File: | src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/CodeGen/InlineSpiller.cpp |
Warning: | line 1348, column 14 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===- InlineSpiller.cpp - Insert spills and restores inline --------------===// | |||
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 | // The inline spiller modifies the machine function directly instead of | |||
10 | // inserting spills and restores in VirtRegMap. | |||
11 | // | |||
12 | //===----------------------------------------------------------------------===// | |||
13 | ||||
14 | #include "SplitKit.h" | |||
15 | #include "llvm/ADT/ArrayRef.h" | |||
16 | #include "llvm/ADT/DenseMap.h" | |||
17 | #include "llvm/ADT/MapVector.h" | |||
18 | #include "llvm/ADT/None.h" | |||
19 | #include "llvm/ADT/STLExtras.h" | |||
20 | #include "llvm/ADT/SetVector.h" | |||
21 | #include "llvm/ADT/SmallPtrSet.h" | |||
22 | #include "llvm/ADT/SmallVector.h" | |||
23 | #include "llvm/ADT/Statistic.h" | |||
24 | #include "llvm/Analysis/AliasAnalysis.h" | |||
25 | #include "llvm/CodeGen/LiveInterval.h" | |||
26 | #include "llvm/CodeGen/LiveIntervalCalc.h" | |||
27 | #include "llvm/CodeGen/LiveIntervals.h" | |||
28 | #include "llvm/CodeGen/LiveRangeEdit.h" | |||
29 | #include "llvm/CodeGen/LiveStacks.h" | |||
30 | #include "llvm/CodeGen/MachineBasicBlock.h" | |||
31 | #include "llvm/CodeGen/MachineBlockFrequencyInfo.h" | |||
32 | #include "llvm/CodeGen/MachineDominators.h" | |||
33 | #include "llvm/CodeGen/MachineFunction.h" | |||
34 | #include "llvm/CodeGen/MachineFunctionPass.h" | |||
35 | #include "llvm/CodeGen/MachineInstr.h" | |||
36 | #include "llvm/CodeGen/MachineInstrBuilder.h" | |||
37 | #include "llvm/CodeGen/MachineInstrBundle.h" | |||
38 | #include "llvm/CodeGen/MachineLoopInfo.h" | |||
39 | #include "llvm/CodeGen/MachineOperand.h" | |||
40 | #include "llvm/CodeGen/MachineRegisterInfo.h" | |||
41 | #include "llvm/CodeGen/SlotIndexes.h" | |||
42 | #include "llvm/CodeGen/Spiller.h" | |||
43 | #include "llvm/CodeGen/StackMaps.h" | |||
44 | #include "llvm/CodeGen/TargetInstrInfo.h" | |||
45 | #include "llvm/CodeGen/TargetOpcodes.h" | |||
46 | #include "llvm/CodeGen/TargetRegisterInfo.h" | |||
47 | #include "llvm/CodeGen/TargetSubtargetInfo.h" | |||
48 | #include "llvm/CodeGen/VirtRegMap.h" | |||
49 | #include "llvm/Config/llvm-config.h" | |||
50 | #include "llvm/Support/BlockFrequency.h" | |||
51 | #include "llvm/Support/BranchProbability.h" | |||
52 | #include "llvm/Support/CommandLine.h" | |||
53 | #include "llvm/Support/Compiler.h" | |||
54 | #include "llvm/Support/Debug.h" | |||
55 | #include "llvm/Support/ErrorHandling.h" | |||
56 | #include "llvm/Support/raw_ostream.h" | |||
57 | #include <cassert> | |||
58 | #include <iterator> | |||
59 | #include <tuple> | |||
60 | #include <utility> | |||
61 | #include <vector> | |||
62 | ||||
63 | using namespace llvm; | |||
64 | ||||
65 | #define DEBUG_TYPE"regalloc" "regalloc" | |||
66 | ||||
67 | STATISTIC(NumSpilledRanges, "Number of spilled live ranges")static llvm::Statistic NumSpilledRanges = {"regalloc", "NumSpilledRanges" , "Number of spilled live ranges"}; | |||
68 | STATISTIC(NumSnippets, "Number of spilled snippets")static llvm::Statistic NumSnippets = {"regalloc", "NumSnippets" , "Number of spilled snippets"}; | |||
69 | STATISTIC(NumSpills, "Number of spills inserted")static llvm::Statistic NumSpills = {"regalloc", "NumSpills", "Number of spills inserted" }; | |||
70 | STATISTIC(NumSpillsRemoved, "Number of spills removed")static llvm::Statistic NumSpillsRemoved = {"regalloc", "NumSpillsRemoved" , "Number of spills removed"}; | |||
71 | STATISTIC(NumReloads, "Number of reloads inserted")static llvm::Statistic NumReloads = {"regalloc", "NumReloads" , "Number of reloads inserted"}; | |||
72 | STATISTIC(NumReloadsRemoved, "Number of reloads removed")static llvm::Statistic NumReloadsRemoved = {"regalloc", "NumReloadsRemoved" , "Number of reloads removed"}; | |||
73 | STATISTIC(NumFolded, "Number of folded stack accesses")static llvm::Statistic NumFolded = {"regalloc", "NumFolded", "Number of folded stack accesses" }; | |||
74 | STATISTIC(NumFoldedLoads, "Number of folded loads")static llvm::Statistic NumFoldedLoads = {"regalloc", "NumFoldedLoads" , "Number of folded loads"}; | |||
75 | STATISTIC(NumRemats, "Number of rematerialized defs for spilling")static llvm::Statistic NumRemats = {"regalloc", "NumRemats", "Number of rematerialized defs for spilling" }; | |||
76 | ||||
77 | static cl::opt<bool> DisableHoisting("disable-spill-hoist", cl::Hidden, | |||
78 | cl::desc("Disable inline spill hoisting")); | |||
79 | static cl::opt<bool> | |||
80 | RestrictStatepointRemat("restrict-statepoint-remat", | |||
81 | cl::init(false), cl::Hidden, | |||
82 | cl::desc("Restrict remat for statepoint operands")); | |||
83 | ||||
84 | namespace { | |||
85 | ||||
86 | class HoistSpillHelper : private LiveRangeEdit::Delegate { | |||
87 | MachineFunction &MF; | |||
88 | LiveIntervals &LIS; | |||
89 | LiveStacks &LSS; | |||
90 | AliasAnalysis *AA; | |||
91 | MachineDominatorTree &MDT; | |||
92 | MachineLoopInfo &Loops; | |||
93 | VirtRegMap &VRM; | |||
94 | MachineRegisterInfo &MRI; | |||
95 | const TargetInstrInfo &TII; | |||
96 | const TargetRegisterInfo &TRI; | |||
97 | const MachineBlockFrequencyInfo &MBFI; | |||
98 | ||||
99 | InsertPointAnalysis IPA; | |||
100 | ||||
101 | // Map from StackSlot to the LiveInterval of the original register. | |||
102 | // Note the LiveInterval of the original register may have been deleted | |||
103 | // after it is spilled. We keep a copy here to track the range where | |||
104 | // spills can be moved. | |||
105 | DenseMap<int, std::unique_ptr<LiveInterval>> StackSlotToOrigLI; | |||
106 | ||||
107 | // Map from pair of (StackSlot and Original VNI) to a set of spills which | |||
108 | // have the same stackslot and have equal values defined by Original VNI. | |||
109 | // These spills are mergeable and are hoist candiates. | |||
110 | using MergeableSpillsMap = | |||
111 | MapVector<std::pair<int, VNInfo *>, SmallPtrSet<MachineInstr *, 16>>; | |||
112 | MergeableSpillsMap MergeableSpills; | |||
113 | ||||
114 | /// This is the map from original register to a set containing all its | |||
115 | /// siblings. To hoist a spill to another BB, we need to find out a live | |||
116 | /// sibling there and use it as the source of the new spill. | |||
117 | DenseMap<Register, SmallSetVector<Register, 16>> Virt2SiblingsMap; | |||
118 | ||||
119 | bool isSpillCandBB(LiveInterval &OrigLI, VNInfo &OrigVNI, | |||
120 | MachineBasicBlock &BB, Register &LiveReg); | |||
121 | ||||
122 | void rmRedundantSpills( | |||
123 | SmallPtrSet<MachineInstr *, 16> &Spills, | |||
124 | SmallVectorImpl<MachineInstr *> &SpillsToRm, | |||
125 | DenseMap<MachineDomTreeNode *, MachineInstr *> &SpillBBToSpill); | |||
126 | ||||
127 | void getVisitOrders( | |||
128 | MachineBasicBlock *Root, SmallPtrSet<MachineInstr *, 16> &Spills, | |||
129 | SmallVectorImpl<MachineDomTreeNode *> &Orders, | |||
130 | SmallVectorImpl<MachineInstr *> &SpillsToRm, | |||
131 | DenseMap<MachineDomTreeNode *, unsigned> &SpillsToKeep, | |||
132 | DenseMap<MachineDomTreeNode *, MachineInstr *> &SpillBBToSpill); | |||
133 | ||||
134 | void runHoistSpills(LiveInterval &OrigLI, VNInfo &OrigVNI, | |||
135 | SmallPtrSet<MachineInstr *, 16> &Spills, | |||
136 | SmallVectorImpl<MachineInstr *> &SpillsToRm, | |||
137 | DenseMap<MachineBasicBlock *, unsigned> &SpillsToIns); | |||
138 | ||||
139 | public: | |||
140 | HoistSpillHelper(MachineFunctionPass &pass, MachineFunction &mf, | |||
141 | VirtRegMap &vrm) | |||
142 | : MF(mf), LIS(pass.getAnalysis<LiveIntervals>()), | |||
143 | LSS(pass.getAnalysis<LiveStacks>()), | |||
144 | AA(&pass.getAnalysis<AAResultsWrapperPass>().getAAResults()), | |||
145 | MDT(pass.getAnalysis<MachineDominatorTree>()), | |||
146 | Loops(pass.getAnalysis<MachineLoopInfo>()), VRM(vrm), | |||
147 | MRI(mf.getRegInfo()), TII(*mf.getSubtarget().getInstrInfo()), | |||
148 | TRI(*mf.getSubtarget().getRegisterInfo()), | |||
149 | MBFI(pass.getAnalysis<MachineBlockFrequencyInfo>()), | |||
150 | IPA(LIS, mf.getNumBlockIDs()) {} | |||
151 | ||||
152 | void addToMergeableSpills(MachineInstr &Spill, int StackSlot, | |||
153 | unsigned Original); | |||
154 | bool rmFromMergeableSpills(MachineInstr &Spill, int StackSlot); | |||
155 | void hoistAllSpills(); | |||
156 | void LRE_DidCloneVirtReg(Register, Register) override; | |||
157 | }; | |||
158 | ||||
159 | class InlineSpiller : public Spiller { | |||
160 | MachineFunction &MF; | |||
161 | LiveIntervals &LIS; | |||
162 | LiveStacks &LSS; | |||
163 | AliasAnalysis *AA; | |||
164 | MachineDominatorTree &MDT; | |||
165 | MachineLoopInfo &Loops; | |||
166 | VirtRegMap &VRM; | |||
167 | MachineRegisterInfo &MRI; | |||
168 | const TargetInstrInfo &TII; | |||
169 | const TargetRegisterInfo &TRI; | |||
170 | const MachineBlockFrequencyInfo &MBFI; | |||
171 | ||||
172 | // Variables that are valid during spill(), but used by multiple methods. | |||
173 | LiveRangeEdit *Edit; | |||
174 | LiveInterval *StackInt; | |||
175 | int StackSlot; | |||
176 | Register Original; | |||
177 | ||||
178 | // All registers to spill to StackSlot, including the main register. | |||
179 | SmallVector<Register, 8> RegsToSpill; | |||
180 | ||||
181 | // All COPY instructions to/from snippets. | |||
182 | // They are ignored since both operands refer to the same stack slot. | |||
183 | SmallPtrSet<MachineInstr*, 8> SnippetCopies; | |||
184 | ||||
185 | // Values that failed to remat at some point. | |||
186 | SmallPtrSet<VNInfo*, 8> UsedValues; | |||
187 | ||||
188 | // Dead defs generated during spilling. | |||
189 | SmallVector<MachineInstr*, 8> DeadDefs; | |||
190 | ||||
191 | // Object records spills information and does the hoisting. | |||
192 | HoistSpillHelper HSpiller; | |||
193 | ||||
194 | // Live range weight calculator. | |||
195 | VirtRegAuxInfo &VRAI; | |||
196 | ||||
197 | ~InlineSpiller() override = default; | |||
198 | ||||
199 | public: | |||
200 | InlineSpiller(MachineFunctionPass &Pass, MachineFunction &MF, VirtRegMap &VRM, | |||
201 | VirtRegAuxInfo &VRAI) | |||
202 | : MF(MF), LIS(Pass.getAnalysis<LiveIntervals>()), | |||
203 | LSS(Pass.getAnalysis<LiveStacks>()), | |||
204 | AA(&Pass.getAnalysis<AAResultsWrapperPass>().getAAResults()), | |||
205 | MDT(Pass.getAnalysis<MachineDominatorTree>()), | |||
206 | Loops(Pass.getAnalysis<MachineLoopInfo>()), VRM(VRM), | |||
207 | MRI(MF.getRegInfo()), TII(*MF.getSubtarget().getInstrInfo()), | |||
208 | TRI(*MF.getSubtarget().getRegisterInfo()), | |||
209 | MBFI(Pass.getAnalysis<MachineBlockFrequencyInfo>()), | |||
210 | HSpiller(Pass, MF, VRM), VRAI(VRAI) {} | |||
211 | ||||
212 | void spill(LiveRangeEdit &) override; | |||
213 | void postOptimization() override; | |||
214 | ||||
215 | private: | |||
216 | bool isSnippet(const LiveInterval &SnipLI); | |||
217 | void collectRegsToSpill(); | |||
218 | ||||
219 | bool isRegToSpill(Register Reg) { return is_contained(RegsToSpill, Reg); } | |||
220 | ||||
221 | bool isSibling(Register Reg); | |||
222 | bool hoistSpillInsideBB(LiveInterval &SpillLI, MachineInstr &CopyMI); | |||
223 | void eliminateRedundantSpills(LiveInterval &LI, VNInfo *VNI); | |||
224 | ||||
225 | void markValueUsed(LiveInterval*, VNInfo*); | |||
226 | bool canGuaranteeAssignmentAfterRemat(Register VReg, MachineInstr &MI); | |||
227 | bool reMaterializeFor(LiveInterval &, MachineInstr &MI); | |||
228 | void reMaterializeAll(); | |||
229 | ||||
230 | bool coalesceStackAccess(MachineInstr *MI, Register Reg); | |||
231 | bool foldMemoryOperand(ArrayRef<std::pair<MachineInstr *, unsigned>>, | |||
232 | MachineInstr *LoadMI = nullptr); | |||
233 | void insertReload(Register VReg, SlotIndex, MachineBasicBlock::iterator MI); | |||
234 | void insertSpill(Register VReg, bool isKill, MachineBasicBlock::iterator MI); | |||
235 | ||||
236 | void spillAroundUses(Register Reg); | |||
237 | void spillAll(); | |||
238 | }; | |||
239 | ||||
240 | } // end anonymous namespace | |||
241 | ||||
242 | Spiller::~Spiller() = default; | |||
243 | ||||
244 | void Spiller::anchor() {} | |||
245 | ||||
246 | Spiller *llvm::createInlineSpiller(MachineFunctionPass &Pass, | |||
247 | MachineFunction &MF, VirtRegMap &VRM, | |||
248 | VirtRegAuxInfo &VRAI) { | |||
249 | return new InlineSpiller(Pass, MF, VRM, VRAI); | |||
250 | } | |||
251 | ||||
252 | //===----------------------------------------------------------------------===// | |||
253 | // Snippets | |||
254 | //===----------------------------------------------------------------------===// | |||
255 | ||||
256 | // When spilling a virtual register, we also spill any snippets it is connected | |||
257 | // to. The snippets are small live ranges that only have a single real use, | |||
258 | // leftovers from live range splitting. Spilling them enables memory operand | |||
259 | // folding or tightens the live range around the single use. | |||
260 | // | |||
261 | // This minimizes register pressure and maximizes the store-to-load distance for | |||
262 | // spill slots which can be important in tight loops. | |||
263 | ||||
264 | /// isFullCopyOf - If MI is a COPY to or from Reg, return the other register, | |||
265 | /// otherwise return 0. | |||
266 | static Register isFullCopyOf(const MachineInstr &MI, Register Reg) { | |||
267 | if (!MI.isFullCopy()) | |||
268 | return Register(); | |||
269 | if (MI.getOperand(0).getReg() == Reg) | |||
270 | return MI.getOperand(1).getReg(); | |||
271 | if (MI.getOperand(1).getReg() == Reg) | |||
272 | return MI.getOperand(0).getReg(); | |||
273 | return Register(); | |||
274 | } | |||
275 | ||||
276 | static void getVDefInterval(const MachineInstr &MI, LiveIntervals &LIS) { | |||
277 | for (unsigned I = 0, E = MI.getNumOperands(); I != E; ++I) { | |||
278 | const MachineOperand &MO = MI.getOperand(I); | |||
279 | if (MO.isReg() && MO.isDef() && Register::isVirtualRegister(MO.getReg())) | |||
280 | LIS.getInterval(MO.getReg()); | |||
281 | } | |||
282 | } | |||
283 | ||||
284 | /// isSnippet - Identify if a live interval is a snippet that should be spilled. | |||
285 | /// It is assumed that SnipLI is a virtual register with the same original as | |||
286 | /// Edit->getReg(). | |||
287 | bool InlineSpiller::isSnippet(const LiveInterval &SnipLI) { | |||
288 | Register Reg = Edit->getReg(); | |||
289 | ||||
290 | // A snippet is a tiny live range with only a single instruction using it | |||
291 | // besides copies to/from Reg or spills/fills. We accept: | |||
292 | // | |||
293 | // %snip = COPY %Reg / FILL fi# | |||
294 | // %snip = USE %snip | |||
295 | // %Reg = COPY %snip / SPILL %snip, fi# | |||
296 | // | |||
297 | if (SnipLI.getNumValNums() > 2 || !LIS.intervalIsInOneMBB(SnipLI)) | |||
298 | return false; | |||
299 | ||||
300 | MachineInstr *UseMI = nullptr; | |||
301 | ||||
302 | // Check that all uses satisfy our criteria. | |||
303 | for (MachineRegisterInfo::reg_instr_nodbg_iterator | |||
304 | RI = MRI.reg_instr_nodbg_begin(SnipLI.reg()), | |||
305 | E = MRI.reg_instr_nodbg_end(); | |||
306 | RI != E;) { | |||
307 | MachineInstr &MI = *RI++; | |||
308 | ||||
309 | // Allow copies to/from Reg. | |||
310 | if (isFullCopyOf(MI, Reg)) | |||
311 | continue; | |||
312 | ||||
313 | // Allow stack slot loads. | |||
314 | int FI; | |||
315 | if (SnipLI.reg() == TII.isLoadFromStackSlot(MI, FI) && FI == StackSlot) | |||
316 | continue; | |||
317 | ||||
318 | // Allow stack slot stores. | |||
319 | if (SnipLI.reg() == TII.isStoreToStackSlot(MI, FI) && FI == StackSlot) | |||
320 | continue; | |||
321 | ||||
322 | // Allow a single additional instruction. | |||
323 | if (UseMI && &MI != UseMI) | |||
324 | return false; | |||
325 | UseMI = &MI; | |||
326 | } | |||
327 | return true; | |||
328 | } | |||
329 | ||||
330 | /// collectRegsToSpill - Collect live range snippets that only have a single | |||
331 | /// real use. | |||
332 | void InlineSpiller::collectRegsToSpill() { | |||
333 | Register Reg = Edit->getReg(); | |||
334 | ||||
335 | // Main register always spills. | |||
336 | RegsToSpill.assign(1, Reg); | |||
337 | SnippetCopies.clear(); | |||
338 | ||||
339 | // Snippets all have the same original, so there can't be any for an original | |||
340 | // register. | |||
341 | if (Original == Reg) | |||
342 | return; | |||
343 | ||||
344 | for (MachineRegisterInfo::reg_instr_iterator | |||
345 | RI = MRI.reg_instr_begin(Reg), E = MRI.reg_instr_end(); RI != E; ) { | |||
346 | MachineInstr &MI = *RI++; | |||
347 | Register SnipReg = isFullCopyOf(MI, Reg); | |||
348 | if (!isSibling(SnipReg)) | |||
349 | continue; | |||
350 | LiveInterval &SnipLI = LIS.getInterval(SnipReg); | |||
351 | if (!isSnippet(SnipLI)) | |||
352 | continue; | |||
353 | SnippetCopies.insert(&MI); | |||
354 | if (isRegToSpill(SnipReg)) | |||
355 | continue; | |||
356 | RegsToSpill.push_back(SnipReg); | |||
357 | LLVM_DEBUG(dbgs() << "\talso spill snippet " << SnipLI << '\n')do { } while (false); | |||
358 | ++NumSnippets; | |||
359 | } | |||
360 | } | |||
361 | ||||
362 | bool InlineSpiller::isSibling(Register Reg) { | |||
363 | return Reg.isVirtual() && VRM.getOriginal(Reg) == Original; | |||
364 | } | |||
365 | ||||
366 | /// It is beneficial to spill to earlier place in the same BB in case | |||
367 | /// as follows: | |||
368 | /// There is an alternative def earlier in the same MBB. | |||
369 | /// Hoist the spill as far as possible in SpillMBB. This can ease | |||
370 | /// register pressure: | |||
371 | /// | |||
372 | /// x = def | |||
373 | /// y = use x | |||
374 | /// s = copy x | |||
375 | /// | |||
376 | /// Hoisting the spill of s to immediately after the def removes the | |||
377 | /// interference between x and y: | |||
378 | /// | |||
379 | /// x = def | |||
380 | /// spill x | |||
381 | /// y = use killed x | |||
382 | /// | |||
383 | /// This hoist only helps when the copy kills its source. | |||
384 | /// | |||
385 | bool InlineSpiller::hoistSpillInsideBB(LiveInterval &SpillLI, | |||
386 | MachineInstr &CopyMI) { | |||
387 | SlotIndex Idx = LIS.getInstructionIndex(CopyMI); | |||
388 | #ifndef NDEBUG1 | |||
389 | VNInfo *VNI = SpillLI.getVNInfoAt(Idx.getRegSlot()); | |||
390 | assert(VNI && VNI->def == Idx.getRegSlot() && "Not defined by copy")((void)0); | |||
391 | #endif | |||
392 | ||||
393 | Register SrcReg = CopyMI.getOperand(1).getReg(); | |||
394 | LiveInterval &SrcLI = LIS.getInterval(SrcReg); | |||
395 | VNInfo *SrcVNI = SrcLI.getVNInfoAt(Idx); | |||
396 | LiveQueryResult SrcQ = SrcLI.Query(Idx); | |||
397 | MachineBasicBlock *DefMBB = LIS.getMBBFromIndex(SrcVNI->def); | |||
398 | if (DefMBB != CopyMI.getParent() || !SrcQ.isKill()) | |||
399 | return false; | |||
400 | ||||
401 | // Conservatively extend the stack slot range to the range of the original | |||
402 | // value. We may be able to do better with stack slot coloring by being more | |||
403 | // careful here. | |||
404 | assert(StackInt && "No stack slot assigned yet.")((void)0); | |||
405 | LiveInterval &OrigLI = LIS.getInterval(Original); | |||
406 | VNInfo *OrigVNI = OrigLI.getVNInfoAt(Idx); | |||
407 | StackInt->MergeValueInAsValue(OrigLI, OrigVNI, StackInt->getValNumInfo(0)); | |||
408 | LLVM_DEBUG(dbgs() << "\tmerged orig valno " << OrigVNI->id << ": "do { } while (false) | |||
409 | << *StackInt << '\n')do { } while (false); | |||
410 | ||||
411 | // We are going to spill SrcVNI immediately after its def, so clear out | |||
412 | // any later spills of the same value. | |||
413 | eliminateRedundantSpills(SrcLI, SrcVNI); | |||
414 | ||||
415 | MachineBasicBlock *MBB = LIS.getMBBFromIndex(SrcVNI->def); | |||
416 | MachineBasicBlock::iterator MII; | |||
417 | if (SrcVNI->isPHIDef()) | |||
418 | MII = MBB->SkipPHIsLabelsAndDebug(MBB->begin()); | |||
419 | else { | |||
420 | MachineInstr *DefMI = LIS.getInstructionFromIndex(SrcVNI->def); | |||
421 | assert(DefMI && "Defining instruction disappeared")((void)0); | |||
422 | MII = DefMI; | |||
423 | ++MII; | |||
424 | } | |||
425 | MachineInstrSpan MIS(MII, MBB); | |||
426 | // Insert spill without kill flag immediately after def. | |||
427 | TII.storeRegToStackSlot(*MBB, MII, SrcReg, false, StackSlot, | |||
428 | MRI.getRegClass(SrcReg), &TRI); | |||
429 | LIS.InsertMachineInstrRangeInMaps(MIS.begin(), MII); | |||
430 | for (const MachineInstr &MI : make_range(MIS.begin(), MII)) | |||
431 | getVDefInterval(MI, LIS); | |||
432 | --MII; // Point to store instruction. | |||
433 | LLVM_DEBUG(dbgs() << "\thoisted: " << SrcVNI->def << '\t' << *MII)do { } while (false); | |||
434 | ||||
435 | // If there is only 1 store instruction is required for spill, add it | |||
436 | // to mergeable list. In X86 AMX, 2 intructions are required to store. | |||
437 | // We disable the merge for this case. | |||
438 | if (MIS.begin() == MII) | |||
439 | HSpiller.addToMergeableSpills(*MII, StackSlot, Original); | |||
440 | ++NumSpills; | |||
441 | return true; | |||
442 | } | |||
443 | ||||
444 | /// eliminateRedundantSpills - SLI:VNI is known to be on the stack. Remove any | |||
445 | /// redundant spills of this value in SLI.reg and sibling copies. | |||
446 | void InlineSpiller::eliminateRedundantSpills(LiveInterval &SLI, VNInfo *VNI) { | |||
447 | assert(VNI && "Missing value")((void)0); | |||
448 | SmallVector<std::pair<LiveInterval*, VNInfo*>, 8> WorkList; | |||
449 | WorkList.push_back(std::make_pair(&SLI, VNI)); | |||
450 | assert(StackInt && "No stack slot assigned yet.")((void)0); | |||
451 | ||||
452 | do { | |||
453 | LiveInterval *LI; | |||
454 | std::tie(LI, VNI) = WorkList.pop_back_val(); | |||
455 | Register Reg = LI->reg(); | |||
456 | LLVM_DEBUG(dbgs() << "Checking redundant spills for " << VNI->id << '@'do { } while (false) | |||
457 | << VNI->def << " in " << *LI << '\n')do { } while (false); | |||
458 | ||||
459 | // Regs to spill are taken care of. | |||
460 | if (isRegToSpill(Reg)) | |||
461 | continue; | |||
462 | ||||
463 | // Add all of VNI's live range to StackInt. | |||
464 | StackInt->MergeValueInAsValue(*LI, VNI, StackInt->getValNumInfo(0)); | |||
465 | LLVM_DEBUG(dbgs() << "Merged to stack int: " << *StackInt << '\n')do { } while (false); | |||
466 | ||||
467 | // Find all spills and copies of VNI. | |||
468 | for (MachineRegisterInfo::use_instr_nodbg_iterator | |||
469 | UI = MRI.use_instr_nodbg_begin(Reg), E = MRI.use_instr_nodbg_end(); | |||
470 | UI != E; ) { | |||
471 | MachineInstr &MI = *UI++; | |||
472 | if (!MI.isCopy() && !MI.mayStore()) | |||
473 | continue; | |||
474 | SlotIndex Idx = LIS.getInstructionIndex(MI); | |||
475 | if (LI->getVNInfoAt(Idx) != VNI) | |||
476 | continue; | |||
477 | ||||
478 | // Follow sibling copies down the dominator tree. | |||
479 | if (Register DstReg = isFullCopyOf(MI, Reg)) { | |||
480 | if (isSibling(DstReg)) { | |||
481 | LiveInterval &DstLI = LIS.getInterval(DstReg); | |||
482 | VNInfo *DstVNI = DstLI.getVNInfoAt(Idx.getRegSlot()); | |||
483 | assert(DstVNI && "Missing defined value")((void)0); | |||
484 | assert(DstVNI->def == Idx.getRegSlot() && "Wrong copy def slot")((void)0); | |||
485 | WorkList.push_back(std::make_pair(&DstLI, DstVNI)); | |||
486 | } | |||
487 | continue; | |||
488 | } | |||
489 | ||||
490 | // Erase spills. | |||
491 | int FI; | |||
492 | if (Reg == TII.isStoreToStackSlot(MI, FI) && FI == StackSlot) { | |||
493 | LLVM_DEBUG(dbgs() << "Redundant spill " << Idx << '\t' << MI)do { } while (false); | |||
494 | // eliminateDeadDefs won't normally remove stores, so switch opcode. | |||
495 | MI.setDesc(TII.get(TargetOpcode::KILL)); | |||
496 | DeadDefs.push_back(&MI); | |||
497 | ++NumSpillsRemoved; | |||
498 | if (HSpiller.rmFromMergeableSpills(MI, StackSlot)) | |||
499 | --NumSpills; | |||
500 | } | |||
501 | } | |||
502 | } while (!WorkList.empty()); | |||
503 | } | |||
504 | ||||
505 | //===----------------------------------------------------------------------===// | |||
506 | // Rematerialization | |||
507 | //===----------------------------------------------------------------------===// | |||
508 | ||||
509 | /// markValueUsed - Remember that VNI failed to rematerialize, so its defining | |||
510 | /// instruction cannot be eliminated. See through snippet copies | |||
511 | void InlineSpiller::markValueUsed(LiveInterval *LI, VNInfo *VNI) { | |||
512 | SmallVector<std::pair<LiveInterval*, VNInfo*>, 8> WorkList; | |||
513 | WorkList.push_back(std::make_pair(LI, VNI)); | |||
514 | do { | |||
515 | std::tie(LI, VNI) = WorkList.pop_back_val(); | |||
516 | if (!UsedValues.insert(VNI).second) | |||
517 | continue; | |||
518 | ||||
519 | if (VNI->isPHIDef()) { | |||
520 | MachineBasicBlock *MBB = LIS.getMBBFromIndex(VNI->def); | |||
521 | for (MachineBasicBlock *P : MBB->predecessors()) { | |||
522 | VNInfo *PVNI = LI->getVNInfoBefore(LIS.getMBBEndIdx(P)); | |||
523 | if (PVNI) | |||
524 | WorkList.push_back(std::make_pair(LI, PVNI)); | |||
525 | } | |||
526 | continue; | |||
527 | } | |||
528 | ||||
529 | // Follow snippet copies. | |||
530 | MachineInstr *MI = LIS.getInstructionFromIndex(VNI->def); | |||
531 | if (!SnippetCopies.count(MI)) | |||
532 | continue; | |||
533 | LiveInterval &SnipLI = LIS.getInterval(MI->getOperand(1).getReg()); | |||
534 | assert(isRegToSpill(SnipLI.reg()) && "Unexpected register in copy")((void)0); | |||
535 | VNInfo *SnipVNI = SnipLI.getVNInfoAt(VNI->def.getRegSlot(true)); | |||
536 | assert(SnipVNI && "Snippet undefined before copy")((void)0); | |||
537 | WorkList.push_back(std::make_pair(&SnipLI, SnipVNI)); | |||
538 | } while (!WorkList.empty()); | |||
539 | } | |||
540 | ||||
541 | bool InlineSpiller::canGuaranteeAssignmentAfterRemat(Register VReg, | |||
542 | MachineInstr &MI) { | |||
543 | if (!RestrictStatepointRemat) | |||
544 | return true; | |||
545 | // Here's a quick explanation of the problem we're trying to handle here: | |||
546 | // * There are some pseudo instructions with more vreg uses than there are | |||
547 | // physical registers on the machine. | |||
548 | // * This is normally handled by spilling the vreg, and folding the reload | |||
549 | // into the user instruction. (Thus decreasing the number of used vregs | |||
550 | // until the remainder can be assigned to physregs.) | |||
551 | // * However, since we may try to spill vregs in any order, we can end up | |||
552 | // trying to spill each operand to the instruction, and then rematting it | |||
553 | // instead. When that happens, the new live intervals (for the remats) are | |||
554 | // expected to be trivially assignable (i.e. RS_Done). However, since we | |||
555 | // may have more remats than physregs, we're guaranteed to fail to assign | |||
556 | // one. | |||
557 | // At the moment, we only handle this for STATEPOINTs since they're the only | |||
558 | // pseudo op where we've seen this. If we start seeing other instructions | |||
559 | // with the same problem, we need to revisit this. | |||
560 | if (MI.getOpcode() != TargetOpcode::STATEPOINT) | |||
561 | return true; | |||
562 | // For STATEPOINTs we allow re-materialization for fixed arguments only hoping | |||
563 | // that number of physical registers is enough to cover all fixed arguments. | |||
564 | // If it is not true we need to revisit it. | |||
565 | for (unsigned Idx = StatepointOpers(&MI).getVarIdx(), | |||
566 | EndIdx = MI.getNumOperands(); | |||
567 | Idx < EndIdx; ++Idx) { | |||
568 | MachineOperand &MO = MI.getOperand(Idx); | |||
569 | if (MO.isReg() && MO.getReg() == VReg) | |||
570 | return false; | |||
571 | } | |||
572 | return true; | |||
573 | } | |||
574 | ||||
575 | /// reMaterializeFor - Attempt to rematerialize before MI instead of reloading. | |||
576 | bool InlineSpiller::reMaterializeFor(LiveInterval &VirtReg, MachineInstr &MI) { | |||
577 | // Analyze instruction | |||
578 | SmallVector<std::pair<MachineInstr *, unsigned>, 8> Ops; | |||
579 | VirtRegInfo RI = AnalyzeVirtRegInBundle(MI, VirtReg.reg(), &Ops); | |||
580 | ||||
581 | if (!RI.Reads) | |||
582 | return false; | |||
583 | ||||
584 | SlotIndex UseIdx = LIS.getInstructionIndex(MI).getRegSlot(true); | |||
585 | VNInfo *ParentVNI = VirtReg.getVNInfoAt(UseIdx.getBaseIndex()); | |||
586 | ||||
587 | if (!ParentVNI) { | |||
588 | LLVM_DEBUG(dbgs() << "\tadding <undef> flags: ")do { } while (false); | |||
589 | for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) { | |||
590 | MachineOperand &MO = MI.getOperand(i); | |||
591 | if (MO.isReg() && MO.isUse() && MO.getReg() == VirtReg.reg()) | |||
592 | MO.setIsUndef(); | |||
593 | } | |||
594 | LLVM_DEBUG(dbgs() << UseIdx << '\t' << MI)do { } while (false); | |||
595 | return true; | |||
596 | } | |||
597 | ||||
598 | if (SnippetCopies.count(&MI)) | |||
599 | return false; | |||
600 | ||||
601 | LiveInterval &OrigLI = LIS.getInterval(Original); | |||
602 | VNInfo *OrigVNI = OrigLI.getVNInfoAt(UseIdx); | |||
603 | LiveRangeEdit::Remat RM(ParentVNI); | |||
604 | RM.OrigMI = LIS.getInstructionFromIndex(OrigVNI->def); | |||
605 | ||||
606 | if (!Edit->canRematerializeAt(RM, OrigVNI, UseIdx, false)) { | |||
607 | markValueUsed(&VirtReg, ParentVNI); | |||
608 | LLVM_DEBUG(dbgs() << "\tcannot remat for " << UseIdx << '\t' << MI)do { } while (false); | |||
609 | return false; | |||
610 | } | |||
611 | ||||
612 | // If the instruction also writes VirtReg.reg, it had better not require the | |||
613 | // same register for uses and defs. | |||
614 | if (RI.Tied) { | |||
615 | markValueUsed(&VirtReg, ParentVNI); | |||
616 | LLVM_DEBUG(dbgs() << "\tcannot remat tied reg: " << UseIdx << '\t' << MI)do { } while (false); | |||
617 | return false; | |||
618 | } | |||
619 | ||||
620 | // Before rematerializing into a register for a single instruction, try to | |||
621 | // fold a load into the instruction. That avoids allocating a new register. | |||
622 | if (RM.OrigMI->canFoldAsLoad() && | |||
623 | foldMemoryOperand(Ops, RM.OrigMI)) { | |||
624 | Edit->markRematerialized(RM.ParentVNI); | |||
625 | ++NumFoldedLoads; | |||
626 | return true; | |||
627 | } | |||
628 | ||||
629 | // If we can't guarantee that we'll be able to actually assign the new vreg, | |||
630 | // we can't remat. | |||
631 | if (!canGuaranteeAssignmentAfterRemat(VirtReg.reg(), MI)) { | |||
632 | markValueUsed(&VirtReg, ParentVNI); | |||
633 | LLVM_DEBUG(dbgs() << "\tcannot remat for " << UseIdx << '\t' << MI)do { } while (false); | |||
634 | return false; | |||
635 | } | |||
636 | ||||
637 | // Allocate a new register for the remat. | |||
638 | Register NewVReg = Edit->createFrom(Original); | |||
639 | ||||
640 | // Finally we can rematerialize OrigMI before MI. | |||
641 | SlotIndex DefIdx = | |||
642 | Edit->rematerializeAt(*MI.getParent(), MI, NewVReg, RM, TRI); | |||
643 | ||||
644 | // We take the DebugLoc from MI, since OrigMI may be attributed to a | |||
645 | // different source location. | |||
646 | auto *NewMI = LIS.getInstructionFromIndex(DefIdx); | |||
647 | NewMI->setDebugLoc(MI.getDebugLoc()); | |||
648 | ||||
649 | (void)DefIdx; | |||
650 | LLVM_DEBUG(dbgs() << "\tremat: " << DefIdx << '\t'do { } while (false) | |||
651 | << *LIS.getInstructionFromIndex(DefIdx))do { } while (false); | |||
652 | ||||
653 | // Replace operands | |||
654 | for (const auto &OpPair : Ops) { | |||
655 | MachineOperand &MO = OpPair.first->getOperand(OpPair.second); | |||
656 | if (MO.isReg() && MO.isUse() && MO.getReg() == VirtReg.reg()) { | |||
657 | MO.setReg(NewVReg); | |||
658 | MO.setIsKill(); | |||
659 | } | |||
660 | } | |||
661 | LLVM_DEBUG(dbgs() << "\t " << UseIdx << '\t' << MI << '\n')do { } while (false); | |||
662 | ||||
663 | ++NumRemats; | |||
664 | return true; | |||
665 | } | |||
666 | ||||
667 | /// reMaterializeAll - Try to rematerialize as many uses as possible, | |||
668 | /// and trim the live ranges after. | |||
669 | void InlineSpiller::reMaterializeAll() { | |||
670 | if (!Edit->anyRematerializable(AA)) | |||
671 | return; | |||
672 | ||||
673 | UsedValues.clear(); | |||
674 | ||||
675 | // Try to remat before all uses of snippets. | |||
676 | bool anyRemat = false; | |||
677 | for (Register Reg : RegsToSpill) { | |||
678 | LiveInterval &LI = LIS.getInterval(Reg); | |||
679 | for (MachineRegisterInfo::reg_bundle_iterator | |||
680 | RegI = MRI.reg_bundle_begin(Reg), E = MRI.reg_bundle_end(); | |||
681 | RegI != E; ) { | |||
682 | MachineInstr &MI = *RegI++; | |||
683 | ||||
684 | // Debug values are not allowed to affect codegen. | |||
685 | if (MI.isDebugValue()) | |||
686 | continue; | |||
687 | ||||
688 | assert(!MI.isDebugInstr() && "Did not expect to find a use in debug "((void)0) | |||
689 | "instruction that isn't a DBG_VALUE")((void)0); | |||
690 | ||||
691 | anyRemat |= reMaterializeFor(LI, MI); | |||
692 | } | |||
693 | } | |||
694 | if (!anyRemat) | |||
695 | return; | |||
696 | ||||
697 | // Remove any values that were completely rematted. | |||
698 | for (Register Reg : RegsToSpill) { | |||
699 | LiveInterval &LI = LIS.getInterval(Reg); | |||
700 | for (LiveInterval::vni_iterator I = LI.vni_begin(), E = LI.vni_end(); | |||
701 | I != E; ++I) { | |||
702 | VNInfo *VNI = *I; | |||
703 | if (VNI->isUnused() || VNI->isPHIDef() || UsedValues.count(VNI)) | |||
704 | continue; | |||
705 | MachineInstr *MI = LIS.getInstructionFromIndex(VNI->def); | |||
706 | MI->addRegisterDead(Reg, &TRI); | |||
707 | if (!MI->allDefsAreDead()) | |||
708 | continue; | |||
709 | LLVM_DEBUG(dbgs() << "All defs dead: " << *MI)do { } while (false); | |||
710 | DeadDefs.push_back(MI); | |||
711 | } | |||
712 | } | |||
713 | ||||
714 | // Eliminate dead code after remat. Note that some snippet copies may be | |||
715 | // deleted here. | |||
716 | if (DeadDefs.empty()) | |||
717 | return; | |||
718 | LLVM_DEBUG(dbgs() << "Remat created " << DeadDefs.size() << " dead defs.\n")do { } while (false); | |||
719 | Edit->eliminateDeadDefs(DeadDefs, RegsToSpill, AA); | |||
720 | ||||
721 | // LiveRangeEdit::eliminateDeadDef is used to remove dead define instructions | |||
722 | // after rematerialization. To remove a VNI for a vreg from its LiveInterval, | |||
723 | // LiveIntervals::removeVRegDefAt is used. However, after non-PHI VNIs are all | |||
724 | // removed, PHI VNI are still left in the LiveInterval. | |||
725 | // So to get rid of unused reg, we need to check whether it has non-dbg | |||
726 | // reference instead of whether it has non-empty interval. | |||
727 | unsigned ResultPos = 0; | |||
728 | for (Register Reg : RegsToSpill) { | |||
729 | if (MRI.reg_nodbg_empty(Reg)) { | |||
730 | Edit->eraseVirtReg(Reg); | |||
731 | continue; | |||
732 | } | |||
733 | ||||
734 | assert(LIS.hasInterval(Reg) &&((void)0) | |||
735 | (!LIS.getInterval(Reg).empty() || !MRI.reg_nodbg_empty(Reg)) &&((void)0) | |||
736 | "Empty and not used live-range?!")((void)0); | |||
737 | ||||
738 | RegsToSpill[ResultPos++] = Reg; | |||
739 | } | |||
740 | RegsToSpill.erase(RegsToSpill.begin() + ResultPos, RegsToSpill.end()); | |||
741 | LLVM_DEBUG(dbgs() << RegsToSpill.size()do { } while (false) | |||
742 | << " registers to spill after remat.\n")do { } while (false); | |||
743 | } | |||
744 | ||||
745 | //===----------------------------------------------------------------------===// | |||
746 | // Spilling | |||
747 | //===----------------------------------------------------------------------===// | |||
748 | ||||
749 | /// If MI is a load or store of StackSlot, it can be removed. | |||
750 | bool InlineSpiller::coalesceStackAccess(MachineInstr *MI, Register Reg) { | |||
751 | int FI = 0; | |||
752 | Register InstrReg = TII.isLoadFromStackSlot(*MI, FI); | |||
753 | bool IsLoad = InstrReg; | |||
754 | if (!IsLoad) | |||
755 | InstrReg = TII.isStoreToStackSlot(*MI, FI); | |||
756 | ||||
757 | // We have a stack access. Is it the right register and slot? | |||
758 | if (InstrReg != Reg || FI != StackSlot) | |||
759 | return false; | |||
760 | ||||
761 | if (!IsLoad) | |||
762 | HSpiller.rmFromMergeableSpills(*MI, StackSlot); | |||
763 | ||||
764 | LLVM_DEBUG(dbgs() << "Coalescing stack access: " << *MI)do { } while (false); | |||
765 | LIS.RemoveMachineInstrFromMaps(*MI); | |||
766 | MI->eraseFromParent(); | |||
767 | ||||
768 | if (IsLoad) { | |||
769 | ++NumReloadsRemoved; | |||
770 | --NumReloads; | |||
771 | } else { | |||
772 | ++NumSpillsRemoved; | |||
773 | --NumSpills; | |||
774 | } | |||
775 | ||||
776 | return true; | |||
777 | } | |||
778 | ||||
779 | #if !defined(NDEBUG1) || defined(LLVM_ENABLE_DUMP) | |||
780 | LLVM_DUMP_METHOD__attribute__((noinline)) | |||
781 | // Dump the range of instructions from B to E with their slot indexes. | |||
782 | static void dumpMachineInstrRangeWithSlotIndex(MachineBasicBlock::iterator B, | |||
783 | MachineBasicBlock::iterator E, | |||
784 | LiveIntervals const &LIS, | |||
785 | const char *const header, | |||
786 | Register VReg = Register()) { | |||
787 | char NextLine = '\n'; | |||
788 | char SlotIndent = '\t'; | |||
789 | ||||
790 | if (std::next(B) == E) { | |||
791 | NextLine = ' '; | |||
792 | SlotIndent = ' '; | |||
793 | } | |||
794 | ||||
795 | dbgs() << '\t' << header << ": " << NextLine; | |||
796 | ||||
797 | for (MachineBasicBlock::iterator I = B; I != E; ++I) { | |||
798 | SlotIndex Idx = LIS.getInstructionIndex(*I).getRegSlot(); | |||
799 | ||||
800 | // If a register was passed in and this instruction has it as a | |||
801 | // destination that is marked as an early clobber, print the | |||
802 | // early-clobber slot index. | |||
803 | if (VReg) { | |||
804 | MachineOperand *MO = I->findRegisterDefOperand(VReg); | |||
805 | if (MO && MO->isEarlyClobber()) | |||
806 | Idx = Idx.getRegSlot(true); | |||
807 | } | |||
808 | ||||
809 | dbgs() << SlotIndent << Idx << '\t' << *I; | |||
810 | } | |||
811 | } | |||
812 | #endif | |||
813 | ||||
814 | /// foldMemoryOperand - Try folding stack slot references in Ops into their | |||
815 | /// instructions. | |||
816 | /// | |||
817 | /// @param Ops Operand indices from AnalyzeVirtRegInBundle(). | |||
818 | /// @param LoadMI Load instruction to use instead of stack slot when non-null. | |||
819 | /// @return True on success. | |||
820 | bool InlineSpiller:: | |||
821 | foldMemoryOperand(ArrayRef<std::pair<MachineInstr *, unsigned>> Ops, | |||
822 | MachineInstr *LoadMI) { | |||
823 | if (Ops.empty()) | |||
824 | return false; | |||
825 | // Don't attempt folding in bundles. | |||
826 | MachineInstr *MI = Ops.front().first; | |||
827 | if (Ops.back().first != MI || MI->isBundled()) | |||
828 | return false; | |||
829 | ||||
830 | bool WasCopy = MI->isCopy(); | |||
831 | Register ImpReg; | |||
832 | ||||
833 | // TII::foldMemoryOperand will do what we need here for statepoint | |||
834 | // (fold load into use and remove corresponding def). We will replace | |||
835 | // uses of removed def with loads (spillAroundUses). | |||
836 | // For that to work we need to untie def and use to pass it through | |||
837 | // foldMemoryOperand and signal foldPatchpoint that it is allowed to | |||
838 | // fold them. | |||
839 | bool UntieRegs = MI->getOpcode() == TargetOpcode::STATEPOINT; | |||
840 | ||||
841 | // Spill subregs if the target allows it. | |||
842 | // We always want to spill subregs for stackmap/patchpoint pseudos. | |||
843 | bool SpillSubRegs = TII.isSubregFoldable() || | |||
844 | MI->getOpcode() == TargetOpcode::STATEPOINT || | |||
845 | MI->getOpcode() == TargetOpcode::PATCHPOINT || | |||
846 | MI->getOpcode() == TargetOpcode::STACKMAP; | |||
847 | ||||
848 | // TargetInstrInfo::foldMemoryOperand only expects explicit, non-tied | |||
849 | // operands. | |||
850 | SmallVector<unsigned, 8> FoldOps; | |||
851 | for (const auto &OpPair : Ops) { | |||
852 | unsigned Idx = OpPair.second; | |||
853 | assert(MI == OpPair.first && "Instruction conflict during operand folding")((void)0); | |||
854 | MachineOperand &MO = MI->getOperand(Idx); | |||
855 | if (MO.isImplicit()) { | |||
856 | ImpReg = MO.getReg(); | |||
857 | continue; | |||
858 | } | |||
859 | ||||
860 | if (!SpillSubRegs && MO.getSubReg()) | |||
861 | return false; | |||
862 | // We cannot fold a load instruction into a def. | |||
863 | if (LoadMI && MO.isDef()) | |||
864 | return false; | |||
865 | // Tied use operands should not be passed to foldMemoryOperand. | |||
866 | if (UntieRegs || !MI->isRegTiedToDefOperand(Idx)) | |||
867 | FoldOps.push_back(Idx); | |||
868 | } | |||
869 | ||||
870 | // If we only have implicit uses, we won't be able to fold that. | |||
871 | // Moreover, TargetInstrInfo::foldMemoryOperand will assert if we try! | |||
872 | if (FoldOps.empty()) | |||
873 | return false; | |||
874 | ||||
875 | MachineInstrSpan MIS(MI, MI->getParent()); | |||
876 | ||||
877 | SmallVector<std::pair<unsigned, unsigned> > TiedOps; | |||
878 | if (UntieRegs) | |||
879 | for (unsigned Idx : FoldOps) { | |||
880 | MachineOperand &MO = MI->getOperand(Idx); | |||
881 | if (!MO.isTied()) | |||
882 | continue; | |||
883 | unsigned Tied = MI->findTiedOperandIdx(Idx); | |||
884 | if (MO.isUse()) | |||
885 | TiedOps.emplace_back(Tied, Idx); | |||
886 | else { | |||
887 | assert(MO.isDef() && "Tied to not use and def?")((void)0); | |||
888 | TiedOps.emplace_back(Idx, Tied); | |||
889 | } | |||
890 | MI->untieRegOperand(Idx); | |||
891 | } | |||
892 | ||||
893 | MachineInstr *FoldMI = | |||
894 | LoadMI ? TII.foldMemoryOperand(*MI, FoldOps, *LoadMI, &LIS) | |||
895 | : TII.foldMemoryOperand(*MI, FoldOps, StackSlot, &LIS, &VRM); | |||
896 | if (!FoldMI) { | |||
897 | // Re-tie operands. | |||
898 | for (auto Tied : TiedOps) | |||
899 | MI->tieOperands(Tied.first, Tied.second); | |||
900 | return false; | |||
901 | } | |||
902 | ||||
903 | // Remove LIS for any dead defs in the original MI not in FoldMI. | |||
904 | for (MIBundleOperands MO(*MI); MO.isValid(); ++MO) { | |||
905 | if (!MO->isReg()) | |||
906 | continue; | |||
907 | Register Reg = MO->getReg(); | |||
908 | if (!Reg || Register::isVirtualRegister(Reg) || MRI.isReserved(Reg)) { | |||
909 | continue; | |||
910 | } | |||
911 | // Skip non-Defs, including undef uses and internal reads. | |||
912 | if (MO->isUse()) | |||
913 | continue; | |||
914 | PhysRegInfo RI = AnalyzePhysRegInBundle(*FoldMI, Reg, &TRI); | |||
915 | if (RI.FullyDefined) | |||
916 | continue; | |||
917 | // FoldMI does not define this physreg. Remove the LI segment. | |||
918 | assert(MO->isDead() && "Cannot fold physreg def")((void)0); | |||
919 | SlotIndex Idx = LIS.getInstructionIndex(*MI).getRegSlot(); | |||
920 | LIS.removePhysRegDefAt(Reg.asMCReg(), Idx); | |||
921 | } | |||
922 | ||||
923 | int FI; | |||
924 | if (TII.isStoreToStackSlot(*MI, FI) && | |||
925 | HSpiller.rmFromMergeableSpills(*MI, FI)) | |||
926 | --NumSpills; | |||
927 | LIS.ReplaceMachineInstrInMaps(*MI, *FoldMI); | |||
928 | // Update the call site info. | |||
929 | if (MI->isCandidateForCallSiteEntry()) | |||
930 | MI->getMF()->moveCallSiteInfo(MI, FoldMI); | |||
931 | MI->eraseFromParent(); | |||
932 | ||||
933 | // Insert any new instructions other than FoldMI into the LIS maps. | |||
934 | assert(!MIS.empty() && "Unexpected empty span of instructions!")((void)0); | |||
935 | for (MachineInstr &MI : MIS) | |||
936 | if (&MI != FoldMI) | |||
937 | LIS.InsertMachineInstrInMaps(MI); | |||
938 | ||||
939 | // TII.foldMemoryOperand may have left some implicit operands on the | |||
940 | // instruction. Strip them. | |||
941 | if (ImpReg) | |||
942 | for (unsigned i = FoldMI->getNumOperands(); i; --i) { | |||
943 | MachineOperand &MO = FoldMI->getOperand(i - 1); | |||
944 | if (!MO.isReg() || !MO.isImplicit()) | |||
945 | break; | |||
946 | if (MO.getReg() == ImpReg) | |||
947 | FoldMI->RemoveOperand(i - 1); | |||
948 | } | |||
949 | ||||
950 | LLVM_DEBUG(dumpMachineInstrRangeWithSlotIndex(MIS.begin(), MIS.end(), LIS,do { } while (false) | |||
951 | "folded"))do { } while (false); | |||
952 | ||||
953 | if (!WasCopy) | |||
954 | ++NumFolded; | |||
955 | else if (Ops.front().second == 0) { | |||
956 | ++NumSpills; | |||
957 | // If there is only 1 store instruction is required for spill, add it | |||
958 | // to mergeable list. In X86 AMX, 2 intructions are required to store. | |||
959 | // We disable the merge for this case. | |||
960 | if (std::distance(MIS.begin(), MIS.end()) <= 1) | |||
961 | HSpiller.addToMergeableSpills(*FoldMI, StackSlot, Original); | |||
962 | } else | |||
963 | ++NumReloads; | |||
964 | return true; | |||
965 | } | |||
966 | ||||
967 | void InlineSpiller::insertReload(Register NewVReg, | |||
968 | SlotIndex Idx, | |||
969 | MachineBasicBlock::iterator MI) { | |||
970 | MachineBasicBlock &MBB = *MI->getParent(); | |||
971 | ||||
972 | MachineInstrSpan MIS(MI, &MBB); | |||
973 | TII.loadRegFromStackSlot(MBB, MI, NewVReg, StackSlot, | |||
974 | MRI.getRegClass(NewVReg), &TRI); | |||
975 | ||||
976 | LIS.InsertMachineInstrRangeInMaps(MIS.begin(), MI); | |||
977 | ||||
978 | LLVM_DEBUG(dumpMachineInstrRangeWithSlotIndex(MIS.begin(), MI, LIS, "reload",do { } while (false) | |||
979 | NewVReg))do { } while (false); | |||
980 | ++NumReloads; | |||
981 | } | |||
982 | ||||
983 | /// Check if \p Def fully defines a VReg with an undefined value. | |||
984 | /// If that's the case, that means the value of VReg is actually | |||
985 | /// not relevant. | |||
986 | static bool isRealSpill(const MachineInstr &Def) { | |||
987 | if (!Def.isImplicitDef()) | |||
988 | return true; | |||
989 | assert(Def.getNumOperands() == 1 &&((void)0) | |||
990 | "Implicit def with more than one definition")((void)0); | |||
991 | // We can say that the VReg defined by Def is undef, only if it is | |||
992 | // fully defined by Def. Otherwise, some of the lanes may not be | |||
993 | // undef and the value of the VReg matters. | |||
994 | return Def.getOperand(0).getSubReg(); | |||
995 | } | |||
996 | ||||
997 | /// insertSpill - Insert a spill of NewVReg after MI. | |||
998 | void InlineSpiller::insertSpill(Register NewVReg, bool isKill, | |||
999 | MachineBasicBlock::iterator MI) { | |||
1000 | // Spill are not terminators, so inserting spills after terminators will | |||
1001 | // violate invariants in MachineVerifier. | |||
1002 | assert(!MI->isTerminator() && "Inserting a spill after a terminator")((void)0); | |||
1003 | MachineBasicBlock &MBB = *MI->getParent(); | |||
1004 | ||||
1005 | MachineInstrSpan MIS(MI, &MBB); | |||
1006 | MachineBasicBlock::iterator SpillBefore = std::next(MI); | |||
1007 | bool IsRealSpill = isRealSpill(*MI); | |||
1008 | ||||
1009 | if (IsRealSpill) | |||
1010 | TII.storeRegToStackSlot(MBB, SpillBefore, NewVReg, isKill, StackSlot, | |||
1011 | MRI.getRegClass(NewVReg), &TRI); | |||
1012 | else | |||
1013 | // Don't spill undef value. | |||
1014 | // Anything works for undef, in particular keeping the memory | |||
1015 | // uninitialized is a viable option and it saves code size and | |||
1016 | // run time. | |||
1017 | BuildMI(MBB, SpillBefore, MI->getDebugLoc(), TII.get(TargetOpcode::KILL)) | |||
1018 | .addReg(NewVReg, getKillRegState(isKill)); | |||
1019 | ||||
1020 | MachineBasicBlock::iterator Spill = std::next(MI); | |||
1021 | LIS.InsertMachineInstrRangeInMaps(Spill, MIS.end()); | |||
1022 | for (const MachineInstr &MI : make_range(Spill, MIS.end())) | |||
1023 | getVDefInterval(MI, LIS); | |||
1024 | ||||
1025 | LLVM_DEBUG(do { } while (false) | |||
1026 | dumpMachineInstrRangeWithSlotIndex(Spill, MIS.end(), LIS, "spill"))do { } while (false); | |||
1027 | ++NumSpills; | |||
1028 | // If there is only 1 store instruction is required for spill, add it | |||
1029 | // to mergeable list. In X86 AMX, 2 intructions are required to store. | |||
1030 | // We disable the merge for this case. | |||
1031 | if (IsRealSpill && std::distance(Spill, MIS.end()) <= 1) | |||
1032 | HSpiller.addToMergeableSpills(*Spill, StackSlot, Original); | |||
1033 | } | |||
1034 | ||||
1035 | /// spillAroundUses - insert spill code around each use of Reg. | |||
1036 | void InlineSpiller::spillAroundUses(Register Reg) { | |||
1037 | LLVM_DEBUG(dbgs() << "spillAroundUses " << printReg(Reg) << '\n')do { } while (false); | |||
1038 | LiveInterval &OldLI = LIS.getInterval(Reg); | |||
1039 | ||||
1040 | // Iterate over instructions using Reg. | |||
1041 | for (MachineRegisterInfo::reg_bundle_iterator | |||
1042 | RegI = MRI.reg_bundle_begin(Reg), E = MRI.reg_bundle_end(); | |||
1043 | RegI != E; ) { | |||
1044 | MachineInstr *MI = &*(RegI++); | |||
1045 | ||||
1046 | // Debug values are not allowed to affect codegen. | |||
1047 | if (MI->isDebugValue()) { | |||
1048 | // Modify DBG_VALUE now that the value is in a spill slot. | |||
1049 | MachineBasicBlock *MBB = MI->getParent(); | |||
1050 | LLVM_DEBUG(dbgs() << "Modifying debug info due to spill:\t" << *MI)do { } while (false); | |||
1051 | buildDbgValueForSpill(*MBB, MI, *MI, StackSlot, Reg); | |||
1052 | MBB->erase(MI); | |||
1053 | continue; | |||
1054 | } | |||
1055 | ||||
1056 | assert(!MI->isDebugInstr() && "Did not expect to find a use in debug "((void)0) | |||
1057 | "instruction that isn't a DBG_VALUE")((void)0); | |||
1058 | ||||
1059 | // Ignore copies to/from snippets. We'll delete them. | |||
1060 | if (SnippetCopies.count(MI)) | |||
1061 | continue; | |||
1062 | ||||
1063 | // Stack slot accesses may coalesce away. | |||
1064 | if (coalesceStackAccess(MI, Reg)) | |||
1065 | continue; | |||
1066 | ||||
1067 | // Analyze instruction. | |||
1068 | SmallVector<std::pair<MachineInstr*, unsigned>, 8> Ops; | |||
1069 | VirtRegInfo RI = AnalyzeVirtRegInBundle(*MI, Reg, &Ops); | |||
1070 | ||||
1071 | // Find the slot index where this instruction reads and writes OldLI. | |||
1072 | // This is usually the def slot, except for tied early clobbers. | |||
1073 | SlotIndex Idx = LIS.getInstructionIndex(*MI).getRegSlot(); | |||
1074 | if (VNInfo *VNI = OldLI.getVNInfoAt(Idx.getRegSlot(true))) | |||
1075 | if (SlotIndex::isSameInstr(Idx, VNI->def)) | |||
1076 | Idx = VNI->def; | |||
1077 | ||||
1078 | // Check for a sibling copy. | |||
1079 | Register SibReg = isFullCopyOf(*MI, Reg); | |||
1080 | if (SibReg && isSibling(SibReg)) { | |||
1081 | // This may actually be a copy between snippets. | |||
1082 | if (isRegToSpill(SibReg)) { | |||
1083 | LLVM_DEBUG(dbgs() << "Found new snippet copy: " << *MI)do { } while (false); | |||
1084 | SnippetCopies.insert(MI); | |||
1085 | continue; | |||
1086 | } | |||
1087 | if (RI.Writes) { | |||
1088 | if (hoistSpillInsideBB(OldLI, *MI)) { | |||
1089 | // This COPY is now dead, the value is already in the stack slot. | |||
1090 | MI->getOperand(0).setIsDead(); | |||
1091 | DeadDefs.push_back(MI); | |||
1092 | continue; | |||
1093 | } | |||
1094 | } else { | |||
1095 | // This is a reload for a sib-reg copy. Drop spills downstream. | |||
1096 | LiveInterval &SibLI = LIS.getInterval(SibReg); | |||
1097 | eliminateRedundantSpills(SibLI, SibLI.getVNInfoAt(Idx)); | |||
1098 | // The COPY will fold to a reload below. | |||
1099 | } | |||
1100 | } | |||
1101 | ||||
1102 | // Attempt to fold memory ops. | |||
1103 | if (foldMemoryOperand(Ops)) | |||
1104 | continue; | |||
1105 | ||||
1106 | // Create a new virtual register for spill/fill. | |||
1107 | // FIXME: Infer regclass from instruction alone. | |||
1108 | Register NewVReg = Edit->createFrom(Reg); | |||
1109 | ||||
1110 | if (RI.Reads) | |||
1111 | insertReload(NewVReg, Idx, MI); | |||
1112 | ||||
1113 | // Rewrite instruction operands. | |||
1114 | bool hasLiveDef = false; | |||
1115 | for (const auto &OpPair : Ops) { | |||
1116 | MachineOperand &MO = OpPair.first->getOperand(OpPair.second); | |||
1117 | MO.setReg(NewVReg); | |||
1118 | if (MO.isUse()) { | |||
1119 | if (!OpPair.first->isRegTiedToDefOperand(OpPair.second)) | |||
1120 | MO.setIsKill(); | |||
1121 | } else { | |||
1122 | if (!MO.isDead()) | |||
1123 | hasLiveDef = true; | |||
1124 | } | |||
1125 | } | |||
1126 | LLVM_DEBUG(dbgs() << "\trewrite: " << Idx << '\t' << *MI << '\n')do { } while (false); | |||
1127 | ||||
1128 | // FIXME: Use a second vreg if instruction has no tied ops. | |||
1129 | if (RI.Writes) | |||
1130 | if (hasLiveDef) | |||
1131 | insertSpill(NewVReg, true, MI); | |||
1132 | } | |||
1133 | } | |||
1134 | ||||
1135 | /// spillAll - Spill all registers remaining after rematerialization. | |||
1136 | void InlineSpiller::spillAll() { | |||
1137 | // Update LiveStacks now that we are committed to spilling. | |||
1138 | if (StackSlot == VirtRegMap::NO_STACK_SLOT) { | |||
1139 | StackSlot = VRM.assignVirt2StackSlot(Original); | |||
1140 | StackInt = &LSS.getOrCreateInterval(StackSlot, MRI.getRegClass(Original)); | |||
1141 | StackInt->getNextValue(SlotIndex(), LSS.getVNInfoAllocator()); | |||
1142 | } else | |||
1143 | StackInt = &LSS.getInterval(StackSlot); | |||
1144 | ||||
1145 | if (Original != Edit->getReg()) | |||
1146 | VRM.assignVirt2StackSlot(Edit->getReg(), StackSlot); | |||
1147 | ||||
1148 | assert(StackInt->getNumValNums() == 1 && "Bad stack interval values")((void)0); | |||
1149 | for (Register Reg : RegsToSpill) | |||
1150 | StackInt->MergeSegmentsInAsValue(LIS.getInterval(Reg), | |||
1151 | StackInt->getValNumInfo(0)); | |||
1152 | LLVM_DEBUG(dbgs() << "Merged spilled regs: " << *StackInt << '\n')do { } while (false); | |||
1153 | ||||
1154 | // Spill around uses of all RegsToSpill. | |||
1155 | for (Register Reg : RegsToSpill) | |||
1156 | spillAroundUses(Reg); | |||
1157 | ||||
1158 | // Hoisted spills may cause dead code. | |||
1159 | if (!DeadDefs.empty()) { | |||
1160 | LLVM_DEBUG(dbgs() << "Eliminating " << DeadDefs.size() << " dead defs\n")do { } while (false); | |||
1161 | Edit->eliminateDeadDefs(DeadDefs, RegsToSpill, AA); | |||
1162 | } | |||
1163 | ||||
1164 | // Finally delete the SnippetCopies. | |||
1165 | for (Register Reg : RegsToSpill) { | |||
1166 | for (MachineRegisterInfo::reg_instr_iterator | |||
1167 | RI = MRI.reg_instr_begin(Reg), E = MRI.reg_instr_end(); | |||
1168 | RI != E; ) { | |||
1169 | MachineInstr &MI = *(RI++); | |||
1170 | assert(SnippetCopies.count(&MI) && "Remaining use wasn't a snippet copy")((void)0); | |||
1171 | // FIXME: Do this with a LiveRangeEdit callback. | |||
1172 | LIS.RemoveMachineInstrFromMaps(MI); | |||
1173 | MI.eraseFromParent(); | |||
1174 | } | |||
1175 | } | |||
1176 | ||||
1177 | // Delete all spilled registers. | |||
1178 | for (Register Reg : RegsToSpill) | |||
1179 | Edit->eraseVirtReg(Reg); | |||
1180 | } | |||
1181 | ||||
1182 | void InlineSpiller::spill(LiveRangeEdit &edit) { | |||
1183 | ++NumSpilledRanges; | |||
1184 | Edit = &edit; | |||
1185 | assert(!Register::isStackSlot(edit.getReg()) &&((void)0) | |||
1186 | "Trying to spill a stack slot.")((void)0); | |||
1187 | // Share a stack slot among all descendants of Original. | |||
1188 | Original = VRM.getOriginal(edit.getReg()); | |||
1189 | StackSlot = VRM.getStackSlot(Original); | |||
1190 | StackInt = nullptr; | |||
1191 | ||||
1192 | LLVM_DEBUG(dbgs() << "Inline spilling "do { } while (false) | |||
1193 | << TRI.getRegClassName(MRI.getRegClass(edit.getReg()))do { } while (false) | |||
1194 | << ':' << edit.getParent() << "\nFrom original "do { } while (false) | |||
1195 | << printReg(Original) << '\n')do { } while (false); | |||
1196 | assert(edit.getParent().isSpillable() &&((void)0) | |||
1197 | "Attempting to spill already spilled value.")((void)0); | |||
1198 | assert(DeadDefs.empty() && "Previous spill didn't remove dead defs")((void)0); | |||
1199 | ||||
1200 | collectRegsToSpill(); | |||
1201 | reMaterializeAll(); | |||
1202 | ||||
1203 | // Remat may handle everything. | |||
1204 | if (!RegsToSpill.empty()) | |||
1205 | spillAll(); | |||
1206 | ||||
1207 | Edit->calculateRegClassAndHint(MF, VRAI); | |||
1208 | } | |||
1209 | ||||
1210 | /// Optimizations after all the reg selections and spills are done. | |||
1211 | void InlineSpiller::postOptimization() { HSpiller.hoistAllSpills(); } | |||
| ||||
1212 | ||||
1213 | /// When a spill is inserted, add the spill to MergeableSpills map. | |||
1214 | void HoistSpillHelper::addToMergeableSpills(MachineInstr &Spill, int StackSlot, | |||
1215 | unsigned Original) { | |||
1216 | BumpPtrAllocator &Allocator = LIS.getVNInfoAllocator(); | |||
1217 | LiveInterval &OrigLI = LIS.getInterval(Original); | |||
1218 | // save a copy of LiveInterval in StackSlotToOrigLI because the original | |||
1219 | // LiveInterval may be cleared after all its references are spilled. | |||
1220 | if (StackSlotToOrigLI.find(StackSlot) == StackSlotToOrigLI.end()) { | |||
1221 | auto LI = std::make_unique<LiveInterval>(OrigLI.reg(), OrigLI.weight()); | |||
1222 | LI->assign(OrigLI, Allocator); | |||
1223 | StackSlotToOrigLI[StackSlot] = std::move(LI); | |||
1224 | } | |||
1225 | SlotIndex Idx = LIS.getInstructionIndex(Spill); | |||
1226 | VNInfo *OrigVNI = StackSlotToOrigLI[StackSlot]->getVNInfoAt(Idx.getRegSlot()); | |||
1227 | std::pair<int, VNInfo *> MIdx = std::make_pair(StackSlot, OrigVNI); | |||
1228 | MergeableSpills[MIdx].insert(&Spill); | |||
1229 | } | |||
1230 | ||||
1231 | /// When a spill is removed, remove the spill from MergeableSpills map. | |||
1232 | /// Return true if the spill is removed successfully. | |||
1233 | bool HoistSpillHelper::rmFromMergeableSpills(MachineInstr &Spill, | |||
1234 | int StackSlot) { | |||
1235 | auto It = StackSlotToOrigLI.find(StackSlot); | |||
1236 | if (It == StackSlotToOrigLI.end()) | |||
1237 | return false; | |||
1238 | SlotIndex Idx = LIS.getInstructionIndex(Spill); | |||
1239 | VNInfo *OrigVNI = It->second->getVNInfoAt(Idx.getRegSlot()); | |||
1240 | std::pair<int, VNInfo *> MIdx = std::make_pair(StackSlot, OrigVNI); | |||
1241 | return MergeableSpills[MIdx].erase(&Spill); | |||
1242 | } | |||
1243 | ||||
1244 | /// Check BB to see if it is a possible target BB to place a hoisted spill, | |||
1245 | /// i.e., there should be a living sibling of OrigReg at the insert point. | |||
1246 | bool HoistSpillHelper::isSpillCandBB(LiveInterval &OrigLI, VNInfo &OrigVNI, | |||
1247 | MachineBasicBlock &BB, Register &LiveReg) { | |||
1248 | SlotIndex Idx = IPA.getLastInsertPoint(OrigLI, BB); | |||
1249 | // The original def could be after the last insert point in the root block, | |||
1250 | // we can't hoist to here. | |||
1251 | if (Idx < OrigVNI.def) { | |||
1252 | // TODO: We could be better here. If LI is not alive in landing pad | |||
1253 | // we could hoist spill after LIP. | |||
1254 | LLVM_DEBUG(dbgs() << "can't spill in root block - def after LIP\n")do { } while (false); | |||
1255 | return false; | |||
1256 | } | |||
1257 | Register OrigReg = OrigLI.reg(); | |||
1258 | SmallSetVector<Register, 16> &Siblings = Virt2SiblingsMap[OrigReg]; | |||
1259 | assert(OrigLI.getVNInfoAt(Idx) == &OrigVNI && "Unexpected VNI")((void)0); | |||
1260 | ||||
1261 | for (const Register &SibReg : Siblings) { | |||
1262 | LiveInterval &LI = LIS.getInterval(SibReg); | |||
1263 | VNInfo *VNI = LI.getVNInfoAt(Idx); | |||
1264 | if (VNI) { | |||
1265 | LiveReg = SibReg; | |||
1266 | return true; | |||
1267 | } | |||
1268 | } | |||
1269 | return false; | |||
1270 | } | |||
1271 | ||||
1272 | /// Remove redundant spills in the same BB. Save those redundant spills in | |||
1273 | /// SpillsToRm, and save the spill to keep and its BB in SpillBBToSpill map. | |||
1274 | void HoistSpillHelper::rmRedundantSpills( | |||
1275 | SmallPtrSet<MachineInstr *, 16> &Spills, | |||
1276 | SmallVectorImpl<MachineInstr *> &SpillsToRm, | |||
1277 | DenseMap<MachineDomTreeNode *, MachineInstr *> &SpillBBToSpill) { | |||
1278 | // For each spill saw, check SpillBBToSpill[] and see if its BB already has | |||
1279 | // another spill inside. If a BB contains more than one spill, only keep the | |||
1280 | // earlier spill with smaller SlotIndex. | |||
1281 | for (const auto CurrentSpill : Spills) { | |||
1282 | MachineBasicBlock *Block = CurrentSpill->getParent(); | |||
1283 | MachineDomTreeNode *Node = MDT.getBase().getNode(Block); | |||
1284 | MachineInstr *PrevSpill = SpillBBToSpill[Node]; | |||
1285 | if (PrevSpill) { | |||
1286 | SlotIndex PIdx = LIS.getInstructionIndex(*PrevSpill); | |||
1287 | SlotIndex CIdx = LIS.getInstructionIndex(*CurrentSpill); | |||
1288 | MachineInstr *SpillToRm = (CIdx > PIdx) ? CurrentSpill : PrevSpill; | |||
1289 | MachineInstr *SpillToKeep = (CIdx > PIdx) ? PrevSpill : CurrentSpill; | |||
1290 | SpillsToRm.push_back(SpillToRm); | |||
1291 | SpillBBToSpill[MDT.getBase().getNode(Block)] = SpillToKeep; | |||
1292 | } else { | |||
1293 | SpillBBToSpill[MDT.getBase().getNode(Block)] = CurrentSpill; | |||
1294 | } | |||
1295 | } | |||
1296 | for (const auto SpillToRm : SpillsToRm) | |||
1297 | Spills.erase(SpillToRm); | |||
1298 | } | |||
1299 | ||||
1300 | /// Starting from \p Root find a top-down traversal order of the dominator | |||
1301 | /// tree to visit all basic blocks containing the elements of \p Spills. | |||
1302 | /// Redundant spills will be found and put into \p SpillsToRm at the same | |||
1303 | /// time. \p SpillBBToSpill will be populated as part of the process and | |||
1304 | /// maps a basic block to the first store occurring in the basic block. | |||
1305 | /// \post SpillsToRm.union(Spills\@post) == Spills\@pre | |||
1306 | void HoistSpillHelper::getVisitOrders( | |||
1307 | MachineBasicBlock *Root, SmallPtrSet<MachineInstr *, 16> &Spills, | |||
1308 | SmallVectorImpl<MachineDomTreeNode *> &Orders, | |||
1309 | SmallVectorImpl<MachineInstr *> &SpillsToRm, | |||
1310 | DenseMap<MachineDomTreeNode *, unsigned> &SpillsToKeep, | |||
1311 | DenseMap<MachineDomTreeNode *, MachineInstr *> &SpillBBToSpill) { | |||
1312 | // The set contains all the possible BB nodes to which we may hoist | |||
1313 | // original spills. | |||
1314 | SmallPtrSet<MachineDomTreeNode *, 8> WorkSet; | |||
1315 | // Save the BB nodes on the path from the first BB node containing | |||
1316 | // non-redundant spill to the Root node. | |||
1317 | SmallPtrSet<MachineDomTreeNode *, 8> NodesOnPath; | |||
1318 | // All the spills to be hoisted must originate from a single def instruction | |||
1319 | // to the OrigReg. It means the def instruction should dominate all the spills | |||
1320 | // to be hoisted. We choose the BB where the def instruction is located as | |||
1321 | // the Root. | |||
1322 | MachineDomTreeNode *RootIDomNode = MDT[Root]->getIDom(); | |||
1323 | // For every node on the dominator tree with spill, walk up on the dominator | |||
1324 | // tree towards the Root node until it is reached. If there is other node | |||
1325 | // containing spill in the middle of the path, the previous spill saw will | |||
1326 | // be redundant and the node containing it will be removed. All the nodes on | |||
1327 | // the path starting from the first node with non-redundant spill to the Root | |||
1328 | // node will be added to the WorkSet, which will contain all the possible | |||
1329 | // locations where spills may be hoisted to after the loop below is done. | |||
1330 | for (const auto Spill : Spills) { | |||
1331 | MachineBasicBlock *Block = Spill->getParent(); | |||
1332 | MachineDomTreeNode *Node = MDT[Block]; | |||
1333 | MachineInstr *SpillToRm = nullptr; | |||
1334 | while (Node != RootIDomNode) { | |||
1335 | // If Node dominates Block, and it already contains a spill, the spill in | |||
1336 | // Block will be redundant. | |||
1337 | if (Node != MDT[Block] && SpillBBToSpill[Node]) { | |||
1338 | SpillToRm = SpillBBToSpill[MDT[Block]]; | |||
1339 | break; | |||
1340 | /// If we see the Node already in WorkSet, the path from the Node to | |||
1341 | /// the Root node must already be traversed by another spill. | |||
1342 | /// Then no need to repeat. | |||
1343 | } else if (WorkSet.count(Node)) { | |||
1344 | break; | |||
1345 | } else { | |||
1346 | NodesOnPath.insert(Node); | |||
1347 | } | |||
1348 | Node = Node->getIDom(); | |||
| ||||
1349 | } | |||
1350 | if (SpillToRm) { | |||
1351 | SpillsToRm.push_back(SpillToRm); | |||
1352 | } else { | |||
1353 | // Add a BB containing the original spills to SpillsToKeep -- i.e., | |||
1354 | // set the initial status before hoisting start. The value of BBs | |||
1355 | // containing original spills is set to 0, in order to descriminate | |||
1356 | // with BBs containing hoisted spills which will be inserted to | |||
1357 | // SpillsToKeep later during hoisting. | |||
1358 | SpillsToKeep[MDT[Block]] = 0; | |||
1359 | WorkSet.insert(NodesOnPath.begin(), NodesOnPath.end()); | |||
1360 | } | |||
1361 | NodesOnPath.clear(); | |||
1362 | } | |||
1363 | ||||
1364 | // Sort the nodes in WorkSet in top-down order and save the nodes | |||
1365 | // in Orders. Orders will be used for hoisting in runHoistSpills. | |||
1366 | unsigned idx = 0; | |||
1367 | Orders.push_back(MDT.getBase().getNode(Root)); | |||
1368 | do { | |||
1369 | MachineDomTreeNode *Node = Orders[idx++]; | |||
1370 | for (MachineDomTreeNode *Child : Node->children()) { | |||
1371 | if (WorkSet.count(Child)) | |||
1372 | Orders.push_back(Child); | |||
1373 | } | |||
1374 | } while (idx != Orders.size()); | |||
1375 | assert(Orders.size() == WorkSet.size() &&((void)0) | |||
1376 | "Orders have different size with WorkSet")((void)0); | |||
1377 | ||||
1378 | #ifndef NDEBUG1 | |||
1379 | LLVM_DEBUG(dbgs() << "Orders size is " << Orders.size() << "\n")do { } while (false); | |||
1380 | SmallVector<MachineDomTreeNode *, 32>::reverse_iterator RIt = Orders.rbegin(); | |||
1381 | for (; RIt != Orders.rend(); RIt++) | |||
1382 | LLVM_DEBUG(dbgs() << "BB" << (*RIt)->getBlock()->getNumber() << ",")do { } while (false); | |||
1383 | LLVM_DEBUG(dbgs() << "\n")do { } while (false); | |||
1384 | #endif | |||
1385 | } | |||
1386 | ||||
1387 | /// Try to hoist spills according to BB hotness. The spills to removed will | |||
1388 | /// be saved in \p SpillsToRm. The spills to be inserted will be saved in | |||
1389 | /// \p SpillsToIns. | |||
1390 | void HoistSpillHelper::runHoistSpills( | |||
1391 | LiveInterval &OrigLI, VNInfo &OrigVNI, | |||
1392 | SmallPtrSet<MachineInstr *, 16> &Spills, | |||
1393 | SmallVectorImpl<MachineInstr *> &SpillsToRm, | |||
1394 | DenseMap<MachineBasicBlock *, unsigned> &SpillsToIns) { | |||
1395 | // Visit order of dominator tree nodes. | |||
1396 | SmallVector<MachineDomTreeNode *, 32> Orders; | |||
1397 | // SpillsToKeep contains all the nodes where spills are to be inserted | |||
1398 | // during hoisting. If the spill to be inserted is an original spill | |||
1399 | // (not a hoisted one), the value of the map entry is 0. If the spill | |||
1400 | // is a hoisted spill, the value of the map entry is the VReg to be used | |||
1401 | // as the source of the spill. | |||
1402 | DenseMap<MachineDomTreeNode *, unsigned> SpillsToKeep; | |||
1403 | // Map from BB to the first spill inside of it. | |||
1404 | DenseMap<MachineDomTreeNode *, MachineInstr *> SpillBBToSpill; | |||
1405 | ||||
1406 | rmRedundantSpills(Spills, SpillsToRm, SpillBBToSpill); | |||
1407 | ||||
1408 | MachineBasicBlock *Root = LIS.getMBBFromIndex(OrigVNI.def); | |||
1409 | getVisitOrders(Root, Spills, Orders, SpillsToRm, SpillsToKeep, | |||
1410 | SpillBBToSpill); | |||
1411 | ||||
1412 | // SpillsInSubTreeMap keeps the map from a dom tree node to a pair of | |||
1413 | // nodes set and the cost of all the spills inside those nodes. | |||
1414 | // The nodes set are the locations where spills are to be inserted | |||
1415 | // in the subtree of current node. | |||
1416 | using NodesCostPair = | |||
1417 | std::pair<SmallPtrSet<MachineDomTreeNode *, 16>, BlockFrequency>; | |||
1418 | DenseMap<MachineDomTreeNode *, NodesCostPair> SpillsInSubTreeMap; | |||
1419 | ||||
1420 | // Iterate Orders set in reverse order, which will be a bottom-up order | |||
1421 | // in the dominator tree. Once we visit a dom tree node, we know its | |||
1422 | // children have already been visited and the spill locations in the | |||
1423 | // subtrees of all the children have been determined. | |||
1424 | SmallVector<MachineDomTreeNode *, 32>::reverse_iterator RIt = Orders.rbegin(); | |||
1425 | for (; RIt != Orders.rend(); RIt++) { | |||
1426 | MachineBasicBlock *Block = (*RIt)->getBlock(); | |||
1427 | ||||
1428 | // If Block contains an original spill, simply continue. | |||
1429 | if (SpillsToKeep.find(*RIt) != SpillsToKeep.end() && !SpillsToKeep[*RIt]) { | |||
1430 | SpillsInSubTreeMap[*RIt].first.insert(*RIt); | |||
1431 | // SpillsInSubTreeMap[*RIt].second contains the cost of spill. | |||
1432 | SpillsInSubTreeMap[*RIt].second = MBFI.getBlockFreq(Block); | |||
1433 | continue; | |||
1434 | } | |||
1435 | ||||
1436 | // Collect spills in subtree of current node (*RIt) to | |||
1437 | // SpillsInSubTreeMap[*RIt].first. | |||
1438 | for (MachineDomTreeNode *Child : (*RIt)->children()) { | |||
1439 | if (SpillsInSubTreeMap.find(Child) == SpillsInSubTreeMap.end()) | |||
1440 | continue; | |||
1441 | // The stmt "SpillsInSubTree = SpillsInSubTreeMap[*RIt].first" below | |||
1442 | // should be placed before getting the begin and end iterators of | |||
1443 | // SpillsInSubTreeMap[Child].first, or else the iterators may be | |||
1444 | // invalidated when SpillsInSubTreeMap[*RIt] is seen the first time | |||
1445 | // and the map grows and then the original buckets in the map are moved. | |||
1446 | SmallPtrSet<MachineDomTreeNode *, 16> &SpillsInSubTree = | |||
1447 | SpillsInSubTreeMap[*RIt].first; | |||
1448 | BlockFrequency &SubTreeCost = SpillsInSubTreeMap[*RIt].second; | |||
1449 | SubTreeCost += SpillsInSubTreeMap[Child].second; | |||
1450 | auto BI = SpillsInSubTreeMap[Child].first.begin(); | |||
1451 | auto EI = SpillsInSubTreeMap[Child].first.end(); | |||
1452 | SpillsInSubTree.insert(BI, EI); | |||
1453 | SpillsInSubTreeMap.erase(Child); | |||
1454 | } | |||
1455 | ||||
1456 | SmallPtrSet<MachineDomTreeNode *, 16> &SpillsInSubTree = | |||
1457 | SpillsInSubTreeMap[*RIt].first; | |||
1458 | BlockFrequency &SubTreeCost = SpillsInSubTreeMap[*RIt].second; | |||
1459 | // No spills in subtree, simply continue. | |||
1460 | if (SpillsInSubTree.empty()) | |||
1461 | continue; | |||
1462 | ||||
1463 | // Check whether Block is a possible candidate to insert spill. | |||
1464 | Register LiveReg; | |||
1465 | if (!isSpillCandBB(OrigLI, OrigVNI, *Block, LiveReg)) | |||
1466 | continue; | |||
1467 | ||||
1468 | // If there are multiple spills that could be merged, bias a little | |||
1469 | // to hoist the spill. | |||
1470 | BranchProbability MarginProb = (SpillsInSubTree.size() > 1) | |||
1471 | ? BranchProbability(9, 10) | |||
1472 | : BranchProbability(1, 1); | |||
1473 | if (SubTreeCost > MBFI.getBlockFreq(Block) * MarginProb) { | |||
1474 | // Hoist: Move spills to current Block. | |||
1475 | for (const auto SpillBB : SpillsInSubTree) { | |||
1476 | // When SpillBB is a BB contains original spill, insert the spill | |||
1477 | // to SpillsToRm. | |||
1478 | if (SpillsToKeep.find(SpillBB) != SpillsToKeep.end() && | |||
1479 | !SpillsToKeep[SpillBB]) { | |||
1480 | MachineInstr *SpillToRm = SpillBBToSpill[SpillBB]; | |||
1481 | SpillsToRm.push_back(SpillToRm); | |||
1482 | } | |||
1483 | // SpillBB will not contain spill anymore, remove it from SpillsToKeep. | |||
1484 | SpillsToKeep.erase(SpillBB); | |||
1485 | } | |||
1486 | // Current Block is the BB containing the new hoisted spill. Add it to | |||
1487 | // SpillsToKeep. LiveReg is the source of the new spill. | |||
1488 | SpillsToKeep[*RIt] = LiveReg; | |||
1489 | LLVM_DEBUG({do { } while (false) | |||
1490 | dbgs() << "spills in BB: ";do { } while (false) | |||
1491 | for (const auto Rspill : SpillsInSubTree)do { } while (false) | |||
1492 | dbgs() << Rspill->getBlock()->getNumber() << " ";do { } while (false) | |||
1493 | dbgs() << "were promoted to BB" << (*RIt)->getBlock()->getNumber()do { } while (false) | |||
1494 | << "\n";do { } while (false) | |||
1495 | })do { } while (false); | |||
1496 | SpillsInSubTree.clear(); | |||
1497 | SpillsInSubTree.insert(*RIt); | |||
1498 | SubTreeCost = MBFI.getBlockFreq(Block); | |||
1499 | } | |||
1500 | } | |||
1501 | // For spills in SpillsToKeep with LiveReg set (i.e., not original spill), | |||
1502 | // save them to SpillsToIns. | |||
1503 | for (const auto &Ent : SpillsToKeep) { | |||
1504 | if (Ent.second) | |||
1505 | SpillsToIns[Ent.first->getBlock()] = Ent.second; | |||
1506 | } | |||
1507 | } | |||
1508 | ||||
1509 | /// For spills with equal values, remove redundant spills and hoist those left | |||
1510 | /// to less hot spots. | |||
1511 | /// | |||
1512 | /// Spills with equal values will be collected into the same set in | |||
1513 | /// MergeableSpills when spill is inserted. These equal spills are originated | |||
1514 | /// from the same defining instruction and are dominated by the instruction. | |||
1515 | /// Before hoisting all the equal spills, redundant spills inside in the same | |||
1516 | /// BB are first marked to be deleted. Then starting from the spills left, walk | |||
1517 | /// up on the dominator tree towards the Root node where the define instruction | |||
1518 | /// is located, mark the dominated spills to be deleted along the way and | |||
1519 | /// collect the BB nodes on the path from non-dominated spills to the define | |||
1520 | /// instruction into a WorkSet. The nodes in WorkSet are the candidate places | |||
1521 | /// where we are considering to hoist the spills. We iterate the WorkSet in | |||
1522 | /// bottom-up order, and for each node, we will decide whether to hoist spills | |||
1523 | /// inside its subtree to that node. In this way, we can get benefit locally | |||
1524 | /// even if hoisting all the equal spills to one cold place is impossible. | |||
1525 | void HoistSpillHelper::hoistAllSpills() { | |||
1526 | SmallVector<Register, 4> NewVRegs; | |||
1527 | LiveRangeEdit Edit(nullptr, NewVRegs, MF, LIS, &VRM, this); | |||
1528 | ||||
1529 | for (unsigned i = 0, e = MRI.getNumVirtRegs(); i != e; ++i) { | |||
1530 | Register Reg = Register::index2VirtReg(i); | |||
1531 | Register Original = VRM.getPreSplitReg(Reg); | |||
1532 | if (!MRI.def_empty(Reg)) | |||
1533 | Virt2SiblingsMap[Original].insert(Reg); | |||
1534 | } | |||
1535 | ||||
1536 | // Each entry in MergeableSpills contains a spill set with equal values. | |||
1537 | for (auto &Ent : MergeableSpills) { | |||
1538 | int Slot = Ent.first.first; | |||
1539 | LiveInterval &OrigLI = *StackSlotToOrigLI[Slot]; | |||
1540 | VNInfo *OrigVNI = Ent.first.second; | |||
1541 | SmallPtrSet<MachineInstr *, 16> &EqValSpills = Ent.second; | |||
1542 | if (Ent.second.empty()) | |||
1543 | continue; | |||
1544 | ||||
1545 | LLVM_DEBUG({do { } while (false) | |||
1546 | dbgs() << "\nFor Slot" << Slot << " and VN" << OrigVNI->id << ":\n"do { } while (false) | |||
1547 | << "Equal spills in BB: ";do { } while (false) | |||
1548 | for (const auto spill : EqValSpills)do { } while (false) | |||
1549 | dbgs() << spill->getParent()->getNumber() << " ";do { } while (false) | |||
1550 | dbgs() << "\n";do { } while (false) | |||
1551 | })do { } while (false); | |||
1552 | ||||
1553 | // SpillsToRm is the spill set to be removed from EqValSpills. | |||
1554 | SmallVector<MachineInstr *, 16> SpillsToRm; | |||
1555 | // SpillsToIns is the spill set to be newly inserted after hoisting. | |||
1556 | DenseMap<MachineBasicBlock *, unsigned> SpillsToIns; | |||
1557 | ||||
1558 | runHoistSpills(OrigLI, *OrigVNI, EqValSpills, SpillsToRm, SpillsToIns); | |||
1559 | ||||
1560 | LLVM_DEBUG({do { } while (false) | |||
1561 | dbgs() << "Finally inserted spills in BB: ";do { } while (false) | |||
1562 | for (const auto &Ispill : SpillsToIns)do { } while (false) | |||
1563 | dbgs() << Ispill.first->getNumber() << " ";do { } while (false) | |||
1564 | dbgs() << "\nFinally removed spills in BB: ";do { } while (false) | |||
1565 | for (const auto Rspill : SpillsToRm)do { } while (false) | |||
1566 | dbgs() << Rspill->getParent()->getNumber() << " ";do { } while (false) | |||
1567 | dbgs() << "\n";do { } while (false) | |||
1568 | })do { } while (false); | |||
1569 | ||||
1570 | // Stack live range update. | |||
1571 | LiveInterval &StackIntvl = LSS.getInterval(Slot); | |||
1572 | if (!SpillsToIns.empty() || !SpillsToRm.empty()) | |||
1573 | StackIntvl.MergeValueInAsValue(OrigLI, OrigVNI, | |||
1574 | StackIntvl.getValNumInfo(0)); | |||
1575 | ||||
1576 | // Insert hoisted spills. | |||
1577 | for (auto const &Insert : SpillsToIns) { | |||
1578 | MachineBasicBlock *BB = Insert.first; | |||
1579 | Register LiveReg = Insert.second; | |||
1580 | MachineBasicBlock::iterator MII = IPA.getLastInsertPointIter(OrigLI, *BB); | |||
1581 | MachineInstrSpan MIS(MII, BB); | |||
1582 | TII.storeRegToStackSlot(*BB, MII, LiveReg, false, Slot, | |||
1583 | MRI.getRegClass(LiveReg), &TRI); | |||
1584 | LIS.InsertMachineInstrRangeInMaps(MIS.begin(), MII); | |||
1585 | for (const MachineInstr &MI : make_range(MIS.begin(), MII)) | |||
1586 | getVDefInterval(MI, LIS); | |||
1587 | ++NumSpills; | |||
1588 | } | |||
1589 | ||||
1590 | // Remove redundant spills or change them to dead instructions. | |||
1591 | NumSpills -= SpillsToRm.size(); | |||
1592 | for (auto const RMEnt : SpillsToRm) { | |||
1593 | RMEnt->setDesc(TII.get(TargetOpcode::KILL)); | |||
1594 | for (unsigned i = RMEnt->getNumOperands(); i; --i) { | |||
1595 | MachineOperand &MO = RMEnt->getOperand(i - 1); | |||
1596 | if (MO.isReg() && MO.isImplicit() && MO.isDef() && !MO.isDead()) | |||
1597 | RMEnt->RemoveOperand(i - 1); | |||
1598 | } | |||
1599 | } | |||
1600 | Edit.eliminateDeadDefs(SpillsToRm, None, AA); | |||
1601 | } | |||
1602 | } | |||
1603 | ||||
1604 | /// For VirtReg clone, the \p New register should have the same physreg or | |||
1605 | /// stackslot as the \p old register. | |||
1606 | void HoistSpillHelper::LRE_DidCloneVirtReg(Register New, Register Old) { | |||
1607 | if (VRM.hasPhys(Old)) | |||
1608 | VRM.assignVirt2Phys(New, VRM.getPhys(Old)); | |||
1609 | else if (VRM.getStackSlot(Old) != VirtRegMap::NO_STACK_SLOT) | |||
1610 | VRM.assignVirt2StackSlot(New, VRM.getStackSlot(Old)); | |||
1611 | else | |||
1612 | llvm_unreachable("VReg should be assigned either physreg or stackslot")__builtin_unreachable(); | |||
1613 | if (VRM.hasShape(Old)) | |||
1614 | VRM.assignVirt2Shape(New, VRM.getShape(Old)); | |||
1615 | } |
1 | //==- llvm/CodeGen/MachineDominators.h - Machine Dom Calculation -*- C++ -*-==// |
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 classes mirroring those in llvm/Analysis/Dominators.h, |
10 | // but for target-specific code rather than target-independent IR. |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #ifndef LLVM_CODEGEN_MACHINEDOMINATORS_H |
15 | #define LLVM_CODEGEN_MACHINEDOMINATORS_H |
16 | |
17 | #include "llvm/ADT/SmallSet.h" |
18 | #include "llvm/ADT/SmallVector.h" |
19 | #include "llvm/CodeGen/MachineBasicBlock.h" |
20 | #include "llvm/CodeGen/MachineFunctionPass.h" |
21 | #include "llvm/CodeGen/MachineInstr.h" |
22 | #include "llvm/Support/GenericDomTree.h" |
23 | #include "llvm/Support/GenericDomTreeConstruction.h" |
24 | #include <cassert> |
25 | #include <memory> |
26 | |
27 | namespace llvm { |
28 | |
29 | template <> |
30 | inline void DominatorTreeBase<MachineBasicBlock, false>::addRoot( |
31 | MachineBasicBlock *MBB) { |
32 | this->Roots.push_back(MBB); |
33 | } |
34 | |
35 | extern template class DomTreeNodeBase<MachineBasicBlock>; |
36 | extern template class DominatorTreeBase<MachineBasicBlock, false>; // DomTree |
37 | extern template class DominatorTreeBase<MachineBasicBlock, true>; // PostDomTree |
38 | |
39 | using MachineDomTreeNode = DomTreeNodeBase<MachineBasicBlock>; |
40 | |
41 | //===------------------------------------- |
42 | /// DominatorTree Class - Concrete subclass of DominatorTreeBase that is used to |
43 | /// compute a normal dominator tree. |
44 | /// |
45 | class MachineDominatorTree : public MachineFunctionPass { |
46 | using DomTreeT = DomTreeBase<MachineBasicBlock>; |
47 | |
48 | /// Helper structure used to hold all the basic blocks |
49 | /// involved in the split of a critical edge. |
50 | struct CriticalEdge { |
51 | MachineBasicBlock *FromBB; |
52 | MachineBasicBlock *ToBB; |
53 | MachineBasicBlock *NewBB; |
54 | }; |
55 | |
56 | /// Pile up all the critical edges to be split. |
57 | /// The splitting of a critical edge is local and thus, it is possible |
58 | /// to apply several of those changes at the same time. |
59 | mutable SmallVector<CriticalEdge, 32> CriticalEdgesToSplit; |
60 | |
61 | /// Remember all the basic blocks that are inserted during |
62 | /// edge splitting. |
63 | /// Invariant: NewBBs == all the basic blocks contained in the NewBB |
64 | /// field of all the elements of CriticalEdgesToSplit. |
65 | /// I.e., forall elt in CriticalEdgesToSplit, it exists BB in NewBBs |
66 | /// such as BB == elt.NewBB. |
67 | mutable SmallSet<MachineBasicBlock *, 32> NewBBs; |
68 | |
69 | /// The DominatorTreeBase that is used to compute a normal dominator tree. |
70 | std::unique_ptr<DomTreeT> DT; |
71 | |
72 | /// Apply all the recorded critical edges to the DT. |
73 | /// This updates the underlying DT information in a way that uses |
74 | /// the fast query path of DT as much as possible. |
75 | /// |
76 | /// \post CriticalEdgesToSplit.empty(). |
77 | void applySplitCriticalEdges() const; |
78 | |
79 | public: |
80 | static char ID; // Pass ID, replacement for typeid |
81 | |
82 | MachineDominatorTree(); |
83 | explicit MachineDominatorTree(MachineFunction &MF) : MachineFunctionPass(ID) { |
84 | calculate(MF); |
85 | } |
86 | |
87 | DomTreeT &getBase() { |
88 | if (!DT) DT.reset(new DomTreeT()); |
89 | applySplitCriticalEdges(); |
90 | return *DT; |
91 | } |
92 | |
93 | void getAnalysisUsage(AnalysisUsage &AU) const override; |
94 | |
95 | MachineBasicBlock *getRoot() const { |
96 | applySplitCriticalEdges(); |
97 | return DT->getRoot(); |
98 | } |
99 | |
100 | MachineDomTreeNode *getRootNode() const { |
101 | applySplitCriticalEdges(); |
102 | return DT->getRootNode(); |
103 | } |
104 | |
105 | bool runOnMachineFunction(MachineFunction &F) override; |
106 | |
107 | void calculate(MachineFunction &F); |
108 | |
109 | bool dominates(const MachineDomTreeNode *A, |
110 | const MachineDomTreeNode *B) const { |
111 | applySplitCriticalEdges(); |
112 | return DT->dominates(A, B); |
113 | } |
114 | |
115 | bool dominates(const MachineBasicBlock *A, const MachineBasicBlock *B) const { |
116 | applySplitCriticalEdges(); |
117 | return DT->dominates(A, B); |
118 | } |
119 | |
120 | // dominates - Return true if A dominates B. This performs the |
121 | // special checks necessary if A and B are in the same basic block. |
122 | bool dominates(const MachineInstr *A, const MachineInstr *B) const { |
123 | applySplitCriticalEdges(); |
124 | const MachineBasicBlock *BBA = A->getParent(), *BBB = B->getParent(); |
125 | if (BBA != BBB) return DT->dominates(BBA, BBB); |
126 | |
127 | // Loop through the basic block until we find A or B. |
128 | MachineBasicBlock::const_iterator I = BBA->begin(); |
129 | for (; &*I != A && &*I != B; ++I) |
130 | /*empty*/ ; |
131 | |
132 | return &*I == A; |
133 | } |
134 | |
135 | bool properlyDominates(const MachineDomTreeNode *A, |
136 | const MachineDomTreeNode *B) const { |
137 | applySplitCriticalEdges(); |
138 | return DT->properlyDominates(A, B); |
139 | } |
140 | |
141 | bool properlyDominates(const MachineBasicBlock *A, |
142 | const MachineBasicBlock *B) const { |
143 | applySplitCriticalEdges(); |
144 | return DT->properlyDominates(A, B); |
145 | } |
146 | |
147 | /// findNearestCommonDominator - Find nearest common dominator basic block |
148 | /// for basic block A and B. If there is no such block then return NULL. |
149 | MachineBasicBlock *findNearestCommonDominator(MachineBasicBlock *A, |
150 | MachineBasicBlock *B) { |
151 | applySplitCriticalEdges(); |
152 | return DT->findNearestCommonDominator(A, B); |
153 | } |
154 | |
155 | MachineDomTreeNode *operator[](MachineBasicBlock *BB) const { |
156 | applySplitCriticalEdges(); |
157 | return DT->getNode(BB); |
158 | } |
159 | |
160 | /// getNode - return the (Post)DominatorTree node for the specified basic |
161 | /// block. This is the same as using operator[] on this class. |
162 | /// |
163 | MachineDomTreeNode *getNode(MachineBasicBlock *BB) const { |
164 | applySplitCriticalEdges(); |
165 | return DT->getNode(BB); |
166 | } |
167 | |
168 | /// addNewBlock - Add a new node to the dominator tree information. This |
169 | /// creates a new node as a child of DomBB dominator node,linking it into |
170 | /// the children list of the immediate dominator. |
171 | MachineDomTreeNode *addNewBlock(MachineBasicBlock *BB, |
172 | MachineBasicBlock *DomBB) { |
173 | applySplitCriticalEdges(); |
174 | return DT->addNewBlock(BB, DomBB); |
175 | } |
176 | |
177 | /// changeImmediateDominator - This method is used to update the dominator |
178 | /// tree information when a node's immediate dominator changes. |
179 | /// |
180 | void changeImmediateDominator(MachineBasicBlock *N, |
181 | MachineBasicBlock *NewIDom) { |
182 | applySplitCriticalEdges(); |
183 | DT->changeImmediateDominator(N, NewIDom); |
184 | } |
185 | |
186 | void changeImmediateDominator(MachineDomTreeNode *N, |
187 | MachineDomTreeNode *NewIDom) { |
188 | applySplitCriticalEdges(); |
189 | DT->changeImmediateDominator(N, NewIDom); |
190 | } |
191 | |
192 | /// eraseNode - Removes a node from the dominator tree. Block must not |
193 | /// dominate any other blocks. Removes node from its immediate dominator's |
194 | /// children list. Deletes dominator node associated with basic block BB. |
195 | void eraseNode(MachineBasicBlock *BB) { |
196 | applySplitCriticalEdges(); |
197 | DT->eraseNode(BB); |
198 | } |
199 | |
200 | /// splitBlock - BB is split and now it has one successor. Update dominator |
201 | /// tree to reflect this change. |
202 | void splitBlock(MachineBasicBlock* NewBB) { |
203 | applySplitCriticalEdges(); |
204 | DT->splitBlock(NewBB); |
205 | } |
206 | |
207 | /// isReachableFromEntry - Return true if A is dominated by the entry |
208 | /// block of the function containing it. |
209 | bool isReachableFromEntry(const MachineBasicBlock *A) { |
210 | applySplitCriticalEdges(); |
211 | return DT->isReachableFromEntry(A); |
212 | } |
213 | |
214 | void releaseMemory() override; |
215 | |
216 | void verifyAnalysis() const override; |
217 | |
218 | void print(raw_ostream &OS, const Module*) const override; |
219 | |
220 | /// Record that the critical edge (FromBB, ToBB) has been |
221 | /// split with NewBB. |
222 | /// This is best to use this method instead of directly update the |
223 | /// underlying information, because this helps mitigating the |
224 | /// number of time the DT information is invalidated. |
225 | /// |
226 | /// \note Do not use this method with regular edges. |
227 | /// |
228 | /// \note To benefit from the compile time improvement incurred by this |
229 | /// method, the users of this method have to limit the queries to the DT |
230 | /// interface between two edges splitting. In other words, they have to |
231 | /// pack the splitting of critical edges as much as possible. |
232 | void recordSplitCriticalEdge(MachineBasicBlock *FromBB, |
233 | MachineBasicBlock *ToBB, |
234 | MachineBasicBlock *NewBB) { |
235 | bool Inserted = NewBBs.insert(NewBB).second; |
236 | (void)Inserted; |
237 | assert(Inserted &&((void)0) |
238 | "A basic block inserted via edge splitting cannot appear twice")((void)0); |
239 | CriticalEdgesToSplit.push_back({FromBB, ToBB, NewBB}); |
240 | } |
241 | }; |
242 | |
243 | //===------------------------------------- |
244 | /// DominatorTree GraphTraits specialization so the DominatorTree can be |
245 | /// iterable by generic graph iterators. |
246 | /// |
247 | |
248 | template <class Node, class ChildIterator> |
249 | struct MachineDomTreeGraphTraitsBase { |
250 | using NodeRef = Node *; |
251 | using ChildIteratorType = ChildIterator; |
252 | |
253 | static NodeRef getEntryNode(NodeRef N) { return N; } |
254 | static ChildIteratorType child_begin(NodeRef N) { return N->begin(); } |
255 | static ChildIteratorType child_end(NodeRef N) { return N->end(); } |
256 | }; |
257 | |
258 | template <class T> struct GraphTraits; |
259 | |
260 | template <> |
261 | struct GraphTraits<MachineDomTreeNode *> |
262 | : public MachineDomTreeGraphTraitsBase<MachineDomTreeNode, |
263 | MachineDomTreeNode::const_iterator> { |
264 | }; |
265 | |
266 | template <> |
267 | struct GraphTraits<const MachineDomTreeNode *> |
268 | : public MachineDomTreeGraphTraitsBase<const MachineDomTreeNode, |
269 | MachineDomTreeNode::const_iterator> { |
270 | }; |
271 | |
272 | template <> struct GraphTraits<MachineDominatorTree*> |
273 | : public GraphTraits<MachineDomTreeNode *> { |
274 | static NodeRef getEntryNode(MachineDominatorTree *DT) { |
275 | return DT->getRootNode(); |
276 | } |
277 | }; |
278 | |
279 | } // end namespace llvm |
280 | |
281 | #endif // LLVM_CODEGEN_MACHINEDOMINATORS_H |
1 | //===- GenericDomTree.h - Generic dominator trees for graphs ----*- C++ -*-===// |
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 | /// \file |
9 | /// |
10 | /// This file defines a set of templates that efficiently compute a dominator |
11 | /// tree over a generic graph. This is used typically in LLVM for fast |
12 | /// dominance queries on the CFG, but is fully generic w.r.t. the underlying |
13 | /// graph types. |
14 | /// |
15 | /// Unlike ADT/* graph algorithms, generic dominator tree has more requirements |
16 | /// on the graph's NodeRef. The NodeRef should be a pointer and, |
17 | /// NodeRef->getParent() must return the parent node that is also a pointer. |
18 | /// |
19 | /// FIXME: Maybe GenericDomTree needs a TreeTraits, instead of GraphTraits. |
20 | /// |
21 | //===----------------------------------------------------------------------===// |
22 | |
23 | #ifndef LLVM_SUPPORT_GENERICDOMTREE_H |
24 | #define LLVM_SUPPORT_GENERICDOMTREE_H |
25 | |
26 | #include "llvm/ADT/DenseMap.h" |
27 | #include "llvm/ADT/GraphTraits.h" |
28 | #include "llvm/ADT/STLExtras.h" |
29 | #include "llvm/ADT/SmallPtrSet.h" |
30 | #include "llvm/ADT/SmallVector.h" |
31 | #include "llvm/Support/CFGDiff.h" |
32 | #include "llvm/Support/CFGUpdate.h" |
33 | #include "llvm/Support/raw_ostream.h" |
34 | #include <algorithm> |
35 | #include <cassert> |
36 | #include <cstddef> |
37 | #include <iterator> |
38 | #include <memory> |
39 | #include <type_traits> |
40 | #include <utility> |
41 | |
42 | namespace llvm { |
43 | |
44 | template <typename NodeT, bool IsPostDom> |
45 | class DominatorTreeBase; |
46 | |
47 | namespace DomTreeBuilder { |
48 | template <typename DomTreeT> |
49 | struct SemiNCAInfo; |
50 | } // namespace DomTreeBuilder |
51 | |
52 | /// Base class for the actual dominator tree node. |
53 | template <class NodeT> class DomTreeNodeBase { |
54 | friend class PostDominatorTree; |
55 | friend class DominatorTreeBase<NodeT, false>; |
56 | friend class DominatorTreeBase<NodeT, true>; |
57 | friend struct DomTreeBuilder::SemiNCAInfo<DominatorTreeBase<NodeT, false>>; |
58 | friend struct DomTreeBuilder::SemiNCAInfo<DominatorTreeBase<NodeT, true>>; |
59 | |
60 | NodeT *TheBB; |
61 | DomTreeNodeBase *IDom; |
62 | unsigned Level; |
63 | SmallVector<DomTreeNodeBase *, 4> Children; |
64 | mutable unsigned DFSNumIn = ~0; |
65 | mutable unsigned DFSNumOut = ~0; |
66 | |
67 | public: |
68 | DomTreeNodeBase(NodeT *BB, DomTreeNodeBase *iDom) |
69 | : TheBB(BB), IDom(iDom), Level(IDom ? IDom->Level + 1 : 0) {} |
70 | |
71 | using iterator = typename SmallVector<DomTreeNodeBase *, 4>::iterator; |
72 | using const_iterator = |
73 | typename SmallVector<DomTreeNodeBase *, 4>::const_iterator; |
74 | |
75 | iterator begin() { return Children.begin(); } |
76 | iterator end() { return Children.end(); } |
77 | const_iterator begin() const { return Children.begin(); } |
78 | const_iterator end() const { return Children.end(); } |
79 | |
80 | DomTreeNodeBase *const &back() const { return Children.back(); } |
81 | DomTreeNodeBase *&back() { return Children.back(); } |
82 | |
83 | iterator_range<iterator> children() { return make_range(begin(), end()); } |
84 | iterator_range<const_iterator> children() const { |
85 | return make_range(begin(), end()); |
86 | } |
87 | |
88 | NodeT *getBlock() const { return TheBB; } |
89 | DomTreeNodeBase *getIDom() const { return IDom; } |
90 | unsigned getLevel() const { return Level; } |
91 | |
92 | std::unique_ptr<DomTreeNodeBase> addChild( |
93 | std::unique_ptr<DomTreeNodeBase> C) { |
94 | Children.push_back(C.get()); |
95 | return C; |
96 | } |
97 | |
98 | bool isLeaf() const { return Children.empty(); } |
99 | size_t getNumChildren() const { return Children.size(); } |
100 | |
101 | void clearAllChildren() { Children.clear(); } |
102 | |
103 | bool compare(const DomTreeNodeBase *Other) const { |
104 | if (getNumChildren() != Other->getNumChildren()) |
105 | return true; |
106 | |
107 | if (Level != Other->Level) return true; |
108 | |
109 | SmallPtrSet<const NodeT *, 4> OtherChildren; |
110 | for (const DomTreeNodeBase *I : *Other) { |
111 | const NodeT *Nd = I->getBlock(); |
112 | OtherChildren.insert(Nd); |
113 | } |
114 | |
115 | for (const DomTreeNodeBase *I : *this) { |
116 | const NodeT *N = I->getBlock(); |
117 | if (OtherChildren.count(N) == 0) |
118 | return true; |
119 | } |
120 | return false; |
121 | } |
122 | |
123 | void setIDom(DomTreeNodeBase *NewIDom) { |
124 | assert(IDom && "No immediate dominator?")((void)0); |
125 | if (IDom == NewIDom) return; |
126 | |
127 | auto I = find(IDom->Children, this); |
128 | assert(I != IDom->Children.end() &&((void)0) |
129 | "Not in immediate dominator children set!")((void)0); |
130 | // I am no longer your child... |
131 | IDom->Children.erase(I); |
132 | |
133 | // Switch to new dominator |
134 | IDom = NewIDom; |
135 | IDom->Children.push_back(this); |
136 | |
137 | UpdateLevel(); |
138 | } |
139 | |
140 | /// getDFSNumIn/getDFSNumOut - These return the DFS visitation order for nodes |
141 | /// in the dominator tree. They are only guaranteed valid if |
142 | /// updateDFSNumbers() has been called. |
143 | unsigned getDFSNumIn() const { return DFSNumIn; } |
144 | unsigned getDFSNumOut() const { return DFSNumOut; } |
145 | |
146 | private: |
147 | // Return true if this node is dominated by other. Use this only if DFS info |
148 | // is valid. |
149 | bool DominatedBy(const DomTreeNodeBase *other) const { |
150 | return this->DFSNumIn >= other->DFSNumIn && |
151 | this->DFSNumOut <= other->DFSNumOut; |
152 | } |
153 | |
154 | void UpdateLevel() { |
155 | assert(IDom)((void)0); |
156 | if (Level == IDom->Level + 1) return; |
157 | |
158 | SmallVector<DomTreeNodeBase *, 64> WorkStack = {this}; |
159 | |
160 | while (!WorkStack.empty()) { |
161 | DomTreeNodeBase *Current = WorkStack.pop_back_val(); |
162 | Current->Level = Current->IDom->Level + 1; |
163 | |
164 | for (DomTreeNodeBase *C : *Current) { |
165 | assert(C->IDom)((void)0); |
166 | if (C->Level != C->IDom->Level + 1) WorkStack.push_back(C); |
167 | } |
168 | } |
169 | } |
170 | }; |
171 | |
172 | template <class NodeT> |
173 | raw_ostream &operator<<(raw_ostream &O, const DomTreeNodeBase<NodeT> *Node) { |
174 | if (Node->getBlock()) |
175 | Node->getBlock()->printAsOperand(O, false); |
176 | else |
177 | O << " <<exit node>>"; |
178 | |
179 | O << " {" << Node->getDFSNumIn() << "," << Node->getDFSNumOut() << "} [" |
180 | << Node->getLevel() << "]\n"; |
181 | |
182 | return O; |
183 | } |
184 | |
185 | template <class NodeT> |
186 | void PrintDomTree(const DomTreeNodeBase<NodeT> *N, raw_ostream &O, |
187 | unsigned Lev) { |
188 | O.indent(2 * Lev) << "[" << Lev << "] " << N; |
189 | for (typename DomTreeNodeBase<NodeT>::const_iterator I = N->begin(), |
190 | E = N->end(); |
191 | I != E; ++I) |
192 | PrintDomTree<NodeT>(*I, O, Lev + 1); |
193 | } |
194 | |
195 | namespace DomTreeBuilder { |
196 | // The routines below are provided in a separate header but referenced here. |
197 | template <typename DomTreeT> |
198 | void Calculate(DomTreeT &DT); |
199 | |
200 | template <typename DomTreeT> |
201 | void CalculateWithUpdates(DomTreeT &DT, |
202 | ArrayRef<typename DomTreeT::UpdateType> Updates); |
203 | |
204 | template <typename DomTreeT> |
205 | void InsertEdge(DomTreeT &DT, typename DomTreeT::NodePtr From, |
206 | typename DomTreeT::NodePtr To); |
207 | |
208 | template <typename DomTreeT> |
209 | void DeleteEdge(DomTreeT &DT, typename DomTreeT::NodePtr From, |
210 | typename DomTreeT::NodePtr To); |
211 | |
212 | template <typename DomTreeT> |
213 | void ApplyUpdates(DomTreeT &DT, |
214 | GraphDiff<typename DomTreeT::NodePtr, |
215 | DomTreeT::IsPostDominator> &PreViewCFG, |
216 | GraphDiff<typename DomTreeT::NodePtr, |
217 | DomTreeT::IsPostDominator> *PostViewCFG); |
218 | |
219 | template <typename DomTreeT> |
220 | bool Verify(const DomTreeT &DT, typename DomTreeT::VerificationLevel VL); |
221 | } // namespace DomTreeBuilder |
222 | |
223 | /// Core dominator tree base class. |
224 | /// |
225 | /// This class is a generic template over graph nodes. It is instantiated for |
226 | /// various graphs in the LLVM IR or in the code generator. |
227 | template <typename NodeT, bool IsPostDom> |
228 | class DominatorTreeBase { |
229 | public: |
230 | static_assert(std::is_pointer<typename GraphTraits<NodeT *>::NodeRef>::value, |
231 | "Currently DominatorTreeBase supports only pointer nodes"); |
232 | using NodeType = NodeT; |
233 | using NodePtr = NodeT *; |
234 | using ParentPtr = decltype(std::declval<NodeT *>()->getParent()); |
235 | static_assert(std::is_pointer<ParentPtr>::value, |
236 | "Currently NodeT's parent must be a pointer type"); |
237 | using ParentType = std::remove_pointer_t<ParentPtr>; |
238 | static constexpr bool IsPostDominator = IsPostDom; |
239 | |
240 | using UpdateType = cfg::Update<NodePtr>; |
241 | using UpdateKind = cfg::UpdateKind; |
242 | static constexpr UpdateKind Insert = UpdateKind::Insert; |
243 | static constexpr UpdateKind Delete = UpdateKind::Delete; |
244 | |
245 | enum class VerificationLevel { Fast, Basic, Full }; |
246 | |
247 | protected: |
248 | // Dominators always have a single root, postdominators can have more. |
249 | SmallVector<NodeT *, IsPostDom ? 4 : 1> Roots; |
250 | |
251 | using DomTreeNodeMapType = |
252 | DenseMap<NodeT *, std::unique_ptr<DomTreeNodeBase<NodeT>>>; |
253 | DomTreeNodeMapType DomTreeNodes; |
254 | DomTreeNodeBase<NodeT> *RootNode = nullptr; |
255 | ParentPtr Parent = nullptr; |
256 | |
257 | mutable bool DFSInfoValid = false; |
258 | mutable unsigned int SlowQueries = 0; |
259 | |
260 | friend struct DomTreeBuilder::SemiNCAInfo<DominatorTreeBase>; |
261 | |
262 | public: |
263 | DominatorTreeBase() {} |
264 | |
265 | DominatorTreeBase(DominatorTreeBase &&Arg) |
266 | : Roots(std::move(Arg.Roots)), |
267 | DomTreeNodes(std::move(Arg.DomTreeNodes)), |
268 | RootNode(Arg.RootNode), |
269 | Parent(Arg.Parent), |
270 | DFSInfoValid(Arg.DFSInfoValid), |
271 | SlowQueries(Arg.SlowQueries) { |
272 | Arg.wipe(); |
273 | } |
274 | |
275 | DominatorTreeBase &operator=(DominatorTreeBase &&RHS) { |
276 | Roots = std::move(RHS.Roots); |
277 | DomTreeNodes = std::move(RHS.DomTreeNodes); |
278 | RootNode = RHS.RootNode; |
279 | Parent = RHS.Parent; |
280 | DFSInfoValid = RHS.DFSInfoValid; |
281 | SlowQueries = RHS.SlowQueries; |
282 | RHS.wipe(); |
283 | return *this; |
284 | } |
285 | |
286 | DominatorTreeBase(const DominatorTreeBase &) = delete; |
287 | DominatorTreeBase &operator=(const DominatorTreeBase &) = delete; |
288 | |
289 | /// Iteration over roots. |
290 | /// |
291 | /// This may include multiple blocks if we are computing post dominators. |
292 | /// For forward dominators, this will always be a single block (the entry |
293 | /// block). |
294 | using root_iterator = typename SmallVectorImpl<NodeT *>::iterator; |
295 | using const_root_iterator = typename SmallVectorImpl<NodeT *>::const_iterator; |
296 | |
297 | root_iterator root_begin() { return Roots.begin(); } |
298 | const_root_iterator root_begin() const { return Roots.begin(); } |
299 | root_iterator root_end() { return Roots.end(); } |
300 | const_root_iterator root_end() const { return Roots.end(); } |
301 | |
302 | size_t root_size() const { return Roots.size(); } |
303 | |
304 | iterator_range<root_iterator> roots() { |
305 | return make_range(root_begin(), root_end()); |
306 | } |
307 | iterator_range<const_root_iterator> roots() const { |
308 | return make_range(root_begin(), root_end()); |
309 | } |
310 | |
311 | /// isPostDominator - Returns true if analysis based of postdoms |
312 | /// |
313 | bool isPostDominator() const { return IsPostDominator; } |
314 | |
315 | /// compare - Return false if the other dominator tree base matches this |
316 | /// dominator tree base. Otherwise return true. |
317 | bool compare(const DominatorTreeBase &Other) const { |
318 | if (Parent != Other.Parent) return true; |
319 | |
320 | if (Roots.size() != Other.Roots.size()) |
321 | return true; |
322 | |
323 | if (!std::is_permutation(Roots.begin(), Roots.end(), Other.Roots.begin())) |
324 | return true; |
325 | |
326 | const DomTreeNodeMapType &OtherDomTreeNodes = Other.DomTreeNodes; |
327 | if (DomTreeNodes.size() != OtherDomTreeNodes.size()) |
328 | return true; |
329 | |
330 | for (const auto &DomTreeNode : DomTreeNodes) { |
331 | NodeT *BB = DomTreeNode.first; |
332 | typename DomTreeNodeMapType::const_iterator OI = |
333 | OtherDomTreeNodes.find(BB); |
334 | if (OI == OtherDomTreeNodes.end()) |
335 | return true; |
336 | |
337 | DomTreeNodeBase<NodeT> &MyNd = *DomTreeNode.second; |
338 | DomTreeNodeBase<NodeT> &OtherNd = *OI->second; |
339 | |
340 | if (MyNd.compare(&OtherNd)) |
341 | return true; |
342 | } |
343 | |
344 | return false; |
345 | } |
346 | |
347 | /// getNode - return the (Post)DominatorTree node for the specified basic |
348 | /// block. This is the same as using operator[] on this class. The result |
349 | /// may (but is not required to) be null for a forward (backwards) |
350 | /// statically unreachable block. |
351 | DomTreeNodeBase<NodeT> *getNode(const NodeT *BB) const { |
352 | auto I = DomTreeNodes.find(BB); |
353 | if (I != DomTreeNodes.end()) |
354 | return I->second.get(); |
355 | return nullptr; |
356 | } |
357 | |
358 | /// See getNode. |
359 | DomTreeNodeBase<NodeT> *operator[](const NodeT *BB) const { |
360 | return getNode(BB); |
361 | } |
362 | |
363 | /// getRootNode - This returns the entry node for the CFG of the function. If |
364 | /// this tree represents the post-dominance relations for a function, however, |
365 | /// this root may be a node with the block == NULL. This is the case when |
366 | /// there are multiple exit nodes from a particular function. Consumers of |
367 | /// post-dominance information must be capable of dealing with this |
368 | /// possibility. |
369 | /// |
370 | DomTreeNodeBase<NodeT> *getRootNode() { return RootNode; } |
371 | const DomTreeNodeBase<NodeT> *getRootNode() const { return RootNode; } |
372 | |
373 | /// Get all nodes dominated by R, including R itself. |
374 | void getDescendants(NodeT *R, SmallVectorImpl<NodeT *> &Result) const { |
375 | Result.clear(); |
376 | const DomTreeNodeBase<NodeT> *RN = getNode(R); |
377 | if (!RN) |
378 | return; // If R is unreachable, it will not be present in the DOM tree. |
379 | SmallVector<const DomTreeNodeBase<NodeT> *, 8> WL; |
380 | WL.push_back(RN); |
381 | |
382 | while (!WL.empty()) { |
383 | const DomTreeNodeBase<NodeT> *N = WL.pop_back_val(); |
384 | Result.push_back(N->getBlock()); |
385 | WL.append(N->begin(), N->end()); |
386 | } |
387 | } |
388 | |
389 | /// properlyDominates - Returns true iff A dominates B and A != B. |
390 | /// Note that this is not a constant time operation! |
391 | /// |
392 | bool properlyDominates(const DomTreeNodeBase<NodeT> *A, |
393 | const DomTreeNodeBase<NodeT> *B) const { |
394 | if (!A || !B) |
395 | return false; |
396 | if (A == B) |
397 | return false; |
398 | return dominates(A, B); |
399 | } |
400 | |
401 | bool properlyDominates(const NodeT *A, const NodeT *B) const; |
402 | |
403 | /// isReachableFromEntry - Return true if A is dominated by the entry |
404 | /// block of the function containing it. |
405 | bool isReachableFromEntry(const NodeT *A) const { |
406 | assert(!this->isPostDominator() &&((void)0) |
407 | "This is not implemented for post dominators")((void)0); |
408 | return isReachableFromEntry(getNode(const_cast<NodeT *>(A))); |
409 | } |
410 | |
411 | bool isReachableFromEntry(const DomTreeNodeBase<NodeT> *A) const { return A; } |
412 | |
413 | /// dominates - Returns true iff A dominates B. Note that this is not a |
414 | /// constant time operation! |
415 | /// |
416 | bool dominates(const DomTreeNodeBase<NodeT> *A, |
417 | const DomTreeNodeBase<NodeT> *B) const { |
418 | // A node trivially dominates itself. |
419 | if (B == A) |
420 | return true; |
421 | |
422 | // An unreachable node is dominated by anything. |
423 | if (!isReachableFromEntry(B)) |
424 | return true; |
425 | |
426 | // And dominates nothing. |
427 | if (!isReachableFromEntry(A)) |
428 | return false; |
429 | |
430 | if (B->getIDom() == A) return true; |
431 | |
432 | if (A->getIDom() == B) return false; |
433 | |
434 | // A can only dominate B if it is higher in the tree. |
435 | if (A->getLevel() >= B->getLevel()) return false; |
436 | |
437 | // Compare the result of the tree walk and the dfs numbers, if expensive |
438 | // checks are enabled. |
439 | #ifdef EXPENSIVE_CHECKS |
440 | assert((!DFSInfoValid ||((void)0) |
441 | (dominatedBySlowTreeWalk(A, B) == B->DominatedBy(A))) &&((void)0) |
442 | "Tree walk disagrees with dfs numbers!")((void)0); |
443 | #endif |
444 | |
445 | if (DFSInfoValid) |
446 | return B->DominatedBy(A); |
447 | |
448 | // If we end up with too many slow queries, just update the |
449 | // DFS numbers on the theory that we are going to keep querying. |
450 | SlowQueries++; |
451 | if (SlowQueries > 32) { |
452 | updateDFSNumbers(); |
453 | return B->DominatedBy(A); |
454 | } |
455 | |
456 | return dominatedBySlowTreeWalk(A, B); |
457 | } |
458 | |
459 | bool dominates(const NodeT *A, const NodeT *B) const; |
460 | |
461 | NodeT *getRoot() const { |
462 | assert(this->Roots.size() == 1 && "Should always have entry node!")((void)0); |
463 | return this->Roots[0]; |
464 | } |
465 | |
466 | /// Find nearest common dominator basic block for basic block A and B. A and B |
467 | /// must have tree nodes. |
468 | NodeT *findNearestCommonDominator(NodeT *A, NodeT *B) const { |
469 | assert(A && B && "Pointers are not valid")((void)0); |
470 | assert(A->getParent() == B->getParent() &&((void)0) |
471 | "Two blocks are not in same function")((void)0); |
472 | |
473 | // If either A or B is a entry block then it is nearest common dominator |
474 | // (for forward-dominators). |
475 | if (!isPostDominator()) { |
476 | NodeT &Entry = A->getParent()->front(); |
477 | if (A == &Entry || B == &Entry) |
478 | return &Entry; |
479 | } |
480 | |
481 | DomTreeNodeBase<NodeT> *NodeA = getNode(A); |
482 | DomTreeNodeBase<NodeT> *NodeB = getNode(B); |
483 | assert(NodeA && "A must be in the tree")((void)0); |
484 | assert(NodeB && "B must be in the tree")((void)0); |
485 | |
486 | // Use level information to go up the tree until the levels match. Then |
487 | // continue going up til we arrive at the same node. |
488 | while (NodeA != NodeB) { |
489 | if (NodeA->getLevel() < NodeB->getLevel()) std::swap(NodeA, NodeB); |
490 | |
491 | NodeA = NodeA->IDom; |
492 | } |
493 | |
494 | return NodeA->getBlock(); |
495 | } |
496 | |
497 | const NodeT *findNearestCommonDominator(const NodeT *A, |
498 | const NodeT *B) const { |
499 | // Cast away the const qualifiers here. This is ok since |
500 | // const is re-introduced on the return type. |
501 | return findNearestCommonDominator(const_cast<NodeT *>(A), |
502 | const_cast<NodeT *>(B)); |
503 | } |
504 | |
505 | bool isVirtualRoot(const DomTreeNodeBase<NodeT> *A) const { |
506 | return isPostDominator() && !A->getBlock(); |
507 | } |
508 | |
509 | //===--------------------------------------------------------------------===// |
510 | // API to update (Post)DominatorTree information based on modifications to |
511 | // the CFG... |
512 | |
513 | /// Inform the dominator tree about a sequence of CFG edge insertions and |
514 | /// deletions and perform a batch update on the tree. |
515 | /// |
516 | /// This function should be used when there were multiple CFG updates after |
517 | /// the last dominator tree update. It takes care of performing the updates |
518 | /// in sync with the CFG and optimizes away the redundant operations that |
519 | /// cancel each other. |
520 | /// The functions expects the sequence of updates to be balanced. Eg.: |
521 | /// - {{Insert, A, B}, {Delete, A, B}, {Insert, A, B}} is fine, because |
522 | /// logically it results in a single insertions. |
523 | /// - {{Insert, A, B}, {Insert, A, B}} is invalid, because it doesn't make |
524 | /// sense to insert the same edge twice. |
525 | /// |
526 | /// What's more, the functions assumes that it's safe to ask every node in the |
527 | /// CFG about its children and inverse children. This implies that deletions |
528 | /// of CFG edges must not delete the CFG nodes before calling this function. |
529 | /// |
530 | /// The applyUpdates function can reorder the updates and remove redundant |
531 | /// ones internally. The batch updater is also able to detect sequences of |
532 | /// zero and exactly one update -- it's optimized to do less work in these |
533 | /// cases. |
534 | /// |
535 | /// Note that for postdominators it automatically takes care of applying |
536 | /// updates on reverse edges internally (so there's no need to swap the |
537 | /// From and To pointers when constructing DominatorTree::UpdateType). |
538 | /// The type of updates is the same for DomTreeBase<T> and PostDomTreeBase<T> |
539 | /// with the same template parameter T. |
540 | /// |
541 | /// \param Updates An unordered sequence of updates to perform. The current |
542 | /// CFG and the reverse of these updates provides the pre-view of the CFG. |
543 | /// |
544 | void applyUpdates(ArrayRef<UpdateType> Updates) { |
545 | GraphDiff<NodePtr, IsPostDominator> PreViewCFG( |
546 | Updates, /*ReverseApplyUpdates=*/true); |
547 | DomTreeBuilder::ApplyUpdates(*this, PreViewCFG, nullptr); |
548 | } |
549 | |
550 | /// \param Updates An unordered sequence of updates to perform. The current |
551 | /// CFG and the reverse of these updates provides the pre-view of the CFG. |
552 | /// \param PostViewUpdates An unordered sequence of update to perform in order |
553 | /// to obtain a post-view of the CFG. The DT will be updated assuming the |
554 | /// obtained PostViewCFG is the desired end state. |
555 | void applyUpdates(ArrayRef<UpdateType> Updates, |
556 | ArrayRef<UpdateType> PostViewUpdates) { |
557 | if (Updates.empty()) { |
558 | GraphDiff<NodePtr, IsPostDom> PostViewCFG(PostViewUpdates); |
559 | DomTreeBuilder::ApplyUpdates(*this, PostViewCFG, &PostViewCFG); |
560 | } else { |
561 | // PreViewCFG needs to merge Updates and PostViewCFG. The updates in |
562 | // Updates need to be reversed, and match the direction in PostViewCFG. |
563 | // The PostViewCFG is created with updates reversed (equivalent to changes |
564 | // made to the CFG), so the PreViewCFG needs all the updates reverse |
565 | // applied. |
566 | SmallVector<UpdateType> AllUpdates(Updates.begin(), Updates.end()); |
567 | append_range(AllUpdates, PostViewUpdates); |
568 | GraphDiff<NodePtr, IsPostDom> PreViewCFG(AllUpdates, |
569 | /*ReverseApplyUpdates=*/true); |
570 | GraphDiff<NodePtr, IsPostDom> PostViewCFG(PostViewUpdates); |
571 | DomTreeBuilder::ApplyUpdates(*this, PreViewCFG, &PostViewCFG); |
572 | } |
573 | } |
574 | |
575 | /// Inform the dominator tree about a CFG edge insertion and update the tree. |
576 | /// |
577 | /// This function has to be called just before or just after making the update |
578 | /// on the actual CFG. There cannot be any other updates that the dominator |
579 | /// tree doesn't know about. |
580 | /// |
581 | /// Note that for postdominators it automatically takes care of inserting |
582 | /// a reverse edge internally (so there's no need to swap the parameters). |
583 | /// |
584 | void insertEdge(NodeT *From, NodeT *To) { |
585 | assert(From)((void)0); |
586 | assert(To)((void)0); |
587 | assert(From->getParent() == Parent)((void)0); |
588 | assert(To->getParent() == Parent)((void)0); |
589 | DomTreeBuilder::InsertEdge(*this, From, To); |
590 | } |
591 | |
592 | /// Inform the dominator tree about a CFG edge deletion and update the tree. |
593 | /// |
594 | /// This function has to be called just after making the update on the actual |
595 | /// CFG. An internal functions checks if the edge doesn't exist in the CFG in |
596 | /// DEBUG mode. There cannot be any other updates that the |
597 | /// dominator tree doesn't know about. |
598 | /// |
599 | /// Note that for postdominators it automatically takes care of deleting |
600 | /// a reverse edge internally (so there's no need to swap the parameters). |
601 | /// |
602 | void deleteEdge(NodeT *From, NodeT *To) { |
603 | assert(From)((void)0); |
604 | assert(To)((void)0); |
605 | assert(From->getParent() == Parent)((void)0); |
606 | assert(To->getParent() == Parent)((void)0); |
607 | DomTreeBuilder::DeleteEdge(*this, From, To); |
608 | } |
609 | |
610 | /// Add a new node to the dominator tree information. |
611 | /// |
612 | /// This creates a new node as a child of DomBB dominator node, linking it |
613 | /// into the children list of the immediate dominator. |
614 | /// |
615 | /// \param BB New node in CFG. |
616 | /// \param DomBB CFG node that is dominator for BB. |
617 | /// \returns New dominator tree node that represents new CFG node. |
618 | /// |
619 | DomTreeNodeBase<NodeT> *addNewBlock(NodeT *BB, NodeT *DomBB) { |
620 | assert(getNode(BB) == nullptr && "Block already in dominator tree!")((void)0); |
621 | DomTreeNodeBase<NodeT> *IDomNode = getNode(DomBB); |
622 | assert(IDomNode && "Not immediate dominator specified for block!")((void)0); |
623 | DFSInfoValid = false; |
624 | return createChild(BB, IDomNode); |
625 | } |
626 | |
627 | /// Add a new node to the forward dominator tree and make it a new root. |
628 | /// |
629 | /// \param BB New node in CFG. |
630 | /// \returns New dominator tree node that represents new CFG node. |
631 | /// |
632 | DomTreeNodeBase<NodeT> *setNewRoot(NodeT *BB) { |
633 | assert(getNode(BB) == nullptr && "Block already in dominator tree!")((void)0); |
634 | assert(!this->isPostDominator() &&((void)0) |
635 | "Cannot change root of post-dominator tree")((void)0); |
636 | DFSInfoValid = false; |
637 | DomTreeNodeBase<NodeT> *NewNode = createNode(BB); |
638 | if (Roots.empty()) { |
639 | addRoot(BB); |
640 | } else { |
641 | assert(Roots.size() == 1)((void)0); |
642 | NodeT *OldRoot = Roots.front(); |
643 | auto &OldNode = DomTreeNodes[OldRoot]; |
644 | OldNode = NewNode->addChild(std::move(DomTreeNodes[OldRoot])); |
645 | OldNode->IDom = NewNode; |
646 | OldNode->UpdateLevel(); |
647 | Roots[0] = BB; |
648 | } |
649 | return RootNode = NewNode; |
650 | } |
651 | |
652 | /// changeImmediateDominator - This method is used to update the dominator |
653 | /// tree information when a node's immediate dominator changes. |
654 | /// |
655 | void changeImmediateDominator(DomTreeNodeBase<NodeT> *N, |
656 | DomTreeNodeBase<NodeT> *NewIDom) { |
657 | assert(N && NewIDom && "Cannot change null node pointers!")((void)0); |
658 | DFSInfoValid = false; |
659 | N->setIDom(NewIDom); |
660 | } |
661 | |
662 | void changeImmediateDominator(NodeT *BB, NodeT *NewBB) { |
663 | changeImmediateDominator(getNode(BB), getNode(NewBB)); |
664 | } |
665 | |
666 | /// eraseNode - Removes a node from the dominator tree. Block must not |
667 | /// dominate any other blocks. Removes node from its immediate dominator's |
668 | /// children list. Deletes dominator node associated with basic block BB. |
669 | void eraseNode(NodeT *BB) { |
670 | DomTreeNodeBase<NodeT> *Node = getNode(BB); |
671 | assert(Node && "Removing node that isn't in dominator tree.")((void)0); |
672 | assert(Node->isLeaf() && "Node is not a leaf node.")((void)0); |
673 | |
674 | DFSInfoValid = false; |
675 | |
676 | // Remove node from immediate dominator's children list. |
677 | DomTreeNodeBase<NodeT> *IDom = Node->getIDom(); |
678 | if (IDom) { |
679 | const auto I = find(IDom->Children, Node); |
680 | assert(I != IDom->Children.end() &&((void)0) |
681 | "Not in immediate dominator children set!")((void)0); |
682 | // I am no longer your child... |
683 | IDom->Children.erase(I); |
684 | } |
685 | |
686 | DomTreeNodes.erase(BB); |
687 | |
688 | if (!IsPostDom) return; |
689 | |
690 | // Remember to update PostDominatorTree roots. |
691 | auto RIt = llvm::find(Roots, BB); |
692 | if (RIt != Roots.end()) { |
693 | std::swap(*RIt, Roots.back()); |
694 | Roots.pop_back(); |
695 | } |
696 | } |
697 | |
698 | /// splitBlock - BB is split and now it has one successor. Update dominator |
699 | /// tree to reflect this change. |
700 | void splitBlock(NodeT *NewBB) { |
701 | if (IsPostDominator) |
702 | Split<Inverse<NodeT *>>(NewBB); |
703 | else |
704 | Split<NodeT *>(NewBB); |
705 | } |
706 | |
707 | /// print - Convert to human readable form |
708 | /// |
709 | void print(raw_ostream &O) const { |
710 | O << "=============================--------------------------------\n"; |
711 | if (IsPostDominator) |
712 | O << "Inorder PostDominator Tree: "; |
713 | else |
714 | O << "Inorder Dominator Tree: "; |
715 | if (!DFSInfoValid) |
716 | O << "DFSNumbers invalid: " << SlowQueries << " slow queries."; |
717 | O << "\n"; |
718 | |
719 | // The postdom tree can have a null root if there are no returns. |
720 | if (getRootNode()) PrintDomTree<NodeT>(getRootNode(), O, 1); |
721 | O << "Roots: "; |
722 | for (const NodePtr Block : Roots) { |
723 | Block->printAsOperand(O, false); |
724 | O << " "; |
725 | } |
726 | O << "\n"; |
727 | } |
728 | |
729 | public: |
730 | /// updateDFSNumbers - Assign In and Out numbers to the nodes while walking |
731 | /// dominator tree in dfs order. |
732 | void updateDFSNumbers() const { |
733 | if (DFSInfoValid) { |
734 | SlowQueries = 0; |
735 | return; |
736 | } |
737 | |
738 | SmallVector<std::pair<const DomTreeNodeBase<NodeT> *, |
739 | typename DomTreeNodeBase<NodeT>::const_iterator>, |
740 | 32> WorkStack; |
741 | |
742 | const DomTreeNodeBase<NodeT> *ThisRoot = getRootNode(); |
743 | assert((!Parent || ThisRoot) && "Empty constructed DomTree")((void)0); |
744 | if (!ThisRoot) |
745 | return; |
746 | |
747 | // Both dominators and postdominators have a single root node. In the case |
748 | // case of PostDominatorTree, this node is a virtual root. |
749 | WorkStack.push_back({ThisRoot, ThisRoot->begin()}); |
750 | |
751 | unsigned DFSNum = 0; |
752 | ThisRoot->DFSNumIn = DFSNum++; |
753 | |
754 | while (!WorkStack.empty()) { |
755 | const DomTreeNodeBase<NodeT> *Node = WorkStack.back().first; |
756 | const auto ChildIt = WorkStack.back().second; |
757 | |
758 | // If we visited all of the children of this node, "recurse" back up the |
759 | // stack setting the DFOutNum. |
760 | if (ChildIt == Node->end()) { |
761 | Node->DFSNumOut = DFSNum++; |
762 | WorkStack.pop_back(); |
763 | } else { |
764 | // Otherwise, recursively visit this child. |
765 | const DomTreeNodeBase<NodeT> *Child = *ChildIt; |
766 | ++WorkStack.back().second; |
767 | |
768 | WorkStack.push_back({Child, Child->begin()}); |
769 | Child->DFSNumIn = DFSNum++; |
770 | } |
771 | } |
772 | |
773 | SlowQueries = 0; |
774 | DFSInfoValid = true; |
775 | } |
776 | |
777 | /// recalculate - compute a dominator tree for the given function |
778 | void recalculate(ParentType &Func) { |
779 | Parent = &Func; |
780 | DomTreeBuilder::Calculate(*this); |
781 | } |
782 | |
783 | void recalculate(ParentType &Func, ArrayRef<UpdateType> Updates) { |
784 | Parent = &Func; |
785 | DomTreeBuilder::CalculateWithUpdates(*this, Updates); |
786 | } |
787 | |
788 | /// verify - checks if the tree is correct. There are 3 level of verification: |
789 | /// - Full -- verifies if the tree is correct by making sure all the |
790 | /// properties (including the parent and the sibling property) |
791 | /// hold. |
792 | /// Takes O(N^3) time. |
793 | /// |
794 | /// - Basic -- checks if the tree is correct, but compares it to a freshly |
795 | /// constructed tree instead of checking the sibling property. |
796 | /// Takes O(N^2) time. |
797 | /// |
798 | /// - Fast -- checks basic tree structure and compares it with a freshly |
799 | /// constructed tree. |
800 | /// Takes O(N^2) time worst case, but is faster in practise (same |
801 | /// as tree construction). |
802 | bool verify(VerificationLevel VL = VerificationLevel::Full) const { |
803 | return DomTreeBuilder::Verify(*this, VL); |
804 | } |
805 | |
806 | void reset() { |
807 | DomTreeNodes.clear(); |
808 | Roots.clear(); |
809 | RootNode = nullptr; |
810 | Parent = nullptr; |
811 | DFSInfoValid = false; |
812 | SlowQueries = 0; |
813 | } |
814 | |
815 | protected: |
816 | void addRoot(NodeT *BB) { this->Roots.push_back(BB); } |
817 | |
818 | DomTreeNodeBase<NodeT> *createChild(NodeT *BB, DomTreeNodeBase<NodeT> *IDom) { |
819 | return (DomTreeNodes[BB] = IDom->addChild( |
820 | std::make_unique<DomTreeNodeBase<NodeT>>(BB, IDom))) |
821 | .get(); |
822 | } |
823 | |
824 | DomTreeNodeBase<NodeT> *createNode(NodeT *BB) { |
825 | return (DomTreeNodes[BB] = |
826 | std::make_unique<DomTreeNodeBase<NodeT>>(BB, nullptr)) |
827 | .get(); |
828 | } |
829 | |
830 | // NewBB is split and now it has one successor. Update dominator tree to |
831 | // reflect this change. |
832 | template <class N> |
833 | void Split(typename GraphTraits<N>::NodeRef NewBB) { |
834 | using GraphT = GraphTraits<N>; |
835 | using NodeRef = typename GraphT::NodeRef; |
836 | assert(std::distance(GraphT::child_begin(NewBB),((void)0) |
837 | GraphT::child_end(NewBB)) == 1 &&((void)0) |
838 | "NewBB should have a single successor!")((void)0); |
839 | NodeRef NewBBSucc = *GraphT::child_begin(NewBB); |
840 | |
841 | SmallVector<NodeRef, 4> PredBlocks(children<Inverse<N>>(NewBB)); |
842 | |
843 | assert(!PredBlocks.empty() && "No predblocks?")((void)0); |
844 | |
845 | bool NewBBDominatesNewBBSucc = true; |
846 | for (auto Pred : children<Inverse<N>>(NewBBSucc)) { |
847 | if (Pred != NewBB && !dominates(NewBBSucc, Pred) && |
848 | isReachableFromEntry(Pred)) { |
849 | NewBBDominatesNewBBSucc = false; |
850 | break; |
851 | } |
852 | } |
853 | |
854 | // Find NewBB's immediate dominator and create new dominator tree node for |
855 | // NewBB. |
856 | NodeT *NewBBIDom = nullptr; |
857 | unsigned i = 0; |
858 | for (i = 0; i < PredBlocks.size(); ++i) |
859 | if (isReachableFromEntry(PredBlocks[i])) { |
860 | NewBBIDom = PredBlocks[i]; |
861 | break; |
862 | } |
863 | |
864 | // It's possible that none of the predecessors of NewBB are reachable; |
865 | // in that case, NewBB itself is unreachable, so nothing needs to be |
866 | // changed. |
867 | if (!NewBBIDom) return; |
868 | |
869 | for (i = i + 1; i < PredBlocks.size(); ++i) { |
870 | if (isReachableFromEntry(PredBlocks[i])) |
871 | NewBBIDom = findNearestCommonDominator(NewBBIDom, PredBlocks[i]); |
872 | } |
873 | |
874 | // Create the new dominator tree node... and set the idom of NewBB. |
875 | DomTreeNodeBase<NodeT> *NewBBNode = addNewBlock(NewBB, NewBBIDom); |
876 | |
877 | // If NewBB strictly dominates other blocks, then it is now the immediate |
878 | // dominator of NewBBSucc. Update the dominator tree as appropriate. |
879 | if (NewBBDominatesNewBBSucc) { |
880 | DomTreeNodeBase<NodeT> *NewBBSuccNode = getNode(NewBBSucc); |
881 | changeImmediateDominator(NewBBSuccNode, NewBBNode); |
882 | } |
883 | } |
884 | |
885 | private: |
886 | bool dominatedBySlowTreeWalk(const DomTreeNodeBase<NodeT> *A, |
887 | const DomTreeNodeBase<NodeT> *B) const { |
888 | assert(A != B)((void)0); |
889 | assert(isReachableFromEntry(B))((void)0); |
890 | assert(isReachableFromEntry(A))((void)0); |
891 | |
892 | const unsigned ALevel = A->getLevel(); |
893 | const DomTreeNodeBase<NodeT> *IDom; |
894 | |
895 | // Don't walk nodes above A's subtree. When we reach A's level, we must |
896 | // either find A or be in some other subtree not dominated by A. |
897 | while ((IDom = B->getIDom()) != nullptr && IDom->getLevel() >= ALevel) |
898 | B = IDom; // Walk up the tree |
899 | |
900 | return B == A; |
901 | } |
902 | |
903 | /// Wipe this tree's state without releasing any resources. |
904 | /// |
905 | /// This is essentially a post-move helper only. It leaves the object in an |
906 | /// assignable and destroyable state, but otherwise invalid. |
907 | void wipe() { |
908 | DomTreeNodes.clear(); |
909 | RootNode = nullptr; |
910 | Parent = nullptr; |
911 | } |
912 | }; |
913 | |
914 | template <typename T> |
915 | using DomTreeBase = DominatorTreeBase<T, false>; |
916 | |
917 | template <typename T> |
918 | using PostDomTreeBase = DominatorTreeBase<T, true>; |
919 | |
920 | // These two functions are declared out of line as a workaround for building |
921 | // with old (< r147295) versions of clang because of pr11642. |
922 | template <typename NodeT, bool IsPostDom> |
923 | bool DominatorTreeBase<NodeT, IsPostDom>::dominates(const NodeT *A, |
924 | const NodeT *B) const { |
925 | if (A == B) |
926 | return true; |
927 | |
928 | // Cast away the const qualifiers here. This is ok since |
929 | // this function doesn't actually return the values returned |
930 | // from getNode. |
931 | return dominates(getNode(const_cast<NodeT *>(A)), |
932 | getNode(const_cast<NodeT *>(B))); |
933 | } |
934 | template <typename NodeT, bool IsPostDom> |
935 | bool DominatorTreeBase<NodeT, IsPostDom>::properlyDominates( |
936 | const NodeT *A, const NodeT *B) const { |
937 | if (A == B) |
938 | return false; |
939 | |
940 | // Cast away the const qualifiers here. This is ok since |
941 | // this function doesn't actually return the values returned |
942 | // from getNode. |
943 | return dominates(getNode(const_cast<NodeT *>(A)), |
944 | getNode(const_cast<NodeT *>(B))); |
945 | } |
946 | |
947 | } // end namespace llvm |
948 | |
949 | #endif // LLVM_SUPPORT_GENERICDOMTREE_H |
1 | //===- llvm/ADT/DenseMap.h - Dense probed hash table ------------*- C++ -*-===// |
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 DenseMap class. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #ifndef LLVM_ADT_DENSEMAP_H |
14 | #define LLVM_ADT_DENSEMAP_H |
15 | |
16 | #include "llvm/ADT/DenseMapInfo.h" |
17 | #include "llvm/ADT/EpochTracker.h" |
18 | #include "llvm/Support/AlignOf.h" |
19 | #include "llvm/Support/Compiler.h" |
20 | #include "llvm/Support/MathExtras.h" |
21 | #include "llvm/Support/MemAlloc.h" |
22 | #include "llvm/Support/ReverseIteration.h" |
23 | #include "llvm/Support/type_traits.h" |
24 | #include <algorithm> |
25 | #include <cassert> |
26 | #include <cstddef> |
27 | #include <cstring> |
28 | #include <initializer_list> |
29 | #include <iterator> |
30 | #include <new> |
31 | #include <type_traits> |
32 | #include <utility> |
33 | |
34 | namespace llvm { |
35 | |
36 | namespace detail { |
37 | |
38 | // We extend a pair to allow users to override the bucket type with their own |
39 | // implementation without requiring two members. |
40 | template <typename KeyT, typename ValueT> |
41 | struct DenseMapPair : public std::pair<KeyT, ValueT> { |
42 | using std::pair<KeyT, ValueT>::pair; |
43 | |
44 | KeyT &getFirst() { return std::pair<KeyT, ValueT>::first; } |
45 | const KeyT &getFirst() const { return std::pair<KeyT, ValueT>::first; } |
46 | ValueT &getSecond() { return std::pair<KeyT, ValueT>::second; } |
47 | const ValueT &getSecond() const { return std::pair<KeyT, ValueT>::second; } |
48 | }; |
49 | |
50 | } // end namespace detail |
51 | |
52 | template <typename KeyT, typename ValueT, |
53 | typename KeyInfoT = DenseMapInfo<KeyT>, |
54 | typename Bucket = llvm::detail::DenseMapPair<KeyT, ValueT>, |
55 | bool IsConst = false> |
56 | class DenseMapIterator; |
57 | |
58 | template <typename DerivedT, typename KeyT, typename ValueT, typename KeyInfoT, |
59 | typename BucketT> |
60 | class DenseMapBase : public DebugEpochBase { |
61 | template <typename T> |
62 | using const_arg_type_t = typename const_pointer_or_const_ref<T>::type; |
63 | |
64 | public: |
65 | using size_type = unsigned; |
66 | using key_type = KeyT; |
67 | using mapped_type = ValueT; |
68 | using value_type = BucketT; |
69 | |
70 | using iterator = DenseMapIterator<KeyT, ValueT, KeyInfoT, BucketT>; |
71 | using const_iterator = |
72 | DenseMapIterator<KeyT, ValueT, KeyInfoT, BucketT, true>; |
73 | |
74 | inline iterator begin() { |
75 | // When the map is empty, avoid the overhead of advancing/retreating past |
76 | // empty buckets. |
77 | if (empty()) |
78 | return end(); |
79 | if (shouldReverseIterate<KeyT>()) |
80 | return makeIterator(getBucketsEnd() - 1, getBuckets(), *this); |
81 | return makeIterator(getBuckets(), getBucketsEnd(), *this); |
82 | } |
83 | inline iterator end() { |
84 | return makeIterator(getBucketsEnd(), getBucketsEnd(), *this, true); |
85 | } |
86 | inline const_iterator begin() const { |
87 | if (empty()) |
88 | return end(); |
89 | if (shouldReverseIterate<KeyT>()) |
90 | return makeConstIterator(getBucketsEnd() - 1, getBuckets(), *this); |
91 | return makeConstIterator(getBuckets(), getBucketsEnd(), *this); |
92 | } |
93 | inline const_iterator end() const { |
94 | return makeConstIterator(getBucketsEnd(), getBucketsEnd(), *this, true); |
95 | } |
96 | |
97 | LLVM_NODISCARD[[clang::warn_unused_result]] bool empty() const { |
98 | return getNumEntries() == 0; |
99 | } |
100 | unsigned size() const { return getNumEntries(); } |
101 | |
102 | /// Grow the densemap so that it can contain at least \p NumEntries items |
103 | /// before resizing again. |
104 | void reserve(size_type NumEntries) { |
105 | auto NumBuckets = getMinBucketToReserveForEntries(NumEntries); |
106 | incrementEpoch(); |
107 | if (NumBuckets > getNumBuckets()) |
108 | grow(NumBuckets); |
109 | } |
110 | |
111 | void clear() { |
112 | incrementEpoch(); |
113 | if (getNumEntries() == 0 && getNumTombstones() == 0) return; |
114 | |
115 | // If the capacity of the array is huge, and the # elements used is small, |
116 | // shrink the array. |
117 | if (getNumEntries() * 4 < getNumBuckets() && getNumBuckets() > 64) { |
118 | shrink_and_clear(); |
119 | return; |
120 | } |
121 | |
122 | const KeyT EmptyKey = getEmptyKey(), TombstoneKey = getTombstoneKey(); |
123 | if (std::is_trivially_destructible<ValueT>::value) { |
124 | // Use a simpler loop when values don't need destruction. |
125 | for (BucketT *P = getBuckets(), *E = getBucketsEnd(); P != E; ++P) |
126 | P->getFirst() = EmptyKey; |
127 | } else { |
128 | unsigned NumEntries = getNumEntries(); |
129 | for (BucketT *P = getBuckets(), *E = getBucketsEnd(); P != E; ++P) { |
130 | if (!KeyInfoT::isEqual(P->getFirst(), EmptyKey)) { |
131 | if (!KeyInfoT::isEqual(P->getFirst(), TombstoneKey)) { |
132 | P->getSecond().~ValueT(); |
133 | --NumEntries; |
134 | } |
135 | P->getFirst() = EmptyKey; |
136 | } |
137 | } |
138 | assert(NumEntries == 0 && "Node count imbalance!")((void)0); |
139 | } |
140 | setNumEntries(0); |
141 | setNumTombstones(0); |
142 | } |
143 | |
144 | /// Return 1 if the specified key is in the map, 0 otherwise. |
145 | size_type count(const_arg_type_t<KeyT> Val) const { |
146 | const BucketT *TheBucket; |
147 | return LookupBucketFor(Val, TheBucket) ? 1 : 0; |
148 | } |
149 | |
150 | iterator find(const_arg_type_t<KeyT> Val) { |
151 | BucketT *TheBucket; |
152 | if (LookupBucketFor(Val, TheBucket)) |
153 | return makeIterator(TheBucket, |
154 | shouldReverseIterate<KeyT>() ? getBuckets() |
155 | : getBucketsEnd(), |
156 | *this, true); |
157 | return end(); |
158 | } |
159 | const_iterator find(const_arg_type_t<KeyT> Val) const { |
160 | const BucketT *TheBucket; |
161 | if (LookupBucketFor(Val, TheBucket)) |
162 | return makeConstIterator(TheBucket, |
163 | shouldReverseIterate<KeyT>() ? getBuckets() |
164 | : getBucketsEnd(), |
165 | *this, true); |
166 | return end(); |
167 | } |
168 | |
169 | /// Alternate version of find() which allows a different, and possibly |
170 | /// less expensive, key type. |
171 | /// The DenseMapInfo is responsible for supplying methods |
172 | /// getHashValue(LookupKeyT) and isEqual(LookupKeyT, KeyT) for each key |
173 | /// type used. |
174 | template<class LookupKeyT> |
175 | iterator find_as(const LookupKeyT &Val) { |
176 | BucketT *TheBucket; |
177 | if (LookupBucketFor(Val, TheBucket)) |
178 | return makeIterator(TheBucket, |
179 | shouldReverseIterate<KeyT>() ? getBuckets() |
180 | : getBucketsEnd(), |
181 | *this, true); |
182 | return end(); |
183 | } |
184 | template<class LookupKeyT> |
185 | const_iterator find_as(const LookupKeyT &Val) const { |
186 | const BucketT *TheBucket; |
187 | if (LookupBucketFor(Val, TheBucket)) |
188 | return makeConstIterator(TheBucket, |
189 | shouldReverseIterate<KeyT>() ? getBuckets() |
190 | : getBucketsEnd(), |
191 | *this, true); |
192 | return end(); |
193 | } |
194 | |
195 | /// lookup - Return the entry for the specified key, or a default |
196 | /// constructed value if no such entry exists. |
197 | ValueT lookup(const_arg_type_t<KeyT> Val) const { |
198 | const BucketT *TheBucket; |
199 | if (LookupBucketFor(Val, TheBucket)) |
200 | return TheBucket->getSecond(); |
201 | return ValueT(); |
202 | } |
203 | |
204 | // Inserts key,value pair into the map if the key isn't already in the map. |
205 | // If the key is already in the map, it returns false and doesn't update the |
206 | // value. |
207 | std::pair<iterator, bool> insert(const std::pair<KeyT, ValueT> &KV) { |
208 | return try_emplace(KV.first, KV.second); |
209 | } |
210 | |
211 | // Inserts key,value pair into the map if the key isn't already in the map. |
212 | // If the key is already in the map, it returns false and doesn't update the |
213 | // value. |
214 | std::pair<iterator, bool> insert(std::pair<KeyT, ValueT> &&KV) { |
215 | return try_emplace(std::move(KV.first), std::move(KV.second)); |
216 | } |
217 | |
218 | // Inserts key,value pair into the map if the key isn't already in the map. |
219 | // The value is constructed in-place if the key is not in the map, otherwise |
220 | // it is not moved. |
221 | template <typename... Ts> |
222 | std::pair<iterator, bool> try_emplace(KeyT &&Key, Ts &&... Args) { |
223 | BucketT *TheBucket; |
224 | if (LookupBucketFor(Key, TheBucket)) |
225 | return std::make_pair(makeIterator(TheBucket, |
226 | shouldReverseIterate<KeyT>() |
227 | ? getBuckets() |
228 | : getBucketsEnd(), |
229 | *this, true), |
230 | false); // Already in map. |
231 | |
232 | // Otherwise, insert the new element. |
233 | TheBucket = |
234 | InsertIntoBucket(TheBucket, std::move(Key), std::forward<Ts>(Args)...); |
235 | return std::make_pair(makeIterator(TheBucket, |
236 | shouldReverseIterate<KeyT>() |
237 | ? getBuckets() |
238 | : getBucketsEnd(), |
239 | *this, true), |
240 | true); |
241 | } |
242 | |
243 | // Inserts key,value pair into the map if the key isn't already in the map. |
244 | // The value is constructed in-place if the key is not in the map, otherwise |
245 | // it is not moved. |
246 | template <typename... Ts> |
247 | std::pair<iterator, bool> try_emplace(const KeyT &Key, Ts &&... Args) { |
248 | BucketT *TheBucket; |
249 | if (LookupBucketFor(Key, TheBucket)) |
250 | return std::make_pair(makeIterator(TheBucket, |
251 | shouldReverseIterate<KeyT>() |
252 | ? getBuckets() |
253 | : getBucketsEnd(), |
254 | *this, true), |
255 | false); // Already in map. |
256 | |
257 | // Otherwise, insert the new element. |
258 | TheBucket = InsertIntoBucket(TheBucket, Key, std::forward<Ts>(Args)...); |
259 | return std::make_pair(makeIterator(TheBucket, |
260 | shouldReverseIterate<KeyT>() |
261 | ? getBuckets() |
262 | : getBucketsEnd(), |
263 | *this, true), |
264 | true); |
265 | } |
266 | |
267 | /// Alternate version of insert() which allows a different, and possibly |
268 | /// less expensive, key type. |
269 | /// The DenseMapInfo is responsible for supplying methods |
270 | /// getHashValue(LookupKeyT) and isEqual(LookupKeyT, KeyT) for each key |
271 | /// type used. |
272 | template <typename LookupKeyT> |
273 | std::pair<iterator, bool> insert_as(std::pair<KeyT, ValueT> &&KV, |
274 | const LookupKeyT &Val) { |
275 | BucketT *TheBucket; |
276 | if (LookupBucketFor(Val, TheBucket)) |
277 | return std::make_pair(makeIterator(TheBucket, |
278 | shouldReverseIterate<KeyT>() |
279 | ? getBuckets() |
280 | : getBucketsEnd(), |
281 | *this, true), |
282 | false); // Already in map. |
283 | |
284 | // Otherwise, insert the new element. |
285 | TheBucket = InsertIntoBucketWithLookup(TheBucket, std::move(KV.first), |
286 | std::move(KV.second), Val); |
287 | return std::make_pair(makeIterator(TheBucket, |
288 | shouldReverseIterate<KeyT>() |
289 | ? getBuckets() |
290 | : getBucketsEnd(), |
291 | *this, true), |
292 | true); |
293 | } |
294 | |
295 | /// insert - Range insertion of pairs. |
296 | template<typename InputIt> |
297 | void insert(InputIt I, InputIt E) { |
298 | for (; I != E; ++I) |
299 | insert(*I); |
300 | } |
301 | |
302 | bool erase(const KeyT &Val) { |
303 | BucketT *TheBucket; |
304 | if (!LookupBucketFor(Val, TheBucket)) |
305 | return false; // not in map. |
306 | |
307 | TheBucket->getSecond().~ValueT(); |
308 | TheBucket->getFirst() = getTombstoneKey(); |
309 | decrementNumEntries(); |
310 | incrementNumTombstones(); |
311 | return true; |
312 | } |
313 | void erase(iterator I) { |
314 | BucketT *TheBucket = &*I; |
315 | TheBucket->getSecond().~ValueT(); |
316 | TheBucket->getFirst() = getTombstoneKey(); |
317 | decrementNumEntries(); |
318 | incrementNumTombstones(); |
319 | } |
320 | |
321 | value_type& FindAndConstruct(const KeyT &Key) { |
322 | BucketT *TheBucket; |
323 | if (LookupBucketFor(Key, TheBucket)) |
324 | return *TheBucket; |
325 | |
326 | return *InsertIntoBucket(TheBucket, Key); |
327 | } |
328 | |
329 | ValueT &operator[](const KeyT &Key) { |
330 | return FindAndConstruct(Key).second; |
331 | } |
332 | |
333 | value_type& FindAndConstruct(KeyT &&Key) { |
334 | BucketT *TheBucket; |
335 | if (LookupBucketFor(Key, TheBucket)) |
336 | return *TheBucket; |
337 | |
338 | return *InsertIntoBucket(TheBucket, std::move(Key)); |
339 | } |
340 | |
341 | ValueT &operator[](KeyT &&Key) { |
342 | return FindAndConstruct(std::move(Key)).second; |
343 | } |
344 | |
345 | /// isPointerIntoBucketsArray - Return true if the specified pointer points |
346 | /// somewhere into the DenseMap's array of buckets (i.e. either to a key or |
347 | /// value in the DenseMap). |
348 | bool isPointerIntoBucketsArray(const void *Ptr) const { |
349 | return Ptr >= getBuckets() && Ptr < getBucketsEnd(); |
350 | } |
351 | |
352 | /// getPointerIntoBucketsArray() - Return an opaque pointer into the buckets |
353 | /// array. In conjunction with the previous method, this can be used to |
354 | /// determine whether an insertion caused the DenseMap to reallocate. |
355 | const void *getPointerIntoBucketsArray() const { return getBuckets(); } |
356 | |
357 | protected: |
358 | DenseMapBase() = default; |
359 | |
360 | void destroyAll() { |
361 | if (getNumBuckets() == 0) // Nothing to do. |
362 | return; |
363 | |
364 | const KeyT EmptyKey = getEmptyKey(), TombstoneKey = getTombstoneKey(); |
365 | for (BucketT *P = getBuckets(), *E = getBucketsEnd(); P != E; ++P) { |
366 | if (!KeyInfoT::isEqual(P->getFirst(), EmptyKey) && |
367 | !KeyInfoT::isEqual(P->getFirst(), TombstoneKey)) |
368 | P->getSecond().~ValueT(); |
369 | P->getFirst().~KeyT(); |
370 | } |
371 | } |
372 | |
373 | void initEmpty() { |
374 | setNumEntries(0); |
375 | setNumTombstones(0); |
376 | |
377 | assert((getNumBuckets() & (getNumBuckets()-1)) == 0 &&((void)0) |
378 | "# initial buckets must be a power of two!")((void)0); |
379 | const KeyT EmptyKey = getEmptyKey(); |
380 | for (BucketT *B = getBuckets(), *E = getBucketsEnd(); B != E; ++B) |
381 | ::new (&B->getFirst()) KeyT(EmptyKey); |
382 | } |
383 | |
384 | /// Returns the number of buckets to allocate to ensure that the DenseMap can |
385 | /// accommodate \p NumEntries without need to grow(). |
386 | unsigned getMinBucketToReserveForEntries(unsigned NumEntries) { |
387 | // Ensure that "NumEntries * 4 < NumBuckets * 3" |
388 | if (NumEntries == 0) |
389 | return 0; |
390 | // +1 is required because of the strict equality. |
391 | // For example if NumEntries is 48, we need to return 401. |
392 | return NextPowerOf2(NumEntries * 4 / 3 + 1); |
393 | } |
394 | |
395 | void moveFromOldBuckets(BucketT *OldBucketsBegin, BucketT *OldBucketsEnd) { |
396 | initEmpty(); |
397 | |
398 | // Insert all the old elements. |
399 | const KeyT EmptyKey = getEmptyKey(); |
400 | const KeyT TombstoneKey = getTombstoneKey(); |
401 | for (BucketT *B = OldBucketsBegin, *E = OldBucketsEnd; B != E; ++B) { |
402 | if (!KeyInfoT::isEqual(B->getFirst(), EmptyKey) && |
403 | !KeyInfoT::isEqual(B->getFirst(), TombstoneKey)) { |
404 | // Insert the key/value into the new table. |
405 | BucketT *DestBucket; |
406 | bool FoundVal = LookupBucketFor(B->getFirst(), DestBucket); |
407 | (void)FoundVal; // silence warning. |
408 | assert(!FoundVal && "Key already in new map?")((void)0); |
409 | DestBucket->getFirst() = std::move(B->getFirst()); |
410 | ::new (&DestBucket->getSecond()) ValueT(std::move(B->getSecond())); |
411 | incrementNumEntries(); |
412 | |
413 | // Free the value. |
414 | B->getSecond().~ValueT(); |
415 | } |
416 | B->getFirst().~KeyT(); |
417 | } |
418 | } |
419 | |
420 | template <typename OtherBaseT> |
421 | void copyFrom( |
422 | const DenseMapBase<OtherBaseT, KeyT, ValueT, KeyInfoT, BucketT> &other) { |
423 | assert(&other != this)((void)0); |
424 | assert(getNumBuckets() == other.getNumBuckets())((void)0); |
425 | |
426 | setNumEntries(other.getNumEntries()); |
427 | setNumTombstones(other.getNumTombstones()); |
428 | |
429 | if (std::is_trivially_copyable<KeyT>::value && |
430 | std::is_trivially_copyable<ValueT>::value) |
431 | memcpy(reinterpret_cast<void *>(getBuckets()), other.getBuckets(), |
432 | getNumBuckets() * sizeof(BucketT)); |
433 | else |
434 | for (size_t i = 0; i < getNumBuckets(); ++i) { |
435 | ::new (&getBuckets()[i].getFirst()) |
436 | KeyT(other.getBuckets()[i].getFirst()); |
437 | if (!KeyInfoT::isEqual(getBuckets()[i].getFirst(), getEmptyKey()) && |
438 | !KeyInfoT::isEqual(getBuckets()[i].getFirst(), getTombstoneKey())) |
439 | ::new (&getBuckets()[i].getSecond()) |
440 | ValueT(other.getBuckets()[i].getSecond()); |
441 | } |
442 | } |
443 | |
444 | static unsigned getHashValue(const KeyT &Val) { |
445 | return KeyInfoT::getHashValue(Val); |
446 | } |
447 | |
448 | template<typename LookupKeyT> |
449 | static unsigned getHashValue(const LookupKeyT &Val) { |
450 | return KeyInfoT::getHashValue(Val); |
451 | } |
452 | |
453 | static const KeyT getEmptyKey() { |
454 | static_assert(std::is_base_of<DenseMapBase, DerivedT>::value, |
455 | "Must pass the derived type to this template!"); |
456 | return KeyInfoT::getEmptyKey(); |
457 | } |
458 | |
459 | static const KeyT getTombstoneKey() { |
460 | return KeyInfoT::getTombstoneKey(); |
461 | } |
462 | |
463 | private: |
464 | iterator makeIterator(BucketT *P, BucketT *E, |
465 | DebugEpochBase &Epoch, |
466 | bool NoAdvance=false) { |
467 | if (shouldReverseIterate<KeyT>()) { |
468 | BucketT *B = P == getBucketsEnd() ? getBuckets() : P + 1; |
469 | return iterator(B, E, Epoch, NoAdvance); |
470 | } |
471 | return iterator(P, E, Epoch, NoAdvance); |
472 | } |
473 | |
474 | const_iterator makeConstIterator(const BucketT *P, const BucketT *E, |
475 | const DebugEpochBase &Epoch, |
476 | const bool NoAdvance=false) const { |
477 | if (shouldReverseIterate<KeyT>()) { |
478 | const BucketT *B = P == getBucketsEnd() ? getBuckets() : P + 1; |
479 | return const_iterator(B, E, Epoch, NoAdvance); |
480 | } |
481 | return const_iterator(P, E, Epoch, NoAdvance); |
482 | } |
483 | |
484 | unsigned getNumEntries() const { |
485 | return static_cast<const DerivedT *>(this)->getNumEntries(); |
486 | } |
487 | |
488 | void setNumEntries(unsigned Num) { |
489 | static_cast<DerivedT *>(this)->setNumEntries(Num); |
490 | } |
491 | |
492 | void incrementNumEntries() { |
493 | setNumEntries(getNumEntries() + 1); |
494 | } |
495 | |
496 | void decrementNumEntries() { |
497 | setNumEntries(getNumEntries() - 1); |
498 | } |
499 | |
500 | unsigned getNumTombstones() const { |
501 | return static_cast<const DerivedT *>(this)->getNumTombstones(); |
502 | } |
503 | |
504 | void setNumTombstones(unsigned Num) { |
505 | static_cast<DerivedT *>(this)->setNumTombstones(Num); |
506 | } |
507 | |
508 | void incrementNumTombstones() { |
509 | setNumTombstones(getNumTombstones() + 1); |
510 | } |
511 | |
512 | void decrementNumTombstones() { |
513 | setNumTombstones(getNumTombstones() - 1); |
514 | } |
515 | |
516 | const BucketT *getBuckets() const { |
517 | return static_cast<const DerivedT *>(this)->getBuckets(); |
518 | } |
519 | |
520 | BucketT *getBuckets() { |
521 | return static_cast<DerivedT *>(this)->getBuckets(); |
522 | } |
523 | |
524 | unsigned getNumBuckets() const { |
525 | return static_cast<const DerivedT *>(this)->getNumBuckets(); |
526 | } |
527 | |
528 | BucketT *getBucketsEnd() { |
529 | return getBuckets() + getNumBuckets(); |
530 | } |
531 | |
532 | const BucketT *getBucketsEnd() const { |
533 | return getBuckets() + getNumBuckets(); |
534 | } |
535 | |
536 | void grow(unsigned AtLeast) { |
537 | static_cast<DerivedT *>(this)->grow(AtLeast); |
538 | } |
539 | |
540 | void shrink_and_clear() { |
541 | static_cast<DerivedT *>(this)->shrink_and_clear(); |
542 | } |
543 | |
544 | template <typename KeyArg, typename... ValueArgs> |
545 | BucketT *InsertIntoBucket(BucketT *TheBucket, KeyArg &&Key, |
546 | ValueArgs &&... Values) { |
547 | TheBucket = InsertIntoBucketImpl(Key, Key, TheBucket); |
548 | |
549 | TheBucket->getFirst() = std::forward<KeyArg>(Key); |
550 | ::new (&TheBucket->getSecond()) ValueT(std::forward<ValueArgs>(Values)...); |
551 | return TheBucket; |
552 | } |
553 | |
554 | template <typename LookupKeyT> |
555 | BucketT *InsertIntoBucketWithLookup(BucketT *TheBucket, KeyT &&Key, |
556 | ValueT &&Value, LookupKeyT &Lookup) { |
557 | TheBucket = InsertIntoBucketImpl(Key, Lookup, TheBucket); |
558 | |
559 | TheBucket->getFirst() = std::move(Key); |
560 | ::new (&TheBucket->getSecond()) ValueT(std::move(Value)); |
561 | return TheBucket; |
562 | } |
563 | |
564 | template <typename LookupKeyT> |
565 | BucketT *InsertIntoBucketImpl(const KeyT &Key, const LookupKeyT &Lookup, |
566 | BucketT *TheBucket) { |
567 | incrementEpoch(); |
568 | |
569 | // If the load of the hash table is more than 3/4, or if fewer than 1/8 of |
570 | // the buckets are empty (meaning that many are filled with tombstones), |
571 | // grow the table. |
572 | // |
573 | // The later case is tricky. For example, if we had one empty bucket with |
574 | // tons of tombstones, failing lookups (e.g. for insertion) would have to |
575 | // probe almost the entire table until it found the empty bucket. If the |
576 | // table completely filled with tombstones, no lookup would ever succeed, |
577 | // causing infinite loops in lookup. |
578 | unsigned NewNumEntries = getNumEntries() + 1; |
579 | unsigned NumBuckets = getNumBuckets(); |
580 | if (LLVM_UNLIKELY(NewNumEntries * 4 >= NumBuckets * 3)__builtin_expect((bool)(NewNumEntries * 4 >= NumBuckets * 3 ), false)) { |
581 | this->grow(NumBuckets * 2); |
582 | LookupBucketFor(Lookup, TheBucket); |
583 | NumBuckets = getNumBuckets(); |
584 | } else if (LLVM_UNLIKELY(NumBuckets-(NewNumEntries+getNumTombstones()) <=__builtin_expect((bool)(NumBuckets-(NewNumEntries+getNumTombstones ()) <= NumBuckets/8), false) |
585 | NumBuckets/8)__builtin_expect((bool)(NumBuckets-(NewNumEntries+getNumTombstones ()) <= NumBuckets/8), false)) { |
586 | this->grow(NumBuckets); |
587 | LookupBucketFor(Lookup, TheBucket); |
588 | } |
589 | assert(TheBucket)((void)0); |
590 | |
591 | // Only update the state after we've grown our bucket space appropriately |
592 | // so that when growing buckets we have self-consistent entry count. |
593 | incrementNumEntries(); |
594 | |
595 | // If we are writing over a tombstone, remember this. |
596 | const KeyT EmptyKey = getEmptyKey(); |
597 | if (!KeyInfoT::isEqual(TheBucket->getFirst(), EmptyKey)) |
598 | decrementNumTombstones(); |
599 | |
600 | return TheBucket; |
601 | } |
602 | |
603 | /// LookupBucketFor - Lookup the appropriate bucket for Val, returning it in |
604 | /// FoundBucket. If the bucket contains the key and a value, this returns |
605 | /// true, otherwise it returns a bucket with an empty marker or tombstone and |
606 | /// returns false. |
607 | template<typename LookupKeyT> |
608 | bool LookupBucketFor(const LookupKeyT &Val, |
609 | const BucketT *&FoundBucket) const { |
610 | const BucketT *BucketsPtr = getBuckets(); |
611 | const unsigned NumBuckets = getNumBuckets(); |
612 | |
613 | if (NumBuckets == 0) { |
614 | FoundBucket = nullptr; |
615 | return false; |
616 | } |
617 | |
618 | // FoundTombstone - Keep track of whether we find a tombstone while probing. |
619 | const BucketT *FoundTombstone = nullptr; |
620 | const KeyT EmptyKey = getEmptyKey(); |
621 | const KeyT TombstoneKey = getTombstoneKey(); |
622 | assert(!KeyInfoT::isEqual(Val, EmptyKey) &&((void)0) |
623 | !KeyInfoT::isEqual(Val, TombstoneKey) &&((void)0) |
624 | "Empty/Tombstone value shouldn't be inserted into map!")((void)0); |
625 | |
626 | unsigned BucketNo = getHashValue(Val) & (NumBuckets-1); |
627 | unsigned ProbeAmt = 1; |
628 | while (true) { |
629 | const BucketT *ThisBucket = BucketsPtr + BucketNo; |
630 | // Found Val's bucket? If so, return it. |
631 | if (LLVM_LIKELY(KeyInfoT::isEqual(Val, ThisBucket->getFirst()))__builtin_expect((bool)(KeyInfoT::isEqual(Val, ThisBucket-> getFirst())), true)) { |
632 | FoundBucket = ThisBucket; |
633 | return true; |
634 | } |
635 | |
636 | // If we found an empty bucket, the key doesn't exist in the set. |
637 | // Insert it and return the default value. |
638 | if (LLVM_LIKELY(KeyInfoT::isEqual(ThisBucket->getFirst(), EmptyKey))__builtin_expect((bool)(KeyInfoT::isEqual(ThisBucket->getFirst (), EmptyKey)), true)) { |
639 | // If we've already seen a tombstone while probing, fill it in instead |
640 | // of the empty bucket we eventually probed to. |
641 | FoundBucket = FoundTombstone ? FoundTombstone : ThisBucket; |
642 | return false; |
643 | } |
644 | |
645 | // If this is a tombstone, remember it. If Val ends up not in the map, we |
646 | // prefer to return it than something that would require more probing. |
647 | if (KeyInfoT::isEqual(ThisBucket->getFirst(), TombstoneKey) && |
648 | !FoundTombstone) |
649 | FoundTombstone = ThisBucket; // Remember the first tombstone found. |
650 | |
651 | // Otherwise, it's a hash collision or a tombstone, continue quadratic |
652 | // probing. |
653 | BucketNo += ProbeAmt++; |
654 | BucketNo &= (NumBuckets-1); |
655 | } |
656 | } |
657 | |
658 | template <typename LookupKeyT> |
659 | bool LookupBucketFor(const LookupKeyT &Val, BucketT *&FoundBucket) { |
660 | const BucketT *ConstFoundBucket; |
661 | bool Result = const_cast<const DenseMapBase *>(this) |
662 | ->LookupBucketFor(Val, ConstFoundBucket); |
663 | FoundBucket = const_cast<BucketT *>(ConstFoundBucket); |
664 | return Result; |
665 | } |
666 | |
667 | public: |
668 | /// Return the approximate size (in bytes) of the actual map. |
669 | /// This is just the raw memory used by DenseMap. |
670 | /// If entries are pointers to objects, the size of the referenced objects |
671 | /// are not included. |
672 | size_t getMemorySize() const { |
673 | return getNumBuckets() * sizeof(BucketT); |
674 | } |
675 | }; |
676 | |
677 | /// Equality comparison for DenseMap. |
678 | /// |
679 | /// Iterates over elements of LHS confirming that each (key, value) pair in LHS |
680 | /// is also in RHS, and that no additional pairs are in RHS. |
681 | /// Equivalent to N calls to RHS.find and N value comparisons. Amortized |
682 | /// complexity is linear, worst case is O(N^2) (if every hash collides). |
683 | template <typename DerivedT, typename KeyT, typename ValueT, typename KeyInfoT, |
684 | typename BucketT> |
685 | bool operator==( |
686 | const DenseMapBase<DerivedT, KeyT, ValueT, KeyInfoT, BucketT> &LHS, |
687 | const DenseMapBase<DerivedT, KeyT, ValueT, KeyInfoT, BucketT> &RHS) { |
688 | if (LHS.size() != RHS.size()) |
689 | return false; |
690 | |
691 | for (auto &KV : LHS) { |
692 | auto I = RHS.find(KV.first); |
693 | if (I == RHS.end() || I->second != KV.second) |
694 | return false; |
695 | } |
696 | |
697 | return true; |
698 | } |
699 | |
700 | /// Inequality comparison for DenseMap. |
701 | /// |
702 | /// Equivalent to !(LHS == RHS). See operator== for performance notes. |
703 | template <typename DerivedT, typename KeyT, typename ValueT, typename KeyInfoT, |
704 | typename BucketT> |
705 | bool operator!=( |
706 | const DenseMapBase<DerivedT, KeyT, ValueT, KeyInfoT, BucketT> &LHS, |
707 | const DenseMapBase<DerivedT, KeyT, ValueT, KeyInfoT, BucketT> &RHS) { |
708 | return !(LHS == RHS); |
709 | } |
710 | |
711 | template <typename KeyT, typename ValueT, |
712 | typename KeyInfoT = DenseMapInfo<KeyT>, |
713 | typename BucketT = llvm::detail::DenseMapPair<KeyT, ValueT>> |
714 | class DenseMap : public DenseMapBase<DenseMap<KeyT, ValueT, KeyInfoT, BucketT>, |
715 | KeyT, ValueT, KeyInfoT, BucketT> { |
716 | friend class DenseMapBase<DenseMap, KeyT, ValueT, KeyInfoT, BucketT>; |
717 | |
718 | // Lift some types from the dependent base class into this class for |
719 | // simplicity of referring to them. |
720 | using BaseT = DenseMapBase<DenseMap, KeyT, ValueT, KeyInfoT, BucketT>; |
721 | |
722 | BucketT *Buckets; |
723 | unsigned NumEntries; |
724 | unsigned NumTombstones; |
725 | unsigned NumBuckets; |
726 | |
727 | public: |
728 | /// Create a DenseMap with an optional \p InitialReserve that guarantee that |
729 | /// this number of elements can be inserted in the map without grow() |
730 | explicit DenseMap(unsigned InitialReserve = 0) { init(InitialReserve); } |
731 | |
732 | DenseMap(const DenseMap &other) : BaseT() { |
733 | init(0); |
734 | copyFrom(other); |
735 | } |
736 | |
737 | DenseMap(DenseMap &&other) : BaseT() { |
738 | init(0); |
739 | swap(other); |
740 | } |
741 | |
742 | template<typename InputIt> |
743 | DenseMap(const InputIt &I, const InputIt &E) { |
744 | init(std::distance(I, E)); |
745 | this->insert(I, E); |
746 | } |
747 | |
748 | DenseMap(std::initializer_list<typename BaseT::value_type> Vals) { |
749 | init(Vals.size()); |
750 | this->insert(Vals.begin(), Vals.end()); |
751 | } |
752 | |
753 | ~DenseMap() { |
754 | this->destroyAll(); |
755 | deallocate_buffer(Buckets, sizeof(BucketT) * NumBuckets, alignof(BucketT)); |
756 | } |
757 | |
758 | void swap(DenseMap& RHS) { |
759 | this->incrementEpoch(); |
760 | RHS.incrementEpoch(); |
761 | std::swap(Buckets, RHS.Buckets); |
762 | std::swap(NumEntries, RHS.NumEntries); |
763 | std::swap(NumTombstones, RHS.NumTombstones); |
764 | std::swap(NumBuckets, RHS.NumBuckets); |
765 | } |
766 | |
767 | DenseMap& operator=(const DenseMap& other) { |
768 | if (&other != this) |
769 | copyFrom(other); |
770 | return *this; |
771 | } |
772 | |
773 | DenseMap& operator=(DenseMap &&other) { |
774 | this->destroyAll(); |
775 | deallocate_buffer(Buckets, sizeof(BucketT) * NumBuckets, alignof(BucketT)); |
776 | init(0); |
777 | swap(other); |
778 | return *this; |
779 | } |
780 | |
781 | void copyFrom(const DenseMap& other) { |
782 | this->destroyAll(); |
783 | deallocate_buffer(Buckets, sizeof(BucketT) * NumBuckets, alignof(BucketT)); |
784 | if (allocateBuckets(other.NumBuckets)) { |
785 | this->BaseT::copyFrom(other); |
786 | } else { |
787 | NumEntries = 0; |
788 | NumTombstones = 0; |
789 | } |
790 | } |
791 | |
792 | void init(unsigned InitNumEntries) { |
793 | auto InitBuckets = BaseT::getMinBucketToReserveForEntries(InitNumEntries); |
794 | if (allocateBuckets(InitBuckets)) { |
795 | this->BaseT::initEmpty(); |
796 | } else { |
797 | NumEntries = 0; |
798 | NumTombstones = 0; |
799 | } |
800 | } |
801 | |
802 | void grow(unsigned AtLeast) { |
803 | unsigned OldNumBuckets = NumBuckets; |
804 | BucketT *OldBuckets = Buckets; |
805 | |
806 | allocateBuckets(std::max<unsigned>(64, static_cast<unsigned>(NextPowerOf2(AtLeast-1)))); |
807 | assert(Buckets)((void)0); |
808 | if (!OldBuckets) { |
809 | this->BaseT::initEmpty(); |
810 | return; |
811 | } |
812 | |
813 | this->moveFromOldBuckets(OldBuckets, OldBuckets+OldNumBuckets); |
814 | |
815 | // Free the old table. |
816 | deallocate_buffer(OldBuckets, sizeof(BucketT) * OldNumBuckets, |
817 | alignof(BucketT)); |
818 | } |
819 | |
820 | void shrink_and_clear() { |
821 | unsigned OldNumBuckets = NumBuckets; |
822 | unsigned OldNumEntries = NumEntries; |
823 | this->destroyAll(); |
824 | |
825 | // Reduce the number of buckets. |
826 | unsigned NewNumBuckets = 0; |
827 | if (OldNumEntries) |
828 | NewNumBuckets = std::max(64, 1 << (Log2_32_Ceil(OldNumEntries) + 1)); |
829 | if (NewNumBuckets == NumBuckets) { |
830 | this->BaseT::initEmpty(); |
831 | return; |
832 | } |
833 | |
834 | deallocate_buffer(Buckets, sizeof(BucketT) * OldNumBuckets, |
835 | alignof(BucketT)); |
836 | init(NewNumBuckets); |
837 | } |
838 | |
839 | private: |
840 | unsigned getNumEntries() const { |
841 | return NumEntries; |
842 | } |
843 | |
844 | void setNumEntries(unsigned Num) { |
845 | NumEntries = Num; |
846 | } |
847 | |
848 | unsigned getNumTombstones() const { |
849 | return NumTombstones; |
850 | } |
851 | |
852 | void setNumTombstones(unsigned Num) { |
853 | NumTombstones = Num; |
854 | } |
855 | |
856 | BucketT *getBuckets() const { |
857 | return Buckets; |
858 | } |
859 | |
860 | unsigned getNumBuckets() const { |
861 | return NumBuckets; |
862 | } |
863 | |
864 | bool allocateBuckets(unsigned Num) { |
865 | NumBuckets = Num; |
866 | if (NumBuckets == 0) { |
867 | Buckets = nullptr; |
868 | return false; |
869 | } |
870 | |
871 | Buckets = static_cast<BucketT *>( |
872 | allocate_buffer(sizeof(BucketT) * NumBuckets, alignof(BucketT))); |
873 | return true; |
874 | } |
875 | }; |
876 | |
877 | template <typename KeyT, typename ValueT, unsigned InlineBuckets = 4, |
878 | typename KeyInfoT = DenseMapInfo<KeyT>, |
879 | typename BucketT = llvm::detail::DenseMapPair<KeyT, ValueT>> |
880 | class SmallDenseMap |
881 | : public DenseMapBase< |
882 | SmallDenseMap<KeyT, ValueT, InlineBuckets, KeyInfoT, BucketT>, KeyT, |
883 | ValueT, KeyInfoT, BucketT> { |
884 | friend class DenseMapBase<SmallDenseMap, KeyT, ValueT, KeyInfoT, BucketT>; |
885 | |
886 | // Lift some types from the dependent base class into this class for |
887 | // simplicity of referring to them. |
888 | using BaseT = DenseMapBase<SmallDenseMap, KeyT, ValueT, KeyInfoT, BucketT>; |
889 | |
890 | static_assert(isPowerOf2_64(InlineBuckets), |
891 | "InlineBuckets must be a power of 2."); |
892 | |
893 | unsigned Small : 1; |
894 | unsigned NumEntries : 31; |
895 | unsigned NumTombstones; |
896 | |
897 | struct LargeRep { |
898 | BucketT *Buckets; |
899 | unsigned NumBuckets; |
900 | }; |
901 | |
902 | /// A "union" of an inline bucket array and the struct representing |
903 | /// a large bucket. This union will be discriminated by the 'Small' bit. |
904 | AlignedCharArrayUnion<BucketT[InlineBuckets], LargeRep> storage; |
905 | |
906 | public: |
907 | explicit SmallDenseMap(unsigned NumInitBuckets = 0) { |
908 | init(NumInitBuckets); |
909 | } |
910 | |
911 | SmallDenseMap(const SmallDenseMap &other) : BaseT() { |
912 | init(0); |
913 | copyFrom(other); |
914 | } |
915 | |
916 | SmallDenseMap(SmallDenseMap &&other) : BaseT() { |
917 | init(0); |
918 | swap(other); |
919 | } |
920 | |
921 | template<typename InputIt> |
922 | SmallDenseMap(const InputIt &I, const InputIt &E) { |
923 | init(NextPowerOf2(std::distance(I, E))); |
924 | this->insert(I, E); |
925 | } |
926 | |
927 | SmallDenseMap(std::initializer_list<typename BaseT::value_type> Vals) |
928 | : SmallDenseMap(Vals.begin(), Vals.end()) {} |
929 | |
930 | ~SmallDenseMap() { |
931 | this->destroyAll(); |
932 | deallocateBuckets(); |
933 | } |
934 | |
935 | void swap(SmallDenseMap& RHS) { |
936 | unsigned TmpNumEntries = RHS.NumEntries; |
937 | RHS.NumEntries = NumEntries; |
938 | NumEntries = TmpNumEntries; |
939 | std::swap(NumTombstones, RHS.NumTombstones); |
940 | |
941 | const KeyT EmptyKey = this->getEmptyKey(); |
942 | const KeyT TombstoneKey = this->getTombstoneKey(); |
943 | if (Small && RHS.Small) { |
944 | // If we're swapping inline bucket arrays, we have to cope with some of |
945 | // the tricky bits of DenseMap's storage system: the buckets are not |
946 | // fully initialized. Thus we swap every key, but we may have |
947 | // a one-directional move of the value. |
948 | for (unsigned i = 0, e = InlineBuckets; i != e; ++i) { |
949 | BucketT *LHSB = &getInlineBuckets()[i], |
950 | *RHSB = &RHS.getInlineBuckets()[i]; |
951 | bool hasLHSValue = (!KeyInfoT::isEqual(LHSB->getFirst(), EmptyKey) && |
952 | !KeyInfoT::isEqual(LHSB->getFirst(), TombstoneKey)); |
953 | bool hasRHSValue = (!KeyInfoT::isEqual(RHSB->getFirst(), EmptyKey) && |
954 | !KeyInfoT::isEqual(RHSB->getFirst(), TombstoneKey)); |
955 | if (hasLHSValue && hasRHSValue) { |
956 | // Swap together if we can... |
957 | std::swap(*LHSB, *RHSB); |
958 | continue; |
959 | } |
960 | // Swap separately and handle any asymmetry. |
961 | std::swap(LHSB->getFirst(), RHSB->getFirst()); |
962 | if (hasLHSValue) { |
963 | ::new (&RHSB->getSecond()) ValueT(std::move(LHSB->getSecond())); |
964 | LHSB->getSecond().~ValueT(); |
965 | } else if (hasRHSValue) { |
966 | ::new (&LHSB->getSecond()) ValueT(std::move(RHSB->getSecond())); |
967 | RHSB->getSecond().~ValueT(); |
968 | } |
969 | } |
970 | return; |
971 | } |
972 | if (!Small && !RHS.Small) { |
973 | std::swap(getLargeRep()->Buckets, RHS.getLargeRep()->Buckets); |
974 | std::swap(getLargeRep()->NumBuckets, RHS.getLargeRep()->NumBuckets); |
975 | return; |
976 | } |
977 | |
978 | SmallDenseMap &SmallSide = Small ? *this : RHS; |
979 | SmallDenseMap &LargeSide = Small ? RHS : *this; |
980 | |
981 | // First stash the large side's rep and move the small side across. |
982 | LargeRep TmpRep = std::move(*LargeSide.getLargeRep()); |
983 | LargeSide.getLargeRep()->~LargeRep(); |
984 | LargeSide.Small = true; |
985 | // This is similar to the standard move-from-old-buckets, but the bucket |
986 | // count hasn't actually rotated in this case. So we have to carefully |
987 | // move construct the keys and values into their new locations, but there |
988 | // is no need to re-hash things. |
989 | for (unsigned i = 0, e = InlineBuckets; i != e; ++i) { |
990 | BucketT *NewB = &LargeSide.getInlineBuckets()[i], |
991 | *OldB = &SmallSide.getInlineBuckets()[i]; |
992 | ::new (&NewB->getFirst()) KeyT(std::move(OldB->getFirst())); |
993 | OldB->getFirst().~KeyT(); |
994 | if (!KeyInfoT::isEqual(NewB->getFirst(), EmptyKey) && |
995 | !KeyInfoT::isEqual(NewB->getFirst(), TombstoneKey)) { |
996 | ::new (&NewB->getSecond()) ValueT(std::move(OldB->getSecond())); |
997 | OldB->getSecond().~ValueT(); |
998 | } |
999 | } |
1000 | |
1001 | // The hard part of moving the small buckets across is done, just move |
1002 | // the TmpRep into its new home. |
1003 | SmallSide.Small = false; |
1004 | new (SmallSide.getLargeRep()) LargeRep(std::move(TmpRep)); |
1005 | } |
1006 | |
1007 | SmallDenseMap& operator=(const SmallDenseMap& other) { |
1008 | if (&other != this) |
1009 | copyFrom(other); |
1010 | return *this; |
1011 | } |
1012 | |
1013 | SmallDenseMap& operator=(SmallDenseMap &&other) { |
1014 | this->destroyAll(); |
1015 | deallocateBuckets(); |
1016 | init(0); |
1017 | swap(other); |
1018 | return *this; |
1019 | } |
1020 | |
1021 | void copyFrom(const SmallDenseMap& other) { |
1022 | this->destroyAll(); |
1023 | deallocateBuckets(); |
1024 | Small = true; |
1025 | if (other.getNumBuckets() > InlineBuckets) { |
1026 | Small = false; |
1027 | new (getLargeRep()) LargeRep(allocateBuckets(other.getNumBuckets())); |
1028 | } |
1029 | this->BaseT::copyFrom(other); |
1030 | } |
1031 | |
1032 | void init(unsigned InitBuckets) { |
1033 | Small = true; |
1034 | if (InitBuckets > InlineBuckets) { |
1035 | Small = false; |
1036 | new (getLargeRep()) LargeRep(allocateBuckets(InitBuckets)); |
1037 | } |
1038 | this->BaseT::initEmpty(); |
1039 | } |
1040 | |
1041 | void grow(unsigned AtLeast) { |
1042 | if (AtLeast > InlineBuckets) |
1043 | AtLeast = std::max<unsigned>(64, NextPowerOf2(AtLeast-1)); |
1044 | |
1045 | if (Small) { |
1046 | // First move the inline buckets into a temporary storage. |
1047 | AlignedCharArrayUnion<BucketT[InlineBuckets]> TmpStorage; |
1048 | BucketT *TmpBegin = reinterpret_cast<BucketT *>(&TmpStorage); |
1049 | BucketT *TmpEnd = TmpBegin; |
1050 | |
1051 | // Loop over the buckets, moving non-empty, non-tombstones into the |
1052 | // temporary storage. Have the loop move the TmpEnd forward as it goes. |
1053 | const KeyT EmptyKey = this->getEmptyKey(); |
1054 | const KeyT TombstoneKey = this->getTombstoneKey(); |
1055 | for (BucketT *P = getBuckets(), *E = P + InlineBuckets; P != E; ++P) { |
1056 | if (!KeyInfoT::isEqual(P->getFirst(), EmptyKey) && |
1057 | !KeyInfoT::isEqual(P->getFirst(), TombstoneKey)) { |
1058 | assert(size_t(TmpEnd - TmpBegin) < InlineBuckets &&((void)0) |
1059 | "Too many inline buckets!")((void)0); |
1060 | ::new (&TmpEnd->getFirst()) KeyT(std::move(P->getFirst())); |
1061 | ::new (&TmpEnd->getSecond()) ValueT(std::move(P->getSecond())); |
1062 | ++TmpEnd; |
1063 | P->getSecond().~ValueT(); |
1064 | } |
1065 | P->getFirst().~KeyT(); |
1066 | } |
1067 | |
1068 | // AtLeast == InlineBuckets can happen if there are many tombstones, |
1069 | // and grow() is used to remove them. Usually we always switch to the |
1070 | // large rep here. |
1071 | if (AtLeast > InlineBuckets) { |
1072 | Small = false; |
1073 | new (getLargeRep()) LargeRep(allocateBuckets(AtLeast)); |
1074 | } |
1075 | this->moveFromOldBuckets(TmpBegin, TmpEnd); |
1076 | return; |
1077 | } |
1078 | |
1079 | LargeRep OldRep = std::move(*getLargeRep()); |
1080 | getLargeRep()->~LargeRep(); |
1081 | if (AtLeast <= InlineBuckets) { |
1082 | Small = true; |
1083 | } else { |
1084 | new (getLargeRep()) LargeRep(allocateBuckets(AtLeast)); |
1085 | } |
1086 | |
1087 | this->moveFromOldBuckets(OldRep.Buckets, OldRep.Buckets+OldRep.NumBuckets); |
1088 | |
1089 | // Free the old table. |
1090 | deallocate_buffer(OldRep.Buckets, sizeof(BucketT) * OldRep.NumBuckets, |
1091 | alignof(BucketT)); |
1092 | } |
1093 | |
1094 | void shrink_and_clear() { |
1095 | unsigned OldSize = this->size(); |
1096 | this->destroyAll(); |
1097 | |
1098 | // Reduce the number of buckets. |
1099 | unsigned NewNumBuckets = 0; |
1100 | if (OldSize) { |
1101 | NewNumBuckets = 1 << (Log2_32_Ceil(OldSize) + 1); |
1102 | if (NewNumBuckets > InlineBuckets && NewNumBuckets < 64u) |
1103 | NewNumBuckets = 64; |
1104 | } |
1105 | if ((Small && NewNumBuckets <= InlineBuckets) || |
1106 | (!Small && NewNumBuckets == getLargeRep()->NumBuckets)) { |
1107 | this->BaseT::initEmpty(); |
1108 | return; |
1109 | } |
1110 | |
1111 | deallocateBuckets(); |
1112 | init(NewNumBuckets); |
1113 | } |
1114 | |
1115 | private: |
1116 | unsigned getNumEntries() const { |
1117 | return NumEntries; |
1118 | } |
1119 | |
1120 | void setNumEntries(unsigned Num) { |
1121 | // NumEntries is hardcoded to be 31 bits wide. |
1122 | assert(Num < (1U << 31) && "Cannot support more than 1<<31 entries")((void)0); |
1123 | NumEntries = Num; |
1124 | } |
1125 | |
1126 | unsigned getNumTombstones() const { |
1127 | return NumTombstones; |
1128 | } |
1129 | |
1130 | void setNumTombstones(unsigned Num) { |
1131 | NumTombstones = Num; |
1132 | } |
1133 | |
1134 | const BucketT *getInlineBuckets() const { |
1135 | assert(Small)((void)0); |
1136 | // Note that this cast does not violate aliasing rules as we assert that |
1137 | // the memory's dynamic type is the small, inline bucket buffer, and the |
1138 | // 'storage' is a POD containing a char buffer. |
1139 | return reinterpret_cast<const BucketT *>(&storage); |
1140 | } |
1141 | |
1142 | BucketT *getInlineBuckets() { |
1143 | return const_cast<BucketT *>( |
1144 | const_cast<const SmallDenseMap *>(this)->getInlineBuckets()); |
1145 | } |
1146 | |
1147 | const LargeRep *getLargeRep() const { |
1148 | assert(!Small)((void)0); |
1149 | // Note, same rule about aliasing as with getInlineBuckets. |
1150 | return reinterpret_cast<const LargeRep *>(&storage); |
1151 | } |
1152 | |
1153 | LargeRep *getLargeRep() { |
1154 | return const_cast<LargeRep *>( |
1155 | const_cast<const SmallDenseMap *>(this)->getLargeRep()); |
1156 | } |
1157 | |
1158 | const BucketT *getBuckets() const { |
1159 | return Small ? getInlineBuckets() : getLargeRep()->Buckets; |
1160 | } |
1161 | |
1162 | BucketT *getBuckets() { |
1163 | return const_cast<BucketT *>( |
1164 | const_cast<const SmallDenseMap *>(this)->getBuckets()); |
1165 | } |
1166 | |
1167 | unsigned getNumBuckets() const { |
1168 | return Small ? InlineBuckets : getLargeRep()->NumBuckets; |
1169 | } |
1170 | |
1171 | void deallocateBuckets() { |
1172 | if (Small) |
1173 | return; |
1174 | |
1175 | deallocate_buffer(getLargeRep()->Buckets, |
1176 | sizeof(BucketT) * getLargeRep()->NumBuckets, |
1177 | alignof(BucketT)); |
1178 | getLargeRep()->~LargeRep(); |
1179 | } |
1180 | |
1181 | LargeRep allocateBuckets(unsigned Num) { |
1182 | assert(Num > InlineBuckets && "Must allocate more buckets than are inline")((void)0); |
1183 | LargeRep Rep = {static_cast<BucketT *>(allocate_buffer( |
1184 | sizeof(BucketT) * Num, alignof(BucketT))), |
1185 | Num}; |
1186 | return Rep; |
1187 | } |
1188 | }; |
1189 | |
1190 | template <typename KeyT, typename ValueT, typename KeyInfoT, typename Bucket, |
1191 | bool IsConst> |
1192 | class DenseMapIterator : DebugEpochBase::HandleBase { |
1193 | friend class DenseMapIterator<KeyT, ValueT, KeyInfoT, Bucket, true>; |
1194 | friend class DenseMapIterator<KeyT, ValueT, KeyInfoT, Bucket, false>; |
1195 | |
1196 | public: |
1197 | using difference_type = ptrdiff_t; |
1198 | using value_type = |
1199 | typename std::conditional<IsConst, const Bucket, Bucket>::type; |
1200 | using pointer = value_type *; |
1201 | using reference = value_type &; |
1202 | using iterator_category = std::forward_iterator_tag; |
1203 | |
1204 | private: |
1205 | pointer Ptr = nullptr; |
1206 | pointer End = nullptr; |
1207 | |
1208 | public: |
1209 | DenseMapIterator() = default; |
1210 | |
1211 | DenseMapIterator(pointer Pos, pointer E, const DebugEpochBase &Epoch, |
1212 | bool NoAdvance = false) |
1213 | : DebugEpochBase::HandleBase(&Epoch), Ptr(Pos), End(E) { |
1214 | assert(isHandleInSync() && "invalid construction!")((void)0); |
1215 | |
1216 | if (NoAdvance) return; |
1217 | if (shouldReverseIterate<KeyT>()) { |
1218 | RetreatPastEmptyBuckets(); |
1219 | return; |
1220 | } |
1221 | AdvancePastEmptyBuckets(); |
1222 | } |
1223 | |
1224 | // Converting ctor from non-const iterators to const iterators. SFINAE'd out |
1225 | // for const iterator destinations so it doesn't end up as a user defined copy |
1226 | // constructor. |
1227 | template <bool IsConstSrc, |
1228 | typename = std::enable_if_t<!IsConstSrc && IsConst>> |
1229 | DenseMapIterator( |
1230 | const DenseMapIterator<KeyT, ValueT, KeyInfoT, Bucket, IsConstSrc> &I) |
1231 | : DebugEpochBase::HandleBase(I), Ptr(I.Ptr), End(I.End) {} |
1232 | |
1233 | reference operator*() const { |
1234 | assert(isHandleInSync() && "invalid iterator access!")((void)0); |
1235 | assert(Ptr != End && "dereferencing end() iterator")((void)0); |
1236 | if (shouldReverseIterate<KeyT>()) |
1237 | return Ptr[-1]; |
1238 | return *Ptr; |
1239 | } |
1240 | pointer operator->() const { |
1241 | assert(isHandleInSync() && "invalid iterator access!")((void)0); |
1242 | assert(Ptr != End && "dereferencing end() iterator")((void)0); |
1243 | if (shouldReverseIterate<KeyT>()) |
1244 | return &(Ptr[-1]); |
1245 | return Ptr; |
1246 | } |
1247 | |
1248 | friend bool operator==(const DenseMapIterator &LHS, |
1249 | const DenseMapIterator &RHS) { |
1250 | assert((!LHS.Ptr || LHS.isHandleInSync()) && "handle not in sync!")((void)0); |
1251 | assert((!RHS.Ptr || RHS.isHandleInSync()) && "handle not in sync!")((void)0); |
1252 | assert(LHS.getEpochAddress() == RHS.getEpochAddress() &&((void)0) |
1253 | "comparing incomparable iterators!")((void)0); |
1254 | return LHS.Ptr == RHS.Ptr; |
1255 | } |
1256 | |
1257 | friend bool operator!=(const DenseMapIterator &LHS, |
1258 | const DenseMapIterator &RHS) { |
1259 | return !(LHS == RHS); |
1260 | } |
1261 | |
1262 | inline DenseMapIterator& operator++() { // Preincrement |
1263 | assert(isHandleInSync() && "invalid iterator access!")((void)0); |
1264 | assert(Ptr != End && "incrementing end() iterator")((void)0); |
1265 | if (shouldReverseIterate<KeyT>()) { |
1266 | --Ptr; |
1267 | RetreatPastEmptyBuckets(); |
1268 | return *this; |
1269 | } |
1270 | ++Ptr; |
1271 | AdvancePastEmptyBuckets(); |
1272 | return *this; |
1273 | } |
1274 | DenseMapIterator operator++(int) { // Postincrement |
1275 | assert(isHandleInSync() && "invalid iterator access!")((void)0); |
1276 | DenseMapIterator tmp = *this; ++*this; return tmp; |
1277 | } |
1278 | |
1279 | private: |
1280 | void AdvancePastEmptyBuckets() { |
1281 | assert(Ptr <= End)((void)0); |
1282 | const KeyT Empty = KeyInfoT::getEmptyKey(); |
1283 | const KeyT Tombstone = KeyInfoT::getTombstoneKey(); |
1284 | |
1285 | while (Ptr != End && (KeyInfoT::isEqual(Ptr->getFirst(), Empty) || |
1286 | KeyInfoT::isEqual(Ptr->getFirst(), Tombstone))) |
1287 | ++Ptr; |
1288 | } |
1289 | |
1290 | void RetreatPastEmptyBuckets() { |
1291 | assert(Ptr >= End)((void)0); |
1292 | const KeyT Empty = KeyInfoT::getEmptyKey(); |
1293 | const KeyT Tombstone = KeyInfoT::getTombstoneKey(); |
1294 | |
1295 | while (Ptr != End && (KeyInfoT::isEqual(Ptr[-1].getFirst(), Empty) || |
1296 | KeyInfoT::isEqual(Ptr[-1].getFirst(), Tombstone))) |
1297 | --Ptr; |
1298 | } |
1299 | }; |
1300 | |
1301 | template <typename KeyT, typename ValueT, typename KeyInfoT> |
1302 | inline size_t capacity_in_bytes(const DenseMap<KeyT, ValueT, KeyInfoT> &X) { |
1303 | return X.getMemorySize(); |
1304 | } |
1305 | |
1306 | } // end namespace llvm |
1307 | |
1308 | #endif // LLVM_ADT_DENSEMAP_H |