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. |
15 | // |
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: |
19 | // |
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. |
23 | // |
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. |
28 | // |
29 | //===----------------------------------------------------------------------===// |
30 | |
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> |
89 | |
90 | using namespace llvm; |
91 | |
92 | #define DEBUG_TYPE "regalloc" |
93 | |
94 | static RegisterRegAlloc |
95 | RegisterPBQPRepAlloc("pbqp" , "PBQP register allocator" , |
96 | createDefaultPBQPRegisterAllocator); |
97 | |
98 | static cl::opt<bool> |
99 | PBQPCoalescing("pbqp-coalescing" , |
100 | cl::desc("Attempt coalescing during PBQP register allocation." ), |
101 | cl::init(Val: false), cl::Hidden); |
102 | |
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 |
109 | |
110 | namespace { |
111 | |
112 | /// |
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; |
119 | |
120 | /// Construct a PBQP register allocator. |
121 | RegAllocPBQP(char *cPassID = nullptr) |
122 | : MachineFunctionPass(ID), customPassID(cPassID) { |
123 | initializeSlotIndexesWrapperPassPass(*PassRegistry::getPassRegistry()); |
124 | initializeLiveIntervalsWrapperPassPass(*PassRegistry::getPassRegistry()); |
125 | initializeLiveStacksPass(*PassRegistry::getPassRegistry()); |
126 | initializeVirtRegMapPass(*PassRegistry::getPassRegistry()); |
127 | } |
128 | |
129 | /// Return the pass name. |
130 | StringRef getPassName() const override { return "PBQP Register Allocator" ; } |
131 | |
132 | /// PBQP analysis usage. |
133 | void getAnalysisUsage(AnalysisUsage &au) const override; |
134 | |
135 | /// Perform register allocation |
136 | bool runOnMachineFunction(MachineFunction &MF) override; |
137 | |
138 | MachineFunctionProperties getRequiredProperties() const override { |
139 | return MachineFunctionProperties().set( |
140 | MachineFunctionProperties::Property::NoPHIs); |
141 | } |
142 | |
143 | MachineFunctionProperties getClearedProperties() const override { |
144 | return MachineFunctionProperties().set( |
145 | MachineFunctionProperties::Property::IsSSA); |
146 | } |
147 | |
148 | private: |
149 | using RegSet = std::set<Register>; |
150 | |
151 | char *customPassID; |
152 | |
153 | RegSet VRegsToAlloc, EmptyIntervalVRegs; |
154 | |
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; |
160 | |
161 | /// Finds the initial set of vreg intervals to allocate. |
162 | void findVRegIntervalsToAlloc(const MachineFunction &MF, LiveIntervals &LIS); |
163 | |
164 | /// Constructs an initial graph. |
165 | void initializeGraph(PBQPRAGraph &G, VirtRegMap &VRM, Spiller &VRegSpiller); |
166 | |
167 | /// Spill the given VReg. |
168 | void spillVReg(Register VReg, SmallVectorImpl<Register> &NewIntervals, |
169 | MachineFunction &MF, LiveIntervals &LIS, VirtRegMap &VRM, |
170 | Spiller &VRegSpiller); |
171 | |
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); |
178 | |
179 | /// Postprocessing before final spilling. Sets basic block "live in" |
180 | /// variables. |
181 | void finalizeAlloc(MachineFunction &MF, LiveIntervals &LIS, |
182 | VirtRegMap &VRM) const; |
183 | |
184 | void postOptimization(Spiller &VRegSpiller, LiveIntervals &LIS); |
185 | }; |
186 | |
187 | char RegAllocPBQP::ID = 0; |
188 | |
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; |
194 | |
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; |
198 | |
199 | for (auto NId : G.nodeIds()) { |
200 | PBQP::PBQPNum SpillCost = |
201 | LIS.getInterval(Reg: 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, Costs: std::move(NodeCosts)); |
209 | } |
210 | } |
211 | }; |
212 | |
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>; |
222 | |
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(NId: MId).getAllowedRegs(); |
228 | |
229 | if (NRegs == MRegs) |
230 | return false; |
231 | |
232 | if (NRegs < MRegs) |
233 | return D.contains(V: IKey(NRegs, MRegs)); |
234 | |
235 | return D.contains(V: IKey(MRegs, NRegs)); |
236 | } |
237 | |
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(NId: MId).getAllowedRegs(); |
243 | |
244 | assert(NRegs != MRegs && "AllowedRegs can not be disjoint with itself" ); |
245 | |
246 | if (NRegs < MRegs) |
247 | D.insert(V: IKey(NRegs, MRegs)); |
248 | else |
249 | D.insert(V: IKey(MRegs, NRegs)); |
250 | } |
251 | |
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>; |
258 | |
259 | static SlotIndex getStartPoint(const IntervalInfo &I) { |
260 | return std::get<0>(t: I)->segments[std::get<1>(t: I)].start; |
261 | } |
262 | |
263 | static SlotIndex getEndPoint(const IntervalInfo &I) { |
264 | return std::get<0>(t: I)->segments[std::get<1>(t: I)].end; |
265 | } |
266 | |
267 | static PBQP::GraphBase::NodeId getNodeId(const IntervalInfo &I) { |
268 | return std::get<2>(t: I); |
269 | } |
270 | |
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(I: I1) > getStartPoint(I: I2); |
276 | } |
277 | |
278 | static bool lowestEndPoint(const IntervalInfo &I1, |
279 | const IntervalInfo &I2) { |
280 | SlotIndex E1 = getEndPoint(I: I1); |
281 | SlotIndex E2 = getEndPoint(I: I2); |
282 | |
283 | if (E1 < E2) |
284 | return true; |
285 | |
286 | if (E1 > E2) |
287 | return false; |
288 | |
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>(t: I1)->reg() < std::get<0>(t: I2)->reg(); |
293 | } |
294 | |
295 | static bool isAtLastSegment(const IntervalInfo &I) { |
296 | return std::get<1>(t: I) == std::get<0>(t: I)->size() - 1; |
297 | } |
298 | |
299 | static IntervalInfo nextSegment(const IntervalInfo &I) { |
300 | return std::make_tuple(args: std::get<0>(t: I), args: std::get<1>(t: I) + 1, args: std::get<2>(t: I)); |
301 | } |
302 | |
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; |
311 | |
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; |
316 | |
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; |
320 | |
321 | // Cache known disjoint allowed registers pairs |
322 | DisjointAllowedRegsCache D; |
323 | |
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); |
330 | |
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(Reg: VReg); |
335 | assert(!LI.empty() && "PBQP graph contains node for empty interval" ); |
336 | Inactive.push(x: std::make_tuple(args: &LI, args: 0, args&: NId)); |
337 | } |
338 | |
339 | while (!Inactive.empty()) { |
340 | // Tentatively grab the "next" interval - this choice may be overriden |
341 | // below. |
342 | IntervalInfo Cur = Inactive.top(); |
343 | |
344 | // Retire any active intervals that end before Cur starts. |
345 | IntervalSet::iterator RetireItr = Active.begin(); |
346 | while (RetireItr != Active.end() && |
347 | (getEndPoint(I: *RetireItr) <= getStartPoint(I: Cur))) { |
348 | // If this interval has subsequent segments, add the next one to the |
349 | // inactive list. |
350 | if (!isAtLastSegment(I: *RetireItr)) |
351 | Inactive.push(x: nextSegment(I: *RetireItr)); |
352 | |
353 | ++RetireItr; |
354 | } |
355 | Active.erase(first: Active.begin(), last: RetireItr); |
356 | |
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(); |
361 | |
362 | // At this point we know that Cur overlaps all active intervals. Add the |
363 | // interference edges. |
364 | PBQP::GraphBase::NodeId NId = getNodeId(I: Cur); |
365 | for (const auto &A : Active) { |
366 | PBQP::GraphBase::NodeId MId = getNodeId(I: A); |
367 | |
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; |
372 | |
373 | // Check that we haven't already added this edge |
374 | IEdgeKey EK(std::min(a: NId, b: MId), std::max(a: NId, b: MId)); |
375 | if (EC.count(V: EK)) |
376 | continue; |
377 | |
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(V: EK); |
383 | } |
384 | |
385 | // Finally, add Cur to the Active set. |
386 | Active.insert(x: Cur); |
387 | } |
388 | } |
389 | |
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(NId: MId).getAllowedRegs(); |
403 | |
404 | // Try looking the edge costs up in the IMatrixCache first. |
405 | IKey K(&NRegs, &MRegs); |
406 | IMatrixCache::iterator I = C.find(Val: K); |
407 | if (I != C.end()) { |
408 | G.addEdgeBypassingCostAllocator(N1Id: NId, N2Id: MId, Costs: I->second); |
409 | return true; |
410 | } |
411 | |
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(RegA: PRegN, RegB: PRegM)) { |
419 | M[I + 1][J + 1] = std::numeric_limits<PBQP::PBQPNum>::infinity(); |
420 | NodesInterfere = true; |
421 | } |
422 | } |
423 | } |
424 | |
425 | if (!NodesInterfere) |
426 | return false; |
427 | |
428 | PBQPRAGraph::EdgeId EId = G.addEdge(N1Id: NId, N2Id: MId, Costs: std::move(M)); |
429 | C[K] = G.getEdgeCostsPtr(EId); |
430 | |
431 | return true; |
432 | } |
433 | }; |
434 | |
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()); |
441 | |
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; |
449 | |
450 | Register DstReg = CP.getDstReg(); |
451 | Register SrcReg = CP.getSrcReg(); |
452 | |
453 | PBQP::PBQPNum CBenefit = MBFI.getBlockFreqRelativeToEntryBlock(MBB: &MBB); |
454 | |
455 | if (CP.isPhys()) { |
456 | if (!MF.getRegInfo().isAllocatable(PhysReg: DstReg)) |
457 | continue; |
458 | |
459 | PBQPRAGraph::NodeId NId = G.getMetadata().getNodeIdForVReg(VReg: SrcReg); |
460 | |
461 | const PBQPRAGraph::NodeMetadata::AllowedRegVector &Allowed = |
462 | G.getNodeMetadata(NId).getAllowedRegs(); |
463 | |
464 | unsigned PRegOpt = 0; |
465 | while (PRegOpt < Allowed.size() && Allowed[PRegOpt].id() != DstReg) |
466 | ++PRegOpt; |
467 | |
468 | if (PRegOpt < Allowed.size()) { |
469 | PBQPRAGraph::RawVector NewCosts(G.getNodeCosts(NId)); |
470 | NewCosts[PRegOpt + 1] -= CBenefit; |
471 | G.setNodeCosts(NId, Costs: std::move(NewCosts)); |
472 | } |
473 | } else { |
474 | PBQPRAGraph::NodeId N1Id = G.getMetadata().getNodeIdForVReg(VReg: DstReg); |
475 | PBQPRAGraph::NodeId N2Id = G.getMetadata().getNodeIdForVReg(VReg: SrcReg); |
476 | const PBQPRAGraph::NodeMetadata::AllowedRegVector *Allowed1 = |
477 | &G.getNodeMetadata(NId: N1Id).getAllowedRegs(); |
478 | const PBQPRAGraph::NodeMetadata::AllowedRegVector *Allowed2 = |
479 | &G.getNodeMetadata(NId: N2Id).getAllowedRegs(); |
480 | |
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(CostMat&: Costs, Allowed1: *Allowed1, Allowed2: *Allowed2, Benefit: CBenefit); |
486 | G.addEdge(N1Id, N2Id, Costs: std::move(Costs)); |
487 | } else { |
488 | if (G.getEdgeNode1Id(EId) == N2Id) { |
489 | std::swap(a&: N1Id, b&: N2Id); |
490 | std::swap(a&: Allowed1, b&: Allowed2); |
491 | } |
492 | PBQPRAGraph::RawMatrix Costs(G.getEdgeCosts(EId)); |
493 | addVirtRegCoalesce(CostMat&: Costs, Allowed1: *Allowed1, Allowed2: *Allowed2, Benefit: CBenefit); |
494 | G.updateEdgeCosts(EId, Costs: std::move(Costs)); |
495 | } |
496 | } |
497 | } |
498 | } |
499 | } |
500 | |
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; |
515 | } |
516 | } |
517 | } |
518 | }; |
519 | |
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, NumInstr: 1); |
526 | } |
527 | |
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) {} |
533 | }; |
534 | } // end anonymous namespace |
535 | |
536 | // Out-of-line destructor/anchor for PBQPRAConstraint. |
537 | PBQPRAConstraint::~PBQPRAConstraint() = default; |
538 | |
539 | void PBQPRAConstraint::anchor() {} |
540 | |
541 | void PBQPRAConstraintList::anchor() {} |
542 | |
543 | void RegAllocPBQP::getAnalysisUsage(AnalysisUsage &au) const { |
544 | au.setPreservesCFG(); |
545 | au.addRequired<AAResultsWrapperPass>(); |
546 | au.addPreserved<AAResultsWrapperPass>(); |
547 | au.addRequired<SlotIndexesWrapperPass>(); |
548 | au.addPreserved<SlotIndexesWrapperPass>(); |
549 | au.addRequired<LiveIntervalsWrapperPass>(); |
550 | au.addPreserved<LiveIntervalsWrapperPass>(); |
551 | //au.addRequiredID(SplitCriticalEdgesID); |
552 | if (customPassID) |
553 | au.addRequiredID(ID&: *customPassID); |
554 | au.addRequired<LiveStacks>(); |
555 | au.addPreserved<LiveStacks>(); |
556 | au.addRequired<MachineBlockFrequencyInfoWrapperPass>(); |
557 | au.addPreserved<MachineBlockFrequencyInfoWrapperPass>(); |
558 | au.addRequired<MachineLoopInfoWrapperPass>(); |
559 | au.addPreserved<MachineLoopInfoWrapperPass>(); |
560 | au.addRequired<MachineDominatorTreeWrapperPass>(); |
561 | au.addPreserved<MachineDominatorTreeWrapperPass>(); |
562 | au.addRequired<VirtRegMap>(); |
563 | au.addPreserved<VirtRegMap>(); |
564 | MachineFunctionPass::getAnalysisUsage(AU&: au); |
565 | } |
566 | |
567 | void RegAllocPBQP::findVRegIntervalsToAlloc(const MachineFunction &MF, |
568 | LiveIntervals &LIS) { |
569 | const MachineRegisterInfo &MRI = MF.getRegInfo(); |
570 | |
571 | // Iterate over all live ranges. |
572 | for (unsigned I = 0, E = MRI.getNumVirtRegs(); I != E; ++I) { |
573 | Register Reg = Register::index2VirtReg(Index: I); |
574 | if (MRI.reg_nodbg_empty(RegNo: Reg)) |
575 | continue; |
576 | VRegsToAlloc.insert(x: Reg); |
577 | } |
578 | } |
579 | |
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(RegA: Reg, RegB: CSR[i])) |
586 | return true; |
587 | return false; |
588 | } |
589 | |
590 | void RegAllocPBQP::initializeGraph(PBQPRAGraph &G, VirtRegMap &VRM, |
591 | Spiller &VRegSpiller) { |
592 | MachineFunction &MF = G.getMetadata().MF; |
593 | |
594 | LiveIntervals &LIS = G.getMetadata().LIS; |
595 | const MachineRegisterInfo &MRI = G.getMetadata().MF.getRegInfo(); |
596 | const TargetRegisterInfo &TRI = |
597 | *G.getMetadata().MF.getSubtarget().getRegisterInfo(); |
598 | |
599 | std::vector<Register> Worklist(VRegsToAlloc.begin(), VRegsToAlloc.end()); |
600 | |
601 | std::map<Register, std::vector<MCRegister>> VRegAllowedMap; |
602 | |
603 | while (!Worklist.empty()) { |
604 | Register VReg = Worklist.back(); |
605 | Worklist.pop_back(); |
606 | |
607 | LiveInterval &VRegLI = LIS.getInterval(Reg: VReg); |
608 | |
609 | // If this is an empty interval move it to the EmptyIntervalVRegs set then |
610 | // continue. |
611 | if (VRegLI.empty()) { |
612 | EmptyIntervalVRegs.insert(x: VRegLI.reg()); |
613 | VRegsToAlloc.erase(x: VRegLI.reg()); |
614 | continue; |
615 | } |
616 | |
617 | const TargetRegisterClass *TRC = MRI.getRegClass(Reg: VReg); |
618 | |
619 | // Record any overlaps with regmask operands. |
620 | BitVector RegMaskOverlaps; |
621 | LIS.checkRegMaskInterference(LI: VRegLI, UsableRegs&: RegMaskOverlaps); |
622 | |
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(PhysReg: PReg)) |
629 | continue; |
630 | |
631 | // vregLI crosses a regmask operand that clobbers preg. |
632 | if (!RegMaskOverlaps.empty() && !RegMaskOverlaps.test(Idx: PReg)) |
633 | continue; |
634 | |
635 | // vregLI overlaps fixed regunit interference. |
636 | bool Interference = false; |
637 | for (MCRegUnit Unit : TRI.regunits(Reg: PReg)) { |
638 | if (VRegLI.overlaps(other: LIS.getRegUnit(Unit))) { |
639 | Interference = true; |
640 | break; |
641 | } |
642 | } |
643 | if (Interference) |
644 | continue; |
645 | |
646 | // preg is usable for this virtual register. |
647 | VRegAllowed.push_back(x: PReg); |
648 | } |
649 | |
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, NewIntervals&: NewVRegs, MF, LIS, VRM, VRegSpiller); |
655 | llvm::append_range(C&: Worklist, R&: NewVRegs); |
656 | continue; |
657 | } |
658 | |
659 | VRegAllowedMap[VReg.id()] = std::move(VRegAllowed); |
660 | } |
661 | |
662 | for (auto &KV : VRegAllowedMap) { |
663 | auto VReg = KV.first; |
664 | |
665 | // Move empty intervals to the EmptyIntervalVReg set. |
666 | if (LIS.getInterval(Reg: VReg).empty()) { |
667 | EmptyIntervalVRegs.insert(x: VReg); |
668 | VRegsToAlloc.erase(x: VReg); |
669 | continue; |
670 | } |
671 | |
672 | auto &VRegAllowed = KV.second; |
673 | |
674 | PBQPRAGraph::RawVector NodeCosts(VRegAllowed.size() + 1, 0); |
675 | |
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(Reg: VRegAllowed[i], TRI, MF)) |
680 | NodeCosts[1 + i] += 1.0; |
681 | |
682 | PBQPRAGraph::NodeId NId = G.addNode(Costs: std::move(NodeCosts)); |
683 | G.getNodeMetadata(NId).setVReg(VReg); |
684 | G.getNodeMetadata(NId).setAllowedRegs( |
685 | G.getMetadata().getAllowedRegs(Allowed: std::move(VRegAllowed))); |
686 | G.getMetadata().setNodeIdForVReg(VReg, NId); |
687 | } |
688 | } |
689 | |
690 | void RegAllocPBQP::spillVReg(Register VReg, |
691 | SmallVectorImpl<Register> &NewIntervals, |
692 | MachineFunction &MF, LiveIntervals &LIS, |
693 | VirtRegMap &VRM, Spiller &VRegSpiller) { |
694 | VRegsToAlloc.erase(x: VReg); |
695 | LiveRangeEdit LRE(&LIS.getInterval(Reg: VReg), NewIntervals, MF, LIS, &VRM, |
696 | nullptr, &DeadRemats); |
697 | VRegSpiller.spill(LRE); |
698 | |
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: " ); |
703 | |
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(Reg: R); |
708 | assert(!LI.empty() && "Empty spill range." ); |
709 | LLVM_DEBUG(dbgs() << printReg(LI.reg(), &TRI) << " " ); |
710 | VRegsToAlloc.insert(x: LI.reg()); |
711 | } |
712 | |
713 | LLVM_DEBUG(dbgs() << ")\n" ); |
714 | } |
715 | |
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; |
724 | |
725 | // Set to true if we have any spills |
726 | bool AnotherRoundNeeded = false; |
727 | |
728 | // Clear the existing allocation. |
729 | VRM.clearAllVirt(); |
730 | |
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(nodeId: NId); |
736 | |
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(virtReg: VReg, physReg: 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, NewIntervals&: NewVRegs, MF, LIS, VRM, VRegSpiller); |
748 | AnotherRoundNeeded |= !NewVRegs.empty(); |
749 | } |
750 | } |
751 | |
752 | return !AnotherRoundNeeded; |
753 | } |
754 | |
755 | void RegAllocPBQP::finalizeAlloc(MachineFunction &MF, |
756 | LiveIntervals &LIS, |
757 | VirtRegMap &VRM) const { |
758 | MachineRegisterInfo &MRI = MF.getRegInfo(); |
759 | |
760 | // First allocate registers for the empty intervals. |
761 | for (const Register &R : EmptyIntervalVRegs) { |
762 | LiveInterval &LI = LIS.getInterval(Reg: R); |
763 | |
764 | Register PReg = MRI.getSimpleHint(VReg: LI.reg()); |
765 | |
766 | if (PReg == 0) { |
767 | const TargetRegisterClass &RC = *MRI.getRegClass(Reg: LI.reg()); |
768 | const ArrayRef<MCPhysReg> RawPRegOrder = RC.getRawAllocationOrder(MF); |
769 | for (MCRegister CandidateReg : RawPRegOrder) { |
770 | if (!VRM.getRegInfo().isReserved(PhysReg: CandidateReg)) { |
771 | PReg = CandidateReg; |
772 | break; |
773 | } |
774 | } |
775 | assert(PReg && |
776 | "No un-reserved physical registers in this register class" ); |
777 | } |
778 | |
779 | VRM.assignVirt2Phys(virtReg: LI.reg(), physReg: PReg); |
780 | } |
781 | } |
782 | |
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(MI&: *DeadInst); |
788 | DeadInst->eraseFromParent(); |
789 | } |
790 | DeadRemats.clear(); |
791 | } |
792 | |
793 | bool RegAllocPBQP::runOnMachineFunction(MachineFunction &MF) { |
794 | LiveIntervals &LIS = getAnalysis<LiveIntervalsWrapperPass>().getLIS(); |
795 | MachineBlockFrequencyInfo &MBFI = |
796 | getAnalysis<MachineBlockFrequencyInfoWrapperPass>().getMBFI(); |
797 | |
798 | VirtRegMap &VRM = getAnalysis<VirtRegMap>(); |
799 | |
800 | PBQPVirtRegAuxInfo VRAI( |
801 | MF, LIS, VRM, getAnalysis<MachineLoopInfoWrapperPass>().getLI(), MBFI); |
802 | VRAI.calculateSpillWeightsAndHints(); |
803 | |
804 | // FIXME: we create DefaultVRAI here to match existing behavior pre-passing |
805 | // the VRAI through the spiller to the live range editor. However, it probably |
806 | // makes more sense to pass the PBQP VRAI. The existing behavior had |
807 | // LiveRangeEdit make its own VirtRegAuxInfo object. |
808 | VirtRegAuxInfo DefaultVRAI( |
809 | MF, LIS, VRM, getAnalysis<MachineLoopInfoWrapperPass>().getLI(), MBFI); |
810 | std::unique_ptr<Spiller> VRegSpiller( |
811 | createInlineSpiller(Pass&: *this, MF, VRM, VRAI&: DefaultVRAI)); |
812 | |
813 | MF.getRegInfo().freezeReservedRegs(); |
814 | |
815 | LLVM_DEBUG(dbgs() << "PBQP Register Allocating for " << MF.getName() << "\n" ); |
816 | |
817 | // Allocator main loop: |
818 | // |
819 | // * Map current regalloc problem to a PBQP problem |
820 | // * Solve the PBQP problem |
821 | // * Map the solution back to a register allocation |
822 | // * Spill if necessary |
823 | // |
824 | // This process is continued till no more spills are generated. |
825 | |
826 | // Find the vreg intervals in need of allocation. |
827 | findVRegIntervalsToAlloc(MF, LIS); |
828 | |
829 | #ifndef NDEBUG |
830 | const Function &F = MF.getFunction(); |
831 | std::string FullyQualifiedName = |
832 | F.getParent()->getModuleIdentifier() + "." + F.getName().str(); |
833 | #endif |
834 | |
835 | // If there are non-empty intervals allocate them using pbqp. |
836 | if (!VRegsToAlloc.empty()) { |
837 | const TargetSubtargetInfo &Subtarget = MF.getSubtarget(); |
838 | std::unique_ptr<PBQPRAConstraintList> ConstraintsRoot = |
839 | std::make_unique<PBQPRAConstraintList>(); |
840 | ConstraintsRoot->addConstraint(C: std::make_unique<SpillCosts>()); |
841 | ConstraintsRoot->addConstraint(C: std::make_unique<Interference>()); |
842 | if (PBQPCoalescing) |
843 | ConstraintsRoot->addConstraint(C: std::make_unique<Coalescing>()); |
844 | ConstraintsRoot->addConstraint(C: Subtarget.getCustomPBQPConstraints()); |
845 | |
846 | bool PBQPAllocComplete = false; |
847 | unsigned Round = 0; |
848 | |
849 | while (!PBQPAllocComplete) { |
850 | LLVM_DEBUG(dbgs() << " PBQP Regalloc round " << Round << ":\n" ); |
851 | (void) Round; |
852 | |
853 | PBQPRAGraph G(PBQPRAGraph::GraphMetadata(MF, LIS, MBFI)); |
854 | initializeGraph(G, VRM, VRegSpiller&: *VRegSpiller); |
855 | ConstraintsRoot->apply(G); |
856 | |
857 | #ifndef NDEBUG |
858 | if (PBQPDumpGraphs) { |
859 | std::ostringstream RS; |
860 | RS << Round; |
861 | std::string GraphFileName = FullyQualifiedName + "." + RS.str() + |
862 | ".pbqpgraph" ; |
863 | std::error_code EC; |
864 | raw_fd_ostream OS(GraphFileName, EC, sys::fs::OF_TextWithCRLF); |
865 | LLVM_DEBUG(dbgs() << "Dumping graph for round " << Round << " to \"" |
866 | << GraphFileName << "\"\n" ); |
867 | G.dump(OS); |
868 | } |
869 | #endif |
870 | |
871 | PBQP::Solution Solution = PBQP::RegAlloc::solve(G); |
872 | PBQPAllocComplete = mapPBQPToRegAlloc(G, Solution, VRM, VRegSpiller&: *VRegSpiller); |
873 | ++Round; |
874 | } |
875 | } |
876 | |
877 | // Finalise allocation, allocate empty ranges. |
878 | finalizeAlloc(MF, LIS, VRM); |
879 | postOptimization(VRegSpiller&: *VRegSpiller, LIS); |
880 | VRegsToAlloc.clear(); |
881 | EmptyIntervalVRegs.clear(); |
882 | |
883 | LLVM_DEBUG(dbgs() << "Post alloc VirtRegMap:\n" << VRM << "\n" ); |
884 | |
885 | return true; |
886 | } |
887 | |
888 | /// Create Printable object for node and register info. |
889 | static Printable PrintNodeInfo(PBQP::RegAlloc::PBQPRAGraph::NodeId NId, |
890 | const PBQP::RegAlloc::PBQPRAGraph &G) { |
891 | return Printable([NId, &G](raw_ostream &OS) { |
892 | const MachineRegisterInfo &MRI = G.getMetadata().MF.getRegInfo(); |
893 | const TargetRegisterInfo *TRI = MRI.getTargetRegisterInfo(); |
894 | Register VReg = G.getNodeMetadata(NId).getVReg(); |
895 | const char *RegClassName = TRI->getRegClassName(Class: MRI.getRegClass(Reg: VReg)); |
896 | OS << NId << " (" << RegClassName << ':' << printReg(Reg: VReg, TRI) << ')'; |
897 | }); |
898 | } |
899 | |
900 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
901 | LLVM_DUMP_METHOD void PBQP::RegAlloc::PBQPRAGraph::dump(raw_ostream &OS) const { |
902 | for (auto NId : nodeIds()) { |
903 | const Vector &Costs = getNodeCosts(NId); |
904 | assert(Costs.getLength() != 0 && "Empty vector in graph." ); |
905 | OS << PrintNodeInfo(NId, *this) << ": " << Costs << '\n'; |
906 | } |
907 | OS << '\n'; |
908 | |
909 | for (auto EId : edgeIds()) { |
910 | NodeId N1Id = getEdgeNode1Id(EId); |
911 | NodeId N2Id = getEdgeNode2Id(EId); |
912 | assert(N1Id != N2Id && "PBQP graphs should not have self-edges." ); |
913 | const Matrix &M = getEdgeCosts(EId); |
914 | assert(M.getRows() != 0 && "No rows in matrix." ); |
915 | assert(M.getCols() != 0 && "No cols in matrix." ); |
916 | OS << PrintNodeInfo(N1Id, *this) << ' ' << M.getRows() << " rows / " ; |
917 | OS << PrintNodeInfo(N2Id, *this) << ' ' << M.getCols() << " cols:\n" ; |
918 | OS << M << '\n'; |
919 | } |
920 | } |
921 | |
922 | LLVM_DUMP_METHOD void PBQP::RegAlloc::PBQPRAGraph::dump() const { |
923 | dump(dbgs()); |
924 | } |
925 | #endif |
926 | |
927 | void PBQP::RegAlloc::PBQPRAGraph::printDot(raw_ostream &OS) const { |
928 | OS << "graph {\n" ; |
929 | for (auto NId : nodeIds()) { |
930 | OS << " node" << NId << " [ label=\"" |
931 | << PrintNodeInfo(NId, G: *this) << "\\n" |
932 | << getNodeCosts(NId) << "\" ]\n" ; |
933 | } |
934 | |
935 | OS << " edge [ len=" << nodeIds().size() << " ]\n" ; |
936 | for (auto EId : edgeIds()) { |
937 | OS << " node" << getEdgeNode1Id(EId) |
938 | << " -- node" << getEdgeNode2Id(EId) |
939 | << " [ label=\"" ; |
940 | const Matrix &EdgeCosts = getEdgeCosts(EId); |
941 | for (unsigned i = 0; i < EdgeCosts.getRows(); ++i) { |
942 | OS << EdgeCosts.getRowAsVector(R: i) << "\\n" ; |
943 | } |
944 | OS << "\" ]\n" ; |
945 | } |
946 | OS << "}\n" ; |
947 | } |
948 | |
949 | FunctionPass *llvm::createPBQPRegisterAllocator(char *customPassID) { |
950 | return new RegAllocPBQP(customPassID); |
951 | } |
952 | |
953 | FunctionPass* llvm::createDefaultPBQPRegisterAllocator() { |
954 | return createPBQPRegisterAllocator(); |
955 | } |
956 | |