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