1 | //===- RegAllocGreedy.cpp - greedy 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 defines the RAGreedy function pass for register allocation in |
10 | // optimized builds. |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #include "RegAllocGreedy.h" |
15 | #include "AllocationOrder.h" |
16 | #include "InterferenceCache.h" |
17 | #include "LiveDebugVariables.h" |
18 | #include "RegAllocBase.h" |
19 | #include "RegAllocEvictionAdvisor.h" |
20 | #include "RegAllocPriorityAdvisor.h" |
21 | #include "SpillPlacement.h" |
22 | #include "SplitKit.h" |
23 | #include "llvm/ADT/ArrayRef.h" |
24 | #include "llvm/ADT/BitVector.h" |
25 | #include "llvm/ADT/IndexedMap.h" |
26 | #include "llvm/ADT/SmallSet.h" |
27 | #include "llvm/ADT/SmallVector.h" |
28 | #include "llvm/ADT/Statistic.h" |
29 | #include "llvm/ADT/StringRef.h" |
30 | #include "llvm/Analysis/AliasAnalysis.h" |
31 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
32 | #include "llvm/CodeGen/CalcSpillWeights.h" |
33 | #include "llvm/CodeGen/EdgeBundles.h" |
34 | #include "llvm/CodeGen/LiveInterval.h" |
35 | #include "llvm/CodeGen/LiveIntervalUnion.h" |
36 | #include "llvm/CodeGen/LiveIntervals.h" |
37 | #include "llvm/CodeGen/LiveRangeEdit.h" |
38 | #include "llvm/CodeGen/LiveRegMatrix.h" |
39 | #include "llvm/CodeGen/LiveStacks.h" |
40 | #include "llvm/CodeGen/MachineBasicBlock.h" |
41 | #include "llvm/CodeGen/MachineBlockFrequencyInfo.h" |
42 | #include "llvm/CodeGen/MachineDominators.h" |
43 | #include "llvm/CodeGen/MachineFrameInfo.h" |
44 | #include "llvm/CodeGen/MachineFunction.h" |
45 | #include "llvm/CodeGen/MachineFunctionPass.h" |
46 | #include "llvm/CodeGen/MachineInstr.h" |
47 | #include "llvm/CodeGen/MachineLoopInfo.h" |
48 | #include "llvm/CodeGen/MachineOperand.h" |
49 | #include "llvm/CodeGen/MachineOptimizationRemarkEmitter.h" |
50 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
51 | #include "llvm/CodeGen/RegAllocRegistry.h" |
52 | #include "llvm/CodeGen/RegisterClassInfo.h" |
53 | #include "llvm/CodeGen/SlotIndexes.h" |
54 | #include "llvm/CodeGen/Spiller.h" |
55 | #include "llvm/CodeGen/TargetInstrInfo.h" |
56 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
57 | #include "llvm/CodeGen/TargetSubtargetInfo.h" |
58 | #include "llvm/CodeGen/VirtRegMap.h" |
59 | #include "llvm/IR/DebugInfoMetadata.h" |
60 | #include "llvm/IR/Function.h" |
61 | #include "llvm/IR/LLVMContext.h" |
62 | #include "llvm/InitializePasses.h" |
63 | #include "llvm/MC/MCRegisterInfo.h" |
64 | #include "llvm/Pass.h" |
65 | #include "llvm/Support/BlockFrequency.h" |
66 | #include "llvm/Support/BranchProbability.h" |
67 | #include "llvm/Support/CommandLine.h" |
68 | #include "llvm/Support/Debug.h" |
69 | #include "llvm/Support/MathExtras.h" |
70 | #include "llvm/Support/Timer.h" |
71 | #include "llvm/Support/raw_ostream.h" |
72 | #include <algorithm> |
73 | #include <cassert> |
74 | #include <cstdint> |
75 | #include <utility> |
76 | |
77 | using namespace llvm; |
78 | |
79 | #define DEBUG_TYPE "regalloc" |
80 | |
81 | STATISTIC(NumGlobalSplits, "Number of split global live ranges" ); |
82 | STATISTIC(NumLocalSplits, "Number of split local live ranges" ); |
83 | STATISTIC(NumEvicted, "Number of interferences evicted" ); |
84 | |
85 | static cl::opt<SplitEditor::ComplementSpillMode> SplitSpillMode( |
86 | "split-spill-mode" , cl::Hidden, |
87 | cl::desc("Spill mode for splitting live ranges" ), |
88 | cl::values(clEnumValN(SplitEditor::SM_Partition, "default" , "Default" ), |
89 | clEnumValN(SplitEditor::SM_Size, "size" , "Optimize for size" ), |
90 | clEnumValN(SplitEditor::SM_Speed, "speed" , "Optimize for speed" )), |
91 | cl::init(Val: SplitEditor::SM_Speed)); |
92 | |
93 | static cl::opt<unsigned> |
94 | LastChanceRecoloringMaxDepth("lcr-max-depth" , cl::Hidden, |
95 | cl::desc("Last chance recoloring max depth" ), |
96 | cl::init(Val: 5)); |
97 | |
98 | static cl::opt<unsigned> LastChanceRecoloringMaxInterference( |
99 | "lcr-max-interf" , cl::Hidden, |
100 | cl::desc("Last chance recoloring maximum number of considered" |
101 | " interference at a time" ), |
102 | cl::init(Val: 8)); |
103 | |
104 | static cl::opt<bool> ExhaustiveSearch( |
105 | "exhaustive-register-search" , cl::NotHidden, |
106 | cl::desc("Exhaustive Search for registers bypassing the depth " |
107 | "and interference cutoffs of last chance recoloring" ), |
108 | cl::Hidden); |
109 | |
110 | static cl::opt<bool> EnableDeferredSpilling( |
111 | "enable-deferred-spilling" , cl::Hidden, |
112 | cl::desc("Instead of spilling a variable right away, defer the actual " |
113 | "code insertion to the end of the allocation. That way the " |
114 | "allocator might still find a suitable coloring for this " |
115 | "variable because of other evicted variables." ), |
116 | cl::init(Val: false)); |
117 | |
118 | // FIXME: Find a good default for this flag and remove the flag. |
119 | static cl::opt<unsigned> |
120 | CSRFirstTimeCost("regalloc-csr-first-time-cost" , |
121 | cl::desc("Cost for first time use of callee-saved register." ), |
122 | cl::init(Val: 0), cl::Hidden); |
123 | |
124 | static cl::opt<unsigned long> GrowRegionComplexityBudget( |
125 | "grow-region-complexity-budget" , |
126 | cl::desc("growRegion() does not scale with the number of BB edges, so " |
127 | "limit its budget and bail out once we reach the limit." ), |
128 | cl::init(Val: 10000), cl::Hidden); |
129 | |
130 | static cl::opt<bool> GreedyRegClassPriorityTrumpsGlobalness( |
131 | "greedy-regclass-priority-trumps-globalness" , |
132 | cl::desc("Change the greedy register allocator's live range priority " |
133 | "calculation to make the AllocationPriority of the register class " |
134 | "more important then whether the range is global" ), |
135 | cl::Hidden); |
136 | |
137 | static cl::opt<bool> GreedyReverseLocalAssignment( |
138 | "greedy-reverse-local-assignment" , |
139 | cl::desc("Reverse allocation order of local live ranges, such that " |
140 | "shorter local live ranges will tend to be allocated first" ), |
141 | cl::Hidden); |
142 | |
143 | static cl::opt<unsigned> SplitThresholdForRegWithHint( |
144 | "split-threshold-for-reg-with-hint" , |
145 | cl::desc("The threshold for splitting a virtual register with a hint, in " |
146 | "percentate" ), |
147 | cl::init(Val: 75), cl::Hidden); |
148 | |
149 | static RegisterRegAlloc greedyRegAlloc("greedy" , "greedy register allocator" , |
150 | createGreedyRegisterAllocator); |
151 | |
152 | char RAGreedy::ID = 0; |
153 | char &llvm::RAGreedyID = RAGreedy::ID; |
154 | |
155 | INITIALIZE_PASS_BEGIN(RAGreedy, "greedy" , |
156 | "Greedy Register Allocator" , false, false) |
157 | INITIALIZE_PASS_DEPENDENCY(LiveDebugVariables) |
158 | INITIALIZE_PASS_DEPENDENCY(SlotIndexes) |
159 | INITIALIZE_PASS_DEPENDENCY(LiveIntervals) |
160 | INITIALIZE_PASS_DEPENDENCY(RegisterCoalescer) |
161 | INITIALIZE_PASS_DEPENDENCY(MachineScheduler) |
162 | INITIALIZE_PASS_DEPENDENCY(LiveStacks) |
163 | INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree) |
164 | INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo) |
165 | INITIALIZE_PASS_DEPENDENCY(VirtRegMap) |
166 | INITIALIZE_PASS_DEPENDENCY(LiveRegMatrix) |
167 | INITIALIZE_PASS_DEPENDENCY(EdgeBundles) |
168 | INITIALIZE_PASS_DEPENDENCY(SpillPlacement) |
169 | INITIALIZE_PASS_DEPENDENCY(MachineOptimizationRemarkEmitterPass) |
170 | INITIALIZE_PASS_DEPENDENCY(RegAllocEvictionAdvisorAnalysis) |
171 | INITIALIZE_PASS_DEPENDENCY(RegAllocPriorityAdvisorAnalysis) |
172 | INITIALIZE_PASS_END(RAGreedy, "greedy" , |
173 | "Greedy Register Allocator" , false, false) |
174 | |
175 | #ifndef NDEBUG |
176 | const char *const RAGreedy::StageName[] = { |
177 | "RS_New" , |
178 | "RS_Assign" , |
179 | "RS_Split" , |
180 | "RS_Split2" , |
181 | "RS_Spill" , |
182 | "RS_Memory" , |
183 | "RS_Done" |
184 | }; |
185 | #endif |
186 | |
187 | // Hysteresis to use when comparing floats. |
188 | // This helps stabilize decisions based on float comparisons. |
189 | const float Hysteresis = (2007 / 2048.0f); // 0.97998046875 |
190 | |
191 | FunctionPass* llvm::createGreedyRegisterAllocator() { |
192 | return new RAGreedy(); |
193 | } |
194 | |
195 | FunctionPass *llvm::createGreedyRegisterAllocator(RegClassFilterFunc Ftor) { |
196 | return new RAGreedy(Ftor); |
197 | } |
198 | |
199 | RAGreedy::RAGreedy(RegClassFilterFunc F): |
200 | MachineFunctionPass(ID), |
201 | RegAllocBase(F) { |
202 | } |
203 | |
204 | void RAGreedy::getAnalysisUsage(AnalysisUsage &AU) const { |
205 | AU.setPreservesCFG(); |
206 | AU.addRequired<MachineBlockFrequencyInfo>(); |
207 | AU.addPreserved<MachineBlockFrequencyInfo>(); |
208 | AU.addRequired<LiveIntervals>(); |
209 | AU.addPreserved<LiveIntervals>(); |
210 | AU.addRequired<SlotIndexes>(); |
211 | AU.addPreserved<SlotIndexes>(); |
212 | AU.addRequired<LiveDebugVariables>(); |
213 | AU.addPreserved<LiveDebugVariables>(); |
214 | AU.addRequired<LiveStacks>(); |
215 | AU.addPreserved<LiveStacks>(); |
216 | AU.addRequired<MachineDominatorTree>(); |
217 | AU.addPreserved<MachineDominatorTree>(); |
218 | AU.addRequired<MachineLoopInfo>(); |
219 | AU.addPreserved<MachineLoopInfo>(); |
220 | AU.addRequired<VirtRegMap>(); |
221 | AU.addPreserved<VirtRegMap>(); |
222 | AU.addRequired<LiveRegMatrix>(); |
223 | AU.addPreserved<LiveRegMatrix>(); |
224 | AU.addRequired<EdgeBundles>(); |
225 | AU.addRequired<SpillPlacement>(); |
226 | AU.addRequired<MachineOptimizationRemarkEmitterPass>(); |
227 | AU.addRequired<RegAllocEvictionAdvisorAnalysis>(); |
228 | AU.addRequired<RegAllocPriorityAdvisorAnalysis>(); |
229 | MachineFunctionPass::getAnalysisUsage(AU); |
230 | } |
231 | |
232 | //===----------------------------------------------------------------------===// |
233 | // LiveRangeEdit delegate methods |
234 | //===----------------------------------------------------------------------===// |
235 | |
236 | bool RAGreedy::LRE_CanEraseVirtReg(Register VirtReg) { |
237 | LiveInterval &LI = LIS->getInterval(Reg: VirtReg); |
238 | if (VRM->hasPhys(virtReg: VirtReg)) { |
239 | Matrix->unassign(VirtReg: LI); |
240 | aboutToRemoveInterval(LI); |
241 | return true; |
242 | } |
243 | // Unassigned virtreg is probably in the priority queue. |
244 | // RegAllocBase will erase it after dequeueing. |
245 | // Nonetheless, clear the live-range so that the debug |
246 | // dump will show the right state for that VirtReg. |
247 | LI.clear(); |
248 | return false; |
249 | } |
250 | |
251 | void RAGreedy::LRE_WillShrinkVirtReg(Register VirtReg) { |
252 | if (!VRM->hasPhys(virtReg: VirtReg)) |
253 | return; |
254 | |
255 | // Register is assigned, put it back on the queue for reassignment. |
256 | LiveInterval &LI = LIS->getInterval(Reg: VirtReg); |
257 | Matrix->unassign(VirtReg: LI); |
258 | RegAllocBase::enqueue(LI: &LI); |
259 | } |
260 | |
261 | void RAGreedy::LRE_DidCloneVirtReg(Register New, Register Old) { |
262 | ExtraInfo->LRE_DidCloneVirtReg(New, Old); |
263 | } |
264 | |
265 | void RAGreedy::ExtraRegInfo::(Register New, Register Old) { |
266 | // Cloning a register we haven't even heard about yet? Just ignore it. |
267 | if (!Info.inBounds(n: Old)) |
268 | return; |
269 | |
270 | // LRE may clone a virtual register because dead code elimination causes it to |
271 | // be split into connected components. The new components are much smaller |
272 | // than the original, so they should get a new chance at being assigned. |
273 | // same stage as the parent. |
274 | Info[Old].Stage = RS_Assign; |
275 | Info.grow(n: New.id()); |
276 | Info[New] = Info[Old]; |
277 | } |
278 | |
279 | void RAGreedy::releaseMemory() { |
280 | SpillerInstance.reset(); |
281 | GlobalCand.clear(); |
282 | } |
283 | |
284 | void RAGreedy::enqueueImpl(const LiveInterval *LI) { enqueue(CurQueue&: Queue, LI); } |
285 | |
286 | void RAGreedy::enqueue(PQueue &CurQueue, const LiveInterval *LI) { |
287 | // Prioritize live ranges by size, assigning larger ranges first. |
288 | // The queue holds (size, reg) pairs. |
289 | const Register Reg = LI->reg(); |
290 | assert(Reg.isVirtual() && "Can only enqueue virtual registers" ); |
291 | |
292 | auto Stage = ExtraInfo->getOrInitStage(Reg); |
293 | if (Stage == RS_New) { |
294 | Stage = RS_Assign; |
295 | ExtraInfo->setStage(Reg, Stage); |
296 | } |
297 | |
298 | unsigned Ret = PriorityAdvisor->getPriority(LI: *LI); |
299 | |
300 | // The virtual register number is a tie breaker for same-sized ranges. |
301 | // Give lower vreg numbers higher priority to assign them first. |
302 | CurQueue.push(x: std::make_pair(x&: Ret, y: ~Reg)); |
303 | } |
304 | |
305 | unsigned DefaultPriorityAdvisor::getPriority(const LiveInterval &LI) const { |
306 | const unsigned Size = LI.getSize(); |
307 | const Register Reg = LI.reg(); |
308 | unsigned Prio; |
309 | LiveRangeStage Stage = RA.getExtraInfo().getStage(VirtReg: LI); |
310 | |
311 | if (Stage == RS_Split) { |
312 | // Unsplit ranges that couldn't be allocated immediately are deferred until |
313 | // everything else has been allocated. |
314 | Prio = Size; |
315 | } else if (Stage == RS_Memory) { |
316 | // Memory operand should be considered last. |
317 | // Change the priority such that Memory operand are assigned in |
318 | // the reverse order that they came in. |
319 | // TODO: Make this a member variable and probably do something about hints. |
320 | static unsigned MemOp = 0; |
321 | Prio = MemOp++; |
322 | } else { |
323 | // Giant live ranges fall back to the global assignment heuristic, which |
324 | // prevents excessive spilling in pathological cases. |
325 | const TargetRegisterClass &RC = *MRI->getRegClass(Reg); |
326 | bool ForceGlobal = RC.GlobalPriority || |
327 | (!ReverseLocalAssignment && |
328 | (Size / SlotIndex::InstrDist) > |
329 | (2 * RegClassInfo.getNumAllocatableRegs(RC: &RC))); |
330 | unsigned GlobalBit = 0; |
331 | |
332 | if (Stage == RS_Assign && !ForceGlobal && !LI.empty() && |
333 | LIS->intervalIsInOneMBB(LI)) { |
334 | // Allocate original local ranges in linear instruction order. Since they |
335 | // are singly defined, this produces optimal coloring in the absence of |
336 | // global interference and other constraints. |
337 | if (!ReverseLocalAssignment) |
338 | Prio = LI.beginIndex().getApproxInstrDistance(other: Indexes->getLastIndex()); |
339 | else { |
340 | // Allocating bottom up may allow many short LRGs to be assigned first |
341 | // to one of the cheap registers. This could be much faster for very |
342 | // large blocks on targets with many physical registers. |
343 | Prio = Indexes->getZeroIndex().getApproxInstrDistance(other: LI.endIndex()); |
344 | } |
345 | } else { |
346 | // Allocate global and split ranges in long->short order. Long ranges that |
347 | // don't fit should be spilled (or split) ASAP so they don't create |
348 | // interference. Mark a bit to prioritize global above local ranges. |
349 | Prio = Size; |
350 | GlobalBit = 1; |
351 | } |
352 | |
353 | // Priority bit layout: |
354 | // 31 RS_Assign priority |
355 | // 30 Preference priority |
356 | // if (RegClassPriorityTrumpsGlobalness) |
357 | // 29-25 AllocPriority |
358 | // 24 GlobalBit |
359 | // else |
360 | // 29 Global bit |
361 | // 28-24 AllocPriority |
362 | // 0-23 Size/Instr distance |
363 | |
364 | // Clamp the size to fit with the priority masking scheme |
365 | Prio = std::min(a: Prio, b: (unsigned)maxUIntN(N: 24)); |
366 | assert(isUInt<5>(RC.AllocationPriority) && "allocation priority overflow" ); |
367 | |
368 | if (RegClassPriorityTrumpsGlobalness) |
369 | Prio |= RC.AllocationPriority << 25 | GlobalBit << 24; |
370 | else |
371 | Prio |= GlobalBit << 29 | RC.AllocationPriority << 24; |
372 | |
373 | // Mark a higher bit to prioritize global and local above RS_Split. |
374 | Prio |= (1u << 31); |
375 | |
376 | // Boost ranges that have a physical register hint. |
377 | if (VRM->hasKnownPreference(VirtReg: Reg)) |
378 | Prio |= (1u << 30); |
379 | } |
380 | |
381 | return Prio; |
382 | } |
383 | |
384 | const LiveInterval *RAGreedy::dequeue() { return dequeue(CurQueue&: Queue); } |
385 | |
386 | const LiveInterval *RAGreedy::dequeue(PQueue &CurQueue) { |
387 | if (CurQueue.empty()) |
388 | return nullptr; |
389 | LiveInterval *LI = &LIS->getInterval(Reg: ~CurQueue.top().second); |
390 | CurQueue.pop(); |
391 | return LI; |
392 | } |
393 | |
394 | //===----------------------------------------------------------------------===// |
395 | // Direct Assignment |
396 | //===----------------------------------------------------------------------===// |
397 | |
398 | /// tryAssign - Try to assign VirtReg to an available register. |
399 | MCRegister RAGreedy::tryAssign(const LiveInterval &VirtReg, |
400 | AllocationOrder &Order, |
401 | SmallVectorImpl<Register> &NewVRegs, |
402 | const SmallVirtRegSet &FixedRegisters) { |
403 | MCRegister PhysReg; |
404 | for (auto I = Order.begin(), E = Order.end(); I != E && !PhysReg; ++I) { |
405 | assert(*I); |
406 | if (!Matrix->checkInterference(VirtReg, PhysReg: *I)) { |
407 | if (I.isHint()) |
408 | return *I; |
409 | else |
410 | PhysReg = *I; |
411 | } |
412 | } |
413 | if (!PhysReg.isValid()) |
414 | return PhysReg; |
415 | |
416 | // PhysReg is available, but there may be a better choice. |
417 | |
418 | // If we missed a simple hint, try to cheaply evict interference from the |
419 | // preferred register. |
420 | if (Register Hint = MRI->getSimpleHint(VReg: VirtReg.reg())) |
421 | if (Order.isHint(Reg: Hint)) { |
422 | MCRegister PhysHint = Hint.asMCReg(); |
423 | LLVM_DEBUG(dbgs() << "missed hint " << printReg(PhysHint, TRI) << '\n'); |
424 | |
425 | if (EvictAdvisor->canEvictHintInterference(VirtReg, PhysReg: PhysHint, |
426 | FixedRegisters)) { |
427 | evictInterference(VirtReg, PhysHint, NewVRegs); |
428 | return PhysHint; |
429 | } |
430 | |
431 | // We can also split the virtual register in cold blocks. |
432 | if (trySplitAroundHintReg(Hint: PhysHint, VirtReg, NewVRegs, Order)) |
433 | return 0; |
434 | |
435 | // Record the missed hint, we may be able to recover |
436 | // at the end if the surrounding allocation changed. |
437 | SetOfBrokenHints.insert(X: &VirtReg); |
438 | } |
439 | |
440 | // Try to evict interference from a cheaper alternative. |
441 | uint8_t Cost = RegCosts[PhysReg]; |
442 | |
443 | // Most registers have 0 additional cost. |
444 | if (!Cost) |
445 | return PhysReg; |
446 | |
447 | LLVM_DEBUG(dbgs() << printReg(PhysReg, TRI) << " is available at cost " |
448 | << (unsigned)Cost << '\n'); |
449 | MCRegister CheapReg = tryEvict(VirtReg, Order, NewVRegs, Cost, FixedRegisters); |
450 | return CheapReg ? CheapReg : PhysReg; |
451 | } |
452 | |
453 | //===----------------------------------------------------------------------===// |
454 | // Interference eviction |
455 | //===----------------------------------------------------------------------===// |
456 | |
457 | bool RegAllocEvictionAdvisor::canReassign(const LiveInterval &VirtReg, |
458 | MCRegister FromReg) const { |
459 | auto HasRegUnitInterference = [&](MCRegUnit Unit) { |
460 | // Instantiate a "subquery", not to be confused with the Queries array. |
461 | LiveIntervalUnion::Query SubQ(VirtReg, Matrix->getLiveUnions()[Unit]); |
462 | return SubQ.checkInterference(); |
463 | }; |
464 | |
465 | for (MCRegister Reg : |
466 | AllocationOrder::create(VirtReg: VirtReg.reg(), VRM: *VRM, RegClassInfo, Matrix)) { |
467 | if (Reg == FromReg) |
468 | continue; |
469 | // If no units have interference, reassignment is possible. |
470 | if (none_of(Range: TRI->regunits(Reg), P: HasRegUnitInterference)) { |
471 | LLVM_DEBUG(dbgs() << "can reassign: " << VirtReg << " from " |
472 | << printReg(FromReg, TRI) << " to " |
473 | << printReg(Reg, TRI) << '\n'); |
474 | return true; |
475 | } |
476 | } |
477 | return false; |
478 | } |
479 | |
480 | /// evictInterference - Evict any interferring registers that prevent VirtReg |
481 | /// from being assigned to Physreg. This assumes that canEvictInterference |
482 | /// returned true. |
483 | void RAGreedy::evictInterference(const LiveInterval &VirtReg, |
484 | MCRegister PhysReg, |
485 | SmallVectorImpl<Register> &NewVRegs) { |
486 | // Make sure that VirtReg has a cascade number, and assign that cascade |
487 | // number to every evicted register. These live ranges than then only be |
488 | // evicted by a newer cascade, preventing infinite loops. |
489 | unsigned Cascade = ExtraInfo->getOrAssignNewCascade(Reg: VirtReg.reg()); |
490 | |
491 | LLVM_DEBUG(dbgs() << "evicting " << printReg(PhysReg, TRI) |
492 | << " interference: Cascade " << Cascade << '\n'); |
493 | |
494 | // Collect all interfering virtregs first. |
495 | SmallVector<const LiveInterval *, 8> Intfs; |
496 | for (MCRegUnit Unit : TRI->regunits(Reg: PhysReg)) { |
497 | LiveIntervalUnion::Query &Q = Matrix->query(LR: VirtReg, RegUnit: Unit); |
498 | // We usually have the interfering VRegs cached so collectInterferingVRegs() |
499 | // should be fast, we may need to recalculate if when different physregs |
500 | // overlap the same register unit so we had different SubRanges queried |
501 | // against it. |
502 | ArrayRef<const LiveInterval *> IVR = Q.interferingVRegs(); |
503 | Intfs.append(in_start: IVR.begin(), in_end: IVR.end()); |
504 | } |
505 | |
506 | // Evict them second. This will invalidate the queries. |
507 | for (const LiveInterval *Intf : Intfs) { |
508 | // The same VirtReg may be present in multiple RegUnits. Skip duplicates. |
509 | if (!VRM->hasPhys(virtReg: Intf->reg())) |
510 | continue; |
511 | |
512 | Matrix->unassign(VirtReg: *Intf); |
513 | assert((ExtraInfo->getCascade(Intf->reg()) < Cascade || |
514 | VirtReg.isSpillable() < Intf->isSpillable()) && |
515 | "Cannot decrease cascade number, illegal eviction" ); |
516 | ExtraInfo->setCascade(Reg: Intf->reg(), Cascade); |
517 | ++NumEvicted; |
518 | NewVRegs.push_back(Elt: Intf->reg()); |
519 | } |
520 | } |
521 | |
522 | /// Returns true if the given \p PhysReg is a callee saved register and has not |
523 | /// been used for allocation yet. |
524 | bool RegAllocEvictionAdvisor::isUnusedCalleeSavedReg(MCRegister PhysReg) const { |
525 | MCRegister CSR = RegClassInfo.getLastCalleeSavedAlias(PhysReg); |
526 | if (!CSR) |
527 | return false; |
528 | |
529 | return !Matrix->isPhysRegUsed(PhysReg); |
530 | } |
531 | |
532 | std::optional<unsigned> |
533 | RegAllocEvictionAdvisor::getOrderLimit(const LiveInterval &VirtReg, |
534 | const AllocationOrder &Order, |
535 | unsigned CostPerUseLimit) const { |
536 | unsigned OrderLimit = Order.getOrder().size(); |
537 | |
538 | if (CostPerUseLimit < uint8_t(~0u)) { |
539 | // Check of any registers in RC are below CostPerUseLimit. |
540 | const TargetRegisterClass *RC = MRI->getRegClass(Reg: VirtReg.reg()); |
541 | uint8_t MinCost = RegClassInfo.getMinCost(RC); |
542 | if (MinCost >= CostPerUseLimit) { |
543 | LLVM_DEBUG(dbgs() << TRI->getRegClassName(RC) << " minimum cost = " |
544 | << MinCost << ", no cheaper registers to be found.\n" ); |
545 | return std::nullopt; |
546 | } |
547 | |
548 | // It is normal for register classes to have a long tail of registers with |
549 | // the same cost. We don't need to look at them if they're too expensive. |
550 | if (RegCosts[Order.getOrder().back()] >= CostPerUseLimit) { |
551 | OrderLimit = RegClassInfo.getLastCostChange(RC); |
552 | LLVM_DEBUG(dbgs() << "Only trying the first " << OrderLimit |
553 | << " regs.\n" ); |
554 | } |
555 | } |
556 | return OrderLimit; |
557 | } |
558 | |
559 | bool RegAllocEvictionAdvisor::canAllocatePhysReg(unsigned CostPerUseLimit, |
560 | MCRegister PhysReg) const { |
561 | if (RegCosts[PhysReg] >= CostPerUseLimit) |
562 | return false; |
563 | // The first use of a callee-saved register in a function has cost 1. |
564 | // Don't start using a CSR when the CostPerUseLimit is low. |
565 | if (CostPerUseLimit == 1 && isUnusedCalleeSavedReg(PhysReg)) { |
566 | LLVM_DEBUG( |
567 | dbgs() << printReg(PhysReg, TRI) << " would clobber CSR " |
568 | << printReg(RegClassInfo.getLastCalleeSavedAlias(PhysReg), TRI) |
569 | << '\n'); |
570 | return false; |
571 | } |
572 | return true; |
573 | } |
574 | |
575 | /// tryEvict - Try to evict all interferences for a physreg. |
576 | /// @param VirtReg Currently unassigned virtual register. |
577 | /// @param Order Physregs to try. |
578 | /// @return Physreg to assign VirtReg, or 0. |
579 | MCRegister RAGreedy::tryEvict(const LiveInterval &VirtReg, |
580 | AllocationOrder &Order, |
581 | SmallVectorImpl<Register> &NewVRegs, |
582 | uint8_t CostPerUseLimit, |
583 | const SmallVirtRegSet &FixedRegisters) { |
584 | NamedRegionTimer T("evict" , "Evict" , TimerGroupName, TimerGroupDescription, |
585 | TimePassesIsEnabled); |
586 | |
587 | MCRegister BestPhys = EvictAdvisor->tryFindEvictionCandidate( |
588 | VirtReg, Order, CostPerUseLimit, FixedRegisters); |
589 | if (BestPhys.isValid()) |
590 | evictInterference(VirtReg, PhysReg: BestPhys, NewVRegs); |
591 | return BestPhys; |
592 | } |
593 | |
594 | //===----------------------------------------------------------------------===// |
595 | // Region Splitting |
596 | //===----------------------------------------------------------------------===// |
597 | |
598 | /// addSplitConstraints - Fill out the SplitConstraints vector based on the |
599 | /// interference pattern in Physreg and its aliases. Add the constraints to |
600 | /// SpillPlacement and return the static cost of this split in Cost, assuming |
601 | /// that all preferences in SplitConstraints are met. |
602 | /// Return false if there are no bundles with positive bias. |
603 | bool RAGreedy::addSplitConstraints(InterferenceCache::Cursor Intf, |
604 | BlockFrequency &Cost) { |
605 | ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks(); |
606 | |
607 | // Reset interference dependent info. |
608 | SplitConstraints.resize(N: UseBlocks.size()); |
609 | BlockFrequency StaticCost = BlockFrequency(0); |
610 | for (unsigned I = 0; I != UseBlocks.size(); ++I) { |
611 | const SplitAnalysis::BlockInfo &BI = UseBlocks[I]; |
612 | SpillPlacement::BlockConstraint &BC = SplitConstraints[I]; |
613 | |
614 | BC.Number = BI.MBB->getNumber(); |
615 | Intf.moveToBlock(MBBNum: BC.Number); |
616 | BC.Entry = BI.LiveIn ? SpillPlacement::PrefReg : SpillPlacement::DontCare; |
617 | BC.Exit = (BI.LiveOut && |
618 | !LIS->getInstructionFromIndex(index: BI.LastInstr)->isImplicitDef()) |
619 | ? SpillPlacement::PrefReg |
620 | : SpillPlacement::DontCare; |
621 | BC.ChangesValue = BI.FirstDef.isValid(); |
622 | |
623 | if (!Intf.hasInterference()) |
624 | continue; |
625 | |
626 | // Number of spill code instructions to insert. |
627 | unsigned Ins = 0; |
628 | |
629 | // Interference for the live-in value. |
630 | if (BI.LiveIn) { |
631 | if (Intf.first() <= Indexes->getMBBStartIdx(Num: BC.Number)) { |
632 | BC.Entry = SpillPlacement::MustSpill; |
633 | ++Ins; |
634 | } else if (Intf.first() < BI.FirstInstr) { |
635 | BC.Entry = SpillPlacement::PrefSpill; |
636 | ++Ins; |
637 | } else if (Intf.first() < BI.LastInstr) { |
638 | ++Ins; |
639 | } |
640 | |
641 | // Abort if the spill cannot be inserted at the MBB' start |
642 | if (((BC.Entry == SpillPlacement::MustSpill) || |
643 | (BC.Entry == SpillPlacement::PrefSpill)) && |
644 | SlotIndex::isEarlierInstr(A: BI.FirstInstr, |
645 | B: SA->getFirstSplitPoint(Num: BC.Number))) |
646 | return false; |
647 | } |
648 | |
649 | // Interference for the live-out value. |
650 | if (BI.LiveOut) { |
651 | if (Intf.last() >= SA->getLastSplitPoint(Num: BC.Number)) { |
652 | BC.Exit = SpillPlacement::MustSpill; |
653 | ++Ins; |
654 | } else if (Intf.last() > BI.LastInstr) { |
655 | BC.Exit = SpillPlacement::PrefSpill; |
656 | ++Ins; |
657 | } else if (Intf.last() > BI.FirstInstr) { |
658 | ++Ins; |
659 | } |
660 | } |
661 | |
662 | // Accumulate the total frequency of inserted spill code. |
663 | while (Ins--) |
664 | StaticCost += SpillPlacer->getBlockFrequency(Number: BC.Number); |
665 | } |
666 | Cost = StaticCost; |
667 | |
668 | // Add constraints for use-blocks. Note that these are the only constraints |
669 | // that may add a positive bias, it is downhill from here. |
670 | SpillPlacer->addConstraints(LiveBlocks: SplitConstraints); |
671 | return SpillPlacer->scanActiveBundles(); |
672 | } |
673 | |
674 | /// addThroughConstraints - Add constraints and links to SpillPlacer from the |
675 | /// live-through blocks in Blocks. |
676 | bool RAGreedy::addThroughConstraints(InterferenceCache::Cursor Intf, |
677 | ArrayRef<unsigned> Blocks) { |
678 | const unsigned GroupSize = 8; |
679 | SpillPlacement::BlockConstraint BCS[GroupSize]; |
680 | unsigned TBS[GroupSize]; |
681 | unsigned B = 0, T = 0; |
682 | |
683 | for (unsigned Number : Blocks) { |
684 | Intf.moveToBlock(MBBNum: Number); |
685 | |
686 | if (!Intf.hasInterference()) { |
687 | assert(T < GroupSize && "Array overflow" ); |
688 | TBS[T] = Number; |
689 | if (++T == GroupSize) { |
690 | SpillPlacer->addLinks(Links: ArrayRef(TBS, T)); |
691 | T = 0; |
692 | } |
693 | continue; |
694 | } |
695 | |
696 | assert(B < GroupSize && "Array overflow" ); |
697 | BCS[B].Number = Number; |
698 | |
699 | // Abort if the spill cannot be inserted at the MBB' start |
700 | MachineBasicBlock *MBB = MF->getBlockNumbered(N: Number); |
701 | auto FirstNonDebugInstr = MBB->getFirstNonDebugInstr(); |
702 | if (FirstNonDebugInstr != MBB->end() && |
703 | SlotIndex::isEarlierInstr(A: LIS->getInstructionIndex(Instr: *FirstNonDebugInstr), |
704 | B: SA->getFirstSplitPoint(Num: Number))) |
705 | return false; |
706 | // Interference for the live-in value. |
707 | if (Intf.first() <= Indexes->getMBBStartIdx(Num: Number)) |
708 | BCS[B].Entry = SpillPlacement::MustSpill; |
709 | else |
710 | BCS[B].Entry = SpillPlacement::PrefSpill; |
711 | |
712 | // Interference for the live-out value. |
713 | if (Intf.last() >= SA->getLastSplitPoint(Num: Number)) |
714 | BCS[B].Exit = SpillPlacement::MustSpill; |
715 | else |
716 | BCS[B].Exit = SpillPlacement::PrefSpill; |
717 | |
718 | if (++B == GroupSize) { |
719 | SpillPlacer->addConstraints(LiveBlocks: ArrayRef(BCS, B)); |
720 | B = 0; |
721 | } |
722 | } |
723 | |
724 | SpillPlacer->addConstraints(LiveBlocks: ArrayRef(BCS, B)); |
725 | SpillPlacer->addLinks(Links: ArrayRef(TBS, T)); |
726 | return true; |
727 | } |
728 | |
729 | bool RAGreedy::growRegion(GlobalSplitCandidate &Cand) { |
730 | // Keep track of through blocks that have not been added to SpillPlacer. |
731 | BitVector Todo = SA->getThroughBlocks(); |
732 | SmallVectorImpl<unsigned> &ActiveBlocks = Cand.ActiveBlocks; |
733 | unsigned AddedTo = 0; |
734 | #ifndef NDEBUG |
735 | unsigned Visited = 0; |
736 | #endif |
737 | |
738 | unsigned long Budget = GrowRegionComplexityBudget; |
739 | while (true) { |
740 | ArrayRef<unsigned> NewBundles = SpillPlacer->getRecentPositive(); |
741 | // Find new through blocks in the periphery of PrefRegBundles. |
742 | for (unsigned Bundle : NewBundles) { |
743 | // Look at all blocks connected to Bundle in the full graph. |
744 | ArrayRef<unsigned> Blocks = Bundles->getBlocks(Bundle); |
745 | // Limit compilation time by bailing out after we use all our budget. |
746 | if (Blocks.size() >= Budget) |
747 | return false; |
748 | Budget -= Blocks.size(); |
749 | for (unsigned Block : Blocks) { |
750 | if (!Todo.test(Idx: Block)) |
751 | continue; |
752 | Todo.reset(Idx: Block); |
753 | // This is a new through block. Add it to SpillPlacer later. |
754 | ActiveBlocks.push_back(Elt: Block); |
755 | #ifndef NDEBUG |
756 | ++Visited; |
757 | #endif |
758 | } |
759 | } |
760 | // Any new blocks to add? |
761 | if (ActiveBlocks.size() == AddedTo) |
762 | break; |
763 | |
764 | // Compute through constraints from the interference, or assume that all |
765 | // through blocks prefer spilling when forming compact regions. |
766 | auto NewBlocks = ArrayRef(ActiveBlocks).slice(N: AddedTo); |
767 | if (Cand.PhysReg) { |
768 | if (!addThroughConstraints(Intf: Cand.Intf, Blocks: NewBlocks)) |
769 | return false; |
770 | } else { |
771 | // Providing that the variable being spilled does not look like a loop |
772 | // induction variable, which is expensive to spill around and better |
773 | // pushed into a condition inside the loop if possible, provide a strong |
774 | // negative bias on through blocks to prevent unwanted liveness on loop |
775 | // backedges. |
776 | bool PrefSpill = true; |
777 | if (SA->looksLikeLoopIV() && NewBlocks.size() >= 2) { |
778 | // Check that the current bundle is adding a Header + start+end of |
779 | // loop-internal blocks. If the block is indeed a header, don't make |
780 | // the NewBlocks as PrefSpill to allow the variable to be live in |
781 | // Header<->Latch. |
782 | MachineLoop *L = Loops->getLoopFor(BB: MF->getBlockNumbered(N: NewBlocks[0])); |
783 | if (L && L->getHeader()->getNumber() == (int)NewBlocks[0] && |
784 | all_of(Range: NewBlocks.drop_front(), P: [&](unsigned Block) { |
785 | return L == Loops->getLoopFor(BB: MF->getBlockNumbered(N: Block)); |
786 | })) |
787 | PrefSpill = false; |
788 | } |
789 | if (PrefSpill) |
790 | SpillPlacer->addPrefSpill(Blocks: NewBlocks, /* Strong= */ true); |
791 | } |
792 | AddedTo = ActiveBlocks.size(); |
793 | |
794 | // Perhaps iterating can enable more bundles? |
795 | SpillPlacer->iterate(); |
796 | } |
797 | LLVM_DEBUG(dbgs() << ", v=" << Visited); |
798 | return true; |
799 | } |
800 | |
801 | /// calcCompactRegion - Compute the set of edge bundles that should be live |
802 | /// when splitting the current live range into compact regions. Compact |
803 | /// regions can be computed without looking at interference. They are the |
804 | /// regions formed by removing all the live-through blocks from the live range. |
805 | /// |
806 | /// Returns false if the current live range is already compact, or if the |
807 | /// compact regions would form single block regions anyway. |
808 | bool RAGreedy::calcCompactRegion(GlobalSplitCandidate &Cand) { |
809 | // Without any through blocks, the live range is already compact. |
810 | if (!SA->getNumThroughBlocks()) |
811 | return false; |
812 | |
813 | // Compact regions don't correspond to any physreg. |
814 | Cand.reset(Cache&: IntfCache, Reg: MCRegister::NoRegister); |
815 | |
816 | LLVM_DEBUG(dbgs() << "Compact region bundles" ); |
817 | |
818 | // Use the spill placer to determine the live bundles. GrowRegion pretends |
819 | // that all the through blocks have interference when PhysReg is unset. |
820 | SpillPlacer->prepare(RegBundles&: Cand.LiveBundles); |
821 | |
822 | // The static split cost will be zero since Cand.Intf reports no interference. |
823 | BlockFrequency Cost; |
824 | if (!addSplitConstraints(Intf: Cand.Intf, Cost)) { |
825 | LLVM_DEBUG(dbgs() << ", none.\n" ); |
826 | return false; |
827 | } |
828 | |
829 | if (!growRegion(Cand)) { |
830 | LLVM_DEBUG(dbgs() << ", cannot spill all interferences.\n" ); |
831 | return false; |
832 | } |
833 | |
834 | SpillPlacer->finish(); |
835 | |
836 | if (!Cand.LiveBundles.any()) { |
837 | LLVM_DEBUG(dbgs() << ", none.\n" ); |
838 | return false; |
839 | } |
840 | |
841 | LLVM_DEBUG({ |
842 | for (int I : Cand.LiveBundles.set_bits()) |
843 | dbgs() << " EB#" << I; |
844 | dbgs() << ".\n" ; |
845 | }); |
846 | return true; |
847 | } |
848 | |
849 | /// calcSpillCost - Compute how expensive it would be to split the live range in |
850 | /// SA around all use blocks instead of forming bundle regions. |
851 | BlockFrequency RAGreedy::calcSpillCost() { |
852 | BlockFrequency Cost = BlockFrequency(0); |
853 | ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks(); |
854 | for (const SplitAnalysis::BlockInfo &BI : UseBlocks) { |
855 | unsigned Number = BI.MBB->getNumber(); |
856 | // We normally only need one spill instruction - a load or a store. |
857 | Cost += SpillPlacer->getBlockFrequency(Number); |
858 | |
859 | // Unless the value is redefined in the block. |
860 | if (BI.LiveIn && BI.LiveOut && BI.FirstDef) |
861 | Cost += SpillPlacer->getBlockFrequency(Number); |
862 | } |
863 | return Cost; |
864 | } |
865 | |
866 | /// calcGlobalSplitCost - Return the global split cost of following the split |
867 | /// pattern in LiveBundles. This cost should be added to the local cost of the |
868 | /// interference pattern in SplitConstraints. |
869 | /// |
870 | BlockFrequency RAGreedy::calcGlobalSplitCost(GlobalSplitCandidate &Cand, |
871 | const AllocationOrder &Order) { |
872 | BlockFrequency GlobalCost = BlockFrequency(0); |
873 | const BitVector &LiveBundles = Cand.LiveBundles; |
874 | ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks(); |
875 | for (unsigned I = 0; I != UseBlocks.size(); ++I) { |
876 | const SplitAnalysis::BlockInfo &BI = UseBlocks[I]; |
877 | SpillPlacement::BlockConstraint &BC = SplitConstraints[I]; |
878 | bool RegIn = LiveBundles[Bundles->getBundle(N: BC.Number, Out: false)]; |
879 | bool RegOut = LiveBundles[Bundles->getBundle(N: BC.Number, Out: true)]; |
880 | unsigned Ins = 0; |
881 | |
882 | Cand.Intf.moveToBlock(MBBNum: BC.Number); |
883 | |
884 | if (BI.LiveIn) |
885 | Ins += RegIn != (BC.Entry == SpillPlacement::PrefReg); |
886 | if (BI.LiveOut) |
887 | Ins += RegOut != (BC.Exit == SpillPlacement::PrefReg); |
888 | while (Ins--) |
889 | GlobalCost += SpillPlacer->getBlockFrequency(Number: BC.Number); |
890 | } |
891 | |
892 | for (unsigned Number : Cand.ActiveBlocks) { |
893 | bool RegIn = LiveBundles[Bundles->getBundle(N: Number, Out: false)]; |
894 | bool RegOut = LiveBundles[Bundles->getBundle(N: Number, Out: true)]; |
895 | if (!RegIn && !RegOut) |
896 | continue; |
897 | if (RegIn && RegOut) { |
898 | // We need double spill code if this block has interference. |
899 | Cand.Intf.moveToBlock(MBBNum: Number); |
900 | if (Cand.Intf.hasInterference()) { |
901 | GlobalCost += SpillPlacer->getBlockFrequency(Number); |
902 | GlobalCost += SpillPlacer->getBlockFrequency(Number); |
903 | } |
904 | continue; |
905 | } |
906 | // live-in / stack-out or stack-in live-out. |
907 | GlobalCost += SpillPlacer->getBlockFrequency(Number); |
908 | } |
909 | return GlobalCost; |
910 | } |
911 | |
912 | /// splitAroundRegion - Split the current live range around the regions |
913 | /// determined by BundleCand and GlobalCand. |
914 | /// |
915 | /// Before calling this function, GlobalCand and BundleCand must be initialized |
916 | /// so each bundle is assigned to a valid candidate, or NoCand for the |
917 | /// stack-bound bundles. The shared SA/SE SplitAnalysis and SplitEditor |
918 | /// objects must be initialized for the current live range, and intervals |
919 | /// created for the used candidates. |
920 | /// |
921 | /// @param LREdit The LiveRangeEdit object handling the current split. |
922 | /// @param UsedCands List of used GlobalCand entries. Every BundleCand value |
923 | /// must appear in this list. |
924 | void RAGreedy::splitAroundRegion(LiveRangeEdit &LREdit, |
925 | ArrayRef<unsigned> UsedCands) { |
926 | // These are the intervals created for new global ranges. We may create more |
927 | // intervals for local ranges. |
928 | const unsigned NumGlobalIntvs = LREdit.size(); |
929 | LLVM_DEBUG(dbgs() << "splitAroundRegion with " << NumGlobalIntvs |
930 | << " globals.\n" ); |
931 | assert(NumGlobalIntvs && "No global intervals configured" ); |
932 | |
933 | // Isolate even single instructions when dealing with a proper sub-class. |
934 | // That guarantees register class inflation for the stack interval because it |
935 | // is all copies. |
936 | Register Reg = SA->getParent().reg(); |
937 | bool SingleInstrs = RegClassInfo.isProperSubClass(RC: MRI->getRegClass(Reg)); |
938 | |
939 | // First handle all the blocks with uses. |
940 | ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks(); |
941 | for (const SplitAnalysis::BlockInfo &BI : UseBlocks) { |
942 | unsigned Number = BI.MBB->getNumber(); |
943 | unsigned IntvIn = 0, IntvOut = 0; |
944 | SlotIndex IntfIn, IntfOut; |
945 | if (BI.LiveIn) { |
946 | unsigned CandIn = BundleCand[Bundles->getBundle(N: Number, Out: false)]; |
947 | if (CandIn != NoCand) { |
948 | GlobalSplitCandidate &Cand = GlobalCand[CandIn]; |
949 | IntvIn = Cand.IntvIdx; |
950 | Cand.Intf.moveToBlock(MBBNum: Number); |
951 | IntfIn = Cand.Intf.first(); |
952 | } |
953 | } |
954 | if (BI.LiveOut) { |
955 | unsigned CandOut = BundleCand[Bundles->getBundle(N: Number, Out: true)]; |
956 | if (CandOut != NoCand) { |
957 | GlobalSplitCandidate &Cand = GlobalCand[CandOut]; |
958 | IntvOut = Cand.IntvIdx; |
959 | Cand.Intf.moveToBlock(MBBNum: Number); |
960 | IntfOut = Cand.Intf.last(); |
961 | } |
962 | } |
963 | |
964 | // Create separate intervals for isolated blocks with multiple uses. |
965 | if (!IntvIn && !IntvOut) { |
966 | LLVM_DEBUG(dbgs() << printMBBReference(*BI.MBB) << " isolated.\n" ); |
967 | if (SA->shouldSplitSingleBlock(BI, SingleInstrs)) |
968 | SE->splitSingleBlock(BI); |
969 | continue; |
970 | } |
971 | |
972 | if (IntvIn && IntvOut) |
973 | SE->splitLiveThroughBlock(MBBNum: Number, IntvIn, LeaveBefore: IntfIn, IntvOut, EnterAfter: IntfOut); |
974 | else if (IntvIn) |
975 | SE->splitRegInBlock(BI, IntvIn, LeaveBefore: IntfIn); |
976 | else |
977 | SE->splitRegOutBlock(BI, IntvOut, EnterAfter: IntfOut); |
978 | } |
979 | |
980 | // Handle live-through blocks. The relevant live-through blocks are stored in |
981 | // the ActiveBlocks list with each candidate. We need to filter out |
982 | // duplicates. |
983 | BitVector Todo = SA->getThroughBlocks(); |
984 | for (unsigned UsedCand : UsedCands) { |
985 | ArrayRef<unsigned> Blocks = GlobalCand[UsedCand].ActiveBlocks; |
986 | for (unsigned Number : Blocks) { |
987 | if (!Todo.test(Idx: Number)) |
988 | continue; |
989 | Todo.reset(Idx: Number); |
990 | |
991 | unsigned IntvIn = 0, IntvOut = 0; |
992 | SlotIndex IntfIn, IntfOut; |
993 | |
994 | unsigned CandIn = BundleCand[Bundles->getBundle(N: Number, Out: false)]; |
995 | if (CandIn != NoCand) { |
996 | GlobalSplitCandidate &Cand = GlobalCand[CandIn]; |
997 | IntvIn = Cand.IntvIdx; |
998 | Cand.Intf.moveToBlock(MBBNum: Number); |
999 | IntfIn = Cand.Intf.first(); |
1000 | } |
1001 | |
1002 | unsigned CandOut = BundleCand[Bundles->getBundle(N: Number, Out: true)]; |
1003 | if (CandOut != NoCand) { |
1004 | GlobalSplitCandidate &Cand = GlobalCand[CandOut]; |
1005 | IntvOut = Cand.IntvIdx; |
1006 | Cand.Intf.moveToBlock(MBBNum: Number); |
1007 | IntfOut = Cand.Intf.last(); |
1008 | } |
1009 | if (!IntvIn && !IntvOut) |
1010 | continue; |
1011 | SE->splitLiveThroughBlock(MBBNum: Number, IntvIn, LeaveBefore: IntfIn, IntvOut, EnterAfter: IntfOut); |
1012 | } |
1013 | } |
1014 | |
1015 | ++NumGlobalSplits; |
1016 | |
1017 | SmallVector<unsigned, 8> IntvMap; |
1018 | SE->finish(LRMap: &IntvMap); |
1019 | DebugVars->splitRegister(OldReg: Reg, NewRegs: LREdit.regs(), LIS&: *LIS); |
1020 | |
1021 | unsigned OrigBlocks = SA->getNumLiveBlocks(); |
1022 | |
1023 | // Sort out the new intervals created by splitting. We get four kinds: |
1024 | // - Remainder intervals should not be split again. |
1025 | // - Candidate intervals can be assigned to Cand.PhysReg. |
1026 | // - Block-local splits are candidates for local splitting. |
1027 | // - DCE leftovers should go back on the queue. |
1028 | for (unsigned I = 0, E = LREdit.size(); I != E; ++I) { |
1029 | const LiveInterval &Reg = LIS->getInterval(Reg: LREdit.get(idx: I)); |
1030 | |
1031 | // Ignore old intervals from DCE. |
1032 | if (ExtraInfo->getOrInitStage(Reg: Reg.reg()) != RS_New) |
1033 | continue; |
1034 | |
1035 | // Remainder interval. Don't try splitting again, spill if it doesn't |
1036 | // allocate. |
1037 | if (IntvMap[I] == 0) { |
1038 | ExtraInfo->setStage(VirtReg: Reg, Stage: RS_Spill); |
1039 | continue; |
1040 | } |
1041 | |
1042 | // Global intervals. Allow repeated splitting as long as the number of live |
1043 | // blocks is strictly decreasing. |
1044 | if (IntvMap[I] < NumGlobalIntvs) { |
1045 | if (SA->countLiveBlocks(li: &Reg) >= OrigBlocks) { |
1046 | LLVM_DEBUG(dbgs() << "Main interval covers the same " << OrigBlocks |
1047 | << " blocks as original.\n" ); |
1048 | // Don't allow repeated splitting as a safe guard against looping. |
1049 | ExtraInfo->setStage(VirtReg: Reg, Stage: RS_Split2); |
1050 | } |
1051 | continue; |
1052 | } |
1053 | |
1054 | // Other intervals are treated as new. This includes local intervals created |
1055 | // for blocks with multiple uses, and anything created by DCE. |
1056 | } |
1057 | |
1058 | if (VerifyEnabled) |
1059 | MF->verify(p: this, Banner: "After splitting live range around region" ); |
1060 | } |
1061 | |
1062 | MCRegister RAGreedy::tryRegionSplit(const LiveInterval &VirtReg, |
1063 | AllocationOrder &Order, |
1064 | SmallVectorImpl<Register> &NewVRegs) { |
1065 | if (!TRI->shouldRegionSplitForVirtReg(MF: *MF, VirtReg)) |
1066 | return MCRegister::NoRegister; |
1067 | unsigned NumCands = 0; |
1068 | BlockFrequency SpillCost = calcSpillCost(); |
1069 | BlockFrequency BestCost; |
1070 | |
1071 | // Check if we can split this live range around a compact region. |
1072 | bool HasCompact = calcCompactRegion(Cand&: GlobalCand.front()); |
1073 | if (HasCompact) { |
1074 | // Yes, keep GlobalCand[0] as the compact region candidate. |
1075 | NumCands = 1; |
1076 | BestCost = BlockFrequency::max(); |
1077 | } else { |
1078 | // No benefit from the compact region, our fallback will be per-block |
1079 | // splitting. Make sure we find a solution that is cheaper than spilling. |
1080 | BestCost = SpillCost; |
1081 | LLVM_DEBUG(dbgs() << "Cost of isolating all blocks = " |
1082 | << printBlockFreq(*MBFI, BestCost) << '\n'); |
1083 | } |
1084 | |
1085 | unsigned BestCand = calculateRegionSplitCost(VirtReg, Order, BestCost, |
1086 | NumCands, IgnoreCSR: false /*IgnoreCSR*/); |
1087 | |
1088 | // No solutions found, fall back to single block splitting. |
1089 | if (!HasCompact && BestCand == NoCand) |
1090 | return MCRegister::NoRegister; |
1091 | |
1092 | return doRegionSplit(VirtReg, BestCand, HasCompact, NewVRegs); |
1093 | } |
1094 | |
1095 | unsigned |
1096 | RAGreedy::calculateRegionSplitCostAroundReg(MCPhysReg PhysReg, |
1097 | AllocationOrder &Order, |
1098 | BlockFrequency &BestCost, |
1099 | unsigned &NumCands, |
1100 | unsigned &BestCand) { |
1101 | // Discard bad candidates before we run out of interference cache cursors. |
1102 | // This will only affect register classes with a lot of registers (>32). |
1103 | if (NumCands == IntfCache.getMaxCursors()) { |
1104 | unsigned WorstCount = ~0u; |
1105 | unsigned Worst = 0; |
1106 | for (unsigned CandIndex = 0; CandIndex != NumCands; ++CandIndex) { |
1107 | if (CandIndex == BestCand || !GlobalCand[CandIndex].PhysReg) |
1108 | continue; |
1109 | unsigned Count = GlobalCand[CandIndex].LiveBundles.count(); |
1110 | if (Count < WorstCount) { |
1111 | Worst = CandIndex; |
1112 | WorstCount = Count; |
1113 | } |
1114 | } |
1115 | --NumCands; |
1116 | GlobalCand[Worst] = GlobalCand[NumCands]; |
1117 | if (BestCand == NumCands) |
1118 | BestCand = Worst; |
1119 | } |
1120 | |
1121 | if (GlobalCand.size() <= NumCands) |
1122 | GlobalCand.resize(N: NumCands+1); |
1123 | GlobalSplitCandidate &Cand = GlobalCand[NumCands]; |
1124 | Cand.reset(Cache&: IntfCache, Reg: PhysReg); |
1125 | |
1126 | SpillPlacer->prepare(RegBundles&: Cand.LiveBundles); |
1127 | BlockFrequency Cost; |
1128 | if (!addSplitConstraints(Intf: Cand.Intf, Cost)) { |
1129 | LLVM_DEBUG(dbgs() << printReg(PhysReg, TRI) << "\tno positive bundles\n" ); |
1130 | return BestCand; |
1131 | } |
1132 | LLVM_DEBUG(dbgs() << printReg(PhysReg, TRI) |
1133 | << "\tstatic = " << printBlockFreq(*MBFI, Cost)); |
1134 | if (Cost >= BestCost) { |
1135 | LLVM_DEBUG({ |
1136 | if (BestCand == NoCand) |
1137 | dbgs() << " worse than no bundles\n" ; |
1138 | else |
1139 | dbgs() << " worse than " |
1140 | << printReg(GlobalCand[BestCand].PhysReg, TRI) << '\n'; |
1141 | }); |
1142 | return BestCand; |
1143 | } |
1144 | if (!growRegion(Cand)) { |
1145 | LLVM_DEBUG(dbgs() << ", cannot spill all interferences.\n" ); |
1146 | return BestCand; |
1147 | } |
1148 | |
1149 | SpillPlacer->finish(); |
1150 | |
1151 | // No live bundles, defer to splitSingleBlocks(). |
1152 | if (!Cand.LiveBundles.any()) { |
1153 | LLVM_DEBUG(dbgs() << " no bundles.\n" ); |
1154 | return BestCand; |
1155 | } |
1156 | |
1157 | Cost += calcGlobalSplitCost(Cand, Order); |
1158 | LLVM_DEBUG({ |
1159 | dbgs() << ", total = " << printBlockFreq(*MBFI, Cost) << " with bundles" ; |
1160 | for (int I : Cand.LiveBundles.set_bits()) |
1161 | dbgs() << " EB#" << I; |
1162 | dbgs() << ".\n" ; |
1163 | }); |
1164 | if (Cost < BestCost) { |
1165 | BestCand = NumCands; |
1166 | BestCost = Cost; |
1167 | } |
1168 | ++NumCands; |
1169 | |
1170 | return BestCand; |
1171 | } |
1172 | |
1173 | unsigned RAGreedy::calculateRegionSplitCost(const LiveInterval &VirtReg, |
1174 | AllocationOrder &Order, |
1175 | BlockFrequency &BestCost, |
1176 | unsigned &NumCands, |
1177 | bool IgnoreCSR) { |
1178 | unsigned BestCand = NoCand; |
1179 | for (MCPhysReg PhysReg : Order) { |
1180 | assert(PhysReg); |
1181 | if (IgnoreCSR && EvictAdvisor->isUnusedCalleeSavedReg(PhysReg)) |
1182 | continue; |
1183 | |
1184 | calculateRegionSplitCostAroundReg(PhysReg, Order, BestCost, NumCands, |
1185 | BestCand); |
1186 | } |
1187 | |
1188 | return BestCand; |
1189 | } |
1190 | |
1191 | unsigned RAGreedy::doRegionSplit(const LiveInterval &VirtReg, unsigned BestCand, |
1192 | bool HasCompact, |
1193 | SmallVectorImpl<Register> &NewVRegs) { |
1194 | SmallVector<unsigned, 8> UsedCands; |
1195 | // Prepare split editor. |
1196 | LiveRangeEdit LREdit(&VirtReg, NewVRegs, *MF, *LIS, VRM, this, &DeadRemats); |
1197 | SE->reset(LREdit, SplitSpillMode); |
1198 | |
1199 | // Assign all edge bundles to the preferred candidate, or NoCand. |
1200 | BundleCand.assign(NumElts: Bundles->getNumBundles(), Elt: NoCand); |
1201 | |
1202 | // Assign bundles for the best candidate region. |
1203 | if (BestCand != NoCand) { |
1204 | GlobalSplitCandidate &Cand = GlobalCand[BestCand]; |
1205 | if (unsigned B = Cand.getBundles(B&: BundleCand, C: BestCand)) { |
1206 | UsedCands.push_back(Elt: BestCand); |
1207 | Cand.IntvIdx = SE->openIntv(); |
1208 | LLVM_DEBUG(dbgs() << "Split for " << printReg(Cand.PhysReg, TRI) << " in " |
1209 | << B << " bundles, intv " << Cand.IntvIdx << ".\n" ); |
1210 | (void)B; |
1211 | } |
1212 | } |
1213 | |
1214 | // Assign bundles for the compact region. |
1215 | if (HasCompact) { |
1216 | GlobalSplitCandidate &Cand = GlobalCand.front(); |
1217 | assert(!Cand.PhysReg && "Compact region has no physreg" ); |
1218 | if (unsigned B = Cand.getBundles(B&: BundleCand, C: 0)) { |
1219 | UsedCands.push_back(Elt: 0); |
1220 | Cand.IntvIdx = SE->openIntv(); |
1221 | LLVM_DEBUG(dbgs() << "Split for compact region in " << B |
1222 | << " bundles, intv " << Cand.IntvIdx << ".\n" ); |
1223 | (void)B; |
1224 | } |
1225 | } |
1226 | |
1227 | splitAroundRegion(LREdit, UsedCands); |
1228 | return 0; |
1229 | } |
1230 | |
1231 | // VirtReg has a physical Hint, this function tries to split VirtReg around |
1232 | // Hint if we can place new COPY instructions in cold blocks. |
1233 | bool RAGreedy::trySplitAroundHintReg(MCPhysReg Hint, |
1234 | const LiveInterval &VirtReg, |
1235 | SmallVectorImpl<Register> &NewVRegs, |
1236 | AllocationOrder &Order) { |
1237 | // Split the VirtReg may generate COPY instructions in multiple cold basic |
1238 | // blocks, and increase code size. So we avoid it when the function is |
1239 | // optimized for size. |
1240 | if (MF->getFunction().hasOptSize()) |
1241 | return false; |
1242 | |
1243 | // Don't allow repeated splitting as a safe guard against looping. |
1244 | if (ExtraInfo->getStage(VirtReg) >= RS_Split2) |
1245 | return false; |
1246 | |
1247 | BlockFrequency Cost = BlockFrequency(0); |
1248 | Register Reg = VirtReg.reg(); |
1249 | |
1250 | // Compute the cost of assigning a non Hint physical register to VirtReg. |
1251 | // We define it as the total frequency of broken COPY instructions to/from |
1252 | // Hint register, and after split, they can be deleted. |
1253 | for (const MachineInstr &Instr : MRI->reg_nodbg_instructions(Reg)) { |
1254 | if (!TII->isFullCopyInstr(MI: Instr)) |
1255 | continue; |
1256 | Register OtherReg = Instr.getOperand(i: 1).getReg(); |
1257 | if (OtherReg == Reg) { |
1258 | OtherReg = Instr.getOperand(i: 0).getReg(); |
1259 | if (OtherReg == Reg) |
1260 | continue; |
1261 | // Check if VirtReg interferes with OtherReg after this COPY instruction. |
1262 | if (VirtReg.liveAt(index: LIS->getInstructionIndex(Instr).getRegSlot())) |
1263 | continue; |
1264 | } |
1265 | MCRegister OtherPhysReg = |
1266 | OtherReg.isPhysical() ? OtherReg.asMCReg() : VRM->getPhys(virtReg: OtherReg); |
1267 | if (OtherPhysReg == Hint) |
1268 | Cost += MBFI->getBlockFreq(MBB: Instr.getParent()); |
1269 | } |
1270 | |
1271 | // Decrease the cost so it will be split in colder blocks. |
1272 | BranchProbability Threshold(SplitThresholdForRegWithHint, 100); |
1273 | Cost *= Threshold; |
1274 | if (Cost == BlockFrequency(0)) |
1275 | return false; |
1276 | |
1277 | unsigned NumCands = 0; |
1278 | unsigned BestCand = NoCand; |
1279 | SA->analyze(li: &VirtReg); |
1280 | calculateRegionSplitCostAroundReg(PhysReg: Hint, Order, BestCost&: Cost, NumCands, BestCand); |
1281 | if (BestCand == NoCand) |
1282 | return false; |
1283 | |
1284 | doRegionSplit(VirtReg, BestCand, HasCompact: false/*HasCompact*/, NewVRegs); |
1285 | return true; |
1286 | } |
1287 | |
1288 | //===----------------------------------------------------------------------===// |
1289 | // Per-Block Splitting |
1290 | //===----------------------------------------------------------------------===// |
1291 | |
1292 | /// tryBlockSplit - Split a global live range around every block with uses. This |
1293 | /// creates a lot of local live ranges, that will be split by tryLocalSplit if |
1294 | /// they don't allocate. |
1295 | unsigned RAGreedy::tryBlockSplit(const LiveInterval &VirtReg, |
1296 | AllocationOrder &Order, |
1297 | SmallVectorImpl<Register> &NewVRegs) { |
1298 | assert(&SA->getParent() == &VirtReg && "Live range wasn't analyzed" ); |
1299 | Register Reg = VirtReg.reg(); |
1300 | bool SingleInstrs = RegClassInfo.isProperSubClass(RC: MRI->getRegClass(Reg)); |
1301 | LiveRangeEdit LREdit(&VirtReg, NewVRegs, *MF, *LIS, VRM, this, &DeadRemats); |
1302 | SE->reset(LREdit, SplitSpillMode); |
1303 | ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks(); |
1304 | for (const SplitAnalysis::BlockInfo &BI : UseBlocks) { |
1305 | if (SA->shouldSplitSingleBlock(BI, SingleInstrs)) |
1306 | SE->splitSingleBlock(BI); |
1307 | } |
1308 | // No blocks were split. |
1309 | if (LREdit.empty()) |
1310 | return 0; |
1311 | |
1312 | // We did split for some blocks. |
1313 | SmallVector<unsigned, 8> IntvMap; |
1314 | SE->finish(LRMap: &IntvMap); |
1315 | |
1316 | // Tell LiveDebugVariables about the new ranges. |
1317 | DebugVars->splitRegister(OldReg: Reg, NewRegs: LREdit.regs(), LIS&: *LIS); |
1318 | |
1319 | // Sort out the new intervals created by splitting. The remainder interval |
1320 | // goes straight to spilling, the new local ranges get to stay RS_New. |
1321 | for (unsigned I = 0, E = LREdit.size(); I != E; ++I) { |
1322 | const LiveInterval &LI = LIS->getInterval(Reg: LREdit.get(idx: I)); |
1323 | if (ExtraInfo->getOrInitStage(Reg: LI.reg()) == RS_New && IntvMap[I] == 0) |
1324 | ExtraInfo->setStage(VirtReg: LI, Stage: RS_Spill); |
1325 | } |
1326 | |
1327 | if (VerifyEnabled) |
1328 | MF->verify(p: this, Banner: "After splitting live range around basic blocks" ); |
1329 | return 0; |
1330 | } |
1331 | |
1332 | //===----------------------------------------------------------------------===// |
1333 | // Per-Instruction Splitting |
1334 | //===----------------------------------------------------------------------===// |
1335 | |
1336 | /// Get the number of allocatable registers that match the constraints of \p Reg |
1337 | /// on \p MI and that are also in \p SuperRC. |
1338 | static unsigned getNumAllocatableRegsForConstraints( |
1339 | const MachineInstr *MI, Register Reg, const TargetRegisterClass *SuperRC, |
1340 | const TargetInstrInfo *TII, const TargetRegisterInfo *TRI, |
1341 | const RegisterClassInfo &RCI) { |
1342 | assert(SuperRC && "Invalid register class" ); |
1343 | |
1344 | const TargetRegisterClass *ConstrainedRC = |
1345 | MI->getRegClassConstraintEffectForVReg(Reg, CurRC: SuperRC, TII, TRI, |
1346 | /* ExploreBundle */ true); |
1347 | if (!ConstrainedRC) |
1348 | return 0; |
1349 | return RCI.getNumAllocatableRegs(RC: ConstrainedRC); |
1350 | } |
1351 | |
1352 | static LaneBitmask getInstReadLaneMask(const MachineRegisterInfo &MRI, |
1353 | const TargetRegisterInfo &TRI, |
1354 | const MachineInstr &FirstMI, |
1355 | Register Reg) { |
1356 | LaneBitmask Mask; |
1357 | SmallVector<std::pair<MachineInstr *, unsigned>, 8> Ops; |
1358 | (void)AnalyzeVirtRegInBundle(MI&: const_cast<MachineInstr &>(FirstMI), Reg, Ops: &Ops); |
1359 | |
1360 | for (auto [MI, OpIdx] : Ops) { |
1361 | const MachineOperand &MO = MI->getOperand(i: OpIdx); |
1362 | assert(MO.isReg() && MO.getReg() == Reg); |
1363 | unsigned SubReg = MO.getSubReg(); |
1364 | if (SubReg == 0 && MO.isUse()) { |
1365 | if (MO.isUndef()) |
1366 | continue; |
1367 | return MRI.getMaxLaneMaskForVReg(Reg); |
1368 | } |
1369 | |
1370 | LaneBitmask SubRegMask = TRI.getSubRegIndexLaneMask(SubIdx: SubReg); |
1371 | if (MO.isDef()) { |
1372 | if (!MO.isUndef()) |
1373 | Mask |= ~SubRegMask; |
1374 | } else |
1375 | Mask |= SubRegMask; |
1376 | } |
1377 | |
1378 | return Mask; |
1379 | } |
1380 | |
1381 | /// Return true if \p MI at \P Use reads a subset of the lanes live in \p |
1382 | /// VirtReg. |
1383 | static bool readsLaneSubset(const MachineRegisterInfo &MRI, |
1384 | const MachineInstr *MI, const LiveInterval &VirtReg, |
1385 | const TargetRegisterInfo *TRI, SlotIndex Use, |
1386 | const TargetInstrInfo *TII) { |
1387 | // Early check the common case. Beware of the semi-formed bundles SplitKit |
1388 | // creates by setting the bundle flag on copies without a matching BUNDLE. |
1389 | |
1390 | auto DestSrc = TII->isCopyInstr(MI: *MI); |
1391 | if (DestSrc && !MI->isBundled() && |
1392 | DestSrc->Destination->getSubReg() == DestSrc->Source->getSubReg()) |
1393 | return false; |
1394 | |
1395 | // FIXME: We're only considering uses, but should be consider defs too? |
1396 | LaneBitmask ReadMask = getInstReadLaneMask(MRI, TRI: *TRI, FirstMI: *MI, Reg: VirtReg.reg()); |
1397 | |
1398 | LaneBitmask LiveAtMask; |
1399 | for (const LiveInterval::SubRange &S : VirtReg.subranges()) { |
1400 | if (S.liveAt(index: Use)) |
1401 | LiveAtMask |= S.LaneMask; |
1402 | } |
1403 | |
1404 | // If the live lanes aren't different from the lanes used by the instruction, |
1405 | // this doesn't help. |
1406 | return (ReadMask & ~(LiveAtMask & TRI->getCoveringLanes())).any(); |
1407 | } |
1408 | |
1409 | /// tryInstructionSplit - Split a live range around individual instructions. |
1410 | /// This is normally not worthwhile since the spiller is doing essentially the |
1411 | /// same thing. However, when the live range is in a constrained register |
1412 | /// class, it may help to insert copies such that parts of the live range can |
1413 | /// be moved to a larger register class. |
1414 | /// |
1415 | /// This is similar to spilling to a larger register class. |
1416 | unsigned RAGreedy::tryInstructionSplit(const LiveInterval &VirtReg, |
1417 | AllocationOrder &Order, |
1418 | SmallVectorImpl<Register> &NewVRegs) { |
1419 | const TargetRegisterClass *CurRC = MRI->getRegClass(Reg: VirtReg.reg()); |
1420 | // There is no point to this if there are no larger sub-classes. |
1421 | |
1422 | bool SplitSubClass = true; |
1423 | if (!RegClassInfo.isProperSubClass(RC: CurRC)) { |
1424 | if (!VirtReg.hasSubRanges()) |
1425 | return 0; |
1426 | SplitSubClass = false; |
1427 | } |
1428 | |
1429 | // Always enable split spill mode, since we're effectively spilling to a |
1430 | // register. |
1431 | LiveRangeEdit LREdit(&VirtReg, NewVRegs, *MF, *LIS, VRM, this, &DeadRemats); |
1432 | SE->reset(LREdit, SplitEditor::SM_Size); |
1433 | |
1434 | ArrayRef<SlotIndex> Uses = SA->getUseSlots(); |
1435 | if (Uses.size() <= 1) |
1436 | return 0; |
1437 | |
1438 | LLVM_DEBUG(dbgs() << "Split around " << Uses.size() |
1439 | << " individual instrs.\n" ); |
1440 | |
1441 | const TargetRegisterClass *SuperRC = |
1442 | TRI->getLargestLegalSuperClass(RC: CurRC, *MF); |
1443 | unsigned SuperRCNumAllocatableRegs = |
1444 | RegClassInfo.getNumAllocatableRegs(RC: SuperRC); |
1445 | // Split around every non-copy instruction if this split will relax |
1446 | // the constraints on the virtual register. |
1447 | // Otherwise, splitting just inserts uncoalescable copies that do not help |
1448 | // the allocation. |
1449 | for (const SlotIndex Use : Uses) { |
1450 | if (const MachineInstr *MI = Indexes->getInstructionFromIndex(index: Use)) { |
1451 | if (TII->isFullCopyInstr(MI: *MI) || |
1452 | (SplitSubClass && |
1453 | SuperRCNumAllocatableRegs == |
1454 | getNumAllocatableRegsForConstraints(MI, Reg: VirtReg.reg(), SuperRC, |
1455 | TII, TRI, RCI: RegClassInfo)) || |
1456 | // TODO: Handle split for subranges with subclass constraints? |
1457 | (!SplitSubClass && VirtReg.hasSubRanges() && |
1458 | !readsLaneSubset(MRI: *MRI, MI, VirtReg, TRI, Use, TII))) { |
1459 | LLVM_DEBUG(dbgs() << " skip:\t" << Use << '\t' << *MI); |
1460 | continue; |
1461 | } |
1462 | } |
1463 | SE->openIntv(); |
1464 | SlotIndex SegStart = SE->enterIntvBefore(Idx: Use); |
1465 | SlotIndex SegStop = SE->leaveIntvAfter(Idx: Use); |
1466 | SE->useIntv(Start: SegStart, End: SegStop); |
1467 | } |
1468 | |
1469 | if (LREdit.empty()) { |
1470 | LLVM_DEBUG(dbgs() << "All uses were copies.\n" ); |
1471 | return 0; |
1472 | } |
1473 | |
1474 | SmallVector<unsigned, 8> IntvMap; |
1475 | SE->finish(LRMap: &IntvMap); |
1476 | DebugVars->splitRegister(OldReg: VirtReg.reg(), NewRegs: LREdit.regs(), LIS&: *LIS); |
1477 | // Assign all new registers to RS_Spill. This was the last chance. |
1478 | ExtraInfo->setStage(Begin: LREdit.begin(), End: LREdit.end(), NewStage: RS_Spill); |
1479 | return 0; |
1480 | } |
1481 | |
1482 | //===----------------------------------------------------------------------===// |
1483 | // Local Splitting |
1484 | //===----------------------------------------------------------------------===// |
1485 | |
1486 | /// calcGapWeights - Compute the maximum spill weight that needs to be evicted |
1487 | /// in order to use PhysReg between two entries in SA->UseSlots. |
1488 | /// |
1489 | /// GapWeight[I] represents the gap between UseSlots[I] and UseSlots[I + 1]. |
1490 | /// |
1491 | void RAGreedy::calcGapWeights(MCRegister PhysReg, |
1492 | SmallVectorImpl<float> &GapWeight) { |
1493 | assert(SA->getUseBlocks().size() == 1 && "Not a local interval" ); |
1494 | const SplitAnalysis::BlockInfo &BI = SA->getUseBlocks().front(); |
1495 | ArrayRef<SlotIndex> Uses = SA->getUseSlots(); |
1496 | const unsigned NumGaps = Uses.size()-1; |
1497 | |
1498 | // Start and end points for the interference check. |
1499 | SlotIndex StartIdx = |
1500 | BI.LiveIn ? BI.FirstInstr.getBaseIndex() : BI.FirstInstr; |
1501 | SlotIndex StopIdx = |
1502 | BI.LiveOut ? BI.LastInstr.getBoundaryIndex() : BI.LastInstr; |
1503 | |
1504 | GapWeight.assign(NumElts: NumGaps, Elt: 0.0f); |
1505 | |
1506 | // Add interference from each overlapping register. |
1507 | for (MCRegUnit Unit : TRI->regunits(Reg: PhysReg)) { |
1508 | if (!Matrix->query(LR: const_cast<LiveInterval &>(SA->getParent()), RegUnit: Unit) |
1509 | .checkInterference()) |
1510 | continue; |
1511 | |
1512 | // We know that VirtReg is a continuous interval from FirstInstr to |
1513 | // LastInstr, so we don't need InterferenceQuery. |
1514 | // |
1515 | // Interference that overlaps an instruction is counted in both gaps |
1516 | // surrounding the instruction. The exception is interference before |
1517 | // StartIdx and after StopIdx. |
1518 | // |
1519 | LiveIntervalUnion::SegmentIter IntI = |
1520 | Matrix->getLiveUnions()[Unit].find(x: StartIdx); |
1521 | for (unsigned Gap = 0; IntI.valid() && IntI.start() < StopIdx; ++IntI) { |
1522 | // Skip the gaps before IntI. |
1523 | while (Uses[Gap+1].getBoundaryIndex() < IntI.start()) |
1524 | if (++Gap == NumGaps) |
1525 | break; |
1526 | if (Gap == NumGaps) |
1527 | break; |
1528 | |
1529 | // Update the gaps covered by IntI. |
1530 | const float weight = IntI.value()->weight(); |
1531 | for (; Gap != NumGaps; ++Gap) { |
1532 | GapWeight[Gap] = std::max(a: GapWeight[Gap], b: weight); |
1533 | if (Uses[Gap+1].getBaseIndex() >= IntI.stop()) |
1534 | break; |
1535 | } |
1536 | if (Gap == NumGaps) |
1537 | break; |
1538 | } |
1539 | } |
1540 | |
1541 | // Add fixed interference. |
1542 | for (MCRegUnit Unit : TRI->regunits(Reg: PhysReg)) { |
1543 | const LiveRange &LR = LIS->getRegUnit(Unit); |
1544 | LiveRange::const_iterator I = LR.find(Pos: StartIdx); |
1545 | LiveRange::const_iterator E = LR.end(); |
1546 | |
1547 | // Same loop as above. Mark any overlapped gaps as HUGE_VALF. |
1548 | for (unsigned Gap = 0; I != E && I->start < StopIdx; ++I) { |
1549 | while (Uses[Gap+1].getBoundaryIndex() < I->start) |
1550 | if (++Gap == NumGaps) |
1551 | break; |
1552 | if (Gap == NumGaps) |
1553 | break; |
1554 | |
1555 | for (; Gap != NumGaps; ++Gap) { |
1556 | GapWeight[Gap] = huge_valf; |
1557 | if (Uses[Gap+1].getBaseIndex() >= I->end) |
1558 | break; |
1559 | } |
1560 | if (Gap == NumGaps) |
1561 | break; |
1562 | } |
1563 | } |
1564 | } |
1565 | |
1566 | /// tryLocalSplit - Try to split VirtReg into smaller intervals inside its only |
1567 | /// basic block. |
1568 | /// |
1569 | unsigned RAGreedy::tryLocalSplit(const LiveInterval &VirtReg, |
1570 | AllocationOrder &Order, |
1571 | SmallVectorImpl<Register> &NewVRegs) { |
1572 | // TODO: the function currently only handles a single UseBlock; it should be |
1573 | // possible to generalize. |
1574 | if (SA->getUseBlocks().size() != 1) |
1575 | return 0; |
1576 | |
1577 | const SplitAnalysis::BlockInfo &BI = SA->getUseBlocks().front(); |
1578 | |
1579 | // Note that it is possible to have an interval that is live-in or live-out |
1580 | // while only covering a single block - A phi-def can use undef values from |
1581 | // predecessors, and the block could be a single-block loop. |
1582 | // We don't bother doing anything clever about such a case, we simply assume |
1583 | // that the interval is continuous from FirstInstr to LastInstr. We should |
1584 | // make sure that we don't do anything illegal to such an interval, though. |
1585 | |
1586 | ArrayRef<SlotIndex> Uses = SA->getUseSlots(); |
1587 | if (Uses.size() <= 2) |
1588 | return 0; |
1589 | const unsigned NumGaps = Uses.size()-1; |
1590 | |
1591 | LLVM_DEBUG({ |
1592 | dbgs() << "tryLocalSplit: " ; |
1593 | for (const auto &Use : Uses) |
1594 | dbgs() << ' ' << Use; |
1595 | dbgs() << '\n'; |
1596 | }); |
1597 | |
1598 | // If VirtReg is live across any register mask operands, compute a list of |
1599 | // gaps with register masks. |
1600 | SmallVector<unsigned, 8> RegMaskGaps; |
1601 | if (Matrix->checkRegMaskInterference(VirtReg)) { |
1602 | // Get regmask slots for the whole block. |
1603 | ArrayRef<SlotIndex> RMS = LIS->getRegMaskSlotsInBlock(MBBNum: BI.MBB->getNumber()); |
1604 | LLVM_DEBUG(dbgs() << RMS.size() << " regmasks in block:" ); |
1605 | // Constrain to VirtReg's live range. |
1606 | unsigned RI = |
1607 | llvm::lower_bound(Range&: RMS, Value: Uses.front().getRegSlot()) - RMS.begin(); |
1608 | unsigned RE = RMS.size(); |
1609 | for (unsigned I = 0; I != NumGaps && RI != RE; ++I) { |
1610 | // Look for Uses[I] <= RMS <= Uses[I + 1]. |
1611 | assert(!SlotIndex::isEarlierInstr(RMS[RI], Uses[I])); |
1612 | if (SlotIndex::isEarlierInstr(A: Uses[I + 1], B: RMS[RI])) |
1613 | continue; |
1614 | // Skip a regmask on the same instruction as the last use. It doesn't |
1615 | // overlap the live range. |
1616 | if (SlotIndex::isSameInstr(A: Uses[I + 1], B: RMS[RI]) && I + 1 == NumGaps) |
1617 | break; |
1618 | LLVM_DEBUG(dbgs() << ' ' << RMS[RI] << ':' << Uses[I] << '-' |
1619 | << Uses[I + 1]); |
1620 | RegMaskGaps.push_back(Elt: I); |
1621 | // Advance ri to the next gap. A regmask on one of the uses counts in |
1622 | // both gaps. |
1623 | while (RI != RE && SlotIndex::isEarlierInstr(A: RMS[RI], B: Uses[I + 1])) |
1624 | ++RI; |
1625 | } |
1626 | LLVM_DEBUG(dbgs() << '\n'); |
1627 | } |
1628 | |
1629 | // Since we allow local split results to be split again, there is a risk of |
1630 | // creating infinite loops. It is tempting to require that the new live |
1631 | // ranges have less instructions than the original. That would guarantee |
1632 | // convergence, but it is too strict. A live range with 3 instructions can be |
1633 | // split 2+3 (including the COPY), and we want to allow that. |
1634 | // |
1635 | // Instead we use these rules: |
1636 | // |
1637 | // 1. Allow any split for ranges with getStage() < RS_Split2. (Except for the |
1638 | // noop split, of course). |
1639 | // 2. Require progress be made for ranges with getStage() == RS_Split2. All |
1640 | // the new ranges must have fewer instructions than before the split. |
1641 | // 3. New ranges with the same number of instructions are marked RS_Split2, |
1642 | // smaller ranges are marked RS_New. |
1643 | // |
1644 | // These rules allow a 3 -> 2+3 split once, which we need. They also prevent |
1645 | // excessive splitting and infinite loops. |
1646 | // |
1647 | bool ProgressRequired = ExtraInfo->getStage(VirtReg) >= RS_Split2; |
1648 | |
1649 | // Best split candidate. |
1650 | unsigned BestBefore = NumGaps; |
1651 | unsigned BestAfter = 0; |
1652 | float BestDiff = 0; |
1653 | |
1654 | const float blockFreq = |
1655 | SpillPlacer->getBlockFrequency(Number: BI.MBB->getNumber()).getFrequency() * |
1656 | (1.0f / MBFI->getEntryFreq().getFrequency()); |
1657 | SmallVector<float, 8> GapWeight; |
1658 | |
1659 | for (MCPhysReg PhysReg : Order) { |
1660 | assert(PhysReg); |
1661 | // Keep track of the largest spill weight that would need to be evicted in |
1662 | // order to make use of PhysReg between UseSlots[I] and UseSlots[I + 1]. |
1663 | calcGapWeights(PhysReg, GapWeight); |
1664 | |
1665 | // Remove any gaps with regmask clobbers. |
1666 | if (Matrix->checkRegMaskInterference(VirtReg, PhysReg)) |
1667 | for (unsigned I = 0, E = RegMaskGaps.size(); I != E; ++I) |
1668 | GapWeight[RegMaskGaps[I]] = huge_valf; |
1669 | |
1670 | // Try to find the best sequence of gaps to close. |
1671 | // The new spill weight must be larger than any gap interference. |
1672 | |
1673 | // We will split before Uses[SplitBefore] and after Uses[SplitAfter]. |
1674 | unsigned SplitBefore = 0, SplitAfter = 1; |
1675 | |
1676 | // MaxGap should always be max(GapWeight[SplitBefore..SplitAfter-1]). |
1677 | // It is the spill weight that needs to be evicted. |
1678 | float MaxGap = GapWeight[0]; |
1679 | |
1680 | while (true) { |
1681 | // Live before/after split? |
1682 | const bool LiveBefore = SplitBefore != 0 || BI.LiveIn; |
1683 | const bool LiveAfter = SplitAfter != NumGaps || BI.LiveOut; |
1684 | |
1685 | LLVM_DEBUG(dbgs() << printReg(PhysReg, TRI) << ' ' << Uses[SplitBefore] |
1686 | << '-' << Uses[SplitAfter] << " I=" << MaxGap); |
1687 | |
1688 | // Stop before the interval gets so big we wouldn't be making progress. |
1689 | if (!LiveBefore && !LiveAfter) { |
1690 | LLVM_DEBUG(dbgs() << " all\n" ); |
1691 | break; |
1692 | } |
1693 | // Should the interval be extended or shrunk? |
1694 | bool Shrink = true; |
1695 | |
1696 | // How many gaps would the new range have? |
1697 | unsigned NewGaps = LiveBefore + SplitAfter - SplitBefore + LiveAfter; |
1698 | |
1699 | // Legally, without causing looping? |
1700 | bool Legal = !ProgressRequired || NewGaps < NumGaps; |
1701 | |
1702 | if (Legal && MaxGap < huge_valf) { |
1703 | // Estimate the new spill weight. Each instruction reads or writes the |
1704 | // register. Conservatively assume there are no read-modify-write |
1705 | // instructions. |
1706 | // |
1707 | // Try to guess the size of the new interval. |
1708 | const float EstWeight = normalizeSpillWeight( |
1709 | UseDefFreq: blockFreq * (NewGaps + 1), |
1710 | Size: Uses[SplitBefore].distance(other: Uses[SplitAfter]) + |
1711 | (LiveBefore + LiveAfter) * SlotIndex::InstrDist, |
1712 | NumInstr: 1); |
1713 | // Would this split be possible to allocate? |
1714 | // Never allocate all gaps, we wouldn't be making progress. |
1715 | LLVM_DEBUG(dbgs() << " w=" << EstWeight); |
1716 | if (EstWeight * Hysteresis >= MaxGap) { |
1717 | Shrink = false; |
1718 | float Diff = EstWeight - MaxGap; |
1719 | if (Diff > BestDiff) { |
1720 | LLVM_DEBUG(dbgs() << " (best)" ); |
1721 | BestDiff = Hysteresis * Diff; |
1722 | BestBefore = SplitBefore; |
1723 | BestAfter = SplitAfter; |
1724 | } |
1725 | } |
1726 | } |
1727 | |
1728 | // Try to shrink. |
1729 | if (Shrink) { |
1730 | if (++SplitBefore < SplitAfter) { |
1731 | LLVM_DEBUG(dbgs() << " shrink\n" ); |
1732 | // Recompute the max when necessary. |
1733 | if (GapWeight[SplitBefore - 1] >= MaxGap) { |
1734 | MaxGap = GapWeight[SplitBefore]; |
1735 | for (unsigned I = SplitBefore + 1; I != SplitAfter; ++I) |
1736 | MaxGap = std::max(a: MaxGap, b: GapWeight[I]); |
1737 | } |
1738 | continue; |
1739 | } |
1740 | MaxGap = 0; |
1741 | } |
1742 | |
1743 | // Try to extend the interval. |
1744 | if (SplitAfter >= NumGaps) { |
1745 | LLVM_DEBUG(dbgs() << " end\n" ); |
1746 | break; |
1747 | } |
1748 | |
1749 | LLVM_DEBUG(dbgs() << " extend\n" ); |
1750 | MaxGap = std::max(a: MaxGap, b: GapWeight[SplitAfter++]); |
1751 | } |
1752 | } |
1753 | |
1754 | // Didn't find any candidates? |
1755 | if (BestBefore == NumGaps) |
1756 | return 0; |
1757 | |
1758 | LLVM_DEBUG(dbgs() << "Best local split range: " << Uses[BestBefore] << '-' |
1759 | << Uses[BestAfter] << ", " << BestDiff << ", " |
1760 | << (BestAfter - BestBefore + 1) << " instrs\n" ); |
1761 | |
1762 | LiveRangeEdit LREdit(&VirtReg, NewVRegs, *MF, *LIS, VRM, this, &DeadRemats); |
1763 | SE->reset(LREdit); |
1764 | |
1765 | SE->openIntv(); |
1766 | SlotIndex SegStart = SE->enterIntvBefore(Idx: Uses[BestBefore]); |
1767 | SlotIndex SegStop = SE->leaveIntvAfter(Idx: Uses[BestAfter]); |
1768 | SE->useIntv(Start: SegStart, End: SegStop); |
1769 | SmallVector<unsigned, 8> IntvMap; |
1770 | SE->finish(LRMap: &IntvMap); |
1771 | DebugVars->splitRegister(OldReg: VirtReg.reg(), NewRegs: LREdit.regs(), LIS&: *LIS); |
1772 | // If the new range has the same number of instructions as before, mark it as |
1773 | // RS_Split2 so the next split will be forced to make progress. Otherwise, |
1774 | // leave the new intervals as RS_New so they can compete. |
1775 | bool LiveBefore = BestBefore != 0 || BI.LiveIn; |
1776 | bool LiveAfter = BestAfter != NumGaps || BI.LiveOut; |
1777 | unsigned NewGaps = LiveBefore + BestAfter - BestBefore + LiveAfter; |
1778 | if (NewGaps >= NumGaps) { |
1779 | LLVM_DEBUG(dbgs() << "Tagging non-progress ranges:" ); |
1780 | assert(!ProgressRequired && "Didn't make progress when it was required." ); |
1781 | for (unsigned I = 0, E = IntvMap.size(); I != E; ++I) |
1782 | if (IntvMap[I] == 1) { |
1783 | ExtraInfo->setStage(VirtReg: LIS->getInterval(Reg: LREdit.get(idx: I)), Stage: RS_Split2); |
1784 | LLVM_DEBUG(dbgs() << ' ' << printReg(LREdit.get(I))); |
1785 | } |
1786 | LLVM_DEBUG(dbgs() << '\n'); |
1787 | } |
1788 | ++NumLocalSplits; |
1789 | |
1790 | return 0; |
1791 | } |
1792 | |
1793 | //===----------------------------------------------------------------------===// |
1794 | // Live Range Splitting |
1795 | //===----------------------------------------------------------------------===// |
1796 | |
1797 | /// trySplit - Try to split VirtReg or one of its interferences, making it |
1798 | /// assignable. |
1799 | /// @return Physreg when VirtReg may be assigned and/or new NewVRegs. |
1800 | unsigned RAGreedy::trySplit(const LiveInterval &VirtReg, AllocationOrder &Order, |
1801 | SmallVectorImpl<Register> &NewVRegs, |
1802 | const SmallVirtRegSet &FixedRegisters) { |
1803 | // Ranges must be Split2 or less. |
1804 | if (ExtraInfo->getStage(VirtReg) >= RS_Spill) |
1805 | return 0; |
1806 | |
1807 | // Local intervals are handled separately. |
1808 | if (LIS->intervalIsInOneMBB(LI: VirtReg)) { |
1809 | NamedRegionTimer T("local_split" , "Local Splitting" , TimerGroupName, |
1810 | TimerGroupDescription, TimePassesIsEnabled); |
1811 | SA->analyze(li: &VirtReg); |
1812 | Register PhysReg = tryLocalSplit(VirtReg, Order, NewVRegs); |
1813 | if (PhysReg || !NewVRegs.empty()) |
1814 | return PhysReg; |
1815 | return tryInstructionSplit(VirtReg, Order, NewVRegs); |
1816 | } |
1817 | |
1818 | NamedRegionTimer T("global_split" , "Global Splitting" , TimerGroupName, |
1819 | TimerGroupDescription, TimePassesIsEnabled); |
1820 | |
1821 | SA->analyze(li: &VirtReg); |
1822 | |
1823 | // First try to split around a region spanning multiple blocks. RS_Split2 |
1824 | // ranges already made dubious progress with region splitting, so they go |
1825 | // straight to single block splitting. |
1826 | if (ExtraInfo->getStage(VirtReg) < RS_Split2) { |
1827 | MCRegister PhysReg = tryRegionSplit(VirtReg, Order, NewVRegs); |
1828 | if (PhysReg || !NewVRegs.empty()) |
1829 | return PhysReg; |
1830 | } |
1831 | |
1832 | // Then isolate blocks. |
1833 | return tryBlockSplit(VirtReg, Order, NewVRegs); |
1834 | } |
1835 | |
1836 | //===----------------------------------------------------------------------===// |
1837 | // Last Chance Recoloring |
1838 | //===----------------------------------------------------------------------===// |
1839 | |
1840 | /// Return true if \p reg has any tied def operand. |
1841 | static bool hasTiedDef(MachineRegisterInfo *MRI, unsigned reg) { |
1842 | for (const MachineOperand &MO : MRI->def_operands(Reg: reg)) |
1843 | if (MO.isTied()) |
1844 | return true; |
1845 | |
1846 | return false; |
1847 | } |
1848 | |
1849 | /// Return true if the existing assignment of \p Intf overlaps, but is not the |
1850 | /// same, as \p PhysReg. |
1851 | static bool assignedRegPartiallyOverlaps(const TargetRegisterInfo &TRI, |
1852 | const VirtRegMap &VRM, |
1853 | MCRegister PhysReg, |
1854 | const LiveInterval &Intf) { |
1855 | MCRegister AssignedReg = VRM.getPhys(virtReg: Intf.reg()); |
1856 | if (PhysReg == AssignedReg) |
1857 | return false; |
1858 | return TRI.regsOverlap(RegA: PhysReg, RegB: AssignedReg); |
1859 | } |
1860 | |
1861 | /// mayRecolorAllInterferences - Check if the virtual registers that |
1862 | /// interfere with \p VirtReg on \p PhysReg (or one of its aliases) may be |
1863 | /// recolored to free \p PhysReg. |
1864 | /// When true is returned, \p RecoloringCandidates has been augmented with all |
1865 | /// the live intervals that need to be recolored in order to free \p PhysReg |
1866 | /// for \p VirtReg. |
1867 | /// \p FixedRegisters contains all the virtual registers that cannot be |
1868 | /// recolored. |
1869 | bool RAGreedy::mayRecolorAllInterferences( |
1870 | MCRegister PhysReg, const LiveInterval &VirtReg, |
1871 | SmallLISet &RecoloringCandidates, const SmallVirtRegSet &FixedRegisters) { |
1872 | const TargetRegisterClass *CurRC = MRI->getRegClass(Reg: VirtReg.reg()); |
1873 | |
1874 | for (MCRegUnit Unit : TRI->regunits(Reg: PhysReg)) { |
1875 | LiveIntervalUnion::Query &Q = Matrix->query(LR: VirtReg, RegUnit: Unit); |
1876 | // If there is LastChanceRecoloringMaxInterference or more interferences, |
1877 | // chances are one would not be recolorable. |
1878 | if (Q.interferingVRegs(MaxInterferingRegs: LastChanceRecoloringMaxInterference).size() >= |
1879 | LastChanceRecoloringMaxInterference && |
1880 | !ExhaustiveSearch) { |
1881 | LLVM_DEBUG(dbgs() << "Early abort: too many interferences.\n" ); |
1882 | CutOffInfo |= CO_Interf; |
1883 | return false; |
1884 | } |
1885 | for (const LiveInterval *Intf : reverse(C: Q.interferingVRegs())) { |
1886 | // If Intf is done and sits on the same register class as VirtReg, it |
1887 | // would not be recolorable as it is in the same state as |
1888 | // VirtReg. However there are at least two exceptions. |
1889 | // |
1890 | // If VirtReg has tied defs and Intf doesn't, then |
1891 | // there is still a point in examining if it can be recolorable. |
1892 | // |
1893 | // Additionally, if the register class has overlapping tuple members, it |
1894 | // may still be recolorable using a different tuple. This is more likely |
1895 | // if the existing assignment aliases with the candidate. |
1896 | // |
1897 | if (((ExtraInfo->getStage(VirtReg: *Intf) == RS_Done && |
1898 | MRI->getRegClass(Reg: Intf->reg()) == CurRC && |
1899 | !assignedRegPartiallyOverlaps(TRI: *TRI, VRM: *VRM, PhysReg, Intf: *Intf)) && |
1900 | !(hasTiedDef(MRI, reg: VirtReg.reg()) && |
1901 | !hasTiedDef(MRI, reg: Intf->reg()))) || |
1902 | FixedRegisters.count(V: Intf->reg())) { |
1903 | LLVM_DEBUG( |
1904 | dbgs() << "Early abort: the interference is not recolorable.\n" ); |
1905 | return false; |
1906 | } |
1907 | RecoloringCandidates.insert(X: Intf); |
1908 | } |
1909 | } |
1910 | return true; |
1911 | } |
1912 | |
1913 | /// tryLastChanceRecoloring - Try to assign a color to \p VirtReg by recoloring |
1914 | /// its interferences. |
1915 | /// Last chance recoloring chooses a color for \p VirtReg and recolors every |
1916 | /// virtual register that was using it. The recoloring process may recursively |
1917 | /// use the last chance recoloring. Therefore, when a virtual register has been |
1918 | /// assigned a color by this mechanism, it is marked as Fixed, i.e., it cannot |
1919 | /// be last-chance-recolored again during this recoloring "session". |
1920 | /// E.g., |
1921 | /// Let |
1922 | /// vA can use {R1, R2 } |
1923 | /// vB can use { R2, R3} |
1924 | /// vC can use {R1 } |
1925 | /// Where vA, vB, and vC cannot be split anymore (they are reloads for |
1926 | /// instance) and they all interfere. |
1927 | /// |
1928 | /// vA is assigned R1 |
1929 | /// vB is assigned R2 |
1930 | /// vC tries to evict vA but vA is already done. |
1931 | /// Regular register allocation fails. |
1932 | /// |
1933 | /// Last chance recoloring kicks in: |
1934 | /// vC does as if vA was evicted => vC uses R1. |
1935 | /// vC is marked as fixed. |
1936 | /// vA needs to find a color. |
1937 | /// None are available. |
1938 | /// vA cannot evict vC: vC is a fixed virtual register now. |
1939 | /// vA does as if vB was evicted => vA uses R2. |
1940 | /// vB needs to find a color. |
1941 | /// R3 is available. |
1942 | /// Recoloring => vC = R1, vA = R2, vB = R3 |
1943 | /// |
1944 | /// \p Order defines the preferred allocation order for \p VirtReg. |
1945 | /// \p NewRegs will contain any new virtual register that have been created |
1946 | /// (split, spill) during the process and that must be assigned. |
1947 | /// \p FixedRegisters contains all the virtual registers that cannot be |
1948 | /// recolored. |
1949 | /// |
1950 | /// \p RecolorStack tracks the original assignments of successfully recolored |
1951 | /// registers. |
1952 | /// |
1953 | /// \p Depth gives the current depth of the last chance recoloring. |
1954 | /// \return a physical register that can be used for VirtReg or ~0u if none |
1955 | /// exists. |
1956 | unsigned RAGreedy::tryLastChanceRecoloring(const LiveInterval &VirtReg, |
1957 | AllocationOrder &Order, |
1958 | SmallVectorImpl<Register> &NewVRegs, |
1959 | SmallVirtRegSet &FixedRegisters, |
1960 | RecoloringStack &RecolorStack, |
1961 | unsigned Depth) { |
1962 | if (!TRI->shouldUseLastChanceRecoloringForVirtReg(MF: *MF, VirtReg)) |
1963 | return ~0u; |
1964 | |
1965 | LLVM_DEBUG(dbgs() << "Try last chance recoloring for " << VirtReg << '\n'); |
1966 | |
1967 | const ssize_t EntryStackSize = RecolorStack.size(); |
1968 | |
1969 | // Ranges must be Done. |
1970 | assert((ExtraInfo->getStage(VirtReg) >= RS_Done || !VirtReg.isSpillable()) && |
1971 | "Last chance recoloring should really be last chance" ); |
1972 | // Set the max depth to LastChanceRecoloringMaxDepth. |
1973 | // We may want to reconsider that if we end up with a too large search space |
1974 | // for target with hundreds of registers. |
1975 | // Indeed, in that case we may want to cut the search space earlier. |
1976 | if (Depth >= LastChanceRecoloringMaxDepth && !ExhaustiveSearch) { |
1977 | LLVM_DEBUG(dbgs() << "Abort because max depth has been reached.\n" ); |
1978 | CutOffInfo |= CO_Depth; |
1979 | return ~0u; |
1980 | } |
1981 | |
1982 | // Set of Live intervals that will need to be recolored. |
1983 | SmallLISet RecoloringCandidates; |
1984 | |
1985 | // Mark VirtReg as fixed, i.e., it will not be recolored pass this point in |
1986 | // this recoloring "session". |
1987 | assert(!FixedRegisters.count(VirtReg.reg())); |
1988 | FixedRegisters.insert(V: VirtReg.reg()); |
1989 | SmallVector<Register, 4> CurrentNewVRegs; |
1990 | |
1991 | for (MCRegister PhysReg : Order) { |
1992 | assert(PhysReg.isValid()); |
1993 | LLVM_DEBUG(dbgs() << "Try to assign: " << VirtReg << " to " |
1994 | << printReg(PhysReg, TRI) << '\n'); |
1995 | RecoloringCandidates.clear(); |
1996 | CurrentNewVRegs.clear(); |
1997 | |
1998 | // It is only possible to recolor virtual register interference. |
1999 | if (Matrix->checkInterference(VirtReg, PhysReg) > |
2000 | LiveRegMatrix::IK_VirtReg) { |
2001 | LLVM_DEBUG( |
2002 | dbgs() << "Some interferences are not with virtual registers.\n" ); |
2003 | |
2004 | continue; |
2005 | } |
2006 | |
2007 | // Early give up on this PhysReg if it is obvious we cannot recolor all |
2008 | // the interferences. |
2009 | if (!mayRecolorAllInterferences(PhysReg, VirtReg, RecoloringCandidates, |
2010 | FixedRegisters)) { |
2011 | LLVM_DEBUG(dbgs() << "Some interferences cannot be recolored.\n" ); |
2012 | continue; |
2013 | } |
2014 | |
2015 | // RecoloringCandidates contains all the virtual registers that interfere |
2016 | // with VirtReg on PhysReg (or one of its aliases). Enqueue them for |
2017 | // recoloring and perform the actual recoloring. |
2018 | PQueue RecoloringQueue; |
2019 | for (const LiveInterval *RC : RecoloringCandidates) { |
2020 | Register ItVirtReg = RC->reg(); |
2021 | enqueue(CurQueue&: RecoloringQueue, LI: RC); |
2022 | assert(VRM->hasPhys(ItVirtReg) && |
2023 | "Interferences are supposed to be with allocated variables" ); |
2024 | |
2025 | // Record the current allocation. |
2026 | RecolorStack.push_back(Elt: std::make_pair(x&: RC, y: VRM->getPhys(virtReg: ItVirtReg))); |
2027 | |
2028 | // unset the related struct. |
2029 | Matrix->unassign(VirtReg: *RC); |
2030 | } |
2031 | |
2032 | // Do as if VirtReg was assigned to PhysReg so that the underlying |
2033 | // recoloring has the right information about the interferes and |
2034 | // available colors. |
2035 | Matrix->assign(VirtReg, PhysReg); |
2036 | |
2037 | // Save the current recoloring state. |
2038 | // If we cannot recolor all the interferences, we will have to start again |
2039 | // at this point for the next physical register. |
2040 | SmallVirtRegSet SaveFixedRegisters(FixedRegisters); |
2041 | if (tryRecoloringCandidates(RecoloringQueue, CurrentNewVRegs, |
2042 | FixedRegisters, RecolorStack, Depth)) { |
2043 | // Push the queued vregs into the main queue. |
2044 | for (Register NewVReg : CurrentNewVRegs) |
2045 | NewVRegs.push_back(Elt: NewVReg); |
2046 | // Do not mess up with the global assignment process. |
2047 | // I.e., VirtReg must be unassigned. |
2048 | Matrix->unassign(VirtReg); |
2049 | return PhysReg; |
2050 | } |
2051 | |
2052 | LLVM_DEBUG(dbgs() << "Fail to assign: " << VirtReg << " to " |
2053 | << printReg(PhysReg, TRI) << '\n'); |
2054 | |
2055 | // The recoloring attempt failed, undo the changes. |
2056 | FixedRegisters = SaveFixedRegisters; |
2057 | Matrix->unassign(VirtReg); |
2058 | |
2059 | // For a newly created vreg which is also in RecoloringCandidates, |
2060 | // don't add it to NewVRegs because its physical register will be restored |
2061 | // below. Other vregs in CurrentNewVRegs are created by calling |
2062 | // selectOrSplit and should be added into NewVRegs. |
2063 | for (Register R : CurrentNewVRegs) { |
2064 | if (RecoloringCandidates.count(key: &LIS->getInterval(Reg: R))) |
2065 | continue; |
2066 | NewVRegs.push_back(Elt: R); |
2067 | } |
2068 | |
2069 | // Roll back our unsuccessful recoloring. Also roll back any successful |
2070 | // recolorings in any recursive recoloring attempts, since it's possible |
2071 | // they would have introduced conflicts with assignments we will be |
2072 | // restoring further up the stack. Perform all unassignments prior to |
2073 | // reassigning, since sub-recolorings may have conflicted with the registers |
2074 | // we are going to restore to their original assignments. |
2075 | for (ssize_t I = RecolorStack.size() - 1; I >= EntryStackSize; --I) { |
2076 | const LiveInterval *LI; |
2077 | MCRegister PhysReg; |
2078 | std::tie(args&: LI, args&: PhysReg) = RecolorStack[I]; |
2079 | |
2080 | if (VRM->hasPhys(virtReg: LI->reg())) |
2081 | Matrix->unassign(VirtReg: *LI); |
2082 | } |
2083 | |
2084 | for (size_t I = EntryStackSize; I != RecolorStack.size(); ++I) { |
2085 | const LiveInterval *LI; |
2086 | MCRegister PhysReg; |
2087 | std::tie(args&: LI, args&: PhysReg) = RecolorStack[I]; |
2088 | if (!LI->empty() && !MRI->reg_nodbg_empty(RegNo: LI->reg())) |
2089 | Matrix->assign(VirtReg: *LI, PhysReg); |
2090 | } |
2091 | |
2092 | // Pop the stack of recoloring attempts. |
2093 | RecolorStack.resize(N: EntryStackSize); |
2094 | } |
2095 | |
2096 | // Last chance recoloring did not worked either, give up. |
2097 | return ~0u; |
2098 | } |
2099 | |
2100 | /// tryRecoloringCandidates - Try to assign a new color to every register |
2101 | /// in \RecoloringQueue. |
2102 | /// \p NewRegs will contain any new virtual register created during the |
2103 | /// recoloring process. |
2104 | /// \p FixedRegisters[in/out] contains all the registers that have been |
2105 | /// recolored. |
2106 | /// \return true if all virtual registers in RecoloringQueue were successfully |
2107 | /// recolored, false otherwise. |
2108 | bool RAGreedy::tryRecoloringCandidates(PQueue &RecoloringQueue, |
2109 | SmallVectorImpl<Register> &NewVRegs, |
2110 | SmallVirtRegSet &FixedRegisters, |
2111 | RecoloringStack &RecolorStack, |
2112 | unsigned Depth) { |
2113 | while (!RecoloringQueue.empty()) { |
2114 | const LiveInterval *LI = dequeue(CurQueue&: RecoloringQueue); |
2115 | LLVM_DEBUG(dbgs() << "Try to recolor: " << *LI << '\n'); |
2116 | MCRegister PhysReg = selectOrSplitImpl(*LI, NewVRegs, FixedRegisters, |
2117 | RecolorStack, Depth + 1); |
2118 | // When splitting happens, the live-range may actually be empty. |
2119 | // In that case, this is okay to continue the recoloring even |
2120 | // if we did not find an alternative color for it. Indeed, |
2121 | // there will not be anything to color for LI in the end. |
2122 | if (PhysReg == ~0u || (!PhysReg && !LI->empty())) |
2123 | return false; |
2124 | |
2125 | if (!PhysReg) { |
2126 | assert(LI->empty() && "Only empty live-range do not require a register" ); |
2127 | LLVM_DEBUG(dbgs() << "Recoloring of " << *LI |
2128 | << " succeeded. Empty LI.\n" ); |
2129 | continue; |
2130 | } |
2131 | LLVM_DEBUG(dbgs() << "Recoloring of " << *LI |
2132 | << " succeeded with: " << printReg(PhysReg, TRI) << '\n'); |
2133 | |
2134 | Matrix->assign(VirtReg: *LI, PhysReg); |
2135 | FixedRegisters.insert(V: LI->reg()); |
2136 | } |
2137 | return true; |
2138 | } |
2139 | |
2140 | //===----------------------------------------------------------------------===// |
2141 | // Main Entry Point |
2142 | //===----------------------------------------------------------------------===// |
2143 | |
2144 | MCRegister RAGreedy::selectOrSplit(const LiveInterval &VirtReg, |
2145 | SmallVectorImpl<Register> &NewVRegs) { |
2146 | CutOffInfo = CO_None; |
2147 | LLVMContext &Ctx = MF->getFunction().getContext(); |
2148 | SmallVirtRegSet FixedRegisters; |
2149 | RecoloringStack RecolorStack; |
2150 | MCRegister Reg = |
2151 | selectOrSplitImpl(VirtReg, NewVRegs, FixedRegisters, RecolorStack); |
2152 | if (Reg == ~0U && (CutOffInfo != CO_None)) { |
2153 | uint8_t CutOffEncountered = CutOffInfo & (CO_Depth | CO_Interf); |
2154 | if (CutOffEncountered == CO_Depth) |
2155 | Ctx.emitError(ErrorStr: "register allocation failed: maximum depth for recoloring " |
2156 | "reached. Use -fexhaustive-register-search to skip " |
2157 | "cutoffs" ); |
2158 | else if (CutOffEncountered == CO_Interf) |
2159 | Ctx.emitError(ErrorStr: "register allocation failed: maximum interference for " |
2160 | "recoloring reached. Use -fexhaustive-register-search " |
2161 | "to skip cutoffs" ); |
2162 | else if (CutOffEncountered == (CO_Depth | CO_Interf)) |
2163 | Ctx.emitError(ErrorStr: "register allocation failed: maximum interference and " |
2164 | "depth for recoloring reached. Use " |
2165 | "-fexhaustive-register-search to skip cutoffs" ); |
2166 | } |
2167 | return Reg; |
2168 | } |
2169 | |
2170 | /// Using a CSR for the first time has a cost because it causes push|pop |
2171 | /// to be added to prologue|epilogue. Splitting a cold section of the live |
2172 | /// range can have lower cost than using the CSR for the first time; |
2173 | /// Spilling a live range in the cold path can have lower cost than using |
2174 | /// the CSR for the first time. Returns the physical register if we decide |
2175 | /// to use the CSR; otherwise return 0. |
2176 | MCRegister RAGreedy::tryAssignCSRFirstTime( |
2177 | const LiveInterval &VirtReg, AllocationOrder &Order, MCRegister PhysReg, |
2178 | uint8_t &CostPerUseLimit, SmallVectorImpl<Register> &NewVRegs) { |
2179 | if (ExtraInfo->getStage(VirtReg) == RS_Spill && VirtReg.isSpillable()) { |
2180 | // We choose spill over using the CSR for the first time if the spill cost |
2181 | // is lower than CSRCost. |
2182 | SA->analyze(li: &VirtReg); |
2183 | if (calcSpillCost() >= CSRCost) |
2184 | return PhysReg; |
2185 | |
2186 | // We are going to spill, set CostPerUseLimit to 1 to make sure that |
2187 | // we will not use a callee-saved register in tryEvict. |
2188 | CostPerUseLimit = 1; |
2189 | return 0; |
2190 | } |
2191 | if (ExtraInfo->getStage(VirtReg) < RS_Split) { |
2192 | // We choose pre-splitting over using the CSR for the first time if |
2193 | // the cost of splitting is lower than CSRCost. |
2194 | SA->analyze(li: &VirtReg); |
2195 | unsigned NumCands = 0; |
2196 | BlockFrequency BestCost = CSRCost; // Don't modify CSRCost. |
2197 | unsigned BestCand = calculateRegionSplitCost(VirtReg, Order, BestCost, |
2198 | NumCands, IgnoreCSR: true /*IgnoreCSR*/); |
2199 | if (BestCand == NoCand) |
2200 | // Use the CSR if we can't find a region split below CSRCost. |
2201 | return PhysReg; |
2202 | |
2203 | // Perform the actual pre-splitting. |
2204 | doRegionSplit(VirtReg, BestCand, HasCompact: false/*HasCompact*/, NewVRegs); |
2205 | return 0; |
2206 | } |
2207 | return PhysReg; |
2208 | } |
2209 | |
2210 | void RAGreedy::aboutToRemoveInterval(const LiveInterval &LI) { |
2211 | // Do not keep invalid information around. |
2212 | SetOfBrokenHints.remove(X: &LI); |
2213 | } |
2214 | |
2215 | void RAGreedy::initializeCSRCost() { |
2216 | // We use the larger one out of the command-line option and the value report |
2217 | // by TRI. |
2218 | CSRCost = BlockFrequency( |
2219 | std::max(a: (unsigned)CSRFirstTimeCost, b: TRI->getCSRFirstUseCost())); |
2220 | if (!CSRCost.getFrequency()) |
2221 | return; |
2222 | |
2223 | // Raw cost is relative to Entry == 2^14; scale it appropriately. |
2224 | uint64_t ActualEntry = MBFI->getEntryFreq().getFrequency(); |
2225 | if (!ActualEntry) { |
2226 | CSRCost = BlockFrequency(0); |
2227 | return; |
2228 | } |
2229 | uint64_t FixedEntry = 1 << 14; |
2230 | if (ActualEntry < FixedEntry) |
2231 | CSRCost *= BranchProbability(ActualEntry, FixedEntry); |
2232 | else if (ActualEntry <= UINT32_MAX) |
2233 | // Invert the fraction and divide. |
2234 | CSRCost /= BranchProbability(FixedEntry, ActualEntry); |
2235 | else |
2236 | // Can't use BranchProbability in general, since it takes 32-bit numbers. |
2237 | CSRCost = |
2238 | BlockFrequency(CSRCost.getFrequency() * (ActualEntry / FixedEntry)); |
2239 | } |
2240 | |
2241 | /// Collect the hint info for \p Reg. |
2242 | /// The results are stored into \p Out. |
2243 | /// \p Out is not cleared before being populated. |
2244 | void RAGreedy::collectHintInfo(Register Reg, HintsInfo &Out) { |
2245 | for (const MachineInstr &Instr : MRI->reg_nodbg_instructions(Reg)) { |
2246 | if (!TII->isFullCopyInstr(MI: Instr)) |
2247 | continue; |
2248 | // Look for the other end of the copy. |
2249 | Register OtherReg = Instr.getOperand(i: 0).getReg(); |
2250 | if (OtherReg == Reg) { |
2251 | OtherReg = Instr.getOperand(i: 1).getReg(); |
2252 | if (OtherReg == Reg) |
2253 | continue; |
2254 | } |
2255 | // Get the current assignment. |
2256 | MCRegister OtherPhysReg = |
2257 | OtherReg.isPhysical() ? OtherReg.asMCReg() : VRM->getPhys(virtReg: OtherReg); |
2258 | // Push the collected information. |
2259 | Out.push_back(Elt: HintInfo(MBFI->getBlockFreq(MBB: Instr.getParent()), OtherReg, |
2260 | OtherPhysReg)); |
2261 | } |
2262 | } |
2263 | |
2264 | /// Using the given \p List, compute the cost of the broken hints if |
2265 | /// \p PhysReg was used. |
2266 | /// \return The cost of \p List for \p PhysReg. |
2267 | BlockFrequency RAGreedy::getBrokenHintFreq(const HintsInfo &List, |
2268 | MCRegister PhysReg) { |
2269 | BlockFrequency Cost = BlockFrequency(0); |
2270 | for (const HintInfo &Info : List) { |
2271 | if (Info.PhysReg != PhysReg) |
2272 | Cost += Info.Freq; |
2273 | } |
2274 | return Cost; |
2275 | } |
2276 | |
2277 | /// Using the register assigned to \p VirtReg, try to recolor |
2278 | /// all the live ranges that are copy-related with \p VirtReg. |
2279 | /// The recoloring is then propagated to all the live-ranges that have |
2280 | /// been recolored and so on, until no more copies can be coalesced or |
2281 | /// it is not profitable. |
2282 | /// For a given live range, profitability is determined by the sum of the |
2283 | /// frequencies of the non-identity copies it would introduce with the old |
2284 | /// and new register. |
2285 | void RAGreedy::tryHintRecoloring(const LiveInterval &VirtReg) { |
2286 | // We have a broken hint, check if it is possible to fix it by |
2287 | // reusing PhysReg for the copy-related live-ranges. Indeed, we evicted |
2288 | // some register and PhysReg may be available for the other live-ranges. |
2289 | SmallSet<Register, 4> Visited; |
2290 | SmallVector<unsigned, 2> RecoloringCandidates; |
2291 | HintsInfo Info; |
2292 | Register Reg = VirtReg.reg(); |
2293 | MCRegister PhysReg = VRM->getPhys(virtReg: Reg); |
2294 | // Start the recoloring algorithm from the input live-interval, then |
2295 | // it will propagate to the ones that are copy-related with it. |
2296 | Visited.insert(V: Reg); |
2297 | RecoloringCandidates.push_back(Elt: Reg); |
2298 | |
2299 | LLVM_DEBUG(dbgs() << "Trying to reconcile hints for: " << printReg(Reg, TRI) |
2300 | << '(' << printReg(PhysReg, TRI) << ")\n" ); |
2301 | |
2302 | do { |
2303 | Reg = RecoloringCandidates.pop_back_val(); |
2304 | |
2305 | // We cannot recolor physical register. |
2306 | if (Reg.isPhysical()) |
2307 | continue; |
2308 | |
2309 | // This may be a skipped class |
2310 | if (!VRM->hasPhys(virtReg: Reg)) { |
2311 | assert(!ShouldAllocateClass(*TRI, *MRI->getRegClass(Reg)) && |
2312 | "We have an unallocated variable which should have been handled" ); |
2313 | continue; |
2314 | } |
2315 | |
2316 | // Get the live interval mapped with this virtual register to be able |
2317 | // to check for the interference with the new color. |
2318 | LiveInterval &LI = LIS->getInterval(Reg); |
2319 | MCRegister CurrPhys = VRM->getPhys(virtReg: Reg); |
2320 | // Check that the new color matches the register class constraints and |
2321 | // that it is free for this live range. |
2322 | if (CurrPhys != PhysReg && (!MRI->getRegClass(Reg)->contains(Reg: PhysReg) || |
2323 | Matrix->checkInterference(VirtReg: LI, PhysReg))) |
2324 | continue; |
2325 | |
2326 | LLVM_DEBUG(dbgs() << printReg(Reg, TRI) << '(' << printReg(CurrPhys, TRI) |
2327 | << ") is recolorable.\n" ); |
2328 | |
2329 | // Gather the hint info. |
2330 | Info.clear(); |
2331 | collectHintInfo(Reg, Out&: Info); |
2332 | // Check if recoloring the live-range will increase the cost of the |
2333 | // non-identity copies. |
2334 | if (CurrPhys != PhysReg) { |
2335 | LLVM_DEBUG(dbgs() << "Checking profitability:\n" ); |
2336 | BlockFrequency OldCopiesCost = getBrokenHintFreq(List: Info, PhysReg: CurrPhys); |
2337 | BlockFrequency NewCopiesCost = getBrokenHintFreq(List: Info, PhysReg); |
2338 | LLVM_DEBUG(dbgs() << "Old Cost: " << printBlockFreq(*MBFI, OldCopiesCost) |
2339 | << "\nNew Cost: " |
2340 | << printBlockFreq(*MBFI, NewCopiesCost) << '\n'); |
2341 | if (OldCopiesCost < NewCopiesCost) { |
2342 | LLVM_DEBUG(dbgs() << "=> Not profitable.\n" ); |
2343 | continue; |
2344 | } |
2345 | // At this point, the cost is either cheaper or equal. If it is |
2346 | // equal, we consider this is profitable because it may expose |
2347 | // more recoloring opportunities. |
2348 | LLVM_DEBUG(dbgs() << "=> Profitable.\n" ); |
2349 | // Recolor the live-range. |
2350 | Matrix->unassign(VirtReg: LI); |
2351 | Matrix->assign(VirtReg: LI, PhysReg); |
2352 | } |
2353 | // Push all copy-related live-ranges to keep reconciling the broken |
2354 | // hints. |
2355 | for (const HintInfo &HI : Info) { |
2356 | if (Visited.insert(V: HI.Reg).second) |
2357 | RecoloringCandidates.push_back(Elt: HI.Reg); |
2358 | } |
2359 | } while (!RecoloringCandidates.empty()); |
2360 | } |
2361 | |
2362 | /// Try to recolor broken hints. |
2363 | /// Broken hints may be repaired by recoloring when an evicted variable |
2364 | /// freed up a register for a larger live-range. |
2365 | /// Consider the following example: |
2366 | /// BB1: |
2367 | /// a = |
2368 | /// b = |
2369 | /// BB2: |
2370 | /// ... |
2371 | /// = b |
2372 | /// = a |
2373 | /// Let us assume b gets split: |
2374 | /// BB1: |
2375 | /// a = |
2376 | /// b = |
2377 | /// BB2: |
2378 | /// c = b |
2379 | /// ... |
2380 | /// d = c |
2381 | /// = d |
2382 | /// = a |
2383 | /// Because of how the allocation work, b, c, and d may be assigned different |
2384 | /// colors. Now, if a gets evicted later: |
2385 | /// BB1: |
2386 | /// a = |
2387 | /// st a, SpillSlot |
2388 | /// b = |
2389 | /// BB2: |
2390 | /// c = b |
2391 | /// ... |
2392 | /// d = c |
2393 | /// = d |
2394 | /// e = ld SpillSlot |
2395 | /// = e |
2396 | /// This is likely that we can assign the same register for b, c, and d, |
2397 | /// getting rid of 2 copies. |
2398 | void RAGreedy::tryHintsRecoloring() { |
2399 | for (const LiveInterval *LI : SetOfBrokenHints) { |
2400 | assert(LI->reg().isVirtual() && |
2401 | "Recoloring is possible only for virtual registers" ); |
2402 | // Some dead defs may be around (e.g., because of debug uses). |
2403 | // Ignore those. |
2404 | if (!VRM->hasPhys(virtReg: LI->reg())) |
2405 | continue; |
2406 | tryHintRecoloring(VirtReg: *LI); |
2407 | } |
2408 | } |
2409 | |
2410 | MCRegister RAGreedy::selectOrSplitImpl(const LiveInterval &VirtReg, |
2411 | SmallVectorImpl<Register> &NewVRegs, |
2412 | SmallVirtRegSet &FixedRegisters, |
2413 | RecoloringStack &RecolorStack, |
2414 | unsigned Depth) { |
2415 | uint8_t CostPerUseLimit = uint8_t(~0u); |
2416 | // First try assigning a free register. |
2417 | auto Order = |
2418 | AllocationOrder::create(VirtReg: VirtReg.reg(), VRM: *VRM, RegClassInfo, Matrix); |
2419 | if (MCRegister PhysReg = |
2420 | tryAssign(VirtReg, Order, NewVRegs, FixedRegisters)) { |
2421 | // When NewVRegs is not empty, we may have made decisions such as evicting |
2422 | // a virtual register, go with the earlier decisions and use the physical |
2423 | // register. |
2424 | if (CSRCost.getFrequency() && |
2425 | EvictAdvisor->isUnusedCalleeSavedReg(PhysReg) && NewVRegs.empty()) { |
2426 | MCRegister CSRReg = tryAssignCSRFirstTime(VirtReg, Order, PhysReg, |
2427 | CostPerUseLimit, NewVRegs); |
2428 | if (CSRReg || !NewVRegs.empty()) |
2429 | // Return now if we decide to use a CSR or create new vregs due to |
2430 | // pre-splitting. |
2431 | return CSRReg; |
2432 | } else |
2433 | return PhysReg; |
2434 | } |
2435 | // Non emtpy NewVRegs means VirtReg has been split. |
2436 | if (!NewVRegs.empty()) |
2437 | return 0; |
2438 | |
2439 | LiveRangeStage Stage = ExtraInfo->getStage(VirtReg); |
2440 | LLVM_DEBUG(dbgs() << StageName[Stage] << " Cascade " |
2441 | << ExtraInfo->getCascade(VirtReg.reg()) << '\n'); |
2442 | |
2443 | // Try to evict a less worthy live range, but only for ranges from the primary |
2444 | // queue. The RS_Split ranges already failed to do this, and they should not |
2445 | // get a second chance until they have been split. |
2446 | if (Stage != RS_Split) |
2447 | if (Register PhysReg = |
2448 | tryEvict(VirtReg, Order, NewVRegs, CostPerUseLimit, |
2449 | FixedRegisters)) { |
2450 | Register Hint = MRI->getSimpleHint(VReg: VirtReg.reg()); |
2451 | // If VirtReg has a hint and that hint is broken record this |
2452 | // virtual register as a recoloring candidate for broken hint. |
2453 | // Indeed, since we evicted a variable in its neighborhood it is |
2454 | // likely we can at least partially recolor some of the |
2455 | // copy-related live-ranges. |
2456 | if (Hint && Hint != PhysReg) |
2457 | SetOfBrokenHints.insert(X: &VirtReg); |
2458 | return PhysReg; |
2459 | } |
2460 | |
2461 | assert((NewVRegs.empty() || Depth) && "Cannot append to existing NewVRegs" ); |
2462 | |
2463 | // The first time we see a live range, don't try to split or spill. |
2464 | // Wait until the second time, when all smaller ranges have been allocated. |
2465 | // This gives a better picture of the interference to split around. |
2466 | if (Stage < RS_Split) { |
2467 | ExtraInfo->setStage(VirtReg, Stage: RS_Split); |
2468 | LLVM_DEBUG(dbgs() << "wait for second round\n" ); |
2469 | NewVRegs.push_back(Elt: VirtReg.reg()); |
2470 | return 0; |
2471 | } |
2472 | |
2473 | if (Stage < RS_Spill) { |
2474 | // Try splitting VirtReg or interferences. |
2475 | unsigned NewVRegSizeBefore = NewVRegs.size(); |
2476 | Register PhysReg = trySplit(VirtReg, Order, NewVRegs, FixedRegisters); |
2477 | if (PhysReg || (NewVRegs.size() - NewVRegSizeBefore)) |
2478 | return PhysReg; |
2479 | } |
2480 | |
2481 | // If we couldn't allocate a register from spilling, there is probably some |
2482 | // invalid inline assembly. The base class will report it. |
2483 | if (Stage >= RS_Done || !VirtReg.isSpillable()) { |
2484 | return tryLastChanceRecoloring(VirtReg, Order, NewVRegs, FixedRegisters, |
2485 | RecolorStack, Depth); |
2486 | } |
2487 | |
2488 | // Finally spill VirtReg itself. |
2489 | if ((EnableDeferredSpilling || |
2490 | TRI->shouldUseDeferredSpillingForVirtReg(MF: *MF, VirtReg)) && |
2491 | ExtraInfo->getStage(VirtReg) < RS_Memory) { |
2492 | // TODO: This is experimental and in particular, we do not model |
2493 | // the live range splitting done by spilling correctly. |
2494 | // We would need a deep integration with the spiller to do the |
2495 | // right thing here. Anyway, that is still good for early testing. |
2496 | ExtraInfo->setStage(VirtReg, Stage: RS_Memory); |
2497 | LLVM_DEBUG(dbgs() << "Do as if this register is in memory\n" ); |
2498 | NewVRegs.push_back(Elt: VirtReg.reg()); |
2499 | } else { |
2500 | NamedRegionTimer T("spill" , "Spiller" , TimerGroupName, |
2501 | TimerGroupDescription, TimePassesIsEnabled); |
2502 | LiveRangeEdit LRE(&VirtReg, NewVRegs, *MF, *LIS, VRM, this, &DeadRemats); |
2503 | spiller().spill(LRE); |
2504 | ExtraInfo->setStage(Begin: NewVRegs.begin(), End: NewVRegs.end(), NewStage: RS_Done); |
2505 | |
2506 | // Tell LiveDebugVariables about the new ranges. Ranges not being covered by |
2507 | // the new regs are kept in LDV (still mapping to the old register), until |
2508 | // we rewrite spilled locations in LDV at a later stage. |
2509 | DebugVars->splitRegister(OldReg: VirtReg.reg(), NewRegs: LRE.regs(), LIS&: *LIS); |
2510 | |
2511 | if (VerifyEnabled) |
2512 | MF->verify(p: this, Banner: "After spilling" ); |
2513 | } |
2514 | |
2515 | // The live virtual register requesting allocation was spilled, so tell |
2516 | // the caller not to allocate anything during this round. |
2517 | return 0; |
2518 | } |
2519 | |
2520 | void RAGreedy::RAGreedyStats::(MachineOptimizationRemarkMissed &R) { |
2521 | using namespace ore; |
2522 | if (Spills) { |
2523 | R << NV("NumSpills" , Spills) << " spills " ; |
2524 | R << NV("TotalSpillsCost" , SpillsCost) << " total spills cost " ; |
2525 | } |
2526 | if (FoldedSpills) { |
2527 | R << NV("NumFoldedSpills" , FoldedSpills) << " folded spills " ; |
2528 | R << NV("TotalFoldedSpillsCost" , FoldedSpillsCost) |
2529 | << " total folded spills cost " ; |
2530 | } |
2531 | if (Reloads) { |
2532 | R << NV("NumReloads" , Reloads) << " reloads " ; |
2533 | R << NV("TotalReloadsCost" , ReloadsCost) << " total reloads cost " ; |
2534 | } |
2535 | if (FoldedReloads) { |
2536 | R << NV("NumFoldedReloads" , FoldedReloads) << " folded reloads " ; |
2537 | R << NV("TotalFoldedReloadsCost" , FoldedReloadsCost) |
2538 | << " total folded reloads cost " ; |
2539 | } |
2540 | if (ZeroCostFoldedReloads) |
2541 | R << NV("NumZeroCostFoldedReloads" , ZeroCostFoldedReloads) |
2542 | << " zero cost folded reloads " ; |
2543 | if (Copies) { |
2544 | R << NV("NumVRCopies" , Copies) << " virtual registers copies " ; |
2545 | R << NV("TotalCopiesCost" , CopiesCost) << " total copies cost " ; |
2546 | } |
2547 | } |
2548 | |
2549 | RAGreedy::RAGreedyStats RAGreedy::computeStats(MachineBasicBlock &MBB) { |
2550 | RAGreedyStats Stats; |
2551 | const MachineFrameInfo &MFI = MF->getFrameInfo(); |
2552 | int FI; |
2553 | |
2554 | auto isSpillSlotAccess = [&MFI](const MachineMemOperand *A) { |
2555 | return MFI.isSpillSlotObjectIndex(ObjectIdx: cast<FixedStackPseudoSourceValue>( |
2556 | Val: A->getPseudoValue())->getFrameIndex()); |
2557 | }; |
2558 | auto isPatchpointInstr = [](const MachineInstr &MI) { |
2559 | return MI.getOpcode() == TargetOpcode::PATCHPOINT || |
2560 | MI.getOpcode() == TargetOpcode::STACKMAP || |
2561 | MI.getOpcode() == TargetOpcode::STATEPOINT; |
2562 | }; |
2563 | for (MachineInstr &MI : MBB) { |
2564 | auto DestSrc = TII->isCopyInstr(MI); |
2565 | if (DestSrc) { |
2566 | const MachineOperand &Dest = *DestSrc->Destination; |
2567 | const MachineOperand &Src = *DestSrc->Source; |
2568 | Register SrcReg = Src.getReg(); |
2569 | Register DestReg = Dest.getReg(); |
2570 | // Only count `COPY`s with a virtual register as source or destination. |
2571 | if (SrcReg.isVirtual() || DestReg.isVirtual()) { |
2572 | if (SrcReg.isVirtual()) { |
2573 | SrcReg = VRM->getPhys(virtReg: SrcReg); |
2574 | if (SrcReg && Src.getSubReg()) |
2575 | SrcReg = TRI->getSubReg(Reg: SrcReg, Idx: Src.getSubReg()); |
2576 | } |
2577 | if (DestReg.isVirtual()) { |
2578 | DestReg = VRM->getPhys(virtReg: DestReg); |
2579 | if (DestReg && Dest.getSubReg()) |
2580 | DestReg = TRI->getSubReg(Reg: DestReg, Idx: Dest.getSubReg()); |
2581 | } |
2582 | if (SrcReg != DestReg) |
2583 | ++Stats.Copies; |
2584 | } |
2585 | continue; |
2586 | } |
2587 | |
2588 | SmallVector<const MachineMemOperand *, 2> Accesses; |
2589 | if (TII->isLoadFromStackSlot(MI, FrameIndex&: FI) && MFI.isSpillSlotObjectIndex(ObjectIdx: FI)) { |
2590 | ++Stats.Reloads; |
2591 | continue; |
2592 | } |
2593 | if (TII->isStoreToStackSlot(MI, FrameIndex&: FI) && MFI.isSpillSlotObjectIndex(ObjectIdx: FI)) { |
2594 | ++Stats.Spills; |
2595 | continue; |
2596 | } |
2597 | if (TII->hasLoadFromStackSlot(MI, Accesses) && |
2598 | llvm::any_of(Range&: Accesses, P: isSpillSlotAccess)) { |
2599 | if (!isPatchpointInstr(MI)) { |
2600 | Stats.FoldedReloads += Accesses.size(); |
2601 | continue; |
2602 | } |
2603 | // For statepoint there may be folded and zero cost folded stack reloads. |
2604 | std::pair<unsigned, unsigned> NonZeroCostRange = |
2605 | TII->getPatchpointUnfoldableRange(MI); |
2606 | SmallSet<unsigned, 16> FoldedReloads; |
2607 | SmallSet<unsigned, 16> ZeroCostFoldedReloads; |
2608 | for (unsigned Idx = 0, E = MI.getNumOperands(); Idx < E; ++Idx) { |
2609 | MachineOperand &MO = MI.getOperand(i: Idx); |
2610 | if (!MO.isFI() || !MFI.isSpillSlotObjectIndex(ObjectIdx: MO.getIndex())) |
2611 | continue; |
2612 | if (Idx >= NonZeroCostRange.first && Idx < NonZeroCostRange.second) |
2613 | FoldedReloads.insert(V: MO.getIndex()); |
2614 | else |
2615 | ZeroCostFoldedReloads.insert(V: MO.getIndex()); |
2616 | } |
2617 | // If stack slot is used in folded reload it is not zero cost then. |
2618 | for (unsigned Slot : FoldedReloads) |
2619 | ZeroCostFoldedReloads.erase(V: Slot); |
2620 | Stats.FoldedReloads += FoldedReloads.size(); |
2621 | Stats.ZeroCostFoldedReloads += ZeroCostFoldedReloads.size(); |
2622 | continue; |
2623 | } |
2624 | Accesses.clear(); |
2625 | if (TII->hasStoreToStackSlot(MI, Accesses) && |
2626 | llvm::any_of(Range&: Accesses, P: isSpillSlotAccess)) { |
2627 | Stats.FoldedSpills += Accesses.size(); |
2628 | } |
2629 | } |
2630 | // Set cost of collected statistic by multiplication to relative frequency of |
2631 | // this basic block. |
2632 | float RelFreq = MBFI->getBlockFreqRelativeToEntryBlock(MBB: &MBB); |
2633 | Stats.ReloadsCost = RelFreq * Stats.Reloads; |
2634 | Stats.FoldedReloadsCost = RelFreq * Stats.FoldedReloads; |
2635 | Stats.SpillsCost = RelFreq * Stats.Spills; |
2636 | Stats.FoldedSpillsCost = RelFreq * Stats.FoldedSpills; |
2637 | Stats.CopiesCost = RelFreq * Stats.Copies; |
2638 | return Stats; |
2639 | } |
2640 | |
2641 | RAGreedy::RAGreedyStats RAGreedy::reportStats(MachineLoop *L) { |
2642 | RAGreedyStats Stats; |
2643 | |
2644 | // Sum up the spill and reloads in subloops. |
2645 | for (MachineLoop *SubLoop : *L) |
2646 | Stats.add(other: reportStats(L: SubLoop)); |
2647 | |
2648 | for (MachineBasicBlock *MBB : L->getBlocks()) |
2649 | // Handle blocks that were not included in subloops. |
2650 | if (Loops->getLoopFor(BB: MBB) == L) |
2651 | Stats.add(other: computeStats(MBB&: *MBB)); |
2652 | |
2653 | if (!Stats.isEmpty()) { |
2654 | using namespace ore; |
2655 | |
2656 | ORE->emit(RemarkBuilder: [&]() { |
2657 | MachineOptimizationRemarkMissed R(DEBUG_TYPE, "LoopSpillReloadCopies" , |
2658 | L->getStartLoc(), L->getHeader()); |
2659 | Stats.report(R); |
2660 | R << "generated in loop" ; |
2661 | return R; |
2662 | }); |
2663 | } |
2664 | return Stats; |
2665 | } |
2666 | |
2667 | void RAGreedy::reportStats() { |
2668 | if (!ORE->allowExtraAnalysis(DEBUG_TYPE)) |
2669 | return; |
2670 | RAGreedyStats Stats; |
2671 | for (MachineLoop *L : *Loops) |
2672 | Stats.add(other: reportStats(L)); |
2673 | // Process non-loop blocks. |
2674 | for (MachineBasicBlock &MBB : *MF) |
2675 | if (!Loops->getLoopFor(BB: &MBB)) |
2676 | Stats.add(other: computeStats(MBB)); |
2677 | if (!Stats.isEmpty()) { |
2678 | using namespace ore; |
2679 | |
2680 | ORE->emit(RemarkBuilder: [&]() { |
2681 | DebugLoc Loc; |
2682 | if (auto *SP = MF->getFunction().getSubprogram()) |
2683 | Loc = DILocation::get(Context&: SP->getContext(), Line: SP->getLine(), Column: 1, Scope: SP); |
2684 | MachineOptimizationRemarkMissed R(DEBUG_TYPE, "SpillReloadCopies" , Loc, |
2685 | &MF->front()); |
2686 | Stats.report(R); |
2687 | R << "generated in function" ; |
2688 | return R; |
2689 | }); |
2690 | } |
2691 | } |
2692 | |
2693 | bool RAGreedy::hasVirtRegAlloc() { |
2694 | for (unsigned I = 0, E = MRI->getNumVirtRegs(); I != E; ++I) { |
2695 | Register Reg = Register::index2VirtReg(Index: I); |
2696 | if (MRI->reg_nodbg_empty(RegNo: Reg)) |
2697 | continue; |
2698 | const TargetRegisterClass *RC = MRI->getRegClass(Reg); |
2699 | if (!RC) |
2700 | continue; |
2701 | if (ShouldAllocateClass(*TRI, *RC)) |
2702 | return true; |
2703 | } |
2704 | |
2705 | return false; |
2706 | } |
2707 | |
2708 | bool RAGreedy::runOnMachineFunction(MachineFunction &mf) { |
2709 | LLVM_DEBUG(dbgs() << "********** GREEDY REGISTER ALLOCATION **********\n" |
2710 | << "********** Function: " << mf.getName() << '\n'); |
2711 | |
2712 | MF = &mf; |
2713 | TII = MF->getSubtarget().getInstrInfo(); |
2714 | |
2715 | if (VerifyEnabled) |
2716 | MF->verify(p: this, Banner: "Before greedy register allocator" ); |
2717 | |
2718 | RegAllocBase::init(vrm&: getAnalysis<VirtRegMap>(), |
2719 | lis&: getAnalysis<LiveIntervals>(), |
2720 | mat&: getAnalysis<LiveRegMatrix>()); |
2721 | |
2722 | // Early return if there is no virtual register to be allocated to a |
2723 | // physical register. |
2724 | if (!hasVirtRegAlloc()) |
2725 | return false; |
2726 | |
2727 | Indexes = &getAnalysis<SlotIndexes>(); |
2728 | // Renumber to get accurate and consistent results from |
2729 | // SlotIndexes::getApproxInstrDistance. |
2730 | Indexes->packIndexes(); |
2731 | MBFI = &getAnalysis<MachineBlockFrequencyInfo>(); |
2732 | DomTree = &getAnalysis<MachineDominatorTree>(); |
2733 | ORE = &getAnalysis<MachineOptimizationRemarkEmitterPass>().getORE(); |
2734 | Loops = &getAnalysis<MachineLoopInfo>(); |
2735 | Bundles = &getAnalysis<EdgeBundles>(); |
2736 | SpillPlacer = &getAnalysis<SpillPlacement>(); |
2737 | DebugVars = &getAnalysis<LiveDebugVariables>(); |
2738 | |
2739 | initializeCSRCost(); |
2740 | |
2741 | RegCosts = TRI->getRegisterCosts(MF: *MF); |
2742 | RegClassPriorityTrumpsGlobalness = |
2743 | GreedyRegClassPriorityTrumpsGlobalness.getNumOccurrences() |
2744 | ? GreedyRegClassPriorityTrumpsGlobalness |
2745 | : TRI->regClassPriorityTrumpsGlobalness(MF: *MF); |
2746 | |
2747 | ReverseLocalAssignment = GreedyReverseLocalAssignment.getNumOccurrences() |
2748 | ? GreedyReverseLocalAssignment |
2749 | : TRI->reverseLocalAssignment(); |
2750 | |
2751 | ExtraInfo.emplace(); |
2752 | EvictAdvisor = |
2753 | getAnalysis<RegAllocEvictionAdvisorAnalysis>().getAdvisor(MF: *MF, RA: *this); |
2754 | PriorityAdvisor = |
2755 | getAnalysis<RegAllocPriorityAdvisorAnalysis>().getAdvisor(MF: *MF, RA: *this); |
2756 | |
2757 | VRAI = std::make_unique<VirtRegAuxInfo>(args&: *MF, args&: *LIS, args&: *VRM, args&: *Loops, args&: *MBFI); |
2758 | SpillerInstance.reset(p: createInlineSpiller(Pass&: *this, MF&: *MF, VRM&: *VRM, VRAI&: *VRAI)); |
2759 | |
2760 | VRAI->calculateSpillWeightsAndHints(); |
2761 | |
2762 | LLVM_DEBUG(LIS->dump()); |
2763 | |
2764 | SA.reset(p: new SplitAnalysis(*VRM, *LIS, *Loops)); |
2765 | SE.reset(p: new SplitEditor(*SA, *LIS, *VRM, *DomTree, *MBFI, *VRAI)); |
2766 | |
2767 | IntfCache.init(mf: MF, liuarray: Matrix->getLiveUnions(), indexes: Indexes, lis: LIS, tri: TRI); |
2768 | GlobalCand.resize(N: 32); // This will grow as needed. |
2769 | SetOfBrokenHints.clear(); |
2770 | |
2771 | allocatePhysRegs(); |
2772 | tryHintsRecoloring(); |
2773 | |
2774 | if (VerifyEnabled) |
2775 | MF->verify(p: this, Banner: "Before post optimization" ); |
2776 | postOptimization(); |
2777 | reportStats(); |
2778 | |
2779 | releaseMemory(); |
2780 | return true; |
2781 | } |
2782 | |