NRegs.size() + 1, MRegs.size() + 1, 0);
413 bool NodesInterfere = false;
414 for (unsigned I = 0; I != NRegs.size(); ++I) {
415 MCRegister PRegN = NRegs[I];
416 for (unsigned J = 0; J != MRegs.size(); ++J) {
417 MCRegister PRegM = MRegs[J];
418 if (TRI.regsOverlap(PRegN, PRegM)) {
419 M[I + 1][J + 1] = std::numeric_limits<PBQP::PBQPNum>::infinity();
420 NodesInterfere = true;
425 if (!NodesInterfere)
426 return false;
428 PBQPRAGraph::EdgeId EId = G.addEdge(NId, MId, std::move(M));
429 C[K] = G.getEdgeCostsPtr(EId);
431 return true;
435 class Coalescing : public PBQPRAConstraint {
436 public:
437 void apply(PBQPRAGraph &G) override {
438 MachineFunction &MF = G.getMetadata().MF;
439 MachineBlockFrequencyInfo &MBFI = G.getMetadata().MBFI;
440 CoalescerPair CP(*MF.getSubtarget().getRegisterInfo());
442 // Scan the machine function and add a coalescing cost whenever CoalescerPair
443 // gives the Ok.
444 for (const auto &MBB : MF) {
445 for (const auto &MI : MBB) {
446 // Skip not-coalescable or already coalesced copies.
447 if (!CP.setRegisters(&MI) || CP.getSrcReg() == CP.getDstReg())
448 continue;
450 Register DstReg = CP.getDstReg();
451 Register SrcReg = CP.getSrcReg();
453 PBQP::PBQPNum CBenefit = MBFI.getBlockFreqRelativeToEntryBlock(&MBB);
455 if (CP.isPhys()) {
456 if (!MF.getRegInfo().isAllocatable(DstReg))
457 continue;
459 PBQPRAGraph::NodeId NId = G.getMetadata().getNodeIdForVReg(SrcReg);
461 const PBQPRAGraph::NodeMetadata::AllowedRegVector &Allowed =
462 G.getNodeMetadata(NId).getAllowedRegs();
464 unsigned PRegOpt = 0;
465 while (PRegOpt < Allowed.size() && Allowed[PRegOpt].id() != DstReg)
466 ++PRegOpt;
468 if (PRegOpt < Allowed.size()) {
469 PBQPRAGraph::RawVector NewCosts(G.getNodeCosts(NId));
470 NewCosts[PRegOpt + 1] -= CBenefit;
471 G.setNodeCosts(NId, std::move(NewCosts));
473 } else {
474 PBQPRAGraph::NodeId N1Id = G.getMetadata().getNodeIdForVReg(DstReg);
475 PBQPRAGraph::NodeId N2Id = G.getMetadata().getNodeIdForVReg(SrcReg);
476 const PBQPRAGraph::NodeMetadata::AllowedRegVector *Allowed1 =
477 &G.getNodeMetadata(N1Id).getAllowedRegs();
478 const PBQPRAGraph::NodeMetadata::AllowedRegVector *Allowed2 =
479 &G.getNodeMetadata(N2Id).getAllowedRegs();
481 PBQPRAGraph::EdgeId EId = G.findEdge(N1Id, N2Id);
482 if (EId == G.invalidEdgeId()) {
483 PBQPRAGraph::RawMatrix Costs(Allowed1->size() + 1,
484 Allowed2->size() + 1, 0);
485 addVirtRegCoalesce(Costs, *Allowed1, *Allowed2, CBenefit);
486 G.addEdge(N1Id, N2Id, std::move(Costs));
487 } else {
488 if (G.getEdgeNode1Id(EId) == N2Id) {
489 std::swap(N1Id, N2Id);
490 std::swap(Allowed1, Allowed2);
492 PBQPRAGraph::RawMatrix Costs(G.getEdgeCosts(EId));
493 addVirtRegCoalesce(Costs, *Allowed1, *Allowed2, CBenefit);
494 G.updateEdgeCosts(EId, std::move(Costs));
501 private:
502 void addVirtRegCoalesce(
503 PBQPRAGraph::RawMatrix &CostMat,
504 const PBQPRAGraph::NodeMetadata::AllowedRegVector &Allowed1,
505 const PBQPRAGraph::NodeMetadata::AllowedRegVector &Allowed2,
506 PBQP::PBQPNum Benefit) {
507 assert(CostMat.getRows() == Allowed1.size() + 1 && "Size mismatch.");
508 assert(CostMat.getCols() == Allowed2.size() + 1 && "Size mismatch.");
509 for (unsigned I = 0; I != Allowed1.size(); ++I) {
510 MCRegister PReg1 = Allowed1[I];
511 for (unsigned J = 0; J != Allowed2.size(); ++J) {
512 MCRegister PReg2 = Allowed2[J];
513 if (PReg1 == PReg2)
514 CostMat[I + 1][J + 1] -= Benefit;
520 /// PBQP-specific implementation of weight normalization.
521 class PBQPVirtRegAuxInfo final : public VirtRegAuxInfo {
522 float normalize(float UseDefFreq, unsigned Size, unsigned NumInstr) override {
523 // All intervals have a spill weight that is mostly proportional to the
524 // number of uses, with uses in loops having a bigger weight.
525 return NumInstr * VirtRegAuxInfo::normalize(UseDefFreq, Size, 1);
528 public:
529 PBQPVirtRegAuxInfo(MachineFunction &MF, LiveIntervals &LIS, VirtRegMap &VRM,
530 const MachineLoopInfo &Loops,
531 const MachineBlockFrequencyInfo &MBFI)
532 : VirtRegAuxInfo(MF, LIS, VRM, Loops, MBFI) {}
534 } // end anonymous namespace
536 // Out-of-line destructor/anchor for PBQPRAConstraint.
537 PBQPRAConstraint::~PBQPRAConstraint() = default;
539 void PBQPRAConstraint::anchor() {}
541 void PBQPRAConstraintList::anchor() {}
543 void RegAllocPBQP::getAnalysisUsage(AnalysisUsage &au) const {
544 au.setPreservesCFG();
545 au.addRequired<AAResultsWrapperPass>();
546 au.addPreserved<AAResultsWrapperPass>();
547 au.addRequired<SlotIndexes>();
548 au.addPreserved<SlotIndexes>();
549 au.addRequired<LiveIntervals>();
550 au.addPreserved<LiveIntervals>();
551 //au.addRequiredID(SplitCriticalEdgesID);
552 if (customPassID)
553 au.addRequiredID(*customPassID);
554 au.addRequired<LiveStacks>();
555 au.addPreserved<LiveStacks>();
556 au.addRequired<MachineBlockFrequencyInfo>();
557 au.addPreserved<MachineBlockFrequencyInfo>();
558 au.addRequired<MachineLoopInfo>();
559 au.addPreserved<MachineLoopInfo>();
560 au.addRequired<MachineDominatorTree>();
561 au.addPreserved<MachineDominatorTree>();
562 au.addRequired<VirtRegMap>();
563 au.addPreserved<VirtRegMap>();
564 MachineFunctionPass::getAnalysisUsage(au);
567 void RegAllocPBQP::findVRegIntervalsToAlloc(const MachineFunction &MF,
568 LiveIntervals &LIS) {
569 const MachineRegisterInfo &MRI = MF.getRegInfo();
571 // Iterate over all live ranges.
572 for (unsigned I = 0, E = MRI.getNumVirtRegs(); I != E; ++I) {
573 Register Reg = Register::index2VirtReg(I);
574 if (MRI.reg_nodbg_empty(Reg))
575 continue;
576 VRegsToAlloc.insert(Reg);
580 static bool isACalleeSavedRegister(MCRegister Reg,
581 const TargetRegisterInfo &TRI,
582 const MachineFunction &MF) {
583 const MCPhysReg *CSR = MF.getRegInfo().getCalleeSavedRegs();
584 for (unsigned i = 0; CSR[i] != 0; ++i)
585 if (TRI.regsOverlap(Reg, CSR[i]))
586 return true;
587 return false;
590 void RegAllocPBQP::initializeGraph(PBQPRAGraph &G, VirtRegMap &VRM,
591 Spiller &VRegSpiller) {
592 MachineFunction &MF = G.getMetadata().MF;
594 LiveIntervals &LIS = G.getMetadata().LIS;
595 const MachineRegisterInfo &MRI = G.getMetadata().MF.getRegInfo();
596 const TargetRegisterInfo &TRI =
597 *G.getMetadata().MF.getSubtarget().getRegisterInfo();
599 std::vector<Register> Worklist(VRegsToAlloc.begin(), VRegsToAlloc.end());
601 std::map<Register, std::vector<MCRegister>> VRegAllowedMap;
603 while (!Worklist.empty()) {
604 Register VReg = Worklist.back();
605 Worklist.pop_back();
607 LiveInterval &VRegLI = LIS.getInterval(VReg);
609 // If this is an empty interval move it to the EmptyIntervalVRegs set then
610 // continue.
611 if (VRegLI.empty()) {
612 EmptyIntervalVRegs.insert(VRegLI.reg());
613 VRegsToAlloc.erase(VRegLI.reg());
614 continue;
617 const TargetRegisterClass *TRC = MRI.getRegClass(VReg);
619 // Record any overlaps with regmask operands.
620 BitVector RegMaskOverlaps;
621 LIS.checkRegMaskInterference(VRegLI, RegMaskOverlaps);
623 // Compute an initial allowed set for the current vreg.
624 std::vector<MCRegister> VRegAllowed;
625 ArrayRef<MCPhysReg> RawPRegOrder = TRC->getRawAllocationOrder(MF);
626 for (MCPhysReg R : RawPRegOrder) {
627 MCRegister PReg(R);
628 if (MRI.isReserved(PReg))
629 continue;
631 // vregLI crosses a regmask operand that clobbers preg.
632 if (!RegMaskOverlaps.empty() && !RegMaskOverlaps.test(PReg))
633 continue;
635 // vregLI overlaps fixed regunit interference.
636 bool Interference = false;
637 for (MCRegUnit Unit : TRI.regunits(PReg)) {
638 if (VRegLI.overlaps(LIS.getRegUnit(Unit))) {
639 Interference = true;
640 break;
643 if (Interference)
644 continue;
646 // preg is usable for this virtual register.
647 VRegAllowed.push_back(PReg);
650 // Check for vregs that have no allowed registers. These should be
651 // pre-spilled and the new vregs added to the worklist.
652 if (VRegAllowed.empty()) {
653 SmallVector<Register, 8> NewVRegs;
654 spillVReg(VReg, NewVRegs, MF, LIS, VRM, VRegSpiller);
655 llvm::append_range(Worklist, NewVRegs);
656 continue;
659 VRegAllowedMap[VReg.id()] = std::move(VRegAllowed);
662 for (auto &KV : VRegAllowedMap) {
663 auto VReg = KV.first;
665 // Move empty intervals to the EmptyIntervalVReg set.
666 if (LIS.getInterval(VReg).empty()) {
667 EmptyIntervalVRegs.insert(VReg);
668 VRegsToAlloc.erase(VReg);
669 continue;
672 auto &VRegAllowed = KV.second;
674 PBQPRAGraph::RawVector NodeCosts(VRegAllowed.size() + 1, 0);
676 // Tweak cost of callee saved registers, as using then force spilling and
677 // restoring them. This would only happen in the prologue / epilogue though.
678 for (unsigned i = 0; i != VRegAllowed.size(); ++i)
679 if (isACalleeSavedRegister(VRegAllowed[i], TRI, MF))
680 NodeCosts[1 + i] += 1.0;
682 PBQPRAGraph::NodeId NId = G.addNode(std::move(NodeCosts));
683 G.getNodeMetadata(NId).setVReg(VReg);
684 G.getNodeMetadata(NId).setAllowedRegs(
685 G.getMetadata().getAllowedRegs(std::move(VRegAllowed)));
686 G.getMetadata().setNodeIdForVReg(VReg, NId);
690 void RegAllocPBQP::spillVReg(Register VReg,
691 SmallVectorImpl<Register> &NewIntervals,
692 MachineFunction &MF, LiveIntervals &LIS,
693 VirtRegMap &VRM, Spiller &VRegSpiller) {
694 VRegsToAlloc.erase(VReg);
695 LiveRangeEdit LRE(&LIS.getInterval(VReg), NewIntervals, MF, LIS, &VRM,
696 nullptr, &DeadRemats);
697 VRegSpiller.spill(LRE);
699 const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
700 (void)TRI;
701 LLVM_DEBUG(dbgs() << "VREG " << printReg(VReg, &TRI) << " -> SPILLED (Cost: "
702 << LRE.getParent().weight() << ", New vregs: ");
704 // Copy any newly inserted live intervals into the list of regs to
705 // allocate.
706 for (const Register &R : LRE) {
707 const LiveInterval &LI = LIS.getInterval(R);
708 assert(!LI.empty() && "Empty spill range.");
709 LLVM_DEBUG(dbgs() << printReg(LI.reg(), &TRI) << " ");
710 VRegsToAlloc.insert(LI.reg());
713 LLVM_DEBUG(dbgs() << ")\n");
716 bool RegAllocPBQP::mapPBQPToRegAlloc(const PBQPRAGraph &G,
717 const PBQP::Solution &Solution,
718 VirtRegMap &VRM,
719 Spiller &VRegSpiller) {
720 MachineFunction &MF = G.getMetadata().MF;
721 LiveIntervals &LIS = G.getMetadata().LIS;
722 const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
723 (void)TRI;
725 // Set to true if we have any spills
726 bool AnotherRoundNeeded = false;
728 // Clear the existing allocation.
729 VRM.clearAllVirt();
731 // Iterate over the nodes mapping the PBQP solution to a register
732 // assignment.
733 for (auto NId : G.nodeIds()) {
734 Register VReg = G.getNodeMetadata(NId).getVReg();
735 unsigned AllocOpt = Solution.getSelection(NId);
737 if (AllocOpt != PBQP::RegAlloc::getSpillOptionIdx()) {
738 MCRegister PReg = G.getNodeMetadata(NId).getAllowedRegs()[AllocOpt - 1];
739 LLVM_DEBUG(dbgs() << "VREG " << printReg(VReg, &TRI) << " -> "
740 << TRI.getName(PReg) << "\n");
741 assert(PReg != 0 && "Invalid preg selected.");
742 VRM.assignVirt2Phys(VReg, PReg);
743 } else {
744 // Spill VReg. If this introduces new intervals we'll need another round
745 // of allocation.
746 SmallVector<Register, 8> NewVRegs;
747 spillVReg(VReg, NewVRegs, MF, LIS, VRM, VRegSpiller);
748 AnotherRoundNeeded |= !NewVRegs.empty();
752 return !AnotherRoundNeeded;
755 void RegAllocPBQP::finalizeAlloc(MachineFunction &MF,
756 LiveIntervals &LIS,
757 VirtRegMap &VRM) const {
758 MachineRegisterInfo &MRI = MF.getRegInfo();
760 // First allocate registers for the empty intervals.
761 for (const Register &R : EmptyIntervalVRegs) {
762 LiveInterval &LI = LIS.getInterval(R);
764 Register PReg = MRI.getSimpleHint(LI.reg());
766 if (PReg == 0) {
767 const TargetRegisterClass &RC = *MRI.getRegClass(LI.reg());
768 const ArrayRef<MCPhysReg> RawPRegOrder = RC.getRawAllocationOrder(MF);
769 for (MCRegister CandidateReg : RawPRegOrder) {
770 if (!VRM.getRegInfo().isReserved(CandidateReg)) {
771 PReg = CandidateReg;
772 break;
775 assert(PReg &&
776 "No un-reserved physical registers in this register class");
779 VRM.assignVirt2Phys(LI.reg(), PReg);
783 void RegAllocPBQP::postOptimization(Spiller &VRegSpiller, LiveIntervals &LIS) {
784 VRegSpiller.postOptimization();
785 /// Remove dead defs because of rematerialization.
786 for (auto *DeadInst : DeadRemats) {
787 LIS.RemoveMachineInstrFromMaps(*DeadInst);
788 DeadInst->eraseFromParent();
790 DeadRemats.clear();
793 bool RegAllocPBQP::runOnMachineFunction(MachineFunction &MF) {
794 LiveIntervals &LIS = getAnalysis<LiveIntervals>();
795 MachineBlockFrequencyInfo &MBFI =
796 getAnalysis<MachineBlockFrequencyInfo>();
798 VirtRegMap &VRM = getAnalysis<VirtRegMap>();
800 PBQPVirtRegAuxInfo VRAI(MF, LIS, VRM, getAnalysis<MachineLoopInfo>(), MBFI);
801 VRAI.calculateSpillWeightsAndHints();
803 // FIXME: we create DefaultVRAI here to match existing behavior pre-passing
804 // the VRAI through the spiller to the live range editor. However, it probably
805 // makes more sense to pass the PBQP VRAI. The existing behavior had
806 // LiveRangeEdit make its own VirtRegAuxInfo object.
807 VirtRegAuxInfo DefaultVRAI(MF, LIS, VRM, getAnalysis<MachineLoopInfo>(),
808 MBFI);
809 std::unique_ptr<Spiller> VRegSpiller(
810 createInlineSpiller(*this, MF, VRM, DefaultVRAI));
812 MF.getRegInfo().freezeReservedRegs(MF);
814 LLVM_DEBUG(dbgs() << "PBQP Register Allocating for " << MF.getName() << "\n");
816 // Allocator main loop:
818 // * Map current regalloc problem to a PBQP problem
819 // * Solve the PBQP problem
820 // * Map the solution back to a register allocation
821 // * Spill if necessary
823 // This process is continued till no more spills are generated.
825 // Find the vreg intervals in need of allocation.
826 findVRegIntervalsToAlloc(MF, LIS);
828 #ifndef NDEBUG
829 const Function &F = MF.getFunction();
830 std::string FullyQualifiedName =
831 F.getParent()->getModuleIdentifier() + "." + F.getName().str();
832 #endif
834 // If there are non-empty intervals allocate them using pbqp.
835 if (!VRegsToAlloc.empty()) {
836 const TargetSubtargetInfo &Subtarget = MF.getSubtarget();
837 std::unique_ptr<PBQPRAConstraintList> ConstraintsRoot =
838 std::make_unique<PBQPRAConstraintList>();
839 ConstraintsRoot->addConstraint(std::make_unique<SpillCosts>());
840 ConstraintsRoot->addConstraint(std::make_unique<Interference>());
841 if (PBQPCoalescing)
842 ConstraintsRoot->addConstraint(std::make_unique<Coalescing>());
843 ConstraintsRoot->addConstraint(Subtarget.getCustomPBQPConstraints());
845 bool PBQPAllocComplete = false;
846 unsigned Round = 0;
848 while (!PBQPAllocComplete) {
849 LLVM_DEBUG(dbgs() << " PBQP Regalloc round " << Round << ":\n");
850 (void) Round;
852 PBQPRAGraph G(PBQPRAGraph::GraphMetadata(MF, LIS, MBFI));
853 initializeGraph(G, VRM, *VRegSpiller);
854 ConstraintsRoot->apply(G);
856 #ifndef NDEBUG
857 if (PBQPDumpGraphs) {
858 std::ostringstream RS;
859 RS << Round;
860 std::string GraphFileName = FullyQualifiedName + "." + RS.str() +
861 ".pbqpgraph";
862 std::error_code EC;
863 raw_fd_ostream OS(GraphFileName, EC, sys::fs::OF_TextWithCRLF);
864 LLVM_DEBUG(dbgs() << "Dumping graph for round " << Round << " to \""
865 << GraphFileName << "\"\n");
866 G.dump(OS);
868 #endif
870 PBQP::Solution Solution = PBQP::RegAlloc::solve(G);
871 PBQPAllocComplete = mapPBQPToRegAlloc(G, Solution, VRM, *VRegSpiller);
872 ++Round;
876 // Finalise allocation, allocate empty ranges.
877 finalizeAlloc(MF, LIS, VRM);
878 postOptimization(*VRegSpiller, LIS);
879 VRegsToAlloc.clear();
880 EmptyIntervalVRegs.clear();
882 LLVM_DEBUG(dbgs() << "Post alloc VirtRegMap:\n" << VRM << "\n");
884 return true;
887 /// Create Printable object for node and register info.
888 static Printable PrintNodeInfo(PBQP::RegAlloc::PBQPRAGraph::NodeId NId,
889 const PBQP::RegAlloc::PBQPRAGraph &G) {
890 return Printable([NId, &G](raw_ostream &OS) {
891 const MachineRegisterInfo &MRI = G.getMetadata().MF.getRegInfo();
892 const TargetRegisterInfo *TRI = MRI.getTargetRegisterInfo();
893 Register VReg = G.getNodeMetadata(NId).getVReg();
894 const char *RegClassName = TRI->getRegClassName(MRI.getRegClass(VReg));
895 OS << NId << " (" << RegClassName << ':' << printReg(VReg, TRI) << ')';
899 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
900 LLVM_DUMP_METHOD void PBQP::RegAlloc::PBQPRAGraph::dump(raw_ostream &OS) const {
901 for (auto NId : nodeIds()) {
902 const Vector &Costs = getNodeCosts(NId);
903 assert(Costs.getLength() != 0 && "Empty vector in graph.");
904 OS << PrintNodeInfo(NId, *this) << ": " << Costs << '\n';
906 OS << '\n';
908 for (auto EId : edgeIds()) {
909 NodeId N1Id = getEdgeNode1Id(EId);
910 NodeId N2Id = getEdgeNode2Id(EId);
911 assert(N1Id != N2Id && "PBQP graphs should not have self-edges.");
912 const Matrix &M = getEdgeCosts(EId);
913 assert(M.getRows() != 0 && "No rows in matrix.");
914 assert(M.getCols() != 0 && "No cols in matrix.");
915 OS << PrintNodeInfo(N1Id, *this) << ' ' << M.getRows() << " rows / ";
916 OS << PrintNodeInfo(N2Id, *this) << ' ' << M.getCols() << " cols:\n";
917 OS << M << '\n';
921 LLVM_DUMP_METHOD void PBQP::RegAlloc::PBQPRAGraph::dump() const {
922 dump(dbgs());
924 #endif
926 void PBQP::RegAlloc::PBQPRAGraph::printDot(raw_ostream &OS) const {
927 OS << "graph {\n";
928 for (auto NId : nodeIds()) {
929 OS << " node" << NId << " [ label=\""
930 << PrintNodeInfo(NId, *this) << "\\n"
931 << getNodeCosts(NId) << "\" ]\n";
934 OS << " edge [ len=" << nodeIds().size() << " ]\n";
935 for (auto EId : edgeIds()) {
936 OS << " node" << getEdgeNode1Id(EId)
937 << " -- node" << getEdgeNode2Id(EId)
938 << " [ label=\"";
939 const Matrix &EdgeCosts = getEdgeCosts(EId);
940 for (unsigned i = 0; i < EdgeCosts.getRows(); ++i) {
941 OS << EdgeCosts.getRowAsVector(i) << "\\n";
943 OS << "\" ]\n";
945 OS << "}\n";
948 FunctionPass *llvm::createPBQPRegisterAllocator(char *customPassID) {
949 return new RegAllocPBQP(customPassID);
952 FunctionPass* llvm::createDefaultPBQPRegisterAllocator() {
953 return createPBQPRegisterAllocator();