1 | //===- InstCombinePHI.cpp -------------------------------------------------===// |
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 implements the visitPHINode function. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "InstCombineInternal.h" |
14 | #include "llvm/ADT/STLExtras.h" |
15 | #include "llvm/ADT/SmallPtrSet.h" |
16 | #include "llvm/ADT/Statistic.h" |
17 | #include "llvm/Analysis/InstructionSimplify.h" |
18 | #include "llvm/Analysis/ValueTracking.h" |
19 | #include "llvm/IR/PatternMatch.h" |
20 | #include "llvm/Support/CommandLine.h" |
21 | #include "llvm/Transforms/InstCombine/InstCombiner.h" |
22 | #include "llvm/Transforms/Utils/Local.h" |
23 | #include <optional> |
24 | |
25 | using namespace llvm; |
26 | using namespace llvm::PatternMatch; |
27 | |
28 | #define DEBUG_TYPE "instcombine" |
29 | |
30 | static cl::opt<unsigned> |
31 | MaxNumPhis("instcombine-max-num-phis" , cl::init(Val: 512), |
32 | cl::desc("Maximum number phis to handle in intptr/ptrint folding" )); |
33 | |
34 | STATISTIC(NumPHIsOfInsertValues, |
35 | "Number of phi-of-insertvalue turned into insertvalue-of-phis" ); |
36 | STATISTIC(, |
37 | "Number of phi-of-extractvalue turned into extractvalue-of-phi" ); |
38 | STATISTIC(NumPHICSEs, "Number of PHI's that got CSE'd" ); |
39 | |
40 | /// The PHI arguments will be folded into a single operation with a PHI node |
41 | /// as input. The debug location of the single operation will be the merged |
42 | /// locations of the original PHI node arguments. |
43 | void InstCombinerImpl::PHIArgMergedDebugLoc(Instruction *Inst, PHINode &PN) { |
44 | auto *FirstInst = cast<Instruction>(Val: PN.getIncomingValue(i: 0)); |
45 | Inst->setDebugLoc(FirstInst->getDebugLoc()); |
46 | // We do not expect a CallInst here, otherwise, N-way merging of DebugLoc |
47 | // will be inefficient. |
48 | assert(!isa<CallInst>(Inst)); |
49 | |
50 | for (Value *V : drop_begin(RangeOrContainer: PN.incoming_values())) { |
51 | auto *I = cast<Instruction>(Val: V); |
52 | Inst->applyMergedLocation(LocA: Inst->getDebugLoc(), LocB: I->getDebugLoc()); |
53 | } |
54 | } |
55 | |
56 | /// If the phi is within a phi web, which is formed by the def-use chain |
57 | /// of phis and all the phis in the web are only used in the other phis. |
58 | /// In this case, these phis are dead and we will remove all of them. |
59 | bool InstCombinerImpl::foldDeadPhiWeb(PHINode &PN) { |
60 | SmallVector<PHINode *, 16> Stack; |
61 | SmallPtrSet<PHINode *, 16> Visited; |
62 | Stack.push_back(Elt: &PN); |
63 | while (!Stack.empty()) { |
64 | PHINode *Phi = Stack.pop_back_val(); |
65 | if (!Visited.insert(Ptr: Phi).second) |
66 | continue; |
67 | // Early stop if the set of PHIs is large |
68 | if (Visited.size() == 16) |
69 | return false; |
70 | for (User *Use : Phi->users()) { |
71 | if (PHINode *PhiUse = dyn_cast<PHINode>(Val: Use)) |
72 | Stack.push_back(Elt: PhiUse); |
73 | else |
74 | return false; |
75 | } |
76 | } |
77 | for (PHINode *Phi : Visited) |
78 | replaceInstUsesWith(I&: *Phi, V: PoisonValue::get(T: Phi->getType())); |
79 | for (PHINode *Phi : Visited) |
80 | eraseInstFromFunction(I&: *Phi); |
81 | return true; |
82 | } |
83 | |
84 | // Replace Integer typed PHI PN if the PHI's value is used as a pointer value. |
85 | // If there is an existing pointer typed PHI that produces the same value as PN, |
86 | // replace PN and the IntToPtr operation with it. Otherwise, synthesize a new |
87 | // PHI node: |
88 | // |
89 | // Case-1: |
90 | // bb1: |
91 | // int_init = PtrToInt(ptr_init) |
92 | // br label %bb2 |
93 | // bb2: |
94 | // int_val = PHI([int_init, %bb1], [int_val_inc, %bb2] |
95 | // ptr_val = PHI([ptr_init, %bb1], [ptr_val_inc, %bb2] |
96 | // ptr_val2 = IntToPtr(int_val) |
97 | // ... |
98 | // use(ptr_val2) |
99 | // ptr_val_inc = ... |
100 | // inc_val_inc = PtrToInt(ptr_val_inc) |
101 | // |
102 | // ==> |
103 | // bb1: |
104 | // br label %bb2 |
105 | // bb2: |
106 | // ptr_val = PHI([ptr_init, %bb1], [ptr_val_inc, %bb2] |
107 | // ... |
108 | // use(ptr_val) |
109 | // ptr_val_inc = ... |
110 | // |
111 | // Case-2: |
112 | // bb1: |
113 | // int_ptr = BitCast(ptr_ptr) |
114 | // int_init = Load(int_ptr) |
115 | // br label %bb2 |
116 | // bb2: |
117 | // int_val = PHI([int_init, %bb1], [int_val_inc, %bb2] |
118 | // ptr_val2 = IntToPtr(int_val) |
119 | // ... |
120 | // use(ptr_val2) |
121 | // ptr_val_inc = ... |
122 | // inc_val_inc = PtrToInt(ptr_val_inc) |
123 | // ==> |
124 | // bb1: |
125 | // ptr_init = Load(ptr_ptr) |
126 | // br label %bb2 |
127 | // bb2: |
128 | // ptr_val = PHI([ptr_init, %bb1], [ptr_val_inc, %bb2] |
129 | // ... |
130 | // use(ptr_val) |
131 | // ptr_val_inc = ... |
132 | // ... |
133 | // |
134 | bool InstCombinerImpl::foldIntegerTypedPHI(PHINode &PN) { |
135 | if (!PN.getType()->isIntegerTy()) |
136 | return false; |
137 | if (!PN.hasOneUse()) |
138 | return false; |
139 | |
140 | auto *IntToPtr = dyn_cast<IntToPtrInst>(Val: PN.user_back()); |
141 | if (!IntToPtr) |
142 | return false; |
143 | |
144 | // Check if the pointer is actually used as pointer: |
145 | auto HasPointerUse = [](Instruction *IIP) { |
146 | for (User *U : IIP->users()) { |
147 | Value *Ptr = nullptr; |
148 | if (LoadInst *LoadI = dyn_cast<LoadInst>(Val: U)) { |
149 | Ptr = LoadI->getPointerOperand(); |
150 | } else if (StoreInst *SI = dyn_cast<StoreInst>(Val: U)) { |
151 | Ptr = SI->getPointerOperand(); |
152 | } else if (GetElementPtrInst *GI = dyn_cast<GetElementPtrInst>(Val: U)) { |
153 | Ptr = GI->getPointerOperand(); |
154 | } |
155 | |
156 | if (Ptr && Ptr == IIP) |
157 | return true; |
158 | } |
159 | return false; |
160 | }; |
161 | |
162 | if (!HasPointerUse(IntToPtr)) |
163 | return false; |
164 | |
165 | if (DL.getPointerSizeInBits(AS: IntToPtr->getAddressSpace()) != |
166 | DL.getTypeSizeInBits(Ty: IntToPtr->getOperand(i_nocapture: 0)->getType())) |
167 | return false; |
168 | |
169 | SmallVector<Value *, 4> AvailablePtrVals; |
170 | for (auto Incoming : zip(t: PN.blocks(), u: PN.incoming_values())) { |
171 | BasicBlock *BB = std::get<0>(t&: Incoming); |
172 | Value *Arg = std::get<1>(t&: Incoming); |
173 | |
174 | // Arg could be a constant, constant expr, etc., which we don't cover here. |
175 | if (!isa<Instruction>(Val: Arg) && !isa<Argument>(Val: Arg)) |
176 | return false; |
177 | |
178 | // First look backward: |
179 | if (auto *PI = dyn_cast<PtrToIntInst>(Val: Arg)) { |
180 | AvailablePtrVals.emplace_back(Args: PI->getOperand(i_nocapture: 0)); |
181 | continue; |
182 | } |
183 | |
184 | // Next look forward: |
185 | Value *ArgIntToPtr = nullptr; |
186 | for (User *U : Arg->users()) { |
187 | if (isa<IntToPtrInst>(Val: U) && U->getType() == IntToPtr->getType() && |
188 | (DT.dominates(Def: cast<Instruction>(Val: U), BB) || |
189 | cast<Instruction>(Val: U)->getParent() == BB)) { |
190 | ArgIntToPtr = U; |
191 | break; |
192 | } |
193 | } |
194 | |
195 | if (ArgIntToPtr) { |
196 | AvailablePtrVals.emplace_back(Args&: ArgIntToPtr); |
197 | continue; |
198 | } |
199 | |
200 | // If Arg is defined by a PHI, allow it. This will also create |
201 | // more opportunities iteratively. |
202 | if (isa<PHINode>(Val: Arg)) { |
203 | AvailablePtrVals.emplace_back(Args&: Arg); |
204 | continue; |
205 | } |
206 | |
207 | // For a single use integer load: |
208 | auto *LoadI = dyn_cast<LoadInst>(Val: Arg); |
209 | if (!LoadI) |
210 | return false; |
211 | |
212 | if (!LoadI->hasOneUse()) |
213 | return false; |
214 | |
215 | // Push the integer typed Load instruction into the available |
216 | // value set, and fix it up later when the pointer typed PHI |
217 | // is synthesized. |
218 | AvailablePtrVals.emplace_back(Args&: LoadI); |
219 | } |
220 | |
221 | // Now search for a matching PHI |
222 | auto *BB = PN.getParent(); |
223 | assert(AvailablePtrVals.size() == PN.getNumIncomingValues() && |
224 | "Not enough available ptr typed incoming values" ); |
225 | PHINode *MatchingPtrPHI = nullptr; |
226 | unsigned NumPhis = 0; |
227 | for (PHINode &PtrPHI : BB->phis()) { |
228 | // FIXME: consider handling this in AggressiveInstCombine |
229 | if (NumPhis++ > MaxNumPhis) |
230 | return false; |
231 | if (&PtrPHI == &PN || PtrPHI.getType() != IntToPtr->getType()) |
232 | continue; |
233 | if (any_of(Range: zip(t: PN.blocks(), u&: AvailablePtrVals), |
234 | P: [&](const auto &BlockAndValue) { |
235 | BasicBlock *BB = std::get<0>(BlockAndValue); |
236 | Value *V = std::get<1>(BlockAndValue); |
237 | return PtrPHI.getIncomingValueForBlock(BB) != V; |
238 | })) |
239 | continue; |
240 | MatchingPtrPHI = &PtrPHI; |
241 | break; |
242 | } |
243 | |
244 | if (MatchingPtrPHI) { |
245 | assert(MatchingPtrPHI->getType() == IntToPtr->getType() && |
246 | "Phi's Type does not match with IntToPtr" ); |
247 | // Explicitly replace the inttoptr (rather than inserting a ptrtoint) here, |
248 | // to make sure another transform can't undo it in the meantime. |
249 | replaceInstUsesWith(I&: *IntToPtr, V: MatchingPtrPHI); |
250 | eraseInstFromFunction(I&: *IntToPtr); |
251 | eraseInstFromFunction(I&: PN); |
252 | return true; |
253 | } |
254 | |
255 | // If it requires a conversion for every PHI operand, do not do it. |
256 | if (all_of(Range&: AvailablePtrVals, P: [&](Value *V) { |
257 | return (V->getType() != IntToPtr->getType()) || isa<IntToPtrInst>(Val: V); |
258 | })) |
259 | return false; |
260 | |
261 | // If any of the operand that requires casting is a terminator |
262 | // instruction, do not do it. Similarly, do not do the transform if the value |
263 | // is PHI in a block with no insertion point, for example, a catchswitch |
264 | // block, since we will not be able to insert a cast after the PHI. |
265 | if (any_of(Range&: AvailablePtrVals, P: [&](Value *V) { |
266 | if (V->getType() == IntToPtr->getType()) |
267 | return false; |
268 | auto *Inst = dyn_cast<Instruction>(Val: V); |
269 | if (!Inst) |
270 | return false; |
271 | if (Inst->isTerminator()) |
272 | return true; |
273 | auto *BB = Inst->getParent(); |
274 | if (isa<PHINode>(Val: Inst) && BB->getFirstInsertionPt() == BB->end()) |
275 | return true; |
276 | return false; |
277 | })) |
278 | return false; |
279 | |
280 | PHINode *NewPtrPHI = PHINode::Create( |
281 | Ty: IntToPtr->getType(), NumReservedValues: PN.getNumIncomingValues(), NameStr: PN.getName() + ".ptr" ); |
282 | |
283 | InsertNewInstBefore(New: NewPtrPHI, Old: PN.getIterator()); |
284 | SmallDenseMap<Value *, Instruction *> Casts; |
285 | for (auto Incoming : zip(t: PN.blocks(), u&: AvailablePtrVals)) { |
286 | auto *IncomingBB = std::get<0>(t&: Incoming); |
287 | auto *IncomingVal = std::get<1>(t&: Incoming); |
288 | |
289 | if (IncomingVal->getType() == IntToPtr->getType()) { |
290 | NewPtrPHI->addIncoming(V: IncomingVal, BB: IncomingBB); |
291 | continue; |
292 | } |
293 | |
294 | #ifndef NDEBUG |
295 | LoadInst *LoadI = dyn_cast<LoadInst>(IncomingVal); |
296 | assert((isa<PHINode>(IncomingVal) || |
297 | IncomingVal->getType()->isPointerTy() || |
298 | (LoadI && LoadI->hasOneUse())) && |
299 | "Can not replace LoadInst with multiple uses" ); |
300 | #endif |
301 | // Need to insert a BitCast. |
302 | // For an integer Load instruction with a single use, the load + IntToPtr |
303 | // cast will be simplified into a pointer load: |
304 | // %v = load i64, i64* %a.ip, align 8 |
305 | // %v.cast = inttoptr i64 %v to float ** |
306 | // ==> |
307 | // %v.ptrp = bitcast i64 * %a.ip to float ** |
308 | // %v.cast = load float *, float ** %v.ptrp, align 8 |
309 | Instruction *&CI = Casts[IncomingVal]; |
310 | if (!CI) { |
311 | CI = CastInst::CreateBitOrPointerCast(S: IncomingVal, Ty: IntToPtr->getType(), |
312 | Name: IncomingVal->getName() + ".ptr" ); |
313 | if (auto *IncomingI = dyn_cast<Instruction>(Val: IncomingVal)) { |
314 | BasicBlock::iterator InsertPos(IncomingI); |
315 | InsertPos++; |
316 | BasicBlock *BB = IncomingI->getParent(); |
317 | if (isa<PHINode>(Val: IncomingI)) |
318 | InsertPos = BB->getFirstInsertionPt(); |
319 | assert(InsertPos != BB->end() && "should have checked above" ); |
320 | InsertNewInstBefore(New: CI, Old: InsertPos); |
321 | } else { |
322 | auto *InsertBB = &IncomingBB->getParent()->getEntryBlock(); |
323 | InsertNewInstBefore(New: CI, Old: InsertBB->getFirstInsertionPt()); |
324 | } |
325 | } |
326 | NewPtrPHI->addIncoming(V: CI, BB: IncomingBB); |
327 | } |
328 | |
329 | // Explicitly replace the inttoptr (rather than inserting a ptrtoint) here, |
330 | // to make sure another transform can't undo it in the meantime. |
331 | replaceInstUsesWith(I&: *IntToPtr, V: NewPtrPHI); |
332 | eraseInstFromFunction(I&: *IntToPtr); |
333 | eraseInstFromFunction(I&: PN); |
334 | return true; |
335 | } |
336 | |
337 | // Remove RoundTrip IntToPtr/PtrToInt Cast on PHI-Operand and |
338 | // fold Phi-operand to bitcast. |
339 | Instruction *InstCombinerImpl::foldPHIArgIntToPtrToPHI(PHINode &PN) { |
340 | // convert ptr2int ( phi[ int2ptr(ptr2int(x))] ) --> ptr2int ( phi [ x ] ) |
341 | // Make sure all uses of phi are ptr2int. |
342 | if (!all_of(Range: PN.users(), P: [](User *U) { return isa<PtrToIntInst>(Val: U); })) |
343 | return nullptr; |
344 | |
345 | // Iterating over all operands to check presence of target pointers for |
346 | // optimization. |
347 | bool OperandWithRoundTripCast = false; |
348 | for (unsigned OpNum = 0; OpNum != PN.getNumIncomingValues(); ++OpNum) { |
349 | if (auto *NewOp = |
350 | simplifyIntToPtrRoundTripCast(Val: PN.getIncomingValue(i: OpNum))) { |
351 | replaceOperand(I&: PN, OpNum, V: NewOp); |
352 | OperandWithRoundTripCast = true; |
353 | } |
354 | } |
355 | if (!OperandWithRoundTripCast) |
356 | return nullptr; |
357 | return &PN; |
358 | } |
359 | |
360 | /// If we have something like phi [insertvalue(a,b,0), insertvalue(c,d,0)], |
361 | /// turn this into a phi[a,c] and phi[b,d] and a single insertvalue. |
362 | Instruction * |
363 | InstCombinerImpl::foldPHIArgInsertValueInstructionIntoPHI(PHINode &PN) { |
364 | auto *FirstIVI = cast<InsertValueInst>(Val: PN.getIncomingValue(i: 0)); |
365 | |
366 | // Scan to see if all operands are `insertvalue`'s with the same indices, |
367 | // and all have a single use. |
368 | for (Value *V : drop_begin(RangeOrContainer: PN.incoming_values())) { |
369 | auto *I = dyn_cast<InsertValueInst>(Val: V); |
370 | if (!I || !I->hasOneUser() || I->getIndices() != FirstIVI->getIndices()) |
371 | return nullptr; |
372 | } |
373 | |
374 | // For each operand of an `insertvalue` |
375 | std::array<PHINode *, 2> NewOperands; |
376 | for (int OpIdx : {0, 1}) { |
377 | auto *&NewOperand = NewOperands[OpIdx]; |
378 | // Create a new PHI node to receive the values the operand has in each |
379 | // incoming basic block. |
380 | NewOperand = PHINode::Create( |
381 | Ty: FirstIVI->getOperand(i_nocapture: OpIdx)->getType(), NumReservedValues: PN.getNumIncomingValues(), |
382 | NameStr: FirstIVI->getOperand(i_nocapture: OpIdx)->getName() + ".pn" ); |
383 | // And populate each operand's PHI with said values. |
384 | for (auto Incoming : zip(t: PN.blocks(), u: PN.incoming_values())) |
385 | NewOperand->addIncoming( |
386 | V: cast<InsertValueInst>(Val&: std::get<1>(t&: Incoming))->getOperand(i_nocapture: OpIdx), |
387 | BB: std::get<0>(t&: Incoming)); |
388 | InsertNewInstBefore(New: NewOperand, Old: PN.getIterator()); |
389 | } |
390 | |
391 | // And finally, create `insertvalue` over the newly-formed PHI nodes. |
392 | auto *NewIVI = InsertValueInst::Create(Agg: NewOperands[0], Val: NewOperands[1], |
393 | Idxs: FirstIVI->getIndices(), NameStr: PN.getName()); |
394 | |
395 | PHIArgMergedDebugLoc(Inst: NewIVI, PN); |
396 | ++NumPHIsOfInsertValues; |
397 | return NewIVI; |
398 | } |
399 | |
400 | /// If we have something like phi [extractvalue(a,0), extractvalue(b,0)], |
401 | /// turn this into a phi[a,b] and a single extractvalue. |
402 | Instruction * |
403 | InstCombinerImpl::(PHINode &PN) { |
404 | auto *FirstEVI = cast<ExtractValueInst>(Val: PN.getIncomingValue(i: 0)); |
405 | |
406 | // Scan to see if all operands are `extractvalue`'s with the same indices, |
407 | // and all have a single use. |
408 | for (Value *V : drop_begin(RangeOrContainer: PN.incoming_values())) { |
409 | auto *I = dyn_cast<ExtractValueInst>(Val: V); |
410 | if (!I || !I->hasOneUser() || I->getIndices() != FirstEVI->getIndices() || |
411 | I->getAggregateOperand()->getType() != |
412 | FirstEVI->getAggregateOperand()->getType()) |
413 | return nullptr; |
414 | } |
415 | |
416 | // Create a new PHI node to receive the values the aggregate operand has |
417 | // in each incoming basic block. |
418 | auto *NewAggregateOperand = PHINode::Create( |
419 | Ty: FirstEVI->getAggregateOperand()->getType(), NumReservedValues: PN.getNumIncomingValues(), |
420 | NameStr: FirstEVI->getAggregateOperand()->getName() + ".pn" ); |
421 | // And populate the PHI with said values. |
422 | for (auto Incoming : zip(t: PN.blocks(), u: PN.incoming_values())) |
423 | NewAggregateOperand->addIncoming( |
424 | V: cast<ExtractValueInst>(Val&: std::get<1>(t&: Incoming))->getAggregateOperand(), |
425 | BB: std::get<0>(t&: Incoming)); |
426 | InsertNewInstBefore(New: NewAggregateOperand, Old: PN.getIterator()); |
427 | |
428 | // And finally, create `extractvalue` over the newly-formed PHI nodes. |
429 | auto *NewEVI = ExtractValueInst::Create(Agg: NewAggregateOperand, |
430 | Idxs: FirstEVI->getIndices(), NameStr: PN.getName()); |
431 | |
432 | PHIArgMergedDebugLoc(Inst: NewEVI, PN); |
433 | ++NumPHIsOfExtractValues; |
434 | return NewEVI; |
435 | } |
436 | |
437 | /// If we have something like phi [add (a,b), add(a,c)] and if a/b/c and the |
438 | /// adds all have a single user, turn this into a phi and a single binop. |
439 | Instruction *InstCombinerImpl::foldPHIArgBinOpIntoPHI(PHINode &PN) { |
440 | Instruction *FirstInst = cast<Instruction>(Val: PN.getIncomingValue(i: 0)); |
441 | assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)); |
442 | unsigned Opc = FirstInst->getOpcode(); |
443 | Value *LHSVal = FirstInst->getOperand(i: 0); |
444 | Value *RHSVal = FirstInst->getOperand(i: 1); |
445 | |
446 | Type *LHSType = LHSVal->getType(); |
447 | Type *RHSType = RHSVal->getType(); |
448 | |
449 | // Scan to see if all operands are the same opcode, and all have one user. |
450 | for (Value *V : drop_begin(RangeOrContainer: PN.incoming_values())) { |
451 | Instruction *I = dyn_cast<Instruction>(Val: V); |
452 | if (!I || I->getOpcode() != Opc || !I->hasOneUser() || |
453 | // Verify type of the LHS matches so we don't fold cmp's of different |
454 | // types. |
455 | I->getOperand(i: 0)->getType() != LHSType || |
456 | I->getOperand(i: 1)->getType() != RHSType) |
457 | return nullptr; |
458 | |
459 | // If they are CmpInst instructions, check their predicates |
460 | if (CmpInst *CI = dyn_cast<CmpInst>(Val: I)) |
461 | if (CI->getPredicate() != cast<CmpInst>(Val: FirstInst)->getPredicate()) |
462 | return nullptr; |
463 | |
464 | // Keep track of which operand needs a phi node. |
465 | if (I->getOperand(i: 0) != LHSVal) LHSVal = nullptr; |
466 | if (I->getOperand(i: 1) != RHSVal) RHSVal = nullptr; |
467 | } |
468 | |
469 | // If both LHS and RHS would need a PHI, don't do this transformation, |
470 | // because it would increase the number of PHIs entering the block, |
471 | // which leads to higher register pressure. This is especially |
472 | // bad when the PHIs are in the header of a loop. |
473 | if (!LHSVal && !RHSVal) |
474 | return nullptr; |
475 | |
476 | // Otherwise, this is safe to transform! |
477 | |
478 | Value *InLHS = FirstInst->getOperand(i: 0); |
479 | Value *InRHS = FirstInst->getOperand(i: 1); |
480 | PHINode *NewLHS = nullptr, *NewRHS = nullptr; |
481 | if (!LHSVal) { |
482 | NewLHS = PHINode::Create(Ty: LHSType, NumReservedValues: PN.getNumIncomingValues(), |
483 | NameStr: FirstInst->getOperand(i: 0)->getName() + ".pn" ); |
484 | NewLHS->addIncoming(V: InLHS, BB: PN.getIncomingBlock(i: 0)); |
485 | InsertNewInstBefore(New: NewLHS, Old: PN.getIterator()); |
486 | LHSVal = NewLHS; |
487 | } |
488 | |
489 | if (!RHSVal) { |
490 | NewRHS = PHINode::Create(Ty: RHSType, NumReservedValues: PN.getNumIncomingValues(), |
491 | NameStr: FirstInst->getOperand(i: 1)->getName() + ".pn" ); |
492 | NewRHS->addIncoming(V: InRHS, BB: PN.getIncomingBlock(i: 0)); |
493 | InsertNewInstBefore(New: NewRHS, Old: PN.getIterator()); |
494 | RHSVal = NewRHS; |
495 | } |
496 | |
497 | // Add all operands to the new PHIs. |
498 | if (NewLHS || NewRHS) { |
499 | for (auto Incoming : drop_begin(RangeOrContainer: zip(t: PN.blocks(), u: PN.incoming_values()))) { |
500 | BasicBlock *InBB = std::get<0>(t&: Incoming); |
501 | Value *InVal = std::get<1>(t&: Incoming); |
502 | Instruction *InInst = cast<Instruction>(Val: InVal); |
503 | if (NewLHS) { |
504 | Value *NewInLHS = InInst->getOperand(i: 0); |
505 | NewLHS->addIncoming(V: NewInLHS, BB: InBB); |
506 | } |
507 | if (NewRHS) { |
508 | Value *NewInRHS = InInst->getOperand(i: 1); |
509 | NewRHS->addIncoming(V: NewInRHS, BB: InBB); |
510 | } |
511 | } |
512 | } |
513 | |
514 | if (CmpInst *CIOp = dyn_cast<CmpInst>(Val: FirstInst)) { |
515 | CmpInst *NewCI = CmpInst::Create(Op: CIOp->getOpcode(), Pred: CIOp->getPredicate(), |
516 | S1: LHSVal, S2: RHSVal); |
517 | PHIArgMergedDebugLoc(Inst: NewCI, PN); |
518 | return NewCI; |
519 | } |
520 | |
521 | BinaryOperator *BinOp = cast<BinaryOperator>(Val: FirstInst); |
522 | BinaryOperator *NewBinOp = |
523 | BinaryOperator::Create(Op: BinOp->getOpcode(), S1: LHSVal, S2: RHSVal); |
524 | |
525 | NewBinOp->copyIRFlags(V: PN.getIncomingValue(i: 0)); |
526 | |
527 | for (Value *V : drop_begin(RangeOrContainer: PN.incoming_values())) |
528 | NewBinOp->andIRFlags(V); |
529 | |
530 | PHIArgMergedDebugLoc(Inst: NewBinOp, PN); |
531 | return NewBinOp; |
532 | } |
533 | |
534 | Instruction *InstCombinerImpl::foldPHIArgGEPIntoPHI(PHINode &PN) { |
535 | GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(Val: PN.getIncomingValue(i: 0)); |
536 | |
537 | SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(), |
538 | FirstInst->op_end()); |
539 | // This is true if all GEP bases are allocas and if all indices into them are |
540 | // constants. |
541 | bool AllBasePointersAreAllocas = true; |
542 | |
543 | // We don't want to replace this phi if the replacement would require |
544 | // more than one phi, which leads to higher register pressure. This is |
545 | // especially bad when the PHIs are in the header of a loop. |
546 | bool NeededPhi = false; |
547 | |
548 | // Remember flags of the first phi-operand getelementptr. |
549 | GEPNoWrapFlags NW = FirstInst->getNoWrapFlags(); |
550 | |
551 | // Scan to see if all operands are the same opcode, and all have one user. |
552 | for (Value *V : drop_begin(RangeOrContainer: PN.incoming_values())) { |
553 | GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Val: V); |
554 | if (!GEP || !GEP->hasOneUser() || |
555 | GEP->getSourceElementType() != FirstInst->getSourceElementType() || |
556 | GEP->getNumOperands() != FirstInst->getNumOperands()) |
557 | return nullptr; |
558 | |
559 | NW &= GEP->getNoWrapFlags(); |
560 | |
561 | // Keep track of whether or not all GEPs are of alloca pointers. |
562 | if (AllBasePointersAreAllocas && |
563 | (!isa<AllocaInst>(Val: GEP->getOperand(i_nocapture: 0)) || |
564 | !GEP->hasAllConstantIndices())) |
565 | AllBasePointersAreAllocas = false; |
566 | |
567 | // Compare the operand lists. |
568 | for (unsigned Op = 0, E = FirstInst->getNumOperands(); Op != E; ++Op) { |
569 | if (FirstInst->getOperand(i_nocapture: Op) == GEP->getOperand(i_nocapture: Op)) |
570 | continue; |
571 | |
572 | // Don't merge two GEPs when two operands differ (introducing phi nodes) |
573 | // if one of the PHIs has a constant for the index. The index may be |
574 | // substantially cheaper to compute for the constants, so making it a |
575 | // variable index could pessimize the path. This also handles the case |
576 | // for struct indices, which must always be constant. |
577 | if (isa<Constant>(Val: FirstInst->getOperand(i_nocapture: Op)) || |
578 | isa<Constant>(Val: GEP->getOperand(i_nocapture: Op))) |
579 | return nullptr; |
580 | |
581 | if (FirstInst->getOperand(i_nocapture: Op)->getType() != |
582 | GEP->getOperand(i_nocapture: Op)->getType()) |
583 | return nullptr; |
584 | |
585 | // If we already needed a PHI for an earlier operand, and another operand |
586 | // also requires a PHI, we'd be introducing more PHIs than we're |
587 | // eliminating, which increases register pressure on entry to the PHI's |
588 | // block. |
589 | if (NeededPhi) |
590 | return nullptr; |
591 | |
592 | FixedOperands[Op] = nullptr; // Needs a PHI. |
593 | NeededPhi = true; |
594 | } |
595 | } |
596 | |
597 | // If all of the base pointers of the PHI'd GEPs are from allocas, don't |
598 | // bother doing this transformation. At best, this will just save a bit of |
599 | // offset calculation, but all the predecessors will have to materialize the |
600 | // stack address into a register anyway. We'd actually rather *clone* the |
601 | // load up into the predecessors so that we have a load of a gep of an alloca, |
602 | // which can usually all be folded into the load. |
603 | if (AllBasePointersAreAllocas) |
604 | return nullptr; |
605 | |
606 | // Otherwise, this is safe to transform. Insert PHI nodes for each operand |
607 | // that is variable. |
608 | SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size()); |
609 | |
610 | bool HasAnyPHIs = false; |
611 | for (unsigned I = 0, E = FixedOperands.size(); I != E; ++I) { |
612 | if (FixedOperands[I]) |
613 | continue; // operand doesn't need a phi. |
614 | Value *FirstOp = FirstInst->getOperand(i_nocapture: I); |
615 | PHINode *NewPN = |
616 | PHINode::Create(Ty: FirstOp->getType(), NumReservedValues: E, NameStr: FirstOp->getName() + ".pn" ); |
617 | InsertNewInstBefore(New: NewPN, Old: PN.getIterator()); |
618 | |
619 | NewPN->addIncoming(V: FirstOp, BB: PN.getIncomingBlock(i: 0)); |
620 | OperandPhis[I] = NewPN; |
621 | FixedOperands[I] = NewPN; |
622 | HasAnyPHIs = true; |
623 | } |
624 | |
625 | // Add all operands to the new PHIs. |
626 | if (HasAnyPHIs) { |
627 | for (auto Incoming : drop_begin(RangeOrContainer: zip(t: PN.blocks(), u: PN.incoming_values()))) { |
628 | BasicBlock *InBB = std::get<0>(t&: Incoming); |
629 | Value *InVal = std::get<1>(t&: Incoming); |
630 | GetElementPtrInst *InGEP = cast<GetElementPtrInst>(Val: InVal); |
631 | |
632 | for (unsigned Op = 0, E = OperandPhis.size(); Op != E; ++Op) |
633 | if (PHINode *OpPhi = OperandPhis[Op]) |
634 | OpPhi->addIncoming(V: InGEP->getOperand(i_nocapture: Op), BB: InBB); |
635 | } |
636 | } |
637 | |
638 | Value *Base = FixedOperands[0]; |
639 | GetElementPtrInst *NewGEP = |
640 | GetElementPtrInst::Create(PointeeType: FirstInst->getSourceElementType(), Ptr: Base, |
641 | IdxList: ArrayRef(FixedOperands).slice(N: 1), NW); |
642 | PHIArgMergedDebugLoc(Inst: NewGEP, PN); |
643 | return NewGEP; |
644 | } |
645 | |
646 | /// Return true if we know that it is safe to sink the load out of the block |
647 | /// that defines it. This means that it must be obvious the value of the load is |
648 | /// not changed from the point of the load to the end of the block it is in. |
649 | /// |
650 | /// Finally, it is safe, but not profitable, to sink a load targeting a |
651 | /// non-address-taken alloca. Doing so will cause us to not promote the alloca |
652 | /// to a register. |
653 | static bool isSafeAndProfitableToSinkLoad(LoadInst *L) { |
654 | BasicBlock::iterator BBI = L->getIterator(), E = L->getParent()->end(); |
655 | |
656 | for (++BBI; BBI != E; ++BBI) |
657 | if (BBI->mayWriteToMemory()) { |
658 | // Calls that only access inaccessible memory do not block sinking the |
659 | // load. |
660 | if (auto *CB = dyn_cast<CallBase>(Val&: BBI)) |
661 | if (CB->onlyAccessesInaccessibleMemory()) |
662 | continue; |
663 | return false; |
664 | } |
665 | |
666 | // Check for non-address taken alloca. If not address-taken already, it isn't |
667 | // profitable to do this xform. |
668 | if (AllocaInst *AI = dyn_cast<AllocaInst>(Val: L->getOperand(i_nocapture: 0))) { |
669 | bool IsAddressTaken = false; |
670 | for (User *U : AI->users()) { |
671 | if (isa<LoadInst>(Val: U)) continue; |
672 | if (StoreInst *SI = dyn_cast<StoreInst>(Val: U)) { |
673 | // If storing TO the alloca, then the address isn't taken. |
674 | if (SI->getOperand(i_nocapture: 1) == AI) continue; |
675 | } |
676 | IsAddressTaken = true; |
677 | break; |
678 | } |
679 | |
680 | if (!IsAddressTaken && AI->isStaticAlloca()) |
681 | return false; |
682 | } |
683 | |
684 | // If this load is a load from a GEP with a constant offset from an alloca, |
685 | // then we don't want to sink it. In its present form, it will be |
686 | // load [constant stack offset]. Sinking it will cause us to have to |
687 | // materialize the stack addresses in each predecessor in a register only to |
688 | // do a shared load from register in the successor. |
689 | if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Val: L->getOperand(i_nocapture: 0))) |
690 | if (AllocaInst *AI = dyn_cast<AllocaInst>(Val: GEP->getOperand(i_nocapture: 0))) |
691 | if (AI->isStaticAlloca() && GEP->hasAllConstantIndices()) |
692 | return false; |
693 | |
694 | return true; |
695 | } |
696 | |
697 | Instruction *InstCombinerImpl::foldPHIArgLoadIntoPHI(PHINode &PN) { |
698 | LoadInst *FirstLI = cast<LoadInst>(Val: PN.getIncomingValue(i: 0)); |
699 | |
700 | // Can't forward swifterror through a phi. |
701 | if (FirstLI->getOperand(i_nocapture: 0)->isSwiftError()) |
702 | return nullptr; |
703 | |
704 | // FIXME: This is overconservative; this transform is allowed in some cases |
705 | // for atomic operations. |
706 | if (FirstLI->isAtomic()) |
707 | return nullptr; |
708 | |
709 | // When processing loads, we need to propagate two bits of information to the |
710 | // sunk load: whether it is volatile, and what its alignment is. |
711 | bool IsVolatile = FirstLI->isVolatile(); |
712 | Align LoadAlignment = FirstLI->getAlign(); |
713 | const unsigned LoadAddrSpace = FirstLI->getPointerAddressSpace(); |
714 | |
715 | // We can't sink the load if the loaded value could be modified between the |
716 | // load and the PHI. |
717 | if (FirstLI->getParent() != PN.getIncomingBlock(i: 0) || |
718 | !isSafeAndProfitableToSinkLoad(L: FirstLI)) |
719 | return nullptr; |
720 | |
721 | // If the PHI is of volatile loads and the load block has multiple |
722 | // successors, sinking it would remove a load of the volatile value from |
723 | // the path through the other successor. |
724 | if (IsVolatile && |
725 | FirstLI->getParent()->getTerminator()->getNumSuccessors() != 1) |
726 | return nullptr; |
727 | |
728 | for (auto Incoming : drop_begin(RangeOrContainer: zip(t: PN.blocks(), u: PN.incoming_values()))) { |
729 | BasicBlock *InBB = std::get<0>(t&: Incoming); |
730 | Value *InVal = std::get<1>(t&: Incoming); |
731 | LoadInst *LI = dyn_cast<LoadInst>(Val: InVal); |
732 | if (!LI || !LI->hasOneUser() || LI->isAtomic()) |
733 | return nullptr; |
734 | |
735 | // Make sure all arguments are the same type of operation. |
736 | if (LI->isVolatile() != IsVolatile || |
737 | LI->getPointerAddressSpace() != LoadAddrSpace) |
738 | return nullptr; |
739 | |
740 | // Can't forward swifterror through a phi. |
741 | if (LI->getOperand(i_nocapture: 0)->isSwiftError()) |
742 | return nullptr; |
743 | |
744 | // We can't sink the load if the loaded value could be modified between |
745 | // the load and the PHI. |
746 | if (LI->getParent() != InBB || !isSafeAndProfitableToSinkLoad(L: LI)) |
747 | return nullptr; |
748 | |
749 | LoadAlignment = std::min(a: LoadAlignment, b: LI->getAlign()); |
750 | |
751 | // If the PHI is of volatile loads and the load block has multiple |
752 | // successors, sinking it would remove a load of the volatile value from |
753 | // the path through the other successor. |
754 | if (IsVolatile && LI->getParent()->getTerminator()->getNumSuccessors() != 1) |
755 | return nullptr; |
756 | } |
757 | |
758 | // Okay, they are all the same operation. Create a new PHI node of the |
759 | // correct type, and PHI together all of the LHS's of the instructions. |
760 | PHINode *NewPN = PHINode::Create(Ty: FirstLI->getOperand(i_nocapture: 0)->getType(), |
761 | NumReservedValues: PN.getNumIncomingValues(), |
762 | NameStr: PN.getName()+".in" ); |
763 | |
764 | Value *InVal = FirstLI->getOperand(i_nocapture: 0); |
765 | NewPN->addIncoming(V: InVal, BB: PN.getIncomingBlock(i: 0)); |
766 | LoadInst *NewLI = |
767 | new LoadInst(FirstLI->getType(), NewPN, "" , IsVolatile, LoadAlignment); |
768 | NewLI->copyMetadata(SrcInst: *FirstLI); |
769 | |
770 | // Add all operands to the new PHI and combine TBAA metadata. |
771 | for (auto Incoming : drop_begin(RangeOrContainer: zip(t: PN.blocks(), u: PN.incoming_values()))) { |
772 | BasicBlock *BB = std::get<0>(t&: Incoming); |
773 | Value *V = std::get<1>(t&: Incoming); |
774 | LoadInst *LI = cast<LoadInst>(Val: V); |
775 | combineMetadataForCSE(K: NewLI, J: LI, DoesKMove: true); |
776 | Value *NewInVal = LI->getOperand(i_nocapture: 0); |
777 | if (NewInVal != InVal) |
778 | InVal = nullptr; |
779 | NewPN->addIncoming(V: NewInVal, BB); |
780 | } |
781 | |
782 | if (InVal) { |
783 | // The new PHI unions all of the same values together. This is really |
784 | // common, so we handle it intelligently here for compile-time speed. |
785 | NewLI->setOperand(i_nocapture: 0, Val_nocapture: InVal); |
786 | delete NewPN; |
787 | } else { |
788 | InsertNewInstBefore(New: NewPN, Old: PN.getIterator()); |
789 | } |
790 | |
791 | // If this was a volatile load that we are merging, make sure to loop through |
792 | // and mark all the input loads as non-volatile. If we don't do this, we will |
793 | // insert a new volatile load and the old ones will not be deletable. |
794 | if (IsVolatile) |
795 | for (Value *IncValue : PN.incoming_values()) |
796 | cast<LoadInst>(Val: IncValue)->setVolatile(false); |
797 | |
798 | PHIArgMergedDebugLoc(Inst: NewLI, PN); |
799 | return NewLI; |
800 | } |
801 | |
802 | /// TODO: This function could handle other cast types, but then it might |
803 | /// require special-casing a cast from the 'i1' type. See the comment in |
804 | /// FoldPHIArgOpIntoPHI() about pessimizing illegal integer types. |
805 | Instruction *InstCombinerImpl::foldPHIArgZextsIntoPHI(PHINode &Phi) { |
806 | // We cannot create a new instruction after the PHI if the terminator is an |
807 | // EHPad because there is no valid insertion point. |
808 | if (Instruction *TI = Phi.getParent()->getTerminator()) |
809 | if (TI->isEHPad()) |
810 | return nullptr; |
811 | |
812 | // Early exit for the common case of a phi with two operands. These are |
813 | // handled elsewhere. See the comment below where we check the count of zexts |
814 | // and constants for more details. |
815 | unsigned NumIncomingValues = Phi.getNumIncomingValues(); |
816 | if (NumIncomingValues < 3) |
817 | return nullptr; |
818 | |
819 | // Find the narrower type specified by the first zext. |
820 | Type *NarrowType = nullptr; |
821 | for (Value *V : Phi.incoming_values()) { |
822 | if (auto *Zext = dyn_cast<ZExtInst>(Val: V)) { |
823 | NarrowType = Zext->getSrcTy(); |
824 | break; |
825 | } |
826 | } |
827 | if (!NarrowType) |
828 | return nullptr; |
829 | |
830 | // Walk the phi operands checking that we only have zexts or constants that |
831 | // we can shrink for free. Store the new operands for the new phi. |
832 | SmallVector<Value *, 4> NewIncoming; |
833 | unsigned NumZexts = 0; |
834 | unsigned NumConsts = 0; |
835 | for (Value *V : Phi.incoming_values()) { |
836 | if (auto *Zext = dyn_cast<ZExtInst>(Val: V)) { |
837 | // All zexts must be identical and have one user. |
838 | if (Zext->getSrcTy() != NarrowType || !Zext->hasOneUser()) |
839 | return nullptr; |
840 | NewIncoming.push_back(Elt: Zext->getOperand(i_nocapture: 0)); |
841 | NumZexts++; |
842 | } else if (auto *C = dyn_cast<Constant>(Val: V)) { |
843 | // Make sure that constants can fit in the new type. |
844 | Constant *Trunc = getLosslessUnsignedTrunc(C, TruncTy: NarrowType); |
845 | if (!Trunc) |
846 | return nullptr; |
847 | NewIncoming.push_back(Elt: Trunc); |
848 | NumConsts++; |
849 | } else { |
850 | // If it's not a cast or a constant, bail out. |
851 | return nullptr; |
852 | } |
853 | } |
854 | |
855 | // The more common cases of a phi with no constant operands or just one |
856 | // variable operand are handled by FoldPHIArgOpIntoPHI() and foldOpIntoPhi() |
857 | // respectively. foldOpIntoPhi() wants to do the opposite transform that is |
858 | // performed here. It tries to replicate a cast in the phi operand's basic |
859 | // block to expose other folding opportunities. Thus, InstCombine will |
860 | // infinite loop without this check. |
861 | if (NumConsts == 0 || NumZexts < 2) |
862 | return nullptr; |
863 | |
864 | // All incoming values are zexts or constants that are safe to truncate. |
865 | // Create a new phi node of the narrow type, phi together all of the new |
866 | // operands, and zext the result back to the original type. |
867 | PHINode *NewPhi = PHINode::Create(Ty: NarrowType, NumReservedValues: NumIncomingValues, |
868 | NameStr: Phi.getName() + ".shrunk" ); |
869 | for (unsigned I = 0; I != NumIncomingValues; ++I) |
870 | NewPhi->addIncoming(V: NewIncoming[I], BB: Phi.getIncomingBlock(i: I)); |
871 | |
872 | InsertNewInstBefore(New: NewPhi, Old: Phi.getIterator()); |
873 | auto *CI = CastInst::CreateZExtOrBitCast(S: NewPhi, Ty: Phi.getType()); |
874 | |
875 | // We use a dropped location here because the new ZExt is necessarily a merge |
876 | // of ZExtInsts and at least one constant from incoming branches; the presence |
877 | // of the constant means we have no viable DebugLoc from that branch, and |
878 | // therefore we must use a dropped location. |
879 | CI->setDebugLoc(DebugLoc::getDropped()); |
880 | return CI; |
881 | } |
882 | |
883 | /// If all operands to a PHI node are the same "unary" operator and they all are |
884 | /// only used by the PHI, PHI together their inputs, and do the operation once, |
885 | /// to the result of the PHI. |
886 | Instruction *InstCombinerImpl::foldPHIArgOpIntoPHI(PHINode &PN) { |
887 | // We cannot create a new instruction after the PHI if the terminator is an |
888 | // EHPad because there is no valid insertion point. |
889 | if (Instruction *TI = PN.getParent()->getTerminator()) |
890 | if (TI->isEHPad()) |
891 | return nullptr; |
892 | |
893 | Instruction *FirstInst = cast<Instruction>(Val: PN.getIncomingValue(i: 0)); |
894 | |
895 | if (isa<GetElementPtrInst>(Val: FirstInst)) |
896 | return foldPHIArgGEPIntoPHI(PN); |
897 | if (isa<LoadInst>(Val: FirstInst)) |
898 | return foldPHIArgLoadIntoPHI(PN); |
899 | if (isa<InsertValueInst>(Val: FirstInst)) |
900 | return foldPHIArgInsertValueInstructionIntoPHI(PN); |
901 | if (isa<ExtractValueInst>(Val: FirstInst)) |
902 | return foldPHIArgExtractValueInstructionIntoPHI(PN); |
903 | |
904 | // Scan the instruction, looking for input operations that can be folded away. |
905 | // If all input operands to the phi are the same instruction (e.g. a cast from |
906 | // the same type or "+42") we can pull the operation through the PHI, reducing |
907 | // code size and simplifying code. |
908 | Constant *ConstantOp = nullptr; |
909 | Type *CastSrcTy = nullptr; |
910 | |
911 | if (isa<CastInst>(Val: FirstInst)) { |
912 | CastSrcTy = FirstInst->getOperand(i: 0)->getType(); |
913 | |
914 | // Be careful about transforming integer PHIs. We don't want to pessimize |
915 | // the code by turning an i32 into an i1293. |
916 | if (PN.getType()->isIntegerTy() && CastSrcTy->isIntegerTy()) { |
917 | if (!shouldChangeType(From: PN.getType(), To: CastSrcTy)) |
918 | return nullptr; |
919 | } |
920 | } else if (isa<BinaryOperator>(Val: FirstInst) || isa<CmpInst>(Val: FirstInst)) { |
921 | // Can fold binop, compare or shift here if the RHS is a constant, |
922 | // otherwise call FoldPHIArgBinOpIntoPHI. |
923 | ConstantOp = dyn_cast<Constant>(Val: FirstInst->getOperand(i: 1)); |
924 | if (!ConstantOp) |
925 | return foldPHIArgBinOpIntoPHI(PN); |
926 | } else { |
927 | return nullptr; // Cannot fold this operation. |
928 | } |
929 | |
930 | // Check to see if all arguments are the same operation. |
931 | for (Value *V : drop_begin(RangeOrContainer: PN.incoming_values())) { |
932 | Instruction *I = dyn_cast<Instruction>(Val: V); |
933 | if (!I || !I->hasOneUser() || !I->isSameOperationAs(I: FirstInst)) |
934 | return nullptr; |
935 | if (CastSrcTy) { |
936 | if (I->getOperand(i: 0)->getType() != CastSrcTy) |
937 | return nullptr; // Cast operation must match. |
938 | } else if (I->getOperand(i: 1) != ConstantOp) { |
939 | return nullptr; |
940 | } |
941 | } |
942 | |
943 | // Okay, they are all the same operation. Create a new PHI node of the |
944 | // correct type, and PHI together all of the LHS's of the instructions. |
945 | PHINode *NewPN = PHINode::Create(Ty: FirstInst->getOperand(i: 0)->getType(), |
946 | NumReservedValues: PN.getNumIncomingValues(), |
947 | NameStr: PN.getName()+".in" ); |
948 | |
949 | Value *InVal = FirstInst->getOperand(i: 0); |
950 | NewPN->addIncoming(V: InVal, BB: PN.getIncomingBlock(i: 0)); |
951 | |
952 | // Add all operands to the new PHI. |
953 | for (auto Incoming : drop_begin(RangeOrContainer: zip(t: PN.blocks(), u: PN.incoming_values()))) { |
954 | BasicBlock *BB = std::get<0>(t&: Incoming); |
955 | Value *V = std::get<1>(t&: Incoming); |
956 | Value *NewInVal = cast<Instruction>(Val: V)->getOperand(i: 0); |
957 | if (NewInVal != InVal) |
958 | InVal = nullptr; |
959 | NewPN->addIncoming(V: NewInVal, BB); |
960 | } |
961 | |
962 | Value *PhiVal; |
963 | if (InVal) { |
964 | // The new PHI unions all of the same values together. This is really |
965 | // common, so we handle it intelligently here for compile-time speed. |
966 | PhiVal = InVal; |
967 | delete NewPN; |
968 | } else { |
969 | InsertNewInstBefore(New: NewPN, Old: PN.getIterator()); |
970 | PhiVal = NewPN; |
971 | } |
972 | |
973 | // Insert and return the new operation. |
974 | if (CastInst *FirstCI = dyn_cast<CastInst>(Val: FirstInst)) { |
975 | CastInst *NewCI = CastInst::Create(FirstCI->getOpcode(), S: PhiVal, |
976 | Ty: PN.getType()); |
977 | PHIArgMergedDebugLoc(Inst: NewCI, PN); |
978 | return NewCI; |
979 | } |
980 | |
981 | if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Val: FirstInst)) { |
982 | BinOp = BinaryOperator::Create(Op: BinOp->getOpcode(), S1: PhiVal, S2: ConstantOp); |
983 | BinOp->copyIRFlags(V: PN.getIncomingValue(i: 0)); |
984 | |
985 | for (Value *V : drop_begin(RangeOrContainer: PN.incoming_values())) |
986 | BinOp->andIRFlags(V); |
987 | |
988 | PHIArgMergedDebugLoc(Inst: BinOp, PN); |
989 | return BinOp; |
990 | } |
991 | |
992 | CmpInst *CIOp = cast<CmpInst>(Val: FirstInst); |
993 | CmpInst *NewCI = CmpInst::Create(Op: CIOp->getOpcode(), Pred: CIOp->getPredicate(), |
994 | S1: PhiVal, S2: ConstantOp); |
995 | PHIArgMergedDebugLoc(Inst: NewCI, PN); |
996 | return NewCI; |
997 | } |
998 | |
999 | /// Return true if this phi node is always equal to NonPhiInVal. |
1000 | /// This happens with mutually cyclic phi nodes like: |
1001 | /// z = some value; x = phi (y, z); y = phi (x, z) |
1002 | static bool PHIsEqualValue(PHINode *PN, Value *&NonPhiInVal, |
1003 | SmallPtrSetImpl<PHINode *> &ValueEqualPHIs) { |
1004 | // See if we already saw this PHI node. |
1005 | if (!ValueEqualPHIs.insert(Ptr: PN).second) |
1006 | return true; |
1007 | |
1008 | // Don't scan crazily complex things. |
1009 | if (ValueEqualPHIs.size() == 16) |
1010 | return false; |
1011 | |
1012 | // Scan the operands to see if they are either phi nodes or are equal to |
1013 | // the value. |
1014 | for (Value *Op : PN->incoming_values()) { |
1015 | if (PHINode *OpPN = dyn_cast<PHINode>(Val: Op)) { |
1016 | if (!PHIsEqualValue(PN: OpPN, NonPhiInVal, ValueEqualPHIs)) { |
1017 | if (NonPhiInVal) |
1018 | return false; |
1019 | NonPhiInVal = OpPN; |
1020 | } |
1021 | } else if (Op != NonPhiInVal) |
1022 | return false; |
1023 | } |
1024 | |
1025 | return true; |
1026 | } |
1027 | |
1028 | /// Return an existing non-zero constant if this phi node has one, otherwise |
1029 | /// return constant 1. |
1030 | static ConstantInt *getAnyNonZeroConstInt(PHINode &PN) { |
1031 | assert(isa<IntegerType>(PN.getType()) && "Expect only integer type phi" ); |
1032 | for (Value *V : PN.operands()) |
1033 | if (auto *ConstVA = dyn_cast<ConstantInt>(Val: V)) |
1034 | if (!ConstVA->isZero()) |
1035 | return ConstVA; |
1036 | return ConstantInt::get(Ty: cast<IntegerType>(Val: PN.getType()), V: 1); |
1037 | } |
1038 | |
1039 | namespace { |
1040 | struct PHIUsageRecord { |
1041 | unsigned PHIId; // The ID # of the PHI (something determinstic to sort on) |
1042 | unsigned Shift; // The amount shifted. |
1043 | Instruction *Inst; // The trunc instruction. |
1044 | |
1045 | PHIUsageRecord(unsigned Pn, unsigned Sh, Instruction *User) |
1046 | : PHIId(Pn), Shift(Sh), Inst(User) {} |
1047 | |
1048 | bool operator<(const PHIUsageRecord &RHS) const { |
1049 | if (PHIId < RHS.PHIId) return true; |
1050 | if (PHIId > RHS.PHIId) return false; |
1051 | if (Shift < RHS.Shift) return true; |
1052 | if (Shift > RHS.Shift) return false; |
1053 | return Inst->getType()->getPrimitiveSizeInBits() < |
1054 | RHS.Inst->getType()->getPrimitiveSizeInBits(); |
1055 | } |
1056 | }; |
1057 | |
1058 | struct LoweredPHIRecord { |
1059 | PHINode *PN; // The PHI that was lowered. |
1060 | unsigned Shift; // The amount shifted. |
1061 | unsigned Width; // The width extracted. |
1062 | |
1063 | LoweredPHIRecord(PHINode *Phi, unsigned Sh, Type *Ty) |
1064 | : PN(Phi), Shift(Sh), Width(Ty->getPrimitiveSizeInBits()) {} |
1065 | |
1066 | // Ctor form used by DenseMap. |
1067 | LoweredPHIRecord(PHINode *Phi, unsigned Sh) : PN(Phi), Shift(Sh), Width(0) {} |
1068 | }; |
1069 | } // namespace |
1070 | |
1071 | namespace llvm { |
1072 | template<> |
1073 | struct DenseMapInfo<LoweredPHIRecord> { |
1074 | static inline LoweredPHIRecord getEmptyKey() { |
1075 | return LoweredPHIRecord(nullptr, 0); |
1076 | } |
1077 | static inline LoweredPHIRecord getTombstoneKey() { |
1078 | return LoweredPHIRecord(nullptr, 1); |
1079 | } |
1080 | static unsigned getHashValue(const LoweredPHIRecord &Val) { |
1081 | return DenseMapInfo<PHINode*>::getHashValue(PtrVal: Val.PN) ^ (Val.Shift>>3) ^ |
1082 | (Val.Width>>3); |
1083 | } |
1084 | static bool isEqual(const LoweredPHIRecord &LHS, |
1085 | const LoweredPHIRecord &RHS) { |
1086 | return LHS.PN == RHS.PN && LHS.Shift == RHS.Shift && |
1087 | LHS.Width == RHS.Width; |
1088 | } |
1089 | }; |
1090 | } // namespace llvm |
1091 | |
1092 | |
1093 | /// This is an integer PHI and we know that it has an illegal type: see if it is |
1094 | /// only used by trunc or trunc(lshr) operations. If so, we split the PHI into |
1095 | /// the various pieces being extracted. This sort of thing is introduced when |
1096 | /// SROA promotes an aggregate to large integer values. |
1097 | /// |
1098 | /// TODO: The user of the trunc may be an bitcast to float/double/vector or an |
1099 | /// inttoptr. We should produce new PHIs in the right type. |
1100 | /// |
1101 | Instruction *InstCombinerImpl::SliceUpIllegalIntegerPHI(PHINode &FirstPhi) { |
1102 | // PHIUsers - Keep track of all of the truncated values extracted from a set |
1103 | // of PHIs, along with their offset. These are the things we want to rewrite. |
1104 | SmallVector<PHIUsageRecord, 16> PHIUsers; |
1105 | |
1106 | // PHIs are often mutually cyclic, so we keep track of a whole set of PHI |
1107 | // nodes which are extracted from. PHIsToSlice is a set we use to avoid |
1108 | // revisiting PHIs, PHIsInspected is a ordered list of PHIs that we need to |
1109 | // check the uses of (to ensure they are all extracts). |
1110 | SmallVector<PHINode*, 8> PHIsToSlice; |
1111 | SmallPtrSet<PHINode*, 8> PHIsInspected; |
1112 | |
1113 | PHIsToSlice.push_back(Elt: &FirstPhi); |
1114 | PHIsInspected.insert(Ptr: &FirstPhi); |
1115 | |
1116 | for (unsigned PHIId = 0; PHIId != PHIsToSlice.size(); ++PHIId) { |
1117 | PHINode *PN = PHIsToSlice[PHIId]; |
1118 | |
1119 | // Scan the input list of the PHI. If any input is an invoke, and if the |
1120 | // input is defined in the predecessor, then we won't be split the critical |
1121 | // edge which is required to insert a truncate. Because of this, we have to |
1122 | // bail out. |
1123 | for (auto Incoming : zip(t: PN->blocks(), u: PN->incoming_values())) { |
1124 | BasicBlock *BB = std::get<0>(t&: Incoming); |
1125 | Value *V = std::get<1>(t&: Incoming); |
1126 | InvokeInst *II = dyn_cast<InvokeInst>(Val: V); |
1127 | if (!II) |
1128 | continue; |
1129 | if (II->getParent() != BB) |
1130 | continue; |
1131 | |
1132 | // If we have a phi, and if it's directly in the predecessor, then we have |
1133 | // a critical edge where we need to put the truncate. Since we can't |
1134 | // split the edge in instcombine, we have to bail out. |
1135 | return nullptr; |
1136 | } |
1137 | |
1138 | // If the incoming value is a PHI node before a catchswitch, we cannot |
1139 | // extract the value within that BB because we cannot insert any non-PHI |
1140 | // instructions in the BB. |
1141 | for (auto *Pred : PN->blocks()) |
1142 | if (Pred->getFirstInsertionPt() == Pred->end()) |
1143 | return nullptr; |
1144 | |
1145 | for (User *U : PN->users()) { |
1146 | Instruction *UserI = cast<Instruction>(Val: U); |
1147 | |
1148 | // If the user is a PHI, inspect its uses recursively. |
1149 | if (PHINode *UserPN = dyn_cast<PHINode>(Val: UserI)) { |
1150 | if (PHIsInspected.insert(Ptr: UserPN).second) |
1151 | PHIsToSlice.push_back(Elt: UserPN); |
1152 | continue; |
1153 | } |
1154 | |
1155 | // Truncates are always ok. |
1156 | if (isa<TruncInst>(Val: UserI)) { |
1157 | PHIUsers.push_back(Elt: PHIUsageRecord(PHIId, 0, UserI)); |
1158 | continue; |
1159 | } |
1160 | |
1161 | // Otherwise it must be a lshr which can only be used by one trunc. |
1162 | if (UserI->getOpcode() != Instruction::LShr || |
1163 | !UserI->hasOneUse() || !isa<TruncInst>(Val: UserI->user_back()) || |
1164 | !isa<ConstantInt>(Val: UserI->getOperand(i: 1))) |
1165 | return nullptr; |
1166 | |
1167 | // Bail on out of range shifts. |
1168 | unsigned SizeInBits = UserI->getType()->getScalarSizeInBits(); |
1169 | if (cast<ConstantInt>(Val: UserI->getOperand(i: 1))->getValue().uge(RHS: SizeInBits)) |
1170 | return nullptr; |
1171 | |
1172 | unsigned Shift = cast<ConstantInt>(Val: UserI->getOperand(i: 1))->getZExtValue(); |
1173 | PHIUsers.push_back(Elt: PHIUsageRecord(PHIId, Shift, UserI->user_back())); |
1174 | } |
1175 | } |
1176 | |
1177 | // If we have no users, they must be all self uses, just nuke the PHI. |
1178 | if (PHIUsers.empty()) |
1179 | return replaceInstUsesWith(I&: FirstPhi, V: PoisonValue::get(T: FirstPhi.getType())); |
1180 | |
1181 | // If this phi node is transformable, create new PHIs for all the pieces |
1182 | // extracted out of it. First, sort the users by their offset and size. |
1183 | array_pod_sort(Start: PHIUsers.begin(), End: PHIUsers.end()); |
1184 | |
1185 | LLVM_DEBUG(dbgs() << "SLICING UP PHI: " << FirstPhi << '\n'; |
1186 | for (unsigned I = 1; I != PHIsToSlice.size(); ++I) dbgs() |
1187 | << "AND USER PHI #" << I << ": " << *PHIsToSlice[I] << '\n'); |
1188 | |
1189 | // PredValues - This is a temporary used when rewriting PHI nodes. It is |
1190 | // hoisted out here to avoid construction/destruction thrashing. |
1191 | DenseMap<BasicBlock*, Value*> PredValues; |
1192 | |
1193 | // ExtractedVals - Each new PHI we introduce is saved here so we don't |
1194 | // introduce redundant PHIs. |
1195 | DenseMap<LoweredPHIRecord, PHINode*> ; |
1196 | |
1197 | for (unsigned UserI = 0, UserE = PHIUsers.size(); UserI != UserE; ++UserI) { |
1198 | unsigned PHIId = PHIUsers[UserI].PHIId; |
1199 | PHINode *PN = PHIsToSlice[PHIId]; |
1200 | unsigned Offset = PHIUsers[UserI].Shift; |
1201 | Type *Ty = PHIUsers[UserI].Inst->getType(); |
1202 | |
1203 | PHINode *EltPHI; |
1204 | |
1205 | // If we've already lowered a user like this, reuse the previously lowered |
1206 | // value. |
1207 | if ((EltPHI = ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)]) == nullptr) { |
1208 | |
1209 | // Otherwise, Create the new PHI node for this user. |
1210 | EltPHI = PHINode::Create(Ty, NumReservedValues: PN->getNumIncomingValues(), |
1211 | NameStr: PN->getName() + ".off" + Twine(Offset), |
1212 | InsertBefore: PN->getIterator()); |
1213 | assert(EltPHI->getType() != PN->getType() && |
1214 | "Truncate didn't shrink phi?" ); |
1215 | |
1216 | for (auto Incoming : zip(t: PN->blocks(), u: PN->incoming_values())) { |
1217 | BasicBlock *Pred = std::get<0>(t&: Incoming); |
1218 | Value *InVal = std::get<1>(t&: Incoming); |
1219 | Value *&PredVal = PredValues[Pred]; |
1220 | |
1221 | // If we already have a value for this predecessor, reuse it. |
1222 | if (PredVal) { |
1223 | EltPHI->addIncoming(V: PredVal, BB: Pred); |
1224 | continue; |
1225 | } |
1226 | |
1227 | // Handle the PHI self-reuse case. |
1228 | if (InVal == PN) { |
1229 | PredVal = EltPHI; |
1230 | EltPHI->addIncoming(V: PredVal, BB: Pred); |
1231 | continue; |
1232 | } |
1233 | |
1234 | if (PHINode *InPHI = dyn_cast<PHINode>(Val: PN)) { |
1235 | // If the incoming value was a PHI, and if it was one of the PHIs we |
1236 | // already rewrote it, just use the lowered value. |
1237 | if (Value *Res = ExtractedVals[LoweredPHIRecord(InPHI, Offset, Ty)]) { |
1238 | PredVal = Res; |
1239 | EltPHI->addIncoming(V: PredVal, BB: Pred); |
1240 | continue; |
1241 | } |
1242 | } |
1243 | |
1244 | // Otherwise, do an extract in the predecessor. |
1245 | Builder.SetInsertPoint(Pred->getTerminator()); |
1246 | Value *Res = InVal; |
1247 | if (Offset) |
1248 | Res = Builder.CreateLShr( |
1249 | LHS: Res, RHS: ConstantInt::get(Ty: InVal->getType(), V: Offset), Name: "extract" ); |
1250 | Res = Builder.CreateTrunc(V: Res, DestTy: Ty, Name: "extract.t" ); |
1251 | PredVal = Res; |
1252 | EltPHI->addIncoming(V: Res, BB: Pred); |
1253 | |
1254 | // If the incoming value was a PHI, and if it was one of the PHIs we are |
1255 | // rewriting, we will ultimately delete the code we inserted. This |
1256 | // means we need to revisit that PHI to make sure we extract out the |
1257 | // needed piece. |
1258 | if (PHINode *OldInVal = dyn_cast<PHINode>(Val: InVal)) |
1259 | if (PHIsInspected.count(Ptr: OldInVal)) { |
1260 | unsigned RefPHIId = |
1261 | find(Range&: PHIsToSlice, Val: OldInVal) - PHIsToSlice.begin(); |
1262 | PHIUsers.push_back( |
1263 | Elt: PHIUsageRecord(RefPHIId, Offset, cast<Instruction>(Val: Res))); |
1264 | ++UserE; |
1265 | } |
1266 | } |
1267 | PredValues.clear(); |
1268 | |
1269 | LLVM_DEBUG(dbgs() << " Made element PHI for offset " << Offset << ": " |
1270 | << *EltPHI << '\n'); |
1271 | ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)] = EltPHI; |
1272 | } |
1273 | |
1274 | // Replace the use of this piece with the PHI node. |
1275 | replaceInstUsesWith(I&: *PHIUsers[UserI].Inst, V: EltPHI); |
1276 | } |
1277 | |
1278 | // Replace all the remaining uses of the PHI nodes (self uses and the lshrs) |
1279 | // with poison. |
1280 | Value *Poison = PoisonValue::get(T: FirstPhi.getType()); |
1281 | for (PHINode *PHI : drop_begin(RangeOrContainer&: PHIsToSlice)) |
1282 | replaceInstUsesWith(I&: *PHI, V: Poison); |
1283 | return replaceInstUsesWith(I&: FirstPhi, V: Poison); |
1284 | } |
1285 | |
1286 | static Value *simplifyUsingControlFlow(InstCombiner &Self, PHINode &PN, |
1287 | const DominatorTree &DT) { |
1288 | // Simplify the following patterns: |
1289 | // if (cond) |
1290 | // / \ |
1291 | // ... ... |
1292 | // \ / |
1293 | // phi [true] [false] |
1294 | // and |
1295 | // switch (cond) |
1296 | // case v1: / \ case v2: |
1297 | // ... ... |
1298 | // \ / |
1299 | // phi [v1] [v2] |
1300 | // Make sure all inputs are constants. |
1301 | if (!all_of(Range: PN.operands(), P: [](Value *V) { return isa<ConstantInt>(Val: V); })) |
1302 | return nullptr; |
1303 | |
1304 | BasicBlock *BB = PN.getParent(); |
1305 | // Do not bother with unreachable instructions. |
1306 | if (!DT.isReachableFromEntry(A: BB)) |
1307 | return nullptr; |
1308 | |
1309 | // Determine which value the condition of the idom has for which successor. |
1310 | LLVMContext &Context = PN.getContext(); |
1311 | auto *IDom = DT.getNode(BB)->getIDom()->getBlock(); |
1312 | Value *Cond; |
1313 | SmallDenseMap<ConstantInt *, BasicBlock *, 8> SuccForValue; |
1314 | SmallDenseMap<BasicBlock *, unsigned, 8> SuccCount; |
1315 | auto AddSucc = [&](ConstantInt *C, BasicBlock *Succ) { |
1316 | SuccForValue[C] = Succ; |
1317 | ++SuccCount[Succ]; |
1318 | }; |
1319 | if (auto *BI = dyn_cast<BranchInst>(Val: IDom->getTerminator())) { |
1320 | if (BI->isUnconditional()) |
1321 | return nullptr; |
1322 | |
1323 | Cond = BI->getCondition(); |
1324 | AddSucc(ConstantInt::getTrue(Context), BI->getSuccessor(i: 0)); |
1325 | AddSucc(ConstantInt::getFalse(Context), BI->getSuccessor(i: 1)); |
1326 | } else if (auto *SI = dyn_cast<SwitchInst>(Val: IDom->getTerminator())) { |
1327 | Cond = SI->getCondition(); |
1328 | ++SuccCount[SI->getDefaultDest()]; |
1329 | for (auto Case : SI->cases()) |
1330 | AddSucc(Case.getCaseValue(), Case.getCaseSuccessor()); |
1331 | } else { |
1332 | return nullptr; |
1333 | } |
1334 | |
1335 | if (Cond->getType() != PN.getType()) |
1336 | return nullptr; |
1337 | |
1338 | // Check that edges outgoing from the idom's terminators dominate respective |
1339 | // inputs of the Phi. |
1340 | std::optional<bool> Invert; |
1341 | for (auto Pair : zip(t: PN.incoming_values(), u: PN.blocks())) { |
1342 | auto *Input = cast<ConstantInt>(Val&: std::get<0>(t&: Pair)); |
1343 | BasicBlock *Pred = std::get<1>(t&: Pair); |
1344 | auto IsCorrectInput = [&](ConstantInt *Input) { |
1345 | // The input needs to be dominated by the corresponding edge of the idom. |
1346 | // This edge cannot be a multi-edge, as that would imply that multiple |
1347 | // different condition values follow the same edge. |
1348 | auto It = SuccForValue.find(Val: Input); |
1349 | return It != SuccForValue.end() && SuccCount[It->second] == 1 && |
1350 | DT.dominates(BBE1: BasicBlockEdge(IDom, It->second), |
1351 | BBE2: BasicBlockEdge(Pred, BB)); |
1352 | }; |
1353 | |
1354 | // Depending on the constant, the condition may need to be inverted. |
1355 | bool NeedsInvert; |
1356 | if (IsCorrectInput(Input)) |
1357 | NeedsInvert = false; |
1358 | else if (IsCorrectInput(cast<ConstantInt>(Val: ConstantExpr::getNot(C: Input)))) |
1359 | NeedsInvert = true; |
1360 | else |
1361 | return nullptr; |
1362 | |
1363 | // Make sure the inversion requirement is always the same. |
1364 | if (Invert && *Invert != NeedsInvert) |
1365 | return nullptr; |
1366 | |
1367 | Invert = NeedsInvert; |
1368 | } |
1369 | |
1370 | if (!*Invert) |
1371 | return Cond; |
1372 | |
1373 | // This Phi is actually opposite to branching condition of IDom. We invert |
1374 | // the condition that will potentially open up some opportunities for |
1375 | // sinking. |
1376 | auto InsertPt = BB->getFirstInsertionPt(); |
1377 | if (InsertPt != BB->end()) { |
1378 | Self.Builder.SetInsertPoint(TheBB: &*BB, IP: InsertPt); |
1379 | return Self.Builder.CreateNot(V: Cond); |
1380 | } |
1381 | |
1382 | return nullptr; |
1383 | } |
1384 | |
1385 | // Fold iv = phi(start, iv.next = iv2.next op start) |
1386 | // where iv2 = phi(iv2.start, iv2.next = iv2 + iv2.step) |
1387 | // and iv2.start op start = start |
1388 | // to iv = iv2 op start |
1389 | static Value *foldDependentIVs(PHINode &PN, IRBuilderBase &Builder) { |
1390 | BasicBlock *BB = PN.getParent(); |
1391 | if (PN.getNumIncomingValues() != 2) |
1392 | return nullptr; |
1393 | |
1394 | Value *Start; |
1395 | Instruction *IvNext; |
1396 | BinaryOperator *Iv2Next; |
1397 | auto MatchOuterIV = [&](Value *V1, Value *V2) { |
1398 | if (match(V: V2, P: m_c_BinOp(L: m_Specific(V: V1), R: m_BinOp(I&: Iv2Next))) || |
1399 | match(V: V2, P: m_GEP(Ops: m_Specific(V: V1), Ops: m_BinOp(I&: Iv2Next)))) { |
1400 | Start = V1; |
1401 | IvNext = cast<Instruction>(Val: V2); |
1402 | return true; |
1403 | } |
1404 | return false; |
1405 | }; |
1406 | |
1407 | if (!MatchOuterIV(PN.getIncomingValue(i: 0), PN.getIncomingValue(i: 1)) && |
1408 | !MatchOuterIV(PN.getIncomingValue(i: 1), PN.getIncomingValue(i: 0))) |
1409 | return nullptr; |
1410 | |
1411 | PHINode *Iv2; |
1412 | Value *Iv2Start, *Iv2Step; |
1413 | if (!matchSimpleRecurrence(I: Iv2Next, P&: Iv2, Start&: Iv2Start, Step&: Iv2Step) || |
1414 | Iv2->getParent() != BB) |
1415 | return nullptr; |
1416 | |
1417 | auto *BO = dyn_cast<BinaryOperator>(Val: IvNext); |
1418 | Constant *Identity = |
1419 | BO ? ConstantExpr::getBinOpIdentity(Opcode: BO->getOpcode(), Ty: Iv2Start->getType()) |
1420 | : Constant::getNullValue(Ty: Iv2Start->getType()); |
1421 | if (Iv2Start != Identity) |
1422 | return nullptr; |
1423 | |
1424 | Builder.SetInsertPoint(TheBB: &*BB, IP: BB->getFirstInsertionPt()); |
1425 | if (!BO) { |
1426 | auto *GEP = cast<GEPOperator>(Val: IvNext); |
1427 | return Builder.CreateGEP(Ty: GEP->getSourceElementType(), Ptr: Start, IdxList: Iv2, Name: "" , |
1428 | NW: cast<GEPOperator>(Val: IvNext)->getNoWrapFlags()); |
1429 | } |
1430 | |
1431 | assert(BO->isCommutative() && "Must be commutative" ); |
1432 | Value *Res = Builder.CreateBinOp(Opc: BO->getOpcode(), LHS: Iv2, RHS: Start); |
1433 | cast<Instruction>(Val: Res)->copyIRFlags(V: BO); |
1434 | return Res; |
1435 | } |
1436 | |
1437 | // PHINode simplification |
1438 | // |
1439 | Instruction *InstCombinerImpl::visitPHINode(PHINode &PN) { |
1440 | if (Value *V = simplifyInstruction(I: &PN, Q: SQ.getWithInstruction(I: &PN))) |
1441 | return replaceInstUsesWith(I&: PN, V); |
1442 | |
1443 | if (Instruction *Result = foldPHIArgZextsIntoPHI(Phi&: PN)) |
1444 | return Result; |
1445 | |
1446 | if (Instruction *Result = foldPHIArgIntToPtrToPHI(PN)) |
1447 | return Result; |
1448 | |
1449 | // If all PHI operands are the same operation, pull them through the PHI, |
1450 | // reducing code size. |
1451 | auto *Inst0 = dyn_cast<Instruction>(Val: PN.getIncomingValue(i: 0)); |
1452 | auto *Inst1 = dyn_cast<Instruction>(Val: PN.getIncomingValue(i: 1)); |
1453 | if (Inst0 && Inst1 && Inst0->getOpcode() == Inst1->getOpcode() && |
1454 | Inst0->hasOneUser()) |
1455 | if (Instruction *Result = foldPHIArgOpIntoPHI(PN)) |
1456 | return Result; |
1457 | |
1458 | // If the incoming values are pointer casts of the same original value, |
1459 | // replace the phi with a single cast iff we can insert a non-PHI instruction. |
1460 | if (PN.getType()->isPointerTy() && |
1461 | PN.getParent()->getFirstInsertionPt() != PN.getParent()->end()) { |
1462 | Value *IV0 = PN.getIncomingValue(i: 0); |
1463 | Value *IV0Stripped = IV0->stripPointerCasts(); |
1464 | // Set to keep track of values known to be equal to IV0Stripped after |
1465 | // stripping pointer casts. |
1466 | SmallPtrSet<Value *, 4> CheckedIVs; |
1467 | CheckedIVs.insert(Ptr: IV0); |
1468 | if (IV0 != IV0Stripped && |
1469 | all_of(Range: PN.incoming_values(), P: [&CheckedIVs, IV0Stripped](Value *IV) { |
1470 | return !CheckedIVs.insert(Ptr: IV).second || |
1471 | IV0Stripped == IV->stripPointerCasts(); |
1472 | })) { |
1473 | return CastInst::CreatePointerCast(S: IV0Stripped, Ty: PN.getType()); |
1474 | } |
1475 | } |
1476 | |
1477 | if (foldDeadPhiWeb(PN)) |
1478 | return nullptr; |
1479 | |
1480 | // Optimization when the phi only has one use |
1481 | if (PN.hasOneUse()) { |
1482 | if (foldIntegerTypedPHI(PN)) |
1483 | return nullptr; |
1484 | |
1485 | // If this phi has a single use, and if that use just computes a value for |
1486 | // the next iteration of a loop, delete the phi. This occurs with unused |
1487 | // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this |
1488 | // common case here is good because the only other things that catch this |
1489 | // are induction variable analysis (sometimes) and ADCE, which is only run |
1490 | // late. |
1491 | Instruction *PHIUser = cast<Instruction>(Val: PN.user_back()); |
1492 | if (PHIUser->hasOneUse() && |
1493 | (isa<BinaryOperator>(Val: PHIUser) || isa<UnaryOperator>(Val: PHIUser) || |
1494 | isa<GetElementPtrInst>(Val: PHIUser)) && |
1495 | PHIUser->user_back() == &PN) { |
1496 | return replaceInstUsesWith(I&: PN, V: PoisonValue::get(T: PN.getType())); |
1497 | } |
1498 | } |
1499 | |
1500 | // When a PHI is used only to be compared with zero, it is safe to replace |
1501 | // an incoming value proved as known nonzero with any non-zero constant. |
1502 | // For example, in the code below, the incoming value %v can be replaced |
1503 | // with any non-zero constant based on the fact that the PHI is only used to |
1504 | // be compared with zero and %v is a known non-zero value: |
1505 | // %v = select %cond, 1, 2 |
1506 | // %p = phi [%v, BB] ... |
1507 | // icmp eq, %p, 0 |
1508 | // FIXME: To be simple, handle only integer type for now. |
1509 | // This handles a small number of uses to keep the complexity down, and an |
1510 | // icmp(or(phi)) can equally be replaced with any non-zero constant as the |
1511 | // "or" will only add bits. |
1512 | if (!PN.hasNUsesOrMore(N: 3)) { |
1513 | SmallVector<Instruction *> DropPoisonFlags; |
1514 | bool AllUsesOfPhiEndsInCmp = all_of(Range: PN.users(), P: [&](User *U) { |
1515 | auto *CmpInst = dyn_cast<ICmpInst>(Val: U); |
1516 | if (!CmpInst) { |
1517 | // This is always correct as OR only add bits and we are checking |
1518 | // against 0. |
1519 | if (U->hasOneUse() && match(V: U, P: m_c_Or(L: m_Specific(V: &PN), R: m_Value()))) { |
1520 | DropPoisonFlags.push_back(Elt: cast<Instruction>(Val: U)); |
1521 | CmpInst = dyn_cast<ICmpInst>(Val: U->user_back()); |
1522 | } |
1523 | } |
1524 | if (!CmpInst || !isa<IntegerType>(Val: PN.getType()) || |
1525 | !CmpInst->isEquality() || !match(V: CmpInst->getOperand(i_nocapture: 1), P: m_Zero())) { |
1526 | return false; |
1527 | } |
1528 | return true; |
1529 | }); |
1530 | // All uses of PHI results in a compare with zero. |
1531 | if (AllUsesOfPhiEndsInCmp) { |
1532 | ConstantInt *NonZeroConst = nullptr; |
1533 | bool MadeChange = false; |
1534 | for (unsigned I = 0, E = PN.getNumIncomingValues(); I != E; ++I) { |
1535 | Instruction *CtxI = PN.getIncomingBlock(i: I)->getTerminator(); |
1536 | Value *VA = PN.getIncomingValue(i: I); |
1537 | if (isKnownNonZero(V: VA, Q: getSimplifyQuery().getWithInstruction(I: CtxI))) { |
1538 | if (!NonZeroConst) |
1539 | NonZeroConst = getAnyNonZeroConstInt(PN); |
1540 | if (NonZeroConst != VA) { |
1541 | replaceOperand(I&: PN, OpNum: I, V: NonZeroConst); |
1542 | // The "disjoint" flag may no longer hold after the transform. |
1543 | for (Instruction *I : DropPoisonFlags) |
1544 | I->dropPoisonGeneratingFlags(); |
1545 | MadeChange = true; |
1546 | } |
1547 | } |
1548 | } |
1549 | if (MadeChange) |
1550 | return &PN; |
1551 | } |
1552 | } |
1553 | |
1554 | // We sometimes end up with phi cycles that non-obviously end up being the |
1555 | // same value, for example: |
1556 | // z = some value; x = phi (y, z); y = phi (x, z) |
1557 | // where the phi nodes don't necessarily need to be in the same block. Do a |
1558 | // quick check to see if the PHI node only contains a single non-phi value, if |
1559 | // so, scan to see if the phi cycle is actually equal to that value. If the |
1560 | // phi has no non-phi values then allow the "NonPhiInVal" to be set later if |
1561 | // one of the phis itself does not have a single input. |
1562 | { |
1563 | unsigned InValNo = 0, NumIncomingVals = PN.getNumIncomingValues(); |
1564 | // Scan for the first non-phi operand. |
1565 | while (InValNo != NumIncomingVals && |
1566 | isa<PHINode>(Val: PN.getIncomingValue(i: InValNo))) |
1567 | ++InValNo; |
1568 | |
1569 | Value *NonPhiInVal = |
1570 | InValNo != NumIncomingVals ? PN.getIncomingValue(i: InValNo) : nullptr; |
1571 | |
1572 | // Scan the rest of the operands to see if there are any conflicts, if so |
1573 | // there is no need to recursively scan other phis. |
1574 | if (NonPhiInVal) |
1575 | for (++InValNo; InValNo != NumIncomingVals; ++InValNo) { |
1576 | Value *OpVal = PN.getIncomingValue(i: InValNo); |
1577 | if (OpVal != NonPhiInVal && !isa<PHINode>(Val: OpVal)) |
1578 | break; |
1579 | } |
1580 | |
1581 | // If we scanned over all operands, then we have one unique value plus |
1582 | // phi values. Scan PHI nodes to see if they all merge in each other or |
1583 | // the value. |
1584 | if (InValNo == NumIncomingVals) { |
1585 | SmallPtrSet<PHINode *, 16> ValueEqualPHIs; |
1586 | if (PHIsEqualValue(PN: &PN, NonPhiInVal, ValueEqualPHIs)) |
1587 | return replaceInstUsesWith(I&: PN, V: NonPhiInVal); |
1588 | } |
1589 | } |
1590 | |
1591 | // If there are multiple PHIs, sort their operands so that they all list |
1592 | // the blocks in the same order. This will help identical PHIs be eliminated |
1593 | // by other passes. Other passes shouldn't depend on this for correctness |
1594 | // however. |
1595 | auto Res = PredOrder.try_emplace(Key: PN.getParent()); |
1596 | if (!Res.second) { |
1597 | const auto &Preds = Res.first->second; |
1598 | for (unsigned I = 0, E = PN.getNumIncomingValues(); I != E; ++I) { |
1599 | BasicBlock *BBA = PN.getIncomingBlock(i: I); |
1600 | BasicBlock *BBB = Preds[I]; |
1601 | if (BBA != BBB) { |
1602 | Value *VA = PN.getIncomingValue(i: I); |
1603 | unsigned J = PN.getBasicBlockIndex(BB: BBB); |
1604 | Value *VB = PN.getIncomingValue(i: J); |
1605 | PN.setIncomingBlock(i: I, BB: BBB); |
1606 | PN.setIncomingValue(i: I, V: VB); |
1607 | PN.setIncomingBlock(i: J, BB: BBA); |
1608 | PN.setIncomingValue(i: J, V: VA); |
1609 | // NOTE: Instcombine normally would want us to "return &PN" if we |
1610 | // modified any of the operands of an instruction. However, since we |
1611 | // aren't adding or removing uses (just rearranging them) we don't do |
1612 | // this in this case. |
1613 | } |
1614 | } |
1615 | } else { |
1616 | // Remember the block order of the first encountered phi node. |
1617 | append_range(C&: Res.first->second, R: PN.blocks()); |
1618 | } |
1619 | |
1620 | // Is there an identical PHI node in this basic block? |
1621 | for (PHINode &IdenticalPN : PN.getParent()->phis()) { |
1622 | // Ignore the PHI node itself. |
1623 | if (&IdenticalPN == &PN) |
1624 | continue; |
1625 | // Note that even though we've just canonicalized this PHI, due to the |
1626 | // worklist visitation order, there are no guarantess that *every* PHI |
1627 | // has been canonicalized, so we can't just compare operands ranges. |
1628 | if (!PN.isIdenticalToWhenDefined(I: &IdenticalPN)) |
1629 | continue; |
1630 | // Just use that PHI instead then. |
1631 | ++NumPHICSEs; |
1632 | return replaceInstUsesWith(I&: PN, V: &IdenticalPN); |
1633 | } |
1634 | |
1635 | // If this is an integer PHI and we know that it has an illegal type, see if |
1636 | // it is only used by trunc or trunc(lshr) operations. If so, we split the |
1637 | // PHI into the various pieces being extracted. This sort of thing is |
1638 | // introduced when SROA promotes an aggregate to a single large integer type. |
1639 | if (PN.getType()->isIntegerTy() && |
1640 | !DL.isLegalInteger(Width: PN.getType()->getPrimitiveSizeInBits())) |
1641 | if (Instruction *Res = SliceUpIllegalIntegerPHI(FirstPhi&: PN)) |
1642 | return Res; |
1643 | |
1644 | // Ultimately, try to replace this Phi with a dominating condition. |
1645 | if (auto *V = simplifyUsingControlFlow(Self&: *this, PN, DT)) |
1646 | return replaceInstUsesWith(I&: PN, V); |
1647 | |
1648 | if (Value *Res = foldDependentIVs(PN, Builder)) |
1649 | return replaceInstUsesWith(I&: PN, V: Res); |
1650 | |
1651 | return nullptr; |
1652 | } |
1653 | |