1 | //===- TailDuplicator.cpp - Duplicate blocks into predecessors' tails -----===// |
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 utility class duplicates basic blocks ending in unconditional branches |
10 | // into the tails of their predecessors. |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #include "llvm/CodeGen/TailDuplicator.h" |
15 | #include "llvm/ADT/DenseMap.h" |
16 | #include "llvm/ADT/DenseSet.h" |
17 | #include "llvm/ADT/STLExtras.h" |
18 | #include "llvm/ADT/SetVector.h" |
19 | #include "llvm/ADT/SmallPtrSet.h" |
20 | #include "llvm/ADT/SmallVector.h" |
21 | #include "llvm/ADT/Statistic.h" |
22 | #include "llvm/CodeGen/MachineBasicBlock.h" |
23 | #include "llvm/CodeGen/MachineBranchProbabilityInfo.h" |
24 | #include "llvm/CodeGen/MachineFunction.h" |
25 | #include "llvm/CodeGen/MachineInstr.h" |
26 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
27 | #include "llvm/CodeGen/MachineOperand.h" |
28 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
29 | #include "llvm/CodeGen/MachineSSAUpdater.h" |
30 | #include "llvm/CodeGen/MachineSizeOpts.h" |
31 | #include "llvm/CodeGen/TargetInstrInfo.h" |
32 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
33 | #include "llvm/CodeGen/TargetSubtargetInfo.h" |
34 | #include "llvm/IR/DebugLoc.h" |
35 | #include "llvm/IR/Function.h" |
36 | #include "llvm/Support/CommandLine.h" |
37 | #include "llvm/Support/Debug.h" |
38 | #include "llvm/Support/ErrorHandling.h" |
39 | #include "llvm/Support/raw_ostream.h" |
40 | #include "llvm/Target/TargetMachine.h" |
41 | #include <algorithm> |
42 | #include <cassert> |
43 | #include <iterator> |
44 | #include <utility> |
45 | |
46 | using namespace llvm; |
47 | |
48 | #define DEBUG_TYPE "tailduplication" |
49 | |
50 | STATISTIC(NumTails, "Number of tails duplicated" ); |
51 | STATISTIC(NumTailDups, "Number of tail duplicated blocks" ); |
52 | STATISTIC(NumTailDupAdded, |
53 | "Number of instructions added due to tail duplication" ); |
54 | STATISTIC(NumTailDupRemoved, |
55 | "Number of instructions removed due to tail duplication" ); |
56 | STATISTIC(NumDeadBlocks, "Number of dead blocks removed" ); |
57 | STATISTIC(NumAddedPHIs, "Number of phis added" ); |
58 | |
59 | // Heuristic for tail duplication. |
60 | static cl::opt<unsigned> TailDuplicateSize( |
61 | "tail-dup-size" , |
62 | cl::desc("Maximum instructions to consider tail duplicating" ), cl::init(Val: 2), |
63 | cl::Hidden); |
64 | |
65 | static cl::opt<unsigned> TailDupIndirectBranchSize( |
66 | "tail-dup-indirect-size" , |
67 | cl::desc("Maximum instructions to consider tail duplicating blocks that " |
68 | "end with indirect branches." ), cl::init(Val: 20), |
69 | cl::Hidden); |
70 | |
71 | static cl::opt<unsigned> |
72 | TailDupPredSize("tail-dup-pred-size" , |
73 | cl::desc("Maximum predecessors (maximum successors at the " |
74 | "same time) to consider tail duplicating blocks." ), |
75 | cl::init(Val: 16), cl::Hidden); |
76 | |
77 | static cl::opt<unsigned> |
78 | TailDupSuccSize("tail-dup-succ-size" , |
79 | cl::desc("Maximum successors (maximum predecessors at the " |
80 | "same time) to consider tail duplicating blocks." ), |
81 | cl::init(Val: 16), cl::Hidden); |
82 | |
83 | static cl::opt<bool> |
84 | TailDupVerify("tail-dup-verify" , |
85 | cl::desc("Verify sanity of PHI instructions during taildup" ), |
86 | cl::init(Val: false), cl::Hidden); |
87 | |
88 | static cl::opt<unsigned> TailDupLimit("tail-dup-limit" , cl::init(Val: ~0U), |
89 | cl::Hidden); |
90 | |
91 | void TailDuplicator::initMF(MachineFunction &MFin, bool PreRegAlloc, |
92 | const MachineBranchProbabilityInfo *MBPIin, |
93 | MBFIWrapper *MBFIin, |
94 | ProfileSummaryInfo *PSIin, |
95 | bool LayoutModeIn, unsigned TailDupSizeIn) { |
96 | MF = &MFin; |
97 | TII = MF->getSubtarget().getInstrInfo(); |
98 | TRI = MF->getSubtarget().getRegisterInfo(); |
99 | MRI = &MF->getRegInfo(); |
100 | MBPI = MBPIin; |
101 | MBFI = MBFIin; |
102 | PSI = PSIin; |
103 | TailDupSize = TailDupSizeIn; |
104 | |
105 | assert(MBPI != nullptr && "Machine Branch Probability Info required" ); |
106 | |
107 | LayoutMode = LayoutModeIn; |
108 | this->PreRegAlloc = PreRegAlloc; |
109 | } |
110 | |
111 | static void VerifyPHIs(MachineFunction &MF, bool ) { |
112 | for (MachineBasicBlock &MBB : llvm::drop_begin(RangeOrContainer&: MF)) { |
113 | SmallSetVector<MachineBasicBlock *, 8> Preds(MBB.pred_begin(), |
114 | MBB.pred_end()); |
115 | MachineBasicBlock::iterator MI = MBB.begin(); |
116 | while (MI != MBB.end()) { |
117 | if (!MI->isPHI()) |
118 | break; |
119 | for (MachineBasicBlock *PredBB : Preds) { |
120 | bool Found = false; |
121 | for (unsigned i = 1, e = MI->getNumOperands(); i != e; i += 2) { |
122 | MachineBasicBlock *PHIBB = MI->getOperand(i: i + 1).getMBB(); |
123 | if (PHIBB == PredBB) { |
124 | Found = true; |
125 | break; |
126 | } |
127 | } |
128 | if (!Found) { |
129 | dbgs() << "Malformed PHI in " << printMBBReference(MBB) << ": " |
130 | << *MI; |
131 | dbgs() << " missing input from predecessor " |
132 | << printMBBReference(MBB: *PredBB) << '\n'; |
133 | llvm_unreachable(nullptr); |
134 | } |
135 | } |
136 | |
137 | for (unsigned i = 1, e = MI->getNumOperands(); i != e; i += 2) { |
138 | MachineBasicBlock *PHIBB = MI->getOperand(i: i + 1).getMBB(); |
139 | if (CheckExtra && !Preds.count(key: PHIBB)) { |
140 | dbgs() << "Warning: malformed PHI in " << printMBBReference(MBB) |
141 | << ": " << *MI; |
142 | dbgs() << " extra input from predecessor " |
143 | << printMBBReference(MBB: *PHIBB) << '\n'; |
144 | llvm_unreachable(nullptr); |
145 | } |
146 | if (PHIBB->getNumber() < 0) { |
147 | dbgs() << "Malformed PHI in " << printMBBReference(MBB) << ": " |
148 | << *MI; |
149 | dbgs() << " non-existing " << printMBBReference(MBB: *PHIBB) << '\n'; |
150 | llvm_unreachable(nullptr); |
151 | } |
152 | } |
153 | ++MI; |
154 | } |
155 | } |
156 | } |
157 | |
158 | /// Tail duplicate the block and cleanup. |
159 | /// \p IsSimple - return value of isSimpleBB |
160 | /// \p MBB - block to be duplicated |
161 | /// \p ForcedLayoutPred - If non-null, treat this block as the layout |
162 | /// predecessor, instead of using the ordering in MF |
163 | /// \p DuplicatedPreds - if non-null, \p DuplicatedPreds will contain a list of |
164 | /// all Preds that received a copy of \p MBB. |
165 | /// \p RemovalCallback - if non-null, called just before MBB is deleted. |
166 | bool TailDuplicator::tailDuplicateAndUpdate( |
167 | bool IsSimple, MachineBasicBlock *MBB, |
168 | MachineBasicBlock *ForcedLayoutPred, |
169 | SmallVectorImpl<MachineBasicBlock*> *DuplicatedPreds, |
170 | function_ref<void(MachineBasicBlock *)> *RemovalCallback, |
171 | SmallVectorImpl<MachineBasicBlock *> *CandidatePtr) { |
172 | // Save the successors list. |
173 | SmallSetVector<MachineBasicBlock *, 8> Succs(MBB->succ_begin(), |
174 | MBB->succ_end()); |
175 | |
176 | SmallVector<MachineBasicBlock *, 8> TDBBs; |
177 | SmallVector<MachineInstr *, 16> Copies; |
178 | if (!tailDuplicate(IsSimple, TailBB: MBB, ForcedLayoutPred, |
179 | TDBBs, Copies, CandidatePtr)) |
180 | return false; |
181 | |
182 | ++NumTails; |
183 | |
184 | SmallVector<MachineInstr *, 8> NewPHIs; |
185 | MachineSSAUpdater SSAUpdate(*MF, &NewPHIs); |
186 | |
187 | // TailBB's immediate successors are now successors of those predecessors |
188 | // which duplicated TailBB. Add the predecessors as sources to the PHI |
189 | // instructions. |
190 | bool isDead = MBB->pred_empty() && !MBB->hasAddressTaken(); |
191 | if (PreRegAlloc) |
192 | updateSuccessorsPHIs(FromBB: MBB, isDead, TDBBs, Succs); |
193 | |
194 | // If it is dead, remove it. |
195 | if (isDead) { |
196 | NumTailDupRemoved += MBB->size(); |
197 | removeDeadBlock(MBB, RemovalCallback); |
198 | ++NumDeadBlocks; |
199 | } |
200 | |
201 | // Update SSA form. |
202 | if (!SSAUpdateVRs.empty()) { |
203 | for (unsigned VReg : SSAUpdateVRs) { |
204 | SSAUpdate.Initialize(V: VReg); |
205 | |
206 | // If the original definition is still around, add it as an available |
207 | // value. |
208 | MachineInstr *DefMI = MRI->getVRegDef(Reg: VReg); |
209 | MachineBasicBlock *DefBB = nullptr; |
210 | if (DefMI) { |
211 | DefBB = DefMI->getParent(); |
212 | SSAUpdate.AddAvailableValue(BB: DefBB, V: VReg); |
213 | } |
214 | |
215 | // Add the new vregs as available values. |
216 | DenseMap<Register, AvailableValsTy>::iterator LI = |
217 | SSAUpdateVals.find(Val: VReg); |
218 | for (std::pair<MachineBasicBlock *, Register> &J : LI->second) { |
219 | MachineBasicBlock *SrcBB = J.first; |
220 | Register SrcReg = J.second; |
221 | SSAUpdate.AddAvailableValue(BB: SrcBB, V: SrcReg); |
222 | } |
223 | |
224 | SmallVector<MachineOperand *> DebugUses; |
225 | // Rewrite uses that are outside of the original def's block. |
226 | for (MachineOperand &UseMO : |
227 | llvm::make_early_inc_range(Range: MRI->use_operands(Reg: VReg))) { |
228 | MachineInstr *UseMI = UseMO.getParent(); |
229 | // Rewrite debug uses last so that they can take advantage of any |
230 | // register mappings introduced by other users in its BB, since we |
231 | // cannot create new register definitions specifically for the debug |
232 | // instruction (as debug instructions should not affect CodeGen). |
233 | if (UseMI->isDebugValue()) { |
234 | DebugUses.push_back(Elt: &UseMO); |
235 | continue; |
236 | } |
237 | if (UseMI->getParent() == DefBB && !UseMI->isPHI()) |
238 | continue; |
239 | SSAUpdate.RewriteUse(U&: UseMO); |
240 | } |
241 | for (auto *UseMO : DebugUses) { |
242 | MachineInstr *UseMI = UseMO->getParent(); |
243 | UseMO->setReg( |
244 | SSAUpdate.GetValueInMiddleOfBlock(BB: UseMI->getParent(), ExistingValueOnly: true)); |
245 | } |
246 | } |
247 | |
248 | SSAUpdateVRs.clear(); |
249 | SSAUpdateVals.clear(); |
250 | } |
251 | |
252 | // Eliminate some of the copies inserted by tail duplication to maintain |
253 | // SSA form. |
254 | for (MachineInstr *Copy : Copies) { |
255 | if (!Copy->isCopy()) |
256 | continue; |
257 | Register Dst = Copy->getOperand(i: 0).getReg(); |
258 | Register Src = Copy->getOperand(i: 1).getReg(); |
259 | if (MRI->hasOneNonDBGUse(RegNo: Src) && |
260 | MRI->constrainRegClass(Reg: Src, RC: MRI->getRegClass(Reg: Dst))) { |
261 | // Copy is the only use. Do trivial copy propagation here. |
262 | MRI->replaceRegWith(FromReg: Dst, ToReg: Src); |
263 | Copy->eraseFromParent(); |
264 | } |
265 | } |
266 | |
267 | if (NewPHIs.size()) |
268 | NumAddedPHIs += NewPHIs.size(); |
269 | |
270 | if (DuplicatedPreds) |
271 | *DuplicatedPreds = std::move(TDBBs); |
272 | |
273 | return true; |
274 | } |
275 | |
276 | /// Look for small blocks that are unconditionally branched to and do not fall |
277 | /// through. Tail-duplicate their instructions into their predecessors to |
278 | /// eliminate (dynamic) branches. |
279 | bool TailDuplicator::tailDuplicateBlocks() { |
280 | bool MadeChange = false; |
281 | |
282 | if (PreRegAlloc && TailDupVerify) { |
283 | LLVM_DEBUG(dbgs() << "\n*** Before tail-duplicating\n" ); |
284 | VerifyPHIs(MF&: *MF, CheckExtra: true); |
285 | } |
286 | |
287 | for (MachineBasicBlock &MBB : |
288 | llvm::make_early_inc_range(Range: llvm::drop_begin(RangeOrContainer&: *MF))) { |
289 | if (NumTails == TailDupLimit) |
290 | break; |
291 | |
292 | bool IsSimple = isSimpleBB(TailBB: &MBB); |
293 | |
294 | if (!shouldTailDuplicate(IsSimple, TailBB&: MBB)) |
295 | continue; |
296 | |
297 | MadeChange |= tailDuplicateAndUpdate(IsSimple, MBB: &MBB, ForcedLayoutPred: nullptr); |
298 | } |
299 | |
300 | if (PreRegAlloc && TailDupVerify) |
301 | VerifyPHIs(MF&: *MF, CheckExtra: false); |
302 | |
303 | return MadeChange; |
304 | } |
305 | |
306 | static bool isDefLiveOut(Register Reg, MachineBasicBlock *BB, |
307 | const MachineRegisterInfo *MRI) { |
308 | for (MachineInstr &UseMI : MRI->use_instructions(Reg)) { |
309 | if (UseMI.isDebugValue()) |
310 | continue; |
311 | if (UseMI.getParent() != BB) |
312 | return true; |
313 | } |
314 | return false; |
315 | } |
316 | |
317 | static unsigned getPHISrcRegOpIdx(MachineInstr *MI, MachineBasicBlock *SrcBB) { |
318 | for (unsigned i = 1, e = MI->getNumOperands(); i != e; i += 2) |
319 | if (MI->getOperand(i: i + 1).getMBB() == SrcBB) |
320 | return i; |
321 | return 0; |
322 | } |
323 | |
324 | // Remember which registers are used by phis in this block. This is |
325 | // used to determine which registers are liveout while modifying the |
326 | // block (which is why we need to copy the information). |
327 | static void getRegsUsedByPHIs(const MachineBasicBlock &BB, |
328 | DenseSet<Register> *UsedByPhi) { |
329 | for (const auto &MI : BB) { |
330 | if (!MI.isPHI()) |
331 | break; |
332 | for (unsigned i = 1, e = MI.getNumOperands(); i != e; i += 2) { |
333 | Register SrcReg = MI.getOperand(i).getReg(); |
334 | UsedByPhi->insert(V: SrcReg); |
335 | } |
336 | } |
337 | } |
338 | |
339 | /// Add a definition and source virtual registers pair for SSA update. |
340 | void TailDuplicator::addSSAUpdateEntry(Register OrigReg, Register NewReg, |
341 | MachineBasicBlock *BB) { |
342 | DenseMap<Register, AvailableValsTy>::iterator LI = |
343 | SSAUpdateVals.find(Val: OrigReg); |
344 | if (LI != SSAUpdateVals.end()) |
345 | LI->second.push_back(x: std::make_pair(x&: BB, y&: NewReg)); |
346 | else { |
347 | AvailableValsTy Vals; |
348 | Vals.push_back(x: std::make_pair(x&: BB, y&: NewReg)); |
349 | SSAUpdateVals.insert(KV: std::make_pair(x&: OrigReg, y&: Vals)); |
350 | SSAUpdateVRs.push_back(Elt: OrigReg); |
351 | } |
352 | } |
353 | |
354 | /// Process PHI node in TailBB by turning it into a copy in PredBB. Remember the |
355 | /// source register that's contributed by PredBB and update SSA update map. |
356 | void TailDuplicator::processPHI( |
357 | MachineInstr *MI, MachineBasicBlock *TailBB, MachineBasicBlock *PredBB, |
358 | DenseMap<Register, RegSubRegPair> &LocalVRMap, |
359 | SmallVectorImpl<std::pair<Register, RegSubRegPair>> &Copies, |
360 | const DenseSet<Register> &RegsUsedByPhi, bool Remove) { |
361 | Register DefReg = MI->getOperand(i: 0).getReg(); |
362 | unsigned SrcOpIdx = getPHISrcRegOpIdx(MI, SrcBB: PredBB); |
363 | assert(SrcOpIdx && "Unable to find matching PHI source?" ); |
364 | Register SrcReg = MI->getOperand(i: SrcOpIdx).getReg(); |
365 | unsigned SrcSubReg = MI->getOperand(i: SrcOpIdx).getSubReg(); |
366 | const TargetRegisterClass *RC = MRI->getRegClass(Reg: DefReg); |
367 | LocalVRMap.insert(KV: std::make_pair(x&: DefReg, y: RegSubRegPair(SrcReg, SrcSubReg))); |
368 | |
369 | // Insert a copy from source to the end of the block. The def register is the |
370 | // available value liveout of the block. |
371 | Register NewDef = MRI->createVirtualRegister(RegClass: RC); |
372 | Copies.push_back(Elt: std::make_pair(x&: NewDef, y: RegSubRegPair(SrcReg, SrcSubReg))); |
373 | if (isDefLiveOut(Reg: DefReg, BB: TailBB, MRI) || RegsUsedByPhi.count(V: DefReg)) |
374 | addSSAUpdateEntry(OrigReg: DefReg, NewReg: NewDef, BB: PredBB); |
375 | |
376 | if (!Remove) |
377 | return; |
378 | |
379 | // Remove PredBB from the PHI node. |
380 | MI->removeOperand(OpNo: SrcOpIdx + 1); |
381 | MI->removeOperand(OpNo: SrcOpIdx); |
382 | if (MI->getNumOperands() == 1 && !TailBB->hasAddressTaken()) |
383 | MI->eraseFromParent(); |
384 | else if (MI->getNumOperands() == 1) |
385 | MI->setDesc(TII->get(Opcode: TargetOpcode::IMPLICIT_DEF)); |
386 | } |
387 | |
388 | /// Duplicate a TailBB instruction to PredBB and update |
389 | /// the source operands due to earlier PHI translation. |
390 | void TailDuplicator::duplicateInstruction( |
391 | MachineInstr *MI, MachineBasicBlock *TailBB, MachineBasicBlock *PredBB, |
392 | DenseMap<Register, RegSubRegPair> &LocalVRMap, |
393 | const DenseSet<Register> &UsedByPhi) { |
394 | // Allow duplication of CFI instructions. |
395 | if (MI->isCFIInstruction()) { |
396 | BuildMI(BB&: *PredBB, I: PredBB->end(), MIMD: PredBB->findDebugLoc(MBBI: PredBB->begin()), |
397 | MCID: TII->get(Opcode: TargetOpcode::CFI_INSTRUCTION)) |
398 | .addCFIIndex(CFIIndex: MI->getOperand(i: 0).getCFIIndex()) |
399 | .setMIFlags(MI->getFlags()); |
400 | return; |
401 | } |
402 | MachineInstr &NewMI = TII->duplicate(MBB&: *PredBB, InsertBefore: PredBB->end(), Orig: *MI); |
403 | if (PreRegAlloc) { |
404 | for (unsigned i = 0, e = NewMI.getNumOperands(); i != e; ++i) { |
405 | MachineOperand &MO = NewMI.getOperand(i); |
406 | if (!MO.isReg()) |
407 | continue; |
408 | Register Reg = MO.getReg(); |
409 | if (!Reg.isVirtual()) |
410 | continue; |
411 | if (MO.isDef()) { |
412 | const TargetRegisterClass *RC = MRI->getRegClass(Reg); |
413 | Register NewReg = MRI->createVirtualRegister(RegClass: RC); |
414 | MO.setReg(NewReg); |
415 | LocalVRMap.insert(KV: std::make_pair(x&: Reg, y: RegSubRegPair(NewReg, 0))); |
416 | if (isDefLiveOut(Reg, BB: TailBB, MRI) || UsedByPhi.count(V: Reg)) |
417 | addSSAUpdateEntry(OrigReg: Reg, NewReg, BB: PredBB); |
418 | } else { |
419 | auto VI = LocalVRMap.find(Val: Reg); |
420 | if (VI != LocalVRMap.end()) { |
421 | // Need to make sure that the register class of the mapped register |
422 | // will satisfy the constraints of the class of the register being |
423 | // replaced. |
424 | auto *OrigRC = MRI->getRegClass(Reg); |
425 | auto *MappedRC = MRI->getRegClass(Reg: VI->second.Reg); |
426 | const TargetRegisterClass *ConstrRC; |
427 | if (VI->second.SubReg != 0) { |
428 | ConstrRC = TRI->getMatchingSuperRegClass(A: MappedRC, B: OrigRC, |
429 | Idx: VI->second.SubReg); |
430 | if (ConstrRC) { |
431 | // The actual constraining (as in "find appropriate new class") |
432 | // is done by getMatchingSuperRegClass, so now we only need to |
433 | // change the class of the mapped register. |
434 | MRI->setRegClass(Reg: VI->second.Reg, RC: ConstrRC); |
435 | } |
436 | } else { |
437 | // For mapped registers that do not have sub-registers, simply |
438 | // restrict their class to match the original one. |
439 | |
440 | // We don't want debug instructions affecting the resulting code so |
441 | // if we're cloning a debug instruction then just use MappedRC |
442 | // rather than constraining the register class further. |
443 | ConstrRC = NewMI.isDebugInstr() |
444 | ? MappedRC |
445 | : MRI->constrainRegClass(Reg: VI->second.Reg, RC: OrigRC); |
446 | } |
447 | |
448 | if (ConstrRC) { |
449 | // If the class constraining succeeded, we can simply replace |
450 | // the old register with the mapped one. |
451 | MO.setReg(VI->second.Reg); |
452 | // We have Reg -> VI.Reg:VI.SubReg, so if Reg is used with a |
453 | // sub-register, we need to compose the sub-register indices. |
454 | MO.setSubReg( |
455 | TRI->composeSubRegIndices(a: VI->second.SubReg, b: MO.getSubReg())); |
456 | } else { |
457 | // The direct replacement is not possible, due to failing register |
458 | // class constraints. An explicit COPY is necessary. Create one |
459 | // that can be reused. |
460 | Register NewReg = MRI->createVirtualRegister(RegClass: OrigRC); |
461 | BuildMI(BB&: *PredBB, I&: NewMI, MIMD: NewMI.getDebugLoc(), |
462 | MCID: TII->get(Opcode: TargetOpcode::COPY), DestReg: NewReg) |
463 | .addReg(RegNo: VI->second.Reg, flags: 0, SubReg: VI->second.SubReg); |
464 | LocalVRMap.erase(I: VI); |
465 | LocalVRMap.insert(KV: std::make_pair(x&: Reg, y: RegSubRegPair(NewReg, 0))); |
466 | MO.setReg(NewReg); |
467 | // The composed VI.Reg:VI.SubReg is replaced with NewReg, which |
468 | // is equivalent to the whole register Reg. Hence, Reg:subreg |
469 | // is same as NewReg:subreg, so keep the sub-register index |
470 | // unchanged. |
471 | } |
472 | // Clear any kill flags from this operand. The new register could |
473 | // have uses after this one, so kills are not valid here. |
474 | MO.setIsKill(false); |
475 | } |
476 | } |
477 | } |
478 | } |
479 | } |
480 | |
481 | /// After FromBB is tail duplicated into its predecessor blocks, the successors |
482 | /// have gained new predecessors. Update the PHI instructions in them |
483 | /// accordingly. |
484 | void TailDuplicator::updateSuccessorsPHIs( |
485 | MachineBasicBlock *FromBB, bool isDead, |
486 | SmallVectorImpl<MachineBasicBlock *> &TDBBs, |
487 | SmallSetVector<MachineBasicBlock *, 8> &Succs) { |
488 | for (MachineBasicBlock *SuccBB : Succs) { |
489 | for (MachineInstr &MI : *SuccBB) { |
490 | if (!MI.isPHI()) |
491 | break; |
492 | MachineInstrBuilder MIB(*FromBB->getParent(), MI); |
493 | unsigned Idx = 0; |
494 | for (unsigned i = 1, e = MI.getNumOperands(); i != e; i += 2) { |
495 | MachineOperand &MO = MI.getOperand(i: i + 1); |
496 | if (MO.getMBB() == FromBB) { |
497 | Idx = i; |
498 | break; |
499 | } |
500 | } |
501 | |
502 | assert(Idx != 0); |
503 | MachineOperand &MO0 = MI.getOperand(i: Idx); |
504 | Register Reg = MO0.getReg(); |
505 | if (isDead) { |
506 | // Folded into the previous BB. |
507 | // There could be duplicate phi source entries. FIXME: Should sdisel |
508 | // or earlier pass fixed this? |
509 | for (unsigned i = MI.getNumOperands() - 2; i != Idx; i -= 2) { |
510 | MachineOperand &MO = MI.getOperand(i: i + 1); |
511 | if (MO.getMBB() == FromBB) { |
512 | MI.removeOperand(OpNo: i + 1); |
513 | MI.removeOperand(OpNo: i); |
514 | } |
515 | } |
516 | } else |
517 | Idx = 0; |
518 | |
519 | // If Idx is set, the operands at Idx and Idx+1 must be removed. |
520 | // We reuse the location to avoid expensive removeOperand calls. |
521 | |
522 | DenseMap<Register, AvailableValsTy>::iterator LI = |
523 | SSAUpdateVals.find(Val: Reg); |
524 | if (LI != SSAUpdateVals.end()) { |
525 | // This register is defined in the tail block. |
526 | for (const std::pair<MachineBasicBlock *, Register> &J : LI->second) { |
527 | MachineBasicBlock *SrcBB = J.first; |
528 | // If we didn't duplicate a bb into a particular predecessor, we |
529 | // might still have added an entry to SSAUpdateVals to correcly |
530 | // recompute SSA. If that case, avoid adding a dummy extra argument |
531 | // this PHI. |
532 | if (!SrcBB->isSuccessor(MBB: SuccBB)) |
533 | continue; |
534 | |
535 | Register SrcReg = J.second; |
536 | if (Idx != 0) { |
537 | MI.getOperand(i: Idx).setReg(SrcReg); |
538 | MI.getOperand(i: Idx + 1).setMBB(SrcBB); |
539 | Idx = 0; |
540 | } else { |
541 | MIB.addReg(RegNo: SrcReg).addMBB(MBB: SrcBB); |
542 | } |
543 | } |
544 | } else { |
545 | // Live in tail block, must also be live in predecessors. |
546 | for (MachineBasicBlock *SrcBB : TDBBs) { |
547 | if (Idx != 0) { |
548 | MI.getOperand(i: Idx).setReg(Reg); |
549 | MI.getOperand(i: Idx + 1).setMBB(SrcBB); |
550 | Idx = 0; |
551 | } else { |
552 | MIB.addReg(RegNo: Reg).addMBB(MBB: SrcBB); |
553 | } |
554 | } |
555 | } |
556 | if (Idx != 0) { |
557 | MI.removeOperand(OpNo: Idx + 1); |
558 | MI.removeOperand(OpNo: Idx); |
559 | } |
560 | } |
561 | } |
562 | } |
563 | |
564 | /// Determine if it is profitable to duplicate this block. |
565 | bool TailDuplicator::shouldTailDuplicate(bool IsSimple, |
566 | MachineBasicBlock &TailBB) { |
567 | // When doing tail-duplication during layout, the block ordering is in flux, |
568 | // so canFallThrough returns a result based on incorrect information and |
569 | // should just be ignored. |
570 | if (!LayoutMode && TailBB.canFallThrough()) |
571 | return false; |
572 | |
573 | // Don't try to tail-duplicate single-block loops. |
574 | if (TailBB.isSuccessor(MBB: &TailBB)) |
575 | return false; |
576 | |
577 | // Duplicating a BB which has both multiple predecessors and successors will |
578 | // result in a complex CFG and also may cause huge amount of PHI nodes. If we |
579 | // want to remove this limitation, we have to address |
580 | // https://github.com/llvm/llvm-project/issues/78578. |
581 | if (TailBB.pred_size() > TailDupPredSize && |
582 | TailBB.succ_size() > TailDupSuccSize) |
583 | return false; |
584 | |
585 | // Set the limit on the cost to duplicate. When optimizing for size, |
586 | // duplicate only one, because one branch instruction can be eliminated to |
587 | // compensate for the duplication. |
588 | unsigned MaxDuplicateCount; |
589 | bool OptForSize = MF->getFunction().hasOptSize() || |
590 | llvm::shouldOptimizeForSize(MBB: &TailBB, PSI, MBFIWrapper: MBFI); |
591 | if (TailDupSize == 0) |
592 | MaxDuplicateCount = TailDuplicateSize; |
593 | else |
594 | MaxDuplicateCount = TailDupSize; |
595 | if (OptForSize) |
596 | MaxDuplicateCount = 1; |
597 | |
598 | // If the block to be duplicated ends in an unanalyzable fallthrough, don't |
599 | // duplicate it. |
600 | // A similar check is necessary in MachineBlockPlacement to make sure pairs of |
601 | // blocks with unanalyzable fallthrough get layed out contiguously. |
602 | MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr; |
603 | SmallVector<MachineOperand, 4> PredCond; |
604 | if (TII->analyzeBranch(MBB&: TailBB, TBB&: PredTBB, FBB&: PredFBB, Cond&: PredCond) && |
605 | TailBB.canFallThrough()) |
606 | return false; |
607 | |
608 | // If the target has hardware branch prediction that can handle indirect |
609 | // branches, duplicating them can often make them predictable when there |
610 | // are common paths through the code. The limit needs to be high enough |
611 | // to allow undoing the effects of tail merging and other optimizations |
612 | // that rearrange the predecessors of the indirect branch. |
613 | |
614 | bool HasIndirectbr = false; |
615 | if (!TailBB.empty()) |
616 | HasIndirectbr = TailBB.back().isIndirectBranch(); |
617 | |
618 | if (HasIndirectbr && PreRegAlloc) |
619 | MaxDuplicateCount = TailDupIndirectBranchSize; |
620 | |
621 | // Check the instructions in the block to determine whether tail-duplication |
622 | // is invalid or unlikely to be profitable. |
623 | unsigned InstrCount = 0; |
624 | for (MachineInstr &MI : TailBB) { |
625 | // Non-duplicable things shouldn't be tail-duplicated. |
626 | // CFI instructions are marked as non-duplicable, because Darwin compact |
627 | // unwind info emission can't handle multiple prologue setups. In case of |
628 | // DWARF, allow them be duplicated, so that their existence doesn't prevent |
629 | // tail duplication of some basic blocks, that would be duplicated otherwise. |
630 | if (MI.isNotDuplicable() && |
631 | (TailBB.getParent()->getTarget().getTargetTriple().isOSDarwin() || |
632 | !MI.isCFIInstruction())) |
633 | return false; |
634 | |
635 | // Convergent instructions can be duplicated only if doing so doesn't add |
636 | // new control dependencies, which is what we're going to do here. |
637 | if (MI.isConvergent()) |
638 | return false; |
639 | |
640 | // Do not duplicate 'return' instructions if this is a pre-regalloc run. |
641 | // A return may expand into a lot more instructions (e.g. reload of callee |
642 | // saved registers) after PEI. |
643 | if (PreRegAlloc && MI.isReturn()) |
644 | return false; |
645 | |
646 | // Avoid duplicating calls before register allocation. Calls presents a |
647 | // barrier to register allocation so duplicating them may end up increasing |
648 | // spills. |
649 | if (PreRegAlloc && MI.isCall()) |
650 | return false; |
651 | |
652 | // TailDuplicator::appendCopies will erroneously place COPYs after |
653 | // INLINEASM_BR instructions after 4b0aa5724fea, which demonstrates the same |
654 | // bug that was fixed in f7a53d82c090. |
655 | // FIXME: Use findPHICopyInsertPoint() to find the correct insertion point |
656 | // for the COPY when replacing PHIs. |
657 | if (MI.getOpcode() == TargetOpcode::INLINEASM_BR) |
658 | return false; |
659 | |
660 | if (MI.isBundle()) |
661 | InstrCount += MI.getBundleSize(); |
662 | else if (!MI.isPHI() && !MI.isMetaInstruction()) |
663 | InstrCount += 1; |
664 | |
665 | if (InstrCount > MaxDuplicateCount) |
666 | return false; |
667 | } |
668 | |
669 | // Check if any of the successors of TailBB has a PHI node in which the |
670 | // value corresponding to TailBB uses a subregister. |
671 | // If a phi node uses a register paired with a subregister, the actual |
672 | // "value type" of the phi may differ from the type of the register without |
673 | // any subregisters. Due to a bug, tail duplication may add a new operand |
674 | // without a necessary subregister, producing an invalid code. This is |
675 | // demonstrated by test/CodeGen/Hexagon/tail-dup-subreg-abort.ll. |
676 | // Disable tail duplication for this case for now, until the problem is |
677 | // fixed. |
678 | for (auto *SB : TailBB.successors()) { |
679 | for (auto &I : *SB) { |
680 | if (!I.isPHI()) |
681 | break; |
682 | unsigned Idx = getPHISrcRegOpIdx(MI: &I, SrcBB: &TailBB); |
683 | assert(Idx != 0); |
684 | MachineOperand &PU = I.getOperand(i: Idx); |
685 | if (PU.getSubReg() != 0) |
686 | return false; |
687 | } |
688 | } |
689 | |
690 | if (HasIndirectbr && PreRegAlloc) |
691 | return true; |
692 | |
693 | if (IsSimple) |
694 | return true; |
695 | |
696 | if (!PreRegAlloc) |
697 | return true; |
698 | |
699 | return canCompletelyDuplicateBB(BB&: TailBB); |
700 | } |
701 | |
702 | /// True if this BB has only one unconditional jump. |
703 | bool TailDuplicator::isSimpleBB(MachineBasicBlock *TailBB) { |
704 | if (TailBB->succ_size() != 1) |
705 | return false; |
706 | if (TailBB->pred_empty()) |
707 | return false; |
708 | MachineBasicBlock::iterator I = TailBB->getFirstNonDebugInstr(SkipPseudoOp: true); |
709 | if (I == TailBB->end()) |
710 | return true; |
711 | return I->isUnconditionalBranch(); |
712 | } |
713 | |
714 | static bool bothUsedInPHI(const MachineBasicBlock &A, |
715 | const SmallPtrSet<MachineBasicBlock *, 8> &SuccsB) { |
716 | for (MachineBasicBlock *BB : A.successors()) |
717 | if (SuccsB.count(Ptr: BB) && !BB->empty() && BB->begin()->isPHI()) |
718 | return true; |
719 | |
720 | return false; |
721 | } |
722 | |
723 | bool TailDuplicator::canCompletelyDuplicateBB(MachineBasicBlock &BB) { |
724 | for (MachineBasicBlock *PredBB : BB.predecessors()) { |
725 | if (PredBB->succ_size() > 1) |
726 | return false; |
727 | |
728 | MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr; |
729 | SmallVector<MachineOperand, 4> PredCond; |
730 | if (TII->analyzeBranch(MBB&: *PredBB, TBB&: PredTBB, FBB&: PredFBB, Cond&: PredCond)) |
731 | return false; |
732 | |
733 | if (!PredCond.empty()) |
734 | return false; |
735 | } |
736 | return true; |
737 | } |
738 | |
739 | bool TailDuplicator::duplicateSimpleBB( |
740 | MachineBasicBlock *TailBB, SmallVectorImpl<MachineBasicBlock *> &TDBBs, |
741 | const DenseSet<Register> &UsedByPhi) { |
742 | SmallPtrSet<MachineBasicBlock *, 8> Succs(TailBB->succ_begin(), |
743 | TailBB->succ_end()); |
744 | SmallVector<MachineBasicBlock *, 8> Preds(TailBB->predecessors()); |
745 | bool Changed = false; |
746 | for (MachineBasicBlock *PredBB : Preds) { |
747 | if (PredBB->hasEHPadSuccessor() || PredBB->mayHaveInlineAsmBr()) |
748 | continue; |
749 | |
750 | if (bothUsedInPHI(A: *PredBB, SuccsB: Succs)) |
751 | continue; |
752 | |
753 | MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr; |
754 | SmallVector<MachineOperand, 4> PredCond; |
755 | if (TII->analyzeBranch(MBB&: *PredBB, TBB&: PredTBB, FBB&: PredFBB, Cond&: PredCond)) |
756 | continue; |
757 | |
758 | Changed = true; |
759 | LLVM_DEBUG(dbgs() << "\nTail-duplicating into PredBB: " << *PredBB |
760 | << "From simple Succ: " << *TailBB); |
761 | |
762 | MachineBasicBlock *NewTarget = *TailBB->succ_begin(); |
763 | MachineBasicBlock *NextBB = PredBB->getNextNode(); |
764 | |
765 | // Make PredFBB explicit. |
766 | if (PredCond.empty()) |
767 | PredFBB = PredTBB; |
768 | |
769 | // Make fall through explicit. |
770 | if (!PredTBB) |
771 | PredTBB = NextBB; |
772 | if (!PredFBB) |
773 | PredFBB = NextBB; |
774 | |
775 | // Redirect |
776 | if (PredFBB == TailBB) |
777 | PredFBB = NewTarget; |
778 | if (PredTBB == TailBB) |
779 | PredTBB = NewTarget; |
780 | |
781 | // Make the branch unconditional if possible |
782 | if (PredTBB == PredFBB) { |
783 | PredCond.clear(); |
784 | PredFBB = nullptr; |
785 | } |
786 | |
787 | // Avoid adding fall through branches. |
788 | if (PredFBB == NextBB) |
789 | PredFBB = nullptr; |
790 | if (PredTBB == NextBB && PredFBB == nullptr) |
791 | PredTBB = nullptr; |
792 | |
793 | auto DL = PredBB->findBranchDebugLoc(); |
794 | TII->removeBranch(MBB&: *PredBB); |
795 | |
796 | if (!PredBB->isSuccessor(MBB: NewTarget)) |
797 | PredBB->replaceSuccessor(Old: TailBB, New: NewTarget); |
798 | else { |
799 | PredBB->removeSuccessor(Succ: TailBB, NormalizeSuccProbs: true); |
800 | assert(PredBB->succ_size() <= 1); |
801 | } |
802 | |
803 | if (PredTBB) |
804 | TII->insertBranch(MBB&: *PredBB, TBB: PredTBB, FBB: PredFBB, Cond: PredCond, DL); |
805 | |
806 | TDBBs.push_back(Elt: PredBB); |
807 | } |
808 | return Changed; |
809 | } |
810 | |
811 | bool TailDuplicator::canTailDuplicate(MachineBasicBlock *TailBB, |
812 | MachineBasicBlock *PredBB) { |
813 | // EH edges are ignored by analyzeBranch. |
814 | if (PredBB->succ_size() > 1) |
815 | return false; |
816 | |
817 | MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr; |
818 | SmallVector<MachineOperand, 4> PredCond; |
819 | if (TII->analyzeBranch(MBB&: *PredBB, TBB&: PredTBB, FBB&: PredFBB, Cond&: PredCond)) |
820 | return false; |
821 | if (!PredCond.empty()) |
822 | return false; |
823 | // FIXME: This is overly conservative; it may be ok to relax this in the |
824 | // future under more specific conditions. If TailBB is an INLINEASM_BR |
825 | // indirect target, we need to see if the edge from PredBB to TailBB is from |
826 | // an INLINEASM_BR in PredBB, and then also if that edge was from the |
827 | // indirect target list, fallthrough/default target, or potentially both. If |
828 | // it's both, TailDuplicator::tailDuplicate will remove the edge, corrupting |
829 | // the successor list in PredBB and predecessor list in TailBB. |
830 | if (TailBB->isInlineAsmBrIndirectTarget()) |
831 | return false; |
832 | return true; |
833 | } |
834 | |
835 | /// If it is profitable, duplicate TailBB's contents in each |
836 | /// of its predecessors. |
837 | /// \p IsSimple result of isSimpleBB |
838 | /// \p TailBB Block to be duplicated. |
839 | /// \p ForcedLayoutPred When non-null, use this block as the layout predecessor |
840 | /// instead of the previous block in MF's order. |
841 | /// \p TDBBs A vector to keep track of all blocks tail-duplicated |
842 | /// into. |
843 | /// \p Copies A vector of copy instructions inserted. Used later to |
844 | /// walk all the inserted copies and remove redundant ones. |
845 | bool TailDuplicator::tailDuplicate(bool IsSimple, MachineBasicBlock *TailBB, |
846 | MachineBasicBlock *ForcedLayoutPred, |
847 | SmallVectorImpl<MachineBasicBlock *> &TDBBs, |
848 | SmallVectorImpl<MachineInstr *> &Copies, |
849 | SmallVectorImpl<MachineBasicBlock *> *CandidatePtr) { |
850 | LLVM_DEBUG(dbgs() << "\n*** Tail-duplicating " << printMBBReference(*TailBB) |
851 | << '\n'); |
852 | |
853 | bool ShouldUpdateTerminators = TailBB->canFallThrough(); |
854 | |
855 | DenseSet<Register> UsedByPhi; |
856 | getRegsUsedByPHIs(BB: *TailBB, UsedByPhi: &UsedByPhi); |
857 | |
858 | if (IsSimple) |
859 | return duplicateSimpleBB(TailBB, TDBBs, UsedByPhi); |
860 | |
861 | // Iterate through all the unique predecessors and tail-duplicate this |
862 | // block into them, if possible. Copying the list ahead of time also |
863 | // avoids trouble with the predecessor list reallocating. |
864 | bool Changed = false; |
865 | SmallSetVector<MachineBasicBlock *, 8> Preds; |
866 | if (CandidatePtr) |
867 | Preds.insert(Start: CandidatePtr->begin(), End: CandidatePtr->end()); |
868 | else |
869 | Preds.insert(Start: TailBB->pred_begin(), End: TailBB->pred_end()); |
870 | |
871 | for (MachineBasicBlock *PredBB : Preds) { |
872 | assert(TailBB != PredBB && |
873 | "Single-block loop should have been rejected earlier!" ); |
874 | |
875 | if (!canTailDuplicate(TailBB, PredBB)) |
876 | continue; |
877 | |
878 | // Don't duplicate into a fall-through predecessor (at least for now). |
879 | // If profile is available, findDuplicateCandidates can choose better |
880 | // fall-through predecessor. |
881 | if (!(MF->getFunction().hasProfileData() && LayoutMode)) { |
882 | bool IsLayoutSuccessor = false; |
883 | if (ForcedLayoutPred) |
884 | IsLayoutSuccessor = (ForcedLayoutPred == PredBB); |
885 | else if (PredBB->isLayoutSuccessor(MBB: TailBB) && PredBB->canFallThrough()) |
886 | IsLayoutSuccessor = true; |
887 | if (IsLayoutSuccessor) |
888 | continue; |
889 | } |
890 | |
891 | LLVM_DEBUG(dbgs() << "\nTail-duplicating into PredBB: " << *PredBB |
892 | << "From Succ: " << *TailBB); |
893 | |
894 | TDBBs.push_back(Elt: PredBB); |
895 | |
896 | // Remove PredBB's unconditional branch. |
897 | TII->removeBranch(MBB&: *PredBB); |
898 | |
899 | // Clone the contents of TailBB into PredBB. |
900 | DenseMap<Register, RegSubRegPair> LocalVRMap; |
901 | SmallVector<std::pair<Register, RegSubRegPair>, 4> CopyInfos; |
902 | for (MachineInstr &MI : llvm::make_early_inc_range(Range&: *TailBB)) { |
903 | if (MI.isPHI()) { |
904 | // Replace the uses of the def of the PHI with the register coming |
905 | // from PredBB. |
906 | processPHI(MI: &MI, TailBB, PredBB, LocalVRMap, Copies&: CopyInfos, RegsUsedByPhi: UsedByPhi, Remove: true); |
907 | } else { |
908 | // Replace def of virtual registers with new registers, and update |
909 | // uses with PHI source register or the new registers. |
910 | duplicateInstruction(MI: &MI, TailBB, PredBB, LocalVRMap, UsedByPhi); |
911 | } |
912 | } |
913 | appendCopies(MBB: PredBB, CopyInfos, Copies); |
914 | |
915 | NumTailDupAdded += TailBB->size() - 1; // subtract one for removed branch |
916 | |
917 | // Update the CFG. |
918 | PredBB->removeSuccessor(I: PredBB->succ_begin()); |
919 | assert(PredBB->succ_empty() && |
920 | "TailDuplicate called on block with multiple successors!" ); |
921 | for (MachineBasicBlock *Succ : TailBB->successors()) |
922 | PredBB->addSuccessor(Succ, Prob: MBPI->getEdgeProbability(Src: TailBB, Dst: Succ)); |
923 | |
924 | // Update branches in pred to jump to tail's layout successor if needed. |
925 | if (ShouldUpdateTerminators) |
926 | PredBB->updateTerminator(PreviousLayoutSuccessor: TailBB->getNextNode()); |
927 | |
928 | Changed = true; |
929 | ++NumTailDups; |
930 | } |
931 | |
932 | // If TailBB was duplicated into all its predecessors except for the prior |
933 | // block, which falls through unconditionally, move the contents of this |
934 | // block into the prior block. |
935 | MachineBasicBlock *PrevBB = ForcedLayoutPred; |
936 | if (!PrevBB) |
937 | PrevBB = &*std::prev(x: TailBB->getIterator()); |
938 | MachineBasicBlock *PriorTBB = nullptr, *PriorFBB = nullptr; |
939 | SmallVector<MachineOperand, 4> PriorCond; |
940 | // This has to check PrevBB->succ_size() because EH edges are ignored by |
941 | // analyzeBranch. |
942 | if (PrevBB->succ_size() == 1 && |
943 | // Layout preds are not always CFG preds. Check. |
944 | *PrevBB->succ_begin() == TailBB && |
945 | !TII->analyzeBranch(MBB&: *PrevBB, TBB&: PriorTBB, FBB&: PriorFBB, Cond&: PriorCond) && |
946 | PriorCond.empty() && |
947 | (!PriorTBB || PriorTBB == TailBB) && |
948 | TailBB->pred_size() == 1 && |
949 | !TailBB->hasAddressTaken()) { |
950 | LLVM_DEBUG(dbgs() << "\nMerging into block: " << *PrevBB |
951 | << "From MBB: " << *TailBB); |
952 | // There may be a branch to the layout successor. This is unlikely but it |
953 | // happens. The correct thing to do is to remove the branch before |
954 | // duplicating the instructions in all cases. |
955 | bool RemovedBranches = TII->removeBranch(MBB&: *PrevBB) != 0; |
956 | |
957 | // If there are still tail instructions, abort the merge |
958 | if (PrevBB->getFirstTerminator() == PrevBB->end()) { |
959 | if (PreRegAlloc) { |
960 | DenseMap<Register, RegSubRegPair> LocalVRMap; |
961 | SmallVector<std::pair<Register, RegSubRegPair>, 4> CopyInfos; |
962 | MachineBasicBlock::iterator I = TailBB->begin(); |
963 | // Process PHI instructions first. |
964 | while (I != TailBB->end() && I->isPHI()) { |
965 | // Replace the uses of the def of the PHI with the register coming |
966 | // from PredBB. |
967 | MachineInstr *MI = &*I++; |
968 | processPHI(MI, TailBB, PredBB: PrevBB, LocalVRMap, Copies&: CopyInfos, RegsUsedByPhi: UsedByPhi, |
969 | Remove: true); |
970 | } |
971 | |
972 | // Now copy the non-PHI instructions. |
973 | while (I != TailBB->end()) { |
974 | // Replace def of virtual registers with new registers, and update |
975 | // uses with PHI source register or the new registers. |
976 | MachineInstr *MI = &*I++; |
977 | assert(!MI->isBundle() && "Not expecting bundles before regalloc!" ); |
978 | duplicateInstruction(MI, TailBB, PredBB: PrevBB, LocalVRMap, UsedByPhi); |
979 | MI->eraseFromParent(); |
980 | } |
981 | appendCopies(MBB: PrevBB, CopyInfos, Copies); |
982 | } else { |
983 | TII->removeBranch(MBB&: *PrevBB); |
984 | // No PHIs to worry about, just splice the instructions over. |
985 | PrevBB->splice(Where: PrevBB->end(), Other: TailBB, From: TailBB->begin(), To: TailBB->end()); |
986 | } |
987 | PrevBB->removeSuccessor(I: PrevBB->succ_begin()); |
988 | assert(PrevBB->succ_empty()); |
989 | PrevBB->transferSuccessors(FromMBB: TailBB); |
990 | |
991 | // Update branches in PrevBB based on Tail's layout successor. |
992 | if (ShouldUpdateTerminators) |
993 | PrevBB->updateTerminator(PreviousLayoutSuccessor: TailBB->getNextNode()); |
994 | |
995 | TDBBs.push_back(Elt: PrevBB); |
996 | Changed = true; |
997 | } else { |
998 | LLVM_DEBUG(dbgs() << "Abort merging blocks, the predecessor still " |
999 | "contains terminator instructions" ); |
1000 | // Return early if no changes were made |
1001 | if (!Changed) |
1002 | return RemovedBranches; |
1003 | } |
1004 | Changed |= RemovedBranches; |
1005 | } |
1006 | |
1007 | // If this is after register allocation, there are no phis to fix. |
1008 | if (!PreRegAlloc) |
1009 | return Changed; |
1010 | |
1011 | // If we made no changes so far, we are safe. |
1012 | if (!Changed) |
1013 | return Changed; |
1014 | |
1015 | // Handle the nasty case in that we duplicated a block that is part of a loop |
1016 | // into some but not all of its predecessors. For example: |
1017 | // 1 -> 2 <-> 3 | |
1018 | // \ | |
1019 | // \---> rest | |
1020 | // if we duplicate 2 into 1 but not into 3, we end up with |
1021 | // 12 -> 3 <-> 2 -> rest | |
1022 | // \ / | |
1023 | // \----->-----/ | |
1024 | // If there was a "var = phi(1, 3)" in 2, it has to be ultimately replaced |
1025 | // with a phi in 3 (which now dominates 2). |
1026 | // What we do here is introduce a copy in 3 of the register defined by the |
1027 | // phi, just like when we are duplicating 2 into 3, but we don't copy any |
1028 | // real instructions or remove the 3 -> 2 edge from the phi in 2. |
1029 | for (MachineBasicBlock *PredBB : Preds) { |
1030 | if (is_contained(Range&: TDBBs, Element: PredBB)) |
1031 | continue; |
1032 | |
1033 | // EH edges |
1034 | if (PredBB->succ_size() != 1) |
1035 | continue; |
1036 | |
1037 | DenseMap<Register, RegSubRegPair> LocalVRMap; |
1038 | SmallVector<std::pair<Register, RegSubRegPair>, 4> CopyInfos; |
1039 | // Process PHI instructions first. |
1040 | for (MachineInstr &MI : make_early_inc_range(Range: TailBB->phis())) { |
1041 | // Replace the uses of the def of the PHI with the register coming |
1042 | // from PredBB. |
1043 | processPHI(MI: &MI, TailBB, PredBB, LocalVRMap, Copies&: CopyInfos, RegsUsedByPhi: UsedByPhi, Remove: false); |
1044 | } |
1045 | appendCopies(MBB: PredBB, CopyInfos, Copies); |
1046 | } |
1047 | |
1048 | return Changed; |
1049 | } |
1050 | |
1051 | /// At the end of the block \p MBB generate COPY instructions between registers |
1052 | /// described by \p CopyInfos. Append resulting instructions to \p Copies. |
1053 | void TailDuplicator::appendCopies(MachineBasicBlock *MBB, |
1054 | SmallVectorImpl<std::pair<Register, RegSubRegPair>> &CopyInfos, |
1055 | SmallVectorImpl<MachineInstr*> &Copies) { |
1056 | MachineBasicBlock::iterator Loc = MBB->getFirstTerminator(); |
1057 | const MCInstrDesc &CopyD = TII->get(Opcode: TargetOpcode::COPY); |
1058 | for (auto &CI : CopyInfos) { |
1059 | auto C = BuildMI(BB&: *MBB, I: Loc, MIMD: DebugLoc(), MCID: CopyD, DestReg: CI.first) |
1060 | .addReg(RegNo: CI.second.Reg, flags: 0, SubReg: CI.second.SubReg); |
1061 | Copies.push_back(Elt: C); |
1062 | } |
1063 | } |
1064 | |
1065 | /// Remove the specified dead machine basic block from the function, updating |
1066 | /// the CFG. |
1067 | void TailDuplicator::removeDeadBlock( |
1068 | MachineBasicBlock *MBB, |
1069 | function_ref<void(MachineBasicBlock *)> *RemovalCallback) { |
1070 | assert(MBB->pred_empty() && "MBB must be dead!" ); |
1071 | LLVM_DEBUG(dbgs() << "\nRemoving MBB: " << *MBB); |
1072 | |
1073 | MachineFunction *MF = MBB->getParent(); |
1074 | // Update the call site info. |
1075 | for (const MachineInstr &MI : *MBB) |
1076 | if (MI.shouldUpdateCallSiteInfo()) |
1077 | MF->eraseCallSiteInfo(MI: &MI); |
1078 | |
1079 | if (RemovalCallback) |
1080 | (*RemovalCallback)(MBB); |
1081 | |
1082 | // Remove all successors. |
1083 | while (!MBB->succ_empty()) |
1084 | MBB->removeSuccessor(I: MBB->succ_end() - 1); |
1085 | |
1086 | // Remove the block. |
1087 | MBB->eraseFromParent(); |
1088 | } |
1089 | |