File: | src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Support/GenericDomTree.h |
Warning: | line 494, column 12 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //====- X86FlagsCopyLowering.cpp - Lowers COPY nodes of EFLAGS ------------===// | |||
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 | /// Lowers COPY nodes of EFLAGS by directly extracting and preserving individual | |||
11 | /// flag bits. | |||
12 | /// | |||
13 | /// We have to do this by carefully analyzing and rewriting the usage of the | |||
14 | /// copied EFLAGS register because there is no general way to rematerialize the | |||
15 | /// entire EFLAGS register safely and efficiently. Using `popf` both forces | |||
16 | /// dynamic stack adjustment and can create correctness issues due to IF, TF, | |||
17 | /// and other non-status flags being overwritten. Using sequences involving | |||
18 | /// SAHF don't work on all x86 processors and are often quite slow compared to | |||
19 | /// directly testing a single status preserved in its own GPR. | |||
20 | /// | |||
21 | //===----------------------------------------------------------------------===// | |||
22 | ||||
23 | #include "X86.h" | |||
24 | #include "X86InstrBuilder.h" | |||
25 | #include "X86InstrInfo.h" | |||
26 | #include "X86Subtarget.h" | |||
27 | #include "llvm/ADT/ArrayRef.h" | |||
28 | #include "llvm/ADT/DenseMap.h" | |||
29 | #include "llvm/ADT/PostOrderIterator.h" | |||
30 | #include "llvm/ADT/STLExtras.h" | |||
31 | #include "llvm/ADT/ScopeExit.h" | |||
32 | #include "llvm/ADT/SmallPtrSet.h" | |||
33 | #include "llvm/ADT/SmallSet.h" | |||
34 | #include "llvm/ADT/SmallVector.h" | |||
35 | #include "llvm/ADT/SparseBitVector.h" | |||
36 | #include "llvm/ADT/Statistic.h" | |||
37 | #include "llvm/CodeGen/MachineBasicBlock.h" | |||
38 | #include "llvm/CodeGen/MachineConstantPool.h" | |||
39 | #include "llvm/CodeGen/MachineDominators.h" | |||
40 | #include "llvm/CodeGen/MachineFunction.h" | |||
41 | #include "llvm/CodeGen/MachineFunctionPass.h" | |||
42 | #include "llvm/CodeGen/MachineInstr.h" | |||
43 | #include "llvm/CodeGen/MachineInstrBuilder.h" | |||
44 | #include "llvm/CodeGen/MachineModuleInfo.h" | |||
45 | #include "llvm/CodeGen/MachineOperand.h" | |||
46 | #include "llvm/CodeGen/MachineRegisterInfo.h" | |||
47 | #include "llvm/CodeGen/MachineSSAUpdater.h" | |||
48 | #include "llvm/CodeGen/TargetInstrInfo.h" | |||
49 | #include "llvm/CodeGen/TargetRegisterInfo.h" | |||
50 | #include "llvm/CodeGen/TargetSchedule.h" | |||
51 | #include "llvm/CodeGen/TargetSubtargetInfo.h" | |||
52 | #include "llvm/IR/DebugLoc.h" | |||
53 | #include "llvm/MC/MCSchedule.h" | |||
54 | #include "llvm/Pass.h" | |||
55 | #include "llvm/Support/CommandLine.h" | |||
56 | #include "llvm/Support/Debug.h" | |||
57 | #include "llvm/Support/raw_ostream.h" | |||
58 | #include <algorithm> | |||
59 | #include <cassert> | |||
60 | #include <iterator> | |||
61 | #include <utility> | |||
62 | ||||
63 | using namespace llvm; | |||
64 | ||||
65 | #define PASS_KEY"x86-flags-copy-lowering" "x86-flags-copy-lowering" | |||
66 | #define DEBUG_TYPE"x86-flags-copy-lowering" PASS_KEY"x86-flags-copy-lowering" | |||
67 | ||||
68 | STATISTIC(NumCopiesEliminated, "Number of copies of EFLAGS eliminated")static llvm::Statistic NumCopiesEliminated = {"x86-flags-copy-lowering" , "NumCopiesEliminated", "Number of copies of EFLAGS eliminated" }; | |||
69 | STATISTIC(NumSetCCsInserted, "Number of setCC instructions inserted")static llvm::Statistic NumSetCCsInserted = {"x86-flags-copy-lowering" , "NumSetCCsInserted", "Number of setCC instructions inserted" }; | |||
70 | STATISTIC(NumTestsInserted, "Number of test instructions inserted")static llvm::Statistic NumTestsInserted = {"x86-flags-copy-lowering" , "NumTestsInserted", "Number of test instructions inserted"}; | |||
71 | STATISTIC(NumAddsInserted, "Number of adds instructions inserted")static llvm::Statistic NumAddsInserted = {"x86-flags-copy-lowering" , "NumAddsInserted", "Number of adds instructions inserted"}; | |||
72 | ||||
73 | namespace { | |||
74 | ||||
75 | // Convenient array type for storing registers associated with each condition. | |||
76 | using CondRegArray = std::array<unsigned, X86::LAST_VALID_COND + 1>; | |||
77 | ||||
78 | class X86FlagsCopyLoweringPass : public MachineFunctionPass { | |||
79 | public: | |||
80 | X86FlagsCopyLoweringPass() : MachineFunctionPass(ID) { } | |||
81 | ||||
82 | StringRef getPassName() const override { return "X86 EFLAGS copy lowering"; } | |||
83 | bool runOnMachineFunction(MachineFunction &MF) override; | |||
84 | void getAnalysisUsage(AnalysisUsage &AU) const override; | |||
85 | ||||
86 | /// Pass identification, replacement for typeid. | |||
87 | static char ID; | |||
88 | ||||
89 | private: | |||
90 | MachineRegisterInfo *MRI = nullptr; | |||
91 | const X86Subtarget *Subtarget = nullptr; | |||
92 | const X86InstrInfo *TII = nullptr; | |||
93 | const TargetRegisterInfo *TRI = nullptr; | |||
94 | const TargetRegisterClass *PromoteRC = nullptr; | |||
95 | MachineDominatorTree *MDT = nullptr; | |||
96 | ||||
97 | CondRegArray collectCondsInRegs(MachineBasicBlock &MBB, | |||
98 | MachineBasicBlock::iterator CopyDefI); | |||
99 | ||||
100 | Register promoteCondToReg(MachineBasicBlock &MBB, | |||
101 | MachineBasicBlock::iterator TestPos, | |||
102 | const DebugLoc &TestLoc, X86::CondCode Cond); | |||
103 | std::pair<unsigned, bool> getCondOrInverseInReg( | |||
104 | MachineBasicBlock &TestMBB, MachineBasicBlock::iterator TestPos, | |||
105 | const DebugLoc &TestLoc, X86::CondCode Cond, CondRegArray &CondRegs); | |||
106 | void insertTest(MachineBasicBlock &MBB, MachineBasicBlock::iterator Pos, | |||
107 | const DebugLoc &Loc, unsigned Reg); | |||
108 | ||||
109 | void rewriteArithmetic(MachineBasicBlock &TestMBB, | |||
110 | MachineBasicBlock::iterator TestPos, | |||
111 | const DebugLoc &TestLoc, MachineInstr &MI, | |||
112 | MachineOperand &FlagUse, CondRegArray &CondRegs); | |||
113 | void rewriteCMov(MachineBasicBlock &TestMBB, | |||
114 | MachineBasicBlock::iterator TestPos, const DebugLoc &TestLoc, | |||
115 | MachineInstr &CMovI, MachineOperand &FlagUse, | |||
116 | CondRegArray &CondRegs); | |||
117 | void rewriteFCMov(MachineBasicBlock &TestMBB, | |||
118 | MachineBasicBlock::iterator TestPos, | |||
119 | const DebugLoc &TestLoc, MachineInstr &CMovI, | |||
120 | MachineOperand &FlagUse, CondRegArray &CondRegs); | |||
121 | void rewriteCondJmp(MachineBasicBlock &TestMBB, | |||
122 | MachineBasicBlock::iterator TestPos, | |||
123 | const DebugLoc &TestLoc, MachineInstr &JmpI, | |||
124 | CondRegArray &CondRegs); | |||
125 | void rewriteCopy(MachineInstr &MI, MachineOperand &FlagUse, | |||
126 | MachineInstr &CopyDefI); | |||
127 | void rewriteSetCC(MachineBasicBlock &TestMBB, | |||
128 | MachineBasicBlock::iterator TestPos, | |||
129 | const DebugLoc &TestLoc, MachineInstr &SetCCI, | |||
130 | MachineOperand &FlagUse, CondRegArray &CondRegs); | |||
131 | }; | |||
132 | ||||
133 | } // end anonymous namespace | |||
134 | ||||
135 | INITIALIZE_PASS_BEGIN(X86FlagsCopyLoweringPass, DEBUG_TYPE,static void *initializeX86FlagsCopyLoweringPassPassOnce(PassRegistry &Registry) { | |||
136 | "X86 EFLAGS copy lowering", false, false)static void *initializeX86FlagsCopyLoweringPassPassOnce(PassRegistry &Registry) { | |||
137 | INITIALIZE_PASS_END(X86FlagsCopyLoweringPass, DEBUG_TYPE,PassInfo *PI = new PassInfo( "X86 EFLAGS copy lowering", "x86-flags-copy-lowering" , &X86FlagsCopyLoweringPass::ID, PassInfo::NormalCtor_t(callDefaultCtor <X86FlagsCopyLoweringPass>), false, false); Registry.registerPass (*PI, true); return PI; } static llvm::once_flag InitializeX86FlagsCopyLoweringPassPassFlag ; void llvm::initializeX86FlagsCopyLoweringPassPass(PassRegistry &Registry) { llvm::call_once(InitializeX86FlagsCopyLoweringPassPassFlag , initializeX86FlagsCopyLoweringPassPassOnce, std::ref(Registry )); } | |||
138 | "X86 EFLAGS copy lowering", false, false)PassInfo *PI = new PassInfo( "X86 EFLAGS copy lowering", "x86-flags-copy-lowering" , &X86FlagsCopyLoweringPass::ID, PassInfo::NormalCtor_t(callDefaultCtor <X86FlagsCopyLoweringPass>), false, false); Registry.registerPass (*PI, true); return PI; } static llvm::once_flag InitializeX86FlagsCopyLoweringPassPassFlag ; void llvm::initializeX86FlagsCopyLoweringPassPass(PassRegistry &Registry) { llvm::call_once(InitializeX86FlagsCopyLoweringPassPassFlag , initializeX86FlagsCopyLoweringPassPassOnce, std::ref(Registry )); } | |||
139 | ||||
140 | FunctionPass *llvm::createX86FlagsCopyLoweringPass() { | |||
141 | return new X86FlagsCopyLoweringPass(); | |||
142 | } | |||
143 | ||||
144 | char X86FlagsCopyLoweringPass::ID = 0; | |||
145 | ||||
146 | void X86FlagsCopyLoweringPass::getAnalysisUsage(AnalysisUsage &AU) const { | |||
147 | AU.addRequired<MachineDominatorTree>(); | |||
148 | MachineFunctionPass::getAnalysisUsage(AU); | |||
149 | } | |||
150 | ||||
151 | namespace { | |||
152 | /// An enumeration of the arithmetic instruction mnemonics which have | |||
153 | /// interesting flag semantics. | |||
154 | /// | |||
155 | /// We can map instruction opcodes into these mnemonics to make it easy to | |||
156 | /// dispatch with specific functionality. | |||
157 | enum class FlagArithMnemonic { | |||
158 | ADC, | |||
159 | ADCX, | |||
160 | ADOX, | |||
161 | RCL, | |||
162 | RCR, | |||
163 | SBB, | |||
164 | SETB, | |||
165 | }; | |||
166 | } // namespace | |||
167 | ||||
168 | static FlagArithMnemonic getMnemonicFromOpcode(unsigned Opcode) { | |||
169 | switch (Opcode) { | |||
170 | default: | |||
171 | report_fatal_error("No support for lowering a copy into EFLAGS when used " | |||
172 | "by this instruction!"); | |||
173 | ||||
174 | #define LLVM_EXPAND_INSTR_SIZES(MNEMONIC, SUFFIX) \ | |||
175 | case X86::MNEMONIC##8##SUFFIX: \ | |||
176 | case X86::MNEMONIC##16##SUFFIX: \ | |||
177 | case X86::MNEMONIC##32##SUFFIX: \ | |||
178 | case X86::MNEMONIC##64##SUFFIX: | |||
179 | ||||
180 | #define LLVM_EXPAND_ADC_SBB_INSTR(MNEMONIC) \ | |||
181 | LLVM_EXPAND_INSTR_SIZES(MNEMONIC, rr) \ | |||
182 | LLVM_EXPAND_INSTR_SIZES(MNEMONIC, rr_REV) \ | |||
183 | LLVM_EXPAND_INSTR_SIZES(MNEMONIC, rm) \ | |||
184 | LLVM_EXPAND_INSTR_SIZES(MNEMONIC, mr) \ | |||
185 | case X86::MNEMONIC##8ri: \ | |||
186 | case X86::MNEMONIC##16ri8: \ | |||
187 | case X86::MNEMONIC##32ri8: \ | |||
188 | case X86::MNEMONIC##64ri8: \ | |||
189 | case X86::MNEMONIC##16ri: \ | |||
190 | case X86::MNEMONIC##32ri: \ | |||
191 | case X86::MNEMONIC##64ri32: \ | |||
192 | case X86::MNEMONIC##8mi: \ | |||
193 | case X86::MNEMONIC##16mi8: \ | |||
194 | case X86::MNEMONIC##32mi8: \ | |||
195 | case X86::MNEMONIC##64mi8: \ | |||
196 | case X86::MNEMONIC##16mi: \ | |||
197 | case X86::MNEMONIC##32mi: \ | |||
198 | case X86::MNEMONIC##64mi32: \ | |||
199 | case X86::MNEMONIC##8i8: \ | |||
200 | case X86::MNEMONIC##16i16: \ | |||
201 | case X86::MNEMONIC##32i32: \ | |||
202 | case X86::MNEMONIC##64i32: | |||
203 | ||||
204 | LLVM_EXPAND_ADC_SBB_INSTR(ADC) | |||
205 | return FlagArithMnemonic::ADC; | |||
206 | ||||
207 | LLVM_EXPAND_ADC_SBB_INSTR(SBB) | |||
208 | return FlagArithMnemonic::SBB; | |||
209 | ||||
210 | #undef LLVM_EXPAND_ADC_SBB_INSTR | |||
211 | ||||
212 | LLVM_EXPAND_INSTR_SIZES(RCL, rCL) | |||
213 | LLVM_EXPAND_INSTR_SIZES(RCL, r1) | |||
214 | LLVM_EXPAND_INSTR_SIZES(RCL, ri) | |||
215 | return FlagArithMnemonic::RCL; | |||
216 | ||||
217 | LLVM_EXPAND_INSTR_SIZES(RCR, rCL) | |||
218 | LLVM_EXPAND_INSTR_SIZES(RCR, r1) | |||
219 | LLVM_EXPAND_INSTR_SIZES(RCR, ri) | |||
220 | return FlagArithMnemonic::RCR; | |||
221 | ||||
222 | #undef LLVM_EXPAND_INSTR_SIZES | |||
223 | ||||
224 | case X86::ADCX32rr: | |||
225 | case X86::ADCX64rr: | |||
226 | case X86::ADCX32rm: | |||
227 | case X86::ADCX64rm: | |||
228 | return FlagArithMnemonic::ADCX; | |||
229 | ||||
230 | case X86::ADOX32rr: | |||
231 | case X86::ADOX64rr: | |||
232 | case X86::ADOX32rm: | |||
233 | case X86::ADOX64rm: | |||
234 | return FlagArithMnemonic::ADOX; | |||
235 | ||||
236 | case X86::SETB_C32r: | |||
237 | case X86::SETB_C64r: | |||
238 | return FlagArithMnemonic::SETB; | |||
239 | } | |||
240 | } | |||
241 | ||||
242 | static MachineBasicBlock &splitBlock(MachineBasicBlock &MBB, | |||
243 | MachineInstr &SplitI, | |||
244 | const X86InstrInfo &TII) { | |||
245 | MachineFunction &MF = *MBB.getParent(); | |||
246 | ||||
247 | assert(SplitI.getParent() == &MBB &&((void)0) | |||
248 | "Split instruction must be in the split block!")((void)0); | |||
249 | assert(SplitI.isBranch() &&((void)0) | |||
250 | "Only designed to split a tail of branch instructions!")((void)0); | |||
251 | assert(X86::getCondFromBranch(SplitI) != X86::COND_INVALID &&((void)0) | |||
252 | "Must split on an actual jCC instruction!")((void)0); | |||
253 | ||||
254 | // Dig out the previous instruction to the split point. | |||
255 | MachineInstr &PrevI = *std::prev(SplitI.getIterator()); | |||
256 | assert(PrevI.isBranch() && "Must split after a branch!")((void)0); | |||
257 | assert(X86::getCondFromBranch(PrevI) != X86::COND_INVALID &&((void)0) | |||
258 | "Must split after an actual jCC instruction!")((void)0); | |||
259 | assert(!std::prev(PrevI.getIterator())->isTerminator() &&((void)0) | |||
260 | "Must only have this one terminator prior to the split!")((void)0); | |||
261 | ||||
262 | // Grab the one successor edge that will stay in `MBB`. | |||
263 | MachineBasicBlock &UnsplitSucc = *PrevI.getOperand(0).getMBB(); | |||
264 | ||||
265 | // Analyze the original block to see if we are actually splitting an edge | |||
266 | // into two edges. This can happen when we have multiple conditional jumps to | |||
267 | // the same successor. | |||
268 | bool IsEdgeSplit = | |||
269 | std::any_of(SplitI.getIterator(), MBB.instr_end(), | |||
270 | [&](MachineInstr &MI) { | |||
271 | assert(MI.isTerminator() &&((void)0) | |||
272 | "Should only have spliced terminators!")((void)0); | |||
273 | return llvm::any_of( | |||
274 | MI.operands(), [&](MachineOperand &MOp) { | |||
275 | return MOp.isMBB() && MOp.getMBB() == &UnsplitSucc; | |||
276 | }); | |||
277 | }) || | |||
278 | MBB.getFallThrough() == &UnsplitSucc; | |||
279 | ||||
280 | MachineBasicBlock &NewMBB = *MF.CreateMachineBasicBlock(); | |||
281 | ||||
282 | // Insert the new block immediately after the current one. Any existing | |||
283 | // fallthrough will be sunk into this new block anyways. | |||
284 | MF.insert(std::next(MachineFunction::iterator(&MBB)), &NewMBB); | |||
285 | ||||
286 | // Splice the tail of instructions into the new block. | |||
287 | NewMBB.splice(NewMBB.end(), &MBB, SplitI.getIterator(), MBB.end()); | |||
288 | ||||
289 | // Copy the necessary succesors (and their probability info) into the new | |||
290 | // block. | |||
291 | for (auto SI = MBB.succ_begin(), SE = MBB.succ_end(); SI != SE; ++SI) | |||
292 | if (IsEdgeSplit || *SI != &UnsplitSucc) | |||
293 | NewMBB.copySuccessor(&MBB, SI); | |||
294 | // Normalize the probabilities if we didn't end up splitting the edge. | |||
295 | if (!IsEdgeSplit) | |||
296 | NewMBB.normalizeSuccProbs(); | |||
297 | ||||
298 | // Now replace all of the moved successors in the original block with the new | |||
299 | // block. This will merge their probabilities. | |||
300 | for (MachineBasicBlock *Succ : NewMBB.successors()) | |||
301 | if (Succ != &UnsplitSucc) | |||
302 | MBB.replaceSuccessor(Succ, &NewMBB); | |||
303 | ||||
304 | // We should always end up replacing at least one successor. | |||
305 | assert(MBB.isSuccessor(&NewMBB) &&((void)0) | |||
306 | "Failed to make the new block a successor!")((void)0); | |||
307 | ||||
308 | // Now update all the PHIs. | |||
309 | for (MachineBasicBlock *Succ : NewMBB.successors()) { | |||
310 | for (MachineInstr &MI : *Succ) { | |||
311 | if (!MI.isPHI()) | |||
312 | break; | |||
313 | ||||
314 | for (int OpIdx = 1, NumOps = MI.getNumOperands(); OpIdx < NumOps; | |||
315 | OpIdx += 2) { | |||
316 | MachineOperand &OpV = MI.getOperand(OpIdx); | |||
317 | MachineOperand &OpMBB = MI.getOperand(OpIdx + 1); | |||
318 | assert(OpMBB.isMBB() && "Block operand to a PHI is not a block!")((void)0); | |||
319 | if (OpMBB.getMBB() != &MBB) | |||
320 | continue; | |||
321 | ||||
322 | // Replace the operand for unsplit successors | |||
323 | if (!IsEdgeSplit || Succ != &UnsplitSucc) { | |||
324 | OpMBB.setMBB(&NewMBB); | |||
325 | ||||
326 | // We have to continue scanning as there may be multiple entries in | |||
327 | // the PHI. | |||
328 | continue; | |||
329 | } | |||
330 | ||||
331 | // When we have split the edge append a new successor. | |||
332 | MI.addOperand(MF, OpV); | |||
333 | MI.addOperand(MF, MachineOperand::CreateMBB(&NewMBB)); | |||
334 | break; | |||
335 | } | |||
336 | } | |||
337 | } | |||
338 | ||||
339 | return NewMBB; | |||
340 | } | |||
341 | ||||
342 | static X86::CondCode getCondFromFCMOV(unsigned Opcode) { | |||
343 | switch (Opcode) { | |||
344 | default: return X86::COND_INVALID; | |||
345 | case X86::CMOVBE_Fp32: case X86::CMOVBE_Fp64: case X86::CMOVBE_Fp80: | |||
346 | return X86::COND_BE; | |||
347 | case X86::CMOVB_Fp32: case X86::CMOVB_Fp64: case X86::CMOVB_Fp80: | |||
348 | return X86::COND_B; | |||
349 | case X86::CMOVE_Fp32: case X86::CMOVE_Fp64: case X86::CMOVE_Fp80: | |||
350 | return X86::COND_E; | |||
351 | case X86::CMOVNBE_Fp32: case X86::CMOVNBE_Fp64: case X86::CMOVNBE_Fp80: | |||
352 | return X86::COND_A; | |||
353 | case X86::CMOVNB_Fp32: case X86::CMOVNB_Fp64: case X86::CMOVNB_Fp80: | |||
354 | return X86::COND_AE; | |||
355 | case X86::CMOVNE_Fp32: case X86::CMOVNE_Fp64: case X86::CMOVNE_Fp80: | |||
356 | return X86::COND_NE; | |||
357 | case X86::CMOVNP_Fp32: case X86::CMOVNP_Fp64: case X86::CMOVNP_Fp80: | |||
358 | return X86::COND_NP; | |||
359 | case X86::CMOVP_Fp32: case X86::CMOVP_Fp64: case X86::CMOVP_Fp80: | |||
360 | return X86::COND_P; | |||
361 | } | |||
362 | } | |||
363 | ||||
364 | bool X86FlagsCopyLoweringPass::runOnMachineFunction(MachineFunction &MF) { | |||
365 | LLVM_DEBUG(dbgs() << "********** " << getPassName() << " : " << MF.getName()do { } while (false) | |||
| ||||
366 | << " **********\n")do { } while (false); | |||
367 | ||||
368 | Subtarget = &MF.getSubtarget<X86Subtarget>(); | |||
369 | MRI = &MF.getRegInfo(); | |||
370 | TII = Subtarget->getInstrInfo(); | |||
371 | TRI = Subtarget->getRegisterInfo(); | |||
372 | MDT = &getAnalysis<MachineDominatorTree>(); | |||
373 | PromoteRC = &X86::GR8RegClass; | |||
374 | ||||
375 | if (MF.begin() == MF.end()) | |||
376 | // Nothing to do for a degenerate empty function... | |||
377 | return false; | |||
378 | ||||
379 | // Collect the copies in RPO so that when there are chains where a copy is in | |||
380 | // turn copied again we visit the first one first. This ensures we can find | |||
381 | // viable locations for testing the original EFLAGS that dominate all the | |||
382 | // uses across complex CFGs. | |||
383 | SmallVector<MachineInstr *, 4> Copies; | |||
384 | ReversePostOrderTraversal<MachineFunction *> RPOT(&MF); | |||
385 | for (MachineBasicBlock *MBB : RPOT) | |||
386 | for (MachineInstr &MI : *MBB) | |||
387 | if (MI.getOpcode() == TargetOpcode::COPY && | |||
388 | MI.getOperand(0).getReg() == X86::EFLAGS) | |||
389 | Copies.push_back(&MI); | |||
390 | ||||
391 | for (MachineInstr *CopyI : Copies) { | |||
392 | MachineBasicBlock &MBB = *CopyI->getParent(); | |||
393 | ||||
394 | MachineOperand &VOp = CopyI->getOperand(1); | |||
395 | assert(VOp.isReg() &&((void)0) | |||
396 | "The input to the copy for EFLAGS should always be a register!")((void)0); | |||
397 | MachineInstr &CopyDefI = *MRI->getVRegDef(VOp.getReg()); | |||
398 | if (CopyDefI.getOpcode() != TargetOpcode::COPY) { | |||
399 | // FIXME: The big likely candidate here are PHI nodes. We could in theory | |||
400 | // handle PHI nodes, but it gets really, really hard. Insanely hard. Hard | |||
401 | // enough that it is probably better to change every other part of LLVM | |||
402 | // to avoid creating them. The issue is that once we have PHIs we won't | |||
403 | // know which original EFLAGS value we need to capture with our setCCs | |||
404 | // below. The end result will be computing a complete set of setCCs that | |||
405 | // we *might* want, computing them in every place where we copy *out* of | |||
406 | // EFLAGS and then doing SSA formation on all of them to insert necessary | |||
407 | // PHI nodes and consume those here. Then hoping that somehow we DCE the | |||
408 | // unnecessary ones. This DCE seems very unlikely to be successful and so | |||
409 | // we will almost certainly end up with a glut of dead setCC | |||
410 | // instructions. Until we have a motivating test case and fail to avoid | |||
411 | // it by changing other parts of LLVM's lowering, we refuse to handle | |||
412 | // this complex case here. | |||
413 | LLVM_DEBUG(do { } while (false) | |||
414 | dbgs() << "ERROR: Encountered unexpected def of an eflags copy: ";do { } while (false) | |||
415 | CopyDefI.dump())do { } while (false); | |||
416 | report_fatal_error( | |||
417 | "Cannot lower EFLAGS copy unless it is defined in turn by a copy!"); | |||
418 | } | |||
419 | ||||
420 | auto Cleanup = make_scope_exit([&] { | |||
421 | // All uses of the EFLAGS copy are now rewritten, kill the copy into | |||
422 | // eflags and if dead the copy from. | |||
423 | CopyI->eraseFromParent(); | |||
424 | if (MRI->use_empty(CopyDefI.getOperand(0).getReg())) | |||
425 | CopyDefI.eraseFromParent(); | |||
426 | ++NumCopiesEliminated; | |||
427 | }); | |||
428 | ||||
429 | MachineOperand &DOp = CopyI->getOperand(0); | |||
430 | assert(DOp.isDef() && "Expected register def!")((void)0); | |||
431 | assert(DOp.getReg() == X86::EFLAGS && "Unexpected copy def register!")((void)0); | |||
432 | if (DOp.isDead()) | |||
433 | continue; | |||
434 | ||||
435 | MachineBasicBlock *TestMBB = CopyDefI.getParent(); | |||
436 | auto TestPos = CopyDefI.getIterator(); | |||
437 | DebugLoc TestLoc = CopyDefI.getDebugLoc(); | |||
438 | ||||
439 | LLVM_DEBUG(dbgs() << "Rewriting copy: "; CopyI->dump())do { } while (false); | |||
440 | ||||
441 | // Walk up across live-in EFLAGS to find where they were actually def'ed. | |||
442 | // | |||
443 | // This copy's def may just be part of a region of blocks covered by | |||
444 | // a single def of EFLAGS and we want to find the top of that region where | |||
445 | // possible. | |||
446 | // | |||
447 | // This is essentially a search for a *candidate* reaching definition | |||
448 | // location. We don't need to ever find the actual reaching definition here, | |||
449 | // but we want to walk up the dominator tree to find the highest point which | |||
450 | // would be viable for such a definition. | |||
451 | auto HasEFLAGSClobber = [&](MachineBasicBlock::iterator Begin, | |||
452 | MachineBasicBlock::iterator End) { | |||
453 | // Scan backwards as we expect these to be relatively short and often find | |||
454 | // a clobber near the end. | |||
455 | return llvm::any_of( | |||
456 | llvm::reverse(llvm::make_range(Begin, End)), [&](MachineInstr &MI) { | |||
457 | // Flag any instruction (other than the copy we are | |||
458 | // currently rewriting) that defs EFLAGS. | |||
459 | return &MI != CopyI && MI.findRegisterDefOperand(X86::EFLAGS); | |||
460 | }); | |||
461 | }; | |||
462 | auto HasEFLAGSClobberPath = [&](MachineBasicBlock *BeginMBB, | |||
463 | MachineBasicBlock *EndMBB) { | |||
464 | assert(MDT->dominates(BeginMBB, EndMBB) &&((void)0) | |||
465 | "Only support paths down the dominator tree!")((void)0); | |||
466 | SmallPtrSet<MachineBasicBlock *, 4> Visited; | |||
467 | SmallVector<MachineBasicBlock *, 4> Worklist; | |||
468 | // We terminate at the beginning. No need to scan it. | |||
469 | Visited.insert(BeginMBB); | |||
470 | Worklist.push_back(EndMBB); | |||
471 | do { | |||
472 | auto *MBB = Worklist.pop_back_val(); | |||
473 | for (auto *PredMBB : MBB->predecessors()) { | |||
474 | if (!Visited.insert(PredMBB).second) | |||
475 | continue; | |||
476 | if (HasEFLAGSClobber(PredMBB->begin(), PredMBB->end())) | |||
477 | return true; | |||
478 | // Enqueue this block to walk its predecessors. | |||
479 | Worklist.push_back(PredMBB); | |||
480 | } | |||
481 | } while (!Worklist.empty()); | |||
482 | // No clobber found along a path from the begin to end. | |||
483 | return false; | |||
484 | }; | |||
485 | while (TestMBB->isLiveIn(X86::EFLAGS) && !TestMBB->pred_empty() && | |||
486 | !HasEFLAGSClobber(TestMBB->begin(), TestPos)) { | |||
487 | // Find the nearest common dominator of the predecessors, as | |||
488 | // that will be the best candidate to hoist into. | |||
489 | MachineBasicBlock *HoistMBB = | |||
490 | std::accumulate(std::next(TestMBB->pred_begin()), TestMBB->pred_end(), | |||
491 | *TestMBB->pred_begin(), | |||
492 | [&](MachineBasicBlock *LHS, MachineBasicBlock *RHS) { | |||
493 | return MDT->findNearestCommonDominator(LHS, RHS); | |||
494 | }); | |||
495 | ||||
496 | // Now we need to scan all predecessors that may be reached along paths to | |||
497 | // the hoist block. A clobber anywhere in any of these blocks the hoist. | |||
498 | // Note that this even handles loops because we require *no* clobbers. | |||
499 | if (HasEFLAGSClobberPath(HoistMBB, TestMBB)) | |||
500 | break; | |||
501 | ||||
502 | // We also need the terminators to not sneakily clobber flags. | |||
503 | if (HasEFLAGSClobber(HoistMBB->getFirstTerminator()->getIterator(), | |||
504 | HoistMBB->instr_end())) | |||
505 | break; | |||
506 | ||||
507 | // We found a viable location, hoist our test position to it. | |||
508 | TestMBB = HoistMBB; | |||
509 | TestPos = TestMBB->getFirstTerminator()->getIterator(); | |||
510 | // Clear the debug location as it would just be confusing after hoisting. | |||
511 | TestLoc = DebugLoc(); | |||
512 | } | |||
513 | LLVM_DEBUG({do { } while (false) | |||
514 | auto DefIt = llvm::find_if(do { } while (false) | |||
515 | llvm::reverse(llvm::make_range(TestMBB->instr_begin(), TestPos)),do { } while (false) | |||
516 | [&](MachineInstr &MI) {do { } while (false) | |||
517 | return MI.findRegisterDefOperand(X86::EFLAGS);do { } while (false) | |||
518 | });do { } while (false) | |||
519 | if (DefIt.base() != TestMBB->instr_begin()) {do { } while (false) | |||
520 | dbgs() << " Using EFLAGS defined by: ";do { } while (false) | |||
521 | DefIt->dump();do { } while (false) | |||
522 | } else {do { } while (false) | |||
523 | dbgs() << " Using live-in flags for BB:\n";do { } while (false) | |||
524 | TestMBB->dump();do { } while (false) | |||
525 | }do { } while (false) | |||
526 | })do { } while (false); | |||
527 | ||||
528 | // While rewriting uses, we buffer jumps and rewrite them in a second pass | |||
529 | // because doing so will perturb the CFG that we are walking to find the | |||
530 | // uses in the first place. | |||
531 | SmallVector<MachineInstr *, 4> JmpIs; | |||
532 | ||||
533 | // Gather the condition flags that have already been preserved in | |||
534 | // registers. We do this from scratch each time as we expect there to be | |||
535 | // very few of them and we expect to not revisit the same copy definition | |||
536 | // many times. If either of those change sufficiently we could build a map | |||
537 | // of these up front instead. | |||
538 | CondRegArray CondRegs = collectCondsInRegs(*TestMBB, TestPos); | |||
539 | ||||
540 | // Collect the basic blocks we need to scan. Typically this will just be | |||
541 | // a single basic block but we may have to scan multiple blocks if the | |||
542 | // EFLAGS copy lives into successors. | |||
543 | SmallVector<MachineBasicBlock *, 2> Blocks; | |||
544 | SmallPtrSet<MachineBasicBlock *, 2> VisitedBlocks; | |||
545 | Blocks.push_back(&MBB); | |||
546 | ||||
547 | do { | |||
548 | MachineBasicBlock &UseMBB = *Blocks.pop_back_val(); | |||
549 | ||||
550 | // Track when if/when we find a kill of the flags in this block. | |||
551 | bool FlagsKilled = false; | |||
552 | ||||
553 | // In most cases, we walk from the beginning to the end of the block. But | |||
554 | // when the block is the same block as the copy is from, we will visit it | |||
555 | // twice. The first time we start from the copy and go to the end. The | |||
556 | // second time we start from the beginning and go to the copy. This lets | |||
557 | // us handle copies inside of cycles. | |||
558 | // FIXME: This loop is *super* confusing. This is at least in part | |||
559 | // a symptom of all of this routine needing to be refactored into | |||
560 | // documentable components. Once done, there may be a better way to write | |||
561 | // this loop. | |||
562 | for (auto MII = (&UseMBB == &MBB && !VisitedBlocks.count(&UseMBB)) | |||
563 | ? std::next(CopyI->getIterator()) | |||
564 | : UseMBB.instr_begin(), | |||
565 | MIE = UseMBB.instr_end(); | |||
566 | MII != MIE;) { | |||
567 | MachineInstr &MI = *MII++; | |||
568 | // If we are in the original copy block and encounter either the copy | |||
569 | // def or the copy itself, break so that we don't re-process any part of | |||
570 | // the block or process the instructions in the range that was copied | |||
571 | // over. | |||
572 | if (&MI == CopyI || &MI == &CopyDefI) { | |||
573 | assert(&UseMBB == &MBB && VisitedBlocks.count(&MBB) &&((void)0) | |||
574 | "Should only encounter these on the second pass over the "((void)0) | |||
575 | "original block.")((void)0); | |||
576 | break; | |||
577 | } | |||
578 | ||||
579 | MachineOperand *FlagUse = MI.findRegisterUseOperand(X86::EFLAGS); | |||
580 | if (!FlagUse) { | |||
581 | if (MI.findRegisterDefOperand(X86::EFLAGS)) { | |||
582 | // If EFLAGS are defined, it's as-if they were killed. We can stop | |||
583 | // scanning here. | |||
584 | // | |||
585 | // NB!!! Many instructions only modify some flags. LLVM currently | |||
586 | // models this as clobbering all flags, but if that ever changes | |||
587 | // this will need to be carefully updated to handle that more | |||
588 | // complex logic. | |||
589 | FlagsKilled = true; | |||
590 | break; | |||
591 | } | |||
592 | continue; | |||
593 | } | |||
594 | ||||
595 | LLVM_DEBUG(dbgs() << " Rewriting use: "; MI.dump())do { } while (false); | |||
596 | ||||
597 | // Check the kill flag before we rewrite as that may change it. | |||
598 | if (FlagUse->isKill()) | |||
599 | FlagsKilled = true; | |||
600 | ||||
601 | // Once we encounter a branch, the rest of the instructions must also be | |||
602 | // branches. We can't rewrite in place here, so we handle them below. | |||
603 | // | |||
604 | // Note that we don't have to handle tail calls here, even conditional | |||
605 | // tail calls, as those are not introduced into the X86 MI until post-RA | |||
606 | // branch folding or black placement. As a consequence, we get to deal | |||
607 | // with the simpler formulation of conditional branches followed by tail | |||
608 | // calls. | |||
609 | if (X86::getCondFromBranch(MI) != X86::COND_INVALID) { | |||
610 | auto JmpIt = MI.getIterator(); | |||
611 | do { | |||
612 | JmpIs.push_back(&*JmpIt); | |||
613 | ++JmpIt; | |||
614 | } while (JmpIt != UseMBB.instr_end() && | |||
615 | X86::getCondFromBranch(*JmpIt) != | |||
616 | X86::COND_INVALID); | |||
617 | break; | |||
618 | } | |||
619 | ||||
620 | // Otherwise we can just rewrite in-place. | |||
621 | if (X86::getCondFromCMov(MI) != X86::COND_INVALID) { | |||
622 | rewriteCMov(*TestMBB, TestPos, TestLoc, MI, *FlagUse, CondRegs); | |||
623 | } else if (getCondFromFCMOV(MI.getOpcode()) != X86::COND_INVALID) { | |||
624 | rewriteFCMov(*TestMBB, TestPos, TestLoc, MI, *FlagUse, CondRegs); | |||
625 | } else if (X86::getCondFromSETCC(MI) != X86::COND_INVALID) { | |||
626 | rewriteSetCC(*TestMBB, TestPos, TestLoc, MI, *FlagUse, CondRegs); | |||
627 | } else if (MI.getOpcode() == TargetOpcode::COPY) { | |||
628 | rewriteCopy(MI, *FlagUse, CopyDefI); | |||
629 | } else { | |||
630 | // We assume all other instructions that use flags also def them. | |||
631 | assert(MI.findRegisterDefOperand(X86::EFLAGS) &&((void)0) | |||
632 | "Expected a def of EFLAGS for this instruction!")((void)0); | |||
633 | ||||
634 | // NB!!! Several arithmetic instructions only *partially* update | |||
635 | // flags. Theoretically, we could generate MI code sequences that | |||
636 | // would rely on this fact and observe different flags independently. | |||
637 | // But currently LLVM models all of these instructions as clobbering | |||
638 | // all the flags in an undef way. We rely on that to simplify the | |||
639 | // logic. | |||
640 | FlagsKilled = true; | |||
641 | ||||
642 | // Generically handle remaining uses as arithmetic instructions. | |||
643 | rewriteArithmetic(*TestMBB, TestPos, TestLoc, MI, *FlagUse, | |||
644 | CondRegs); | |||
645 | } | |||
646 | ||||
647 | // If this was the last use of the flags, we're done. | |||
648 | if (FlagsKilled) | |||
649 | break; | |||
650 | } | |||
651 | ||||
652 | // If the flags were killed, we're done with this block. | |||
653 | if (FlagsKilled) | |||
654 | continue; | |||
655 | ||||
656 | // Otherwise we need to scan successors for ones where the flags live-in | |||
657 | // and queue those up for processing. | |||
658 | for (MachineBasicBlock *SuccMBB : UseMBB.successors()) | |||
659 | if (SuccMBB->isLiveIn(X86::EFLAGS) && | |||
660 | VisitedBlocks.insert(SuccMBB).second) { | |||
661 | // We currently don't do any PHI insertion and so we require that the | |||
662 | // test basic block dominates all of the use basic blocks. Further, we | |||
663 | // can't have a cycle from the test block back to itself as that would | |||
664 | // create a cycle requiring a PHI to break it. | |||
665 | // | |||
666 | // We could in theory do PHI insertion here if it becomes useful by | |||
667 | // just taking undef values in along every edge that we don't trace | |||
668 | // this EFLAGS copy along. This isn't as bad as fully general PHI | |||
669 | // insertion, but still seems like a great deal of complexity. | |||
670 | // | |||
671 | // Because it is theoretically possible that some earlier MI pass or | |||
672 | // other lowering transformation could induce this to happen, we do | |||
673 | // a hard check even in non-debug builds here. | |||
674 | if (SuccMBB == TestMBB || !MDT->dominates(TestMBB, SuccMBB)) { | |||
675 | LLVM_DEBUG({do { } while (false) | |||
676 | dbgs()do { } while (false) | |||
677 | << "ERROR: Encountered use that is not dominated by our test "do { } while (false) | |||
678 | "basic block! Rewriting this would require inserting PHI "do { } while (false) | |||
679 | "nodes to track the flag state across the CFG.\n\nTest "do { } while (false) | |||
680 | "block:\n";do { } while (false) | |||
681 | TestMBB->dump();do { } while (false) | |||
682 | dbgs() << "Use block:\n";do { } while (false) | |||
683 | SuccMBB->dump();do { } while (false) | |||
684 | })do { } while (false); | |||
685 | report_fatal_error( | |||
686 | "Cannot lower EFLAGS copy when original copy def " | |||
687 | "does not dominate all uses."); | |||
688 | } | |||
689 | ||||
690 | Blocks.push_back(SuccMBB); | |||
691 | ||||
692 | // After this, EFLAGS will be recreated before each use. | |||
693 | SuccMBB->removeLiveIn(X86::EFLAGS); | |||
694 | } | |||
695 | } while (!Blocks.empty()); | |||
696 | ||||
697 | // Now rewrite the jumps that use the flags. These we handle specially | |||
698 | // because if there are multiple jumps in a single basic block we'll have | |||
699 | // to do surgery on the CFG. | |||
700 | MachineBasicBlock *LastJmpMBB = nullptr; | |||
701 | for (MachineInstr *JmpI : JmpIs) { | |||
702 | // Past the first jump within a basic block we need to split the blocks | |||
703 | // apart. | |||
704 | if (JmpI->getParent() == LastJmpMBB) | |||
705 | splitBlock(*JmpI->getParent(), *JmpI, *TII); | |||
706 | else | |||
707 | LastJmpMBB = JmpI->getParent(); | |||
708 | ||||
709 | rewriteCondJmp(*TestMBB, TestPos, TestLoc, *JmpI, CondRegs); | |||
710 | } | |||
711 | ||||
712 | // FIXME: Mark the last use of EFLAGS before the copy's def as a kill if | |||
713 | // the copy's def operand is itself a kill. | |||
714 | } | |||
715 | ||||
716 | #ifndef NDEBUG1 | |||
717 | for (MachineBasicBlock &MBB : MF) | |||
718 | for (MachineInstr &MI : MBB) | |||
719 | if (MI.getOpcode() == TargetOpcode::COPY && | |||
720 | (MI.getOperand(0).getReg() == X86::EFLAGS || | |||
721 | MI.getOperand(1).getReg() == X86::EFLAGS)) { | |||
722 | LLVM_DEBUG(dbgs() << "ERROR: Found a COPY involving EFLAGS: ";do { } while (false) | |||
723 | MI.dump())do { } while (false); | |||
724 | llvm_unreachable("Unlowered EFLAGS copy!")__builtin_unreachable(); | |||
725 | } | |||
726 | #endif | |||
727 | ||||
728 | return true; | |||
729 | } | |||
730 | ||||
731 | /// Collect any conditions that have already been set in registers so that we | |||
732 | /// can re-use them rather than adding duplicates. | |||
733 | CondRegArray X86FlagsCopyLoweringPass::collectCondsInRegs( | |||
734 | MachineBasicBlock &MBB, MachineBasicBlock::iterator TestPos) { | |||
735 | CondRegArray CondRegs = {}; | |||
736 | ||||
737 | // Scan backwards across the range of instructions with live EFLAGS. | |||
738 | for (MachineInstr &MI : | |||
739 | llvm::reverse(llvm::make_range(MBB.begin(), TestPos))) { | |||
740 | X86::CondCode Cond = X86::getCondFromSETCC(MI); | |||
741 | if (Cond != X86::COND_INVALID && !MI.mayStore() && | |||
742 | MI.getOperand(0).isReg() && MI.getOperand(0).getReg().isVirtual()) { | |||
743 | assert(MI.getOperand(0).isDef() &&((void)0) | |||
744 | "A non-storing SETcc should always define a register!")((void)0); | |||
745 | CondRegs[Cond] = MI.getOperand(0).getReg(); | |||
746 | } | |||
747 | ||||
748 | // Stop scanning when we see the first definition of the EFLAGS as prior to | |||
749 | // this we would potentially capture the wrong flag state. | |||
750 | if (MI.findRegisterDefOperand(X86::EFLAGS)) | |||
751 | break; | |||
752 | } | |||
753 | return CondRegs; | |||
754 | } | |||
755 | ||||
756 | Register X86FlagsCopyLoweringPass::promoteCondToReg( | |||
757 | MachineBasicBlock &TestMBB, MachineBasicBlock::iterator TestPos, | |||
758 | const DebugLoc &TestLoc, X86::CondCode Cond) { | |||
759 | Register Reg = MRI->createVirtualRegister(PromoteRC); | |||
760 | auto SetI = BuildMI(TestMBB, TestPos, TestLoc, | |||
761 | TII->get(X86::SETCCr), Reg).addImm(Cond); | |||
762 | (void)SetI; | |||
763 | LLVM_DEBUG(dbgs() << " save cond: "; SetI->dump())do { } while (false); | |||
764 | ++NumSetCCsInserted; | |||
765 | return Reg; | |||
766 | } | |||
767 | ||||
768 | std::pair<unsigned, bool> X86FlagsCopyLoweringPass::getCondOrInverseInReg( | |||
769 | MachineBasicBlock &TestMBB, MachineBasicBlock::iterator TestPos, | |||
770 | const DebugLoc &TestLoc, X86::CondCode Cond, CondRegArray &CondRegs) { | |||
771 | unsigned &CondReg = CondRegs[Cond]; | |||
772 | unsigned &InvCondReg = CondRegs[X86::GetOppositeBranchCondition(Cond)]; | |||
773 | if (!CondReg && !InvCondReg) | |||
774 | CondReg = promoteCondToReg(TestMBB, TestPos, TestLoc, Cond); | |||
775 | ||||
776 | if (CondReg) | |||
777 | return {CondReg, false}; | |||
778 | else | |||
779 | return {InvCondReg, true}; | |||
780 | } | |||
781 | ||||
782 | void X86FlagsCopyLoweringPass::insertTest(MachineBasicBlock &MBB, | |||
783 | MachineBasicBlock::iterator Pos, | |||
784 | const DebugLoc &Loc, unsigned Reg) { | |||
785 | auto TestI = | |||
786 | BuildMI(MBB, Pos, Loc, TII->get(X86::TEST8rr)).addReg(Reg).addReg(Reg); | |||
787 | (void)TestI; | |||
788 | LLVM_DEBUG(dbgs() << " test cond: "; TestI->dump())do { } while (false); | |||
789 | ++NumTestsInserted; | |||
790 | } | |||
791 | ||||
792 | void X86FlagsCopyLoweringPass::rewriteArithmetic( | |||
793 | MachineBasicBlock &TestMBB, MachineBasicBlock::iterator TestPos, | |||
794 | const DebugLoc &TestLoc, MachineInstr &MI, MachineOperand &FlagUse, | |||
795 | CondRegArray &CondRegs) { | |||
796 | // Arithmetic is either reading CF or OF. Figure out which condition we need | |||
797 | // to preserve in a register. | |||
798 | X86::CondCode Cond = X86::COND_INVALID; | |||
799 | ||||
800 | // The addend to use to reset CF or OF when added to the flag value. | |||
801 | int Addend = 0; | |||
802 | ||||
803 | switch (getMnemonicFromOpcode(MI.getOpcode())) { | |||
804 | case FlagArithMnemonic::ADC: | |||
805 | case FlagArithMnemonic::ADCX: | |||
806 | case FlagArithMnemonic::RCL: | |||
807 | case FlagArithMnemonic::RCR: | |||
808 | case FlagArithMnemonic::SBB: | |||
809 | case FlagArithMnemonic::SETB: | |||
810 | Cond = X86::COND_B; // CF == 1 | |||
811 | // Set up an addend that when one is added will need a carry due to not | |||
812 | // having a higher bit available. | |||
813 | Addend = 255; | |||
814 | break; | |||
815 | ||||
816 | case FlagArithMnemonic::ADOX: | |||
817 | Cond = X86::COND_O; // OF == 1 | |||
818 | // Set up an addend that when one is added will turn from positive to | |||
819 | // negative and thus overflow in the signed domain. | |||
820 | Addend = 127; | |||
821 | break; | |||
822 | } | |||
823 | ||||
824 | // Now get a register that contains the value of the flag input to the | |||
825 | // arithmetic. We require exactly this flag to simplify the arithmetic | |||
826 | // required to materialize it back into the flag. | |||
827 | unsigned &CondReg = CondRegs[Cond]; | |||
828 | if (!CondReg) | |||
829 | CondReg = promoteCondToReg(TestMBB, TestPos, TestLoc, Cond); | |||
830 | ||||
831 | MachineBasicBlock &MBB = *MI.getParent(); | |||
832 | ||||
833 | // Insert an instruction that will set the flag back to the desired value. | |||
834 | Register TmpReg = MRI->createVirtualRegister(PromoteRC); | |||
835 | auto AddI = | |||
836 | BuildMI(MBB, MI.getIterator(), MI.getDebugLoc(), TII->get(X86::ADD8ri)) | |||
837 | .addDef(TmpReg, RegState::Dead) | |||
838 | .addReg(CondReg) | |||
839 | .addImm(Addend); | |||
840 | (void)AddI; | |||
841 | LLVM_DEBUG(dbgs() << " add cond: "; AddI->dump())do { } while (false); | |||
842 | ++NumAddsInserted; | |||
843 | FlagUse.setIsKill(true); | |||
844 | } | |||
845 | ||||
846 | void X86FlagsCopyLoweringPass::rewriteCMov(MachineBasicBlock &TestMBB, | |||
847 | MachineBasicBlock::iterator TestPos, | |||
848 | const DebugLoc &TestLoc, | |||
849 | MachineInstr &CMovI, | |||
850 | MachineOperand &FlagUse, | |||
851 | CondRegArray &CondRegs) { | |||
852 | // First get the register containing this specific condition. | |||
853 | X86::CondCode Cond = X86::getCondFromCMov(CMovI); | |||
854 | unsigned CondReg; | |||
855 | bool Inverted; | |||
856 | std::tie(CondReg, Inverted) = | |||
857 | getCondOrInverseInReg(TestMBB, TestPos, TestLoc, Cond, CondRegs); | |||
858 | ||||
859 | MachineBasicBlock &MBB = *CMovI.getParent(); | |||
860 | ||||
861 | // Insert a direct test of the saved register. | |||
862 | insertTest(MBB, CMovI.getIterator(), CMovI.getDebugLoc(), CondReg); | |||
863 | ||||
864 | // Rewrite the CMov to use the !ZF flag from the test, and then kill its use | |||
865 | // of the flags afterward. | |||
866 | CMovI.getOperand(CMovI.getDesc().getNumOperands() - 1) | |||
867 | .setImm(Inverted ? X86::COND_E : X86::COND_NE); | |||
868 | FlagUse.setIsKill(true); | |||
869 | LLVM_DEBUG(dbgs() << " fixed cmov: "; CMovI.dump())do { } while (false); | |||
870 | } | |||
871 | ||||
872 | void X86FlagsCopyLoweringPass::rewriteFCMov(MachineBasicBlock &TestMBB, | |||
873 | MachineBasicBlock::iterator TestPos, | |||
874 | const DebugLoc &TestLoc, | |||
875 | MachineInstr &CMovI, | |||
876 | MachineOperand &FlagUse, | |||
877 | CondRegArray &CondRegs) { | |||
878 | // First get the register containing this specific condition. | |||
879 | X86::CondCode Cond = getCondFromFCMOV(CMovI.getOpcode()); | |||
880 | unsigned CondReg; | |||
881 | bool Inverted; | |||
882 | std::tie(CondReg, Inverted) = | |||
883 | getCondOrInverseInReg(TestMBB, TestPos, TestLoc, Cond, CondRegs); | |||
884 | ||||
885 | MachineBasicBlock &MBB = *CMovI.getParent(); | |||
886 | ||||
887 | // Insert a direct test of the saved register. | |||
888 | insertTest(MBB, CMovI.getIterator(), CMovI.getDebugLoc(), CondReg); | |||
889 | ||||
890 | auto getFCMOVOpcode = [](unsigned Opcode, bool Inverted) { | |||
891 | switch (Opcode) { | |||
892 | default: llvm_unreachable("Unexpected opcode!")__builtin_unreachable(); | |||
893 | case X86::CMOVBE_Fp32: case X86::CMOVNBE_Fp32: | |||
894 | case X86::CMOVB_Fp32: case X86::CMOVNB_Fp32: | |||
895 | case X86::CMOVE_Fp32: case X86::CMOVNE_Fp32: | |||
896 | case X86::CMOVP_Fp32: case X86::CMOVNP_Fp32: | |||
897 | return Inverted ? X86::CMOVE_Fp32 : X86::CMOVNE_Fp32; | |||
898 | case X86::CMOVBE_Fp64: case X86::CMOVNBE_Fp64: | |||
899 | case X86::CMOVB_Fp64: case X86::CMOVNB_Fp64: | |||
900 | case X86::CMOVE_Fp64: case X86::CMOVNE_Fp64: | |||
901 | case X86::CMOVP_Fp64: case X86::CMOVNP_Fp64: | |||
902 | return Inverted ? X86::CMOVE_Fp64 : X86::CMOVNE_Fp64; | |||
903 | case X86::CMOVBE_Fp80: case X86::CMOVNBE_Fp80: | |||
904 | case X86::CMOVB_Fp80: case X86::CMOVNB_Fp80: | |||
905 | case X86::CMOVE_Fp80: case X86::CMOVNE_Fp80: | |||
906 | case X86::CMOVP_Fp80: case X86::CMOVNP_Fp80: | |||
907 | return Inverted ? X86::CMOVE_Fp80 : X86::CMOVNE_Fp80; | |||
908 | } | |||
909 | }; | |||
910 | ||||
911 | // Rewrite the CMov to use the !ZF flag from the test. | |||
912 | CMovI.setDesc(TII->get(getFCMOVOpcode(CMovI.getOpcode(), Inverted))); | |||
913 | FlagUse.setIsKill(true); | |||
914 | LLVM_DEBUG(dbgs() << " fixed fcmov: "; CMovI.dump())do { } while (false); | |||
915 | } | |||
916 | ||||
917 | void X86FlagsCopyLoweringPass::rewriteCondJmp( | |||
918 | MachineBasicBlock &TestMBB, MachineBasicBlock::iterator TestPos, | |||
919 | const DebugLoc &TestLoc, MachineInstr &JmpI, CondRegArray &CondRegs) { | |||
920 | // First get the register containing this specific condition. | |||
921 | X86::CondCode Cond = X86::getCondFromBranch(JmpI); | |||
922 | unsigned CondReg; | |||
923 | bool Inverted; | |||
924 | std::tie(CondReg, Inverted) = | |||
925 | getCondOrInverseInReg(TestMBB, TestPos, TestLoc, Cond, CondRegs); | |||
926 | ||||
927 | MachineBasicBlock &JmpMBB = *JmpI.getParent(); | |||
928 | ||||
929 | // Insert a direct test of the saved register. | |||
930 | insertTest(JmpMBB, JmpI.getIterator(), JmpI.getDebugLoc(), CondReg); | |||
931 | ||||
932 | // Rewrite the jump to use the !ZF flag from the test, and kill its use of | |||
933 | // flags afterward. | |||
934 | JmpI.getOperand(1).setImm(Inverted ? X86::COND_E : X86::COND_NE); | |||
935 | JmpI.findRegisterUseOperand(X86::EFLAGS)->setIsKill(true); | |||
936 | LLVM_DEBUG(dbgs() << " fixed jCC: "; JmpI.dump())do { } while (false); | |||
937 | } | |||
938 | ||||
939 | void X86FlagsCopyLoweringPass::rewriteCopy(MachineInstr &MI, | |||
940 | MachineOperand &FlagUse, | |||
941 | MachineInstr &CopyDefI) { | |||
942 | // Just replace this copy with the original copy def. | |||
943 | MRI->replaceRegWith(MI.getOperand(0).getReg(), | |||
944 | CopyDefI.getOperand(0).getReg()); | |||
945 | MI.eraseFromParent(); | |||
946 | } | |||
947 | ||||
948 | void X86FlagsCopyLoweringPass::rewriteSetCC(MachineBasicBlock &TestMBB, | |||
949 | MachineBasicBlock::iterator TestPos, | |||
950 | const DebugLoc &TestLoc, | |||
951 | MachineInstr &SetCCI, | |||
952 | MachineOperand &FlagUse, | |||
953 | CondRegArray &CondRegs) { | |||
954 | X86::CondCode Cond = X86::getCondFromSETCC(SetCCI); | |||
955 | // Note that we can't usefully rewrite this to the inverse without complex | |||
956 | // analysis of the users of the setCC. Largely we rely on duplicates which | |||
957 | // could have been avoided already being avoided here. | |||
958 | unsigned &CondReg = CondRegs[Cond]; | |||
959 | if (!CondReg) | |||
960 | CondReg = promoteCondToReg(TestMBB, TestPos, TestLoc, Cond); | |||
961 | ||||
962 | // Rewriting a register def is trivial: we just replace the register and | |||
963 | // remove the setcc. | |||
964 | if (!SetCCI.mayStore()) { | |||
965 | assert(SetCCI.getOperand(0).isReg() &&((void)0) | |||
966 | "Cannot have a non-register defined operand to SETcc!")((void)0); | |||
967 | MRI->replaceRegWith(SetCCI.getOperand(0).getReg(), CondReg); | |||
968 | SetCCI.eraseFromParent(); | |||
969 | return; | |||
970 | } | |||
971 | ||||
972 | // Otherwise, we need to emit a store. | |||
973 | auto MIB = BuildMI(*SetCCI.getParent(), SetCCI.getIterator(), | |||
974 | SetCCI.getDebugLoc(), TII->get(X86::MOV8mr)); | |||
975 | // Copy the address operands. | |||
976 | for (int i = 0; i < X86::AddrNumOperands; ++i) | |||
977 | MIB.add(SetCCI.getOperand(i)); | |||
978 | ||||
979 | MIB.addReg(CondReg); | |||
980 | ||||
981 | MIB.setMemRefs(SetCCI.memoperands()); | |||
982 | ||||
983 | SetCCI.eraseFromParent(); | |||
984 | } |
1 | // -*- C++ -*- |
2 | //===---------------------------- numeric ---------------------------------===// |
3 | // |
4 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
5 | // See https://llvm.org/LICENSE.txt for license information. |
6 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
7 | // |
8 | //===----------------------------------------------------------------------===// |
9 | |
10 | #ifndef _LIBCPP_NUMERIC |
11 | #define _LIBCPP_NUMERIC |
12 | |
13 | /* |
14 | numeric synopsis |
15 | |
16 | namespace std |
17 | { |
18 | |
19 | template <class InputIterator, class T> |
20 | constexpr T // constexpr since C++20 |
21 | accumulate(InputIterator first, InputIterator last, T init); |
22 | |
23 | template <class InputIterator, class T, class BinaryOperation> |
24 | constexpr T // constexpr since C++20 |
25 | accumulate(InputIterator first, InputIterator last, T init, BinaryOperation binary_op); |
26 | |
27 | template<class InputIterator> |
28 | constexpr typename iterator_traits<InputIterator>::value_type // constexpr since C++20 |
29 | reduce(InputIterator first, InputIterator last); // C++17 |
30 | |
31 | template<class InputIterator, class T> |
32 | constexpr T // constexpr since C++20 |
33 | reduce(InputIterator first, InputIterator last, T init); // C++17 |
34 | |
35 | template<class InputIterator, class T, class BinaryOperation> |
36 | constexpr T // constexpr since C++20 |
37 | reduce(InputIterator first, InputIterator last, T init, BinaryOperation binary_op); // C++17 |
38 | |
39 | template <class InputIterator1, class InputIterator2, class T> |
40 | constexpr T // constexpr since C++20 |
41 | inner_product(InputIterator1 first1, InputIterator1 last1, InputIterator2 first2, T init); |
42 | |
43 | template <class InputIterator1, class InputIterator2, class T, class BinaryOperation1, class BinaryOperation2> |
44 | constexpr T // constexpr since C++20 |
45 | inner_product(InputIterator1 first1, InputIterator1 last1, InputIterator2 first2, |
46 | T init, BinaryOperation1 binary_op1, BinaryOperation2 binary_op2); |
47 | |
48 | |
49 | template<class InputIterator1, class InputIterator2, class T> |
50 | constexpr T // constexpr since C++20 |
51 | transform_reduce(InputIterator1 first1, InputIterator1 last1, |
52 | InputIterator2 first2, T init); // C++17 |
53 | |
54 | template<class InputIterator1, class InputIterator2, class T, class BinaryOperation1, class BinaryOperation2> |
55 | constexpr T // constexpr since C++20 |
56 | transform_reduce(InputIterator1 first1, InputIterator1 last1, |
57 | InputIterator2 first2, T init, |
58 | BinaryOperation1 binary_op1, BinaryOperation2 binary_op2); // C++17 |
59 | |
60 | template<class InputIterator, class T, class BinaryOperation, class UnaryOperation> |
61 | constexpr T // constexpr since C++20 |
62 | transform_reduce(InputIterator first, InputIterator last, T init, |
63 | BinaryOperation binary_op, UnaryOperation unary_op); // C++17 |
64 | |
65 | template <class InputIterator, class OutputIterator> |
66 | constexpr OutputIterator // constexpr since C++20 |
67 | partial_sum(InputIterator first, InputIterator last, OutputIterator result); |
68 | |
69 | template <class InputIterator, class OutputIterator, class BinaryOperation> |
70 | constexpr OutputIterator // constexpr since C++20 |
71 | partial_sum(InputIterator first, InputIterator last, OutputIterator result, BinaryOperation binary_op); |
72 | |
73 | template<class InputIterator, class OutputIterator, class T> |
74 | constexpr OutputIterator // constexpr since C++20 |
75 | exclusive_scan(InputIterator first, InputIterator last, |
76 | OutputIterator result, T init); // C++17 |
77 | |
78 | template<class InputIterator, class OutputIterator, class T, class BinaryOperation> |
79 | constexpr OutputIterator // constexpr since C++20 |
80 | exclusive_scan(InputIterator first, InputIterator last, |
81 | OutputIterator result, T init, BinaryOperation binary_op); // C++17 |
82 | |
83 | template<class InputIterator, class OutputIterator> |
84 | constexpr OutputIterator // constexpr since C++20 |
85 | inclusive_scan(InputIterator first, InputIterator last, OutputIterator result); // C++17 |
86 | |
87 | template<class InputIterator, class OutputIterator, class BinaryOperation> |
88 | constexpr OutputIterator // constexpr since C++20 |
89 | inclusive_scan(InputIterator first, InputIterator last, |
90 | OutputIterator result, BinaryOperation binary_op); // C++17 |
91 | |
92 | template<class InputIterator, class OutputIterator, class BinaryOperation, class T> |
93 | constexpr OutputIterator // constexpr since C++20 |
94 | inclusive_scan(InputIterator first, InputIterator last, |
95 | OutputIterator result, BinaryOperation binary_op, T init); // C++17 |
96 | |
97 | template<class InputIterator, class OutputIterator, class T, |
98 | class BinaryOperation, class UnaryOperation> |
99 | constexpr OutputIterator // constexpr since C++20 |
100 | transform_exclusive_scan(InputIterator first, InputIterator last, |
101 | OutputIterator result, T init, |
102 | BinaryOperation binary_op, UnaryOperation unary_op); // C++17 |
103 | |
104 | template<class InputIterator, class OutputIterator, |
105 | class BinaryOperation, class UnaryOperation> |
106 | constexpr OutputIterator // constexpr since C++20 |
107 | transform_inclusive_scan(InputIterator first, InputIterator last, |
108 | OutputIterator result, |
109 | BinaryOperation binary_op, UnaryOperation unary_op); // C++17 |
110 | |
111 | template<class InputIterator, class OutputIterator, |
112 | class BinaryOperation, class UnaryOperation, class T> |
113 | constexpr OutputIterator // constexpr since C++20 |
114 | transform_inclusive_scan(InputIterator first, InputIterator last, |
115 | OutputIterator result, |
116 | BinaryOperation binary_op, UnaryOperation unary_op, |
117 | T init); // C++17 |
118 | |
119 | template <class InputIterator, class OutputIterator> |
120 | constexpr OutputIterator // constexpr since C++20 |
121 | adjacent_difference(InputIterator first, InputIterator last, OutputIterator result); |
122 | |
123 | template <class InputIterator, class OutputIterator, class BinaryOperation> |
124 | constexpr OutputIterator // constexpr since C++20 |
125 | adjacent_difference(InputIterator first, InputIterator last, OutputIterator result, BinaryOperation binary_op); |
126 | |
127 | template <class ForwardIterator, class T> |
128 | constexpr void // constexpr since C++20 |
129 | iota(ForwardIterator first, ForwardIterator last, T value); |
130 | |
131 | template <class M, class N> |
132 | constexpr common_type_t<M,N> gcd(M m, N n); // C++17 |
133 | |
134 | template <class M, class N> |
135 | constexpr common_type_t<M,N> lcm(M m, N n); // C++17 |
136 | |
137 | template<class T> |
138 | constexpr T midpoint(T a, T b) noexcept; // C++20 |
139 | |
140 | template<class T> |
141 | constexpr T* midpoint(T* a, T* b); // C++20 |
142 | |
143 | } // std |
144 | |
145 | */ |
146 | |
147 | #include <__config> |
148 | #include <__debug> |
149 | #include <cmath> // for isnormal |
150 | #include <functional> |
151 | #include <iterator> |
152 | #include <limits> // for numeric_limits |
153 | #include <version> |
154 | |
155 | #if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER) |
156 | #pragma GCC system_header |
157 | #endif |
158 | |
159 | _LIBCPP_PUSH_MACROSpush_macro("min") push_macro("max") |
160 | #include <__undef_macros> |
161 | |
162 | _LIBCPP_BEGIN_NAMESPACE_STDnamespace std { inline namespace __1 { |
163 | |
164 | template <class _InputIterator, class _Tp> |
165 | _LIBCPP_INLINE_VISIBILITY__attribute__ ((__visibility__("hidden"))) __attribute__ ((__exclude_from_explicit_instantiation__ )) _LIBCPP_CONSTEXPR_AFTER_CXX17 |
166 | _Tp |
167 | accumulate(_InputIterator __first, _InputIterator __last, _Tp __init) |
168 | { |
169 | for (; __first != __last; ++__first) |
170 | #if _LIBCPP_STD_VER14 > 17 |
171 | __init = _VSTDstd::__1::move(__init) + *__first; |
172 | #else |
173 | __init = __init + *__first; |
174 | #endif |
175 | return __init; |
176 | } |
177 | |
178 | template <class _InputIterator, class _Tp, class _BinaryOperation> |
179 | _LIBCPP_INLINE_VISIBILITY__attribute__ ((__visibility__("hidden"))) __attribute__ ((__exclude_from_explicit_instantiation__ )) _LIBCPP_CONSTEXPR_AFTER_CXX17 |
180 | _Tp |
181 | accumulate(_InputIterator __first, _InputIterator __last, _Tp __init, _BinaryOperation __binary_op) |
182 | { |
183 | for (; __first != __last; ++__first) |
184 | #if _LIBCPP_STD_VER14 > 17 |
185 | __init = __binary_op(_VSTDstd::__1::move(__init), *__first); |
186 | #else |
187 | __init = __binary_op(__init, *__first); |
188 | #endif |
189 | return __init; |
190 | } |
191 | |
192 | #if _LIBCPP_STD_VER14 > 14 |
193 | template <class _InputIterator, class _Tp, class _BinaryOp> |
194 | _LIBCPP_INLINE_VISIBILITY__attribute__ ((__visibility__("hidden"))) __attribute__ ((__exclude_from_explicit_instantiation__ )) _LIBCPP_CONSTEXPR_AFTER_CXX17 |
195 | _Tp |
196 | reduce(_InputIterator __first, _InputIterator __last, _Tp __init, _BinaryOp __b) |
197 | { |
198 | for (; __first != __last; ++__first) |
199 | __init = __b(__init, *__first); |
200 | return __init; |
201 | } |
202 | |
203 | template <class _InputIterator, class _Tp> |
204 | _LIBCPP_INLINE_VISIBILITY__attribute__ ((__visibility__("hidden"))) __attribute__ ((__exclude_from_explicit_instantiation__ )) _LIBCPP_CONSTEXPR_AFTER_CXX17 |
205 | _Tp |
206 | reduce(_InputIterator __first, _InputIterator __last, _Tp __init) |
207 | { |
208 | return _VSTDstd::__1::reduce(__first, __last, __init, _VSTDstd::__1::plus<>()); |
209 | } |
210 | |
211 | template <class _InputIterator> |
212 | _LIBCPP_INLINE_VISIBILITY__attribute__ ((__visibility__("hidden"))) __attribute__ ((__exclude_from_explicit_instantiation__ )) _LIBCPP_CONSTEXPR_AFTER_CXX17 |
213 | typename iterator_traits<_InputIterator>::value_type |
214 | reduce(_InputIterator __first, _InputIterator __last) |
215 | { |
216 | return _VSTDstd::__1::reduce(__first, __last, |
217 | typename iterator_traits<_InputIterator>::value_type{}); |
218 | } |
219 | #endif |
220 | |
221 | template <class _InputIterator1, class _InputIterator2, class _Tp> |
222 | _LIBCPP_INLINE_VISIBILITY__attribute__ ((__visibility__("hidden"))) __attribute__ ((__exclude_from_explicit_instantiation__ )) _LIBCPP_CONSTEXPR_AFTER_CXX17 |
223 | _Tp |
224 | inner_product(_InputIterator1 __first1, _InputIterator1 __last1, _InputIterator2 __first2, _Tp __init) |
225 | { |
226 | for (; __first1 != __last1; ++__first1, (void) ++__first2) |
227 | #if _LIBCPP_STD_VER14 > 17 |
228 | __init = _VSTDstd::__1::move(__init) + *__first1 * *__first2; |
229 | #else |
230 | __init = __init + *__first1 * *__first2; |
231 | #endif |
232 | return __init; |
233 | } |
234 | |
235 | template <class _InputIterator1, class _InputIterator2, class _Tp, class _BinaryOperation1, class _BinaryOperation2> |
236 | _LIBCPP_INLINE_VISIBILITY__attribute__ ((__visibility__("hidden"))) __attribute__ ((__exclude_from_explicit_instantiation__ )) _LIBCPP_CONSTEXPR_AFTER_CXX17 |
237 | _Tp |
238 | inner_product(_InputIterator1 __first1, _InputIterator1 __last1, _InputIterator2 __first2, |
239 | _Tp __init, _BinaryOperation1 __binary_op1, _BinaryOperation2 __binary_op2) |
240 | { |
241 | for (; __first1 != __last1; ++__first1, (void) ++__first2) |
242 | #if _LIBCPP_STD_VER14 > 17 |
243 | __init = __binary_op1(_VSTDstd::__1::move(__init), __binary_op2(*__first1, *__first2)); |
244 | #else |
245 | __init = __binary_op1(__init, __binary_op2(*__first1, *__first2)); |
246 | #endif |
247 | return __init; |
248 | } |
249 | |
250 | #if _LIBCPP_STD_VER14 > 14 |
251 | template <class _InputIterator, class _Tp, class _BinaryOp, class _UnaryOp> |
252 | _LIBCPP_INLINE_VISIBILITY__attribute__ ((__visibility__("hidden"))) __attribute__ ((__exclude_from_explicit_instantiation__ )) _LIBCPP_CONSTEXPR_AFTER_CXX17 |
253 | _Tp |
254 | transform_reduce(_InputIterator __first, _InputIterator __last, |
255 | _Tp __init, _BinaryOp __b, _UnaryOp __u) |
256 | { |
257 | for (; __first != __last; ++__first) |
258 | __init = __b(__init, __u(*__first)); |
259 | return __init; |
260 | } |
261 | |
262 | template <class _InputIterator1, class _InputIterator2, |
263 | class _Tp, class _BinaryOp1, class _BinaryOp2> |
264 | _LIBCPP_INLINE_VISIBILITY__attribute__ ((__visibility__("hidden"))) __attribute__ ((__exclude_from_explicit_instantiation__ )) _LIBCPP_CONSTEXPR_AFTER_CXX17 |
265 | _Tp |
266 | transform_reduce(_InputIterator1 __first1, _InputIterator1 __last1, |
267 | _InputIterator2 __first2, _Tp __init, _BinaryOp1 __b1, _BinaryOp2 __b2) |
268 | { |
269 | for (; __first1 != __last1; ++__first1, (void) ++__first2) |
270 | __init = __b1(__init, __b2(*__first1, *__first2)); |
271 | return __init; |
272 | } |
273 | |
274 | template <class _InputIterator1, class _InputIterator2, class _Tp> |
275 | _LIBCPP_INLINE_VISIBILITY__attribute__ ((__visibility__("hidden"))) __attribute__ ((__exclude_from_explicit_instantiation__ )) _LIBCPP_CONSTEXPR_AFTER_CXX17 |
276 | _Tp |
277 | transform_reduce(_InputIterator1 __first1, _InputIterator1 __last1, |
278 | _InputIterator2 __first2, _Tp __init) |
279 | { |
280 | return _VSTDstd::__1::transform_reduce(__first1, __last1, __first2, _VSTDstd::__1::move(__init), |
281 | _VSTDstd::__1::plus<>(), _VSTDstd::__1::multiplies<>()); |
282 | } |
283 | #endif |
284 | |
285 | template <class _InputIterator, class _OutputIterator> |
286 | _LIBCPP_INLINE_VISIBILITY__attribute__ ((__visibility__("hidden"))) __attribute__ ((__exclude_from_explicit_instantiation__ )) _LIBCPP_CONSTEXPR_AFTER_CXX17 |
287 | _OutputIterator |
288 | partial_sum(_InputIterator __first, _InputIterator __last, _OutputIterator __result) |
289 | { |
290 | if (__first != __last) |
291 | { |
292 | typename iterator_traits<_InputIterator>::value_type __t(*__first); |
293 | *__result = __t; |
294 | for (++__first, (void) ++__result; __first != __last; ++__first, (void) ++__result) |
295 | { |
296 | #if _LIBCPP_STD_VER14 > 17 |
297 | __t = _VSTDstd::__1::move(__t) + *__first; |
298 | #else |
299 | __t = __t + *__first; |
300 | #endif |
301 | *__result = __t; |
302 | } |
303 | } |
304 | return __result; |
305 | } |
306 | |
307 | template <class _InputIterator, class _OutputIterator, class _BinaryOperation> |
308 | _LIBCPP_INLINE_VISIBILITY__attribute__ ((__visibility__("hidden"))) __attribute__ ((__exclude_from_explicit_instantiation__ )) _LIBCPP_CONSTEXPR_AFTER_CXX17 |
309 | _OutputIterator |
310 | partial_sum(_InputIterator __first, _InputIterator __last, _OutputIterator __result, |
311 | _BinaryOperation __binary_op) |
312 | { |
313 | if (__first != __last) |
314 | { |
315 | typename iterator_traits<_InputIterator>::value_type __t(*__first); |
316 | *__result = __t; |
317 | for (++__first, (void) ++__result; __first != __last; ++__first, (void) ++__result) |
318 | { |
319 | #if _LIBCPP_STD_VER14 > 17 |
320 | __t = __binary_op(_VSTDstd::__1::move(__t), *__first); |
321 | #else |
322 | __t = __binary_op(__t, *__first); |
323 | #endif |
324 | *__result = __t; |
325 | } |
326 | } |
327 | return __result; |
328 | } |
329 | |
330 | #if _LIBCPP_STD_VER14 > 14 |
331 | template <class _InputIterator, class _OutputIterator, class _Tp, class _BinaryOp> |
332 | _LIBCPP_INLINE_VISIBILITY__attribute__ ((__visibility__("hidden"))) __attribute__ ((__exclude_from_explicit_instantiation__ )) _LIBCPP_CONSTEXPR_AFTER_CXX17 |
333 | _OutputIterator |
334 | exclusive_scan(_InputIterator __first, _InputIterator __last, |
335 | _OutputIterator __result, _Tp __init, _BinaryOp __b) |
336 | { |
337 | if (__first != __last) |
338 | { |
339 | _Tp __tmp(__b(__init, *__first)); |
340 | while (true) |
341 | { |
342 | *__result = _VSTDstd::__1::move(__init); |
343 | ++__result; |
344 | ++__first; |
345 | if (__first == __last) |
346 | break; |
347 | __init = _VSTDstd::__1::move(__tmp); |
348 | __tmp = __b(__init, *__first); |
349 | } |
350 | } |
351 | return __result; |
352 | } |
353 | |
354 | template <class _InputIterator, class _OutputIterator, class _Tp> |
355 | _LIBCPP_INLINE_VISIBILITY__attribute__ ((__visibility__("hidden"))) __attribute__ ((__exclude_from_explicit_instantiation__ )) _LIBCPP_CONSTEXPR_AFTER_CXX17 |
356 | _OutputIterator |
357 | exclusive_scan(_InputIterator __first, _InputIterator __last, |
358 | _OutputIterator __result, _Tp __init) |
359 | { |
360 | return _VSTDstd::__1::exclusive_scan(__first, __last, __result, __init, _VSTDstd::__1::plus<>()); |
361 | } |
362 | |
363 | template <class _InputIterator, class _OutputIterator, class _Tp, class _BinaryOp> |
364 | _LIBCPP_INLINE_VISIBILITY__attribute__ ((__visibility__("hidden"))) __attribute__ ((__exclude_from_explicit_instantiation__ )) _LIBCPP_CONSTEXPR_AFTER_CXX17 |
365 | _OutputIterator inclusive_scan(_InputIterator __first, _InputIterator __last, |
366 | _OutputIterator __result, _BinaryOp __b, _Tp __init) |
367 | { |
368 | for (; __first != __last; ++__first, (void) ++__result) { |
369 | __init = __b(__init, *__first); |
370 | *__result = __init; |
371 | } |
372 | return __result; |
373 | } |
374 | |
375 | template <class _InputIterator, class _OutputIterator, class _BinaryOp> |
376 | _LIBCPP_INLINE_VISIBILITY__attribute__ ((__visibility__("hidden"))) __attribute__ ((__exclude_from_explicit_instantiation__ )) _LIBCPP_CONSTEXPR_AFTER_CXX17 |
377 | _OutputIterator inclusive_scan(_InputIterator __first, _InputIterator __last, |
378 | _OutputIterator __result, _BinaryOp __b) |
379 | { |
380 | if (__first != __last) { |
381 | typename iterator_traits<_InputIterator>::value_type __init = *__first; |
382 | *__result++ = __init; |
383 | if (++__first != __last) |
384 | return _VSTDstd::__1::inclusive_scan(__first, __last, __result, __b, __init); |
385 | } |
386 | |
387 | return __result; |
388 | } |
389 | |
390 | template <class _InputIterator, class _OutputIterator> |
391 | _LIBCPP_INLINE_VISIBILITY__attribute__ ((__visibility__("hidden"))) __attribute__ ((__exclude_from_explicit_instantiation__ )) _LIBCPP_CONSTEXPR_AFTER_CXX17 |
392 | _OutputIterator inclusive_scan(_InputIterator __first, _InputIterator __last, |
393 | _OutputIterator __result) |
394 | { |
395 | return _VSTDstd::__1::inclusive_scan(__first, __last, __result, _VSTDstd::__1::plus<>()); |
396 | } |
397 | |
398 | template <class _InputIterator, class _OutputIterator, class _Tp, |
399 | class _BinaryOp, class _UnaryOp> |
400 | _LIBCPP_INLINE_VISIBILITY__attribute__ ((__visibility__("hidden"))) __attribute__ ((__exclude_from_explicit_instantiation__ )) _LIBCPP_CONSTEXPR_AFTER_CXX17 |
401 | _OutputIterator |
402 | transform_exclusive_scan(_InputIterator __first, _InputIterator __last, |
403 | _OutputIterator __result, _Tp __init, |
404 | _BinaryOp __b, _UnaryOp __u) |
405 | { |
406 | if (__first != __last) |
407 | { |
408 | _Tp __saved = __init; |
409 | do |
410 | { |
411 | __init = __b(__init, __u(*__first)); |
412 | *__result = __saved; |
413 | __saved = __init; |
414 | ++__result; |
415 | } while (++__first != __last); |
416 | } |
417 | return __result; |
418 | } |
419 | |
420 | template <class _InputIterator, class _OutputIterator, class _Tp, class _BinaryOp, class _UnaryOp> |
421 | _LIBCPP_INLINE_VISIBILITY__attribute__ ((__visibility__("hidden"))) __attribute__ ((__exclude_from_explicit_instantiation__ )) _LIBCPP_CONSTEXPR_AFTER_CXX17 |
422 | _OutputIterator |
423 | transform_inclusive_scan(_InputIterator __first, _InputIterator __last, |
424 | _OutputIterator __result, _BinaryOp __b, _UnaryOp __u, _Tp __init) |
425 | { |
426 | for (; __first != __last; ++__first, (void) ++__result) { |
427 | __init = __b(__init, __u(*__first)); |
428 | *__result = __init; |
429 | } |
430 | |
431 | return __result; |
432 | } |
433 | |
434 | template <class _InputIterator, class _OutputIterator, class _BinaryOp, class _UnaryOp> |
435 | _LIBCPP_INLINE_VISIBILITY__attribute__ ((__visibility__("hidden"))) __attribute__ ((__exclude_from_explicit_instantiation__ )) _LIBCPP_CONSTEXPR_AFTER_CXX17 |
436 | _OutputIterator |
437 | transform_inclusive_scan(_InputIterator __first, _InputIterator __last, |
438 | _OutputIterator __result, _BinaryOp __b, _UnaryOp __u) |
439 | { |
440 | if (__first != __last) { |
441 | typename iterator_traits<_InputIterator>::value_type __init = __u(*__first); |
442 | *__result++ = __init; |
443 | if (++__first != __last) |
444 | return _VSTDstd::__1::transform_inclusive_scan(__first, __last, __result, __b, __u, __init); |
445 | } |
446 | |
447 | return __result; |
448 | } |
449 | #endif |
450 | |
451 | template <class _InputIterator, class _OutputIterator> |
452 | _LIBCPP_INLINE_VISIBILITY__attribute__ ((__visibility__("hidden"))) __attribute__ ((__exclude_from_explicit_instantiation__ )) _LIBCPP_CONSTEXPR_AFTER_CXX17 |
453 | _OutputIterator |
454 | adjacent_difference(_InputIterator __first, _InputIterator __last, _OutputIterator __result) |
455 | { |
456 | if (__first != __last) |
457 | { |
458 | typename iterator_traits<_InputIterator>::value_type __acc(*__first); |
459 | *__result = __acc; |
460 | for (++__first, (void) ++__result; __first != __last; ++__first, (void) ++__result) |
461 | { |
462 | typename iterator_traits<_InputIterator>::value_type __val(*__first); |
463 | #if _LIBCPP_STD_VER14 > 17 |
464 | *__result = __val - _VSTDstd::__1::move(__acc); |
465 | #else |
466 | *__result = __val - __acc; |
467 | #endif |
468 | __acc = _VSTDstd::__1::move(__val); |
469 | } |
470 | } |
471 | return __result; |
472 | } |
473 | |
474 | template <class _InputIterator, class _OutputIterator, class _BinaryOperation> |
475 | _LIBCPP_INLINE_VISIBILITY__attribute__ ((__visibility__("hidden"))) __attribute__ ((__exclude_from_explicit_instantiation__ )) _LIBCPP_CONSTEXPR_AFTER_CXX17 |
476 | _OutputIterator |
477 | adjacent_difference(_InputIterator __first, _InputIterator __last, _OutputIterator __result, |
478 | _BinaryOperation __binary_op) |
479 | { |
480 | if (__first != __last) |
481 | { |
482 | typename iterator_traits<_InputIterator>::value_type __acc(*__first); |
483 | *__result = __acc; |
484 | for (++__first, (void) ++__result; __first != __last; ++__first, (void) ++__result) |
485 | { |
486 | typename iterator_traits<_InputIterator>::value_type __val(*__first); |
487 | #if _LIBCPP_STD_VER14 > 17 |
488 | *__result = __binary_op(__val, _VSTDstd::__1::move(__acc)); |
489 | #else |
490 | *__result = __binary_op(__val, __acc); |
491 | #endif |
492 | __acc = _VSTDstd::__1::move(__val); |
493 | } |
494 | } |
495 | return __result; |
496 | } |
497 | |
498 | template <class _ForwardIterator, class _Tp> |
499 | _LIBCPP_INLINE_VISIBILITY__attribute__ ((__visibility__("hidden"))) __attribute__ ((__exclude_from_explicit_instantiation__ )) _LIBCPP_CONSTEXPR_AFTER_CXX17 |
500 | void |
501 | iota(_ForwardIterator __first, _ForwardIterator __last, _Tp __value_) |
502 | { |
503 | for (; __first != __last; ++__first, (void) ++__value_) |
504 | *__first = __value_; |
505 | } |
506 | |
507 | |
508 | #if _LIBCPP_STD_VER14 > 14 |
509 | template <typename _Result, typename _Source, bool _IsSigned = is_signed<_Source>::value> struct __ct_abs; |
510 | |
511 | template <typename _Result, typename _Source> |
512 | struct __ct_abs<_Result, _Source, true> { |
513 | _LIBCPP_CONSTEXPRconstexpr _LIBCPP_INLINE_VISIBILITY__attribute__ ((__visibility__("hidden"))) __attribute__ ((__exclude_from_explicit_instantiation__ )) |
514 | _Result operator()(_Source __t) const noexcept |
515 | { |
516 | if (__t >= 0) return __t; |
517 | if (__t == numeric_limits<_Source>::min()) return -static_cast<_Result>(__t); |
518 | return -__t; |
519 | } |
520 | }; |
521 | |
522 | template <typename _Result, typename _Source> |
523 | struct __ct_abs<_Result, _Source, false> { |
524 | _LIBCPP_CONSTEXPRconstexpr _LIBCPP_INLINE_VISIBILITY__attribute__ ((__visibility__("hidden"))) __attribute__ ((__exclude_from_explicit_instantiation__ )) |
525 | _Result operator()(_Source __t) const noexcept { return __t; } |
526 | }; |
527 | |
528 | |
529 | template<class _Tp> |
530 | _LIBCPP_CONSTEXPRconstexpr _LIBCPP_HIDDEN__attribute__ ((__visibility__("hidden"))) |
531 | _Tp __gcd(_Tp __m, _Tp __n) |
532 | { |
533 | static_assert((!is_signed<_Tp>::value), ""); |
534 | return __n == 0 ? __m : _VSTDstd::__1::__gcd<_Tp>(__n, __m % __n); |
535 | } |
536 | |
537 | |
538 | template<class _Tp, class _Up> |
539 | _LIBCPP_CONSTEXPRconstexpr _LIBCPP_INLINE_VISIBILITY__attribute__ ((__visibility__("hidden"))) __attribute__ ((__exclude_from_explicit_instantiation__ )) |
540 | common_type_t<_Tp,_Up> |
541 | gcd(_Tp __m, _Up __n) |
542 | { |
543 | static_assert((is_integral<_Tp>::value && is_integral<_Up>::value), "Arguments to gcd must be integer types"); |
544 | static_assert((!is_same<typename remove_cv<_Tp>::type, bool>::value), "First argument to gcd cannot be bool" ); |
545 | static_assert((!is_same<typename remove_cv<_Up>::type, bool>::value), "Second argument to gcd cannot be bool" ); |
546 | using _Rp = common_type_t<_Tp,_Up>; |
547 | using _Wp = make_unsigned_t<_Rp>; |
548 | return static_cast<_Rp>(_VSTDstd::__1::__gcd( |
549 | static_cast<_Wp>(__ct_abs<_Rp, _Tp>()(__m)), |
550 | static_cast<_Wp>(__ct_abs<_Rp, _Up>()(__n)))); |
551 | } |
552 | |
553 | template<class _Tp, class _Up> |
554 | _LIBCPP_CONSTEXPRconstexpr _LIBCPP_INLINE_VISIBILITY__attribute__ ((__visibility__("hidden"))) __attribute__ ((__exclude_from_explicit_instantiation__ )) |
555 | common_type_t<_Tp,_Up> |
556 | lcm(_Tp __m, _Up __n) |
557 | { |
558 | static_assert((is_integral<_Tp>::value && is_integral<_Up>::value), "Arguments to lcm must be integer types"); |
559 | static_assert((!is_same<typename remove_cv<_Tp>::type, bool>::value), "First argument to lcm cannot be bool" ); |
560 | static_assert((!is_same<typename remove_cv<_Up>::type, bool>::value), "Second argument to lcm cannot be bool" ); |
561 | if (__m == 0 || __n == 0) |
562 | return 0; |
563 | |
564 | using _Rp = common_type_t<_Tp,_Up>; |
565 | _Rp __val1 = __ct_abs<_Rp, _Tp>()(__m) / _VSTDstd::__1::gcd(__m, __n); |
566 | _Rp __val2 = __ct_abs<_Rp, _Up>()(__n); |
567 | _LIBCPP_ASSERT((numeric_limits<_Rp>::max() / __val1 > __val2), "Overflow in lcm")((void)0); |
568 | return __val1 * __val2; |
569 | } |
570 | |
571 | #endif /* _LIBCPP_STD_VER > 14 */ |
572 | |
573 | #if _LIBCPP_STD_VER14 > 17 |
574 | template <class _Tp> |
575 | _LIBCPP_INLINE_VISIBILITY__attribute__ ((__visibility__("hidden"))) __attribute__ ((__exclude_from_explicit_instantiation__ )) constexpr |
576 | enable_if_t<is_integral_v<_Tp> && !is_same_v<bool, _Tp> && !is_null_pointer_v<_Tp>, _Tp> |
577 | midpoint(_Tp __a, _Tp __b) noexcept |
578 | _LIBCPP_DISABLE_UBSAN_UNSIGNED_INTEGER_CHECK__attribute__((__no_sanitize__("unsigned-integer-overflow"))) |
579 | { |
580 | using _Up = make_unsigned_t<_Tp>; |
581 | constexpr _Up __bitshift = numeric_limits<_Up>::digits - 1; |
582 | |
583 | _Up __diff = _Up(__b) - _Up(__a); |
584 | _Up __sign_bit = __b < __a; |
585 | |
586 | _Up __half_diff = (__diff / 2) + (__sign_bit << __bitshift) + (__sign_bit & __diff); |
587 | |
588 | return __a + __half_diff; |
589 | } |
590 | |
591 | |
592 | template <class _TPtr> |
593 | _LIBCPP_INLINE_VISIBILITY__attribute__ ((__visibility__("hidden"))) __attribute__ ((__exclude_from_explicit_instantiation__ )) constexpr |
594 | enable_if_t<is_pointer_v<_TPtr> |
595 | && is_object_v<remove_pointer_t<_TPtr>> |
596 | && ! is_void_v<remove_pointer_t<_TPtr>> |
597 | && (sizeof(remove_pointer_t<_TPtr>) > 0), _TPtr> |
598 | midpoint(_TPtr __a, _TPtr __b) noexcept |
599 | { |
600 | return __a + _VSTDstd::__1::midpoint(ptrdiff_t(0), __b - __a); |
601 | } |
602 | |
603 | |
604 | template <typename _Tp> |
605 | constexpr int __sign(_Tp __val) { |
606 | return (_Tp(0) < __val) - (__val < _Tp(0)); |
607 | } |
608 | |
609 | template <typename _Fp> |
610 | constexpr _Fp __fp_abs(_Fp __f) { return __f >= 0 ? __f : -__f; } |
611 | |
612 | template <class _Fp> |
613 | _LIBCPP_INLINE_VISIBILITY__attribute__ ((__visibility__("hidden"))) __attribute__ ((__exclude_from_explicit_instantiation__ )) constexpr |
614 | enable_if_t<is_floating_point_v<_Fp>, _Fp> |
615 | midpoint(_Fp __a, _Fp __b) noexcept |
616 | { |
617 | constexpr _Fp __lo = numeric_limits<_Fp>::min()*2; |
618 | constexpr _Fp __hi = numeric_limits<_Fp>::max()/2; |
619 | return __fp_abs(__a) <= __hi && __fp_abs(__b) <= __hi ? // typical case: overflow is impossible |
620 | (__a + __b)/2 : // always correctly rounded |
621 | __fp_abs(__a) < __lo ? __a + __b/2 : // not safe to halve a |
622 | __fp_abs(__b) < __lo ? __a/2 + __b : // not safe to halve b |
623 | __a/2 + __b/2; // otherwise correctly rounded |
624 | } |
625 | |
626 | #endif // _LIBCPP_STD_VER > 17 |
627 | |
628 | _LIBCPP_END_NAMESPACE_STD} } |
629 | |
630 | _LIBCPP_POP_MACROSpop_macro("min") pop_macro("max") |
631 | |
632 | #if defined(_LIBCPP_HAS_PARALLEL_ALGORITHMS) && _LIBCPP_STD_VER14 >= 17 |
633 | # include <__pstl_numeric> |
634 | #endif |
635 | |
636 | #endif // _LIBCPP_NUMERIC |
1 | //==- llvm/CodeGen/MachineDominators.h - Machine Dom Calculation -*- C++ -*-==// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file defines classes mirroring those in llvm/Analysis/Dominators.h, |
10 | // but for target-specific code rather than target-independent IR. |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #ifndef LLVM_CODEGEN_MACHINEDOMINATORS_H |
15 | #define LLVM_CODEGEN_MACHINEDOMINATORS_H |
16 | |
17 | #include "llvm/ADT/SmallSet.h" |
18 | #include "llvm/ADT/SmallVector.h" |
19 | #include "llvm/CodeGen/MachineBasicBlock.h" |
20 | #include "llvm/CodeGen/MachineFunctionPass.h" |
21 | #include "llvm/CodeGen/MachineInstr.h" |
22 | #include "llvm/Support/GenericDomTree.h" |
23 | #include "llvm/Support/GenericDomTreeConstruction.h" |
24 | #include <cassert> |
25 | #include <memory> |
26 | |
27 | namespace llvm { |
28 | |
29 | template <> |
30 | inline void DominatorTreeBase<MachineBasicBlock, false>::addRoot( |
31 | MachineBasicBlock *MBB) { |
32 | this->Roots.push_back(MBB); |
33 | } |
34 | |
35 | extern template class DomTreeNodeBase<MachineBasicBlock>; |
36 | extern template class DominatorTreeBase<MachineBasicBlock, false>; // DomTree |
37 | extern template class DominatorTreeBase<MachineBasicBlock, true>; // PostDomTree |
38 | |
39 | using MachineDomTreeNode = DomTreeNodeBase<MachineBasicBlock>; |
40 | |
41 | //===------------------------------------- |
42 | /// DominatorTree Class - Concrete subclass of DominatorTreeBase that is used to |
43 | /// compute a normal dominator tree. |
44 | /// |
45 | class MachineDominatorTree : public MachineFunctionPass { |
46 | using DomTreeT = DomTreeBase<MachineBasicBlock>; |
47 | |
48 | /// Helper structure used to hold all the basic blocks |
49 | /// involved in the split of a critical edge. |
50 | struct CriticalEdge { |
51 | MachineBasicBlock *FromBB; |
52 | MachineBasicBlock *ToBB; |
53 | MachineBasicBlock *NewBB; |
54 | }; |
55 | |
56 | /// Pile up all the critical edges to be split. |
57 | /// The splitting of a critical edge is local and thus, it is possible |
58 | /// to apply several of those changes at the same time. |
59 | mutable SmallVector<CriticalEdge, 32> CriticalEdgesToSplit; |
60 | |
61 | /// Remember all the basic blocks that are inserted during |
62 | /// edge splitting. |
63 | /// Invariant: NewBBs == all the basic blocks contained in the NewBB |
64 | /// field of all the elements of CriticalEdgesToSplit. |
65 | /// I.e., forall elt in CriticalEdgesToSplit, it exists BB in NewBBs |
66 | /// such as BB == elt.NewBB. |
67 | mutable SmallSet<MachineBasicBlock *, 32> NewBBs; |
68 | |
69 | /// The DominatorTreeBase that is used to compute a normal dominator tree. |
70 | std::unique_ptr<DomTreeT> DT; |
71 | |
72 | /// Apply all the recorded critical edges to the DT. |
73 | /// This updates the underlying DT information in a way that uses |
74 | /// the fast query path of DT as much as possible. |
75 | /// |
76 | /// \post CriticalEdgesToSplit.empty(). |
77 | void applySplitCriticalEdges() const; |
78 | |
79 | public: |
80 | static char ID; // Pass ID, replacement for typeid |
81 | |
82 | MachineDominatorTree(); |
83 | explicit MachineDominatorTree(MachineFunction &MF) : MachineFunctionPass(ID) { |
84 | calculate(MF); |
85 | } |
86 | |
87 | DomTreeT &getBase() { |
88 | if (!DT) DT.reset(new DomTreeT()); |
89 | applySplitCriticalEdges(); |
90 | return *DT; |
91 | } |
92 | |
93 | void getAnalysisUsage(AnalysisUsage &AU) const override; |
94 | |
95 | MachineBasicBlock *getRoot() const { |
96 | applySplitCriticalEdges(); |
97 | return DT->getRoot(); |
98 | } |
99 | |
100 | MachineDomTreeNode *getRootNode() const { |
101 | applySplitCriticalEdges(); |
102 | return DT->getRootNode(); |
103 | } |
104 | |
105 | bool runOnMachineFunction(MachineFunction &F) override; |
106 | |
107 | void calculate(MachineFunction &F); |
108 | |
109 | bool dominates(const MachineDomTreeNode *A, |
110 | const MachineDomTreeNode *B) const { |
111 | applySplitCriticalEdges(); |
112 | return DT->dominates(A, B); |
113 | } |
114 | |
115 | bool dominates(const MachineBasicBlock *A, const MachineBasicBlock *B) const { |
116 | applySplitCriticalEdges(); |
117 | return DT->dominates(A, B); |
118 | } |
119 | |
120 | // dominates - Return true if A dominates B. This performs the |
121 | // special checks necessary if A and B are in the same basic block. |
122 | bool dominates(const MachineInstr *A, const MachineInstr *B) const { |
123 | applySplitCriticalEdges(); |
124 | const MachineBasicBlock *BBA = A->getParent(), *BBB = B->getParent(); |
125 | if (BBA != BBB) return DT->dominates(BBA, BBB); |
126 | |
127 | // Loop through the basic block until we find A or B. |
128 | MachineBasicBlock::const_iterator I = BBA->begin(); |
129 | for (; &*I != A && &*I != B; ++I) |
130 | /*empty*/ ; |
131 | |
132 | return &*I == A; |
133 | } |
134 | |
135 | bool properlyDominates(const MachineDomTreeNode *A, |
136 | const MachineDomTreeNode *B) const { |
137 | applySplitCriticalEdges(); |
138 | return DT->properlyDominates(A, B); |
139 | } |
140 | |
141 | bool properlyDominates(const MachineBasicBlock *A, |
142 | const MachineBasicBlock *B) const { |
143 | applySplitCriticalEdges(); |
144 | return DT->properlyDominates(A, B); |
145 | } |
146 | |
147 | /// findNearestCommonDominator - Find nearest common dominator basic block |
148 | /// for basic block A and B. If there is no such block then return NULL. |
149 | MachineBasicBlock *findNearestCommonDominator(MachineBasicBlock *A, |
150 | MachineBasicBlock *B) { |
151 | applySplitCriticalEdges(); |
152 | return DT->findNearestCommonDominator(A, B); |
153 | } |
154 | |
155 | MachineDomTreeNode *operator[](MachineBasicBlock *BB) const { |
156 | applySplitCriticalEdges(); |
157 | return DT->getNode(BB); |
158 | } |
159 | |
160 | /// getNode - return the (Post)DominatorTree node for the specified basic |
161 | /// block. This is the same as using operator[] on this class. |
162 | /// |
163 | MachineDomTreeNode *getNode(MachineBasicBlock *BB) const { |
164 | applySplitCriticalEdges(); |
165 | return DT->getNode(BB); |
166 | } |
167 | |
168 | /// addNewBlock - Add a new node to the dominator tree information. This |
169 | /// creates a new node as a child of DomBB dominator node,linking it into |
170 | /// the children list of the immediate dominator. |
171 | MachineDomTreeNode *addNewBlock(MachineBasicBlock *BB, |
172 | MachineBasicBlock *DomBB) { |
173 | applySplitCriticalEdges(); |
174 | return DT->addNewBlock(BB, DomBB); |
175 | } |
176 | |
177 | /// changeImmediateDominator - This method is used to update the dominator |
178 | /// tree information when a node's immediate dominator changes. |
179 | /// |
180 | void changeImmediateDominator(MachineBasicBlock *N, |
181 | MachineBasicBlock *NewIDom) { |
182 | applySplitCriticalEdges(); |
183 | DT->changeImmediateDominator(N, NewIDom); |
184 | } |
185 | |
186 | void changeImmediateDominator(MachineDomTreeNode *N, |
187 | MachineDomTreeNode *NewIDom) { |
188 | applySplitCriticalEdges(); |
189 | DT->changeImmediateDominator(N, NewIDom); |
190 | } |
191 | |
192 | /// eraseNode - Removes a node from the dominator tree. Block must not |
193 | /// dominate any other blocks. Removes node from its immediate dominator's |
194 | /// children list. Deletes dominator node associated with basic block BB. |
195 | void eraseNode(MachineBasicBlock *BB) { |
196 | applySplitCriticalEdges(); |
197 | DT->eraseNode(BB); |
198 | } |
199 | |
200 | /// splitBlock - BB is split and now it has one successor. Update dominator |
201 | /// tree to reflect this change. |
202 | void splitBlock(MachineBasicBlock* NewBB) { |
203 | applySplitCriticalEdges(); |
204 | DT->splitBlock(NewBB); |
205 | } |
206 | |
207 | /// isReachableFromEntry - Return true if A is dominated by the entry |
208 | /// block of the function containing it. |
209 | bool isReachableFromEntry(const MachineBasicBlock *A) { |
210 | applySplitCriticalEdges(); |
211 | return DT->isReachableFromEntry(A); |
212 | } |
213 | |
214 | void releaseMemory() override; |
215 | |
216 | void verifyAnalysis() const override; |
217 | |
218 | void print(raw_ostream &OS, const Module*) const override; |
219 | |
220 | /// Record that the critical edge (FromBB, ToBB) has been |
221 | /// split with NewBB. |
222 | /// This is best to use this method instead of directly update the |
223 | /// underlying information, because this helps mitigating the |
224 | /// number of time the DT information is invalidated. |
225 | /// |
226 | /// \note Do not use this method with regular edges. |
227 | /// |
228 | /// \note To benefit from the compile time improvement incurred by this |
229 | /// method, the users of this method have to limit the queries to the DT |
230 | /// interface between two edges splitting. In other words, they have to |
231 | /// pack the splitting of critical edges as much as possible. |
232 | void recordSplitCriticalEdge(MachineBasicBlock *FromBB, |
233 | MachineBasicBlock *ToBB, |
234 | MachineBasicBlock *NewBB) { |
235 | bool Inserted = NewBBs.insert(NewBB).second; |
236 | (void)Inserted; |
237 | assert(Inserted &&((void)0) |
238 | "A basic block inserted via edge splitting cannot appear twice")((void)0); |
239 | CriticalEdgesToSplit.push_back({FromBB, ToBB, NewBB}); |
240 | } |
241 | }; |
242 | |
243 | //===------------------------------------- |
244 | /// DominatorTree GraphTraits specialization so the DominatorTree can be |
245 | /// iterable by generic graph iterators. |
246 | /// |
247 | |
248 | template <class Node, class ChildIterator> |
249 | struct MachineDomTreeGraphTraitsBase { |
250 | using NodeRef = Node *; |
251 | using ChildIteratorType = ChildIterator; |
252 | |
253 | static NodeRef getEntryNode(NodeRef N) { return N; } |
254 | static ChildIteratorType child_begin(NodeRef N) { return N->begin(); } |
255 | static ChildIteratorType child_end(NodeRef N) { return N->end(); } |
256 | }; |
257 | |
258 | template <class T> struct GraphTraits; |
259 | |
260 | template <> |
261 | struct GraphTraits<MachineDomTreeNode *> |
262 | : public MachineDomTreeGraphTraitsBase<MachineDomTreeNode, |
263 | MachineDomTreeNode::const_iterator> { |
264 | }; |
265 | |
266 | template <> |
267 | struct GraphTraits<const MachineDomTreeNode *> |
268 | : public MachineDomTreeGraphTraitsBase<const MachineDomTreeNode, |
269 | MachineDomTreeNode::const_iterator> { |
270 | }; |
271 | |
272 | template <> struct GraphTraits<MachineDominatorTree*> |
273 | : public GraphTraits<MachineDomTreeNode *> { |
274 | static NodeRef getEntryNode(MachineDominatorTree *DT) { |
275 | return DT->getRootNode(); |
276 | } |
277 | }; |
278 | |
279 | } // end namespace llvm |
280 | |
281 | #endif // LLVM_CODEGEN_MACHINEDOMINATORS_H |
1 | //===- GenericDomTree.h - Generic dominator trees for graphs ----*- C++ -*-===// | |||
2 | // | |||
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. | |||
4 | // See https://llvm.org/LICENSE.txt for license information. | |||
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception | |||
6 | // | |||
7 | //===----------------------------------------------------------------------===// | |||
8 | /// \file | |||
9 | /// | |||
10 | /// This file defines a set of templates that efficiently compute a dominator | |||
11 | /// tree over a generic graph. This is used typically in LLVM for fast | |||
12 | /// dominance queries on the CFG, but is fully generic w.r.t. the underlying | |||
13 | /// graph types. | |||
14 | /// | |||
15 | /// Unlike ADT/* graph algorithms, generic dominator tree has more requirements | |||
16 | /// on the graph's NodeRef. The NodeRef should be a pointer and, | |||
17 | /// NodeRef->getParent() must return the parent node that is also a pointer. | |||
18 | /// | |||
19 | /// FIXME: Maybe GenericDomTree needs a TreeTraits, instead of GraphTraits. | |||
20 | /// | |||
21 | //===----------------------------------------------------------------------===// | |||
22 | ||||
23 | #ifndef LLVM_SUPPORT_GENERICDOMTREE_H | |||
24 | #define LLVM_SUPPORT_GENERICDOMTREE_H | |||
25 | ||||
26 | #include "llvm/ADT/DenseMap.h" | |||
27 | #include "llvm/ADT/GraphTraits.h" | |||
28 | #include "llvm/ADT/STLExtras.h" | |||
29 | #include "llvm/ADT/SmallPtrSet.h" | |||
30 | #include "llvm/ADT/SmallVector.h" | |||
31 | #include "llvm/Support/CFGDiff.h" | |||
32 | #include "llvm/Support/CFGUpdate.h" | |||
33 | #include "llvm/Support/raw_ostream.h" | |||
34 | #include <algorithm> | |||
35 | #include <cassert> | |||
36 | #include <cstddef> | |||
37 | #include <iterator> | |||
38 | #include <memory> | |||
39 | #include <type_traits> | |||
40 | #include <utility> | |||
41 | ||||
42 | namespace llvm { | |||
43 | ||||
44 | template <typename NodeT, bool IsPostDom> | |||
45 | class DominatorTreeBase; | |||
46 | ||||
47 | namespace DomTreeBuilder { | |||
48 | template <typename DomTreeT> | |||
49 | struct SemiNCAInfo; | |||
50 | } // namespace DomTreeBuilder | |||
51 | ||||
52 | /// Base class for the actual dominator tree node. | |||
53 | template <class NodeT> class DomTreeNodeBase { | |||
54 | friend class PostDominatorTree; | |||
55 | friend class DominatorTreeBase<NodeT, false>; | |||
56 | friend class DominatorTreeBase<NodeT, true>; | |||
57 | friend struct DomTreeBuilder::SemiNCAInfo<DominatorTreeBase<NodeT, false>>; | |||
58 | friend struct DomTreeBuilder::SemiNCAInfo<DominatorTreeBase<NodeT, true>>; | |||
59 | ||||
60 | NodeT *TheBB; | |||
61 | DomTreeNodeBase *IDom; | |||
62 | unsigned Level; | |||
63 | SmallVector<DomTreeNodeBase *, 4> Children; | |||
64 | mutable unsigned DFSNumIn = ~0; | |||
65 | mutable unsigned DFSNumOut = ~0; | |||
66 | ||||
67 | public: | |||
68 | DomTreeNodeBase(NodeT *BB, DomTreeNodeBase *iDom) | |||
69 | : TheBB(BB), IDom(iDom), Level(IDom ? IDom->Level + 1 : 0) {} | |||
70 | ||||
71 | using iterator = typename SmallVector<DomTreeNodeBase *, 4>::iterator; | |||
72 | using const_iterator = | |||
73 | typename SmallVector<DomTreeNodeBase *, 4>::const_iterator; | |||
74 | ||||
75 | iterator begin() { return Children.begin(); } | |||
76 | iterator end() { return Children.end(); } | |||
77 | const_iterator begin() const { return Children.begin(); } | |||
78 | const_iterator end() const { return Children.end(); } | |||
79 | ||||
80 | DomTreeNodeBase *const &back() const { return Children.back(); } | |||
81 | DomTreeNodeBase *&back() { return Children.back(); } | |||
82 | ||||
83 | iterator_range<iterator> children() { return make_range(begin(), end()); } | |||
84 | iterator_range<const_iterator> children() const { | |||
85 | return make_range(begin(), end()); | |||
86 | } | |||
87 | ||||
88 | NodeT *getBlock() const { return TheBB; } | |||
89 | DomTreeNodeBase *getIDom() const { return IDom; } | |||
90 | unsigned getLevel() const { return Level; } | |||
91 | ||||
92 | std::unique_ptr<DomTreeNodeBase> addChild( | |||
93 | std::unique_ptr<DomTreeNodeBase> C) { | |||
94 | Children.push_back(C.get()); | |||
95 | return C; | |||
96 | } | |||
97 | ||||
98 | bool isLeaf() const { return Children.empty(); } | |||
99 | size_t getNumChildren() const { return Children.size(); } | |||
100 | ||||
101 | void clearAllChildren() { Children.clear(); } | |||
102 | ||||
103 | bool compare(const DomTreeNodeBase *Other) const { | |||
104 | if (getNumChildren() != Other->getNumChildren()) | |||
105 | return true; | |||
106 | ||||
107 | if (Level != Other->Level) return true; | |||
108 | ||||
109 | SmallPtrSet<const NodeT *, 4> OtherChildren; | |||
110 | for (const DomTreeNodeBase *I : *Other) { | |||
111 | const NodeT *Nd = I->getBlock(); | |||
112 | OtherChildren.insert(Nd); | |||
113 | } | |||
114 | ||||
115 | for (const DomTreeNodeBase *I : *this) { | |||
116 | const NodeT *N = I->getBlock(); | |||
117 | if (OtherChildren.count(N) == 0) | |||
118 | return true; | |||
119 | } | |||
120 | return false; | |||
121 | } | |||
122 | ||||
123 | void setIDom(DomTreeNodeBase *NewIDom) { | |||
124 | assert(IDom && "No immediate dominator?")((void)0); | |||
125 | if (IDom == NewIDom) return; | |||
126 | ||||
127 | auto I = find(IDom->Children, this); | |||
128 | assert(I != IDom->Children.end() &&((void)0) | |||
129 | "Not in immediate dominator children set!")((void)0); | |||
130 | // I am no longer your child... | |||
131 | IDom->Children.erase(I); | |||
132 | ||||
133 | // Switch to new dominator | |||
134 | IDom = NewIDom; | |||
135 | IDom->Children.push_back(this); | |||
136 | ||||
137 | UpdateLevel(); | |||
138 | } | |||
139 | ||||
140 | /// getDFSNumIn/getDFSNumOut - These return the DFS visitation order for nodes | |||
141 | /// in the dominator tree. They are only guaranteed valid if | |||
142 | /// updateDFSNumbers() has been called. | |||
143 | unsigned getDFSNumIn() const { return DFSNumIn; } | |||
144 | unsigned getDFSNumOut() const { return DFSNumOut; } | |||
145 | ||||
146 | private: | |||
147 | // Return true if this node is dominated by other. Use this only if DFS info | |||
148 | // is valid. | |||
149 | bool DominatedBy(const DomTreeNodeBase *other) const { | |||
150 | return this->DFSNumIn >= other->DFSNumIn && | |||
151 | this->DFSNumOut <= other->DFSNumOut; | |||
152 | } | |||
153 | ||||
154 | void UpdateLevel() { | |||
155 | assert(IDom)((void)0); | |||
156 | if (Level == IDom->Level + 1) return; | |||
157 | ||||
158 | SmallVector<DomTreeNodeBase *, 64> WorkStack = {this}; | |||
159 | ||||
160 | while (!WorkStack.empty()) { | |||
161 | DomTreeNodeBase *Current = WorkStack.pop_back_val(); | |||
162 | Current->Level = Current->IDom->Level + 1; | |||
163 | ||||
164 | for (DomTreeNodeBase *C : *Current) { | |||
165 | assert(C->IDom)((void)0); | |||
166 | if (C->Level != C->IDom->Level + 1) WorkStack.push_back(C); | |||
167 | } | |||
168 | } | |||
169 | } | |||
170 | }; | |||
171 | ||||
172 | template <class NodeT> | |||
173 | raw_ostream &operator<<(raw_ostream &O, const DomTreeNodeBase<NodeT> *Node) { | |||
174 | if (Node->getBlock()) | |||
175 | Node->getBlock()->printAsOperand(O, false); | |||
176 | else | |||
177 | O << " <<exit node>>"; | |||
178 | ||||
179 | O << " {" << Node->getDFSNumIn() << "," << Node->getDFSNumOut() << "} [" | |||
180 | << Node->getLevel() << "]\n"; | |||
181 | ||||
182 | return O; | |||
183 | } | |||
184 | ||||
185 | template <class NodeT> | |||
186 | void PrintDomTree(const DomTreeNodeBase<NodeT> *N, raw_ostream &O, | |||
187 | unsigned Lev) { | |||
188 | O.indent(2 * Lev) << "[" << Lev << "] " << N; | |||
189 | for (typename DomTreeNodeBase<NodeT>::const_iterator I = N->begin(), | |||
190 | E = N->end(); | |||
191 | I != E; ++I) | |||
192 | PrintDomTree<NodeT>(*I, O, Lev + 1); | |||
193 | } | |||
194 | ||||
195 | namespace DomTreeBuilder { | |||
196 | // The routines below are provided in a separate header but referenced here. | |||
197 | template <typename DomTreeT> | |||
198 | void Calculate(DomTreeT &DT); | |||
199 | ||||
200 | template <typename DomTreeT> | |||
201 | void CalculateWithUpdates(DomTreeT &DT, | |||
202 | ArrayRef<typename DomTreeT::UpdateType> Updates); | |||
203 | ||||
204 | template <typename DomTreeT> | |||
205 | void InsertEdge(DomTreeT &DT, typename DomTreeT::NodePtr From, | |||
206 | typename DomTreeT::NodePtr To); | |||
207 | ||||
208 | template <typename DomTreeT> | |||
209 | void DeleteEdge(DomTreeT &DT, typename DomTreeT::NodePtr From, | |||
210 | typename DomTreeT::NodePtr To); | |||
211 | ||||
212 | template <typename DomTreeT> | |||
213 | void ApplyUpdates(DomTreeT &DT, | |||
214 | GraphDiff<typename DomTreeT::NodePtr, | |||
215 | DomTreeT::IsPostDominator> &PreViewCFG, | |||
216 | GraphDiff<typename DomTreeT::NodePtr, | |||
217 | DomTreeT::IsPostDominator> *PostViewCFG); | |||
218 | ||||
219 | template <typename DomTreeT> | |||
220 | bool Verify(const DomTreeT &DT, typename DomTreeT::VerificationLevel VL); | |||
221 | } // namespace DomTreeBuilder | |||
222 | ||||
223 | /// Core dominator tree base class. | |||
224 | /// | |||
225 | /// This class is a generic template over graph nodes. It is instantiated for | |||
226 | /// various graphs in the LLVM IR or in the code generator. | |||
227 | template <typename NodeT, bool IsPostDom> | |||
228 | class DominatorTreeBase { | |||
229 | public: | |||
230 | static_assert(std::is_pointer<typename GraphTraits<NodeT *>::NodeRef>::value, | |||
231 | "Currently DominatorTreeBase supports only pointer nodes"); | |||
232 | using NodeType = NodeT; | |||
233 | using NodePtr = NodeT *; | |||
234 | using ParentPtr = decltype(std::declval<NodeT *>()->getParent()); | |||
235 | static_assert(std::is_pointer<ParentPtr>::value, | |||
236 | "Currently NodeT's parent must be a pointer type"); | |||
237 | using ParentType = std::remove_pointer_t<ParentPtr>; | |||
238 | static constexpr bool IsPostDominator = IsPostDom; | |||
239 | ||||
240 | using UpdateType = cfg::Update<NodePtr>; | |||
241 | using UpdateKind = cfg::UpdateKind; | |||
242 | static constexpr UpdateKind Insert = UpdateKind::Insert; | |||
243 | static constexpr UpdateKind Delete = UpdateKind::Delete; | |||
244 | ||||
245 | enum class VerificationLevel { Fast, Basic, Full }; | |||
246 | ||||
247 | protected: | |||
248 | // Dominators always have a single root, postdominators can have more. | |||
249 | SmallVector<NodeT *, IsPostDom ? 4 : 1> Roots; | |||
250 | ||||
251 | using DomTreeNodeMapType = | |||
252 | DenseMap<NodeT *, std::unique_ptr<DomTreeNodeBase<NodeT>>>; | |||
253 | DomTreeNodeMapType DomTreeNodes; | |||
254 | DomTreeNodeBase<NodeT> *RootNode = nullptr; | |||
255 | ParentPtr Parent = nullptr; | |||
256 | ||||
257 | mutable bool DFSInfoValid = false; | |||
258 | mutable unsigned int SlowQueries = 0; | |||
259 | ||||
260 | friend struct DomTreeBuilder::SemiNCAInfo<DominatorTreeBase>; | |||
261 | ||||
262 | public: | |||
263 | DominatorTreeBase() {} | |||
264 | ||||
265 | DominatorTreeBase(DominatorTreeBase &&Arg) | |||
266 | : Roots(std::move(Arg.Roots)), | |||
267 | DomTreeNodes(std::move(Arg.DomTreeNodes)), | |||
268 | RootNode(Arg.RootNode), | |||
269 | Parent(Arg.Parent), | |||
270 | DFSInfoValid(Arg.DFSInfoValid), | |||
271 | SlowQueries(Arg.SlowQueries) { | |||
272 | Arg.wipe(); | |||
273 | } | |||
274 | ||||
275 | DominatorTreeBase &operator=(DominatorTreeBase &&RHS) { | |||
276 | Roots = std::move(RHS.Roots); | |||
277 | DomTreeNodes = std::move(RHS.DomTreeNodes); | |||
278 | RootNode = RHS.RootNode; | |||
279 | Parent = RHS.Parent; | |||
280 | DFSInfoValid = RHS.DFSInfoValid; | |||
281 | SlowQueries = RHS.SlowQueries; | |||
282 | RHS.wipe(); | |||
283 | return *this; | |||
284 | } | |||
285 | ||||
286 | DominatorTreeBase(const DominatorTreeBase &) = delete; | |||
287 | DominatorTreeBase &operator=(const DominatorTreeBase &) = delete; | |||
288 | ||||
289 | /// Iteration over roots. | |||
290 | /// | |||
291 | /// This may include multiple blocks if we are computing post dominators. | |||
292 | /// For forward dominators, this will always be a single block (the entry | |||
293 | /// block). | |||
294 | using root_iterator = typename SmallVectorImpl<NodeT *>::iterator; | |||
295 | using const_root_iterator = typename SmallVectorImpl<NodeT *>::const_iterator; | |||
296 | ||||
297 | root_iterator root_begin() { return Roots.begin(); } | |||
298 | const_root_iterator root_begin() const { return Roots.begin(); } | |||
299 | root_iterator root_end() { return Roots.end(); } | |||
300 | const_root_iterator root_end() const { return Roots.end(); } | |||
301 | ||||
302 | size_t root_size() const { return Roots.size(); } | |||
303 | ||||
304 | iterator_range<root_iterator> roots() { | |||
305 | return make_range(root_begin(), root_end()); | |||
306 | } | |||
307 | iterator_range<const_root_iterator> roots() const { | |||
308 | return make_range(root_begin(), root_end()); | |||
309 | } | |||
310 | ||||
311 | /// isPostDominator - Returns true if analysis based of postdoms | |||
312 | /// | |||
313 | bool isPostDominator() const { return IsPostDominator; } | |||
314 | ||||
315 | /// compare - Return false if the other dominator tree base matches this | |||
316 | /// dominator tree base. Otherwise return true. | |||
317 | bool compare(const DominatorTreeBase &Other) const { | |||
318 | if (Parent != Other.Parent) return true; | |||
319 | ||||
320 | if (Roots.size() != Other.Roots.size()) | |||
321 | return true; | |||
322 | ||||
323 | if (!std::is_permutation(Roots.begin(), Roots.end(), Other.Roots.begin())) | |||
324 | return true; | |||
325 | ||||
326 | const DomTreeNodeMapType &OtherDomTreeNodes = Other.DomTreeNodes; | |||
327 | if (DomTreeNodes.size() != OtherDomTreeNodes.size()) | |||
328 | return true; | |||
329 | ||||
330 | for (const auto &DomTreeNode : DomTreeNodes) { | |||
331 | NodeT *BB = DomTreeNode.first; | |||
332 | typename DomTreeNodeMapType::const_iterator OI = | |||
333 | OtherDomTreeNodes.find(BB); | |||
334 | if (OI == OtherDomTreeNodes.end()) | |||
335 | return true; | |||
336 | ||||
337 | DomTreeNodeBase<NodeT> &MyNd = *DomTreeNode.second; | |||
338 | DomTreeNodeBase<NodeT> &OtherNd = *OI->second; | |||
339 | ||||
340 | if (MyNd.compare(&OtherNd)) | |||
341 | return true; | |||
342 | } | |||
343 | ||||
344 | return false; | |||
345 | } | |||
346 | ||||
347 | /// getNode - return the (Post)DominatorTree node for the specified basic | |||
348 | /// block. This is the same as using operator[] on this class. The result | |||
349 | /// may (but is not required to) be null for a forward (backwards) | |||
350 | /// statically unreachable block. | |||
351 | DomTreeNodeBase<NodeT> *getNode(const NodeT *BB) const { | |||
352 | auto I = DomTreeNodes.find(BB); | |||
353 | if (I != DomTreeNodes.end()) | |||
354 | return I->second.get(); | |||
355 | return nullptr; | |||
356 | } | |||
357 | ||||
358 | /// See getNode. | |||
359 | DomTreeNodeBase<NodeT> *operator[](const NodeT *BB) const { | |||
360 | return getNode(BB); | |||
361 | } | |||
362 | ||||
363 | /// getRootNode - This returns the entry node for the CFG of the function. If | |||
364 | /// this tree represents the post-dominance relations for a function, however, | |||
365 | /// this root may be a node with the block == NULL. This is the case when | |||
366 | /// there are multiple exit nodes from a particular function. Consumers of | |||
367 | /// post-dominance information must be capable of dealing with this | |||
368 | /// possibility. | |||
369 | /// | |||
370 | DomTreeNodeBase<NodeT> *getRootNode() { return RootNode; } | |||
371 | const DomTreeNodeBase<NodeT> *getRootNode() const { return RootNode; } | |||
372 | ||||
373 | /// Get all nodes dominated by R, including R itself. | |||
374 | void getDescendants(NodeT *R, SmallVectorImpl<NodeT *> &Result) const { | |||
375 | Result.clear(); | |||
376 | const DomTreeNodeBase<NodeT> *RN = getNode(R); | |||
377 | if (!RN) | |||
378 | return; // If R is unreachable, it will not be present in the DOM tree. | |||
379 | SmallVector<const DomTreeNodeBase<NodeT> *, 8> WL; | |||
380 | WL.push_back(RN); | |||
381 | ||||
382 | while (!WL.empty()) { | |||
383 | const DomTreeNodeBase<NodeT> *N = WL.pop_back_val(); | |||
384 | Result.push_back(N->getBlock()); | |||
385 | WL.append(N->begin(), N->end()); | |||
386 | } | |||
387 | } | |||
388 | ||||
389 | /// properlyDominates - Returns true iff A dominates B and A != B. | |||
390 | /// Note that this is not a constant time operation! | |||
391 | /// | |||
392 | bool properlyDominates(const DomTreeNodeBase<NodeT> *A, | |||
393 | const DomTreeNodeBase<NodeT> *B) const { | |||
394 | if (!A || !B) | |||
395 | return false; | |||
396 | if (A == B) | |||
397 | return false; | |||
398 | return dominates(A, B); | |||
399 | } | |||
400 | ||||
401 | bool properlyDominates(const NodeT *A, const NodeT *B) const; | |||
402 | ||||
403 | /// isReachableFromEntry - Return true if A is dominated by the entry | |||
404 | /// block of the function containing it. | |||
405 | bool isReachableFromEntry(const NodeT *A) const { | |||
406 | assert(!this->isPostDominator() &&((void)0) | |||
407 | "This is not implemented for post dominators")((void)0); | |||
408 | return isReachableFromEntry(getNode(const_cast<NodeT *>(A))); | |||
409 | } | |||
410 | ||||
411 | bool isReachableFromEntry(const DomTreeNodeBase<NodeT> *A) const { return A; } | |||
412 | ||||
413 | /// dominates - Returns true iff A dominates B. Note that this is not a | |||
414 | /// constant time operation! | |||
415 | /// | |||
416 | bool dominates(const DomTreeNodeBase<NodeT> *A, | |||
417 | const DomTreeNodeBase<NodeT> *B) const { | |||
418 | // A node trivially dominates itself. | |||
419 | if (B == A) | |||
420 | return true; | |||
421 | ||||
422 | // An unreachable node is dominated by anything. | |||
423 | if (!isReachableFromEntry(B)) | |||
424 | return true; | |||
425 | ||||
426 | // And dominates nothing. | |||
427 | if (!isReachableFromEntry(A)) | |||
428 | return false; | |||
429 | ||||
430 | if (B->getIDom() == A) return true; | |||
431 | ||||
432 | if (A->getIDom() == B) return false; | |||
433 | ||||
434 | // A can only dominate B if it is higher in the tree. | |||
435 | if (A->getLevel() >= B->getLevel()) return false; | |||
436 | ||||
437 | // Compare the result of the tree walk and the dfs numbers, if expensive | |||
438 | // checks are enabled. | |||
439 | #ifdef EXPENSIVE_CHECKS | |||
440 | assert((!DFSInfoValid ||((void)0) | |||
441 | (dominatedBySlowTreeWalk(A, B) == B->DominatedBy(A))) &&((void)0) | |||
442 | "Tree walk disagrees with dfs numbers!")((void)0); | |||
443 | #endif | |||
444 | ||||
445 | if (DFSInfoValid) | |||
446 | return B->DominatedBy(A); | |||
447 | ||||
448 | // If we end up with too many slow queries, just update the | |||
449 | // DFS numbers on the theory that we are going to keep querying. | |||
450 | SlowQueries++; | |||
451 | if (SlowQueries > 32) { | |||
452 | updateDFSNumbers(); | |||
453 | return B->DominatedBy(A); | |||
454 | } | |||
455 | ||||
456 | return dominatedBySlowTreeWalk(A, B); | |||
457 | } | |||
458 | ||||
459 | bool dominates(const NodeT *A, const NodeT *B) const; | |||
460 | ||||
461 | NodeT *getRoot() const { | |||
462 | assert(this->Roots.size() == 1 && "Should always have entry node!")((void)0); | |||
463 | return this->Roots[0]; | |||
464 | } | |||
465 | ||||
466 | /// Find nearest common dominator basic block for basic block A and B. A and B | |||
467 | /// must have tree nodes. | |||
468 | NodeT *findNearestCommonDominator(NodeT *A, NodeT *B) const { | |||
469 | assert(A && B && "Pointers are not valid")((void)0); | |||
470 | assert(A->getParent() == B->getParent() &&((void)0) | |||
471 | "Two blocks are not in same function")((void)0); | |||
472 | ||||
473 | // If either A or B is a entry block then it is nearest common dominator | |||
474 | // (for forward-dominators). | |||
475 | if (!isPostDominator()) { | |||
476 | NodeT &Entry = A->getParent()->front(); | |||
477 | if (A == &Entry || B == &Entry) | |||
478 | return &Entry; | |||
479 | } | |||
480 | ||||
481 | DomTreeNodeBase<NodeT> *NodeA = getNode(A); | |||
482 | DomTreeNodeBase<NodeT> *NodeB = getNode(B); | |||
483 | assert(NodeA && "A must be in the tree")((void)0); | |||
484 | assert(NodeB && "B must be in the tree")((void)0); | |||
485 | ||||
486 | // Use level information to go up the tree until the levels match. Then | |||
487 | // continue going up til we arrive at the same node. | |||
488 | while (NodeA != NodeB) { | |||
489 | if (NodeA->getLevel() < NodeB->getLevel()) std::swap(NodeA, NodeB); | |||
490 | ||||
491 | NodeA = NodeA->IDom; | |||
492 | } | |||
493 | ||||
494 | return NodeA->getBlock(); | |||
| ||||
495 | } | |||
496 | ||||
497 | const NodeT *findNearestCommonDominator(const NodeT *A, | |||
498 | const NodeT *B) const { | |||
499 | // Cast away the const qualifiers here. This is ok since | |||
500 | // const is re-introduced on the return type. | |||
501 | return findNearestCommonDominator(const_cast<NodeT *>(A), | |||
502 | const_cast<NodeT *>(B)); | |||
503 | } | |||
504 | ||||
505 | bool isVirtualRoot(const DomTreeNodeBase<NodeT> *A) const { | |||
506 | return isPostDominator() && !A->getBlock(); | |||
507 | } | |||
508 | ||||
509 | //===--------------------------------------------------------------------===// | |||
510 | // API to update (Post)DominatorTree information based on modifications to | |||
511 | // the CFG... | |||
512 | ||||
513 | /// Inform the dominator tree about a sequence of CFG edge insertions and | |||
514 | /// deletions and perform a batch update on the tree. | |||
515 | /// | |||
516 | /// This function should be used when there were multiple CFG updates after | |||
517 | /// the last dominator tree update. It takes care of performing the updates | |||
518 | /// in sync with the CFG and optimizes away the redundant operations that | |||
519 | /// cancel each other. | |||
520 | /// The functions expects the sequence of updates to be balanced. Eg.: | |||
521 | /// - {{Insert, A, B}, {Delete, A, B}, {Insert, A, B}} is fine, because | |||
522 | /// logically it results in a single insertions. | |||
523 | /// - {{Insert, A, B}, {Insert, A, B}} is invalid, because it doesn't make | |||
524 | /// sense to insert the same edge twice. | |||
525 | /// | |||
526 | /// What's more, the functions assumes that it's safe to ask every node in the | |||
527 | /// CFG about its children and inverse children. This implies that deletions | |||
528 | /// of CFG edges must not delete the CFG nodes before calling this function. | |||
529 | /// | |||
530 | /// The applyUpdates function can reorder the updates and remove redundant | |||
531 | /// ones internally. The batch updater is also able to detect sequences of | |||
532 | /// zero and exactly one update -- it's optimized to do less work in these | |||
533 | /// cases. | |||
534 | /// | |||
535 | /// Note that for postdominators it automatically takes care of applying | |||
536 | /// updates on reverse edges internally (so there's no need to swap the | |||
537 | /// From and To pointers when constructing DominatorTree::UpdateType). | |||
538 | /// The type of updates is the same for DomTreeBase<T> and PostDomTreeBase<T> | |||
539 | /// with the same template parameter T. | |||
540 | /// | |||
541 | /// \param Updates An unordered sequence of updates to perform. The current | |||
542 | /// CFG and the reverse of these updates provides the pre-view of the CFG. | |||
543 | /// | |||
544 | void applyUpdates(ArrayRef<UpdateType> Updates) { | |||
545 | GraphDiff<NodePtr, IsPostDominator> PreViewCFG( | |||
546 | Updates, /*ReverseApplyUpdates=*/true); | |||
547 | DomTreeBuilder::ApplyUpdates(*this, PreViewCFG, nullptr); | |||
548 | } | |||
549 | ||||
550 | /// \param Updates An unordered sequence of updates to perform. The current | |||
551 | /// CFG and the reverse of these updates provides the pre-view of the CFG. | |||
552 | /// \param PostViewUpdates An unordered sequence of update to perform in order | |||
553 | /// to obtain a post-view of the CFG. The DT will be updated assuming the | |||
554 | /// obtained PostViewCFG is the desired end state. | |||
555 | void applyUpdates(ArrayRef<UpdateType> Updates, | |||
556 | ArrayRef<UpdateType> PostViewUpdates) { | |||
557 | if (Updates.empty()) { | |||
558 | GraphDiff<NodePtr, IsPostDom> PostViewCFG(PostViewUpdates); | |||
559 | DomTreeBuilder::ApplyUpdates(*this, PostViewCFG, &PostViewCFG); | |||
560 | } else { | |||
561 | // PreViewCFG needs to merge Updates and PostViewCFG. The updates in | |||
562 | // Updates need to be reversed, and match the direction in PostViewCFG. | |||
563 | // The PostViewCFG is created with updates reversed (equivalent to changes | |||
564 | // made to the CFG), so the PreViewCFG needs all the updates reverse | |||
565 | // applied. | |||
566 | SmallVector<UpdateType> AllUpdates(Updates.begin(), Updates.end()); | |||
567 | append_range(AllUpdates, PostViewUpdates); | |||
568 | GraphDiff<NodePtr, IsPostDom> PreViewCFG(AllUpdates, | |||
569 | /*ReverseApplyUpdates=*/true); | |||
570 | GraphDiff<NodePtr, IsPostDom> PostViewCFG(PostViewUpdates); | |||
571 | DomTreeBuilder::ApplyUpdates(*this, PreViewCFG, &PostViewCFG); | |||
572 | } | |||
573 | } | |||
574 | ||||
575 | /// Inform the dominator tree about a CFG edge insertion and update the tree. | |||
576 | /// | |||
577 | /// This function has to be called just before or just after making the update | |||
578 | /// on the actual CFG. There cannot be any other updates that the dominator | |||
579 | /// tree doesn't know about. | |||
580 | /// | |||
581 | /// Note that for postdominators it automatically takes care of inserting | |||
582 | /// a reverse edge internally (so there's no need to swap the parameters). | |||
583 | /// | |||
584 | void insertEdge(NodeT *From, NodeT *To) { | |||
585 | assert(From)((void)0); | |||
586 | assert(To)((void)0); | |||
587 | assert(From->getParent() == Parent)((void)0); | |||
588 | assert(To->getParent() == Parent)((void)0); | |||
589 | DomTreeBuilder::InsertEdge(*this, From, To); | |||
590 | } | |||
591 | ||||
592 | /// Inform the dominator tree about a CFG edge deletion and update the tree. | |||
593 | /// | |||
594 | /// This function has to be called just after making the update on the actual | |||
595 | /// CFG. An internal functions checks if the edge doesn't exist in the CFG in | |||
596 | /// DEBUG mode. There cannot be any other updates that the | |||
597 | /// dominator tree doesn't know about. | |||
598 | /// | |||
599 | /// Note that for postdominators it automatically takes care of deleting | |||
600 | /// a reverse edge internally (so there's no need to swap the parameters). | |||
601 | /// | |||
602 | void deleteEdge(NodeT *From, NodeT *To) { | |||
603 | assert(From)((void)0); | |||
604 | assert(To)((void)0); | |||
605 | assert(From->getParent() == Parent)((void)0); | |||
606 | assert(To->getParent() == Parent)((void)0); | |||
607 | DomTreeBuilder::DeleteEdge(*this, From, To); | |||
608 | } | |||
609 | ||||
610 | /// Add a new node to the dominator tree information. | |||
611 | /// | |||
612 | /// This creates a new node as a child of DomBB dominator node, linking it | |||
613 | /// into the children list of the immediate dominator. | |||
614 | /// | |||
615 | /// \param BB New node in CFG. | |||
616 | /// \param DomBB CFG node that is dominator for BB. | |||
617 | /// \returns New dominator tree node that represents new CFG node. | |||
618 | /// | |||
619 | DomTreeNodeBase<NodeT> *addNewBlock(NodeT *BB, NodeT *DomBB) { | |||
620 | assert(getNode(BB) == nullptr && "Block already in dominator tree!")((void)0); | |||
621 | DomTreeNodeBase<NodeT> *IDomNode = getNode(DomBB); | |||
622 | assert(IDomNode && "Not immediate dominator specified for block!")((void)0); | |||
623 | DFSInfoValid = false; | |||
624 | return createChild(BB, IDomNode); | |||
625 | } | |||
626 | ||||
627 | /// Add a new node to the forward dominator tree and make it a new root. | |||
628 | /// | |||
629 | /// \param BB New node in CFG. | |||
630 | /// \returns New dominator tree node that represents new CFG node. | |||
631 | /// | |||
632 | DomTreeNodeBase<NodeT> *setNewRoot(NodeT *BB) { | |||
633 | assert(getNode(BB) == nullptr && "Block already in dominator tree!")((void)0); | |||
634 | assert(!this->isPostDominator() &&((void)0) | |||
635 | "Cannot change root of post-dominator tree")((void)0); | |||
636 | DFSInfoValid = false; | |||
637 | DomTreeNodeBase<NodeT> *NewNode = createNode(BB); | |||
638 | if (Roots.empty()) { | |||
639 | addRoot(BB); | |||
640 | } else { | |||
641 | assert(Roots.size() == 1)((void)0); | |||
642 | NodeT *OldRoot = Roots.front(); | |||
643 | auto &OldNode = DomTreeNodes[OldRoot]; | |||
644 | OldNode = NewNode->addChild(std::move(DomTreeNodes[OldRoot])); | |||
645 | OldNode->IDom = NewNode; | |||
646 | OldNode->UpdateLevel(); | |||
647 | Roots[0] = BB; | |||
648 | } | |||
649 | return RootNode = NewNode; | |||
650 | } | |||
651 | ||||
652 | /// changeImmediateDominator - This method is used to update the dominator | |||
653 | /// tree information when a node's immediate dominator changes. | |||
654 | /// | |||
655 | void changeImmediateDominator(DomTreeNodeBase<NodeT> *N, | |||
656 | DomTreeNodeBase<NodeT> *NewIDom) { | |||
657 | assert(N && NewIDom && "Cannot change null node pointers!")((void)0); | |||
658 | DFSInfoValid = false; | |||
659 | N->setIDom(NewIDom); | |||
660 | } | |||
661 | ||||
662 | void changeImmediateDominator(NodeT *BB, NodeT *NewBB) { | |||
663 | changeImmediateDominator(getNode(BB), getNode(NewBB)); | |||
664 | } | |||
665 | ||||
666 | /// eraseNode - Removes a node from the dominator tree. Block must not | |||
667 | /// dominate any other blocks. Removes node from its immediate dominator's | |||
668 | /// children list. Deletes dominator node associated with basic block BB. | |||
669 | void eraseNode(NodeT *BB) { | |||
670 | DomTreeNodeBase<NodeT> *Node = getNode(BB); | |||
671 | assert(Node && "Removing node that isn't in dominator tree.")((void)0); | |||
672 | assert(Node->isLeaf() && "Node is not a leaf node.")((void)0); | |||
673 | ||||
674 | DFSInfoValid = false; | |||
675 | ||||
676 | // Remove node from immediate dominator's children list. | |||
677 | DomTreeNodeBase<NodeT> *IDom = Node->getIDom(); | |||
678 | if (IDom) { | |||
679 | const auto I = find(IDom->Children, Node); | |||
680 | assert(I != IDom->Children.end() &&((void)0) | |||
681 | "Not in immediate dominator children set!")((void)0); | |||
682 | // I am no longer your child... | |||
683 | IDom->Children.erase(I); | |||
684 | } | |||
685 | ||||
686 | DomTreeNodes.erase(BB); | |||
687 | ||||
688 | if (!IsPostDom) return; | |||
689 | ||||
690 | // Remember to update PostDominatorTree roots. | |||
691 | auto RIt = llvm::find(Roots, BB); | |||
692 | if (RIt != Roots.end()) { | |||
693 | std::swap(*RIt, Roots.back()); | |||
694 | Roots.pop_back(); | |||
695 | } | |||
696 | } | |||
697 | ||||
698 | /// splitBlock - BB is split and now it has one successor. Update dominator | |||
699 | /// tree to reflect this change. | |||
700 | void splitBlock(NodeT *NewBB) { | |||
701 | if (IsPostDominator) | |||
702 | Split<Inverse<NodeT *>>(NewBB); | |||
703 | else | |||
704 | Split<NodeT *>(NewBB); | |||
705 | } | |||
706 | ||||
707 | /// print - Convert to human readable form | |||
708 | /// | |||
709 | void print(raw_ostream &O) const { | |||
710 | O << "=============================--------------------------------\n"; | |||
711 | if (IsPostDominator) | |||
712 | O << "Inorder PostDominator Tree: "; | |||
713 | else | |||
714 | O << "Inorder Dominator Tree: "; | |||
715 | if (!DFSInfoValid) | |||
716 | O << "DFSNumbers invalid: " << SlowQueries << " slow queries."; | |||
717 | O << "\n"; | |||
718 | ||||
719 | // The postdom tree can have a null root if there are no returns. | |||
720 | if (getRootNode()) PrintDomTree<NodeT>(getRootNode(), O, 1); | |||
721 | O << "Roots: "; | |||
722 | for (const NodePtr Block : Roots) { | |||
723 | Block->printAsOperand(O, false); | |||
724 | O << " "; | |||
725 | } | |||
726 | O << "\n"; | |||
727 | } | |||
728 | ||||
729 | public: | |||
730 | /// updateDFSNumbers - Assign In and Out numbers to the nodes while walking | |||
731 | /// dominator tree in dfs order. | |||
732 | void updateDFSNumbers() const { | |||
733 | if (DFSInfoValid) { | |||
734 | SlowQueries = 0; | |||
735 | return; | |||
736 | } | |||
737 | ||||
738 | SmallVector<std::pair<const DomTreeNodeBase<NodeT> *, | |||
739 | typename DomTreeNodeBase<NodeT>::const_iterator>, | |||
740 | 32> WorkStack; | |||
741 | ||||
742 | const DomTreeNodeBase<NodeT> *ThisRoot = getRootNode(); | |||
743 | assert((!Parent || ThisRoot) && "Empty constructed DomTree")((void)0); | |||
744 | if (!ThisRoot) | |||
745 | return; | |||
746 | ||||
747 | // Both dominators and postdominators have a single root node. In the case | |||
748 | // case of PostDominatorTree, this node is a virtual root. | |||
749 | WorkStack.push_back({ThisRoot, ThisRoot->begin()}); | |||
750 | ||||
751 | unsigned DFSNum = 0; | |||
752 | ThisRoot->DFSNumIn = DFSNum++; | |||
753 | ||||
754 | while (!WorkStack.empty()) { | |||
755 | const DomTreeNodeBase<NodeT> *Node = WorkStack.back().first; | |||
756 | const auto ChildIt = WorkStack.back().second; | |||
757 | ||||
758 | // If we visited all of the children of this node, "recurse" back up the | |||
759 | // stack setting the DFOutNum. | |||
760 | if (ChildIt == Node->end()) { | |||
761 | Node->DFSNumOut = DFSNum++; | |||
762 | WorkStack.pop_back(); | |||
763 | } else { | |||
764 | // Otherwise, recursively visit this child. | |||
765 | const DomTreeNodeBase<NodeT> *Child = *ChildIt; | |||
766 | ++WorkStack.back().second; | |||
767 | ||||
768 | WorkStack.push_back({Child, Child->begin()}); | |||
769 | Child->DFSNumIn = DFSNum++; | |||
770 | } | |||
771 | } | |||
772 | ||||
773 | SlowQueries = 0; | |||
774 | DFSInfoValid = true; | |||
775 | } | |||
776 | ||||
777 | /// recalculate - compute a dominator tree for the given function | |||
778 | void recalculate(ParentType &Func) { | |||
779 | Parent = &Func; | |||
780 | DomTreeBuilder::Calculate(*this); | |||
781 | } | |||
782 | ||||
783 | void recalculate(ParentType &Func, ArrayRef<UpdateType> Updates) { | |||
784 | Parent = &Func; | |||
785 | DomTreeBuilder::CalculateWithUpdates(*this, Updates); | |||
786 | } | |||
787 | ||||
788 | /// verify - checks if the tree is correct. There are 3 level of verification: | |||
789 | /// - Full -- verifies if the tree is correct by making sure all the | |||
790 | /// properties (including the parent and the sibling property) | |||
791 | /// hold. | |||
792 | /// Takes O(N^3) time. | |||
793 | /// | |||
794 | /// - Basic -- checks if the tree is correct, but compares it to a freshly | |||
795 | /// constructed tree instead of checking the sibling property. | |||
796 | /// Takes O(N^2) time. | |||
797 | /// | |||
798 | /// - Fast -- checks basic tree structure and compares it with a freshly | |||
799 | /// constructed tree. | |||
800 | /// Takes O(N^2) time worst case, but is faster in practise (same | |||
801 | /// as tree construction). | |||
802 | bool verify(VerificationLevel VL = VerificationLevel::Full) const { | |||
803 | return DomTreeBuilder::Verify(*this, VL); | |||
804 | } | |||
805 | ||||
806 | void reset() { | |||
807 | DomTreeNodes.clear(); | |||
808 | Roots.clear(); | |||
809 | RootNode = nullptr; | |||
810 | Parent = nullptr; | |||
811 | DFSInfoValid = false; | |||
812 | SlowQueries = 0; | |||
813 | } | |||
814 | ||||
815 | protected: | |||
816 | void addRoot(NodeT *BB) { this->Roots.push_back(BB); } | |||
817 | ||||
818 | DomTreeNodeBase<NodeT> *createChild(NodeT *BB, DomTreeNodeBase<NodeT> *IDom) { | |||
819 | return (DomTreeNodes[BB] = IDom->addChild( | |||
820 | std::make_unique<DomTreeNodeBase<NodeT>>(BB, IDom))) | |||
821 | .get(); | |||
822 | } | |||
823 | ||||
824 | DomTreeNodeBase<NodeT> *createNode(NodeT *BB) { | |||
825 | return (DomTreeNodes[BB] = | |||
826 | std::make_unique<DomTreeNodeBase<NodeT>>(BB, nullptr)) | |||
827 | .get(); | |||
828 | } | |||
829 | ||||
830 | // NewBB is split and now it has one successor. Update dominator tree to | |||
831 | // reflect this change. | |||
832 | template <class N> | |||
833 | void Split(typename GraphTraits<N>::NodeRef NewBB) { | |||
834 | using GraphT = GraphTraits<N>; | |||
835 | using NodeRef = typename GraphT::NodeRef; | |||
836 | assert(std::distance(GraphT::child_begin(NewBB),((void)0) | |||
837 | GraphT::child_end(NewBB)) == 1 &&((void)0) | |||
838 | "NewBB should have a single successor!")((void)0); | |||
839 | NodeRef NewBBSucc = *GraphT::child_begin(NewBB); | |||
840 | ||||
841 | SmallVector<NodeRef, 4> PredBlocks(children<Inverse<N>>(NewBB)); | |||
842 | ||||
843 | assert(!PredBlocks.empty() && "No predblocks?")((void)0); | |||
844 | ||||
845 | bool NewBBDominatesNewBBSucc = true; | |||
846 | for (auto Pred : children<Inverse<N>>(NewBBSucc)) { | |||
847 | if (Pred != NewBB && !dominates(NewBBSucc, Pred) && | |||
848 | isReachableFromEntry(Pred)) { | |||
849 | NewBBDominatesNewBBSucc = false; | |||
850 | break; | |||
851 | } | |||
852 | } | |||
853 | ||||
854 | // Find NewBB's immediate dominator and create new dominator tree node for | |||
855 | // NewBB. | |||
856 | NodeT *NewBBIDom = nullptr; | |||
857 | unsigned i = 0; | |||
858 | for (i = 0; i < PredBlocks.size(); ++i) | |||
859 | if (isReachableFromEntry(PredBlocks[i])) { | |||
860 | NewBBIDom = PredBlocks[i]; | |||
861 | break; | |||
862 | } | |||
863 | ||||
864 | // It's possible that none of the predecessors of NewBB are reachable; | |||
865 | // in that case, NewBB itself is unreachable, so nothing needs to be | |||
866 | // changed. | |||
867 | if (!NewBBIDom) return; | |||
868 | ||||
869 | for (i = i + 1; i < PredBlocks.size(); ++i) { | |||
870 | if (isReachableFromEntry(PredBlocks[i])) | |||
871 | NewBBIDom = findNearestCommonDominator(NewBBIDom, PredBlocks[i]); | |||
872 | } | |||
873 | ||||
874 | // Create the new dominator tree node... and set the idom of NewBB. | |||
875 | DomTreeNodeBase<NodeT> *NewBBNode = addNewBlock(NewBB, NewBBIDom); | |||
876 | ||||
877 | // If NewBB strictly dominates other blocks, then it is now the immediate | |||
878 | // dominator of NewBBSucc. Update the dominator tree as appropriate. | |||
879 | if (NewBBDominatesNewBBSucc) { | |||
880 | DomTreeNodeBase<NodeT> *NewBBSuccNode = getNode(NewBBSucc); | |||
881 | changeImmediateDominator(NewBBSuccNode, NewBBNode); | |||
882 | } | |||
883 | } | |||
884 | ||||
885 | private: | |||
886 | bool dominatedBySlowTreeWalk(const DomTreeNodeBase<NodeT> *A, | |||
887 | const DomTreeNodeBase<NodeT> *B) const { | |||
888 | assert(A != B)((void)0); | |||
889 | assert(isReachableFromEntry(B))((void)0); | |||
890 | assert(isReachableFromEntry(A))((void)0); | |||
891 | ||||
892 | const unsigned ALevel = A->getLevel(); | |||
893 | const DomTreeNodeBase<NodeT> *IDom; | |||
894 | ||||
895 | // Don't walk nodes above A's subtree. When we reach A's level, we must | |||
896 | // either find A or be in some other subtree not dominated by A. | |||
897 | while ((IDom = B->getIDom()) != nullptr && IDom->getLevel() >= ALevel) | |||
898 | B = IDom; // Walk up the tree | |||
899 | ||||
900 | return B == A; | |||
901 | } | |||
902 | ||||
903 | /// Wipe this tree's state without releasing any resources. | |||
904 | /// | |||
905 | /// This is essentially a post-move helper only. It leaves the object in an | |||
906 | /// assignable and destroyable state, but otherwise invalid. | |||
907 | void wipe() { | |||
908 | DomTreeNodes.clear(); | |||
909 | RootNode = nullptr; | |||
910 | Parent = nullptr; | |||
911 | } | |||
912 | }; | |||
913 | ||||
914 | template <typename T> | |||
915 | using DomTreeBase = DominatorTreeBase<T, false>; | |||
916 | ||||
917 | template <typename T> | |||
918 | using PostDomTreeBase = DominatorTreeBase<T, true>; | |||
919 | ||||
920 | // These two functions are declared out of line as a workaround for building | |||
921 | // with old (< r147295) versions of clang because of pr11642. | |||
922 | template <typename NodeT, bool IsPostDom> | |||
923 | bool DominatorTreeBase<NodeT, IsPostDom>::dominates(const NodeT *A, | |||
924 | const NodeT *B) const { | |||
925 | if (A == B) | |||
926 | return true; | |||
927 | ||||
928 | // Cast away the const qualifiers here. This is ok since | |||
929 | // this function doesn't actually return the values returned | |||
930 | // from getNode. | |||
931 | return dominates(getNode(const_cast<NodeT *>(A)), | |||
932 | getNode(const_cast<NodeT *>(B))); | |||
933 | } | |||
934 | template <typename NodeT, bool IsPostDom> | |||
935 | bool DominatorTreeBase<NodeT, IsPostDom>::properlyDominates( | |||
936 | const NodeT *A, const NodeT *B) const { | |||
937 | if (A == B) | |||
938 | return false; | |||
939 | ||||
940 | // Cast away the const qualifiers here. This is ok since | |||
941 | // this function doesn't actually return the values returned | |||
942 | // from getNode. | |||
943 | return dominates(getNode(const_cast<NodeT *>(A)), | |||
944 | getNode(const_cast<NodeT *>(B))); | |||
945 | } | |||
946 | ||||
947 | } // end namespace llvm | |||
948 | ||||
949 | #endif // LLVM_SUPPORT_GENERICDOMTREE_H |
1 | //===- llvm/ADT/DenseMap.h - Dense probed hash table ------------*- C++ -*-===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file defines the DenseMap class. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #ifndef LLVM_ADT_DENSEMAP_H |
14 | #define LLVM_ADT_DENSEMAP_H |
15 | |
16 | #include "llvm/ADT/DenseMapInfo.h" |
17 | #include "llvm/ADT/EpochTracker.h" |
18 | #include "llvm/Support/AlignOf.h" |
19 | #include "llvm/Support/Compiler.h" |
20 | #include "llvm/Support/MathExtras.h" |
21 | #include "llvm/Support/MemAlloc.h" |
22 | #include "llvm/Support/ReverseIteration.h" |
23 | #include "llvm/Support/type_traits.h" |
24 | #include <algorithm> |
25 | #include <cassert> |
26 | #include <cstddef> |
27 | #include <cstring> |
28 | #include <initializer_list> |
29 | #include <iterator> |
30 | #include <new> |
31 | #include <type_traits> |
32 | #include <utility> |
33 | |
34 | namespace llvm { |
35 | |
36 | namespace detail { |
37 | |
38 | // We extend a pair to allow users to override the bucket type with their own |
39 | // implementation without requiring two members. |
40 | template <typename KeyT, typename ValueT> |
41 | struct DenseMapPair : public std::pair<KeyT, ValueT> { |
42 | using std::pair<KeyT, ValueT>::pair; |
43 | |
44 | KeyT &getFirst() { return std::pair<KeyT, ValueT>::first; } |
45 | const KeyT &getFirst() const { return std::pair<KeyT, ValueT>::first; } |
46 | ValueT &getSecond() { return std::pair<KeyT, ValueT>::second; } |
47 | const ValueT &getSecond() const { return std::pair<KeyT, ValueT>::second; } |
48 | }; |
49 | |
50 | } // end namespace detail |
51 | |
52 | template <typename KeyT, typename ValueT, |
53 | typename KeyInfoT = DenseMapInfo<KeyT>, |
54 | typename Bucket = llvm::detail::DenseMapPair<KeyT, ValueT>, |
55 | bool IsConst = false> |
56 | class DenseMapIterator; |
57 | |
58 | template <typename DerivedT, typename KeyT, typename ValueT, typename KeyInfoT, |
59 | typename BucketT> |
60 | class DenseMapBase : public DebugEpochBase { |
61 | template <typename T> |
62 | using const_arg_type_t = typename const_pointer_or_const_ref<T>::type; |
63 | |
64 | public: |
65 | using size_type = unsigned; |
66 | using key_type = KeyT; |
67 | using mapped_type = ValueT; |
68 | using value_type = BucketT; |
69 | |
70 | using iterator = DenseMapIterator<KeyT, ValueT, KeyInfoT, BucketT>; |
71 | using const_iterator = |
72 | DenseMapIterator<KeyT, ValueT, KeyInfoT, BucketT, true>; |
73 | |
74 | inline iterator begin() { |
75 | // When the map is empty, avoid the overhead of advancing/retreating past |
76 | // empty buckets. |
77 | if (empty()) |
78 | return end(); |
79 | if (shouldReverseIterate<KeyT>()) |
80 | return makeIterator(getBucketsEnd() - 1, getBuckets(), *this); |
81 | return makeIterator(getBuckets(), getBucketsEnd(), *this); |
82 | } |
83 | inline iterator end() { |
84 | return makeIterator(getBucketsEnd(), getBucketsEnd(), *this, true); |
85 | } |
86 | inline const_iterator begin() const { |
87 | if (empty()) |
88 | return end(); |
89 | if (shouldReverseIterate<KeyT>()) |
90 | return makeConstIterator(getBucketsEnd() - 1, getBuckets(), *this); |
91 | return makeConstIterator(getBuckets(), getBucketsEnd(), *this); |
92 | } |
93 | inline const_iterator end() const { |
94 | return makeConstIterator(getBucketsEnd(), getBucketsEnd(), *this, true); |
95 | } |
96 | |
97 | LLVM_NODISCARD[[clang::warn_unused_result]] bool empty() const { |
98 | return getNumEntries() == 0; |
99 | } |
100 | unsigned size() const { return getNumEntries(); } |
101 | |
102 | /// Grow the densemap so that it can contain at least \p NumEntries items |
103 | /// before resizing again. |
104 | void reserve(size_type NumEntries) { |
105 | auto NumBuckets = getMinBucketToReserveForEntries(NumEntries); |
106 | incrementEpoch(); |
107 | if (NumBuckets > getNumBuckets()) |
108 | grow(NumBuckets); |
109 | } |
110 | |
111 | void clear() { |
112 | incrementEpoch(); |
113 | if (getNumEntries() == 0 && getNumTombstones() == 0) return; |
114 | |
115 | // If the capacity of the array is huge, and the # elements used is small, |
116 | // shrink the array. |
117 | if (getNumEntries() * 4 < getNumBuckets() && getNumBuckets() > 64) { |
118 | shrink_and_clear(); |
119 | return; |
120 | } |
121 | |
122 | const KeyT EmptyKey = getEmptyKey(), TombstoneKey = getTombstoneKey(); |
123 | if (std::is_trivially_destructible<ValueT>::value) { |
124 | // Use a simpler loop when values don't need destruction. |
125 | for (BucketT *P = getBuckets(), *E = getBucketsEnd(); P != E; ++P) |
126 | P->getFirst() = EmptyKey; |
127 | } else { |
128 | unsigned NumEntries = getNumEntries(); |
129 | for (BucketT *P = getBuckets(), *E = getBucketsEnd(); P != E; ++P) { |
130 | if (!KeyInfoT::isEqual(P->getFirst(), EmptyKey)) { |
131 | if (!KeyInfoT::isEqual(P->getFirst(), TombstoneKey)) { |
132 | P->getSecond().~ValueT(); |
133 | --NumEntries; |
134 | } |
135 | P->getFirst() = EmptyKey; |
136 | } |
137 | } |
138 | assert(NumEntries == 0 && "Node count imbalance!")((void)0); |
139 | } |
140 | setNumEntries(0); |
141 | setNumTombstones(0); |
142 | } |
143 | |
144 | /// Return 1 if the specified key is in the map, 0 otherwise. |
145 | size_type count(const_arg_type_t<KeyT> Val) const { |
146 | const BucketT *TheBucket; |
147 | return LookupBucketFor(Val, TheBucket) ? 1 : 0; |
148 | } |
149 | |
150 | iterator find(const_arg_type_t<KeyT> Val) { |
151 | BucketT *TheBucket; |
152 | if (LookupBucketFor(Val, TheBucket)) |
153 | return makeIterator(TheBucket, |
154 | shouldReverseIterate<KeyT>() ? getBuckets() |
155 | : getBucketsEnd(), |
156 | *this, true); |
157 | return end(); |
158 | } |
159 | const_iterator find(const_arg_type_t<KeyT> Val) const { |
160 | const BucketT *TheBucket; |
161 | if (LookupBucketFor(Val, TheBucket)) |
162 | return makeConstIterator(TheBucket, |
163 | shouldReverseIterate<KeyT>() ? getBuckets() |
164 | : getBucketsEnd(), |
165 | *this, true); |
166 | return end(); |
167 | } |
168 | |
169 | /// Alternate version of find() which allows a different, and possibly |
170 | /// less expensive, key type. |
171 | /// The DenseMapInfo is responsible for supplying methods |
172 | /// getHashValue(LookupKeyT) and isEqual(LookupKeyT, KeyT) for each key |
173 | /// type used. |
174 | template<class LookupKeyT> |
175 | iterator find_as(const LookupKeyT &Val) { |
176 | BucketT *TheBucket; |
177 | if (LookupBucketFor(Val, TheBucket)) |
178 | return makeIterator(TheBucket, |
179 | shouldReverseIterate<KeyT>() ? getBuckets() |
180 | : getBucketsEnd(), |
181 | *this, true); |
182 | return end(); |
183 | } |
184 | template<class LookupKeyT> |
185 | const_iterator find_as(const LookupKeyT &Val) const { |
186 | const BucketT *TheBucket; |
187 | if (LookupBucketFor(Val, TheBucket)) |
188 | return makeConstIterator(TheBucket, |
189 | shouldReverseIterate<KeyT>() ? getBuckets() |
190 | : getBucketsEnd(), |
191 | *this, true); |
192 | return end(); |
193 | } |
194 | |
195 | /// lookup - Return the entry for the specified key, or a default |
196 | /// constructed value if no such entry exists. |
197 | ValueT lookup(const_arg_type_t<KeyT> Val) const { |
198 | const BucketT *TheBucket; |
199 | if (LookupBucketFor(Val, TheBucket)) |
200 | return TheBucket->getSecond(); |
201 | return ValueT(); |
202 | } |
203 | |
204 | // Inserts key,value pair into the map if the key isn't already in the map. |
205 | // If the key is already in the map, it returns false and doesn't update the |
206 | // value. |
207 | std::pair<iterator, bool> insert(const std::pair<KeyT, ValueT> &KV) { |
208 | return try_emplace(KV.first, KV.second); |
209 | } |
210 | |
211 | // Inserts key,value pair into the map if the key isn't already in the map. |
212 | // If the key is already in the map, it returns false and doesn't update the |
213 | // value. |
214 | std::pair<iterator, bool> insert(std::pair<KeyT, ValueT> &&KV) { |
215 | return try_emplace(std::move(KV.first), std::move(KV.second)); |
216 | } |
217 | |
218 | // Inserts key,value pair into the map if the key isn't already in the map. |
219 | // The value is constructed in-place if the key is not in the map, otherwise |
220 | // it is not moved. |
221 | template <typename... Ts> |
222 | std::pair<iterator, bool> try_emplace(KeyT &&Key, Ts &&... Args) { |
223 | BucketT *TheBucket; |
224 | if (LookupBucketFor(Key, TheBucket)) |
225 | return std::make_pair(makeIterator(TheBucket, |
226 | shouldReverseIterate<KeyT>() |
227 | ? getBuckets() |
228 | : getBucketsEnd(), |
229 | *this, true), |
230 | false); // Already in map. |
231 | |
232 | // Otherwise, insert the new element. |
233 | TheBucket = |
234 | InsertIntoBucket(TheBucket, std::move(Key), std::forward<Ts>(Args)...); |
235 | return std::make_pair(makeIterator(TheBucket, |
236 | shouldReverseIterate<KeyT>() |
237 | ? getBuckets() |
238 | : getBucketsEnd(), |
239 | *this, true), |
240 | true); |
241 | } |
242 | |
243 | // Inserts key,value pair into the map if the key isn't already in the map. |
244 | // The value is constructed in-place if the key is not in the map, otherwise |
245 | // it is not moved. |
246 | template <typename... Ts> |
247 | std::pair<iterator, bool> try_emplace(const KeyT &Key, Ts &&... Args) { |
248 | BucketT *TheBucket; |
249 | if (LookupBucketFor(Key, TheBucket)) |
250 | return std::make_pair(makeIterator(TheBucket, |
251 | shouldReverseIterate<KeyT>() |
252 | ? getBuckets() |
253 | : getBucketsEnd(), |
254 | *this, true), |
255 | false); // Already in map. |
256 | |
257 | // Otherwise, insert the new element. |
258 | TheBucket = InsertIntoBucket(TheBucket, Key, std::forward<Ts>(Args)...); |
259 | return std::make_pair(makeIterator(TheBucket, |
260 | shouldReverseIterate<KeyT>() |
261 | ? getBuckets() |
262 | : getBucketsEnd(), |
263 | *this, true), |
264 | true); |
265 | } |
266 | |
267 | /// Alternate version of insert() which allows a different, and possibly |
268 | /// less expensive, key type. |
269 | /// The DenseMapInfo is responsible for supplying methods |
270 | /// getHashValue(LookupKeyT) and isEqual(LookupKeyT, KeyT) for each key |
271 | /// type used. |
272 | template <typename LookupKeyT> |
273 | std::pair<iterator, bool> insert_as(std::pair<KeyT, ValueT> &&KV, |
274 | const LookupKeyT &Val) { |
275 | BucketT *TheBucket; |
276 | if (LookupBucketFor(Val, TheBucket)) |
277 | return std::make_pair(makeIterator(TheBucket, |
278 | shouldReverseIterate<KeyT>() |
279 | ? getBuckets() |
280 | : getBucketsEnd(), |
281 | *this, true), |
282 | false); // Already in map. |
283 | |
284 | // Otherwise, insert the new element. |
285 | TheBucket = InsertIntoBucketWithLookup(TheBucket, std::move(KV.first), |
286 | std::move(KV.second), Val); |
287 | return std::make_pair(makeIterator(TheBucket, |
288 | shouldReverseIterate<KeyT>() |
289 | ? getBuckets() |
290 | : getBucketsEnd(), |
291 | *this, true), |
292 | true); |
293 | } |
294 | |
295 | /// insert - Range insertion of pairs. |
296 | template<typename InputIt> |
297 | void insert(InputIt I, InputIt E) { |
298 | for (; I != E; ++I) |
299 | insert(*I); |
300 | } |
301 | |
302 | bool erase(const KeyT &Val) { |
303 | BucketT *TheBucket; |
304 | if (!LookupBucketFor(Val, TheBucket)) |
305 | return false; // not in map. |
306 | |
307 | TheBucket->getSecond().~ValueT(); |
308 | TheBucket->getFirst() = getTombstoneKey(); |
309 | decrementNumEntries(); |
310 | incrementNumTombstones(); |
311 | return true; |
312 | } |
313 | void erase(iterator I) { |
314 | BucketT *TheBucket = &*I; |
315 | TheBucket->getSecond().~ValueT(); |
316 | TheBucket->getFirst() = getTombstoneKey(); |
317 | decrementNumEntries(); |
318 | incrementNumTombstones(); |
319 | } |
320 | |
321 | value_type& FindAndConstruct(const KeyT &Key) { |
322 | BucketT *TheBucket; |
323 | if (LookupBucketFor(Key, TheBucket)) |
324 | return *TheBucket; |
325 | |
326 | return *InsertIntoBucket(TheBucket, Key); |
327 | } |
328 | |
329 | ValueT &operator[](const KeyT &Key) { |
330 | return FindAndConstruct(Key).second; |
331 | } |
332 | |
333 | value_type& FindAndConstruct(KeyT &&Key) { |
334 | BucketT *TheBucket; |
335 | if (LookupBucketFor(Key, TheBucket)) |
336 | return *TheBucket; |
337 | |
338 | return *InsertIntoBucket(TheBucket, std::move(Key)); |
339 | } |
340 | |
341 | ValueT &operator[](KeyT &&Key) { |
342 | return FindAndConstruct(std::move(Key)).second; |
343 | } |
344 | |
345 | /// isPointerIntoBucketsArray - Return true if the specified pointer points |
346 | /// somewhere into the DenseMap's array of buckets (i.e. either to a key or |
347 | /// value in the DenseMap). |
348 | bool isPointerIntoBucketsArray(const void *Ptr) const { |
349 | return Ptr >= getBuckets() && Ptr < getBucketsEnd(); |
350 | } |
351 | |
352 | /// getPointerIntoBucketsArray() - Return an opaque pointer into the buckets |
353 | /// array. In conjunction with the previous method, this can be used to |
354 | /// determine whether an insertion caused the DenseMap to reallocate. |
355 | const void *getPointerIntoBucketsArray() const { return getBuckets(); } |
356 | |
357 | protected: |
358 | DenseMapBase() = default; |
359 | |
360 | void destroyAll() { |
361 | if (getNumBuckets() == 0) // Nothing to do. |
362 | return; |
363 | |
364 | const KeyT EmptyKey = getEmptyKey(), TombstoneKey = getTombstoneKey(); |
365 | for (BucketT *P = getBuckets(), *E = getBucketsEnd(); P != E; ++P) { |
366 | if (!KeyInfoT::isEqual(P->getFirst(), EmptyKey) && |
367 | !KeyInfoT::isEqual(P->getFirst(), TombstoneKey)) |
368 | P->getSecond().~ValueT(); |
369 | P->getFirst().~KeyT(); |
370 | } |
371 | } |
372 | |
373 | void initEmpty() { |
374 | setNumEntries(0); |
375 | setNumTombstones(0); |
376 | |
377 | assert((getNumBuckets() & (getNumBuckets()-1)) == 0 &&((void)0) |
378 | "# initial buckets must be a power of two!")((void)0); |
379 | const KeyT EmptyKey = getEmptyKey(); |
380 | for (BucketT *B = getBuckets(), *E = getBucketsEnd(); B != E; ++B) |
381 | ::new (&B->getFirst()) KeyT(EmptyKey); |
382 | } |
383 | |
384 | /// Returns the number of buckets to allocate to ensure that the DenseMap can |
385 | /// accommodate \p NumEntries without need to grow(). |
386 | unsigned getMinBucketToReserveForEntries(unsigned NumEntries) { |
387 | // Ensure that "NumEntries * 4 < NumBuckets * 3" |
388 | if (NumEntries == 0) |
389 | return 0; |
390 | // +1 is required because of the strict equality. |
391 | // For example if NumEntries is 48, we need to return 401. |
392 | return NextPowerOf2(NumEntries * 4 / 3 + 1); |
393 | } |
394 | |
395 | void moveFromOldBuckets(BucketT *OldBucketsBegin, BucketT *OldBucketsEnd) { |
396 | initEmpty(); |
397 | |
398 | // Insert all the old elements. |
399 | const KeyT EmptyKey = getEmptyKey(); |
400 | const KeyT TombstoneKey = getTombstoneKey(); |
401 | for (BucketT *B = OldBucketsBegin, *E = OldBucketsEnd; B != E; ++B) { |
402 | if (!KeyInfoT::isEqual(B->getFirst(), EmptyKey) && |
403 | !KeyInfoT::isEqual(B->getFirst(), TombstoneKey)) { |
404 | // Insert the key/value into the new table. |
405 | BucketT *DestBucket; |
406 | bool FoundVal = LookupBucketFor(B->getFirst(), DestBucket); |
407 | (void)FoundVal; // silence warning. |
408 | assert(!FoundVal && "Key already in new map?")((void)0); |
409 | DestBucket->getFirst() = std::move(B->getFirst()); |
410 | ::new (&DestBucket->getSecond()) ValueT(std::move(B->getSecond())); |
411 | incrementNumEntries(); |
412 | |
413 | // Free the value. |
414 | B->getSecond().~ValueT(); |
415 | } |
416 | B->getFirst().~KeyT(); |
417 | } |
418 | } |
419 | |
420 | template <typename OtherBaseT> |
421 | void copyFrom( |
422 | const DenseMapBase<OtherBaseT, KeyT, ValueT, KeyInfoT, BucketT> &other) { |
423 | assert(&other != this)((void)0); |
424 | assert(getNumBuckets() == other.getNumBuckets())((void)0); |
425 | |
426 | setNumEntries(other.getNumEntries()); |
427 | setNumTombstones(other.getNumTombstones()); |
428 | |
429 | if (std::is_trivially_copyable<KeyT>::value && |
430 | std::is_trivially_copyable<ValueT>::value) |
431 | memcpy(reinterpret_cast<void *>(getBuckets()), other.getBuckets(), |
432 | getNumBuckets() * sizeof(BucketT)); |
433 | else |
434 | for (size_t i = 0; i < getNumBuckets(); ++i) { |
435 | ::new (&getBuckets()[i].getFirst()) |
436 | KeyT(other.getBuckets()[i].getFirst()); |
437 | if (!KeyInfoT::isEqual(getBuckets()[i].getFirst(), getEmptyKey()) && |
438 | !KeyInfoT::isEqual(getBuckets()[i].getFirst(), getTombstoneKey())) |
439 | ::new (&getBuckets()[i].getSecond()) |
440 | ValueT(other.getBuckets()[i].getSecond()); |
441 | } |
442 | } |
443 | |
444 | static unsigned getHashValue(const KeyT &Val) { |
445 | return KeyInfoT::getHashValue(Val); |
446 | } |
447 | |
448 | template<typename LookupKeyT> |
449 | static unsigned getHashValue(const LookupKeyT &Val) { |
450 | return KeyInfoT::getHashValue(Val); |
451 | } |
452 | |
453 | static const KeyT getEmptyKey() { |
454 | static_assert(std::is_base_of<DenseMapBase, DerivedT>::value, |
455 | "Must pass the derived type to this template!"); |
456 | return KeyInfoT::getEmptyKey(); |
457 | } |
458 | |
459 | static const KeyT getTombstoneKey() { |
460 | return KeyInfoT::getTombstoneKey(); |
461 | } |
462 | |
463 | private: |
464 | iterator makeIterator(BucketT *P, BucketT *E, |
465 | DebugEpochBase &Epoch, |
466 | bool NoAdvance=false) { |
467 | if (shouldReverseIterate<KeyT>()) { |
468 | BucketT *B = P == getBucketsEnd() ? getBuckets() : P + 1; |
469 | return iterator(B, E, Epoch, NoAdvance); |
470 | } |
471 | return iterator(P, E, Epoch, NoAdvance); |
472 | } |
473 | |
474 | const_iterator makeConstIterator(const BucketT *P, const BucketT *E, |
475 | const DebugEpochBase &Epoch, |
476 | const bool NoAdvance=false) const { |
477 | if (shouldReverseIterate<KeyT>()) { |
478 | const BucketT *B = P == getBucketsEnd() ? getBuckets() : P + 1; |
479 | return const_iterator(B, E, Epoch, NoAdvance); |
480 | } |
481 | return const_iterator(P, E, Epoch, NoAdvance); |
482 | } |
483 | |
484 | unsigned getNumEntries() const { |
485 | return static_cast<const DerivedT *>(this)->getNumEntries(); |
486 | } |
487 | |
488 | void setNumEntries(unsigned Num) { |
489 | static_cast<DerivedT *>(this)->setNumEntries(Num); |
490 | } |
491 | |
492 | void incrementNumEntries() { |
493 | setNumEntries(getNumEntries() + 1); |
494 | } |
495 | |
496 | void decrementNumEntries() { |
497 | setNumEntries(getNumEntries() - 1); |
498 | } |
499 | |
500 | unsigned getNumTombstones() const { |
501 | return static_cast<const DerivedT *>(this)->getNumTombstones(); |
502 | } |
503 | |
504 | void setNumTombstones(unsigned Num) { |
505 | static_cast<DerivedT *>(this)->setNumTombstones(Num); |
506 | } |
507 | |
508 | void incrementNumTombstones() { |
509 | setNumTombstones(getNumTombstones() + 1); |
510 | } |
511 | |
512 | void decrementNumTombstones() { |
513 | setNumTombstones(getNumTombstones() - 1); |
514 | } |
515 | |
516 | const BucketT *getBuckets() const { |
517 | return static_cast<const DerivedT *>(this)->getBuckets(); |
518 | } |
519 | |
520 | BucketT *getBuckets() { |
521 | return static_cast<DerivedT *>(this)->getBuckets(); |
522 | } |
523 | |
524 | unsigned getNumBuckets() const { |
525 | return static_cast<const DerivedT *>(this)->getNumBuckets(); |
526 | } |
527 | |
528 | BucketT *getBucketsEnd() { |
529 | return getBuckets() + getNumBuckets(); |
530 | } |
531 | |
532 | const BucketT *getBucketsEnd() const { |
533 | return getBuckets() + getNumBuckets(); |
534 | } |
535 | |
536 | void grow(unsigned AtLeast) { |
537 | static_cast<DerivedT *>(this)->grow(AtLeast); |
538 | } |
539 | |
540 | void shrink_and_clear() { |
541 | static_cast<DerivedT *>(this)->shrink_and_clear(); |
542 | } |
543 | |
544 | template <typename KeyArg, typename... ValueArgs> |
545 | BucketT *InsertIntoBucket(BucketT *TheBucket, KeyArg &&Key, |
546 | ValueArgs &&... Values) { |
547 | TheBucket = InsertIntoBucketImpl(Key, Key, TheBucket); |
548 | |
549 | TheBucket->getFirst() = std::forward<KeyArg>(Key); |
550 | ::new (&TheBucket->getSecond()) ValueT(std::forward<ValueArgs>(Values)...); |
551 | return TheBucket; |
552 | } |
553 | |
554 | template <typename LookupKeyT> |
555 | BucketT *InsertIntoBucketWithLookup(BucketT *TheBucket, KeyT &&Key, |
556 | ValueT &&Value, LookupKeyT &Lookup) { |
557 | TheBucket = InsertIntoBucketImpl(Key, Lookup, TheBucket); |
558 | |
559 | TheBucket->getFirst() = std::move(Key); |
560 | ::new (&TheBucket->getSecond()) ValueT(std::move(Value)); |
561 | return TheBucket; |
562 | } |
563 | |
564 | template <typename LookupKeyT> |
565 | BucketT *InsertIntoBucketImpl(const KeyT &Key, const LookupKeyT &Lookup, |
566 | BucketT *TheBucket) { |
567 | incrementEpoch(); |
568 | |
569 | // If the load of the hash table is more than 3/4, or if fewer than 1/8 of |
570 | // the buckets are empty (meaning that many are filled with tombstones), |
571 | // grow the table. |
572 | // |
573 | // The later case is tricky. For example, if we had one empty bucket with |
574 | // tons of tombstones, failing lookups (e.g. for insertion) would have to |
575 | // probe almost the entire table until it found the empty bucket. If the |
576 | // table completely filled with tombstones, no lookup would ever succeed, |
577 | // causing infinite loops in lookup. |
578 | unsigned NewNumEntries = getNumEntries() + 1; |
579 | unsigned NumBuckets = getNumBuckets(); |
580 | if (LLVM_UNLIKELY(NewNumEntries * 4 >= NumBuckets * 3)__builtin_expect((bool)(NewNumEntries * 4 >= NumBuckets * 3 ), false)) { |
581 | this->grow(NumBuckets * 2); |
582 | LookupBucketFor(Lookup, TheBucket); |
583 | NumBuckets = getNumBuckets(); |
584 | } else if (LLVM_UNLIKELY(NumBuckets-(NewNumEntries+getNumTombstones()) <=__builtin_expect((bool)(NumBuckets-(NewNumEntries+getNumTombstones ()) <= NumBuckets/8), false) |
585 | NumBuckets/8)__builtin_expect((bool)(NumBuckets-(NewNumEntries+getNumTombstones ()) <= NumBuckets/8), false)) { |
586 | this->grow(NumBuckets); |
587 | LookupBucketFor(Lookup, TheBucket); |
588 | } |
589 | assert(TheBucket)((void)0); |
590 | |
591 | // Only update the state after we've grown our bucket space appropriately |
592 | // so that when growing buckets we have self-consistent entry count. |
593 | incrementNumEntries(); |
594 | |
595 | // If we are writing over a tombstone, remember this. |
596 | const KeyT EmptyKey = getEmptyKey(); |
597 | if (!KeyInfoT::isEqual(TheBucket->getFirst(), EmptyKey)) |
598 | decrementNumTombstones(); |
599 | |
600 | return TheBucket; |
601 | } |
602 | |
603 | /// LookupBucketFor - Lookup the appropriate bucket for Val, returning it in |
604 | /// FoundBucket. If the bucket contains the key and a value, this returns |
605 | /// true, otherwise it returns a bucket with an empty marker or tombstone and |
606 | /// returns false. |
607 | template<typename LookupKeyT> |
608 | bool LookupBucketFor(const LookupKeyT &Val, |
609 | const BucketT *&FoundBucket) const { |
610 | const BucketT *BucketsPtr = getBuckets(); |
611 | const unsigned NumBuckets = getNumBuckets(); |
612 | |
613 | if (NumBuckets == 0) { |
614 | FoundBucket = nullptr; |
615 | return false; |
616 | } |
617 | |
618 | // FoundTombstone - Keep track of whether we find a tombstone while probing. |
619 | const BucketT *FoundTombstone = nullptr; |
620 | const KeyT EmptyKey = getEmptyKey(); |
621 | const KeyT TombstoneKey = getTombstoneKey(); |
622 | assert(!KeyInfoT::isEqual(Val, EmptyKey) &&((void)0) |
623 | !KeyInfoT::isEqual(Val, TombstoneKey) &&((void)0) |
624 | "Empty/Tombstone value shouldn't be inserted into map!")((void)0); |
625 | |
626 | unsigned BucketNo = getHashValue(Val) & (NumBuckets-1); |
627 | unsigned ProbeAmt = 1; |
628 | while (true) { |
629 | const BucketT *ThisBucket = BucketsPtr + BucketNo; |
630 | // Found Val's bucket? If so, return it. |
631 | if (LLVM_LIKELY(KeyInfoT::isEqual(Val, ThisBucket->getFirst()))__builtin_expect((bool)(KeyInfoT::isEqual(Val, ThisBucket-> getFirst())), true)) { |
632 | FoundBucket = ThisBucket; |
633 | return true; |
634 | } |
635 | |
636 | // If we found an empty bucket, the key doesn't exist in the set. |
637 | // Insert it and return the default value. |
638 | if (LLVM_LIKELY(KeyInfoT::isEqual(ThisBucket->getFirst(), EmptyKey))__builtin_expect((bool)(KeyInfoT::isEqual(ThisBucket->getFirst (), EmptyKey)), true)) { |
639 | // If we've already seen a tombstone while probing, fill it in instead |
640 | // of the empty bucket we eventually probed to. |
641 | FoundBucket = FoundTombstone ? FoundTombstone : ThisBucket; |
642 | return false; |
643 | } |
644 | |
645 | // If this is a tombstone, remember it. If Val ends up not in the map, we |
646 | // prefer to return it than something that would require more probing. |
647 | if (KeyInfoT::isEqual(ThisBucket->getFirst(), TombstoneKey) && |
648 | !FoundTombstone) |
649 | FoundTombstone = ThisBucket; // Remember the first tombstone found. |
650 | |
651 | // Otherwise, it's a hash collision or a tombstone, continue quadratic |
652 | // probing. |
653 | BucketNo += ProbeAmt++; |
654 | BucketNo &= (NumBuckets-1); |
655 | } |
656 | } |
657 | |
658 | template <typename LookupKeyT> |
659 | bool LookupBucketFor(const LookupKeyT &Val, BucketT *&FoundBucket) { |
660 | const BucketT *ConstFoundBucket; |
661 | bool Result = const_cast<const DenseMapBase *>(this) |
662 | ->LookupBucketFor(Val, ConstFoundBucket); |
663 | FoundBucket = const_cast<BucketT *>(ConstFoundBucket); |
664 | return Result; |
665 | } |
666 | |
667 | public: |
668 | /// Return the approximate size (in bytes) of the actual map. |
669 | /// This is just the raw memory used by DenseMap. |
670 | /// If entries are pointers to objects, the size of the referenced objects |
671 | /// are not included. |
672 | size_t getMemorySize() const { |
673 | return getNumBuckets() * sizeof(BucketT); |
674 | } |
675 | }; |
676 | |
677 | /// Equality comparison for DenseMap. |
678 | /// |
679 | /// Iterates over elements of LHS confirming that each (key, value) pair in LHS |
680 | /// is also in RHS, and that no additional pairs are in RHS. |
681 | /// Equivalent to N calls to RHS.find and N value comparisons. Amortized |
682 | /// complexity is linear, worst case is O(N^2) (if every hash collides). |
683 | template <typename DerivedT, typename KeyT, typename ValueT, typename KeyInfoT, |
684 | typename BucketT> |
685 | bool operator==( |
686 | const DenseMapBase<DerivedT, KeyT, ValueT, KeyInfoT, BucketT> &LHS, |
687 | const DenseMapBase<DerivedT, KeyT, ValueT, KeyInfoT, BucketT> &RHS) { |
688 | if (LHS.size() != RHS.size()) |
689 | return false; |
690 | |
691 | for (auto &KV : LHS) { |
692 | auto I = RHS.find(KV.first); |
693 | if (I == RHS.end() || I->second != KV.second) |
694 | return false; |
695 | } |
696 | |
697 | return true; |
698 | } |
699 | |
700 | /// Inequality comparison for DenseMap. |
701 | /// |
702 | /// Equivalent to !(LHS == RHS). See operator== for performance notes. |
703 | template <typename DerivedT, typename KeyT, typename ValueT, typename KeyInfoT, |
704 | typename BucketT> |
705 | bool operator!=( |
706 | const DenseMapBase<DerivedT, KeyT, ValueT, KeyInfoT, BucketT> &LHS, |
707 | const DenseMapBase<DerivedT, KeyT, ValueT, KeyInfoT, BucketT> &RHS) { |
708 | return !(LHS == RHS); |
709 | } |
710 | |
711 | template <typename KeyT, typename ValueT, |
712 | typename KeyInfoT = DenseMapInfo<KeyT>, |
713 | typename BucketT = llvm::detail::DenseMapPair<KeyT, ValueT>> |
714 | class DenseMap : public DenseMapBase<DenseMap<KeyT, ValueT, KeyInfoT, BucketT>, |
715 | KeyT, ValueT, KeyInfoT, BucketT> { |
716 | friend class DenseMapBase<DenseMap, KeyT, ValueT, KeyInfoT, BucketT>; |
717 | |
718 | // Lift some types from the dependent base class into this class for |
719 | // simplicity of referring to them. |
720 | using BaseT = DenseMapBase<DenseMap, KeyT, ValueT, KeyInfoT, BucketT>; |
721 | |
722 | BucketT *Buckets; |
723 | unsigned NumEntries; |
724 | unsigned NumTombstones; |
725 | unsigned NumBuckets; |
726 | |
727 | public: |
728 | /// Create a DenseMap with an optional \p InitialReserve that guarantee that |
729 | /// this number of elements can be inserted in the map without grow() |
730 | explicit DenseMap(unsigned InitialReserve = 0) { init(InitialReserve); } |
731 | |
732 | DenseMap(const DenseMap &other) : BaseT() { |
733 | init(0); |
734 | copyFrom(other); |
735 | } |
736 | |
737 | DenseMap(DenseMap &&other) : BaseT() { |
738 | init(0); |
739 | swap(other); |
740 | } |
741 | |
742 | template<typename InputIt> |
743 | DenseMap(const InputIt &I, const InputIt &E) { |
744 | init(std::distance(I, E)); |
745 | this->insert(I, E); |
746 | } |
747 | |
748 | DenseMap(std::initializer_list<typename BaseT::value_type> Vals) { |
749 | init(Vals.size()); |
750 | this->insert(Vals.begin(), Vals.end()); |
751 | } |
752 | |
753 | ~DenseMap() { |
754 | this->destroyAll(); |
755 | deallocate_buffer(Buckets, sizeof(BucketT) * NumBuckets, alignof(BucketT)); |
756 | } |
757 | |
758 | void swap(DenseMap& RHS) { |
759 | this->incrementEpoch(); |
760 | RHS.incrementEpoch(); |
761 | std::swap(Buckets, RHS.Buckets); |
762 | std::swap(NumEntries, RHS.NumEntries); |
763 | std::swap(NumTombstones, RHS.NumTombstones); |
764 | std::swap(NumBuckets, RHS.NumBuckets); |
765 | } |
766 | |
767 | DenseMap& operator=(const DenseMap& other) { |
768 | if (&other != this) |
769 | copyFrom(other); |
770 | return *this; |
771 | } |
772 | |
773 | DenseMap& operator=(DenseMap &&other) { |
774 | this->destroyAll(); |
775 | deallocate_buffer(Buckets, sizeof(BucketT) * NumBuckets, alignof(BucketT)); |
776 | init(0); |
777 | swap(other); |
778 | return *this; |
779 | } |
780 | |
781 | void copyFrom(const DenseMap& other) { |
782 | this->destroyAll(); |
783 | deallocate_buffer(Buckets, sizeof(BucketT) * NumBuckets, alignof(BucketT)); |
784 | if (allocateBuckets(other.NumBuckets)) { |
785 | this->BaseT::copyFrom(other); |
786 | } else { |
787 | NumEntries = 0; |
788 | NumTombstones = 0; |
789 | } |
790 | } |
791 | |
792 | void init(unsigned InitNumEntries) { |
793 | auto InitBuckets = BaseT::getMinBucketToReserveForEntries(InitNumEntries); |
794 | if (allocateBuckets(InitBuckets)) { |
795 | this->BaseT::initEmpty(); |
796 | } else { |
797 | NumEntries = 0; |
798 | NumTombstones = 0; |
799 | } |
800 | } |
801 | |
802 | void grow(unsigned AtLeast) { |
803 | unsigned OldNumBuckets = NumBuckets; |
804 | BucketT *OldBuckets = Buckets; |
805 | |
806 | allocateBuckets(std::max<unsigned>(64, static_cast<unsigned>(NextPowerOf2(AtLeast-1)))); |
807 | assert(Buckets)((void)0); |
808 | if (!OldBuckets) { |
809 | this->BaseT::initEmpty(); |
810 | return; |
811 | } |
812 | |
813 | this->moveFromOldBuckets(OldBuckets, OldBuckets+OldNumBuckets); |
814 | |
815 | // Free the old table. |
816 | deallocate_buffer(OldBuckets, sizeof(BucketT) * OldNumBuckets, |
817 | alignof(BucketT)); |
818 | } |
819 | |
820 | void shrink_and_clear() { |
821 | unsigned OldNumBuckets = NumBuckets; |
822 | unsigned OldNumEntries = NumEntries; |
823 | this->destroyAll(); |
824 | |
825 | // Reduce the number of buckets. |
826 | unsigned NewNumBuckets = 0; |
827 | if (OldNumEntries) |
828 | NewNumBuckets = std::max(64, 1 << (Log2_32_Ceil(OldNumEntries) + 1)); |
829 | if (NewNumBuckets == NumBuckets) { |
830 | this->BaseT::initEmpty(); |
831 | return; |
832 | } |
833 | |
834 | deallocate_buffer(Buckets, sizeof(BucketT) * OldNumBuckets, |
835 | alignof(BucketT)); |
836 | init(NewNumBuckets); |
837 | } |
838 | |
839 | private: |
840 | unsigned getNumEntries() const { |
841 | return NumEntries; |
842 | } |
843 | |
844 | void setNumEntries(unsigned Num) { |
845 | NumEntries = Num; |
846 | } |
847 | |
848 | unsigned getNumTombstones() const { |
849 | return NumTombstones; |
850 | } |
851 | |
852 | void setNumTombstones(unsigned Num) { |
853 | NumTombstones = Num; |
854 | } |
855 | |
856 | BucketT *getBuckets() const { |
857 | return Buckets; |
858 | } |
859 | |
860 | unsigned getNumBuckets() const { |
861 | return NumBuckets; |
862 | } |
863 | |
864 | bool allocateBuckets(unsigned Num) { |
865 | NumBuckets = Num; |
866 | if (NumBuckets == 0) { |
867 | Buckets = nullptr; |
868 | return false; |
869 | } |
870 | |
871 | Buckets = static_cast<BucketT *>( |
872 | allocate_buffer(sizeof(BucketT) * NumBuckets, alignof(BucketT))); |
873 | return true; |
874 | } |
875 | }; |
876 | |
877 | template <typename KeyT, typename ValueT, unsigned InlineBuckets = 4, |
878 | typename KeyInfoT = DenseMapInfo<KeyT>, |
879 | typename BucketT = llvm::detail::DenseMapPair<KeyT, ValueT>> |
880 | class SmallDenseMap |
881 | : public DenseMapBase< |
882 | SmallDenseMap<KeyT, ValueT, InlineBuckets, KeyInfoT, BucketT>, KeyT, |
883 | ValueT, KeyInfoT, BucketT> { |
884 | friend class DenseMapBase<SmallDenseMap, KeyT, ValueT, KeyInfoT, BucketT>; |
885 | |
886 | // Lift some types from the dependent base class into this class for |
887 | // simplicity of referring to them. |
888 | using BaseT = DenseMapBase<SmallDenseMap, KeyT, ValueT, KeyInfoT, BucketT>; |
889 | |
890 | static_assert(isPowerOf2_64(InlineBuckets), |
891 | "InlineBuckets must be a power of 2."); |
892 | |
893 | unsigned Small : 1; |
894 | unsigned NumEntries : 31; |
895 | unsigned NumTombstones; |
896 | |
897 | struct LargeRep { |
898 | BucketT *Buckets; |
899 | unsigned NumBuckets; |
900 | }; |
901 | |
902 | /// A "union" of an inline bucket array and the struct representing |
903 | /// a large bucket. This union will be discriminated by the 'Small' bit. |
904 | AlignedCharArrayUnion<BucketT[InlineBuckets], LargeRep> storage; |
905 | |
906 | public: |
907 | explicit SmallDenseMap(unsigned NumInitBuckets = 0) { |
908 | init(NumInitBuckets); |
909 | } |
910 | |
911 | SmallDenseMap(const SmallDenseMap &other) : BaseT() { |
912 | init(0); |
913 | copyFrom(other); |
914 | } |
915 | |
916 | SmallDenseMap(SmallDenseMap &&other) : BaseT() { |
917 | init(0); |
918 | swap(other); |
919 | } |
920 | |
921 | template<typename InputIt> |
922 | SmallDenseMap(const InputIt &I, const InputIt &E) { |
923 | init(NextPowerOf2(std::distance(I, E))); |
924 | this->insert(I, E); |
925 | } |
926 | |
927 | SmallDenseMap(std::initializer_list<typename BaseT::value_type> Vals) |
928 | : SmallDenseMap(Vals.begin(), Vals.end()) {} |
929 | |
930 | ~SmallDenseMap() { |
931 | this->destroyAll(); |
932 | deallocateBuckets(); |
933 | } |
934 | |
935 | void swap(SmallDenseMap& RHS) { |
936 | unsigned TmpNumEntries = RHS.NumEntries; |
937 | RHS.NumEntries = NumEntries; |
938 | NumEntries = TmpNumEntries; |
939 | std::swap(NumTombstones, RHS.NumTombstones); |
940 | |
941 | const KeyT EmptyKey = this->getEmptyKey(); |
942 | const KeyT TombstoneKey = this->getTombstoneKey(); |
943 | if (Small && RHS.Small) { |
944 | // If we're swapping inline bucket arrays, we have to cope with some of |
945 | // the tricky bits of DenseMap's storage system: the buckets are not |
946 | // fully initialized. Thus we swap every key, but we may have |
947 | // a one-directional move of the value. |
948 | for (unsigned i = 0, e = InlineBuckets; i != e; ++i) { |
949 | BucketT *LHSB = &getInlineBuckets()[i], |
950 | *RHSB = &RHS.getInlineBuckets()[i]; |
951 | bool hasLHSValue = (!KeyInfoT::isEqual(LHSB->getFirst(), EmptyKey) && |
952 | !KeyInfoT::isEqual(LHSB->getFirst(), TombstoneKey)); |
953 | bool hasRHSValue = (!KeyInfoT::isEqual(RHSB->getFirst(), EmptyKey) && |
954 | !KeyInfoT::isEqual(RHSB->getFirst(), TombstoneKey)); |
955 | if (hasLHSValue && hasRHSValue) { |
956 | // Swap together if we can... |
957 | std::swap(*LHSB, *RHSB); |
958 | continue; |
959 | } |
960 | // Swap separately and handle any asymmetry. |
961 | std::swap(LHSB->getFirst(), RHSB->getFirst()); |
962 | if (hasLHSValue) { |
963 | ::new (&RHSB->getSecond()) ValueT(std::move(LHSB->getSecond())); |
964 | LHSB->getSecond().~ValueT(); |
965 | } else if (hasRHSValue) { |
966 | ::new (&LHSB->getSecond()) ValueT(std::move(RHSB->getSecond())); |
967 | RHSB->getSecond().~ValueT(); |
968 | } |
969 | } |
970 | return; |
971 | } |
972 | if (!Small && !RHS.Small) { |
973 | std::swap(getLargeRep()->Buckets, RHS.getLargeRep()->Buckets); |
974 | std::swap(getLargeRep()->NumBuckets, RHS.getLargeRep()->NumBuckets); |
975 | return; |
976 | } |
977 | |
978 | SmallDenseMap &SmallSide = Small ? *this : RHS; |
979 | SmallDenseMap &LargeSide = Small ? RHS : *this; |
980 | |
981 | // First stash the large side's rep and move the small side across. |
982 | LargeRep TmpRep = std::move(*LargeSide.getLargeRep()); |
983 | LargeSide.getLargeRep()->~LargeRep(); |
984 | LargeSide.Small = true; |
985 | // This is similar to the standard move-from-old-buckets, but the bucket |
986 | // count hasn't actually rotated in this case. So we have to carefully |
987 | // move construct the keys and values into their new locations, but there |
988 | // is no need to re-hash things. |
989 | for (unsigned i = 0, e = InlineBuckets; i != e; ++i) { |
990 | BucketT *NewB = &LargeSide.getInlineBuckets()[i], |
991 | *OldB = &SmallSide.getInlineBuckets()[i]; |
992 | ::new (&NewB->getFirst()) KeyT(std::move(OldB->getFirst())); |
993 | OldB->getFirst().~KeyT(); |
994 | if (!KeyInfoT::isEqual(NewB->getFirst(), EmptyKey) && |
995 | !KeyInfoT::isEqual(NewB->getFirst(), TombstoneKey)) { |
996 | ::new (&NewB->getSecond()) ValueT(std::move(OldB->getSecond())); |
997 | OldB->getSecond().~ValueT(); |
998 | } |
999 | } |
1000 | |
1001 | // The hard part of moving the small buckets across is done, just move |
1002 | // the TmpRep into its new home. |
1003 | SmallSide.Small = false; |
1004 | new (SmallSide.getLargeRep()) LargeRep(std::move(TmpRep)); |
1005 | } |
1006 | |
1007 | SmallDenseMap& operator=(const SmallDenseMap& other) { |
1008 | if (&other != this) |
1009 | copyFrom(other); |
1010 | return *this; |
1011 | } |
1012 | |
1013 | SmallDenseMap& operator=(SmallDenseMap &&other) { |
1014 | this->destroyAll(); |
1015 | deallocateBuckets(); |
1016 | init(0); |
1017 | swap(other); |
1018 | return *this; |
1019 | } |
1020 | |
1021 | void copyFrom(const SmallDenseMap& other) { |
1022 | this->destroyAll(); |
1023 | deallocateBuckets(); |
1024 | Small = true; |
1025 | if (other.getNumBuckets() > InlineBuckets) { |
1026 | Small = false; |
1027 | new (getLargeRep()) LargeRep(allocateBuckets(other.getNumBuckets())); |
1028 | } |
1029 | this->BaseT::copyFrom(other); |
1030 | } |
1031 | |
1032 | void init(unsigned InitBuckets) { |
1033 | Small = true; |
1034 | if (InitBuckets > InlineBuckets) { |
1035 | Small = false; |
1036 | new (getLargeRep()) LargeRep(allocateBuckets(InitBuckets)); |
1037 | } |
1038 | this->BaseT::initEmpty(); |
1039 | } |
1040 | |
1041 | void grow(unsigned AtLeast) { |
1042 | if (AtLeast > InlineBuckets) |
1043 | AtLeast = std::max<unsigned>(64, NextPowerOf2(AtLeast-1)); |
1044 | |
1045 | if (Small) { |
1046 | // First move the inline buckets into a temporary storage. |
1047 | AlignedCharArrayUnion<BucketT[InlineBuckets]> TmpStorage; |
1048 | BucketT *TmpBegin = reinterpret_cast<BucketT *>(&TmpStorage); |
1049 | BucketT *TmpEnd = TmpBegin; |
1050 | |
1051 | // Loop over the buckets, moving non-empty, non-tombstones into the |
1052 | // temporary storage. Have the loop move the TmpEnd forward as it goes. |
1053 | const KeyT EmptyKey = this->getEmptyKey(); |
1054 | const KeyT TombstoneKey = this->getTombstoneKey(); |
1055 | for (BucketT *P = getBuckets(), *E = P + InlineBuckets; P != E; ++P) { |
1056 | if (!KeyInfoT::isEqual(P->getFirst(), EmptyKey) && |
1057 | !KeyInfoT::isEqual(P->getFirst(), TombstoneKey)) { |
1058 | assert(size_t(TmpEnd - TmpBegin) < InlineBuckets &&((void)0) |
1059 | "Too many inline buckets!")((void)0); |
1060 | ::new (&TmpEnd->getFirst()) KeyT(std::move(P->getFirst())); |
1061 | ::new (&TmpEnd->getSecond()) ValueT(std::move(P->getSecond())); |
1062 | ++TmpEnd; |
1063 | P->getSecond().~ValueT(); |
1064 | } |
1065 | P->getFirst().~KeyT(); |
1066 | } |
1067 | |
1068 | // AtLeast == InlineBuckets can happen if there are many tombstones, |
1069 | // and grow() is used to remove them. Usually we always switch to the |
1070 | // large rep here. |
1071 | if (AtLeast > InlineBuckets) { |
1072 | Small = false; |
1073 | new (getLargeRep()) LargeRep(allocateBuckets(AtLeast)); |
1074 | } |
1075 | this->moveFromOldBuckets(TmpBegin, TmpEnd); |
1076 | return; |
1077 | } |
1078 | |
1079 | LargeRep OldRep = std::move(*getLargeRep()); |
1080 | getLargeRep()->~LargeRep(); |
1081 | if (AtLeast <= InlineBuckets) { |
1082 | Small = true; |
1083 | } else { |
1084 | new (getLargeRep()) LargeRep(allocateBuckets(AtLeast)); |
1085 | } |
1086 | |
1087 | this->moveFromOldBuckets(OldRep.Buckets, OldRep.Buckets+OldRep.NumBuckets); |
1088 | |
1089 | // Free the old table. |
1090 | deallocate_buffer(OldRep.Buckets, sizeof(BucketT) * OldRep.NumBuckets, |
1091 | alignof(BucketT)); |
1092 | } |
1093 | |
1094 | void shrink_and_clear() { |
1095 | unsigned OldSize = this->size(); |
1096 | this->destroyAll(); |
1097 | |
1098 | // Reduce the number of buckets. |
1099 | unsigned NewNumBuckets = 0; |
1100 | if (OldSize) { |
1101 | NewNumBuckets = 1 << (Log2_32_Ceil(OldSize) + 1); |
1102 | if (NewNumBuckets > InlineBuckets && NewNumBuckets < 64u) |
1103 | NewNumBuckets = 64; |
1104 | } |
1105 | if ((Small && NewNumBuckets <= InlineBuckets) || |
1106 | (!Small && NewNumBuckets == getLargeRep()->NumBuckets)) { |
1107 | this->BaseT::initEmpty(); |
1108 | return; |
1109 | } |
1110 | |
1111 | deallocateBuckets(); |
1112 | init(NewNumBuckets); |
1113 | } |
1114 | |
1115 | private: |
1116 | unsigned getNumEntries() const { |
1117 | return NumEntries; |
1118 | } |
1119 | |
1120 | void setNumEntries(unsigned Num) { |
1121 | // NumEntries is hardcoded to be 31 bits wide. |
1122 | assert(Num < (1U << 31) && "Cannot support more than 1<<31 entries")((void)0); |
1123 | NumEntries = Num; |
1124 | } |
1125 | |
1126 | unsigned getNumTombstones() const { |
1127 | return NumTombstones; |
1128 | } |
1129 | |
1130 | void setNumTombstones(unsigned Num) { |
1131 | NumTombstones = Num; |
1132 | } |
1133 | |
1134 | const BucketT *getInlineBuckets() const { |
1135 | assert(Small)((void)0); |
1136 | // Note that this cast does not violate aliasing rules as we assert that |
1137 | // the memory's dynamic type is the small, inline bucket buffer, and the |
1138 | // 'storage' is a POD containing a char buffer. |
1139 | return reinterpret_cast<const BucketT *>(&storage); |
1140 | } |
1141 | |
1142 | BucketT *getInlineBuckets() { |
1143 | return const_cast<BucketT *>( |
1144 | const_cast<const SmallDenseMap *>(this)->getInlineBuckets()); |
1145 | } |
1146 | |
1147 | const LargeRep *getLargeRep() const { |
1148 | assert(!Small)((void)0); |
1149 | // Note, same rule about aliasing as with getInlineBuckets. |
1150 | return reinterpret_cast<const LargeRep *>(&storage); |
1151 | } |
1152 | |
1153 | LargeRep *getLargeRep() { |
1154 | return const_cast<LargeRep *>( |
1155 | const_cast<const SmallDenseMap *>(this)->getLargeRep()); |
1156 | } |
1157 | |
1158 | const BucketT *getBuckets() const { |
1159 | return Small ? getInlineBuckets() : getLargeRep()->Buckets; |
1160 | } |
1161 | |
1162 | BucketT *getBuckets() { |
1163 | return const_cast<BucketT *>( |
1164 | const_cast<const SmallDenseMap *>(this)->getBuckets()); |
1165 | } |
1166 | |
1167 | unsigned getNumBuckets() const { |
1168 | return Small ? InlineBuckets : getLargeRep()->NumBuckets; |
1169 | } |
1170 | |
1171 | void deallocateBuckets() { |
1172 | if (Small) |
1173 | return; |
1174 | |
1175 | deallocate_buffer(getLargeRep()->Buckets, |
1176 | sizeof(BucketT) * getLargeRep()->NumBuckets, |
1177 | alignof(BucketT)); |
1178 | getLargeRep()->~LargeRep(); |
1179 | } |
1180 | |
1181 | LargeRep allocateBuckets(unsigned Num) { |
1182 | assert(Num > InlineBuckets && "Must allocate more buckets than are inline")((void)0); |
1183 | LargeRep Rep = {static_cast<BucketT *>(allocate_buffer( |
1184 | sizeof(BucketT) * Num, alignof(BucketT))), |
1185 | Num}; |
1186 | return Rep; |
1187 | } |
1188 | }; |
1189 | |
1190 | template <typename KeyT, typename ValueT, typename KeyInfoT, typename Bucket, |
1191 | bool IsConst> |
1192 | class DenseMapIterator : DebugEpochBase::HandleBase { |
1193 | friend class DenseMapIterator<KeyT, ValueT, KeyInfoT, Bucket, true>; |
1194 | friend class DenseMapIterator<KeyT, ValueT, KeyInfoT, Bucket, false>; |
1195 | |
1196 | public: |
1197 | using difference_type = ptrdiff_t; |
1198 | using value_type = |
1199 | typename std::conditional<IsConst, const Bucket, Bucket>::type; |
1200 | using pointer = value_type *; |
1201 | using reference = value_type &; |
1202 | using iterator_category = std::forward_iterator_tag; |
1203 | |
1204 | private: |
1205 | pointer Ptr = nullptr; |
1206 | pointer End = nullptr; |
1207 | |
1208 | public: |
1209 | DenseMapIterator() = default; |
1210 | |
1211 | DenseMapIterator(pointer Pos, pointer E, const DebugEpochBase &Epoch, |
1212 | bool NoAdvance = false) |
1213 | : DebugEpochBase::HandleBase(&Epoch), Ptr(Pos), End(E) { |
1214 | assert(isHandleInSync() && "invalid construction!")((void)0); |
1215 | |
1216 | if (NoAdvance) return; |
1217 | if (shouldReverseIterate<KeyT>()) { |
1218 | RetreatPastEmptyBuckets(); |
1219 | return; |
1220 | } |
1221 | AdvancePastEmptyBuckets(); |
1222 | } |
1223 | |
1224 | // Converting ctor from non-const iterators to const iterators. SFINAE'd out |
1225 | // for const iterator destinations so it doesn't end up as a user defined copy |
1226 | // constructor. |
1227 | template <bool IsConstSrc, |
1228 | typename = std::enable_if_t<!IsConstSrc && IsConst>> |
1229 | DenseMapIterator( |
1230 | const DenseMapIterator<KeyT, ValueT, KeyInfoT, Bucket, IsConstSrc> &I) |
1231 | : DebugEpochBase::HandleBase(I), Ptr(I.Ptr), End(I.End) {} |
1232 | |
1233 | reference operator*() const { |
1234 | assert(isHandleInSync() && "invalid iterator access!")((void)0); |
1235 | assert(Ptr != End && "dereferencing end() iterator")((void)0); |
1236 | if (shouldReverseIterate<KeyT>()) |
1237 | return Ptr[-1]; |
1238 | return *Ptr; |
1239 | } |
1240 | pointer operator->() const { |
1241 | assert(isHandleInSync() && "invalid iterator access!")((void)0); |
1242 | assert(Ptr != End && "dereferencing end() iterator")((void)0); |
1243 | if (shouldReverseIterate<KeyT>()) |
1244 | return &(Ptr[-1]); |
1245 | return Ptr; |
1246 | } |
1247 | |
1248 | friend bool operator==(const DenseMapIterator &LHS, |
1249 | const DenseMapIterator &RHS) { |
1250 | assert((!LHS.Ptr || LHS.isHandleInSync()) && "handle not in sync!")((void)0); |
1251 | assert((!RHS.Ptr || RHS.isHandleInSync()) && "handle not in sync!")((void)0); |
1252 | assert(LHS.getEpochAddress() == RHS.getEpochAddress() &&((void)0) |
1253 | "comparing incomparable iterators!")((void)0); |
1254 | return LHS.Ptr == RHS.Ptr; |
1255 | } |
1256 | |
1257 | friend bool operator!=(const DenseMapIterator &LHS, |
1258 | const DenseMapIterator &RHS) { |
1259 | return !(LHS == RHS); |
1260 | } |
1261 | |
1262 | inline DenseMapIterator& operator++() { // Preincrement |
1263 | assert(isHandleInSync() && "invalid iterator access!")((void)0); |
1264 | assert(Ptr != End && "incrementing end() iterator")((void)0); |
1265 | if (shouldReverseIterate<KeyT>()) { |
1266 | --Ptr; |
1267 | RetreatPastEmptyBuckets(); |
1268 | return *this; |
1269 | } |
1270 | ++Ptr; |
1271 | AdvancePastEmptyBuckets(); |
1272 | return *this; |
1273 | } |
1274 | DenseMapIterator operator++(int) { // Postincrement |
1275 | assert(isHandleInSync() && "invalid iterator access!")((void)0); |
1276 | DenseMapIterator tmp = *this; ++*this; return tmp; |
1277 | } |
1278 | |
1279 | private: |
1280 | void AdvancePastEmptyBuckets() { |
1281 | assert(Ptr <= End)((void)0); |
1282 | const KeyT Empty = KeyInfoT::getEmptyKey(); |
1283 | const KeyT Tombstone = KeyInfoT::getTombstoneKey(); |
1284 | |
1285 | while (Ptr != End && (KeyInfoT::isEqual(Ptr->getFirst(), Empty) || |
1286 | KeyInfoT::isEqual(Ptr->getFirst(), Tombstone))) |
1287 | ++Ptr; |
1288 | } |
1289 | |
1290 | void RetreatPastEmptyBuckets() { |
1291 | assert(Ptr >= End)((void)0); |
1292 | const KeyT Empty = KeyInfoT::getEmptyKey(); |
1293 | const KeyT Tombstone = KeyInfoT::getTombstoneKey(); |
1294 | |
1295 | while (Ptr != End && (KeyInfoT::isEqual(Ptr[-1].getFirst(), Empty) || |
1296 | KeyInfoT::isEqual(Ptr[-1].getFirst(), Tombstone))) |
1297 | --Ptr; |
1298 | } |
1299 | }; |
1300 | |
1301 | template <typename KeyT, typename ValueT, typename KeyInfoT> |
1302 | inline size_t capacity_in_bytes(const DenseMap<KeyT, ValueT, KeyInfoT> &X) { |
1303 | return X.getMemorySize(); |
1304 | } |
1305 | |
1306 | } // end namespace llvm |
1307 | |
1308 | #endif // LLVM_ADT_DENSEMAP_H |