File: | build/source/llvm/lib/CodeGen/InlineSpiller.cpp |
Warning: | line 1381, 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/STLExtras.h" | |||
19 | #include "llvm/ADT/SetVector.h" | |||
20 | #include "llvm/ADT/SmallPtrSet.h" | |||
21 | #include "llvm/ADT/SmallVector.h" | |||
22 | #include "llvm/ADT/Statistic.h" | |||
23 | #include "llvm/Analysis/AliasAnalysis.h" | |||
24 | #include "llvm/CodeGen/LiveInterval.h" | |||
25 | #include "llvm/CodeGen/LiveIntervals.h" | |||
26 | #include "llvm/CodeGen/LiveRangeEdit.h" | |||
27 | #include "llvm/CodeGen/LiveStacks.h" | |||
28 | #include "llvm/CodeGen/MachineBasicBlock.h" | |||
29 | #include "llvm/CodeGen/MachineBlockFrequencyInfo.h" | |||
30 | #include "llvm/CodeGen/MachineDominators.h" | |||
31 | #include "llvm/CodeGen/MachineFunction.h" | |||
32 | #include "llvm/CodeGen/MachineFunctionPass.h" | |||
33 | #include "llvm/CodeGen/MachineInstr.h" | |||
34 | #include "llvm/CodeGen/MachineInstrBuilder.h" | |||
35 | #include "llvm/CodeGen/MachineInstrBundle.h" | |||
36 | #include "llvm/CodeGen/MachineLoopInfo.h" | |||
37 | #include "llvm/CodeGen/MachineOperand.h" | |||
38 | #include "llvm/CodeGen/MachineRegisterInfo.h" | |||
39 | #include "llvm/CodeGen/SlotIndexes.h" | |||
40 | #include "llvm/CodeGen/Spiller.h" | |||
41 | #include "llvm/CodeGen/StackMaps.h" | |||
42 | #include "llvm/CodeGen/TargetInstrInfo.h" | |||
43 | #include "llvm/CodeGen/TargetOpcodes.h" | |||
44 | #include "llvm/CodeGen/TargetRegisterInfo.h" | |||
45 | #include "llvm/CodeGen/TargetSubtargetInfo.h" | |||
46 | #include "llvm/CodeGen/VirtRegMap.h" | |||
47 | #include "llvm/Config/llvm-config.h" | |||
48 | #include "llvm/Support/BlockFrequency.h" | |||
49 | #include "llvm/Support/BranchProbability.h" | |||
50 | #include "llvm/Support/CommandLine.h" | |||
51 | #include "llvm/Support/Compiler.h" | |||
52 | #include "llvm/Support/Debug.h" | |||
53 | #include "llvm/Support/ErrorHandling.h" | |||
54 | #include "llvm/Support/raw_ostream.h" | |||
55 | #include <cassert> | |||
56 | #include <iterator> | |||
57 | #include <tuple> | |||
58 | #include <utility> | |||
59 | #include <vector> | |||
60 | ||||
61 | using namespace llvm; | |||
62 | ||||
63 | #define DEBUG_TYPE"regalloc" "regalloc" | |||
64 | ||||
65 | STATISTIC(NumSpilledRanges, "Number of spilled live ranges")static llvm::Statistic NumSpilledRanges = {"regalloc", "NumSpilledRanges" , "Number of spilled live ranges"}; | |||
66 | STATISTIC(NumSnippets, "Number of spilled snippets")static llvm::Statistic NumSnippets = {"regalloc", "NumSnippets" , "Number of spilled snippets"}; | |||
67 | STATISTIC(NumSpills, "Number of spills inserted")static llvm::Statistic NumSpills = {"regalloc", "NumSpills", "Number of spills inserted" }; | |||
68 | STATISTIC(NumSpillsRemoved, "Number of spills removed")static llvm::Statistic NumSpillsRemoved = {"regalloc", "NumSpillsRemoved" , "Number of spills removed"}; | |||
69 | STATISTIC(NumReloads, "Number of reloads inserted")static llvm::Statistic NumReloads = {"regalloc", "NumReloads" , "Number of reloads inserted"}; | |||
70 | STATISTIC(NumReloadsRemoved, "Number of reloads removed")static llvm::Statistic NumReloadsRemoved = {"regalloc", "NumReloadsRemoved" , "Number of reloads removed"}; | |||
71 | STATISTIC(NumFolded, "Number of folded stack accesses")static llvm::Statistic NumFolded = {"regalloc", "NumFolded", "Number of folded stack accesses" }; | |||
72 | STATISTIC(NumFoldedLoads, "Number of folded loads")static llvm::Statistic NumFoldedLoads = {"regalloc", "NumFoldedLoads" , "Number of folded loads"}; | |||
73 | STATISTIC(NumRemats, "Number of rematerialized defs for spilling")static llvm::Statistic NumRemats = {"regalloc", "NumRemats", "Number of rematerialized defs for spilling" }; | |||
74 | ||||
75 | static cl::opt<bool> DisableHoisting("disable-spill-hoist", cl::Hidden, | |||
76 | cl::desc("Disable inline spill hoisting")); | |||
77 | static cl::opt<bool> | |||
78 | RestrictStatepointRemat("restrict-statepoint-remat", | |||
79 | cl::init(false), cl::Hidden, | |||
80 | cl::desc("Restrict remat for statepoint operands")); | |||
81 | ||||
82 | namespace { | |||
83 | ||||
84 | class HoistSpillHelper : private LiveRangeEdit::Delegate { | |||
85 | MachineFunction &MF; | |||
86 | LiveIntervals &LIS; | |||
87 | LiveStacks &LSS; | |||
88 | MachineDominatorTree &MDT; | |||
89 | MachineLoopInfo &Loops; | |||
90 | VirtRegMap &VRM; | |||
91 | MachineRegisterInfo &MRI; | |||
92 | const TargetInstrInfo &TII; | |||
93 | const TargetRegisterInfo &TRI; | |||
94 | const MachineBlockFrequencyInfo &MBFI; | |||
95 | ||||
96 | InsertPointAnalysis IPA; | |||
97 | ||||
98 | // Map from StackSlot to the LiveInterval of the original register. | |||
99 | // Note the LiveInterval of the original register may have been deleted | |||
100 | // after it is spilled. We keep a copy here to track the range where | |||
101 | // spills can be moved. | |||
102 | DenseMap<int, std::unique_ptr<LiveInterval>> StackSlotToOrigLI; | |||
103 | ||||
104 | // Map from pair of (StackSlot and Original VNI) to a set of spills which | |||
105 | // have the same stackslot and have equal values defined by Original VNI. | |||
106 | // These spills are mergeable and are hoist candidates. | |||
107 | using MergeableSpillsMap = | |||
108 | MapVector<std::pair<int, VNInfo *>, SmallPtrSet<MachineInstr *, 16>>; | |||
109 | MergeableSpillsMap MergeableSpills; | |||
110 | ||||
111 | /// This is the map from original register to a set containing all its | |||
112 | /// siblings. To hoist a spill to another BB, we need to find out a live | |||
113 | /// sibling there and use it as the source of the new spill. | |||
114 | DenseMap<Register, SmallSetVector<Register, 16>> Virt2SiblingsMap; | |||
115 | ||||
116 | bool isSpillCandBB(LiveInterval &OrigLI, VNInfo &OrigVNI, | |||
117 | MachineBasicBlock &BB, Register &LiveReg); | |||
118 | ||||
119 | void rmRedundantSpills( | |||
120 | SmallPtrSet<MachineInstr *, 16> &Spills, | |||
121 | SmallVectorImpl<MachineInstr *> &SpillsToRm, | |||
122 | DenseMap<MachineDomTreeNode *, MachineInstr *> &SpillBBToSpill); | |||
123 | ||||
124 | void getVisitOrders( | |||
125 | MachineBasicBlock *Root, SmallPtrSet<MachineInstr *, 16> &Spills, | |||
126 | SmallVectorImpl<MachineDomTreeNode *> &Orders, | |||
127 | SmallVectorImpl<MachineInstr *> &SpillsToRm, | |||
128 | DenseMap<MachineDomTreeNode *, unsigned> &SpillsToKeep, | |||
129 | DenseMap<MachineDomTreeNode *, MachineInstr *> &SpillBBToSpill); | |||
130 | ||||
131 | void runHoistSpills(LiveInterval &OrigLI, VNInfo &OrigVNI, | |||
132 | SmallPtrSet<MachineInstr *, 16> &Spills, | |||
133 | SmallVectorImpl<MachineInstr *> &SpillsToRm, | |||
134 | DenseMap<MachineBasicBlock *, unsigned> &SpillsToIns); | |||
135 | ||||
136 | public: | |||
137 | HoistSpillHelper(MachineFunctionPass &pass, MachineFunction &mf, | |||
138 | VirtRegMap &vrm) | |||
139 | : MF(mf), LIS(pass.getAnalysis<LiveIntervals>()), | |||
140 | LSS(pass.getAnalysis<LiveStacks>()), | |||
141 | MDT(pass.getAnalysis<MachineDominatorTree>()), | |||
142 | Loops(pass.getAnalysis<MachineLoopInfo>()), VRM(vrm), | |||
143 | MRI(mf.getRegInfo()), TII(*mf.getSubtarget().getInstrInfo()), | |||
144 | TRI(*mf.getSubtarget().getRegisterInfo()), | |||
145 | MBFI(pass.getAnalysis<MachineBlockFrequencyInfo>()), | |||
146 | IPA(LIS, mf.getNumBlockIDs()) {} | |||
147 | ||||
148 | void addToMergeableSpills(MachineInstr &Spill, int StackSlot, | |||
149 | unsigned Original); | |||
150 | bool rmFromMergeableSpills(MachineInstr &Spill, int StackSlot); | |||
151 | void hoistAllSpills(); | |||
152 | void LRE_DidCloneVirtReg(Register, Register) override; | |||
153 | }; | |||
154 | ||||
155 | class InlineSpiller : public Spiller { | |||
156 | MachineFunction &MF; | |||
157 | LiveIntervals &LIS; | |||
158 | LiveStacks &LSS; | |||
159 | MachineDominatorTree &MDT; | |||
160 | MachineLoopInfo &Loops; | |||
161 | VirtRegMap &VRM; | |||
162 | MachineRegisterInfo &MRI; | |||
163 | const TargetInstrInfo &TII; | |||
164 | const TargetRegisterInfo &TRI; | |||
165 | const MachineBlockFrequencyInfo &MBFI; | |||
166 | ||||
167 | // Variables that are valid during spill(), but used by multiple methods. | |||
168 | LiveRangeEdit *Edit; | |||
169 | LiveInterval *StackInt; | |||
170 | int StackSlot; | |||
171 | Register Original; | |||
172 | ||||
173 | // All registers to spill to StackSlot, including the main register. | |||
174 | SmallVector<Register, 8> RegsToSpill; | |||
175 | ||||
176 | // All COPY instructions to/from snippets. | |||
177 | // They are ignored since both operands refer to the same stack slot. | |||
178 | SmallPtrSet<MachineInstr*, 8> SnippetCopies; | |||
179 | ||||
180 | // Values that failed to remat at some point. | |||
181 | SmallPtrSet<VNInfo*, 8> UsedValues; | |||
182 | ||||
183 | // Dead defs generated during spilling. | |||
184 | SmallVector<MachineInstr*, 8> DeadDefs; | |||
185 | ||||
186 | // Object records spills information and does the hoisting. | |||
187 | HoistSpillHelper HSpiller; | |||
188 | ||||
189 | // Live range weight calculator. | |||
190 | VirtRegAuxInfo &VRAI; | |||
191 | ||||
192 | ~InlineSpiller() override = default; | |||
193 | ||||
194 | public: | |||
195 | InlineSpiller(MachineFunctionPass &Pass, MachineFunction &MF, VirtRegMap &VRM, | |||
196 | VirtRegAuxInfo &VRAI) | |||
197 | : MF(MF), LIS(Pass.getAnalysis<LiveIntervals>()), | |||
198 | LSS(Pass.getAnalysis<LiveStacks>()), | |||
199 | MDT(Pass.getAnalysis<MachineDominatorTree>()), | |||
200 | Loops(Pass.getAnalysis<MachineLoopInfo>()), VRM(VRM), | |||
201 | MRI(MF.getRegInfo()), TII(*MF.getSubtarget().getInstrInfo()), | |||
202 | TRI(*MF.getSubtarget().getRegisterInfo()), | |||
203 | MBFI(Pass.getAnalysis<MachineBlockFrequencyInfo>()), | |||
204 | HSpiller(Pass, MF, VRM), VRAI(VRAI) {} | |||
205 | ||||
206 | void spill(LiveRangeEdit &) override; | |||
207 | void postOptimization() override; | |||
208 | ||||
209 | private: | |||
210 | bool isSnippet(const LiveInterval &SnipLI); | |||
211 | void collectRegsToSpill(); | |||
212 | ||||
213 | bool isRegToSpill(Register Reg) { return is_contained(RegsToSpill, Reg); } | |||
214 | ||||
215 | bool isSibling(Register Reg); | |||
216 | bool hoistSpillInsideBB(LiveInterval &SpillLI, MachineInstr &CopyMI); | |||
217 | void eliminateRedundantSpills(LiveInterval &LI, VNInfo *VNI); | |||
218 | ||||
219 | void markValueUsed(LiveInterval*, VNInfo*); | |||
220 | bool canGuaranteeAssignmentAfterRemat(Register VReg, MachineInstr &MI); | |||
221 | bool reMaterializeFor(LiveInterval &, MachineInstr &MI); | |||
222 | void reMaterializeAll(); | |||
223 | ||||
224 | bool coalesceStackAccess(MachineInstr *MI, Register Reg); | |||
225 | bool foldMemoryOperand(ArrayRef<std::pair<MachineInstr *, unsigned>>, | |||
226 | MachineInstr *LoadMI = nullptr); | |||
227 | void insertReload(Register VReg, SlotIndex, MachineBasicBlock::iterator MI); | |||
228 | void insertSpill(Register VReg, bool isKill, MachineBasicBlock::iterator MI); | |||
229 | ||||
230 | void spillAroundUses(Register Reg); | |||
231 | void spillAll(); | |||
232 | }; | |||
233 | ||||
234 | } // end anonymous namespace | |||
235 | ||||
236 | Spiller::~Spiller() = default; | |||
237 | ||||
238 | void Spiller::anchor() {} | |||
239 | ||||
240 | Spiller *llvm::createInlineSpiller(MachineFunctionPass &Pass, | |||
241 | MachineFunction &MF, VirtRegMap &VRM, | |||
242 | VirtRegAuxInfo &VRAI) { | |||
243 | return new InlineSpiller(Pass, MF, VRM, VRAI); | |||
244 | } | |||
245 | ||||
246 | //===----------------------------------------------------------------------===// | |||
247 | // Snippets | |||
248 | //===----------------------------------------------------------------------===// | |||
249 | ||||
250 | // When spilling a virtual register, we also spill any snippets it is connected | |||
251 | // to. The snippets are small live ranges that only have a single real use, | |||
252 | // leftovers from live range splitting. Spilling them enables memory operand | |||
253 | // folding or tightens the live range around the single use. | |||
254 | // | |||
255 | // This minimizes register pressure and maximizes the store-to-load distance for | |||
256 | // spill slots which can be important in tight loops. | |||
257 | ||||
258 | /// isFullCopyOf - If MI is a COPY to or from Reg, return the other register, | |||
259 | /// otherwise return 0. | |||
260 | static Register isFullCopyOf(const MachineInstr &MI, Register Reg) { | |||
261 | if (!MI.isFullCopy()) | |||
262 | return Register(); | |||
263 | if (MI.getOperand(0).getReg() == Reg) | |||
264 | return MI.getOperand(1).getReg(); | |||
265 | if (MI.getOperand(1).getReg() == Reg) | |||
266 | return MI.getOperand(0).getReg(); | |||
267 | return Register(); | |||
268 | } | |||
269 | ||||
270 | static void getVDefInterval(const MachineInstr &MI, LiveIntervals &LIS) { | |||
271 | for (const MachineOperand &MO : MI.operands()) | |||
272 | if (MO.isReg() && MO.isDef() && MO.getReg().isVirtual()) | |||
273 | LIS.getInterval(MO.getReg()); | |||
274 | } | |||
275 | ||||
276 | /// isSnippet - Identify if a live interval is a snippet that should be spilled. | |||
277 | /// It is assumed that SnipLI is a virtual register with the same original as | |||
278 | /// Edit->getReg(). | |||
279 | bool InlineSpiller::isSnippet(const LiveInterval &SnipLI) { | |||
280 | Register Reg = Edit->getReg(); | |||
281 | ||||
282 | // A snippet is a tiny live range with only a single instruction using it | |||
283 | // besides copies to/from Reg or spills/fills. | |||
284 | // Exception is done for statepoint instructions which will fold fills | |||
285 | // into their operands. | |||
286 | // We accept: | |||
287 | // | |||
288 | // %snip = COPY %Reg / FILL fi# | |||
289 | // %snip = USE %snip | |||
290 | // %snip = STATEPOINT %snip in var arg area | |||
291 | // %Reg = COPY %snip / SPILL %snip, fi# | |||
292 | // | |||
293 | if (!LIS.intervalIsInOneMBB(SnipLI)) | |||
294 | return false; | |||
295 | ||||
296 | // Number of defs should not exceed 2 not accounting defs coming from | |||
297 | // statepoint instructions. | |||
298 | unsigned NumValNums = SnipLI.getNumValNums(); | |||
299 | for (auto *VNI : SnipLI.vnis()) { | |||
300 | MachineInstr *MI = LIS.getInstructionFromIndex(VNI->def); | |||
301 | if (MI->getOpcode() == TargetOpcode::STATEPOINT) | |||
302 | --NumValNums; | |||
303 | } | |||
304 | if (NumValNums > 2) | |||
305 | return false; | |||
306 | ||||
307 | MachineInstr *UseMI = nullptr; | |||
308 | ||||
309 | // Check that all uses satisfy our criteria. | |||
310 | for (MachineRegisterInfo::reg_instr_nodbg_iterator | |||
311 | RI = MRI.reg_instr_nodbg_begin(SnipLI.reg()), | |||
312 | E = MRI.reg_instr_nodbg_end(); | |||
313 | RI != E;) { | |||
314 | MachineInstr &MI = *RI++; | |||
315 | ||||
316 | // Allow copies to/from Reg. | |||
317 | if (isFullCopyOf(MI, Reg)) | |||
318 | continue; | |||
319 | ||||
320 | // Allow stack slot loads. | |||
321 | int FI; | |||
322 | if (SnipLI.reg() == TII.isLoadFromStackSlot(MI, FI) && FI == StackSlot) | |||
323 | continue; | |||
324 | ||||
325 | // Allow stack slot stores. | |||
326 | if (SnipLI.reg() == TII.isStoreToStackSlot(MI, FI) && FI == StackSlot) | |||
327 | continue; | |||
328 | ||||
329 | if (StatepointOpers::isFoldableReg(&MI, SnipLI.reg())) | |||
330 | continue; | |||
331 | ||||
332 | // Allow a single additional instruction. | |||
333 | if (UseMI && &MI != UseMI) | |||
334 | return false; | |||
335 | UseMI = &MI; | |||
336 | } | |||
337 | return true; | |||
338 | } | |||
339 | ||||
340 | /// collectRegsToSpill - Collect live range snippets that only have a single | |||
341 | /// real use. | |||
342 | void InlineSpiller::collectRegsToSpill() { | |||
343 | Register Reg = Edit->getReg(); | |||
344 | ||||
345 | // Main register always spills. | |||
346 | RegsToSpill.assign(1, Reg); | |||
347 | SnippetCopies.clear(); | |||
348 | ||||
349 | // Snippets all have the same original, so there can't be any for an original | |||
350 | // register. | |||
351 | if (Original == Reg) | |||
352 | return; | |||
353 | ||||
354 | for (MachineInstr &MI : | |||
355 | llvm::make_early_inc_range(MRI.reg_instructions(Reg))) { | |||
356 | Register SnipReg = isFullCopyOf(MI, Reg); | |||
357 | if (!isSibling(SnipReg)) | |||
358 | continue; | |||
359 | LiveInterval &SnipLI = LIS.getInterval(SnipReg); | |||
360 | if (!isSnippet(SnipLI)) | |||
361 | continue; | |||
362 | SnippetCopies.insert(&MI); | |||
363 | if (isRegToSpill(SnipReg)) | |||
364 | continue; | |||
365 | RegsToSpill.push_back(SnipReg); | |||
366 | LLVM_DEBUG(dbgs() << "\talso spill snippet " << SnipLI << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << "\talso spill snippet " << SnipLI << '\n'; } } while (false); | |||
367 | ++NumSnippets; | |||
368 | } | |||
369 | } | |||
370 | ||||
371 | bool InlineSpiller::isSibling(Register Reg) { | |||
372 | return Reg.isVirtual() && VRM.getOriginal(Reg) == Original; | |||
373 | } | |||
374 | ||||
375 | /// It is beneficial to spill to earlier place in the same BB in case | |||
376 | /// as follows: | |||
377 | /// There is an alternative def earlier in the same MBB. | |||
378 | /// Hoist the spill as far as possible in SpillMBB. This can ease | |||
379 | /// register pressure: | |||
380 | /// | |||
381 | /// x = def | |||
382 | /// y = use x | |||
383 | /// s = copy x | |||
384 | /// | |||
385 | /// Hoisting the spill of s to immediately after the def removes the | |||
386 | /// interference between x and y: | |||
387 | /// | |||
388 | /// x = def | |||
389 | /// spill x | |||
390 | /// y = use killed x | |||
391 | /// | |||
392 | /// This hoist only helps when the copy kills its source. | |||
393 | /// | |||
394 | bool InlineSpiller::hoistSpillInsideBB(LiveInterval &SpillLI, | |||
395 | MachineInstr &CopyMI) { | |||
396 | SlotIndex Idx = LIS.getInstructionIndex(CopyMI); | |||
397 | #ifndef NDEBUG | |||
398 | VNInfo *VNI = SpillLI.getVNInfoAt(Idx.getRegSlot()); | |||
399 | assert(VNI && VNI->def == Idx.getRegSlot() && "Not defined by copy")(static_cast <bool> (VNI && VNI->def == Idx. getRegSlot() && "Not defined by copy") ? void (0) : __assert_fail ("VNI && VNI->def == Idx.getRegSlot() && \"Not defined by copy\"" , "llvm/lib/CodeGen/InlineSpiller.cpp", 399, __extension__ __PRETTY_FUNCTION__ )); | |||
400 | #endif | |||
401 | ||||
402 | Register SrcReg = CopyMI.getOperand(1).getReg(); | |||
403 | LiveInterval &SrcLI = LIS.getInterval(SrcReg); | |||
404 | VNInfo *SrcVNI = SrcLI.getVNInfoAt(Idx); | |||
405 | LiveQueryResult SrcQ = SrcLI.Query(Idx); | |||
406 | MachineBasicBlock *DefMBB = LIS.getMBBFromIndex(SrcVNI->def); | |||
407 | if (DefMBB != CopyMI.getParent() || !SrcQ.isKill()) | |||
408 | return false; | |||
409 | ||||
410 | // Conservatively extend the stack slot range to the range of the original | |||
411 | // value. We may be able to do better with stack slot coloring by being more | |||
412 | // careful here. | |||
413 | assert(StackInt && "No stack slot assigned yet.")(static_cast <bool> (StackInt && "No stack slot assigned yet." ) ? void (0) : __assert_fail ("StackInt && \"No stack slot assigned yet.\"" , "llvm/lib/CodeGen/InlineSpiller.cpp", 413, __extension__ __PRETTY_FUNCTION__ )); | |||
414 | LiveInterval &OrigLI = LIS.getInterval(Original); | |||
415 | VNInfo *OrigVNI = OrigLI.getVNInfoAt(Idx); | |||
416 | StackInt->MergeValueInAsValue(OrigLI, OrigVNI, StackInt->getValNumInfo(0)); | |||
417 | LLVM_DEBUG(dbgs() << "\tmerged orig valno " << OrigVNI->id << ": "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << "\tmerged orig valno " << OrigVNI->id << ": " << *StackInt << '\n' ; } } while (false) | |||
418 | << *StackInt << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << "\tmerged orig valno " << OrigVNI->id << ": " << *StackInt << '\n' ; } } while (false); | |||
419 | ||||
420 | // We are going to spill SrcVNI immediately after its def, so clear out | |||
421 | // any later spills of the same value. | |||
422 | eliminateRedundantSpills(SrcLI, SrcVNI); | |||
423 | ||||
424 | MachineBasicBlock *MBB = LIS.getMBBFromIndex(SrcVNI->def); | |||
425 | MachineBasicBlock::iterator MII; | |||
426 | if (SrcVNI->isPHIDef()) | |||
427 | MII = MBB->SkipPHIsLabelsAndDebug(MBB->begin()); | |||
428 | else { | |||
429 | MachineInstr *DefMI = LIS.getInstructionFromIndex(SrcVNI->def); | |||
430 | assert(DefMI && "Defining instruction disappeared")(static_cast <bool> (DefMI && "Defining instruction disappeared" ) ? void (0) : __assert_fail ("DefMI && \"Defining instruction disappeared\"" , "llvm/lib/CodeGen/InlineSpiller.cpp", 430, __extension__ __PRETTY_FUNCTION__ )); | |||
431 | MII = DefMI; | |||
432 | ++MII; | |||
433 | } | |||
434 | MachineInstrSpan MIS(MII, MBB); | |||
435 | // Insert spill without kill flag immediately after def. | |||
436 | TII.storeRegToStackSlot(*MBB, MII, SrcReg, false, StackSlot, | |||
437 | MRI.getRegClass(SrcReg), &TRI, Register()); | |||
438 | LIS.InsertMachineInstrRangeInMaps(MIS.begin(), MII); | |||
439 | for (const MachineInstr &MI : make_range(MIS.begin(), MII)) | |||
440 | getVDefInterval(MI, LIS); | |||
441 | --MII; // Point to store instruction. | |||
442 | LLVM_DEBUG(dbgs() << "\thoisted: " << SrcVNI->def << '\t' << *MII)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << "\thoisted: " << SrcVNI ->def << '\t' << *MII; } } while (false); | |||
443 | ||||
444 | // If there is only 1 store instruction is required for spill, add it | |||
445 | // to mergeable list. In X86 AMX, 2 intructions are required to store. | |||
446 | // We disable the merge for this case. | |||
447 | if (MIS.begin() == MII) | |||
448 | HSpiller.addToMergeableSpills(*MII, StackSlot, Original); | |||
449 | ++NumSpills; | |||
450 | return true; | |||
451 | } | |||
452 | ||||
453 | /// eliminateRedundantSpills - SLI:VNI is known to be on the stack. Remove any | |||
454 | /// redundant spills of this value in SLI.reg and sibling copies. | |||
455 | void InlineSpiller::eliminateRedundantSpills(LiveInterval &SLI, VNInfo *VNI) { | |||
456 | assert(VNI && "Missing value")(static_cast <bool> (VNI && "Missing value") ? void (0) : __assert_fail ("VNI && \"Missing value\"", "llvm/lib/CodeGen/InlineSpiller.cpp" , 456, __extension__ __PRETTY_FUNCTION__)); | |||
457 | SmallVector<std::pair<LiveInterval*, VNInfo*>, 8> WorkList; | |||
458 | WorkList.push_back(std::make_pair(&SLI, VNI)); | |||
459 | assert(StackInt && "No stack slot assigned yet.")(static_cast <bool> (StackInt && "No stack slot assigned yet." ) ? void (0) : __assert_fail ("StackInt && \"No stack slot assigned yet.\"" , "llvm/lib/CodeGen/InlineSpiller.cpp", 459, __extension__ __PRETTY_FUNCTION__ )); | |||
460 | ||||
461 | do { | |||
462 | LiveInterval *LI; | |||
463 | std::tie(LI, VNI) = WorkList.pop_back_val(); | |||
464 | Register Reg = LI->reg(); | |||
465 | LLVM_DEBUG(dbgs() << "Checking redundant spills for " << VNI->id << '@'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << "Checking redundant spills for " << VNI->id << '@' << VNI->def << " in " << *LI << '\n'; } } while (false) | |||
466 | << VNI->def << " in " << *LI << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << "Checking redundant spills for " << VNI->id << '@' << VNI->def << " in " << *LI << '\n'; } } while (false); | |||
467 | ||||
468 | // Regs to spill are taken care of. | |||
469 | if (isRegToSpill(Reg)) | |||
470 | continue; | |||
471 | ||||
472 | // Add all of VNI's live range to StackInt. | |||
473 | StackInt->MergeValueInAsValue(*LI, VNI, StackInt->getValNumInfo(0)); | |||
474 | LLVM_DEBUG(dbgs() << "Merged to stack int: " << *StackInt << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << "Merged to stack int: " << *StackInt << '\n'; } } while (false); | |||
475 | ||||
476 | // Find all spills and copies of VNI. | |||
477 | for (MachineInstr &MI : | |||
478 | llvm::make_early_inc_range(MRI.use_nodbg_instructions(Reg))) { | |||
479 | if (!MI.isCopy() && !MI.mayStore()) | |||
480 | continue; | |||
481 | SlotIndex Idx = LIS.getInstructionIndex(MI); | |||
482 | if (LI->getVNInfoAt(Idx) != VNI) | |||
483 | continue; | |||
484 | ||||
485 | // Follow sibling copies down the dominator tree. | |||
486 | if (Register DstReg = isFullCopyOf(MI, Reg)) { | |||
487 | if (isSibling(DstReg)) { | |||
488 | LiveInterval &DstLI = LIS.getInterval(DstReg); | |||
489 | VNInfo *DstVNI = DstLI.getVNInfoAt(Idx.getRegSlot()); | |||
490 | assert(DstVNI && "Missing defined value")(static_cast <bool> (DstVNI && "Missing defined value" ) ? void (0) : __assert_fail ("DstVNI && \"Missing defined value\"" , "llvm/lib/CodeGen/InlineSpiller.cpp", 490, __extension__ __PRETTY_FUNCTION__ )); | |||
491 | assert(DstVNI->def == Idx.getRegSlot() && "Wrong copy def slot")(static_cast <bool> (DstVNI->def == Idx.getRegSlot() && "Wrong copy def slot") ? void (0) : __assert_fail ("DstVNI->def == Idx.getRegSlot() && \"Wrong copy def slot\"" , "llvm/lib/CodeGen/InlineSpiller.cpp", 491, __extension__ __PRETTY_FUNCTION__ )); | |||
492 | WorkList.push_back(std::make_pair(&DstLI, DstVNI)); | |||
493 | } | |||
494 | continue; | |||
495 | } | |||
496 | ||||
497 | // Erase spills. | |||
498 | int FI; | |||
499 | if (Reg == TII.isStoreToStackSlot(MI, FI) && FI == StackSlot) { | |||
500 | LLVM_DEBUG(dbgs() << "Redundant spill " << Idx << '\t' << MI)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << "Redundant spill " << Idx << '\t' << MI; } } while (false); | |||
501 | // eliminateDeadDefs won't normally remove stores, so switch opcode. | |||
502 | MI.setDesc(TII.get(TargetOpcode::KILL)); | |||
503 | DeadDefs.push_back(&MI); | |||
504 | ++NumSpillsRemoved; | |||
505 | if (HSpiller.rmFromMergeableSpills(MI, StackSlot)) | |||
506 | --NumSpills; | |||
507 | } | |||
508 | } | |||
509 | } while (!WorkList.empty()); | |||
510 | } | |||
511 | ||||
512 | //===----------------------------------------------------------------------===// | |||
513 | // Rematerialization | |||
514 | //===----------------------------------------------------------------------===// | |||
515 | ||||
516 | /// markValueUsed - Remember that VNI failed to rematerialize, so its defining | |||
517 | /// instruction cannot be eliminated. See through snippet copies | |||
518 | void InlineSpiller::markValueUsed(LiveInterval *LI, VNInfo *VNI) { | |||
519 | SmallVector<std::pair<LiveInterval*, VNInfo*>, 8> WorkList; | |||
520 | WorkList.push_back(std::make_pair(LI, VNI)); | |||
521 | do { | |||
522 | std::tie(LI, VNI) = WorkList.pop_back_val(); | |||
523 | if (!UsedValues.insert(VNI).second) | |||
524 | continue; | |||
525 | ||||
526 | if (VNI->isPHIDef()) { | |||
527 | MachineBasicBlock *MBB = LIS.getMBBFromIndex(VNI->def); | |||
528 | for (MachineBasicBlock *P : MBB->predecessors()) { | |||
529 | VNInfo *PVNI = LI->getVNInfoBefore(LIS.getMBBEndIdx(P)); | |||
530 | if (PVNI) | |||
531 | WorkList.push_back(std::make_pair(LI, PVNI)); | |||
532 | } | |||
533 | continue; | |||
534 | } | |||
535 | ||||
536 | // Follow snippet copies. | |||
537 | MachineInstr *MI = LIS.getInstructionFromIndex(VNI->def); | |||
538 | if (!SnippetCopies.count(MI)) | |||
539 | continue; | |||
540 | LiveInterval &SnipLI = LIS.getInterval(MI->getOperand(1).getReg()); | |||
541 | assert(isRegToSpill(SnipLI.reg()) && "Unexpected register in copy")(static_cast <bool> (isRegToSpill(SnipLI.reg()) && "Unexpected register in copy") ? void (0) : __assert_fail ("isRegToSpill(SnipLI.reg()) && \"Unexpected register in copy\"" , "llvm/lib/CodeGen/InlineSpiller.cpp", 541, __extension__ __PRETTY_FUNCTION__ )); | |||
542 | VNInfo *SnipVNI = SnipLI.getVNInfoAt(VNI->def.getRegSlot(true)); | |||
543 | assert(SnipVNI && "Snippet undefined before copy")(static_cast <bool> (SnipVNI && "Snippet undefined before copy" ) ? void (0) : __assert_fail ("SnipVNI && \"Snippet undefined before copy\"" , "llvm/lib/CodeGen/InlineSpiller.cpp", 543, __extension__ __PRETTY_FUNCTION__ )); | |||
544 | WorkList.push_back(std::make_pair(&SnipLI, SnipVNI)); | |||
545 | } while (!WorkList.empty()); | |||
546 | } | |||
547 | ||||
548 | bool InlineSpiller::canGuaranteeAssignmentAfterRemat(Register VReg, | |||
549 | MachineInstr &MI) { | |||
550 | if (!RestrictStatepointRemat) | |||
551 | return true; | |||
552 | // Here's a quick explanation of the problem we're trying to handle here: | |||
553 | // * There are some pseudo instructions with more vreg uses than there are | |||
554 | // physical registers on the machine. | |||
555 | // * This is normally handled by spilling the vreg, and folding the reload | |||
556 | // into the user instruction. (Thus decreasing the number of used vregs | |||
557 | // until the remainder can be assigned to physregs.) | |||
558 | // * However, since we may try to spill vregs in any order, we can end up | |||
559 | // trying to spill each operand to the instruction, and then rematting it | |||
560 | // instead. When that happens, the new live intervals (for the remats) are | |||
561 | // expected to be trivially assignable (i.e. RS_Done). However, since we | |||
562 | // may have more remats than physregs, we're guaranteed to fail to assign | |||
563 | // one. | |||
564 | // At the moment, we only handle this for STATEPOINTs since they're the only | |||
565 | // pseudo op where we've seen this. If we start seeing other instructions | |||
566 | // with the same problem, we need to revisit this. | |||
567 | if (MI.getOpcode() != TargetOpcode::STATEPOINT) | |||
568 | return true; | |||
569 | // For STATEPOINTs we allow re-materialization for fixed arguments only hoping | |||
570 | // that number of physical registers is enough to cover all fixed arguments. | |||
571 | // If it is not true we need to revisit it. | |||
572 | for (unsigned Idx = StatepointOpers(&MI).getVarIdx(), | |||
573 | EndIdx = MI.getNumOperands(); | |||
574 | Idx < EndIdx; ++Idx) { | |||
575 | MachineOperand &MO = MI.getOperand(Idx); | |||
576 | if (MO.isReg() && MO.getReg() == VReg) | |||
577 | return false; | |||
578 | } | |||
579 | return true; | |||
580 | } | |||
581 | ||||
582 | /// reMaterializeFor - Attempt to rematerialize before MI instead of reloading. | |||
583 | bool InlineSpiller::reMaterializeFor(LiveInterval &VirtReg, MachineInstr &MI) { | |||
584 | // Analyze instruction | |||
585 | SmallVector<std::pair<MachineInstr *, unsigned>, 8> Ops; | |||
586 | VirtRegInfo RI = AnalyzeVirtRegInBundle(MI, VirtReg.reg(), &Ops); | |||
587 | ||||
588 | if (!RI.Reads) | |||
589 | return false; | |||
590 | ||||
591 | SlotIndex UseIdx = LIS.getInstructionIndex(MI).getRegSlot(true); | |||
592 | VNInfo *ParentVNI = VirtReg.getVNInfoAt(UseIdx.getBaseIndex()); | |||
593 | ||||
594 | if (!ParentVNI) { | |||
595 | LLVM_DEBUG(dbgs() << "\tadding <undef> flags: ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << "\tadding <undef> flags: " ; } } while (false); | |||
596 | for (MachineOperand &MO : MI.operands()) | |||
597 | if (MO.isReg() && MO.isUse() && MO.getReg() == VirtReg.reg()) | |||
598 | MO.setIsUndef(); | |||
599 | LLVM_DEBUG(dbgs() << UseIdx << '\t' << MI)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << UseIdx << '\t' << MI; } } while (false); | |||
600 | return true; | |||
601 | } | |||
602 | ||||
603 | if (SnippetCopies.count(&MI)) | |||
604 | return false; | |||
605 | ||||
606 | LiveInterval &OrigLI = LIS.getInterval(Original); | |||
607 | VNInfo *OrigVNI = OrigLI.getVNInfoAt(UseIdx); | |||
608 | LiveRangeEdit::Remat RM(ParentVNI); | |||
609 | RM.OrigMI = LIS.getInstructionFromIndex(OrigVNI->def); | |||
610 | ||||
611 | if (!Edit->canRematerializeAt(RM, OrigVNI, UseIdx, false)) { | |||
612 | markValueUsed(&VirtReg, ParentVNI); | |||
613 | LLVM_DEBUG(dbgs() << "\tcannot remat for " << UseIdx << '\t' << MI)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << "\tcannot remat for " << UseIdx << '\t' << MI; } } while (false); | |||
614 | return false; | |||
615 | } | |||
616 | ||||
617 | // If the instruction also writes VirtReg.reg, it had better not require the | |||
618 | // same register for uses and defs. | |||
619 | if (RI.Tied) { | |||
620 | markValueUsed(&VirtReg, ParentVNI); | |||
621 | LLVM_DEBUG(dbgs() << "\tcannot remat tied reg: " << UseIdx << '\t' << MI)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << "\tcannot remat tied reg: " << UseIdx << '\t' << MI; } } while (false); | |||
622 | return false; | |||
623 | } | |||
624 | ||||
625 | // Before rematerializing into a register for a single instruction, try to | |||
626 | // fold a load into the instruction. That avoids allocating a new register. | |||
627 | if (RM.OrigMI->canFoldAsLoad() && | |||
628 | foldMemoryOperand(Ops, RM.OrigMI)) { | |||
629 | Edit->markRematerialized(RM.ParentVNI); | |||
630 | ++NumFoldedLoads; | |||
631 | return true; | |||
632 | } | |||
633 | ||||
634 | // If we can't guarantee that we'll be able to actually assign the new vreg, | |||
635 | // we can't remat. | |||
636 | if (!canGuaranteeAssignmentAfterRemat(VirtReg.reg(), MI)) { | |||
637 | markValueUsed(&VirtReg, ParentVNI); | |||
638 | LLVM_DEBUG(dbgs() << "\tcannot remat for " << UseIdx << '\t' << MI)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << "\tcannot remat for " << UseIdx << '\t' << MI; } } while (false); | |||
639 | return false; | |||
640 | } | |||
641 | ||||
642 | // Allocate a new register for the remat. | |||
643 | Register NewVReg = Edit->createFrom(Original); | |||
644 | ||||
645 | // Finally we can rematerialize OrigMI before MI. | |||
646 | SlotIndex DefIdx = | |||
647 | Edit->rematerializeAt(*MI.getParent(), MI, NewVReg, RM, TRI); | |||
648 | ||||
649 | // We take the DebugLoc from MI, since OrigMI may be attributed to a | |||
650 | // different source location. | |||
651 | auto *NewMI = LIS.getInstructionFromIndex(DefIdx); | |||
652 | NewMI->setDebugLoc(MI.getDebugLoc()); | |||
653 | ||||
654 | (void)DefIdx; | |||
655 | LLVM_DEBUG(dbgs() << "\tremat: " << DefIdx << '\t'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << "\tremat: " << DefIdx << '\t' << *LIS.getInstructionFromIndex(DefIdx); } } while (false) | |||
656 | << *LIS.getInstructionFromIndex(DefIdx))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << "\tremat: " << DefIdx << '\t' << *LIS.getInstructionFromIndex(DefIdx); } } while (false); | |||
657 | ||||
658 | // Replace operands | |||
659 | for (const auto &OpPair : Ops) { | |||
660 | MachineOperand &MO = OpPair.first->getOperand(OpPair.second); | |||
661 | if (MO.isReg() && MO.isUse() && MO.getReg() == VirtReg.reg()) { | |||
662 | MO.setReg(NewVReg); | |||
663 | MO.setIsKill(); | |||
664 | } | |||
665 | } | |||
666 | LLVM_DEBUG(dbgs() << "\t " << UseIdx << '\t' << MI << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << "\t " << UseIdx << '\t' << MI << '\n'; } } while (false); | |||
667 | ||||
668 | ++NumRemats; | |||
669 | return true; | |||
670 | } | |||
671 | ||||
672 | /// reMaterializeAll - Try to rematerialize as many uses as possible, | |||
673 | /// and trim the live ranges after. | |||
674 | void InlineSpiller::reMaterializeAll() { | |||
675 | if (!Edit->anyRematerializable()) | |||
676 | return; | |||
677 | ||||
678 | UsedValues.clear(); | |||
679 | ||||
680 | // Try to remat before all uses of snippets. | |||
681 | bool anyRemat = false; | |||
682 | for (Register Reg : RegsToSpill) { | |||
683 | LiveInterval &LI = LIS.getInterval(Reg); | |||
684 | for (MachineInstr &MI : llvm::make_early_inc_range(MRI.reg_bundles(Reg))) { | |||
685 | // Debug values are not allowed to affect codegen. | |||
686 | if (MI.isDebugValue()) | |||
687 | continue; | |||
688 | ||||
689 | assert(!MI.isDebugInstr() && "Did not expect to find a use in debug "(static_cast <bool> (!MI.isDebugInstr() && "Did not expect to find a use in debug " "instruction that isn't a DBG_VALUE") ? void (0) : __assert_fail ("!MI.isDebugInstr() && \"Did not expect to find a use in debug \" \"instruction that isn't a DBG_VALUE\"" , "llvm/lib/CodeGen/InlineSpiller.cpp", 690, __extension__ __PRETTY_FUNCTION__ )) | |||
690 | "instruction that isn't a DBG_VALUE")(static_cast <bool> (!MI.isDebugInstr() && "Did not expect to find a use in debug " "instruction that isn't a DBG_VALUE") ? void (0) : __assert_fail ("!MI.isDebugInstr() && \"Did not expect to find a use in debug \" \"instruction that isn't a DBG_VALUE\"" , "llvm/lib/CodeGen/InlineSpiller.cpp", 690, __extension__ __PRETTY_FUNCTION__ )); | |||
691 | ||||
692 | anyRemat |= reMaterializeFor(LI, MI); | |||
693 | } | |||
694 | } | |||
695 | if (!anyRemat) | |||
696 | return; | |||
697 | ||||
698 | // Remove any values that were completely rematted. | |||
699 | for (Register Reg : RegsToSpill) { | |||
700 | LiveInterval &LI = LIS.getInterval(Reg); | |||
701 | for (VNInfo *VNI : LI.vnis()) { | |||
702 | if (VNI->isUnused() || VNI->isPHIDef() || UsedValues.count(VNI)) | |||
703 | continue; | |||
704 | MachineInstr *MI = LIS.getInstructionFromIndex(VNI->def); | |||
705 | MI->addRegisterDead(Reg, &TRI); | |||
706 | if (!MI->allDefsAreDead()) | |||
707 | continue; | |||
708 | LLVM_DEBUG(dbgs() << "All defs dead: " << *MI)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << "All defs dead: " << *MI ; } } while (false); | |||
709 | DeadDefs.push_back(MI); | |||
710 | } | |||
711 | } | |||
712 | ||||
713 | // Eliminate dead code after remat. Note that some snippet copies may be | |||
714 | // deleted here. | |||
715 | if (DeadDefs.empty()) | |||
716 | return; | |||
717 | LLVM_DEBUG(dbgs() << "Remat created " << DeadDefs.size() << " dead defs.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << "Remat created " << DeadDefs .size() << " dead defs.\n"; } } while (false); | |||
718 | Edit->eliminateDeadDefs(DeadDefs, RegsToSpill); | |||
719 | ||||
720 | // LiveRangeEdit::eliminateDeadDef is used to remove dead define instructions | |||
721 | // after rematerialization. To remove a VNI for a vreg from its LiveInterval, | |||
722 | // LiveIntervals::removeVRegDefAt is used. However, after non-PHI VNIs are all | |||
723 | // removed, PHI VNI are still left in the LiveInterval. | |||
724 | // So to get rid of unused reg, we need to check whether it has non-dbg | |||
725 | // reference instead of whether it has non-empty interval. | |||
726 | unsigned ResultPos = 0; | |||
727 | for (Register Reg : RegsToSpill) { | |||
728 | if (MRI.reg_nodbg_empty(Reg)) { | |||
729 | Edit->eraseVirtReg(Reg); | |||
730 | continue; | |||
731 | } | |||
732 | ||||
733 | assert(LIS.hasInterval(Reg) &&(static_cast <bool> (LIS.hasInterval(Reg) && (! LIS.getInterval(Reg).empty() || !MRI.reg_nodbg_empty(Reg)) && "Empty and not used live-range?!") ? void (0) : __assert_fail ("LIS.hasInterval(Reg) && (!LIS.getInterval(Reg).empty() || !MRI.reg_nodbg_empty(Reg)) && \"Empty and not used live-range?!\"" , "llvm/lib/CodeGen/InlineSpiller.cpp", 735, __extension__ __PRETTY_FUNCTION__ )) | |||
734 | (!LIS.getInterval(Reg).empty() || !MRI.reg_nodbg_empty(Reg)) &&(static_cast <bool> (LIS.hasInterval(Reg) && (! LIS.getInterval(Reg).empty() || !MRI.reg_nodbg_empty(Reg)) && "Empty and not used live-range?!") ? void (0) : __assert_fail ("LIS.hasInterval(Reg) && (!LIS.getInterval(Reg).empty() || !MRI.reg_nodbg_empty(Reg)) && \"Empty and not used live-range?!\"" , "llvm/lib/CodeGen/InlineSpiller.cpp", 735, __extension__ __PRETTY_FUNCTION__ )) | |||
735 | "Empty and not used live-range?!")(static_cast <bool> (LIS.hasInterval(Reg) && (! LIS.getInterval(Reg).empty() || !MRI.reg_nodbg_empty(Reg)) && "Empty and not used live-range?!") ? void (0) : __assert_fail ("LIS.hasInterval(Reg) && (!LIS.getInterval(Reg).empty() || !MRI.reg_nodbg_empty(Reg)) && \"Empty and not used live-range?!\"" , "llvm/lib/CodeGen/InlineSpiller.cpp", 735, __extension__ __PRETTY_FUNCTION__ )); | |||
736 | ||||
737 | RegsToSpill[ResultPos++] = Reg; | |||
738 | } | |||
739 | RegsToSpill.erase(RegsToSpill.begin() + ResultPos, RegsToSpill.end()); | |||
740 | LLVM_DEBUG(dbgs() << RegsToSpill.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << RegsToSpill.size() << " registers to spill after remat.\n" ; } } while (false) | |||
741 | << " registers to spill after remat.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << RegsToSpill.size() << " registers to spill after remat.\n" ; } } while (false); | |||
742 | } | |||
743 | ||||
744 | //===----------------------------------------------------------------------===// | |||
745 | // Spilling | |||
746 | //===----------------------------------------------------------------------===// | |||
747 | ||||
748 | /// If MI is a load or store of StackSlot, it can be removed. | |||
749 | bool InlineSpiller::coalesceStackAccess(MachineInstr *MI, Register Reg) { | |||
750 | int FI = 0; | |||
751 | Register InstrReg = TII.isLoadFromStackSlot(*MI, FI); | |||
752 | bool IsLoad = InstrReg; | |||
753 | if (!IsLoad) | |||
754 | InstrReg = TII.isStoreToStackSlot(*MI, FI); | |||
755 | ||||
756 | // We have a stack access. Is it the right register and slot? | |||
757 | if (InstrReg != Reg || FI != StackSlot) | |||
758 | return false; | |||
759 | ||||
760 | if (!IsLoad) | |||
761 | HSpiller.rmFromMergeableSpills(*MI, StackSlot); | |||
762 | ||||
763 | LLVM_DEBUG(dbgs() << "Coalescing stack access: " << *MI)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << "Coalescing stack access: " << *MI; } } while (false); | |||
764 | LIS.RemoveMachineInstrFromMaps(*MI); | |||
765 | MI->eraseFromParent(); | |||
766 | ||||
767 | if (IsLoad) { | |||
768 | ++NumReloadsRemoved; | |||
769 | --NumReloads; | |||
770 | } else { | |||
771 | ++NumSpillsRemoved; | |||
772 | --NumSpills; | |||
773 | } | |||
774 | ||||
775 | return true; | |||
776 | } | |||
777 | ||||
778 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | |||
779 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) | |||
780 | // Dump the range of instructions from B to E with their slot indexes. | |||
781 | static void dumpMachineInstrRangeWithSlotIndex(MachineBasicBlock::iterator B, | |||
782 | MachineBasicBlock::iterator E, | |||
783 | LiveIntervals const &LIS, | |||
784 | const char *const header, | |||
785 | Register VReg = Register()) { | |||
786 | char NextLine = '\n'; | |||
787 | char SlotIndent = '\t'; | |||
788 | ||||
789 | if (std::next(B) == E) { | |||
790 | NextLine = ' '; | |||
791 | SlotIndent = ' '; | |||
792 | } | |||
793 | ||||
794 | dbgs() << '\t' << header << ": " << NextLine; | |||
795 | ||||
796 | for (MachineBasicBlock::iterator I = B; I != E; ++I) { | |||
797 | SlotIndex Idx = LIS.getInstructionIndex(*I).getRegSlot(); | |||
798 | ||||
799 | // If a register was passed in and this instruction has it as a | |||
800 | // destination that is marked as an early clobber, print the | |||
801 | // early-clobber slot index. | |||
802 | if (VReg) { | |||
803 | MachineOperand *MO = I->findRegisterDefOperand(VReg); | |||
804 | if (MO && MO->isEarlyClobber()) | |||
805 | Idx = Idx.getRegSlot(true); | |||
806 | } | |||
807 | ||||
808 | dbgs() << SlotIndent << Idx << '\t' << *I; | |||
809 | } | |||
810 | } | |||
811 | #endif | |||
812 | ||||
813 | /// foldMemoryOperand - Try folding stack slot references in Ops into their | |||
814 | /// instructions. | |||
815 | /// | |||
816 | /// @param Ops Operand indices from AnalyzeVirtRegInBundle(). | |||
817 | /// @param LoadMI Load instruction to use instead of stack slot when non-null. | |||
818 | /// @return True on success. | |||
819 | bool InlineSpiller:: | |||
820 | foldMemoryOperand(ArrayRef<std::pair<MachineInstr *, unsigned>> Ops, | |||
821 | MachineInstr *LoadMI) { | |||
822 | if (Ops.empty()) | |||
823 | return false; | |||
824 | // Don't attempt folding in bundles. | |||
825 | MachineInstr *MI = Ops.front().first; | |||
826 | if (Ops.back().first != MI || MI->isBundled()) | |||
827 | return false; | |||
828 | ||||
829 | bool WasCopy = MI->isCopy(); | |||
830 | Register ImpReg; | |||
831 | ||||
832 | // TII::foldMemoryOperand will do what we need here for statepoint | |||
833 | // (fold load into use and remove corresponding def). We will replace | |||
834 | // uses of removed def with loads (spillAroundUses). | |||
835 | // For that to work we need to untie def and use to pass it through | |||
836 | // foldMemoryOperand and signal foldPatchpoint that it is allowed to | |||
837 | // fold them. | |||
838 | bool UntieRegs = MI->getOpcode() == TargetOpcode::STATEPOINT; | |||
839 | ||||
840 | // Spill subregs if the target allows it. | |||
841 | // We always want to spill subregs for stackmap/patchpoint pseudos. | |||
842 | bool SpillSubRegs = TII.isSubregFoldable() || | |||
843 | MI->getOpcode() == TargetOpcode::STATEPOINT || | |||
844 | MI->getOpcode() == TargetOpcode::PATCHPOINT || | |||
845 | MI->getOpcode() == TargetOpcode::STACKMAP; | |||
846 | ||||
847 | // TargetInstrInfo::foldMemoryOperand only expects explicit, non-tied | |||
848 | // operands. | |||
849 | SmallVector<unsigned, 8> FoldOps; | |||
850 | for (const auto &OpPair : Ops) { | |||
851 | unsigned Idx = OpPair.second; | |||
852 | assert(MI == OpPair.first && "Instruction conflict during operand folding")(static_cast <bool> (MI == OpPair.first && "Instruction conflict during operand folding" ) ? void (0) : __assert_fail ("MI == OpPair.first && \"Instruction conflict during operand folding\"" , "llvm/lib/CodeGen/InlineSpiller.cpp", 852, __extension__ __PRETTY_FUNCTION__ )); | |||
853 | MachineOperand &MO = MI->getOperand(Idx); | |||
854 | ||||
855 | // No point restoring an undef read, and we'll produce an invalid live | |||
856 | // interval. | |||
857 | // TODO: Is this really the correct way to handle undef tied uses? | |||
858 | if (MO.isUse() && !MO.readsReg() && !MO.isTied()) | |||
859 | continue; | |||
860 | ||||
861 | if (MO.isImplicit()) { | |||
862 | ImpReg = MO.getReg(); | |||
863 | continue; | |||
864 | } | |||
865 | ||||
866 | if (!SpillSubRegs && MO.getSubReg()) | |||
867 | return false; | |||
868 | // We cannot fold a load instruction into a def. | |||
869 | if (LoadMI && MO.isDef()) | |||
870 | return false; | |||
871 | // Tied use operands should not be passed to foldMemoryOperand. | |||
872 | if (UntieRegs || !MI->isRegTiedToDefOperand(Idx)) | |||
873 | FoldOps.push_back(Idx); | |||
874 | } | |||
875 | ||||
876 | // If we only have implicit uses, we won't be able to fold that. | |||
877 | // Moreover, TargetInstrInfo::foldMemoryOperand will assert if we try! | |||
878 | if (FoldOps.empty()) | |||
879 | return false; | |||
880 | ||||
881 | MachineInstrSpan MIS(MI, MI->getParent()); | |||
882 | ||||
883 | SmallVector<std::pair<unsigned, unsigned> > TiedOps; | |||
884 | if (UntieRegs) | |||
885 | for (unsigned Idx : FoldOps) { | |||
886 | MachineOperand &MO = MI->getOperand(Idx); | |||
887 | if (!MO.isTied()) | |||
888 | continue; | |||
889 | unsigned Tied = MI->findTiedOperandIdx(Idx); | |||
890 | if (MO.isUse()) | |||
891 | TiedOps.emplace_back(Tied, Idx); | |||
892 | else { | |||
893 | assert(MO.isDef() && "Tied to not use and def?")(static_cast <bool> (MO.isDef() && "Tied to not use and def?" ) ? void (0) : __assert_fail ("MO.isDef() && \"Tied to not use and def?\"" , "llvm/lib/CodeGen/InlineSpiller.cpp", 893, __extension__ __PRETTY_FUNCTION__ )); | |||
894 | TiedOps.emplace_back(Idx, Tied); | |||
895 | } | |||
896 | MI->untieRegOperand(Idx); | |||
897 | } | |||
898 | ||||
899 | MachineInstr *FoldMI = | |||
900 | LoadMI ? TII.foldMemoryOperand(*MI, FoldOps, *LoadMI, &LIS) | |||
901 | : TII.foldMemoryOperand(*MI, FoldOps, StackSlot, &LIS, &VRM); | |||
902 | if (!FoldMI) { | |||
903 | // Re-tie operands. | |||
904 | for (auto Tied : TiedOps) | |||
905 | MI->tieOperands(Tied.first, Tied.second); | |||
906 | return false; | |||
907 | } | |||
908 | ||||
909 | // Remove LIS for any dead defs in the original MI not in FoldMI. | |||
910 | for (MIBundleOperands MO(*MI); MO.isValid(); ++MO) { | |||
911 | if (!MO->isReg()) | |||
912 | continue; | |||
913 | Register Reg = MO->getReg(); | |||
914 | if (!Reg || Reg.isVirtual() || MRI.isReserved(Reg)) { | |||
915 | continue; | |||
916 | } | |||
917 | // Skip non-Defs, including undef uses and internal reads. | |||
918 | if (MO->isUse()) | |||
919 | continue; | |||
920 | PhysRegInfo RI = AnalyzePhysRegInBundle(*FoldMI, Reg, &TRI); | |||
921 | if (RI.FullyDefined) | |||
922 | continue; | |||
923 | // FoldMI does not define this physreg. Remove the LI segment. | |||
924 | assert(MO->isDead() && "Cannot fold physreg def")(static_cast <bool> (MO->isDead() && "Cannot fold physreg def" ) ? void (0) : __assert_fail ("MO->isDead() && \"Cannot fold physreg def\"" , "llvm/lib/CodeGen/InlineSpiller.cpp", 924, __extension__ __PRETTY_FUNCTION__ )); | |||
925 | SlotIndex Idx = LIS.getInstructionIndex(*MI).getRegSlot(); | |||
926 | LIS.removePhysRegDefAt(Reg.asMCReg(), Idx); | |||
927 | } | |||
928 | ||||
929 | int FI; | |||
930 | if (TII.isStoreToStackSlot(*MI, FI) && | |||
931 | HSpiller.rmFromMergeableSpills(*MI, FI)) | |||
932 | --NumSpills; | |||
933 | LIS.ReplaceMachineInstrInMaps(*MI, *FoldMI); | |||
934 | // Update the call site info. | |||
935 | if (MI->isCandidateForCallSiteEntry()) | |||
936 | MI->getMF()->moveCallSiteInfo(MI, FoldMI); | |||
937 | ||||
938 | // If we've folded a store into an instruction labelled with debug-info, | |||
939 | // record a substitution from the old operand to the memory operand. Handle | |||
940 | // the simple common case where operand 0 is the one being folded, plus when | |||
941 | // the destination operand is also a tied def. More values could be | |||
942 | // substituted / preserved with more analysis. | |||
943 | if (MI->peekDebugInstrNum() && Ops[0].second == 0) { | |||
944 | // Helper lambda. | |||
945 | auto MakeSubstitution = [this,FoldMI,MI,&Ops]() { | |||
946 | // Substitute old operand zero to the new instructions memory operand. | |||
947 | unsigned OldOperandNum = Ops[0].second; | |||
948 | unsigned NewNum = FoldMI->getDebugInstrNum(); | |||
949 | unsigned OldNum = MI->getDebugInstrNum(); | |||
950 | MF.makeDebugValueSubstitution({OldNum, OldOperandNum}, | |||
951 | {NewNum, MachineFunction::DebugOperandMemNumber}); | |||
952 | }; | |||
953 | ||||
954 | const MachineOperand &Op0 = MI->getOperand(Ops[0].second); | |||
955 | if (Ops.size() == 1 && Op0.isDef()) { | |||
956 | MakeSubstitution(); | |||
957 | } else if (Ops.size() == 2 && Op0.isDef() && MI->getOperand(1).isTied() && | |||
958 | Op0.getReg() == MI->getOperand(1).getReg()) { | |||
959 | MakeSubstitution(); | |||
960 | } | |||
961 | } else if (MI->peekDebugInstrNum()) { | |||
962 | // This is a debug-labelled instruction, but the operand being folded isn't | |||
963 | // at operand zero. Most likely this means it's a load being folded in. | |||
964 | // Substitute any register defs from operand zero up to the one being | |||
965 | // folded -- past that point, we don't know what the new operand indexes | |||
966 | // will be. | |||
967 | MF.substituteDebugValuesForInst(*MI, *FoldMI, Ops[0].second); | |||
968 | } | |||
969 | ||||
970 | MI->eraseFromParent(); | |||
971 | ||||
972 | // Insert any new instructions other than FoldMI into the LIS maps. | |||
973 | assert(!MIS.empty() && "Unexpected empty span of instructions!")(static_cast <bool> (!MIS.empty() && "Unexpected empty span of instructions!" ) ? void (0) : __assert_fail ("!MIS.empty() && \"Unexpected empty span of instructions!\"" , "llvm/lib/CodeGen/InlineSpiller.cpp", 973, __extension__ __PRETTY_FUNCTION__ )); | |||
974 | for (MachineInstr &MI : MIS) | |||
975 | if (&MI != FoldMI) | |||
976 | LIS.InsertMachineInstrInMaps(MI); | |||
977 | ||||
978 | // TII.foldMemoryOperand may have left some implicit operands on the | |||
979 | // instruction. Strip them. | |||
980 | if (ImpReg) | |||
981 | for (unsigned i = FoldMI->getNumOperands(); i; --i) { | |||
982 | MachineOperand &MO = FoldMI->getOperand(i - 1); | |||
983 | if (!MO.isReg() || !MO.isImplicit()) | |||
984 | break; | |||
985 | if (MO.getReg() == ImpReg) | |||
986 | FoldMI->removeOperand(i - 1); | |||
987 | } | |||
988 | ||||
989 | LLVM_DEBUG(dumpMachineInstrRangeWithSlotIndex(MIS.begin(), MIS.end(), LIS,do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dumpMachineInstrRangeWithSlotIndex(MIS.begin( ), MIS.end(), LIS, "folded"); } } while (false) | |||
990 | "folded"))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dumpMachineInstrRangeWithSlotIndex(MIS.begin( ), MIS.end(), LIS, "folded"); } } while (false); | |||
991 | ||||
992 | if (!WasCopy) | |||
993 | ++NumFolded; | |||
994 | else if (Ops.front().second == 0) { | |||
995 | ++NumSpills; | |||
996 | // If there is only 1 store instruction is required for spill, add it | |||
997 | // to mergeable list. In X86 AMX, 2 intructions are required to store. | |||
998 | // We disable the merge for this case. | |||
999 | if (std::distance(MIS.begin(), MIS.end()) <= 1) | |||
1000 | HSpiller.addToMergeableSpills(*FoldMI, StackSlot, Original); | |||
1001 | } else | |||
1002 | ++NumReloads; | |||
1003 | return true; | |||
1004 | } | |||
1005 | ||||
1006 | void InlineSpiller::insertReload(Register NewVReg, | |||
1007 | SlotIndex Idx, | |||
1008 | MachineBasicBlock::iterator MI) { | |||
1009 | MachineBasicBlock &MBB = *MI->getParent(); | |||
1010 | ||||
1011 | MachineInstrSpan MIS(MI, &MBB); | |||
1012 | TII.loadRegFromStackSlot(MBB, MI, NewVReg, StackSlot, | |||
1013 | MRI.getRegClass(NewVReg), &TRI, Register()); | |||
1014 | ||||
1015 | LIS.InsertMachineInstrRangeInMaps(MIS.begin(), MI); | |||
1016 | ||||
1017 | LLVM_DEBUG(dumpMachineInstrRangeWithSlotIndex(MIS.begin(), MI, LIS, "reload",do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dumpMachineInstrRangeWithSlotIndex(MIS.begin( ), MI, LIS, "reload", NewVReg); } } while (false) | |||
1018 | NewVReg))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dumpMachineInstrRangeWithSlotIndex(MIS.begin( ), MI, LIS, "reload", NewVReg); } } while (false); | |||
1019 | ++NumReloads; | |||
1020 | } | |||
1021 | ||||
1022 | /// Check if \p Def fully defines a VReg with an undefined value. | |||
1023 | /// If that's the case, that means the value of VReg is actually | |||
1024 | /// not relevant. | |||
1025 | static bool isRealSpill(const MachineInstr &Def) { | |||
1026 | if (!Def.isImplicitDef()) | |||
1027 | return true; | |||
1028 | assert(Def.getNumOperands() == 1 &&(static_cast <bool> (Def.getNumOperands() == 1 && "Implicit def with more than one definition") ? void (0) : __assert_fail ("Def.getNumOperands() == 1 && \"Implicit def with more than one definition\"" , "llvm/lib/CodeGen/InlineSpiller.cpp", 1029, __extension__ __PRETTY_FUNCTION__ )) | |||
1029 | "Implicit def with more than one definition")(static_cast <bool> (Def.getNumOperands() == 1 && "Implicit def with more than one definition") ? void (0) : __assert_fail ("Def.getNumOperands() == 1 && \"Implicit def with more than one definition\"" , "llvm/lib/CodeGen/InlineSpiller.cpp", 1029, __extension__ __PRETTY_FUNCTION__ )); | |||
1030 | // We can say that the VReg defined by Def is undef, only if it is | |||
1031 | // fully defined by Def. Otherwise, some of the lanes may not be | |||
1032 | // undef and the value of the VReg matters. | |||
1033 | return Def.getOperand(0).getSubReg(); | |||
1034 | } | |||
1035 | ||||
1036 | /// insertSpill - Insert a spill of NewVReg after MI. | |||
1037 | void InlineSpiller::insertSpill(Register NewVReg, bool isKill, | |||
1038 | MachineBasicBlock::iterator MI) { | |||
1039 | // Spill are not terminators, so inserting spills after terminators will | |||
1040 | // violate invariants in MachineVerifier. | |||
1041 | assert(!MI->isTerminator() && "Inserting a spill after a terminator")(static_cast <bool> (!MI->isTerminator() && "Inserting a spill after a terminator" ) ? void (0) : __assert_fail ("!MI->isTerminator() && \"Inserting a spill after a terminator\"" , "llvm/lib/CodeGen/InlineSpiller.cpp", 1041, __extension__ __PRETTY_FUNCTION__ )); | |||
1042 | MachineBasicBlock &MBB = *MI->getParent(); | |||
1043 | ||||
1044 | MachineInstrSpan MIS(MI, &MBB); | |||
1045 | MachineBasicBlock::iterator SpillBefore = std::next(MI); | |||
1046 | bool IsRealSpill = isRealSpill(*MI); | |||
1047 | ||||
1048 | if (IsRealSpill) | |||
1049 | TII.storeRegToStackSlot(MBB, SpillBefore, NewVReg, isKill, StackSlot, | |||
1050 | MRI.getRegClass(NewVReg), &TRI, Register()); | |||
1051 | else | |||
1052 | // Don't spill undef value. | |||
1053 | // Anything works for undef, in particular keeping the memory | |||
1054 | // uninitialized is a viable option and it saves code size and | |||
1055 | // run time. | |||
1056 | BuildMI(MBB, SpillBefore, MI->getDebugLoc(), TII.get(TargetOpcode::KILL)) | |||
1057 | .addReg(NewVReg, getKillRegState(isKill)); | |||
1058 | ||||
1059 | MachineBasicBlock::iterator Spill = std::next(MI); | |||
1060 | LIS.InsertMachineInstrRangeInMaps(Spill, MIS.end()); | |||
1061 | for (const MachineInstr &MI : make_range(Spill, MIS.end())) | |||
1062 | getVDefInterval(MI, LIS); | |||
1063 | ||||
1064 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dumpMachineInstrRangeWithSlotIndex(Spill, MIS .end(), LIS, "spill"); } } while (false) | |||
1065 | dumpMachineInstrRangeWithSlotIndex(Spill, MIS.end(), LIS, "spill"))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dumpMachineInstrRangeWithSlotIndex(Spill, MIS .end(), LIS, "spill"); } } while (false); | |||
1066 | ++NumSpills; | |||
1067 | // If there is only 1 store instruction is required for spill, add it | |||
1068 | // to mergeable list. In X86 AMX, 2 intructions are required to store. | |||
1069 | // We disable the merge for this case. | |||
1070 | if (IsRealSpill && std::distance(Spill, MIS.end()) <= 1) | |||
1071 | HSpiller.addToMergeableSpills(*Spill, StackSlot, Original); | |||
1072 | } | |||
1073 | ||||
1074 | /// spillAroundUses - insert spill code around each use of Reg. | |||
1075 | void InlineSpiller::spillAroundUses(Register Reg) { | |||
1076 | LLVM_DEBUG(dbgs() << "spillAroundUses " << printReg(Reg) << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << "spillAroundUses " << printReg (Reg) << '\n'; } } while (false); | |||
1077 | LiveInterval &OldLI = LIS.getInterval(Reg); | |||
1078 | ||||
1079 | // Iterate over instructions using Reg. | |||
1080 | for (MachineInstr &MI : llvm::make_early_inc_range(MRI.reg_bundles(Reg))) { | |||
1081 | // Debug values are not allowed to affect codegen. | |||
1082 | if (MI.isDebugValue()) { | |||
1083 | // Modify DBG_VALUE now that the value is in a spill slot. | |||
1084 | MachineBasicBlock *MBB = MI.getParent(); | |||
1085 | LLVM_DEBUG(dbgs() << "Modifying debug info due to spill:\t" << MI)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << "Modifying debug info due to spill:\t" << MI; } } while (false); | |||
1086 | buildDbgValueForSpill(*MBB, &MI, MI, StackSlot, Reg); | |||
1087 | MBB->erase(MI); | |||
1088 | continue; | |||
1089 | } | |||
1090 | ||||
1091 | assert(!MI.isDebugInstr() && "Did not expect to find a use in debug "(static_cast <bool> (!MI.isDebugInstr() && "Did not expect to find a use in debug " "instruction that isn't a DBG_VALUE") ? void (0) : __assert_fail ("!MI.isDebugInstr() && \"Did not expect to find a use in debug \" \"instruction that isn't a DBG_VALUE\"" , "llvm/lib/CodeGen/InlineSpiller.cpp", 1092, __extension__ __PRETTY_FUNCTION__ )) | |||
1092 | "instruction that isn't a DBG_VALUE")(static_cast <bool> (!MI.isDebugInstr() && "Did not expect to find a use in debug " "instruction that isn't a DBG_VALUE") ? void (0) : __assert_fail ("!MI.isDebugInstr() && \"Did not expect to find a use in debug \" \"instruction that isn't a DBG_VALUE\"" , "llvm/lib/CodeGen/InlineSpiller.cpp", 1092, __extension__ __PRETTY_FUNCTION__ )); | |||
1093 | ||||
1094 | // Ignore copies to/from snippets. We'll delete them. | |||
1095 | if (SnippetCopies.count(&MI)) | |||
1096 | continue; | |||
1097 | ||||
1098 | // Stack slot accesses may coalesce away. | |||
1099 | if (coalesceStackAccess(&MI, Reg)) | |||
1100 | continue; | |||
1101 | ||||
1102 | // Analyze instruction. | |||
1103 | SmallVector<std::pair<MachineInstr*, unsigned>, 8> Ops; | |||
1104 | VirtRegInfo RI = AnalyzeVirtRegInBundle(MI, Reg, &Ops); | |||
1105 | ||||
1106 | // Find the slot index where this instruction reads and writes OldLI. | |||
1107 | // This is usually the def slot, except for tied early clobbers. | |||
1108 | SlotIndex Idx = LIS.getInstructionIndex(MI).getRegSlot(); | |||
1109 | if (VNInfo *VNI = OldLI.getVNInfoAt(Idx.getRegSlot(true))) | |||
1110 | if (SlotIndex::isSameInstr(Idx, VNI->def)) | |||
1111 | Idx = VNI->def; | |||
1112 | ||||
1113 | // Check for a sibling copy. | |||
1114 | Register SibReg = isFullCopyOf(MI, Reg); | |||
1115 | if (SibReg && isSibling(SibReg)) { | |||
1116 | // This may actually be a copy between snippets. | |||
1117 | if (isRegToSpill(SibReg)) { | |||
1118 | LLVM_DEBUG(dbgs() << "Found new snippet copy: " << MI)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << "Found new snippet copy: " << MI; } } while (false); | |||
1119 | SnippetCopies.insert(&MI); | |||
1120 | continue; | |||
1121 | } | |||
1122 | if (RI.Writes) { | |||
1123 | if (hoistSpillInsideBB(OldLI, MI)) { | |||
1124 | // This COPY is now dead, the value is already in the stack slot. | |||
1125 | MI.getOperand(0).setIsDead(); | |||
1126 | DeadDefs.push_back(&MI); | |||
1127 | continue; | |||
1128 | } | |||
1129 | } else { | |||
1130 | // This is a reload for a sib-reg copy. Drop spills downstream. | |||
1131 | LiveInterval &SibLI = LIS.getInterval(SibReg); | |||
1132 | eliminateRedundantSpills(SibLI, SibLI.getVNInfoAt(Idx)); | |||
1133 | // The COPY will fold to a reload below. | |||
1134 | } | |||
1135 | } | |||
1136 | ||||
1137 | // Attempt to fold memory ops. | |||
1138 | if (foldMemoryOperand(Ops)) | |||
1139 | continue; | |||
1140 | ||||
1141 | // Create a new virtual register for spill/fill. | |||
1142 | // FIXME: Infer regclass from instruction alone. | |||
1143 | Register NewVReg = Edit->createFrom(Reg); | |||
1144 | ||||
1145 | if (RI.Reads) | |||
1146 | insertReload(NewVReg, Idx, &MI); | |||
1147 | ||||
1148 | // Rewrite instruction operands. | |||
1149 | bool hasLiveDef = false; | |||
1150 | for (const auto &OpPair : Ops) { | |||
1151 | MachineOperand &MO = OpPair.first->getOperand(OpPair.second); | |||
1152 | MO.setReg(NewVReg); | |||
1153 | if (MO.isUse()) { | |||
1154 | if (!OpPair.first->isRegTiedToDefOperand(OpPair.second)) | |||
1155 | MO.setIsKill(); | |||
1156 | } else { | |||
1157 | if (!MO.isDead()) | |||
1158 | hasLiveDef = true; | |||
1159 | } | |||
1160 | } | |||
1161 | LLVM_DEBUG(dbgs() << "\trewrite: " << Idx << '\t' << MI << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << "\trewrite: " << Idx << '\t' << MI << '\n'; } } while (false); | |||
1162 | ||||
1163 | // FIXME: Use a second vreg if instruction has no tied ops. | |||
1164 | if (RI.Writes) | |||
1165 | if (hasLiveDef) | |||
1166 | insertSpill(NewVReg, true, &MI); | |||
1167 | } | |||
1168 | } | |||
1169 | ||||
1170 | /// spillAll - Spill all registers remaining after rematerialization. | |||
1171 | void InlineSpiller::spillAll() { | |||
1172 | // Update LiveStacks now that we are committed to spilling. | |||
1173 | if (StackSlot == VirtRegMap::NO_STACK_SLOT) { | |||
1174 | StackSlot = VRM.assignVirt2StackSlot(Original); | |||
1175 | StackInt = &LSS.getOrCreateInterval(StackSlot, MRI.getRegClass(Original)); | |||
1176 | StackInt->getNextValue(SlotIndex(), LSS.getVNInfoAllocator()); | |||
1177 | } else | |||
1178 | StackInt = &LSS.getInterval(StackSlot); | |||
1179 | ||||
1180 | if (Original != Edit->getReg()) | |||
1181 | VRM.assignVirt2StackSlot(Edit->getReg(), StackSlot); | |||
1182 | ||||
1183 | assert(StackInt->getNumValNums() == 1 && "Bad stack interval values")(static_cast <bool> (StackInt->getNumValNums() == 1 && "Bad stack interval values") ? void (0) : __assert_fail ("StackInt->getNumValNums() == 1 && \"Bad stack interval values\"" , "llvm/lib/CodeGen/InlineSpiller.cpp", 1183, __extension__ __PRETTY_FUNCTION__ )); | |||
1184 | for (Register Reg : RegsToSpill) | |||
1185 | StackInt->MergeSegmentsInAsValue(LIS.getInterval(Reg), | |||
1186 | StackInt->getValNumInfo(0)); | |||
1187 | LLVM_DEBUG(dbgs() << "Merged spilled regs: " << *StackInt << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << "Merged spilled regs: " << *StackInt << '\n'; } } while (false); | |||
1188 | ||||
1189 | // Spill around uses of all RegsToSpill. | |||
1190 | for (Register Reg : RegsToSpill) | |||
1191 | spillAroundUses(Reg); | |||
1192 | ||||
1193 | // Hoisted spills may cause dead code. | |||
1194 | if (!DeadDefs.empty()) { | |||
1195 | LLVM_DEBUG(dbgs() << "Eliminating " << DeadDefs.size() << " dead defs\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << "Eliminating " << DeadDefs .size() << " dead defs\n"; } } while (false); | |||
1196 | Edit->eliminateDeadDefs(DeadDefs, RegsToSpill); | |||
1197 | } | |||
1198 | ||||
1199 | // Finally delete the SnippetCopies. | |||
1200 | for (Register Reg : RegsToSpill) { | |||
1201 | for (MachineInstr &MI : | |||
1202 | llvm::make_early_inc_range(MRI.reg_instructions(Reg))) { | |||
1203 | assert(SnippetCopies.count(&MI) && "Remaining use wasn't a snippet copy")(static_cast <bool> (SnippetCopies.count(&MI) && "Remaining use wasn't a snippet copy") ? void (0) : __assert_fail ("SnippetCopies.count(&MI) && \"Remaining use wasn't a snippet copy\"" , "llvm/lib/CodeGen/InlineSpiller.cpp", 1203, __extension__ __PRETTY_FUNCTION__ )); | |||
1204 | // FIXME: Do this with a LiveRangeEdit callback. | |||
1205 | LIS.RemoveMachineInstrFromMaps(MI); | |||
1206 | MI.eraseFromParent(); | |||
1207 | } | |||
1208 | } | |||
1209 | ||||
1210 | // Delete all spilled registers. | |||
1211 | for (Register Reg : RegsToSpill) | |||
1212 | Edit->eraseVirtReg(Reg); | |||
1213 | } | |||
1214 | ||||
1215 | void InlineSpiller::spill(LiveRangeEdit &edit) { | |||
1216 | ++NumSpilledRanges; | |||
1217 | Edit = &edit; | |||
1218 | assert(!Register::isStackSlot(edit.getReg()) &&(static_cast <bool> (!Register::isStackSlot(edit.getReg ()) && "Trying to spill a stack slot.") ? void (0) : __assert_fail ("!Register::isStackSlot(edit.getReg()) && \"Trying to spill a stack slot.\"" , "llvm/lib/CodeGen/InlineSpiller.cpp", 1219, __extension__ __PRETTY_FUNCTION__ )) | |||
1219 | "Trying to spill a stack slot.")(static_cast <bool> (!Register::isStackSlot(edit.getReg ()) && "Trying to spill a stack slot.") ? void (0) : __assert_fail ("!Register::isStackSlot(edit.getReg()) && \"Trying to spill a stack slot.\"" , "llvm/lib/CodeGen/InlineSpiller.cpp", 1219, __extension__ __PRETTY_FUNCTION__ )); | |||
1220 | // Share a stack slot among all descendants of Original. | |||
1221 | Original = VRM.getOriginal(edit.getReg()); | |||
1222 | StackSlot = VRM.getStackSlot(Original); | |||
1223 | StackInt = nullptr; | |||
1224 | ||||
1225 | LLVM_DEBUG(dbgs() << "Inline spilling "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << "Inline spilling " << TRI .getRegClassName(MRI.getRegClass(edit.getReg())) << ':' << edit.getParent() << "\nFrom original " << printReg(Original) << '\n'; } } while (false) | |||
1226 | << TRI.getRegClassName(MRI.getRegClass(edit.getReg()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << "Inline spilling " << TRI .getRegClassName(MRI.getRegClass(edit.getReg())) << ':' << edit.getParent() << "\nFrom original " << printReg(Original) << '\n'; } } while (false) | |||
1227 | << ':' << edit.getParent() << "\nFrom original "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << "Inline spilling " << TRI .getRegClassName(MRI.getRegClass(edit.getReg())) << ':' << edit.getParent() << "\nFrom original " << printReg(Original) << '\n'; } } while (false) | |||
1228 | << printReg(Original) << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << "Inline spilling " << TRI .getRegClassName(MRI.getRegClass(edit.getReg())) << ':' << edit.getParent() << "\nFrom original " << printReg(Original) << '\n'; } } while (false); | |||
1229 | assert(edit.getParent().isSpillable() &&(static_cast <bool> (edit.getParent().isSpillable() && "Attempting to spill already spilled value.") ? void (0) : __assert_fail ("edit.getParent().isSpillable() && \"Attempting to spill already spilled value.\"" , "llvm/lib/CodeGen/InlineSpiller.cpp", 1230, __extension__ __PRETTY_FUNCTION__ )) | |||
1230 | "Attempting to spill already spilled value.")(static_cast <bool> (edit.getParent().isSpillable() && "Attempting to spill already spilled value.") ? void (0) : __assert_fail ("edit.getParent().isSpillable() && \"Attempting to spill already spilled value.\"" , "llvm/lib/CodeGen/InlineSpiller.cpp", 1230, __extension__ __PRETTY_FUNCTION__ )); | |||
1231 | assert(DeadDefs.empty() && "Previous spill didn't remove dead defs")(static_cast <bool> (DeadDefs.empty() && "Previous spill didn't remove dead defs" ) ? void (0) : __assert_fail ("DeadDefs.empty() && \"Previous spill didn't remove dead defs\"" , "llvm/lib/CodeGen/InlineSpiller.cpp", 1231, __extension__ __PRETTY_FUNCTION__ )); | |||
1232 | ||||
1233 | collectRegsToSpill(); | |||
1234 | reMaterializeAll(); | |||
1235 | ||||
1236 | // Remat may handle everything. | |||
1237 | if (!RegsToSpill.empty()) | |||
1238 | spillAll(); | |||
1239 | ||||
1240 | Edit->calculateRegClassAndHint(MF, VRAI); | |||
1241 | } | |||
1242 | ||||
1243 | /// Optimizations after all the reg selections and spills are done. | |||
1244 | void InlineSpiller::postOptimization() { HSpiller.hoistAllSpills(); } | |||
| ||||
1245 | ||||
1246 | /// When a spill is inserted, add the spill to MergeableSpills map. | |||
1247 | void HoistSpillHelper::addToMergeableSpills(MachineInstr &Spill, int StackSlot, | |||
1248 | unsigned Original) { | |||
1249 | BumpPtrAllocator &Allocator = LIS.getVNInfoAllocator(); | |||
1250 | LiveInterval &OrigLI = LIS.getInterval(Original); | |||
1251 | // save a copy of LiveInterval in StackSlotToOrigLI because the original | |||
1252 | // LiveInterval may be cleared after all its references are spilled. | |||
1253 | if (StackSlotToOrigLI.find(StackSlot) == StackSlotToOrigLI.end()) { | |||
1254 | auto LI = std::make_unique<LiveInterval>(OrigLI.reg(), OrigLI.weight()); | |||
1255 | LI->assign(OrigLI, Allocator); | |||
1256 | StackSlotToOrigLI[StackSlot] = std::move(LI); | |||
1257 | } | |||
1258 | SlotIndex Idx = LIS.getInstructionIndex(Spill); | |||
1259 | VNInfo *OrigVNI = StackSlotToOrigLI[StackSlot]->getVNInfoAt(Idx.getRegSlot()); | |||
1260 | std::pair<int, VNInfo *> MIdx = std::make_pair(StackSlot, OrigVNI); | |||
1261 | MergeableSpills[MIdx].insert(&Spill); | |||
1262 | } | |||
1263 | ||||
1264 | /// When a spill is removed, remove the spill from MergeableSpills map. | |||
1265 | /// Return true if the spill is removed successfully. | |||
1266 | bool HoistSpillHelper::rmFromMergeableSpills(MachineInstr &Spill, | |||
1267 | int StackSlot) { | |||
1268 | auto It = StackSlotToOrigLI.find(StackSlot); | |||
1269 | if (It == StackSlotToOrigLI.end()) | |||
1270 | return false; | |||
1271 | SlotIndex Idx = LIS.getInstructionIndex(Spill); | |||
1272 | VNInfo *OrigVNI = It->second->getVNInfoAt(Idx.getRegSlot()); | |||
1273 | std::pair<int, VNInfo *> MIdx = std::make_pair(StackSlot, OrigVNI); | |||
1274 | return MergeableSpills[MIdx].erase(&Spill); | |||
1275 | } | |||
1276 | ||||
1277 | /// Check BB to see if it is a possible target BB to place a hoisted spill, | |||
1278 | /// i.e., there should be a living sibling of OrigReg at the insert point. | |||
1279 | bool HoistSpillHelper::isSpillCandBB(LiveInterval &OrigLI, VNInfo &OrigVNI, | |||
1280 | MachineBasicBlock &BB, Register &LiveReg) { | |||
1281 | SlotIndex Idx = IPA.getLastInsertPoint(OrigLI, BB); | |||
1282 | // The original def could be after the last insert point in the root block, | |||
1283 | // we can't hoist to here. | |||
1284 | if (Idx < OrigVNI.def) { | |||
1285 | // TODO: We could be better here. If LI is not alive in landing pad | |||
1286 | // we could hoist spill after LIP. | |||
1287 | LLVM_DEBUG(dbgs() << "can't spill in root block - def after LIP\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << "can't spill in root block - def after LIP\n" ; } } while (false); | |||
1288 | return false; | |||
1289 | } | |||
1290 | Register OrigReg = OrigLI.reg(); | |||
1291 | SmallSetVector<Register, 16> &Siblings = Virt2SiblingsMap[OrigReg]; | |||
1292 | assert(OrigLI.getVNInfoAt(Idx) == &OrigVNI && "Unexpected VNI")(static_cast <bool> (OrigLI.getVNInfoAt(Idx) == &OrigVNI && "Unexpected VNI") ? void (0) : __assert_fail ("OrigLI.getVNInfoAt(Idx) == &OrigVNI && \"Unexpected VNI\"" , "llvm/lib/CodeGen/InlineSpiller.cpp", 1292, __extension__ __PRETTY_FUNCTION__ )); | |||
1293 | ||||
1294 | for (const Register &SibReg : Siblings) { | |||
1295 | LiveInterval &LI = LIS.getInterval(SibReg); | |||
1296 | VNInfo *VNI = LI.getVNInfoAt(Idx); | |||
1297 | if (VNI) { | |||
1298 | LiveReg = SibReg; | |||
1299 | return true; | |||
1300 | } | |||
1301 | } | |||
1302 | return false; | |||
1303 | } | |||
1304 | ||||
1305 | /// Remove redundant spills in the same BB. Save those redundant spills in | |||
1306 | /// SpillsToRm, and save the spill to keep and its BB in SpillBBToSpill map. | |||
1307 | void HoistSpillHelper::rmRedundantSpills( | |||
1308 | SmallPtrSet<MachineInstr *, 16> &Spills, | |||
1309 | SmallVectorImpl<MachineInstr *> &SpillsToRm, | |||
1310 | DenseMap<MachineDomTreeNode *, MachineInstr *> &SpillBBToSpill) { | |||
1311 | // For each spill saw, check SpillBBToSpill[] and see if its BB already has | |||
1312 | // another spill inside. If a BB contains more than one spill, only keep the | |||
1313 | // earlier spill with smaller SlotIndex. | |||
1314 | for (auto *const CurrentSpill : Spills) { | |||
1315 | MachineBasicBlock *Block = CurrentSpill->getParent(); | |||
1316 | MachineDomTreeNode *Node = MDT.getBase().getNode(Block); | |||
1317 | MachineInstr *PrevSpill = SpillBBToSpill[Node]; | |||
1318 | if (PrevSpill) { | |||
1319 | SlotIndex PIdx = LIS.getInstructionIndex(*PrevSpill); | |||
1320 | SlotIndex CIdx = LIS.getInstructionIndex(*CurrentSpill); | |||
1321 | MachineInstr *SpillToRm = (CIdx > PIdx) ? CurrentSpill : PrevSpill; | |||
1322 | MachineInstr *SpillToKeep = (CIdx > PIdx) ? PrevSpill : CurrentSpill; | |||
1323 | SpillsToRm.push_back(SpillToRm); | |||
1324 | SpillBBToSpill[MDT.getBase().getNode(Block)] = SpillToKeep; | |||
1325 | } else { | |||
1326 | SpillBBToSpill[MDT.getBase().getNode(Block)] = CurrentSpill; | |||
1327 | } | |||
1328 | } | |||
1329 | for (auto *const SpillToRm : SpillsToRm) | |||
1330 | Spills.erase(SpillToRm); | |||
1331 | } | |||
1332 | ||||
1333 | /// Starting from \p Root find a top-down traversal order of the dominator | |||
1334 | /// tree to visit all basic blocks containing the elements of \p Spills. | |||
1335 | /// Redundant spills will be found and put into \p SpillsToRm at the same | |||
1336 | /// time. \p SpillBBToSpill will be populated as part of the process and | |||
1337 | /// maps a basic block to the first store occurring in the basic block. | |||
1338 | /// \post SpillsToRm.union(Spills\@post) == Spills\@pre | |||
1339 | void HoistSpillHelper::getVisitOrders( | |||
1340 | MachineBasicBlock *Root, SmallPtrSet<MachineInstr *, 16> &Spills, | |||
1341 | SmallVectorImpl<MachineDomTreeNode *> &Orders, | |||
1342 | SmallVectorImpl<MachineInstr *> &SpillsToRm, | |||
1343 | DenseMap<MachineDomTreeNode *, unsigned> &SpillsToKeep, | |||
1344 | DenseMap<MachineDomTreeNode *, MachineInstr *> &SpillBBToSpill) { | |||
1345 | // The set contains all the possible BB nodes to which we may hoist | |||
1346 | // original spills. | |||
1347 | SmallPtrSet<MachineDomTreeNode *, 8> WorkSet; | |||
1348 | // Save the BB nodes on the path from the first BB node containing | |||
1349 | // non-redundant spill to the Root node. | |||
1350 | SmallPtrSet<MachineDomTreeNode *, 8> NodesOnPath; | |||
1351 | // All the spills to be hoisted must originate from a single def instruction | |||
1352 | // to the OrigReg. It means the def instruction should dominate all the spills | |||
1353 | // to be hoisted. We choose the BB where the def instruction is located as | |||
1354 | // the Root. | |||
1355 | MachineDomTreeNode *RootIDomNode = MDT[Root]->getIDom(); | |||
1356 | // For every node on the dominator tree with spill, walk up on the dominator | |||
1357 | // tree towards the Root node until it is reached. If there is other node | |||
1358 | // containing spill in the middle of the path, the previous spill saw will | |||
1359 | // be redundant and the node containing it will be removed. All the nodes on | |||
1360 | // the path starting from the first node with non-redundant spill to the Root | |||
1361 | // node will be added to the WorkSet, which will contain all the possible | |||
1362 | // locations where spills may be hoisted to after the loop below is done. | |||
1363 | for (auto *const Spill : Spills) { | |||
1364 | MachineBasicBlock *Block = Spill->getParent(); | |||
1365 | MachineDomTreeNode *Node = MDT[Block]; | |||
1366 | MachineInstr *SpillToRm = nullptr; | |||
1367 | while (Node != RootIDomNode) { | |||
1368 | // If Node dominates Block, and it already contains a spill, the spill in | |||
1369 | // Block will be redundant. | |||
1370 | if (Node != MDT[Block] && SpillBBToSpill[Node]) { | |||
1371 | SpillToRm = SpillBBToSpill[MDT[Block]]; | |||
1372 | break; | |||
1373 | /// If we see the Node already in WorkSet, the path from the Node to | |||
1374 | /// the Root node must already be traversed by another spill. | |||
1375 | /// Then no need to repeat. | |||
1376 | } else if (WorkSet.count(Node)) { | |||
1377 | break; | |||
1378 | } else { | |||
1379 | NodesOnPath.insert(Node); | |||
1380 | } | |||
1381 | Node = Node->getIDom(); | |||
| ||||
1382 | } | |||
1383 | if (SpillToRm) { | |||
1384 | SpillsToRm.push_back(SpillToRm); | |||
1385 | } else { | |||
1386 | // Add a BB containing the original spills to SpillsToKeep -- i.e., | |||
1387 | // set the initial status before hoisting start. The value of BBs | |||
1388 | // containing original spills is set to 0, in order to descriminate | |||
1389 | // with BBs containing hoisted spills which will be inserted to | |||
1390 | // SpillsToKeep later during hoisting. | |||
1391 | SpillsToKeep[MDT[Block]] = 0; | |||
1392 | WorkSet.insert(NodesOnPath.begin(), NodesOnPath.end()); | |||
1393 | } | |||
1394 | NodesOnPath.clear(); | |||
1395 | } | |||
1396 | ||||
1397 | // Sort the nodes in WorkSet in top-down order and save the nodes | |||
1398 | // in Orders. Orders will be used for hoisting in runHoistSpills. | |||
1399 | unsigned idx = 0; | |||
1400 | Orders.push_back(MDT.getBase().getNode(Root)); | |||
1401 | do { | |||
1402 | MachineDomTreeNode *Node = Orders[idx++]; | |||
1403 | for (MachineDomTreeNode *Child : Node->children()) { | |||
1404 | if (WorkSet.count(Child)) | |||
1405 | Orders.push_back(Child); | |||
1406 | } | |||
1407 | } while (idx != Orders.size()); | |||
1408 | assert(Orders.size() == WorkSet.size() &&(static_cast <bool> (Orders.size() == WorkSet.size() && "Orders have different size with WorkSet") ? void (0) : __assert_fail ("Orders.size() == WorkSet.size() && \"Orders have different size with WorkSet\"" , "llvm/lib/CodeGen/InlineSpiller.cpp", 1409, __extension__ __PRETTY_FUNCTION__ )) | |||
1409 | "Orders have different size with WorkSet")(static_cast <bool> (Orders.size() == WorkSet.size() && "Orders have different size with WorkSet") ? void (0) : __assert_fail ("Orders.size() == WorkSet.size() && \"Orders have different size with WorkSet\"" , "llvm/lib/CodeGen/InlineSpiller.cpp", 1409, __extension__ __PRETTY_FUNCTION__ )); | |||
1410 | ||||
1411 | #ifndef NDEBUG | |||
1412 | LLVM_DEBUG(dbgs() << "Orders size is " << Orders.size() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << "Orders size is " << Orders .size() << "\n"; } } while (false); | |||
1413 | SmallVector<MachineDomTreeNode *, 32>::reverse_iterator RIt = Orders.rbegin(); | |||
1414 | for (; RIt != Orders.rend(); RIt++) | |||
1415 | LLVM_DEBUG(dbgs() << "BB" << (*RIt)->getBlock()->getNumber() << ",")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << "BB" << (*RIt)->getBlock ()->getNumber() << ","; } } while (false); | |||
1416 | LLVM_DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { dbgs() << "\n"; } } while (false); | |||
1417 | #endif | |||
1418 | } | |||
1419 | ||||
1420 | /// Try to hoist spills according to BB hotness. The spills to removed will | |||
1421 | /// be saved in \p SpillsToRm. The spills to be inserted will be saved in | |||
1422 | /// \p SpillsToIns. | |||
1423 | void HoistSpillHelper::runHoistSpills( | |||
1424 | LiveInterval &OrigLI, VNInfo &OrigVNI, | |||
1425 | SmallPtrSet<MachineInstr *, 16> &Spills, | |||
1426 | SmallVectorImpl<MachineInstr *> &SpillsToRm, | |||
1427 | DenseMap<MachineBasicBlock *, unsigned> &SpillsToIns) { | |||
1428 | // Visit order of dominator tree nodes. | |||
1429 | SmallVector<MachineDomTreeNode *, 32> Orders; | |||
1430 | // SpillsToKeep contains all the nodes where spills are to be inserted | |||
1431 | // during hoisting. If the spill to be inserted is an original spill | |||
1432 | // (not a hoisted one), the value of the map entry is 0. If the spill | |||
1433 | // is a hoisted spill, the value of the map entry is the VReg to be used | |||
1434 | // as the source of the spill. | |||
1435 | DenseMap<MachineDomTreeNode *, unsigned> SpillsToKeep; | |||
1436 | // Map from BB to the first spill inside of it. | |||
1437 | DenseMap<MachineDomTreeNode *, MachineInstr *> SpillBBToSpill; | |||
1438 | ||||
1439 | rmRedundantSpills(Spills, SpillsToRm, SpillBBToSpill); | |||
1440 | ||||
1441 | MachineBasicBlock *Root = LIS.getMBBFromIndex(OrigVNI.def); | |||
1442 | getVisitOrders(Root, Spills, Orders, SpillsToRm, SpillsToKeep, | |||
1443 | SpillBBToSpill); | |||
1444 | ||||
1445 | // SpillsInSubTreeMap keeps the map from a dom tree node to a pair of | |||
1446 | // nodes set and the cost of all the spills inside those nodes. | |||
1447 | // The nodes set are the locations where spills are to be inserted | |||
1448 | // in the subtree of current node. | |||
1449 | using NodesCostPair = | |||
1450 | std::pair<SmallPtrSet<MachineDomTreeNode *, 16>, BlockFrequency>; | |||
1451 | DenseMap<MachineDomTreeNode *, NodesCostPair> SpillsInSubTreeMap; | |||
1452 | ||||
1453 | // Iterate Orders set in reverse order, which will be a bottom-up order | |||
1454 | // in the dominator tree. Once we visit a dom tree node, we know its | |||
1455 | // children have already been visited and the spill locations in the | |||
1456 | // subtrees of all the children have been determined. | |||
1457 | SmallVector<MachineDomTreeNode *, 32>::reverse_iterator RIt = Orders.rbegin(); | |||
1458 | for (; RIt != Orders.rend(); RIt++) { | |||
1459 | MachineBasicBlock *Block = (*RIt)->getBlock(); | |||
1460 | ||||
1461 | // If Block contains an original spill, simply continue. | |||
1462 | if (SpillsToKeep.find(*RIt) != SpillsToKeep.end() && !SpillsToKeep[*RIt]) { | |||
1463 | SpillsInSubTreeMap[*RIt].first.insert(*RIt); | |||
1464 | // SpillsInSubTreeMap[*RIt].second contains the cost of spill. | |||
1465 | SpillsInSubTreeMap[*RIt].second = MBFI.getBlockFreq(Block); | |||
1466 | continue; | |||
1467 | } | |||
1468 | ||||
1469 | // Collect spills in subtree of current node (*RIt) to | |||
1470 | // SpillsInSubTreeMap[*RIt].first. | |||
1471 | for (MachineDomTreeNode *Child : (*RIt)->children()) { | |||
1472 | if (SpillsInSubTreeMap.find(Child) == SpillsInSubTreeMap.end()) | |||
1473 | continue; | |||
1474 | // The stmt "SpillsInSubTree = SpillsInSubTreeMap[*RIt].first" below | |||
1475 | // should be placed before getting the begin and end iterators of | |||
1476 | // SpillsInSubTreeMap[Child].first, or else the iterators may be | |||
1477 | // invalidated when SpillsInSubTreeMap[*RIt] is seen the first time | |||
1478 | // and the map grows and then the original buckets in the map are moved. | |||
1479 | SmallPtrSet<MachineDomTreeNode *, 16> &SpillsInSubTree = | |||
1480 | SpillsInSubTreeMap[*RIt].first; | |||
1481 | BlockFrequency &SubTreeCost = SpillsInSubTreeMap[*RIt].second; | |||
1482 | SubTreeCost += SpillsInSubTreeMap[Child].second; | |||
1483 | auto BI = SpillsInSubTreeMap[Child].first.begin(); | |||
1484 | auto EI = SpillsInSubTreeMap[Child].first.end(); | |||
1485 | SpillsInSubTree.insert(BI, EI); | |||
1486 | SpillsInSubTreeMap.erase(Child); | |||
1487 | } | |||
1488 | ||||
1489 | SmallPtrSet<MachineDomTreeNode *, 16> &SpillsInSubTree = | |||
1490 | SpillsInSubTreeMap[*RIt].first; | |||
1491 | BlockFrequency &SubTreeCost = SpillsInSubTreeMap[*RIt].second; | |||
1492 | // No spills in subtree, simply continue. | |||
1493 | if (SpillsInSubTree.empty()) | |||
1494 | continue; | |||
1495 | ||||
1496 | // Check whether Block is a possible candidate to insert spill. | |||
1497 | Register LiveReg; | |||
1498 | if (!isSpillCandBB(OrigLI, OrigVNI, *Block, LiveReg)) | |||
1499 | continue; | |||
1500 | ||||
1501 | // If there are multiple spills that could be merged, bias a little | |||
1502 | // to hoist the spill. | |||
1503 | BranchProbability MarginProb = (SpillsInSubTree.size() > 1) | |||
1504 | ? BranchProbability(9, 10) | |||
1505 | : BranchProbability(1, 1); | |||
1506 | if (SubTreeCost > MBFI.getBlockFreq(Block) * MarginProb) { | |||
1507 | // Hoist: Move spills to current Block. | |||
1508 | for (auto *const SpillBB : SpillsInSubTree) { | |||
1509 | // When SpillBB is a BB contains original spill, insert the spill | |||
1510 | // to SpillsToRm. | |||
1511 | if (SpillsToKeep.find(SpillBB) != SpillsToKeep.end() && | |||
1512 | !SpillsToKeep[SpillBB]) { | |||
1513 | MachineInstr *SpillToRm = SpillBBToSpill[SpillBB]; | |||
1514 | SpillsToRm.push_back(SpillToRm); | |||
1515 | } | |||
1516 | // SpillBB will not contain spill anymore, remove it from SpillsToKeep. | |||
1517 | SpillsToKeep.erase(SpillBB); | |||
1518 | } | |||
1519 | // Current Block is the BB containing the new hoisted spill. Add it to | |||
1520 | // SpillsToKeep. LiveReg is the source of the new spill. | |||
1521 | SpillsToKeep[*RIt] = LiveReg; | |||
1522 | LLVM_DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { { dbgs() << "spills in BB: "; for (const auto Rspill : SpillsInSubTree) dbgs() << Rspill->getBlock ()->getNumber() << " "; dbgs() << "were promoted to BB" << (*RIt)->getBlock()->getNumber() << "\n" ; }; } } while (false) | |||
1523 | dbgs() << "spills in BB: ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { { dbgs() << "spills in BB: "; for (const auto Rspill : SpillsInSubTree) dbgs() << Rspill->getBlock ()->getNumber() << " "; dbgs() << "were promoted to BB" << (*RIt)->getBlock()->getNumber() << "\n" ; }; } } while (false) | |||
1524 | for (const auto Rspill : SpillsInSubTree)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { { dbgs() << "spills in BB: "; for (const auto Rspill : SpillsInSubTree) dbgs() << Rspill->getBlock ()->getNumber() << " "; dbgs() << "were promoted to BB" << (*RIt)->getBlock()->getNumber() << "\n" ; }; } } while (false) | |||
1525 | dbgs() << Rspill->getBlock()->getNumber() << " ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { { dbgs() << "spills in BB: "; for (const auto Rspill : SpillsInSubTree) dbgs() << Rspill->getBlock ()->getNumber() << " "; dbgs() << "were promoted to BB" << (*RIt)->getBlock()->getNumber() << "\n" ; }; } } while (false) | |||
1526 | dbgs() << "were promoted to BB" << (*RIt)->getBlock()->getNumber()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { { dbgs() << "spills in BB: "; for (const auto Rspill : SpillsInSubTree) dbgs() << Rspill->getBlock ()->getNumber() << " "; dbgs() << "were promoted to BB" << (*RIt)->getBlock()->getNumber() << "\n" ; }; } } while (false) | |||
1527 | << "\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { { dbgs() << "spills in BB: "; for (const auto Rspill : SpillsInSubTree) dbgs() << Rspill->getBlock ()->getNumber() << " "; dbgs() << "were promoted to BB" << (*RIt)->getBlock()->getNumber() << "\n" ; }; } } while (false) | |||
1528 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { { dbgs() << "spills in BB: "; for (const auto Rspill : SpillsInSubTree) dbgs() << Rspill->getBlock ()->getNumber() << " "; dbgs() << "were promoted to BB" << (*RIt)->getBlock()->getNumber() << "\n" ; }; } } while (false); | |||
1529 | SpillsInSubTree.clear(); | |||
1530 | SpillsInSubTree.insert(*RIt); | |||
1531 | SubTreeCost = MBFI.getBlockFreq(Block); | |||
1532 | } | |||
1533 | } | |||
1534 | // For spills in SpillsToKeep with LiveReg set (i.e., not original spill), | |||
1535 | // save them to SpillsToIns. | |||
1536 | for (const auto &Ent : SpillsToKeep) { | |||
1537 | if (Ent.second) | |||
1538 | SpillsToIns[Ent.first->getBlock()] = Ent.second; | |||
1539 | } | |||
1540 | } | |||
1541 | ||||
1542 | /// For spills with equal values, remove redundant spills and hoist those left | |||
1543 | /// to less hot spots. | |||
1544 | /// | |||
1545 | /// Spills with equal values will be collected into the same set in | |||
1546 | /// MergeableSpills when spill is inserted. These equal spills are originated | |||
1547 | /// from the same defining instruction and are dominated by the instruction. | |||
1548 | /// Before hoisting all the equal spills, redundant spills inside in the same | |||
1549 | /// BB are first marked to be deleted. Then starting from the spills left, walk | |||
1550 | /// up on the dominator tree towards the Root node where the define instruction | |||
1551 | /// is located, mark the dominated spills to be deleted along the way and | |||
1552 | /// collect the BB nodes on the path from non-dominated spills to the define | |||
1553 | /// instruction into a WorkSet. The nodes in WorkSet are the candidate places | |||
1554 | /// where we are considering to hoist the spills. We iterate the WorkSet in | |||
1555 | /// bottom-up order, and for each node, we will decide whether to hoist spills | |||
1556 | /// inside its subtree to that node. In this way, we can get benefit locally | |||
1557 | /// even if hoisting all the equal spills to one cold place is impossible. | |||
1558 | void HoistSpillHelper::hoistAllSpills() { | |||
1559 | SmallVector<Register, 4> NewVRegs; | |||
1560 | LiveRangeEdit Edit(nullptr, NewVRegs, MF, LIS, &VRM, this); | |||
1561 | ||||
1562 | for (unsigned i = 0, e = MRI.getNumVirtRegs(); i != e; ++i) { | |||
1563 | Register Reg = Register::index2VirtReg(i); | |||
1564 | Register Original = VRM.getPreSplitReg(Reg); | |||
1565 | if (!MRI.def_empty(Reg)) | |||
1566 | Virt2SiblingsMap[Original].insert(Reg); | |||
1567 | } | |||
1568 | ||||
1569 | // Each entry in MergeableSpills contains a spill set with equal values. | |||
1570 | for (auto &Ent : MergeableSpills) { | |||
1571 | int Slot = Ent.first.first; | |||
1572 | LiveInterval &OrigLI = *StackSlotToOrigLI[Slot]; | |||
1573 | VNInfo *OrigVNI = Ent.first.second; | |||
1574 | SmallPtrSet<MachineInstr *, 16> &EqValSpills = Ent.second; | |||
1575 | if (Ent.second.empty()) | |||
1576 | continue; | |||
1577 | ||||
1578 | LLVM_DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { { dbgs() << "\nFor Slot" << Slot << " and VN" << OrigVNI->id << ":\n" << "Equal spills in BB: " ; for (const auto spill : EqValSpills) dbgs() << spill-> getParent()->getNumber() << " "; dbgs() << "\n" ; }; } } while (false) | |||
1579 | dbgs() << "\nFor Slot" << Slot << " and VN" << OrigVNI->id << ":\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { { dbgs() << "\nFor Slot" << Slot << " and VN" << OrigVNI->id << ":\n" << "Equal spills in BB: " ; for (const auto spill : EqValSpills) dbgs() << spill-> getParent()->getNumber() << " "; dbgs() << "\n" ; }; } } while (false) | |||
1580 | << "Equal spills in BB: ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { { dbgs() << "\nFor Slot" << Slot << " and VN" << OrigVNI->id << ":\n" << "Equal spills in BB: " ; for (const auto spill : EqValSpills) dbgs() << spill-> getParent()->getNumber() << " "; dbgs() << "\n" ; }; } } while (false) | |||
1581 | for (const auto spill : EqValSpills)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { { dbgs() << "\nFor Slot" << Slot << " and VN" << OrigVNI->id << ":\n" << "Equal spills in BB: " ; for (const auto spill : EqValSpills) dbgs() << spill-> getParent()->getNumber() << " "; dbgs() << "\n" ; }; } } while (false) | |||
1582 | dbgs() << spill->getParent()->getNumber() << " ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { { dbgs() << "\nFor Slot" << Slot << " and VN" << OrigVNI->id << ":\n" << "Equal spills in BB: " ; for (const auto spill : EqValSpills) dbgs() << spill-> getParent()->getNumber() << " "; dbgs() << "\n" ; }; } } while (false) | |||
1583 | dbgs() << "\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { { dbgs() << "\nFor Slot" << Slot << " and VN" << OrigVNI->id << ":\n" << "Equal spills in BB: " ; for (const auto spill : EqValSpills) dbgs() << spill-> getParent()->getNumber() << " "; dbgs() << "\n" ; }; } } while (false) | |||
1584 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { { dbgs() << "\nFor Slot" << Slot << " and VN" << OrigVNI->id << ":\n" << "Equal spills in BB: " ; for (const auto spill : EqValSpills) dbgs() << spill-> getParent()->getNumber() << " "; dbgs() << "\n" ; }; } } while (false); | |||
1585 | ||||
1586 | // SpillsToRm is the spill set to be removed from EqValSpills. | |||
1587 | SmallVector<MachineInstr *, 16> SpillsToRm; | |||
1588 | // SpillsToIns is the spill set to be newly inserted after hoisting. | |||
1589 | DenseMap<MachineBasicBlock *, unsigned> SpillsToIns; | |||
1590 | ||||
1591 | runHoistSpills(OrigLI, *OrigVNI, EqValSpills, SpillsToRm, SpillsToIns); | |||
1592 | ||||
1593 | LLVM_DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { { dbgs() << "Finally inserted spills in BB: " ; for (const auto &Ispill : SpillsToIns) dbgs() << Ispill .first->getNumber() << " "; dbgs() << "\nFinally removed spills in BB: " ; for (const auto Rspill : SpillsToRm) dbgs() << Rspill ->getParent()->getNumber() << " "; dbgs() << "\n"; }; } } while (false) | |||
1594 | dbgs() << "Finally inserted spills in BB: ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { { dbgs() << "Finally inserted spills in BB: " ; for (const auto &Ispill : SpillsToIns) dbgs() << Ispill .first->getNumber() << " "; dbgs() << "\nFinally removed spills in BB: " ; for (const auto Rspill : SpillsToRm) dbgs() << Rspill ->getParent()->getNumber() << " "; dbgs() << "\n"; }; } } while (false) | |||
1595 | for (const auto &Ispill : SpillsToIns)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { { dbgs() << "Finally inserted spills in BB: " ; for (const auto &Ispill : SpillsToIns) dbgs() << Ispill .first->getNumber() << " "; dbgs() << "\nFinally removed spills in BB: " ; for (const auto Rspill : SpillsToRm) dbgs() << Rspill ->getParent()->getNumber() << " "; dbgs() << "\n"; }; } } while (false) | |||
1596 | dbgs() << Ispill.first->getNumber() << " ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { { dbgs() << "Finally inserted spills in BB: " ; for (const auto &Ispill : SpillsToIns) dbgs() << Ispill .first->getNumber() << " "; dbgs() << "\nFinally removed spills in BB: " ; for (const auto Rspill : SpillsToRm) dbgs() << Rspill ->getParent()->getNumber() << " "; dbgs() << "\n"; }; } } while (false) | |||
1597 | dbgs() << "\nFinally removed spills in BB: ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { { dbgs() << "Finally inserted spills in BB: " ; for (const auto &Ispill : SpillsToIns) dbgs() << Ispill .first->getNumber() << " "; dbgs() << "\nFinally removed spills in BB: " ; for (const auto Rspill : SpillsToRm) dbgs() << Rspill ->getParent()->getNumber() << " "; dbgs() << "\n"; }; } } while (false) | |||
1598 | for (const auto Rspill : SpillsToRm)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { { dbgs() << "Finally inserted spills in BB: " ; for (const auto &Ispill : SpillsToIns) dbgs() << Ispill .first->getNumber() << " "; dbgs() << "\nFinally removed spills in BB: " ; for (const auto Rspill : SpillsToRm) dbgs() << Rspill ->getParent()->getNumber() << " "; dbgs() << "\n"; }; } } while (false) | |||
1599 | dbgs() << Rspill->getParent()->getNumber() << " ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { { dbgs() << "Finally inserted spills in BB: " ; for (const auto &Ispill : SpillsToIns) dbgs() << Ispill .first->getNumber() << " "; dbgs() << "\nFinally removed spills in BB: " ; for (const auto Rspill : SpillsToRm) dbgs() << Rspill ->getParent()->getNumber() << " "; dbgs() << "\n"; }; } } while (false) | |||
1600 | dbgs() << "\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { { dbgs() << "Finally inserted spills in BB: " ; for (const auto &Ispill : SpillsToIns) dbgs() << Ispill .first->getNumber() << " "; dbgs() << "\nFinally removed spills in BB: " ; for (const auto Rspill : SpillsToRm) dbgs() << Rspill ->getParent()->getNumber() << " "; dbgs() << "\n"; }; } } while (false) | |||
1601 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("regalloc")) { { dbgs() << "Finally inserted spills in BB: " ; for (const auto &Ispill : SpillsToIns) dbgs() << Ispill .first->getNumber() << " "; dbgs() << "\nFinally removed spills in BB: " ; for (const auto Rspill : SpillsToRm) dbgs() << Rspill ->getParent()->getNumber() << " "; dbgs() << "\n"; }; } } while (false); | |||
1602 | ||||
1603 | // Stack live range update. | |||
1604 | LiveInterval &StackIntvl = LSS.getInterval(Slot); | |||
1605 | if (!SpillsToIns.empty() || !SpillsToRm.empty()) | |||
1606 | StackIntvl.MergeValueInAsValue(OrigLI, OrigVNI, | |||
1607 | StackIntvl.getValNumInfo(0)); | |||
1608 | ||||
1609 | // Insert hoisted spills. | |||
1610 | for (auto const &Insert : SpillsToIns) { | |||
1611 | MachineBasicBlock *BB = Insert.first; | |||
1612 | Register LiveReg = Insert.second; | |||
1613 | MachineBasicBlock::iterator MII = IPA.getLastInsertPointIter(OrigLI, *BB); | |||
1614 | MachineInstrSpan MIS(MII, BB); | |||
1615 | TII.storeRegToStackSlot(*BB, MII, LiveReg, false, Slot, | |||
1616 | MRI.getRegClass(LiveReg), &TRI, Register()); | |||
1617 | LIS.InsertMachineInstrRangeInMaps(MIS.begin(), MII); | |||
1618 | for (const MachineInstr &MI : make_range(MIS.begin(), MII)) | |||
1619 | getVDefInterval(MI, LIS); | |||
1620 | ++NumSpills; | |||
1621 | } | |||
1622 | ||||
1623 | // Remove redundant spills or change them to dead instructions. | |||
1624 | NumSpills -= SpillsToRm.size(); | |||
1625 | for (auto *const RMEnt : SpillsToRm) { | |||
1626 | RMEnt->setDesc(TII.get(TargetOpcode::KILL)); | |||
1627 | for (unsigned i = RMEnt->getNumOperands(); i; --i) { | |||
1628 | MachineOperand &MO = RMEnt->getOperand(i - 1); | |||
1629 | if (MO.isReg() && MO.isImplicit() && MO.isDef() && !MO.isDead()) | |||
1630 | RMEnt->removeOperand(i - 1); | |||
1631 | } | |||
1632 | } | |||
1633 | Edit.eliminateDeadDefs(SpillsToRm, std::nullopt); | |||
1634 | } | |||
1635 | } | |||
1636 | ||||
1637 | /// For VirtReg clone, the \p New register should have the same physreg or | |||
1638 | /// stackslot as the \p old register. | |||
1639 | void HoistSpillHelper::LRE_DidCloneVirtReg(Register New, Register Old) { | |||
1640 | if (VRM.hasPhys(Old)) | |||
1641 | VRM.assignVirt2Phys(New, VRM.getPhys(Old)); | |||
1642 | else if (VRM.getStackSlot(Old) != VirtRegMap::NO_STACK_SLOT) | |||
1643 | VRM.assignVirt2StackSlot(New, VRM.getStackSlot(Old)); | |||
1644 | else | |||
1645 | llvm_unreachable("VReg should be assigned either physreg or stackslot")::llvm::llvm_unreachable_internal("VReg should be assigned either physreg or stackslot" , "llvm/lib/CodeGen/InlineSpiller.cpp", 1645); | |||
1646 | if (VRM.hasShape(Old)) | |||
1647 | VRM.assignVirt2Shape(New, VRM.getShape(Old)); | |||
1648 | } |
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/CodeGen/MachineInstrBundleIterator.h" |
23 | #include "llvm/Support/GenericDomTree.h" |
24 | #include "llvm/Support/GenericDomTreeConstruction.h" |
25 | #include <cassert> |
26 | #include <memory> |
27 | |
28 | namespace llvm { |
29 | class AnalysisUsage; |
30 | class MachineFunction; |
31 | class Module; |
32 | class raw_ostream; |
33 | |
34 | template <> |
35 | inline void DominatorTreeBase<MachineBasicBlock, false>::addRoot( |
36 | MachineBasicBlock *MBB) { |
37 | this->Roots.push_back(MBB); |
38 | } |
39 | |
40 | extern template class DomTreeNodeBase<MachineBasicBlock>; |
41 | extern template class DominatorTreeBase<MachineBasicBlock, false>; // DomTree |
42 | extern template class DominatorTreeBase<MachineBasicBlock, true>; // PostDomTree |
43 | |
44 | using MachineDomTree = DomTreeBase<MachineBasicBlock>; |
45 | using MachineDomTreeNode = DomTreeNodeBase<MachineBasicBlock>; |
46 | |
47 | //===------------------------------------- |
48 | /// DominatorTree Class - Concrete subclass of DominatorTreeBase that is used to |
49 | /// compute a normal dominator tree. |
50 | /// |
51 | class MachineDominatorTree : public MachineFunctionPass { |
52 | /// Helper structure used to hold all the basic blocks |
53 | /// involved in the split of a critical edge. |
54 | struct CriticalEdge { |
55 | MachineBasicBlock *FromBB; |
56 | MachineBasicBlock *ToBB; |
57 | MachineBasicBlock *NewBB; |
58 | }; |
59 | |
60 | /// Pile up all the critical edges to be split. |
61 | /// The splitting of a critical edge is local and thus, it is possible |
62 | /// to apply several of those changes at the same time. |
63 | mutable SmallVector<CriticalEdge, 32> CriticalEdgesToSplit; |
64 | |
65 | /// Remember all the basic blocks that are inserted during |
66 | /// edge splitting. |
67 | /// Invariant: NewBBs == all the basic blocks contained in the NewBB |
68 | /// field of all the elements of CriticalEdgesToSplit. |
69 | /// I.e., forall elt in CriticalEdgesToSplit, it exists BB in NewBBs |
70 | /// such as BB == elt.NewBB. |
71 | mutable SmallSet<MachineBasicBlock *, 32> NewBBs; |
72 | |
73 | /// The DominatorTreeBase that is used to compute a normal dominator tree. |
74 | std::unique_ptr<MachineDomTree> DT; |
75 | |
76 | /// Apply all the recorded critical edges to the DT. |
77 | /// This updates the underlying DT information in a way that uses |
78 | /// the fast query path of DT as much as possible. |
79 | /// |
80 | /// \post CriticalEdgesToSplit.empty(). |
81 | void applySplitCriticalEdges() const; |
82 | |
83 | public: |
84 | static char ID; // Pass ID, replacement for typeid |
85 | |
86 | MachineDominatorTree(); |
87 | explicit MachineDominatorTree(MachineFunction &MF) : MachineFunctionPass(ID) { |
88 | calculate(MF); |
89 | } |
90 | |
91 | MachineDomTree &getBase() { |
92 | if (!DT) |
93 | DT.reset(new MachineDomTree()); |
94 | applySplitCriticalEdges(); |
95 | return *DT; |
96 | } |
97 | |
98 | void getAnalysisUsage(AnalysisUsage &AU) const override; |
99 | |
100 | MachineBasicBlock *getRoot() const { |
101 | applySplitCriticalEdges(); |
102 | return DT->getRoot(); |
103 | } |
104 | |
105 | MachineDomTreeNode *getRootNode() const { |
106 | applySplitCriticalEdges(); |
107 | return DT->getRootNode(); |
108 | } |
109 | |
110 | bool runOnMachineFunction(MachineFunction &F) override; |
111 | |
112 | void calculate(MachineFunction &F); |
113 | |
114 | bool dominates(const MachineDomTreeNode *A, |
115 | const MachineDomTreeNode *B) const { |
116 | applySplitCriticalEdges(); |
117 | return DT->dominates(A, B); |
118 | } |
119 | |
120 | void getDescendants(MachineBasicBlock *A, |
121 | SmallVectorImpl<MachineBasicBlock *> &Result) { |
122 | applySplitCriticalEdges(); |
123 | DT->getDescendants(A, Result); |
124 | } |
125 | |
126 | bool dominates(const MachineBasicBlock *A, const MachineBasicBlock *B) const { |
127 | applySplitCriticalEdges(); |
128 | return DT->dominates(A, B); |
129 | } |
130 | |
131 | // dominates - Return true if A dominates B. This performs the |
132 | // special checks necessary if A and B are in the same basic block. |
133 | bool dominates(const MachineInstr *A, const MachineInstr *B) const { |
134 | applySplitCriticalEdges(); |
135 | const MachineBasicBlock *BBA = A->getParent(), *BBB = B->getParent(); |
136 | if (BBA != BBB) return DT->dominates(BBA, BBB); |
137 | |
138 | // Loop through the basic block until we find A or B. |
139 | MachineBasicBlock::const_iterator I = BBA->begin(); |
140 | for (; &*I != A && &*I != B; ++I) |
141 | /*empty*/ ; |
142 | |
143 | return &*I == A; |
144 | } |
145 | |
146 | bool properlyDominates(const MachineDomTreeNode *A, |
147 | const MachineDomTreeNode *B) const { |
148 | applySplitCriticalEdges(); |
149 | return DT->properlyDominates(A, B); |
150 | } |
151 | |
152 | bool properlyDominates(const MachineBasicBlock *A, |
153 | const MachineBasicBlock *B) const { |
154 | applySplitCriticalEdges(); |
155 | return DT->properlyDominates(A, B); |
156 | } |
157 | |
158 | /// findNearestCommonDominator - Find nearest common dominator basic block |
159 | /// for basic block A and B. If there is no such block then return NULL. |
160 | MachineBasicBlock *findNearestCommonDominator(MachineBasicBlock *A, |
161 | MachineBasicBlock *B) { |
162 | applySplitCriticalEdges(); |
163 | return DT->findNearestCommonDominator(A, B); |
164 | } |
165 | |
166 | MachineDomTreeNode *operator[](MachineBasicBlock *BB) const { |
167 | applySplitCriticalEdges(); |
168 | return DT->getNode(BB); |
169 | } |
170 | |
171 | /// getNode - return the (Post)DominatorTree node for the specified basic |
172 | /// block. This is the same as using operator[] on this class. |
173 | /// |
174 | MachineDomTreeNode *getNode(MachineBasicBlock *BB) const { |
175 | applySplitCriticalEdges(); |
176 | return DT->getNode(BB); |
177 | } |
178 | |
179 | /// addNewBlock - Add a new node to the dominator tree information. This |
180 | /// creates a new node as a child of DomBB dominator node,linking it into |
181 | /// the children list of the immediate dominator. |
182 | MachineDomTreeNode *addNewBlock(MachineBasicBlock *BB, |
183 | MachineBasicBlock *DomBB) { |
184 | applySplitCriticalEdges(); |
185 | return DT->addNewBlock(BB, DomBB); |
186 | } |
187 | |
188 | /// changeImmediateDominator - This method is used to update the dominator |
189 | /// tree information when a node's immediate dominator changes. |
190 | /// |
191 | void changeImmediateDominator(MachineBasicBlock *N, |
192 | MachineBasicBlock *NewIDom) { |
193 | applySplitCriticalEdges(); |
194 | DT->changeImmediateDominator(N, NewIDom); |
195 | } |
196 | |
197 | void changeImmediateDominator(MachineDomTreeNode *N, |
198 | MachineDomTreeNode *NewIDom) { |
199 | applySplitCriticalEdges(); |
200 | DT->changeImmediateDominator(N, NewIDom); |
201 | } |
202 | |
203 | /// eraseNode - Removes a node from the dominator tree. Block must not |
204 | /// dominate any other blocks. Removes node from its immediate dominator's |
205 | /// children list. Deletes dominator node associated with basic block BB. |
206 | void eraseNode(MachineBasicBlock *BB) { |
207 | applySplitCriticalEdges(); |
208 | DT->eraseNode(BB); |
209 | } |
210 | |
211 | /// splitBlock - BB is split and now it has one successor. Update dominator |
212 | /// tree to reflect this change. |
213 | void splitBlock(MachineBasicBlock* NewBB) { |
214 | applySplitCriticalEdges(); |
215 | DT->splitBlock(NewBB); |
216 | } |
217 | |
218 | /// isReachableFromEntry - Return true if A is dominated by the entry |
219 | /// block of the function containing it. |
220 | bool isReachableFromEntry(const MachineBasicBlock *A) { |
221 | applySplitCriticalEdges(); |
222 | return DT->isReachableFromEntry(A); |
223 | } |
224 | |
225 | void releaseMemory() override; |
226 | |
227 | void verifyAnalysis() const override; |
228 | |
229 | void print(raw_ostream &OS, const Module*) const override; |
230 | |
231 | /// Record that the critical edge (FromBB, ToBB) has been |
232 | /// split with NewBB. |
233 | /// This is best to use this method instead of directly update the |
234 | /// underlying information, because this helps mitigating the |
235 | /// number of time the DT information is invalidated. |
236 | /// |
237 | /// \note Do not use this method with regular edges. |
238 | /// |
239 | /// \note To benefit from the compile time improvement incurred by this |
240 | /// method, the users of this method have to limit the queries to the DT |
241 | /// interface between two edges splitting. In other words, they have to |
242 | /// pack the splitting of critical edges as much as possible. |
243 | void recordSplitCriticalEdge(MachineBasicBlock *FromBB, |
244 | MachineBasicBlock *ToBB, |
245 | MachineBasicBlock *NewBB) { |
246 | bool Inserted = NewBBs.insert(NewBB).second; |
247 | (void)Inserted; |
248 | assert(Inserted &&(static_cast <bool> (Inserted && "A basic block inserted via edge splitting cannot appear twice" ) ? void (0) : __assert_fail ("Inserted && \"A basic block inserted via edge splitting cannot appear twice\"" , "llvm/include/llvm/CodeGen/MachineDominators.h", 249, __extension__ __PRETTY_FUNCTION__)) |
249 | "A basic block inserted via edge splitting cannot appear twice")(static_cast <bool> (Inserted && "A basic block inserted via edge splitting cannot appear twice" ) ? void (0) : __assert_fail ("Inserted && \"A basic block inserted via edge splitting cannot appear twice\"" , "llvm/include/llvm/CodeGen/MachineDominators.h", 249, __extension__ __PRETTY_FUNCTION__)); |
250 | CriticalEdgesToSplit.push_back({FromBB, ToBB, NewBB}); |
251 | } |
252 | }; |
253 | |
254 | //===------------------------------------- |
255 | /// DominatorTree GraphTraits specialization so the DominatorTree can be |
256 | /// iterable by generic graph iterators. |
257 | /// |
258 | |
259 | template <class Node, class ChildIterator> |
260 | struct MachineDomTreeGraphTraitsBase { |
261 | using NodeRef = Node *; |
262 | using ChildIteratorType = ChildIterator; |
263 | |
264 | static NodeRef getEntryNode(NodeRef N) { return N; } |
265 | static ChildIteratorType child_begin(NodeRef N) { return N->begin(); } |
266 | static ChildIteratorType child_end(NodeRef N) { return N->end(); } |
267 | }; |
268 | |
269 | template <class T> struct GraphTraits; |
270 | |
271 | template <> |
272 | struct GraphTraits<MachineDomTreeNode *> |
273 | : public MachineDomTreeGraphTraitsBase<MachineDomTreeNode, |
274 | MachineDomTreeNode::const_iterator> { |
275 | }; |
276 | |
277 | template <> |
278 | struct GraphTraits<const MachineDomTreeNode *> |
279 | : public MachineDomTreeGraphTraitsBase<const MachineDomTreeNode, |
280 | MachineDomTreeNode::const_iterator> { |
281 | }; |
282 | |
283 | template <> struct GraphTraits<MachineDominatorTree*> |
284 | : public GraphTraits<MachineDomTreeNode *> { |
285 | static NodeRef getEntryNode(MachineDominatorTree *DT) { |
286 | return DT->getRootNode(); |
287 | } |
288 | }; |
289 | |
290 | } // end namespace llvm |
291 | |
292 | #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 | /// either NodeRef->getParent() must return the parent node that is also a |
18 | /// pointer or DomTreeNodeTraits needs to be specialized. |
19 | /// |
20 | /// FIXME: Maybe GenericDomTree needs a TreeTraits, instead of GraphTraits. |
21 | /// |
22 | //===----------------------------------------------------------------------===// |
23 | |
24 | #ifndef LLVM_SUPPORT_GENERICDOMTREE_H |
25 | #define LLVM_SUPPORT_GENERICDOMTREE_H |
26 | |
27 | #include "llvm/ADT/DenseMap.h" |
28 | #include "llvm/ADT/GraphTraits.h" |
29 | #include "llvm/ADT/STLExtras.h" |
30 | #include "llvm/ADT/SmallPtrSet.h" |
31 | #include "llvm/ADT/SmallVector.h" |
32 | #include "llvm/Support/CFGDiff.h" |
33 | #include "llvm/Support/CFGUpdate.h" |
34 | #include "llvm/Support/raw_ostream.h" |
35 | #include <algorithm> |
36 | #include <cassert> |
37 | #include <cstddef> |
38 | #include <iterator> |
39 | #include <memory> |
40 | #include <type_traits> |
41 | #include <utility> |
42 | |
43 | namespace llvm { |
44 | |
45 | template <typename NodeT, bool IsPostDom> |
46 | class DominatorTreeBase; |
47 | |
48 | namespace DomTreeBuilder { |
49 | template <typename DomTreeT> |
50 | struct SemiNCAInfo; |
51 | } // namespace DomTreeBuilder |
52 | |
53 | /// Base class for the actual dominator tree node. |
54 | template <class NodeT> class DomTreeNodeBase { |
55 | friend class PostDominatorTree; |
56 | friend class DominatorTreeBase<NodeT, false>; |
57 | friend class DominatorTreeBase<NodeT, true>; |
58 | friend struct DomTreeBuilder::SemiNCAInfo<DominatorTreeBase<NodeT, false>>; |
59 | friend struct DomTreeBuilder::SemiNCAInfo<DominatorTreeBase<NodeT, true>>; |
60 | |
61 | NodeT *TheBB; |
62 | DomTreeNodeBase *IDom; |
63 | unsigned Level; |
64 | SmallVector<DomTreeNodeBase *, 4> Children; |
65 | mutable unsigned DFSNumIn = ~0; |
66 | mutable unsigned DFSNumOut = ~0; |
67 | |
68 | public: |
69 | DomTreeNodeBase(NodeT *BB, DomTreeNodeBase *iDom) |
70 | : TheBB(BB), IDom(iDom), Level(IDom ? IDom->Level + 1 : 0) {} |
71 | |
72 | using iterator = typename SmallVector<DomTreeNodeBase *, 4>::iterator; |
73 | using const_iterator = |
74 | typename SmallVector<DomTreeNodeBase *, 4>::const_iterator; |
75 | |
76 | iterator begin() { return Children.begin(); } |
77 | iterator end() { return Children.end(); } |
78 | const_iterator begin() const { return Children.begin(); } |
79 | const_iterator end() const { return Children.end(); } |
80 | |
81 | DomTreeNodeBase *const &back() const { return Children.back(); } |
82 | DomTreeNodeBase *&back() { return Children.back(); } |
83 | |
84 | iterator_range<iterator> children() { return make_range(begin(), end()); } |
85 | iterator_range<const_iterator> children() const { |
86 | return make_range(begin(), end()); |
87 | } |
88 | |
89 | NodeT *getBlock() const { return TheBB; } |
90 | DomTreeNodeBase *getIDom() const { return IDom; } |
91 | unsigned getLevel() const { return Level; } |
92 | |
93 | std::unique_ptr<DomTreeNodeBase> addChild( |
94 | std::unique_ptr<DomTreeNodeBase> C) { |
95 | Children.push_back(C.get()); |
96 | return C; |
97 | } |
98 | |
99 | bool isLeaf() const { return Children.empty(); } |
100 | size_t getNumChildren() const { return Children.size(); } |
101 | |
102 | void clearAllChildren() { Children.clear(); } |
103 | |
104 | bool compare(const DomTreeNodeBase *Other) const { |
105 | if (getNumChildren() != Other->getNumChildren()) |
106 | return true; |
107 | |
108 | if (Level != Other->Level) return true; |
109 | |
110 | SmallPtrSet<const NodeT *, 4> OtherChildren; |
111 | for (const DomTreeNodeBase *I : *Other) { |
112 | const NodeT *Nd = I->getBlock(); |
113 | OtherChildren.insert(Nd); |
114 | } |
115 | |
116 | for (const DomTreeNodeBase *I : *this) { |
117 | const NodeT *N = I->getBlock(); |
118 | if (OtherChildren.count(N) == 0) |
119 | return true; |
120 | } |
121 | return false; |
122 | } |
123 | |
124 | void setIDom(DomTreeNodeBase *NewIDom) { |
125 | assert(IDom && "No immediate dominator?")(static_cast <bool> (IDom && "No immediate dominator?" ) ? void (0) : __assert_fail ("IDom && \"No immediate dominator?\"" , "llvm/include/llvm/Support/GenericDomTree.h", 125, __extension__ __PRETTY_FUNCTION__)); |
126 | if (IDom == NewIDom) return; |
127 | |
128 | auto I = find(IDom->Children, this); |
129 | assert(I != IDom->Children.end() &&(static_cast <bool> (I != IDom->Children.end() && "Not in immediate dominator children set!") ? void (0) : __assert_fail ("I != IDom->Children.end() && \"Not in immediate dominator children set!\"" , "llvm/include/llvm/Support/GenericDomTree.h", 130, __extension__ __PRETTY_FUNCTION__)) |
130 | "Not in immediate dominator children set!")(static_cast <bool> (I != IDom->Children.end() && "Not in immediate dominator children set!") ? void (0) : __assert_fail ("I != IDom->Children.end() && \"Not in immediate dominator children set!\"" , "llvm/include/llvm/Support/GenericDomTree.h", 130, __extension__ __PRETTY_FUNCTION__)); |
131 | // I am no longer your child... |
132 | IDom->Children.erase(I); |
133 | |
134 | // Switch to new dominator |
135 | IDom = NewIDom; |
136 | IDom->Children.push_back(this); |
137 | |
138 | UpdateLevel(); |
139 | } |
140 | |
141 | /// getDFSNumIn/getDFSNumOut - These return the DFS visitation order for nodes |
142 | /// in the dominator tree. They are only guaranteed valid if |
143 | /// updateDFSNumbers() has been called. |
144 | unsigned getDFSNumIn() const { return DFSNumIn; } |
145 | unsigned getDFSNumOut() const { return DFSNumOut; } |
146 | |
147 | private: |
148 | // Return true if this node is dominated by other. Use this only if DFS info |
149 | // is valid. |
150 | bool DominatedBy(const DomTreeNodeBase *other) const { |
151 | return this->DFSNumIn >= other->DFSNumIn && |
152 | this->DFSNumOut <= other->DFSNumOut; |
153 | } |
154 | |
155 | void UpdateLevel() { |
156 | assert(IDom)(static_cast <bool> (IDom) ? void (0) : __assert_fail ( "IDom", "llvm/include/llvm/Support/GenericDomTree.h", 156, __extension__ __PRETTY_FUNCTION__)); |
157 | if (Level == IDom->Level + 1) return; |
158 | |
159 | SmallVector<DomTreeNodeBase *, 64> WorkStack = {this}; |
160 | |
161 | while (!WorkStack.empty()) { |
162 | DomTreeNodeBase *Current = WorkStack.pop_back_val(); |
163 | Current->Level = Current->IDom->Level + 1; |
164 | |
165 | for (DomTreeNodeBase *C : *Current) { |
166 | assert(C->IDom)(static_cast <bool> (C->IDom) ? void (0) : __assert_fail ("C->IDom", "llvm/include/llvm/Support/GenericDomTree.h", 166, __extension__ __PRETTY_FUNCTION__)); |
167 | if (C->Level != C->IDom->Level + 1) WorkStack.push_back(C); |
168 | } |
169 | } |
170 | } |
171 | }; |
172 | |
173 | template <class NodeT> |
174 | raw_ostream &operator<<(raw_ostream &O, const DomTreeNodeBase<NodeT> *Node) { |
175 | if (Node->getBlock()) |
176 | Node->getBlock()->printAsOperand(O, false); |
177 | else |
178 | O << " <<exit node>>"; |
179 | |
180 | O << " {" << Node->getDFSNumIn() << "," << Node->getDFSNumOut() << "} [" |
181 | << Node->getLevel() << "]\n"; |
182 | |
183 | return O; |
184 | } |
185 | |
186 | template <class NodeT> |
187 | void PrintDomTree(const DomTreeNodeBase<NodeT> *N, raw_ostream &O, |
188 | unsigned Lev) { |
189 | O.indent(2 * Lev) << "[" << Lev << "] " << N; |
190 | for (typename DomTreeNodeBase<NodeT>::const_iterator I = N->begin(), |
191 | E = N->end(); |
192 | I != E; ++I) |
193 | PrintDomTree<NodeT>(*I, O, Lev + 1); |
194 | } |
195 | |
196 | namespace DomTreeBuilder { |
197 | // The routines below are provided in a separate header but referenced here. |
198 | template <typename DomTreeT> |
199 | void Calculate(DomTreeT &DT); |
200 | |
201 | template <typename DomTreeT> |
202 | void CalculateWithUpdates(DomTreeT &DT, |
203 | ArrayRef<typename DomTreeT::UpdateType> Updates); |
204 | |
205 | template <typename DomTreeT> |
206 | void InsertEdge(DomTreeT &DT, typename DomTreeT::NodePtr From, |
207 | typename DomTreeT::NodePtr To); |
208 | |
209 | template <typename DomTreeT> |
210 | void DeleteEdge(DomTreeT &DT, typename DomTreeT::NodePtr From, |
211 | typename DomTreeT::NodePtr To); |
212 | |
213 | template <typename DomTreeT> |
214 | void ApplyUpdates(DomTreeT &DT, |
215 | GraphDiff<typename DomTreeT::NodePtr, |
216 | DomTreeT::IsPostDominator> &PreViewCFG, |
217 | GraphDiff<typename DomTreeT::NodePtr, |
218 | DomTreeT::IsPostDominator> *PostViewCFG); |
219 | |
220 | template <typename DomTreeT> |
221 | bool Verify(const DomTreeT &DT, typename DomTreeT::VerificationLevel VL); |
222 | } // namespace DomTreeBuilder |
223 | |
224 | /// Default DomTreeNode traits for NodeT. The default implementation assume a |
225 | /// Function-like NodeT. Can be specialized to support different node types. |
226 | template <typename NodeT> struct DomTreeNodeTraits { |
227 | using NodeType = NodeT; |
228 | using NodePtr = NodeT *; |
229 | using ParentPtr = decltype(std::declval<NodePtr>()->getParent()); |
230 | static_assert(std::is_pointer<ParentPtr>::value, |
231 | "Currently NodeT's parent must be a pointer type"); |
232 | using ParentType = std::remove_pointer_t<ParentPtr>; |
233 | |
234 | static NodeT *getEntryNode(ParentPtr Parent) { return &Parent->front(); } |
235 | static ParentPtr getParent(NodePtr BB) { return BB->getParent(); } |
236 | }; |
237 | |
238 | /// Core dominator tree base class. |
239 | /// |
240 | /// This class is a generic template over graph nodes. It is instantiated for |
241 | /// various graphs in the LLVM IR or in the code generator. |
242 | template <typename NodeT, bool IsPostDom> |
243 | class DominatorTreeBase { |
244 | public: |
245 | static_assert(std::is_pointer<typename GraphTraits<NodeT *>::NodeRef>::value, |
246 | "Currently DominatorTreeBase supports only pointer nodes"); |
247 | using NodeTrait = DomTreeNodeTraits<NodeT>; |
248 | using NodeType = typename NodeTrait::NodeType; |
249 | using NodePtr = typename NodeTrait::NodePtr; |
250 | using ParentPtr = typename NodeTrait::ParentPtr; |
251 | static_assert(std::is_pointer<ParentPtr>::value, |
252 | "Currently NodeT's parent must be a pointer type"); |
253 | using ParentType = std::remove_pointer_t<ParentPtr>; |
254 | static constexpr bool IsPostDominator = IsPostDom; |
255 | |
256 | using UpdateType = cfg::Update<NodePtr>; |
257 | using UpdateKind = cfg::UpdateKind; |
258 | static constexpr UpdateKind Insert = UpdateKind::Insert; |
259 | static constexpr UpdateKind Delete = UpdateKind::Delete; |
260 | |
261 | enum class VerificationLevel { Fast, Basic, Full }; |
262 | |
263 | protected: |
264 | // Dominators always have a single root, postdominators can have more. |
265 | SmallVector<NodeT *, IsPostDom ? 4 : 1> Roots; |
266 | |
267 | using DomTreeNodeMapType = |
268 | DenseMap<NodeT *, std::unique_ptr<DomTreeNodeBase<NodeT>>>; |
269 | DomTreeNodeMapType DomTreeNodes; |
270 | DomTreeNodeBase<NodeT> *RootNode = nullptr; |
271 | ParentPtr Parent = nullptr; |
272 | |
273 | mutable bool DFSInfoValid = false; |
274 | mutable unsigned int SlowQueries = 0; |
275 | |
276 | friend struct DomTreeBuilder::SemiNCAInfo<DominatorTreeBase>; |
277 | |
278 | public: |
279 | DominatorTreeBase() = default; |
280 | |
281 | DominatorTreeBase(DominatorTreeBase &&Arg) |
282 | : Roots(std::move(Arg.Roots)), |
283 | DomTreeNodes(std::move(Arg.DomTreeNodes)), |
284 | RootNode(Arg.RootNode), |
285 | Parent(Arg.Parent), |
286 | DFSInfoValid(Arg.DFSInfoValid), |
287 | SlowQueries(Arg.SlowQueries) { |
288 | Arg.wipe(); |
289 | } |
290 | |
291 | DominatorTreeBase &operator=(DominatorTreeBase &&RHS) { |
292 | Roots = std::move(RHS.Roots); |
293 | DomTreeNodes = std::move(RHS.DomTreeNodes); |
294 | RootNode = RHS.RootNode; |
295 | Parent = RHS.Parent; |
296 | DFSInfoValid = RHS.DFSInfoValid; |
297 | SlowQueries = RHS.SlowQueries; |
298 | RHS.wipe(); |
299 | return *this; |
300 | } |
301 | |
302 | DominatorTreeBase(const DominatorTreeBase &) = delete; |
303 | DominatorTreeBase &operator=(const DominatorTreeBase &) = delete; |
304 | |
305 | /// Iteration over roots. |
306 | /// |
307 | /// This may include multiple blocks if we are computing post dominators. |
308 | /// For forward dominators, this will always be a single block (the entry |
309 | /// block). |
310 | using root_iterator = typename SmallVectorImpl<NodeT *>::iterator; |
311 | using const_root_iterator = typename SmallVectorImpl<NodeT *>::const_iterator; |
312 | |
313 | root_iterator root_begin() { return Roots.begin(); } |
314 | const_root_iterator root_begin() const { return Roots.begin(); } |
315 | root_iterator root_end() { return Roots.end(); } |
316 | const_root_iterator root_end() const { return Roots.end(); } |
317 | |
318 | size_t root_size() const { return Roots.size(); } |
319 | |
320 | iterator_range<root_iterator> roots() { |
321 | return make_range(root_begin(), root_end()); |
322 | } |
323 | iterator_range<const_root_iterator> roots() const { |
324 | return make_range(root_begin(), root_end()); |
325 | } |
326 | |
327 | /// isPostDominator - Returns true if analysis based of postdoms |
328 | /// |
329 | bool isPostDominator() const { return IsPostDominator; } |
330 | |
331 | /// compare - Return false if the other dominator tree base matches this |
332 | /// dominator tree base. Otherwise return true. |
333 | bool compare(const DominatorTreeBase &Other) const { |
334 | if (Parent != Other.Parent) return true; |
335 | |
336 | if (Roots.size() != Other.Roots.size()) |
337 | return true; |
338 | |
339 | if (!std::is_permutation(Roots.begin(), Roots.end(), Other.Roots.begin())) |
340 | return true; |
341 | |
342 | const DomTreeNodeMapType &OtherDomTreeNodes = Other.DomTreeNodes; |
343 | if (DomTreeNodes.size() != OtherDomTreeNodes.size()) |
344 | return true; |
345 | |
346 | for (const auto &DomTreeNode : DomTreeNodes) { |
347 | NodeT *BB = DomTreeNode.first; |
348 | typename DomTreeNodeMapType::const_iterator OI = |
349 | OtherDomTreeNodes.find(BB); |
350 | if (OI == OtherDomTreeNodes.end()) |
351 | return true; |
352 | |
353 | DomTreeNodeBase<NodeT> &MyNd = *DomTreeNode.second; |
354 | DomTreeNodeBase<NodeT> &OtherNd = *OI->second; |
355 | |
356 | if (MyNd.compare(&OtherNd)) |
357 | return true; |
358 | } |
359 | |
360 | return false; |
361 | } |
362 | |
363 | /// getNode - return the (Post)DominatorTree node for the specified basic |
364 | /// block. This is the same as using operator[] on this class. The result |
365 | /// may (but is not required to) be null for a forward (backwards) |
366 | /// statically unreachable block. |
367 | DomTreeNodeBase<NodeT> *getNode(const NodeT *BB) const { |
368 | auto I = DomTreeNodes.find(BB); |
369 | if (I != DomTreeNodes.end()) |
370 | return I->second.get(); |
371 | return nullptr; |
372 | } |
373 | |
374 | /// See getNode. |
375 | DomTreeNodeBase<NodeT> *operator[](const NodeT *BB) const { |
376 | return getNode(BB); |
377 | } |
378 | |
379 | /// getRootNode - This returns the entry node for the CFG of the function. If |
380 | /// this tree represents the post-dominance relations for a function, however, |
381 | /// this root may be a node with the block == NULL. This is the case when |
382 | /// there are multiple exit nodes from a particular function. Consumers of |
383 | /// post-dominance information must be capable of dealing with this |
384 | /// possibility. |
385 | /// |
386 | DomTreeNodeBase<NodeT> *getRootNode() { return RootNode; } |
387 | const DomTreeNodeBase<NodeT> *getRootNode() const { return RootNode; } |
388 | |
389 | /// Get all nodes dominated by R, including R itself. |
390 | void getDescendants(NodeT *R, SmallVectorImpl<NodeT *> &Result) const { |
391 | Result.clear(); |
392 | const DomTreeNodeBase<NodeT> *RN = getNode(R); |
393 | if (!RN) |
394 | return; // If R is unreachable, it will not be present in the DOM tree. |
395 | SmallVector<const DomTreeNodeBase<NodeT> *, 8> WL; |
396 | WL.push_back(RN); |
397 | |
398 | while (!WL.empty()) { |
399 | const DomTreeNodeBase<NodeT> *N = WL.pop_back_val(); |
400 | Result.push_back(N->getBlock()); |
401 | WL.append(N->begin(), N->end()); |
402 | } |
403 | } |
404 | |
405 | /// properlyDominates - Returns true iff A dominates B and A != B. |
406 | /// Note that this is not a constant time operation! |
407 | /// |
408 | bool properlyDominates(const DomTreeNodeBase<NodeT> *A, |
409 | const DomTreeNodeBase<NodeT> *B) const { |
410 | if (!A || !B) |
411 | return false; |
412 | if (A == B) |
413 | return false; |
414 | return dominates(A, B); |
415 | } |
416 | |
417 | bool properlyDominates(const NodeT *A, const NodeT *B) const; |
418 | |
419 | /// isReachableFromEntry - Return true if A is dominated by the entry |
420 | /// block of the function containing it. |
421 | bool isReachableFromEntry(const NodeT *A) const { |
422 | assert(!this->isPostDominator() &&(static_cast <bool> (!this->isPostDominator() && "This is not implemented for post dominators") ? void (0) : __assert_fail ("!this->isPostDominator() && \"This is not implemented for post dominators\"" , "llvm/include/llvm/Support/GenericDomTree.h", 423, __extension__ __PRETTY_FUNCTION__)) |
423 | "This is not implemented for post dominators")(static_cast <bool> (!this->isPostDominator() && "This is not implemented for post dominators") ? void (0) : __assert_fail ("!this->isPostDominator() && \"This is not implemented for post dominators\"" , "llvm/include/llvm/Support/GenericDomTree.h", 423, __extension__ __PRETTY_FUNCTION__)); |
424 | return isReachableFromEntry(getNode(const_cast<NodeT *>(A))); |
425 | } |
426 | |
427 | bool isReachableFromEntry(const DomTreeNodeBase<NodeT> *A) const { return A; } |
428 | |
429 | /// dominates - Returns true iff A dominates B. Note that this is not a |
430 | /// constant time operation! |
431 | /// |
432 | bool dominates(const DomTreeNodeBase<NodeT> *A, |
433 | const DomTreeNodeBase<NodeT> *B) const { |
434 | // A node trivially dominates itself. |
435 | if (B == A) |
436 | return true; |
437 | |
438 | // An unreachable node is dominated by anything. |
439 | if (!isReachableFromEntry(B)) |
440 | return true; |
441 | |
442 | // And dominates nothing. |
443 | if (!isReachableFromEntry(A)) |
444 | return false; |
445 | |
446 | if (B->getIDom() == A) return true; |
447 | |
448 | if (A->getIDom() == B) return false; |
449 | |
450 | // A can only dominate B if it is higher in the tree. |
451 | if (A->getLevel() >= B->getLevel()) return false; |
452 | |
453 | // Compare the result of the tree walk and the dfs numbers, if expensive |
454 | // checks are enabled. |
455 | #ifdef EXPENSIVE_CHECKS |
456 | assert((!DFSInfoValid ||(static_cast <bool> ((!DFSInfoValid || (dominatedBySlowTreeWalk (A, B) == B->DominatedBy(A))) && "Tree walk disagrees with dfs numbers!" ) ? void (0) : __assert_fail ("(!DFSInfoValid || (dominatedBySlowTreeWalk(A, B) == B->DominatedBy(A))) && \"Tree walk disagrees with dfs numbers!\"" , "llvm/include/llvm/Support/GenericDomTree.h", 458, __extension__ __PRETTY_FUNCTION__)) |
457 | (dominatedBySlowTreeWalk(A, B) == B->DominatedBy(A))) &&(static_cast <bool> ((!DFSInfoValid || (dominatedBySlowTreeWalk (A, B) == B->DominatedBy(A))) && "Tree walk disagrees with dfs numbers!" ) ? void (0) : __assert_fail ("(!DFSInfoValid || (dominatedBySlowTreeWalk(A, B) == B->DominatedBy(A))) && \"Tree walk disagrees with dfs numbers!\"" , "llvm/include/llvm/Support/GenericDomTree.h", 458, __extension__ __PRETTY_FUNCTION__)) |
458 | "Tree walk disagrees with dfs numbers!")(static_cast <bool> ((!DFSInfoValid || (dominatedBySlowTreeWalk (A, B) == B->DominatedBy(A))) && "Tree walk disagrees with dfs numbers!" ) ? void (0) : __assert_fail ("(!DFSInfoValid || (dominatedBySlowTreeWalk(A, B) == B->DominatedBy(A))) && \"Tree walk disagrees with dfs numbers!\"" , "llvm/include/llvm/Support/GenericDomTree.h", 458, __extension__ __PRETTY_FUNCTION__)); |
459 | #endif |
460 | |
461 | if (DFSInfoValid) |
462 | return B->DominatedBy(A); |
463 | |
464 | // If we end up with too many slow queries, just update the |
465 | // DFS numbers on the theory that we are going to keep querying. |
466 | SlowQueries++; |
467 | if (SlowQueries > 32) { |
468 | updateDFSNumbers(); |
469 | return B->DominatedBy(A); |
470 | } |
471 | |
472 | return dominatedBySlowTreeWalk(A, B); |
473 | } |
474 | |
475 | bool dominates(const NodeT *A, const NodeT *B) const; |
476 | |
477 | NodeT *getRoot() const { |
478 | assert(this->Roots.size() == 1 && "Should always have entry node!")(static_cast <bool> (this->Roots.size() == 1 && "Should always have entry node!") ? void (0) : __assert_fail ("this->Roots.size() == 1 && \"Should always have entry node!\"" , "llvm/include/llvm/Support/GenericDomTree.h", 478, __extension__ __PRETTY_FUNCTION__)); |
479 | return this->Roots[0]; |
480 | } |
481 | |
482 | /// Find nearest common dominator basic block for basic block A and B. A and B |
483 | /// must have tree nodes. |
484 | NodeT *findNearestCommonDominator(NodeT *A, NodeT *B) const { |
485 | assert(A && B && "Pointers are not valid")(static_cast <bool> (A && B && "Pointers are not valid" ) ? void (0) : __assert_fail ("A && B && \"Pointers are not valid\"" , "llvm/include/llvm/Support/GenericDomTree.h", 485, __extension__ __PRETTY_FUNCTION__)); |
486 | assert(NodeTrait::getParent(A) == NodeTrait::getParent(B) &&(static_cast <bool> (NodeTrait::getParent(A) == NodeTrait ::getParent(B) && "Two blocks are not in same function" ) ? void (0) : __assert_fail ("NodeTrait::getParent(A) == NodeTrait::getParent(B) && \"Two blocks are not in same function\"" , "llvm/include/llvm/Support/GenericDomTree.h", 487, __extension__ __PRETTY_FUNCTION__)) |
487 | "Two blocks are not in same function")(static_cast <bool> (NodeTrait::getParent(A) == NodeTrait ::getParent(B) && "Two blocks are not in same function" ) ? void (0) : __assert_fail ("NodeTrait::getParent(A) == NodeTrait::getParent(B) && \"Two blocks are not in same function\"" , "llvm/include/llvm/Support/GenericDomTree.h", 487, __extension__ __PRETTY_FUNCTION__)); |
488 | |
489 | // If either A or B is a entry block then it is nearest common dominator |
490 | // (for forward-dominators). |
491 | if (!isPostDominator()) { |
492 | NodeT &Entry = |
493 | *DomTreeNodeTraits<NodeT>::getEntryNode(NodeTrait::getParent(A)); |
494 | if (A == &Entry || B == &Entry) |
495 | return &Entry; |
496 | } |
497 | |
498 | DomTreeNodeBase<NodeT> *NodeA = getNode(A); |
499 | DomTreeNodeBase<NodeT> *NodeB = getNode(B); |
500 | assert(NodeA && "A must be in the tree")(static_cast <bool> (NodeA && "A must be in the tree" ) ? void (0) : __assert_fail ("NodeA && \"A must be in the tree\"" , "llvm/include/llvm/Support/GenericDomTree.h", 500, __extension__ __PRETTY_FUNCTION__)); |
501 | assert(NodeB && "B must be in the tree")(static_cast <bool> (NodeB && "B must be in the tree" ) ? void (0) : __assert_fail ("NodeB && \"B must be in the tree\"" , "llvm/include/llvm/Support/GenericDomTree.h", 501, __extension__ __PRETTY_FUNCTION__)); |
502 | |
503 | // Use level information to go up the tree until the levels match. Then |
504 | // continue going up til we arrive at the same node. |
505 | while (NodeA != NodeB) { |
506 | if (NodeA->getLevel() < NodeB->getLevel()) std::swap(NodeA, NodeB); |
507 | |
508 | NodeA = NodeA->IDom; |
509 | } |
510 | |
511 | return NodeA->getBlock(); |
512 | } |
513 | |
514 | const NodeT *findNearestCommonDominator(const NodeT *A, |
515 | const NodeT *B) const { |
516 | // Cast away the const qualifiers here. This is ok since |
517 | // const is re-introduced on the return type. |
518 | return findNearestCommonDominator(const_cast<NodeT *>(A), |
519 | const_cast<NodeT *>(B)); |
520 | } |
521 | |
522 | bool isVirtualRoot(const DomTreeNodeBase<NodeT> *A) const { |
523 | return isPostDominator() && !A->getBlock(); |
524 | } |
525 | |
526 | //===--------------------------------------------------------------------===// |
527 | // API to update (Post)DominatorTree information based on modifications to |
528 | // the CFG... |
529 | |
530 | /// Inform the dominator tree about a sequence of CFG edge insertions and |
531 | /// deletions and perform a batch update on the tree. |
532 | /// |
533 | /// This function should be used when there were multiple CFG updates after |
534 | /// the last dominator tree update. It takes care of performing the updates |
535 | /// in sync with the CFG and optimizes away the redundant operations that |
536 | /// cancel each other. |
537 | /// The functions expects the sequence of updates to be balanced. Eg.: |
538 | /// - {{Insert, A, B}, {Delete, A, B}, {Insert, A, B}} is fine, because |
539 | /// logically it results in a single insertions. |
540 | /// - {{Insert, A, B}, {Insert, A, B}} is invalid, because it doesn't make |
541 | /// sense to insert the same edge twice. |
542 | /// |
543 | /// What's more, the functions assumes that it's safe to ask every node in the |
544 | /// CFG about its children and inverse children. This implies that deletions |
545 | /// of CFG edges must not delete the CFG nodes before calling this function. |
546 | /// |
547 | /// The applyUpdates function can reorder the updates and remove redundant |
548 | /// ones internally (as long as it is done in a deterministic fashion). The |
549 | /// batch updater is also able to detect sequences of zero and exactly one |
550 | /// update -- it's optimized to do less work in these cases. |
551 | /// |
552 | /// Note that for postdominators it automatically takes care of applying |
553 | /// updates on reverse edges internally (so there's no need to swap the |
554 | /// From and To pointers when constructing DominatorTree::UpdateType). |
555 | /// The type of updates is the same for DomTreeBase<T> and PostDomTreeBase<T> |
556 | /// with the same template parameter T. |
557 | /// |
558 | /// \param Updates An ordered sequence of updates to perform. The current CFG |
559 | /// and the reverse of these updates provides the pre-view of the CFG. |
560 | /// |
561 | void applyUpdates(ArrayRef<UpdateType> Updates) { |
562 | GraphDiff<NodePtr, IsPostDominator> PreViewCFG( |
563 | Updates, /*ReverseApplyUpdates=*/true); |
564 | DomTreeBuilder::ApplyUpdates(*this, PreViewCFG, nullptr); |
565 | } |
566 | |
567 | /// \param Updates An ordered sequence of updates to perform. The current CFG |
568 | /// and the reverse of these updates provides the pre-view of the CFG. |
569 | /// \param PostViewUpdates An ordered sequence of update to perform in order |
570 | /// to obtain a post-view of the CFG. The DT will be updated assuming the |
571 | /// obtained PostViewCFG is the desired end state. |
572 | void applyUpdates(ArrayRef<UpdateType> Updates, |
573 | ArrayRef<UpdateType> PostViewUpdates) { |
574 | if (Updates.empty()) { |
575 | GraphDiff<NodePtr, IsPostDom> PostViewCFG(PostViewUpdates); |
576 | DomTreeBuilder::ApplyUpdates(*this, PostViewCFG, &PostViewCFG); |
577 | } else { |
578 | // PreViewCFG needs to merge Updates and PostViewCFG. The updates in |
579 | // Updates need to be reversed, and match the direction in PostViewCFG. |
580 | // The PostViewCFG is created with updates reversed (equivalent to changes |
581 | // made to the CFG), so the PreViewCFG needs all the updates reverse |
582 | // applied. |
583 | SmallVector<UpdateType> AllUpdates(Updates.begin(), Updates.end()); |
584 | append_range(AllUpdates, PostViewUpdates); |
585 | GraphDiff<NodePtr, IsPostDom> PreViewCFG(AllUpdates, |
586 | /*ReverseApplyUpdates=*/true); |
587 | GraphDiff<NodePtr, IsPostDom> PostViewCFG(PostViewUpdates); |
588 | DomTreeBuilder::ApplyUpdates(*this, PreViewCFG, &PostViewCFG); |
589 | } |
590 | } |
591 | |
592 | /// Inform the dominator tree about a CFG edge insertion and update the tree. |
593 | /// |
594 | /// This function has to be called just before or just after making the update |
595 | /// on the actual CFG. There cannot be any other updates that the dominator |
596 | /// tree doesn't know about. |
597 | /// |
598 | /// Note that for postdominators it automatically takes care of inserting |
599 | /// a reverse edge internally (so there's no need to swap the parameters). |
600 | /// |
601 | void insertEdge(NodeT *From, NodeT *To) { |
602 | assert(From)(static_cast <bool> (From) ? void (0) : __assert_fail ( "From", "llvm/include/llvm/Support/GenericDomTree.h", 602, __extension__ __PRETTY_FUNCTION__)); |
603 | assert(To)(static_cast <bool> (To) ? void (0) : __assert_fail ("To" , "llvm/include/llvm/Support/GenericDomTree.h", 603, __extension__ __PRETTY_FUNCTION__)); |
604 | assert(NodeTrait::getParent(From) == Parent)(static_cast <bool> (NodeTrait::getParent(From) == Parent ) ? void (0) : __assert_fail ("NodeTrait::getParent(From) == Parent" , "llvm/include/llvm/Support/GenericDomTree.h", 604, __extension__ __PRETTY_FUNCTION__)); |
605 | assert(NodeTrait::getParent(To) == Parent)(static_cast <bool> (NodeTrait::getParent(To) == Parent ) ? void (0) : __assert_fail ("NodeTrait::getParent(To) == Parent" , "llvm/include/llvm/Support/GenericDomTree.h", 605, __extension__ __PRETTY_FUNCTION__)); |
606 | DomTreeBuilder::InsertEdge(*this, From, To); |
607 | } |
608 | |
609 | /// Inform the dominator tree about a CFG edge deletion and update the tree. |
610 | /// |
611 | /// This function has to be called just after making the update on the actual |
612 | /// CFG. An internal functions checks if the edge doesn't exist in the CFG in |
613 | /// DEBUG mode. There cannot be any other updates that the |
614 | /// dominator tree doesn't know about. |
615 | /// |
616 | /// Note that for postdominators it automatically takes care of deleting |
617 | /// a reverse edge internally (so there's no need to swap the parameters). |
618 | /// |
619 | void deleteEdge(NodeT *From, NodeT *To) { |
620 | assert(From)(static_cast <bool> (From) ? void (0) : __assert_fail ( "From", "llvm/include/llvm/Support/GenericDomTree.h", 620, __extension__ __PRETTY_FUNCTION__)); |
621 | assert(To)(static_cast <bool> (To) ? void (0) : __assert_fail ("To" , "llvm/include/llvm/Support/GenericDomTree.h", 621, __extension__ __PRETTY_FUNCTION__)); |
622 | assert(NodeTrait::getParent(From) == Parent)(static_cast <bool> (NodeTrait::getParent(From) == Parent ) ? void (0) : __assert_fail ("NodeTrait::getParent(From) == Parent" , "llvm/include/llvm/Support/GenericDomTree.h", 622, __extension__ __PRETTY_FUNCTION__)); |
623 | assert(NodeTrait::getParent(To) == Parent)(static_cast <bool> (NodeTrait::getParent(To) == Parent ) ? void (0) : __assert_fail ("NodeTrait::getParent(To) == Parent" , "llvm/include/llvm/Support/GenericDomTree.h", 623, __extension__ __PRETTY_FUNCTION__)); |
624 | DomTreeBuilder::DeleteEdge(*this, From, To); |
625 | } |
626 | |
627 | /// Add a new node to the dominator tree information. |
628 | /// |
629 | /// This creates a new node as a child of DomBB dominator node, linking it |
630 | /// into the children list of the immediate dominator. |
631 | /// |
632 | /// \param BB New node in CFG. |
633 | /// \param DomBB CFG node that is dominator for BB. |
634 | /// \returns New dominator tree node that represents new CFG node. |
635 | /// |
636 | DomTreeNodeBase<NodeT> *addNewBlock(NodeT *BB, NodeT *DomBB) { |
637 | assert(getNode(BB) == nullptr && "Block already in dominator tree!")(static_cast <bool> (getNode(BB) == nullptr && "Block already in dominator tree!" ) ? void (0) : __assert_fail ("getNode(BB) == nullptr && \"Block already in dominator tree!\"" , "llvm/include/llvm/Support/GenericDomTree.h", 637, __extension__ __PRETTY_FUNCTION__)); |
638 | DomTreeNodeBase<NodeT> *IDomNode = getNode(DomBB); |
639 | assert(IDomNode && "Not immediate dominator specified for block!")(static_cast <bool> (IDomNode && "Not immediate dominator specified for block!" ) ? void (0) : __assert_fail ("IDomNode && \"Not immediate dominator specified for block!\"" , "llvm/include/llvm/Support/GenericDomTree.h", 639, __extension__ __PRETTY_FUNCTION__)); |
640 | DFSInfoValid = false; |
641 | return createChild(BB, IDomNode); |
642 | } |
643 | |
644 | /// Add a new node to the forward dominator tree and make it a new root. |
645 | /// |
646 | /// \param BB New node in CFG. |
647 | /// \returns New dominator tree node that represents new CFG node. |
648 | /// |
649 | DomTreeNodeBase<NodeT> *setNewRoot(NodeT *BB) { |
650 | assert(getNode(BB) == nullptr && "Block already in dominator tree!")(static_cast <bool> (getNode(BB) == nullptr && "Block already in dominator tree!" ) ? void (0) : __assert_fail ("getNode(BB) == nullptr && \"Block already in dominator tree!\"" , "llvm/include/llvm/Support/GenericDomTree.h", 650, __extension__ __PRETTY_FUNCTION__)); |
651 | assert(!this->isPostDominator() &&(static_cast <bool> (!this->isPostDominator() && "Cannot change root of post-dominator tree") ? void (0) : __assert_fail ("!this->isPostDominator() && \"Cannot change root of post-dominator tree\"" , "llvm/include/llvm/Support/GenericDomTree.h", 652, __extension__ __PRETTY_FUNCTION__)) |
652 | "Cannot change root of post-dominator tree")(static_cast <bool> (!this->isPostDominator() && "Cannot change root of post-dominator tree") ? void (0) : __assert_fail ("!this->isPostDominator() && \"Cannot change root of post-dominator tree\"" , "llvm/include/llvm/Support/GenericDomTree.h", 652, __extension__ __PRETTY_FUNCTION__)); |
653 | DFSInfoValid = false; |
654 | DomTreeNodeBase<NodeT> *NewNode = createNode(BB); |
655 | if (Roots.empty()) { |
656 | addRoot(BB); |
657 | } else { |
658 | assert(Roots.size() == 1)(static_cast <bool> (Roots.size() == 1) ? void (0) : __assert_fail ("Roots.size() == 1", "llvm/include/llvm/Support/GenericDomTree.h" , 658, __extension__ __PRETTY_FUNCTION__)); |
659 | NodeT *OldRoot = Roots.front(); |
660 | auto &OldNode = DomTreeNodes[OldRoot]; |
661 | OldNode = NewNode->addChild(std::move(DomTreeNodes[OldRoot])); |
662 | OldNode->IDom = NewNode; |
663 | OldNode->UpdateLevel(); |
664 | Roots[0] = BB; |
665 | } |
666 | return RootNode = NewNode; |
667 | } |
668 | |
669 | /// changeImmediateDominator - This method is used to update the dominator |
670 | /// tree information when a node's immediate dominator changes. |
671 | /// |
672 | void changeImmediateDominator(DomTreeNodeBase<NodeT> *N, |
673 | DomTreeNodeBase<NodeT> *NewIDom) { |
674 | assert(N && NewIDom && "Cannot change null node pointers!")(static_cast <bool> (N && NewIDom && "Cannot change null node pointers!" ) ? void (0) : __assert_fail ("N && NewIDom && \"Cannot change null node pointers!\"" , "llvm/include/llvm/Support/GenericDomTree.h", 674, __extension__ __PRETTY_FUNCTION__)); |
675 | DFSInfoValid = false; |
676 | N->setIDom(NewIDom); |
677 | } |
678 | |
679 | void changeImmediateDominator(NodeT *BB, NodeT *NewBB) { |
680 | changeImmediateDominator(getNode(BB), getNode(NewBB)); |
681 | } |
682 | |
683 | /// eraseNode - Removes a node from the dominator tree. Block must not |
684 | /// dominate any other blocks. Removes node from its immediate dominator's |
685 | /// children list. Deletes dominator node associated with basic block BB. |
686 | void eraseNode(NodeT *BB) { |
687 | DomTreeNodeBase<NodeT> *Node = getNode(BB); |
688 | assert(Node && "Removing node that isn't in dominator tree.")(static_cast <bool> (Node && "Removing node that isn't in dominator tree." ) ? void (0) : __assert_fail ("Node && \"Removing node that isn't in dominator tree.\"" , "llvm/include/llvm/Support/GenericDomTree.h", 688, __extension__ __PRETTY_FUNCTION__)); |
689 | assert(Node->isLeaf() && "Node is not a leaf node.")(static_cast <bool> (Node->isLeaf() && "Node is not a leaf node." ) ? void (0) : __assert_fail ("Node->isLeaf() && \"Node is not a leaf node.\"" , "llvm/include/llvm/Support/GenericDomTree.h", 689, __extension__ __PRETTY_FUNCTION__)); |
690 | |
691 | DFSInfoValid = false; |
692 | |
693 | // Remove node from immediate dominator's children list. |
694 | DomTreeNodeBase<NodeT> *IDom = Node->getIDom(); |
695 | if (IDom) { |
696 | const auto I = find(IDom->Children, Node); |
697 | assert(I != IDom->Children.end() &&(static_cast <bool> (I != IDom->Children.end() && "Not in immediate dominator children set!") ? void (0) : __assert_fail ("I != IDom->Children.end() && \"Not in immediate dominator children set!\"" , "llvm/include/llvm/Support/GenericDomTree.h", 698, __extension__ __PRETTY_FUNCTION__)) |
698 | "Not in immediate dominator children set!")(static_cast <bool> (I != IDom->Children.end() && "Not in immediate dominator children set!") ? void (0) : __assert_fail ("I != IDom->Children.end() && \"Not in immediate dominator children set!\"" , "llvm/include/llvm/Support/GenericDomTree.h", 698, __extension__ __PRETTY_FUNCTION__)); |
699 | // I am no longer your child... |
700 | IDom->Children.erase(I); |
701 | } |
702 | |
703 | DomTreeNodes.erase(BB); |
704 | |
705 | if (!IsPostDom) return; |
706 | |
707 | // Remember to update PostDominatorTree roots. |
708 | auto RIt = llvm::find(Roots, BB); |
709 | if (RIt != Roots.end()) { |
710 | std::swap(*RIt, Roots.back()); |
711 | Roots.pop_back(); |
712 | } |
713 | } |
714 | |
715 | /// splitBlock - BB is split and now it has one successor. Update dominator |
716 | /// tree to reflect this change. |
717 | void splitBlock(NodeT *NewBB) { |
718 | if (IsPostDominator) |
719 | Split<Inverse<NodeT *>>(NewBB); |
720 | else |
721 | Split<NodeT *>(NewBB); |
722 | } |
723 | |
724 | /// print - Convert to human readable form |
725 | /// |
726 | void print(raw_ostream &O) const { |
727 | O << "=============================--------------------------------\n"; |
728 | if (IsPostDominator) |
729 | O << "Inorder PostDominator Tree: "; |
730 | else |
731 | O << "Inorder Dominator Tree: "; |
732 | if (!DFSInfoValid) |
733 | O << "DFSNumbers invalid: " << SlowQueries << " slow queries."; |
734 | O << "\n"; |
735 | |
736 | // The postdom tree can have a null root if there are no returns. |
737 | if (getRootNode()) PrintDomTree<NodeT>(getRootNode(), O, 1); |
738 | O << "Roots: "; |
739 | for (const NodePtr Block : Roots) { |
740 | Block->printAsOperand(O, false); |
741 | O << " "; |
742 | } |
743 | O << "\n"; |
744 | } |
745 | |
746 | public: |
747 | /// updateDFSNumbers - Assign In and Out numbers to the nodes while walking |
748 | /// dominator tree in dfs order. |
749 | void updateDFSNumbers() const { |
750 | if (DFSInfoValid) { |
751 | SlowQueries = 0; |
752 | return; |
753 | } |
754 | |
755 | SmallVector<std::pair<const DomTreeNodeBase<NodeT> *, |
756 | typename DomTreeNodeBase<NodeT>::const_iterator>, |
757 | 32> WorkStack; |
758 | |
759 | const DomTreeNodeBase<NodeT> *ThisRoot = getRootNode(); |
760 | assert((!Parent || ThisRoot) && "Empty constructed DomTree")(static_cast <bool> ((!Parent || ThisRoot) && "Empty constructed DomTree" ) ? void (0) : __assert_fail ("(!Parent || ThisRoot) && \"Empty constructed DomTree\"" , "llvm/include/llvm/Support/GenericDomTree.h", 760, __extension__ __PRETTY_FUNCTION__)); |
761 | if (!ThisRoot) |
762 | return; |
763 | |
764 | // Both dominators and postdominators have a single root node. In the case |
765 | // case of PostDominatorTree, this node is a virtual root. |
766 | WorkStack.push_back({ThisRoot, ThisRoot->begin()}); |
767 | |
768 | unsigned DFSNum = 0; |
769 | ThisRoot->DFSNumIn = DFSNum++; |
770 | |
771 | while (!WorkStack.empty()) { |
772 | const DomTreeNodeBase<NodeT> *Node = WorkStack.back().first; |
773 | const auto ChildIt = WorkStack.back().second; |
774 | |
775 | // If we visited all of the children of this node, "recurse" back up the |
776 | // stack setting the DFOutNum. |
777 | if (ChildIt == Node->end()) { |
778 | Node->DFSNumOut = DFSNum++; |
779 | WorkStack.pop_back(); |
780 | } else { |
781 | // Otherwise, recursively visit this child. |
782 | const DomTreeNodeBase<NodeT> *Child = *ChildIt; |
783 | ++WorkStack.back().second; |
784 | |
785 | WorkStack.push_back({Child, Child->begin()}); |
786 | Child->DFSNumIn = DFSNum++; |
787 | } |
788 | } |
789 | |
790 | SlowQueries = 0; |
791 | DFSInfoValid = true; |
792 | } |
793 | |
794 | /// recalculate - compute a dominator tree for the given function |
795 | void recalculate(ParentType &Func) { |
796 | Parent = &Func; |
797 | DomTreeBuilder::Calculate(*this); |
798 | } |
799 | |
800 | void recalculate(ParentType &Func, ArrayRef<UpdateType> Updates) { |
801 | Parent = &Func; |
802 | DomTreeBuilder::CalculateWithUpdates(*this, Updates); |
803 | } |
804 | |
805 | /// verify - checks if the tree is correct. There are 3 level of verification: |
806 | /// - Full -- verifies if the tree is correct by making sure all the |
807 | /// properties (including the parent and the sibling property) |
808 | /// hold. |
809 | /// Takes O(N^3) time. |
810 | /// |
811 | /// - Basic -- checks if the tree is correct, but compares it to a freshly |
812 | /// constructed tree instead of checking the sibling property. |
813 | /// Takes O(N^2) time. |
814 | /// |
815 | /// - Fast -- checks basic tree structure and compares it with a freshly |
816 | /// constructed tree. |
817 | /// Takes O(N^2) time worst case, but is faster in practise (same |
818 | /// as tree construction). |
819 | bool verify(VerificationLevel VL = VerificationLevel::Full) const { |
820 | return DomTreeBuilder::Verify(*this, VL); |
821 | } |
822 | |
823 | void reset() { |
824 | DomTreeNodes.clear(); |
825 | Roots.clear(); |
826 | RootNode = nullptr; |
827 | Parent = nullptr; |
828 | DFSInfoValid = false; |
829 | SlowQueries = 0; |
830 | } |
831 | |
832 | protected: |
833 | void addRoot(NodeT *BB) { this->Roots.push_back(BB); } |
834 | |
835 | DomTreeNodeBase<NodeT> *createChild(NodeT *BB, DomTreeNodeBase<NodeT> *IDom) { |
836 | return (DomTreeNodes[BB] = IDom->addChild( |
837 | std::make_unique<DomTreeNodeBase<NodeT>>(BB, IDom))) |
838 | .get(); |
839 | } |
840 | |
841 | DomTreeNodeBase<NodeT> *createNode(NodeT *BB) { |
842 | return (DomTreeNodes[BB] = |
843 | std::make_unique<DomTreeNodeBase<NodeT>>(BB, nullptr)) |
844 | .get(); |
845 | } |
846 | |
847 | // NewBB is split and now it has one successor. Update dominator tree to |
848 | // reflect this change. |
849 | template <class N> |
850 | void Split(typename GraphTraits<N>::NodeRef NewBB) { |
851 | using GraphT = GraphTraits<N>; |
852 | using NodeRef = typename GraphT::NodeRef; |
853 | assert(std::distance(GraphT::child_begin(NewBB),(static_cast <bool> (std::distance(GraphT::child_begin( NewBB), GraphT::child_end(NewBB)) == 1 && "NewBB should have a single successor!" ) ? void (0) : __assert_fail ("std::distance(GraphT::child_begin(NewBB), GraphT::child_end(NewBB)) == 1 && \"NewBB should have a single successor!\"" , "llvm/include/llvm/Support/GenericDomTree.h", 855, __extension__ __PRETTY_FUNCTION__)) |
854 | GraphT::child_end(NewBB)) == 1 &&(static_cast <bool> (std::distance(GraphT::child_begin( NewBB), GraphT::child_end(NewBB)) == 1 && "NewBB should have a single successor!" ) ? void (0) : __assert_fail ("std::distance(GraphT::child_begin(NewBB), GraphT::child_end(NewBB)) == 1 && \"NewBB should have a single successor!\"" , "llvm/include/llvm/Support/GenericDomTree.h", 855, __extension__ __PRETTY_FUNCTION__)) |
855 | "NewBB should have a single successor!")(static_cast <bool> (std::distance(GraphT::child_begin( NewBB), GraphT::child_end(NewBB)) == 1 && "NewBB should have a single successor!" ) ? void (0) : __assert_fail ("std::distance(GraphT::child_begin(NewBB), GraphT::child_end(NewBB)) == 1 && \"NewBB should have a single successor!\"" , "llvm/include/llvm/Support/GenericDomTree.h", 855, __extension__ __PRETTY_FUNCTION__)); |
856 | NodeRef NewBBSucc = *GraphT::child_begin(NewBB); |
857 | |
858 | SmallVector<NodeRef, 4> PredBlocks(children<Inverse<N>>(NewBB)); |
859 | |
860 | assert(!PredBlocks.empty() && "No predblocks?")(static_cast <bool> (!PredBlocks.empty() && "No predblocks?" ) ? void (0) : __assert_fail ("!PredBlocks.empty() && \"No predblocks?\"" , "llvm/include/llvm/Support/GenericDomTree.h", 860, __extension__ __PRETTY_FUNCTION__)); |
861 | |
862 | bool NewBBDominatesNewBBSucc = true; |
863 | for (auto *Pred : children<Inverse<N>>(NewBBSucc)) { |
864 | if (Pred != NewBB && !dominates(NewBBSucc, Pred) && |
865 | isReachableFromEntry(Pred)) { |
866 | NewBBDominatesNewBBSucc = false; |
867 | break; |
868 | } |
869 | } |
870 | |
871 | // Find NewBB's immediate dominator and create new dominator tree node for |
872 | // NewBB. |
873 | NodeT *NewBBIDom = nullptr; |
874 | unsigned i = 0; |
875 | for (i = 0; i < PredBlocks.size(); ++i) |
876 | if (isReachableFromEntry(PredBlocks[i])) { |
877 | NewBBIDom = PredBlocks[i]; |
878 | break; |
879 | } |
880 | |
881 | // It's possible that none of the predecessors of NewBB are reachable; |
882 | // in that case, NewBB itself is unreachable, so nothing needs to be |
883 | // changed. |
884 | if (!NewBBIDom) return; |
885 | |
886 | for (i = i + 1; i < PredBlocks.size(); ++i) { |
887 | if (isReachableFromEntry(PredBlocks[i])) |
888 | NewBBIDom = findNearestCommonDominator(NewBBIDom, PredBlocks[i]); |
889 | } |
890 | |
891 | // Create the new dominator tree node... and set the idom of NewBB. |
892 | DomTreeNodeBase<NodeT> *NewBBNode = addNewBlock(NewBB, NewBBIDom); |
893 | |
894 | // If NewBB strictly dominates other blocks, then it is now the immediate |
895 | // dominator of NewBBSucc. Update the dominator tree as appropriate. |
896 | if (NewBBDominatesNewBBSucc) { |
897 | DomTreeNodeBase<NodeT> *NewBBSuccNode = getNode(NewBBSucc); |
898 | changeImmediateDominator(NewBBSuccNode, NewBBNode); |
899 | } |
900 | } |
901 | |
902 | private: |
903 | bool dominatedBySlowTreeWalk(const DomTreeNodeBase<NodeT> *A, |
904 | const DomTreeNodeBase<NodeT> *B) const { |
905 | assert(A != B)(static_cast <bool> (A != B) ? void (0) : __assert_fail ("A != B", "llvm/include/llvm/Support/GenericDomTree.h", 905 , __extension__ __PRETTY_FUNCTION__)); |
906 | assert(isReachableFromEntry(B))(static_cast <bool> (isReachableFromEntry(B)) ? void (0 ) : __assert_fail ("isReachableFromEntry(B)", "llvm/include/llvm/Support/GenericDomTree.h" , 906, __extension__ __PRETTY_FUNCTION__)); |
907 | assert(isReachableFromEntry(A))(static_cast <bool> (isReachableFromEntry(A)) ? void (0 ) : __assert_fail ("isReachableFromEntry(A)", "llvm/include/llvm/Support/GenericDomTree.h" , 907, __extension__ __PRETTY_FUNCTION__)); |
908 | |
909 | const unsigned ALevel = A->getLevel(); |
910 | const DomTreeNodeBase<NodeT> *IDom; |
911 | |
912 | // Don't walk nodes above A's subtree. When we reach A's level, we must |
913 | // either find A or be in some other subtree not dominated by A. |
914 | while ((IDom = B->getIDom()) != nullptr && IDom->getLevel() >= ALevel) |
915 | B = IDom; // Walk up the tree |
916 | |
917 | return B == A; |
918 | } |
919 | |
920 | /// Wipe this tree's state without releasing any resources. |
921 | /// |
922 | /// This is essentially a post-move helper only. It leaves the object in an |
923 | /// assignable and destroyable state, but otherwise invalid. |
924 | void wipe() { |
925 | DomTreeNodes.clear(); |
926 | RootNode = nullptr; |
927 | Parent = nullptr; |
928 | } |
929 | }; |
930 | |
931 | template <typename T> |
932 | using DomTreeBase = DominatorTreeBase<T, false>; |
933 | |
934 | template <typename T> |
935 | using PostDomTreeBase = DominatorTreeBase<T, true>; |
936 | |
937 | // These two functions are declared out of line as a workaround for building |
938 | // with old (< r147295) versions of clang because of pr11642. |
939 | template <typename NodeT, bool IsPostDom> |
940 | bool DominatorTreeBase<NodeT, IsPostDom>::dominates(const NodeT *A, |
941 | const NodeT *B) const { |
942 | if (A == B) |
943 | return true; |
944 | |
945 | // Cast away the const qualifiers here. This is ok since |
946 | // this function doesn't actually return the values returned |
947 | // from getNode. |
948 | return dominates(getNode(const_cast<NodeT *>(A)), |
949 | getNode(const_cast<NodeT *>(B))); |
950 | } |
951 | template <typename NodeT, bool IsPostDom> |
952 | bool DominatorTreeBase<NodeT, IsPostDom>::properlyDominates( |
953 | const NodeT *A, const NodeT *B) const { |
954 | if (A == B) |
955 | return false; |
956 | |
957 | // Cast away the const qualifiers here. This is ok since |
958 | // this function doesn't actually return the values returned |
959 | // from getNode. |
960 | return dominates(getNode(const_cast<NodeT *>(A)), |
961 | getNode(const_cast<NodeT *>(B))); |
962 | } |
963 | |
964 | } // end namespace llvm |
965 | |
966 | #endif // LLVM_SUPPORT_GENERICDOMTREE_H |