1 //===-- AtomicExpandPass.cpp - Expand atomic instructions -------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file contains a pass (at IR level) to replace atomic instructions with
11 // __atomic_* library calls, or target specific instruction which implement the
12 // same semantics in a way which better fits the target backend. This can
13 // include the use of (intrinsic-based) load-linked/store-conditional loops,
14 // AtomicCmpXchg, or type coercions.
15 //
16 //===----------------------------------------------------------------------===//
17
18 #include "llvm/CodeGen/AtomicExpandUtils.h"
19 #include "llvm/CodeGen/Passes.h"
20 #include "llvm/IR/Function.h"
21 #include "llvm/IR/IRBuilder.h"
22 #include "llvm/IR/InstIterator.h"
23 #include "llvm/IR/Instructions.h"
24 #include "llvm/IR/Intrinsics.h"
25 #include "llvm/IR/Module.h"
26 #include "llvm/Support/Debug.h"
27 #include "llvm/Support/raw_ostream.h"
28 #include "llvm/Target/TargetLowering.h"
29 #include "llvm/Target/TargetMachine.h"
30 #include "llvm/Target/TargetSubtargetInfo.h"
31
32 using namespace llvm;
33
34 #define DEBUG_TYPE "atomic-expand"
35
36 namespace {
37 class AtomicExpand: public FunctionPass {
38 const TargetMachine *TM;
39 const TargetLowering *TLI;
40 public:
41 static char ID; // Pass identification, replacement for typeid
AtomicExpand(const TargetMachine * TM=nullptr)42 explicit AtomicExpand(const TargetMachine *TM = nullptr)
43 : FunctionPass(ID), TM(TM), TLI(nullptr) {
44 initializeAtomicExpandPass(*PassRegistry::getPassRegistry());
45 }
46
47 bool runOnFunction(Function &F) override;
48
49 private:
50 bool bracketInstWithFences(Instruction *I, AtomicOrdering Order,
51 bool IsStore, bool IsLoad);
52 IntegerType *getCorrespondingIntegerType(Type *T, const DataLayout &DL);
53 LoadInst *convertAtomicLoadToIntegerType(LoadInst *LI);
54 bool tryExpandAtomicLoad(LoadInst *LI);
55 bool expandAtomicLoadToLL(LoadInst *LI);
56 bool expandAtomicLoadToCmpXchg(LoadInst *LI);
57 StoreInst *convertAtomicStoreToIntegerType(StoreInst *SI);
58 bool expandAtomicStore(StoreInst *SI);
59 bool tryExpandAtomicRMW(AtomicRMWInst *AI);
60 Value *
61 insertRMWLLSCLoop(IRBuilder<> &Builder, Type *ResultTy, Value *Addr,
62 AtomicOrdering MemOpOrder,
63 function_ref<Value *(IRBuilder<> &, Value *)> PerformOp);
64 void expandAtomicOpToLLSC(
65 Instruction *I, Type *ResultTy, Value *Addr, AtomicOrdering MemOpOrder,
66 function_ref<Value *(IRBuilder<> &, Value *)> PerformOp);
67 void expandPartwordAtomicRMW(
68 AtomicRMWInst *I,
69 TargetLoweringBase::AtomicExpansionKind ExpansionKind);
70 void expandPartwordCmpXchg(AtomicCmpXchgInst *I);
71
72 AtomicCmpXchgInst *convertCmpXchgToIntegerType(AtomicCmpXchgInst *CI);
73 static Value *insertRMWCmpXchgLoop(
74 IRBuilder<> &Builder, Type *ResultType, Value *Addr,
75 AtomicOrdering MemOpOrder,
76 function_ref<Value *(IRBuilder<> &, Value *)> PerformOp,
77 CreateCmpXchgInstFun CreateCmpXchg);
78
79 bool expandAtomicCmpXchg(AtomicCmpXchgInst *CI);
80 bool isIdempotentRMW(AtomicRMWInst *AI);
81 bool simplifyIdempotentRMW(AtomicRMWInst *AI);
82
83 bool expandAtomicOpToLibcall(Instruction *I, unsigned Size, unsigned Align,
84 Value *PointerOperand, Value *ValueOperand,
85 Value *CASExpected, AtomicOrdering Ordering,
86 AtomicOrdering Ordering2,
87 ArrayRef<RTLIB::Libcall> Libcalls);
88 void expandAtomicLoadToLibcall(LoadInst *LI);
89 void expandAtomicStoreToLibcall(StoreInst *LI);
90 void expandAtomicRMWToLibcall(AtomicRMWInst *I);
91 void expandAtomicCASToLibcall(AtomicCmpXchgInst *I);
92
93 friend bool
94 llvm::expandAtomicRMWToCmpXchg(AtomicRMWInst *AI,
95 CreateCmpXchgInstFun CreateCmpXchg);
96 };
97 }
98
99 char AtomicExpand::ID = 0;
100 char &llvm::AtomicExpandID = AtomicExpand::ID;
101 INITIALIZE_TM_PASS(AtomicExpand, "atomic-expand", "Expand Atomic instructions",
102 false, false)
103
createAtomicExpandPass(const TargetMachine * TM)104 FunctionPass *llvm::createAtomicExpandPass(const TargetMachine *TM) {
105 return new AtomicExpand(TM);
106 }
107
108 namespace {
109 // Helper functions to retrieve the size of atomic instructions.
getAtomicOpSize(LoadInst * LI)110 unsigned getAtomicOpSize(LoadInst *LI) {
111 const DataLayout &DL = LI->getModule()->getDataLayout();
112 return DL.getTypeStoreSize(LI->getType());
113 }
114
getAtomicOpSize(StoreInst * SI)115 unsigned getAtomicOpSize(StoreInst *SI) {
116 const DataLayout &DL = SI->getModule()->getDataLayout();
117 return DL.getTypeStoreSize(SI->getValueOperand()->getType());
118 }
119
getAtomicOpSize(AtomicRMWInst * RMWI)120 unsigned getAtomicOpSize(AtomicRMWInst *RMWI) {
121 const DataLayout &DL = RMWI->getModule()->getDataLayout();
122 return DL.getTypeStoreSize(RMWI->getValOperand()->getType());
123 }
124
getAtomicOpSize(AtomicCmpXchgInst * CASI)125 unsigned getAtomicOpSize(AtomicCmpXchgInst *CASI) {
126 const DataLayout &DL = CASI->getModule()->getDataLayout();
127 return DL.getTypeStoreSize(CASI->getCompareOperand()->getType());
128 }
129
130 // Helper functions to retrieve the alignment of atomic instructions.
getAtomicOpAlign(LoadInst * LI)131 unsigned getAtomicOpAlign(LoadInst *LI) {
132 unsigned Align = LI->getAlignment();
133 // In the future, if this IR restriction is relaxed, we should
134 // return DataLayout::getABITypeAlignment when there's no align
135 // value.
136 assert(Align != 0 && "An atomic LoadInst always has an explicit alignment");
137 return Align;
138 }
139
getAtomicOpAlign(StoreInst * SI)140 unsigned getAtomicOpAlign(StoreInst *SI) {
141 unsigned Align = SI->getAlignment();
142 // In the future, if this IR restriction is relaxed, we should
143 // return DataLayout::getABITypeAlignment when there's no align
144 // value.
145 assert(Align != 0 && "An atomic StoreInst always has an explicit alignment");
146 return Align;
147 }
148
getAtomicOpAlign(AtomicRMWInst * RMWI)149 unsigned getAtomicOpAlign(AtomicRMWInst *RMWI) {
150 // TODO(PR27168): This instruction has no alignment attribute, but unlike the
151 // default alignment for load/store, the default here is to assume
152 // it has NATURAL alignment, not DataLayout-specified alignment.
153 const DataLayout &DL = RMWI->getModule()->getDataLayout();
154 return DL.getTypeStoreSize(RMWI->getValOperand()->getType());
155 }
156
getAtomicOpAlign(AtomicCmpXchgInst * CASI)157 unsigned getAtomicOpAlign(AtomicCmpXchgInst *CASI) {
158 // TODO(PR27168): same comment as above.
159 const DataLayout &DL = CASI->getModule()->getDataLayout();
160 return DL.getTypeStoreSize(CASI->getCompareOperand()->getType());
161 }
162
163 // Determine if a particular atomic operation has a supported size,
164 // and is of appropriate alignment, to be passed through for target
165 // lowering. (Versus turning into a __atomic libcall)
166 template <typename Inst>
atomicSizeSupported(const TargetLowering * TLI,Inst * I)167 bool atomicSizeSupported(const TargetLowering *TLI, Inst *I) {
168 unsigned Size = getAtomicOpSize(I);
169 unsigned Align = getAtomicOpAlign(I);
170 return Align >= Size && Size <= TLI->getMaxAtomicSizeInBitsSupported() / 8;
171 }
172
173 } // end anonymous namespace
174
runOnFunction(Function & F)175 bool AtomicExpand::runOnFunction(Function &F) {
176 if (!TM || !TM->getSubtargetImpl(F)->enableAtomicExpand())
177 return false;
178 TLI = TM->getSubtargetImpl(F)->getTargetLowering();
179
180 SmallVector<Instruction *, 1> AtomicInsts;
181
182 // Changing control-flow while iterating through it is a bad idea, so gather a
183 // list of all atomic instructions before we start.
184 for (inst_iterator II = inst_begin(F), E = inst_end(F); II != E; ++II) {
185 Instruction *I = &*II;
186 if (I->isAtomic() && !isa<FenceInst>(I))
187 AtomicInsts.push_back(I);
188 }
189
190 bool MadeChange = false;
191 for (auto I : AtomicInsts) {
192 auto LI = dyn_cast<LoadInst>(I);
193 auto SI = dyn_cast<StoreInst>(I);
194 auto RMWI = dyn_cast<AtomicRMWInst>(I);
195 auto CASI = dyn_cast<AtomicCmpXchgInst>(I);
196 assert((LI || SI || RMWI || CASI) && "Unknown atomic instruction");
197
198 // If the Size/Alignment is not supported, replace with a libcall.
199 if (LI) {
200 if (!atomicSizeSupported(TLI, LI)) {
201 expandAtomicLoadToLibcall(LI);
202 MadeChange = true;
203 continue;
204 }
205 } else if (SI) {
206 if (!atomicSizeSupported(TLI, SI)) {
207 expandAtomicStoreToLibcall(SI);
208 MadeChange = true;
209 continue;
210 }
211 } else if (RMWI) {
212 if (!atomicSizeSupported(TLI, RMWI)) {
213 expandAtomicRMWToLibcall(RMWI);
214 MadeChange = true;
215 continue;
216 }
217 } else if (CASI) {
218 if (!atomicSizeSupported(TLI, CASI)) {
219 expandAtomicCASToLibcall(CASI);
220 MadeChange = true;
221 continue;
222 }
223 }
224
225 if (TLI->shouldInsertFencesForAtomic(I)) {
226 auto FenceOrdering = AtomicOrdering::Monotonic;
227 bool IsStore, IsLoad;
228 if (LI && isAcquireOrStronger(LI->getOrdering())) {
229 FenceOrdering = LI->getOrdering();
230 LI->setOrdering(AtomicOrdering::Monotonic);
231 IsStore = false;
232 IsLoad = true;
233 } else if (SI && isReleaseOrStronger(SI->getOrdering())) {
234 FenceOrdering = SI->getOrdering();
235 SI->setOrdering(AtomicOrdering::Monotonic);
236 IsStore = true;
237 IsLoad = false;
238 } else if (RMWI && (isReleaseOrStronger(RMWI->getOrdering()) ||
239 isAcquireOrStronger(RMWI->getOrdering()))) {
240 FenceOrdering = RMWI->getOrdering();
241 RMWI->setOrdering(AtomicOrdering::Monotonic);
242 IsStore = IsLoad = true;
243 } else if (CASI && !TLI->shouldExpandAtomicCmpXchgInIR(CASI) &&
244 (isReleaseOrStronger(CASI->getSuccessOrdering()) ||
245 isAcquireOrStronger(CASI->getSuccessOrdering()))) {
246 // If a compare and swap is lowered to LL/SC, we can do smarter fence
247 // insertion, with a stronger one on the success path than on the
248 // failure path. As a result, fence insertion is directly done by
249 // expandAtomicCmpXchg in that case.
250 FenceOrdering = CASI->getSuccessOrdering();
251 CASI->setSuccessOrdering(AtomicOrdering::Monotonic);
252 CASI->setFailureOrdering(AtomicOrdering::Monotonic);
253 IsStore = IsLoad = true;
254 }
255
256 if (FenceOrdering != AtomicOrdering::Monotonic) {
257 MadeChange |= bracketInstWithFences(I, FenceOrdering, IsStore, IsLoad);
258 }
259 }
260
261 if (LI) {
262 if (LI->getType()->isFloatingPointTy()) {
263 // TODO: add a TLI hook to control this so that each target can
264 // convert to lowering the original type one at a time.
265 LI = convertAtomicLoadToIntegerType(LI);
266 assert(LI->getType()->isIntegerTy() && "invariant broken");
267 MadeChange = true;
268 }
269
270 MadeChange |= tryExpandAtomicLoad(LI);
271 } else if (SI) {
272 if (SI->getValueOperand()->getType()->isFloatingPointTy()) {
273 // TODO: add a TLI hook to control this so that each target can
274 // convert to lowering the original type one at a time.
275 SI = convertAtomicStoreToIntegerType(SI);
276 assert(SI->getValueOperand()->getType()->isIntegerTy() &&
277 "invariant broken");
278 MadeChange = true;
279 }
280
281 if (TLI->shouldExpandAtomicStoreInIR(SI))
282 MadeChange |= expandAtomicStore(SI);
283 } else if (RMWI) {
284 // There are two different ways of expanding RMW instructions:
285 // - into a load if it is idempotent
286 // - into a Cmpxchg/LL-SC loop otherwise
287 // we try them in that order.
288
289 if (isIdempotentRMW(RMWI) && simplifyIdempotentRMW(RMWI)) {
290 MadeChange = true;
291 } else {
292 MadeChange |= tryExpandAtomicRMW(RMWI);
293 }
294 } else if (CASI) {
295 // TODO: when we're ready to make the change at the IR level, we can
296 // extend convertCmpXchgToInteger for floating point too.
297 assert(!CASI->getCompareOperand()->getType()->isFloatingPointTy() &&
298 "unimplemented - floating point not legal at IR level");
299 if (CASI->getCompareOperand()->getType()->isPointerTy() ) {
300 // TODO: add a TLI hook to control this so that each target can
301 // convert to lowering the original type one at a time.
302 CASI = convertCmpXchgToIntegerType(CASI);
303 assert(CASI->getCompareOperand()->getType()->isIntegerTy() &&
304 "invariant broken");
305 MadeChange = true;
306 }
307
308 unsigned MinCASSize = TLI->getMinCmpXchgSizeInBits() / 8;
309 unsigned ValueSize = getAtomicOpSize(CASI);
310 if (ValueSize < MinCASSize) {
311 assert(!TLI->shouldExpandAtomicCmpXchgInIR(CASI) &&
312 "MinCmpXchgSizeInBits not yet supported for LL/SC expansions.");
313 expandPartwordCmpXchg(CASI);
314 } else {
315 if (TLI->shouldExpandAtomicCmpXchgInIR(CASI))
316 MadeChange |= expandAtomicCmpXchg(CASI);
317 }
318 }
319 }
320 return MadeChange;
321 }
322
bracketInstWithFences(Instruction * I,AtomicOrdering Order,bool IsStore,bool IsLoad)323 bool AtomicExpand::bracketInstWithFences(Instruction *I, AtomicOrdering Order,
324 bool IsStore, bool IsLoad) {
325 IRBuilder<> Builder(I);
326
327 auto LeadingFence = TLI->emitLeadingFence(Builder, Order, IsStore, IsLoad);
328
329 auto TrailingFence = TLI->emitTrailingFence(Builder, Order, IsStore, IsLoad);
330 // The trailing fence is emitted before the instruction instead of after
331 // because there is no easy way of setting Builder insertion point after
332 // an instruction. So we must erase it from the BB, and insert it back
333 // in the right place.
334 // We have a guard here because not every atomic operation generates a
335 // trailing fence.
336 if (TrailingFence) {
337 TrailingFence->removeFromParent();
338 TrailingFence->insertAfter(I);
339 }
340
341 return (LeadingFence || TrailingFence);
342 }
343
344 /// Get the iX type with the same bitwidth as T.
getCorrespondingIntegerType(Type * T,const DataLayout & DL)345 IntegerType *AtomicExpand::getCorrespondingIntegerType(Type *T,
346 const DataLayout &DL) {
347 EVT VT = TLI->getValueType(DL, T);
348 unsigned BitWidth = VT.getStoreSizeInBits();
349 assert(BitWidth == VT.getSizeInBits() && "must be a power of two");
350 return IntegerType::get(T->getContext(), BitWidth);
351 }
352
353 /// Convert an atomic load of a non-integral type to an integer load of the
354 /// equivalent bitwidth. See the function comment on
355 /// convertAtomicStoreToIntegerType for background.
convertAtomicLoadToIntegerType(LoadInst * LI)356 LoadInst *AtomicExpand::convertAtomicLoadToIntegerType(LoadInst *LI) {
357 auto *M = LI->getModule();
358 Type *NewTy = getCorrespondingIntegerType(LI->getType(),
359 M->getDataLayout());
360
361 IRBuilder<> Builder(LI);
362
363 Value *Addr = LI->getPointerOperand();
364 Type *PT = PointerType::get(NewTy,
365 Addr->getType()->getPointerAddressSpace());
366 Value *NewAddr = Builder.CreateBitCast(Addr, PT);
367
368 auto *NewLI = Builder.CreateLoad(NewAddr);
369 NewLI->setAlignment(LI->getAlignment());
370 NewLI->setVolatile(LI->isVolatile());
371 NewLI->setAtomic(LI->getOrdering(), LI->getSynchScope());
372 DEBUG(dbgs() << "Replaced " << *LI << " with " << *NewLI << "\n");
373
374 Value *NewVal = Builder.CreateBitCast(NewLI, LI->getType());
375 LI->replaceAllUsesWith(NewVal);
376 LI->eraseFromParent();
377 return NewLI;
378 }
379
tryExpandAtomicLoad(LoadInst * LI)380 bool AtomicExpand::tryExpandAtomicLoad(LoadInst *LI) {
381 switch (TLI->shouldExpandAtomicLoadInIR(LI)) {
382 case TargetLoweringBase::AtomicExpansionKind::None:
383 return false;
384 case TargetLoweringBase::AtomicExpansionKind::LLSC:
385 expandAtomicOpToLLSC(
386 LI, LI->getType(), LI->getPointerOperand(), LI->getOrdering(),
387 [](IRBuilder<> &Builder, Value *Loaded) { return Loaded; });
388 return true;
389 case TargetLoweringBase::AtomicExpansionKind::LLOnly:
390 return expandAtomicLoadToLL(LI);
391 case TargetLoweringBase::AtomicExpansionKind::CmpXChg:
392 return expandAtomicLoadToCmpXchg(LI);
393 }
394 llvm_unreachable("Unhandled case in tryExpandAtomicLoad");
395 }
396
expandAtomicLoadToLL(LoadInst * LI)397 bool AtomicExpand::expandAtomicLoadToLL(LoadInst *LI) {
398 IRBuilder<> Builder(LI);
399
400 // On some architectures, load-linked instructions are atomic for larger
401 // sizes than normal loads. For example, the only 64-bit load guaranteed
402 // to be single-copy atomic by ARM is an ldrexd (A3.5.3).
403 Value *Val =
404 TLI->emitLoadLinked(Builder, LI->getPointerOperand(), LI->getOrdering());
405 TLI->emitAtomicCmpXchgNoStoreLLBalance(Builder);
406
407 LI->replaceAllUsesWith(Val);
408 LI->eraseFromParent();
409
410 return true;
411 }
412
expandAtomicLoadToCmpXchg(LoadInst * LI)413 bool AtomicExpand::expandAtomicLoadToCmpXchg(LoadInst *LI) {
414 IRBuilder<> Builder(LI);
415 AtomicOrdering Order = LI->getOrdering();
416 Value *Addr = LI->getPointerOperand();
417 Type *Ty = cast<PointerType>(Addr->getType())->getElementType();
418 Constant *DummyVal = Constant::getNullValue(Ty);
419
420 Value *Pair = Builder.CreateAtomicCmpXchg(
421 Addr, DummyVal, DummyVal, Order,
422 AtomicCmpXchgInst::getStrongestFailureOrdering(Order));
423 Value *Loaded = Builder.CreateExtractValue(Pair, 0, "loaded");
424
425 LI->replaceAllUsesWith(Loaded);
426 LI->eraseFromParent();
427
428 return true;
429 }
430
431 /// Convert an atomic store of a non-integral type to an integer store of the
432 /// equivalent bitwidth. We used to not support floating point or vector
433 /// atomics in the IR at all. The backends learned to deal with the bitcast
434 /// idiom because that was the only way of expressing the notion of a atomic
435 /// float or vector store. The long term plan is to teach each backend to
436 /// instruction select from the original atomic store, but as a migration
437 /// mechanism, we convert back to the old format which the backends understand.
438 /// Each backend will need individual work to recognize the new format.
convertAtomicStoreToIntegerType(StoreInst * SI)439 StoreInst *AtomicExpand::convertAtomicStoreToIntegerType(StoreInst *SI) {
440 IRBuilder<> Builder(SI);
441 auto *M = SI->getModule();
442 Type *NewTy = getCorrespondingIntegerType(SI->getValueOperand()->getType(),
443 M->getDataLayout());
444 Value *NewVal = Builder.CreateBitCast(SI->getValueOperand(), NewTy);
445
446 Value *Addr = SI->getPointerOperand();
447 Type *PT = PointerType::get(NewTy,
448 Addr->getType()->getPointerAddressSpace());
449 Value *NewAddr = Builder.CreateBitCast(Addr, PT);
450
451 StoreInst *NewSI = Builder.CreateStore(NewVal, NewAddr);
452 NewSI->setAlignment(SI->getAlignment());
453 NewSI->setVolatile(SI->isVolatile());
454 NewSI->setAtomic(SI->getOrdering(), SI->getSynchScope());
455 DEBUG(dbgs() << "Replaced " << *SI << " with " << *NewSI << "\n");
456 SI->eraseFromParent();
457 return NewSI;
458 }
459
expandAtomicStore(StoreInst * SI)460 bool AtomicExpand::expandAtomicStore(StoreInst *SI) {
461 // This function is only called on atomic stores that are too large to be
462 // atomic if implemented as a native store. So we replace them by an
463 // atomic swap, that can be implemented for example as a ldrex/strex on ARM
464 // or lock cmpxchg8/16b on X86, as these are atomic for larger sizes.
465 // It is the responsibility of the target to only signal expansion via
466 // shouldExpandAtomicRMW in cases where this is required and possible.
467 IRBuilder<> Builder(SI);
468 AtomicRMWInst *AI =
469 Builder.CreateAtomicRMW(AtomicRMWInst::Xchg, SI->getPointerOperand(),
470 SI->getValueOperand(), SI->getOrdering());
471 SI->eraseFromParent();
472
473 // Now we have an appropriate swap instruction, lower it as usual.
474 return tryExpandAtomicRMW(AI);
475 }
476
createCmpXchgInstFun(IRBuilder<> & Builder,Value * Addr,Value * Loaded,Value * NewVal,AtomicOrdering MemOpOrder,Value * & Success,Value * & NewLoaded)477 static void createCmpXchgInstFun(IRBuilder<> &Builder, Value *Addr,
478 Value *Loaded, Value *NewVal,
479 AtomicOrdering MemOpOrder,
480 Value *&Success, Value *&NewLoaded) {
481 Value* Pair = Builder.CreateAtomicCmpXchg(
482 Addr, Loaded, NewVal, MemOpOrder,
483 AtomicCmpXchgInst::getStrongestFailureOrdering(MemOpOrder));
484 Success = Builder.CreateExtractValue(Pair, 1, "success");
485 NewLoaded = Builder.CreateExtractValue(Pair, 0, "newloaded");
486 }
487
488 /// Emit IR to implement the given atomicrmw operation on values in registers,
489 /// returning the new value.
performAtomicOp(AtomicRMWInst::BinOp Op,IRBuilder<> & Builder,Value * Loaded,Value * Inc)490 static Value *performAtomicOp(AtomicRMWInst::BinOp Op, IRBuilder<> &Builder,
491 Value *Loaded, Value *Inc) {
492 Value *NewVal;
493 switch (Op) {
494 case AtomicRMWInst::Xchg:
495 return Inc;
496 case AtomicRMWInst::Add:
497 return Builder.CreateAdd(Loaded, Inc, "new");
498 case AtomicRMWInst::Sub:
499 return Builder.CreateSub(Loaded, Inc, "new");
500 case AtomicRMWInst::And:
501 return Builder.CreateAnd(Loaded, Inc, "new");
502 case AtomicRMWInst::Nand:
503 return Builder.CreateNot(Builder.CreateAnd(Loaded, Inc), "new");
504 case AtomicRMWInst::Or:
505 return Builder.CreateOr(Loaded, Inc, "new");
506 case AtomicRMWInst::Xor:
507 return Builder.CreateXor(Loaded, Inc, "new");
508 case AtomicRMWInst::Max:
509 NewVal = Builder.CreateICmpSGT(Loaded, Inc);
510 return Builder.CreateSelect(NewVal, Loaded, Inc, "new");
511 case AtomicRMWInst::Min:
512 NewVal = Builder.CreateICmpSLE(Loaded, Inc);
513 return Builder.CreateSelect(NewVal, Loaded, Inc, "new");
514 case AtomicRMWInst::UMax:
515 NewVal = Builder.CreateICmpUGT(Loaded, Inc);
516 return Builder.CreateSelect(NewVal, Loaded, Inc, "new");
517 case AtomicRMWInst::UMin:
518 NewVal = Builder.CreateICmpULE(Loaded, Inc);
519 return Builder.CreateSelect(NewVal, Loaded, Inc, "new");
520 default:
521 llvm_unreachable("Unknown atomic op");
522 }
523 }
524
tryExpandAtomicRMW(AtomicRMWInst * AI)525 bool AtomicExpand::tryExpandAtomicRMW(AtomicRMWInst *AI) {
526 switch (TLI->shouldExpandAtomicRMWInIR(AI)) {
527 case TargetLoweringBase::AtomicExpansionKind::None:
528 return false;
529 case TargetLoweringBase::AtomicExpansionKind::LLSC: {
530 unsigned MinCASSize = TLI->getMinCmpXchgSizeInBits() / 8;
531 unsigned ValueSize = getAtomicOpSize(AI);
532 if (ValueSize < MinCASSize) {
533 llvm_unreachable(
534 "MinCmpXchgSizeInBits not yet supported for LL/SC architectures.");
535 } else {
536 auto PerformOp = [&](IRBuilder<> &Builder, Value *Loaded) {
537 return performAtomicOp(AI->getOperation(), Builder, Loaded,
538 AI->getValOperand());
539 };
540 expandAtomicOpToLLSC(AI, AI->getType(), AI->getPointerOperand(),
541 AI->getOrdering(), PerformOp);
542 }
543 return true;
544 }
545 case TargetLoweringBase::AtomicExpansionKind::CmpXChg: {
546 unsigned MinCASSize = TLI->getMinCmpXchgSizeInBits() / 8;
547 unsigned ValueSize = getAtomicOpSize(AI);
548 if (ValueSize < MinCASSize) {
549 expandPartwordAtomicRMW(AI,
550 TargetLoweringBase::AtomicExpansionKind::CmpXChg);
551 } else {
552 expandAtomicRMWToCmpXchg(AI, createCmpXchgInstFun);
553 }
554 return true;
555 }
556 default:
557 llvm_unreachable("Unhandled case in tryExpandAtomicRMW");
558 }
559 }
560
561 namespace {
562
563 /// Result values from createMaskInstrs helper.
564 struct PartwordMaskValues {
565 Type *WordType;
566 Type *ValueType;
567 Value *AlignedAddr;
568 Value *ShiftAmt;
569 Value *Mask;
570 Value *Inv_Mask;
571 };
572 } // end anonymous namespace
573
574 /// This is a helper function which builds instructions to provide
575 /// values necessary for partword atomic operations. It takes an
576 /// incoming address, Addr, and ValueType, and constructs the address,
577 /// shift-amounts and masks needed to work with a larger value of size
578 /// WordSize.
579 ///
580 /// AlignedAddr: Addr rounded down to a multiple of WordSize
581 ///
582 /// ShiftAmt: Number of bits to right-shift a WordSize value loaded
583 /// from AlignAddr for it to have the same value as if
584 /// ValueType was loaded from Addr.
585 ///
586 /// Mask: Value to mask with the value loaded from AlignAddr to
587 /// include only the part that would've been loaded from Addr.
588 ///
589 /// Inv_Mask: The inverse of Mask.
590
createMaskInstrs(IRBuilder<> & Builder,Instruction * I,Type * ValueType,Value * Addr,unsigned WordSize)591 static PartwordMaskValues createMaskInstrs(IRBuilder<> &Builder, Instruction *I,
592 Type *ValueType, Value *Addr,
593 unsigned WordSize) {
594 PartwordMaskValues Ret;
595
596 BasicBlock *BB = I->getParent();
597 Function *F = BB->getParent();
598 Module *M = I->getModule();
599
600 LLVMContext &Ctx = F->getContext();
601 const DataLayout &DL = M->getDataLayout();
602
603 unsigned ValueSize = DL.getTypeStoreSize(ValueType);
604
605 assert(ValueSize < WordSize);
606
607 Ret.ValueType = ValueType;
608 Ret.WordType = Type::getIntNTy(Ctx, WordSize * 8);
609
610 Type *WordPtrType =
611 Ret.WordType->getPointerTo(Addr->getType()->getPointerAddressSpace());
612
613 Value *AddrInt = Builder.CreatePtrToInt(Addr, DL.getIntPtrType(Ctx));
614 Ret.AlignedAddr = Builder.CreateIntToPtr(
615 Builder.CreateAnd(AddrInt, ~(uint64_t)(WordSize - 1)), WordPtrType,
616 "AlignedAddr");
617
618 Value *PtrLSB = Builder.CreateAnd(AddrInt, WordSize - 1, "PtrLSB");
619 if (DL.isLittleEndian()) {
620 // turn bytes into bits
621 Ret.ShiftAmt = Builder.CreateShl(PtrLSB, 3);
622 } else {
623 // turn bytes into bits, and count from the other side.
624 Ret.ShiftAmt =
625 Builder.CreateShl(Builder.CreateXor(PtrLSB, WordSize - ValueSize), 3);
626 }
627
628 Ret.ShiftAmt = Builder.CreateTrunc(Ret.ShiftAmt, Ret.WordType, "ShiftAmt");
629 Ret.Mask = Builder.CreateShl(
630 ConstantInt::get(Ret.WordType, (1 << ValueSize * 8) - 1), Ret.ShiftAmt,
631 "Mask");
632 Ret.Inv_Mask = Builder.CreateNot(Ret.Mask, "Inv_Mask");
633
634 return Ret;
635 }
636
637 /// Emit IR to implement a masked version of a given atomicrmw
638 /// operation. (That is, only the bits under the Mask should be
639 /// affected by the operation)
performMaskedAtomicOp(AtomicRMWInst::BinOp Op,IRBuilder<> & Builder,Value * Loaded,Value * Shifted_Inc,Value * Inc,const PartwordMaskValues & PMV)640 static Value *performMaskedAtomicOp(AtomicRMWInst::BinOp Op,
641 IRBuilder<> &Builder, Value *Loaded,
642 Value *Shifted_Inc, Value *Inc,
643 const PartwordMaskValues &PMV) {
644 switch (Op) {
645 case AtomicRMWInst::Xchg: {
646 Value *Loaded_MaskOut = Builder.CreateAnd(Loaded, PMV.Inv_Mask);
647 Value *FinalVal = Builder.CreateOr(Loaded_MaskOut, Shifted_Inc);
648 return FinalVal;
649 }
650 case AtomicRMWInst::Or:
651 case AtomicRMWInst::Xor:
652 // Or/Xor won't affect any other bits, so can just be done
653 // directly.
654 return performAtomicOp(Op, Builder, Loaded, Shifted_Inc);
655 case AtomicRMWInst::Add:
656 case AtomicRMWInst::Sub:
657 case AtomicRMWInst::And:
658 case AtomicRMWInst::Nand: {
659 // The other arithmetic ops need to be masked into place.
660 Value *NewVal = performAtomicOp(Op, Builder, Loaded, Shifted_Inc);
661 Value *NewVal_Masked = Builder.CreateAnd(NewVal, PMV.Mask);
662 Value *Loaded_MaskOut = Builder.CreateAnd(Loaded, PMV.Inv_Mask);
663 Value *FinalVal = Builder.CreateOr(Loaded_MaskOut, NewVal_Masked);
664 return FinalVal;
665 }
666 case AtomicRMWInst::Max:
667 case AtomicRMWInst::Min:
668 case AtomicRMWInst::UMax:
669 case AtomicRMWInst::UMin: {
670 // Finally, comparison ops will operate on the full value, so
671 // truncate down to the original size, and expand out again after
672 // doing the operation.
673 Value *Loaded_Shiftdown = Builder.CreateTrunc(
674 Builder.CreateLShr(Loaded, PMV.ShiftAmt), PMV.ValueType);
675 Value *NewVal = performAtomicOp(Op, Builder, Loaded_Shiftdown, Inc);
676 Value *NewVal_Shiftup = Builder.CreateShl(
677 Builder.CreateZExt(NewVal, PMV.WordType), PMV.ShiftAmt);
678 Value *Loaded_MaskOut = Builder.CreateAnd(Loaded, PMV.Inv_Mask);
679 Value *FinalVal = Builder.CreateOr(Loaded_MaskOut, NewVal_Shiftup);
680 return FinalVal;
681 }
682 default:
683 llvm_unreachable("Unknown atomic op");
684 }
685 }
686
687 /// Expand a sub-word atomicrmw operation into an appropriate
688 /// word-sized operation.
689 ///
690 /// It will create an LL/SC or cmpxchg loop, as appropriate, the same
691 /// way as a typical atomicrmw expansion. The only difference here is
692 /// that the operation inside of the loop must operate only upon a
693 /// part of the value.
expandPartwordAtomicRMW(AtomicRMWInst * AI,TargetLoweringBase::AtomicExpansionKind ExpansionKind)694 void AtomicExpand::expandPartwordAtomicRMW(
695 AtomicRMWInst *AI, TargetLoweringBase::AtomicExpansionKind ExpansionKind) {
696
697 assert(ExpansionKind == TargetLoweringBase::AtomicExpansionKind::CmpXChg);
698
699 AtomicOrdering MemOpOrder = AI->getOrdering();
700
701 IRBuilder<> Builder(AI);
702
703 PartwordMaskValues PMV =
704 createMaskInstrs(Builder, AI, AI->getType(), AI->getPointerOperand(),
705 TLI->getMinCmpXchgSizeInBits() / 8);
706
707 Value *ValOperand_Shifted =
708 Builder.CreateShl(Builder.CreateZExt(AI->getValOperand(), PMV.WordType),
709 PMV.ShiftAmt, "ValOperand_Shifted");
710
711 auto PerformPartwordOp = [&](IRBuilder<> &Builder, Value *Loaded) {
712 return performMaskedAtomicOp(AI->getOperation(), Builder, Loaded,
713 ValOperand_Shifted, AI->getValOperand(), PMV);
714 };
715
716 // TODO: When we're ready to support LLSC conversions too, use
717 // insertRMWLLSCLoop here for ExpansionKind==LLSC.
718 Value *OldResult =
719 insertRMWCmpXchgLoop(Builder, PMV.WordType, PMV.AlignedAddr, MemOpOrder,
720 PerformPartwordOp, createCmpXchgInstFun);
721 Value *FinalOldResult = Builder.CreateTrunc(
722 Builder.CreateLShr(OldResult, PMV.ShiftAmt), PMV.ValueType);
723 AI->replaceAllUsesWith(FinalOldResult);
724 AI->eraseFromParent();
725 }
726
expandPartwordCmpXchg(AtomicCmpXchgInst * CI)727 void AtomicExpand::expandPartwordCmpXchg(AtomicCmpXchgInst *CI) {
728 // The basic idea here is that we're expanding a cmpxchg of a
729 // smaller memory size up to a word-sized cmpxchg. To do this, we
730 // need to add a retry-loop for strong cmpxchg, so that
731 // modifications to other parts of the word don't cause a spurious
732 // failure.
733
734 // This generates code like the following:
735 // [[Setup mask values PMV.*]]
736 // %NewVal_Shifted = shl i32 %NewVal, %PMV.ShiftAmt
737 // %Cmp_Shifted = shl i32 %Cmp, %PMV.ShiftAmt
738 // %InitLoaded = load i32* %addr
739 // %InitLoaded_MaskOut = and i32 %InitLoaded, %PMV.Inv_Mask
740 // br partword.cmpxchg.loop
741 // partword.cmpxchg.loop:
742 // %Loaded_MaskOut = phi i32 [ %InitLoaded_MaskOut, %entry ],
743 // [ %OldVal_MaskOut, %partword.cmpxchg.failure ]
744 // %FullWord_NewVal = or i32 %Loaded_MaskOut, %NewVal_Shifted
745 // %FullWord_Cmp = or i32 %Loaded_MaskOut, %Cmp_Shifted
746 // %NewCI = cmpxchg i32* %PMV.AlignedAddr, i32 %FullWord_Cmp,
747 // i32 %FullWord_NewVal success_ordering failure_ordering
748 // %OldVal = extractvalue { i32, i1 } %NewCI, 0
749 // %Success = extractvalue { i32, i1 } %NewCI, 1
750 // br i1 %Success, label %partword.cmpxchg.end,
751 // label %partword.cmpxchg.failure
752 // partword.cmpxchg.failure:
753 // %OldVal_MaskOut = and i32 %OldVal, %PMV.Inv_Mask
754 // %ShouldContinue = icmp ne i32 %Loaded_MaskOut, %OldVal_MaskOut
755 // br i1 %ShouldContinue, label %partword.cmpxchg.loop,
756 // label %partword.cmpxchg.end
757 // partword.cmpxchg.end:
758 // %tmp1 = lshr i32 %OldVal, %PMV.ShiftAmt
759 // %FinalOldVal = trunc i32 %tmp1 to i8
760 // %tmp2 = insertvalue { i8, i1 } undef, i8 %FinalOldVal, 0
761 // %Res = insertvalue { i8, i1 } %25, i1 %Success, 1
762
763 Value *Addr = CI->getPointerOperand();
764 Value *Cmp = CI->getCompareOperand();
765 Value *NewVal = CI->getNewValOperand();
766
767 BasicBlock *BB = CI->getParent();
768 Function *F = BB->getParent();
769 IRBuilder<> Builder(CI);
770 LLVMContext &Ctx = Builder.getContext();
771
772 const int WordSize = TLI->getMinCmpXchgSizeInBits() / 8;
773
774 BasicBlock *EndBB =
775 BB->splitBasicBlock(CI->getIterator(), "partword.cmpxchg.end");
776 auto FailureBB =
777 BasicBlock::Create(Ctx, "partword.cmpxchg.failure", F, EndBB);
778 auto LoopBB = BasicBlock::Create(Ctx, "partword.cmpxchg.loop", F, FailureBB);
779
780 // The split call above "helpfully" added a branch at the end of BB
781 // (to the wrong place).
782 std::prev(BB->end())->eraseFromParent();
783 Builder.SetInsertPoint(BB);
784
785 PartwordMaskValues PMV = createMaskInstrs(
786 Builder, CI, CI->getCompareOperand()->getType(), Addr, WordSize);
787
788 // Shift the incoming values over, into the right location in the word.
789 Value *NewVal_Shifted =
790 Builder.CreateShl(Builder.CreateZExt(NewVal, PMV.WordType), PMV.ShiftAmt);
791 Value *Cmp_Shifted =
792 Builder.CreateShl(Builder.CreateZExt(Cmp, PMV.WordType), PMV.ShiftAmt);
793
794 // Load the entire current word, and mask into place the expected and new
795 // values
796 LoadInst *InitLoaded = Builder.CreateLoad(PMV.WordType, PMV.AlignedAddr);
797 InitLoaded->setVolatile(CI->isVolatile());
798 Value *InitLoaded_MaskOut = Builder.CreateAnd(InitLoaded, PMV.Inv_Mask);
799 Builder.CreateBr(LoopBB);
800
801 // partword.cmpxchg.loop:
802 Builder.SetInsertPoint(LoopBB);
803 PHINode *Loaded_MaskOut = Builder.CreatePHI(PMV.WordType, 2);
804 Loaded_MaskOut->addIncoming(InitLoaded_MaskOut, BB);
805
806 // Mask/Or the expected and new values into place in the loaded word.
807 Value *FullWord_NewVal = Builder.CreateOr(Loaded_MaskOut, NewVal_Shifted);
808 Value *FullWord_Cmp = Builder.CreateOr(Loaded_MaskOut, Cmp_Shifted);
809 AtomicCmpXchgInst *NewCI = Builder.CreateAtomicCmpXchg(
810 PMV.AlignedAddr, FullWord_Cmp, FullWord_NewVal, CI->getSuccessOrdering(),
811 CI->getFailureOrdering(), CI->getSynchScope());
812 NewCI->setVolatile(CI->isVolatile());
813 // When we're building a strong cmpxchg, we need a loop, so you
814 // might think we could use a weak cmpxchg inside. But, using strong
815 // allows the below comparison for ShouldContinue, and we're
816 // expecting the underlying cmpxchg to be a machine instruction,
817 // which is strong anyways.
818 NewCI->setWeak(CI->isWeak());
819
820 Value *OldVal = Builder.CreateExtractValue(NewCI, 0);
821 Value *Success = Builder.CreateExtractValue(NewCI, 1);
822
823 if (CI->isWeak())
824 Builder.CreateBr(EndBB);
825 else
826 Builder.CreateCondBr(Success, EndBB, FailureBB);
827
828 // partword.cmpxchg.failure:
829 Builder.SetInsertPoint(FailureBB);
830 // Upon failure, verify that the masked-out part of the loaded value
831 // has been modified. If it didn't, abort the cmpxchg, since the
832 // masked-in part must've.
833 Value *OldVal_MaskOut = Builder.CreateAnd(OldVal, PMV.Inv_Mask);
834 Value *ShouldContinue = Builder.CreateICmpNE(Loaded_MaskOut, OldVal_MaskOut);
835 Builder.CreateCondBr(ShouldContinue, LoopBB, EndBB);
836
837 // Add the second value to the phi from above
838 Loaded_MaskOut->addIncoming(OldVal_MaskOut, FailureBB);
839
840 // partword.cmpxchg.end:
841 Builder.SetInsertPoint(CI);
842
843 Value *FinalOldVal = Builder.CreateTrunc(
844 Builder.CreateLShr(OldVal, PMV.ShiftAmt), PMV.ValueType);
845 Value *Res = UndefValue::get(CI->getType());
846 Res = Builder.CreateInsertValue(Res, FinalOldVal, 0);
847 Res = Builder.CreateInsertValue(Res, Success, 1);
848
849 CI->replaceAllUsesWith(Res);
850 CI->eraseFromParent();
851 }
852
expandAtomicOpToLLSC(Instruction * I,Type * ResultType,Value * Addr,AtomicOrdering MemOpOrder,function_ref<Value * (IRBuilder<> &,Value *)> PerformOp)853 void AtomicExpand::expandAtomicOpToLLSC(
854 Instruction *I, Type *ResultType, Value *Addr, AtomicOrdering MemOpOrder,
855 function_ref<Value *(IRBuilder<> &, Value *)> PerformOp) {
856 IRBuilder<> Builder(I);
857 Value *Loaded =
858 insertRMWLLSCLoop(Builder, ResultType, Addr, MemOpOrder, PerformOp);
859
860 I->replaceAllUsesWith(Loaded);
861 I->eraseFromParent();
862 }
863
insertRMWLLSCLoop(IRBuilder<> & Builder,Type * ResultTy,Value * Addr,AtomicOrdering MemOpOrder,function_ref<Value * (IRBuilder<> &,Value *)> PerformOp)864 Value *AtomicExpand::insertRMWLLSCLoop(
865 IRBuilder<> &Builder, Type *ResultTy, Value *Addr,
866 AtomicOrdering MemOpOrder,
867 function_ref<Value *(IRBuilder<> &, Value *)> PerformOp) {
868 LLVMContext &Ctx = Builder.getContext();
869 BasicBlock *BB = Builder.GetInsertBlock();
870 Function *F = BB->getParent();
871
872 // Given: atomicrmw some_op iN* %addr, iN %incr ordering
873 //
874 // The standard expansion we produce is:
875 // [...]
876 // atomicrmw.start:
877 // %loaded = @load.linked(%addr)
878 // %new = some_op iN %loaded, %incr
879 // %stored = @store_conditional(%new, %addr)
880 // %try_again = icmp i32 ne %stored, 0
881 // br i1 %try_again, label %loop, label %atomicrmw.end
882 // atomicrmw.end:
883 // [...]
884 BasicBlock *ExitBB =
885 BB->splitBasicBlock(Builder.GetInsertPoint(), "atomicrmw.end");
886 BasicBlock *LoopBB = BasicBlock::Create(Ctx, "atomicrmw.start", F, ExitBB);
887
888 // The split call above "helpfully" added a branch at the end of BB (to the
889 // wrong place).
890 std::prev(BB->end())->eraseFromParent();
891 Builder.SetInsertPoint(BB);
892 Builder.CreateBr(LoopBB);
893
894 // Start the main loop block now that we've taken care of the preliminaries.
895 Builder.SetInsertPoint(LoopBB);
896 Value *Loaded = TLI->emitLoadLinked(Builder, Addr, MemOpOrder);
897
898 Value *NewVal = PerformOp(Builder, Loaded);
899
900 Value *StoreSuccess =
901 TLI->emitStoreConditional(Builder, NewVal, Addr, MemOpOrder);
902 Value *TryAgain = Builder.CreateICmpNE(
903 StoreSuccess, ConstantInt::get(IntegerType::get(Ctx, 32), 0), "tryagain");
904 Builder.CreateCondBr(TryAgain, LoopBB, ExitBB);
905
906 Builder.SetInsertPoint(ExitBB, ExitBB->begin());
907 return Loaded;
908 }
909
910 /// Convert an atomic cmpxchg of a non-integral type to an integer cmpxchg of
911 /// the equivalent bitwidth. We used to not support pointer cmpxchg in the
912 /// IR. As a migration step, we convert back to what use to be the standard
913 /// way to represent a pointer cmpxchg so that we can update backends one by
914 /// one.
convertCmpXchgToIntegerType(AtomicCmpXchgInst * CI)915 AtomicCmpXchgInst *AtomicExpand::convertCmpXchgToIntegerType(AtomicCmpXchgInst *CI) {
916 auto *M = CI->getModule();
917 Type *NewTy = getCorrespondingIntegerType(CI->getCompareOperand()->getType(),
918 M->getDataLayout());
919
920 IRBuilder<> Builder(CI);
921
922 Value *Addr = CI->getPointerOperand();
923 Type *PT = PointerType::get(NewTy,
924 Addr->getType()->getPointerAddressSpace());
925 Value *NewAddr = Builder.CreateBitCast(Addr, PT);
926
927 Value *NewCmp = Builder.CreatePtrToInt(CI->getCompareOperand(), NewTy);
928 Value *NewNewVal = Builder.CreatePtrToInt(CI->getNewValOperand(), NewTy);
929
930
931 auto *NewCI = Builder.CreateAtomicCmpXchg(NewAddr, NewCmp, NewNewVal,
932 CI->getSuccessOrdering(),
933 CI->getFailureOrdering(),
934 CI->getSynchScope());
935 NewCI->setVolatile(CI->isVolatile());
936 NewCI->setWeak(CI->isWeak());
937 DEBUG(dbgs() << "Replaced " << *CI << " with " << *NewCI << "\n");
938
939 Value *OldVal = Builder.CreateExtractValue(NewCI, 0);
940 Value *Succ = Builder.CreateExtractValue(NewCI, 1);
941
942 OldVal = Builder.CreateIntToPtr(OldVal, CI->getCompareOperand()->getType());
943
944 Value *Res = UndefValue::get(CI->getType());
945 Res = Builder.CreateInsertValue(Res, OldVal, 0);
946 Res = Builder.CreateInsertValue(Res, Succ, 1);
947
948 CI->replaceAllUsesWith(Res);
949 CI->eraseFromParent();
950 return NewCI;
951 }
952
953
expandAtomicCmpXchg(AtomicCmpXchgInst * CI)954 bool AtomicExpand::expandAtomicCmpXchg(AtomicCmpXchgInst *CI) {
955 AtomicOrdering SuccessOrder = CI->getSuccessOrdering();
956 AtomicOrdering FailureOrder = CI->getFailureOrdering();
957 Value *Addr = CI->getPointerOperand();
958 BasicBlock *BB = CI->getParent();
959 Function *F = BB->getParent();
960 LLVMContext &Ctx = F->getContext();
961 // If shouldInsertFencesForAtomic() returns true, then the target does not
962 // want to deal with memory orders, and emitLeading/TrailingFence should take
963 // care of everything. Otherwise, emitLeading/TrailingFence are no-op and we
964 // should preserve the ordering.
965 bool ShouldInsertFencesForAtomic = TLI->shouldInsertFencesForAtomic(CI);
966 AtomicOrdering MemOpOrder =
967 ShouldInsertFencesForAtomic ? AtomicOrdering::Monotonic : SuccessOrder;
968
969 // In implementations which use a barrier to achieve release semantics, we can
970 // delay emitting this barrier until we know a store is actually going to be
971 // attempted. The cost of this delay is that we need 2 copies of the block
972 // emitting the load-linked, affecting code size.
973 //
974 // Ideally, this logic would be unconditional except for the minsize check
975 // since in other cases the extra blocks naturally collapse down to the
976 // minimal loop. Unfortunately, this puts too much stress on later
977 // optimisations so we avoid emitting the extra logic in those cases too.
978 bool HasReleasedLoadBB = !CI->isWeak() && ShouldInsertFencesForAtomic &&
979 SuccessOrder != AtomicOrdering::Monotonic &&
980 SuccessOrder != AtomicOrdering::Acquire &&
981 !F->optForMinSize();
982
983 // There's no overhead for sinking the release barrier in a weak cmpxchg, so
984 // do it even on minsize.
985 bool UseUnconditionalReleaseBarrier = F->optForMinSize() && !CI->isWeak();
986
987 // Given: cmpxchg some_op iN* %addr, iN %desired, iN %new success_ord fail_ord
988 //
989 // The full expansion we produce is:
990 // [...]
991 // cmpxchg.start:
992 // %unreleasedload = @load.linked(%addr)
993 // %should_store = icmp eq %unreleasedload, %desired
994 // br i1 %should_store, label %cmpxchg.fencedstore,
995 // label %cmpxchg.nostore
996 // cmpxchg.releasingstore:
997 // fence?
998 // br label cmpxchg.trystore
999 // cmpxchg.trystore:
1000 // %loaded.trystore = phi [%unreleasedload, %releasingstore],
1001 // [%releasedload, %cmpxchg.releasedload]
1002 // %stored = @store_conditional(%new, %addr)
1003 // %success = icmp eq i32 %stored, 0
1004 // br i1 %success, label %cmpxchg.success,
1005 // label %cmpxchg.releasedload/%cmpxchg.failure
1006 // cmpxchg.releasedload:
1007 // %releasedload = @load.linked(%addr)
1008 // %should_store = icmp eq %releasedload, %desired
1009 // br i1 %should_store, label %cmpxchg.trystore,
1010 // label %cmpxchg.failure
1011 // cmpxchg.success:
1012 // fence?
1013 // br label %cmpxchg.end
1014 // cmpxchg.nostore:
1015 // %loaded.nostore = phi [%unreleasedload, %cmpxchg.start],
1016 // [%releasedload,
1017 // %cmpxchg.releasedload/%cmpxchg.trystore]
1018 // @load_linked_fail_balance()?
1019 // br label %cmpxchg.failure
1020 // cmpxchg.failure:
1021 // fence?
1022 // br label %cmpxchg.end
1023 // cmpxchg.end:
1024 // %loaded = phi [%loaded.nostore, %cmpxchg.failure],
1025 // [%loaded.trystore, %cmpxchg.trystore]
1026 // %success = phi i1 [true, %cmpxchg.success], [false, %cmpxchg.failure]
1027 // %restmp = insertvalue { iN, i1 } undef, iN %loaded, 0
1028 // %res = insertvalue { iN, i1 } %restmp, i1 %success, 1
1029 // [...]
1030 BasicBlock *ExitBB = BB->splitBasicBlock(CI->getIterator(), "cmpxchg.end");
1031 auto FailureBB = BasicBlock::Create(Ctx, "cmpxchg.failure", F, ExitBB);
1032 auto NoStoreBB = BasicBlock::Create(Ctx, "cmpxchg.nostore", F, FailureBB);
1033 auto SuccessBB = BasicBlock::Create(Ctx, "cmpxchg.success", F, NoStoreBB);
1034 auto ReleasedLoadBB =
1035 BasicBlock::Create(Ctx, "cmpxchg.releasedload", F, SuccessBB);
1036 auto TryStoreBB =
1037 BasicBlock::Create(Ctx, "cmpxchg.trystore", F, ReleasedLoadBB);
1038 auto ReleasingStoreBB =
1039 BasicBlock::Create(Ctx, "cmpxchg.fencedstore", F, TryStoreBB);
1040 auto StartBB = BasicBlock::Create(Ctx, "cmpxchg.start", F, ReleasingStoreBB);
1041
1042 // This grabs the DebugLoc from CI
1043 IRBuilder<> Builder(CI);
1044
1045 // The split call above "helpfully" added a branch at the end of BB (to the
1046 // wrong place), but we might want a fence too. It's easiest to just remove
1047 // the branch entirely.
1048 std::prev(BB->end())->eraseFromParent();
1049 Builder.SetInsertPoint(BB);
1050 if (ShouldInsertFencesForAtomic && UseUnconditionalReleaseBarrier)
1051 TLI->emitLeadingFence(Builder, SuccessOrder, /*IsStore=*/true,
1052 /*IsLoad=*/true);
1053 Builder.CreateBr(StartBB);
1054
1055 // Start the main loop block now that we've taken care of the preliminaries.
1056 Builder.SetInsertPoint(StartBB);
1057 Value *UnreleasedLoad = TLI->emitLoadLinked(Builder, Addr, MemOpOrder);
1058 Value *ShouldStore = Builder.CreateICmpEQ(
1059 UnreleasedLoad, CI->getCompareOperand(), "should_store");
1060
1061 // If the cmpxchg doesn't actually need any ordering when it fails, we can
1062 // jump straight past that fence instruction (if it exists).
1063 Builder.CreateCondBr(ShouldStore, ReleasingStoreBB, NoStoreBB);
1064
1065 Builder.SetInsertPoint(ReleasingStoreBB);
1066 if (ShouldInsertFencesForAtomic && !UseUnconditionalReleaseBarrier)
1067 TLI->emitLeadingFence(Builder, SuccessOrder, /*IsStore=*/true,
1068 /*IsLoad=*/true);
1069 Builder.CreateBr(TryStoreBB);
1070
1071 Builder.SetInsertPoint(TryStoreBB);
1072 Value *StoreSuccess = TLI->emitStoreConditional(
1073 Builder, CI->getNewValOperand(), Addr, MemOpOrder);
1074 StoreSuccess = Builder.CreateICmpEQ(
1075 StoreSuccess, ConstantInt::get(Type::getInt32Ty(Ctx), 0), "success");
1076 BasicBlock *RetryBB = HasReleasedLoadBB ? ReleasedLoadBB : StartBB;
1077 Builder.CreateCondBr(StoreSuccess, SuccessBB,
1078 CI->isWeak() ? FailureBB : RetryBB);
1079
1080 Builder.SetInsertPoint(ReleasedLoadBB);
1081 Value *SecondLoad;
1082 if (HasReleasedLoadBB) {
1083 SecondLoad = TLI->emitLoadLinked(Builder, Addr, MemOpOrder);
1084 ShouldStore = Builder.CreateICmpEQ(SecondLoad, CI->getCompareOperand(),
1085 "should_store");
1086
1087 // If the cmpxchg doesn't actually need any ordering when it fails, we can
1088 // jump straight past that fence instruction (if it exists).
1089 Builder.CreateCondBr(ShouldStore, TryStoreBB, NoStoreBB);
1090 } else
1091 Builder.CreateUnreachable();
1092
1093 // Make sure later instructions don't get reordered with a fence if
1094 // necessary.
1095 Builder.SetInsertPoint(SuccessBB);
1096 if (ShouldInsertFencesForAtomic)
1097 TLI->emitTrailingFence(Builder, SuccessOrder, /*IsStore=*/true,
1098 /*IsLoad=*/true);
1099 Builder.CreateBr(ExitBB);
1100
1101 Builder.SetInsertPoint(NoStoreBB);
1102 // In the failing case, where we don't execute the store-conditional, the
1103 // target might want to balance out the load-linked with a dedicated
1104 // instruction (e.g., on ARM, clearing the exclusive monitor).
1105 TLI->emitAtomicCmpXchgNoStoreLLBalance(Builder);
1106 Builder.CreateBr(FailureBB);
1107
1108 Builder.SetInsertPoint(FailureBB);
1109 if (ShouldInsertFencesForAtomic)
1110 TLI->emitTrailingFence(Builder, FailureOrder, /*IsStore=*/true,
1111 /*IsLoad=*/true);
1112 Builder.CreateBr(ExitBB);
1113
1114 // Finally, we have control-flow based knowledge of whether the cmpxchg
1115 // succeeded or not. We expose this to later passes by converting any
1116 // subsequent "icmp eq/ne %loaded, %oldval" into a use of an appropriate
1117 // PHI.
1118 Builder.SetInsertPoint(ExitBB, ExitBB->begin());
1119 PHINode *Success = Builder.CreatePHI(Type::getInt1Ty(Ctx), 2);
1120 Success->addIncoming(ConstantInt::getTrue(Ctx), SuccessBB);
1121 Success->addIncoming(ConstantInt::getFalse(Ctx), FailureBB);
1122
1123 // Setup the builder so we can create any PHIs we need.
1124 Value *Loaded;
1125 if (!HasReleasedLoadBB)
1126 Loaded = UnreleasedLoad;
1127 else {
1128 Builder.SetInsertPoint(TryStoreBB, TryStoreBB->begin());
1129 PHINode *TryStoreLoaded = Builder.CreatePHI(UnreleasedLoad->getType(), 2);
1130 TryStoreLoaded->addIncoming(UnreleasedLoad, ReleasingStoreBB);
1131 TryStoreLoaded->addIncoming(SecondLoad, ReleasedLoadBB);
1132
1133 Builder.SetInsertPoint(NoStoreBB, NoStoreBB->begin());
1134 PHINode *NoStoreLoaded = Builder.CreatePHI(UnreleasedLoad->getType(), 2);
1135 NoStoreLoaded->addIncoming(UnreleasedLoad, StartBB);
1136 NoStoreLoaded->addIncoming(SecondLoad, ReleasedLoadBB);
1137
1138 Builder.SetInsertPoint(ExitBB, ++ExitBB->begin());
1139 PHINode *ExitLoaded = Builder.CreatePHI(UnreleasedLoad->getType(), 2);
1140 ExitLoaded->addIncoming(TryStoreLoaded, SuccessBB);
1141 ExitLoaded->addIncoming(NoStoreLoaded, FailureBB);
1142
1143 Loaded = ExitLoaded;
1144 }
1145
1146 // Look for any users of the cmpxchg that are just comparing the loaded value
1147 // against the desired one, and replace them with the CFG-derived version.
1148 SmallVector<ExtractValueInst *, 2> PrunedInsts;
1149 for (auto User : CI->users()) {
1150 ExtractValueInst *EV = dyn_cast<ExtractValueInst>(User);
1151 if (!EV)
1152 continue;
1153
1154 assert(EV->getNumIndices() == 1 && EV->getIndices()[0] <= 1 &&
1155 "weird extraction from { iN, i1 }");
1156
1157 if (EV->getIndices()[0] == 0)
1158 EV->replaceAllUsesWith(Loaded);
1159 else
1160 EV->replaceAllUsesWith(Success);
1161
1162 PrunedInsts.push_back(EV);
1163 }
1164
1165 // We can remove the instructions now we're no longer iterating through them.
1166 for (auto EV : PrunedInsts)
1167 EV->eraseFromParent();
1168
1169 if (!CI->use_empty()) {
1170 // Some use of the full struct return that we don't understand has happened,
1171 // so we've got to reconstruct it properly.
1172 Value *Res;
1173 Res = Builder.CreateInsertValue(UndefValue::get(CI->getType()), Loaded, 0);
1174 Res = Builder.CreateInsertValue(Res, Success, 1);
1175
1176 CI->replaceAllUsesWith(Res);
1177 }
1178
1179 CI->eraseFromParent();
1180 return true;
1181 }
1182
isIdempotentRMW(AtomicRMWInst * RMWI)1183 bool AtomicExpand::isIdempotentRMW(AtomicRMWInst* RMWI) {
1184 auto C = dyn_cast<ConstantInt>(RMWI->getValOperand());
1185 if(!C)
1186 return false;
1187
1188 AtomicRMWInst::BinOp Op = RMWI->getOperation();
1189 switch(Op) {
1190 case AtomicRMWInst::Add:
1191 case AtomicRMWInst::Sub:
1192 case AtomicRMWInst::Or:
1193 case AtomicRMWInst::Xor:
1194 return C->isZero();
1195 case AtomicRMWInst::And:
1196 return C->isMinusOne();
1197 // FIXME: we could also treat Min/Max/UMin/UMax by the INT_MIN/INT_MAX/...
1198 default:
1199 return false;
1200 }
1201 }
1202
simplifyIdempotentRMW(AtomicRMWInst * RMWI)1203 bool AtomicExpand::simplifyIdempotentRMW(AtomicRMWInst* RMWI) {
1204 if (auto ResultingLoad = TLI->lowerIdempotentRMWIntoFencedLoad(RMWI)) {
1205 tryExpandAtomicLoad(ResultingLoad);
1206 return true;
1207 }
1208 return false;
1209 }
1210
insertRMWCmpXchgLoop(IRBuilder<> & Builder,Type * ResultTy,Value * Addr,AtomicOrdering MemOpOrder,function_ref<Value * (IRBuilder<> &,Value *)> PerformOp,CreateCmpXchgInstFun CreateCmpXchg)1211 Value *AtomicExpand::insertRMWCmpXchgLoop(
1212 IRBuilder<> &Builder, Type *ResultTy, Value *Addr,
1213 AtomicOrdering MemOpOrder,
1214 function_ref<Value *(IRBuilder<> &, Value *)> PerformOp,
1215 CreateCmpXchgInstFun CreateCmpXchg) {
1216 LLVMContext &Ctx = Builder.getContext();
1217 BasicBlock *BB = Builder.GetInsertBlock();
1218 Function *F = BB->getParent();
1219
1220 // Given: atomicrmw some_op iN* %addr, iN %incr ordering
1221 //
1222 // The standard expansion we produce is:
1223 // [...]
1224 // %init_loaded = load atomic iN* %addr
1225 // br label %loop
1226 // loop:
1227 // %loaded = phi iN [ %init_loaded, %entry ], [ %new_loaded, %loop ]
1228 // %new = some_op iN %loaded, %incr
1229 // %pair = cmpxchg iN* %addr, iN %loaded, iN %new
1230 // %new_loaded = extractvalue { iN, i1 } %pair, 0
1231 // %success = extractvalue { iN, i1 } %pair, 1
1232 // br i1 %success, label %atomicrmw.end, label %loop
1233 // atomicrmw.end:
1234 // [...]
1235 BasicBlock *ExitBB =
1236 BB->splitBasicBlock(Builder.GetInsertPoint(), "atomicrmw.end");
1237 BasicBlock *LoopBB = BasicBlock::Create(Ctx, "atomicrmw.start", F, ExitBB);
1238
1239 // The split call above "helpfully" added a branch at the end of BB (to the
1240 // wrong place), but we want a load. It's easiest to just remove
1241 // the branch entirely.
1242 std::prev(BB->end())->eraseFromParent();
1243 Builder.SetInsertPoint(BB);
1244 LoadInst *InitLoaded = Builder.CreateLoad(ResultTy, Addr);
1245 // Atomics require at least natural alignment.
1246 InitLoaded->setAlignment(ResultTy->getPrimitiveSizeInBits() / 8);
1247 Builder.CreateBr(LoopBB);
1248
1249 // Start the main loop block now that we've taken care of the preliminaries.
1250 Builder.SetInsertPoint(LoopBB);
1251 PHINode *Loaded = Builder.CreatePHI(ResultTy, 2, "loaded");
1252 Loaded->addIncoming(InitLoaded, BB);
1253
1254 Value *NewVal = PerformOp(Builder, Loaded);
1255
1256 Value *NewLoaded = nullptr;
1257 Value *Success = nullptr;
1258
1259 CreateCmpXchg(Builder, Addr, Loaded, NewVal,
1260 MemOpOrder == AtomicOrdering::Unordered
1261 ? AtomicOrdering::Monotonic
1262 : MemOpOrder,
1263 Success, NewLoaded);
1264 assert(Success && NewLoaded);
1265
1266 Loaded->addIncoming(NewLoaded, LoopBB);
1267
1268 Builder.CreateCondBr(Success, ExitBB, LoopBB);
1269
1270 Builder.SetInsertPoint(ExitBB, ExitBB->begin());
1271 return NewLoaded;
1272 }
1273
1274 // Note: This function is exposed externally by AtomicExpandUtils.h
expandAtomicRMWToCmpXchg(AtomicRMWInst * AI,CreateCmpXchgInstFun CreateCmpXchg)1275 bool llvm::expandAtomicRMWToCmpXchg(AtomicRMWInst *AI,
1276 CreateCmpXchgInstFun CreateCmpXchg) {
1277 IRBuilder<> Builder(AI);
1278 Value *Loaded = AtomicExpand::insertRMWCmpXchgLoop(
1279 Builder, AI->getType(), AI->getPointerOperand(), AI->getOrdering(),
1280 [&](IRBuilder<> &Builder, Value *Loaded) {
1281 return performAtomicOp(AI->getOperation(), Builder, Loaded,
1282 AI->getValOperand());
1283 },
1284 CreateCmpXchg);
1285
1286 AI->replaceAllUsesWith(Loaded);
1287 AI->eraseFromParent();
1288 return true;
1289 }
1290
1291 // In order to use one of the sized library calls such as
1292 // __atomic_fetch_add_4, the alignment must be sufficient, the size
1293 // must be one of the potentially-specialized sizes, and the value
1294 // type must actually exist in C on the target (otherwise, the
1295 // function wouldn't actually be defined.)
canUseSizedAtomicCall(unsigned Size,unsigned Align,const DataLayout & DL)1296 static bool canUseSizedAtomicCall(unsigned Size, unsigned Align,
1297 const DataLayout &DL) {
1298 // TODO: "LargestSize" is an approximation for "largest type that
1299 // you can express in C". It seems to be the case that int128 is
1300 // supported on all 64-bit platforms, otherwise only up to 64-bit
1301 // integers are supported. If we get this wrong, then we'll try to
1302 // call a sized libcall that doesn't actually exist. There should
1303 // really be some more reliable way in LLVM of determining integer
1304 // sizes which are valid in the target's C ABI...
1305 unsigned LargestSize = DL.getLargestLegalIntTypeSizeInBits() >= 64 ? 16 : 8;
1306 return Align >= Size &&
1307 (Size == 1 || Size == 2 || Size == 4 || Size == 8 || Size == 16) &&
1308 Size <= LargestSize;
1309 }
1310
expandAtomicLoadToLibcall(LoadInst * I)1311 void AtomicExpand::expandAtomicLoadToLibcall(LoadInst *I) {
1312 static const RTLIB::Libcall Libcalls[6] = {
1313 RTLIB::ATOMIC_LOAD, RTLIB::ATOMIC_LOAD_1, RTLIB::ATOMIC_LOAD_2,
1314 RTLIB::ATOMIC_LOAD_4, RTLIB::ATOMIC_LOAD_8, RTLIB::ATOMIC_LOAD_16};
1315 unsigned Size = getAtomicOpSize(I);
1316 unsigned Align = getAtomicOpAlign(I);
1317
1318 bool expanded = expandAtomicOpToLibcall(
1319 I, Size, Align, I->getPointerOperand(), nullptr, nullptr,
1320 I->getOrdering(), AtomicOrdering::NotAtomic, Libcalls);
1321 (void)expanded;
1322 assert(expanded && "expandAtomicOpToLibcall shouldn't fail tor Load");
1323 }
1324
expandAtomicStoreToLibcall(StoreInst * I)1325 void AtomicExpand::expandAtomicStoreToLibcall(StoreInst *I) {
1326 static const RTLIB::Libcall Libcalls[6] = {
1327 RTLIB::ATOMIC_STORE, RTLIB::ATOMIC_STORE_1, RTLIB::ATOMIC_STORE_2,
1328 RTLIB::ATOMIC_STORE_4, RTLIB::ATOMIC_STORE_8, RTLIB::ATOMIC_STORE_16};
1329 unsigned Size = getAtomicOpSize(I);
1330 unsigned Align = getAtomicOpAlign(I);
1331
1332 bool expanded = expandAtomicOpToLibcall(
1333 I, Size, Align, I->getPointerOperand(), I->getValueOperand(), nullptr,
1334 I->getOrdering(), AtomicOrdering::NotAtomic, Libcalls);
1335 (void)expanded;
1336 assert(expanded && "expandAtomicOpToLibcall shouldn't fail tor Store");
1337 }
1338
expandAtomicCASToLibcall(AtomicCmpXchgInst * I)1339 void AtomicExpand::expandAtomicCASToLibcall(AtomicCmpXchgInst *I) {
1340 static const RTLIB::Libcall Libcalls[6] = {
1341 RTLIB::ATOMIC_COMPARE_EXCHANGE, RTLIB::ATOMIC_COMPARE_EXCHANGE_1,
1342 RTLIB::ATOMIC_COMPARE_EXCHANGE_2, RTLIB::ATOMIC_COMPARE_EXCHANGE_4,
1343 RTLIB::ATOMIC_COMPARE_EXCHANGE_8, RTLIB::ATOMIC_COMPARE_EXCHANGE_16};
1344 unsigned Size = getAtomicOpSize(I);
1345 unsigned Align = getAtomicOpAlign(I);
1346
1347 bool expanded = expandAtomicOpToLibcall(
1348 I, Size, Align, I->getPointerOperand(), I->getNewValOperand(),
1349 I->getCompareOperand(), I->getSuccessOrdering(), I->getFailureOrdering(),
1350 Libcalls);
1351 (void)expanded;
1352 assert(expanded && "expandAtomicOpToLibcall shouldn't fail tor CAS");
1353 }
1354
GetRMWLibcall(AtomicRMWInst::BinOp Op)1355 static ArrayRef<RTLIB::Libcall> GetRMWLibcall(AtomicRMWInst::BinOp Op) {
1356 static const RTLIB::Libcall LibcallsXchg[6] = {
1357 RTLIB::ATOMIC_EXCHANGE, RTLIB::ATOMIC_EXCHANGE_1,
1358 RTLIB::ATOMIC_EXCHANGE_2, RTLIB::ATOMIC_EXCHANGE_4,
1359 RTLIB::ATOMIC_EXCHANGE_8, RTLIB::ATOMIC_EXCHANGE_16};
1360 static const RTLIB::Libcall LibcallsAdd[6] = {
1361 RTLIB::UNKNOWN_LIBCALL, RTLIB::ATOMIC_FETCH_ADD_1,
1362 RTLIB::ATOMIC_FETCH_ADD_2, RTLIB::ATOMIC_FETCH_ADD_4,
1363 RTLIB::ATOMIC_FETCH_ADD_8, RTLIB::ATOMIC_FETCH_ADD_16};
1364 static const RTLIB::Libcall LibcallsSub[6] = {
1365 RTLIB::UNKNOWN_LIBCALL, RTLIB::ATOMIC_FETCH_SUB_1,
1366 RTLIB::ATOMIC_FETCH_SUB_2, RTLIB::ATOMIC_FETCH_SUB_4,
1367 RTLIB::ATOMIC_FETCH_SUB_8, RTLIB::ATOMIC_FETCH_SUB_16};
1368 static const RTLIB::Libcall LibcallsAnd[6] = {
1369 RTLIB::UNKNOWN_LIBCALL, RTLIB::ATOMIC_FETCH_AND_1,
1370 RTLIB::ATOMIC_FETCH_AND_2, RTLIB::ATOMIC_FETCH_AND_4,
1371 RTLIB::ATOMIC_FETCH_AND_8, RTLIB::ATOMIC_FETCH_AND_16};
1372 static const RTLIB::Libcall LibcallsOr[6] = {
1373 RTLIB::UNKNOWN_LIBCALL, RTLIB::ATOMIC_FETCH_OR_1,
1374 RTLIB::ATOMIC_FETCH_OR_2, RTLIB::ATOMIC_FETCH_OR_4,
1375 RTLIB::ATOMIC_FETCH_OR_8, RTLIB::ATOMIC_FETCH_OR_16};
1376 static const RTLIB::Libcall LibcallsXor[6] = {
1377 RTLIB::UNKNOWN_LIBCALL, RTLIB::ATOMIC_FETCH_XOR_1,
1378 RTLIB::ATOMIC_FETCH_XOR_2, RTLIB::ATOMIC_FETCH_XOR_4,
1379 RTLIB::ATOMIC_FETCH_XOR_8, RTLIB::ATOMIC_FETCH_XOR_16};
1380 static const RTLIB::Libcall LibcallsNand[6] = {
1381 RTLIB::UNKNOWN_LIBCALL, RTLIB::ATOMIC_FETCH_NAND_1,
1382 RTLIB::ATOMIC_FETCH_NAND_2, RTLIB::ATOMIC_FETCH_NAND_4,
1383 RTLIB::ATOMIC_FETCH_NAND_8, RTLIB::ATOMIC_FETCH_NAND_16};
1384
1385 switch (Op) {
1386 case AtomicRMWInst::BAD_BINOP:
1387 llvm_unreachable("Should not have BAD_BINOP.");
1388 case AtomicRMWInst::Xchg:
1389 return makeArrayRef(LibcallsXchg);
1390 case AtomicRMWInst::Add:
1391 return makeArrayRef(LibcallsAdd);
1392 case AtomicRMWInst::Sub:
1393 return makeArrayRef(LibcallsSub);
1394 case AtomicRMWInst::And:
1395 return makeArrayRef(LibcallsAnd);
1396 case AtomicRMWInst::Or:
1397 return makeArrayRef(LibcallsOr);
1398 case AtomicRMWInst::Xor:
1399 return makeArrayRef(LibcallsXor);
1400 case AtomicRMWInst::Nand:
1401 return makeArrayRef(LibcallsNand);
1402 case AtomicRMWInst::Max:
1403 case AtomicRMWInst::Min:
1404 case AtomicRMWInst::UMax:
1405 case AtomicRMWInst::UMin:
1406 // No atomic libcalls are available for max/min/umax/umin.
1407 return {};
1408 }
1409 llvm_unreachable("Unexpected AtomicRMW operation.");
1410 }
1411
expandAtomicRMWToLibcall(AtomicRMWInst * I)1412 void AtomicExpand::expandAtomicRMWToLibcall(AtomicRMWInst *I) {
1413 ArrayRef<RTLIB::Libcall> Libcalls = GetRMWLibcall(I->getOperation());
1414
1415 unsigned Size = getAtomicOpSize(I);
1416 unsigned Align = getAtomicOpAlign(I);
1417
1418 bool Success = false;
1419 if (!Libcalls.empty())
1420 Success = expandAtomicOpToLibcall(
1421 I, Size, Align, I->getPointerOperand(), I->getValOperand(), nullptr,
1422 I->getOrdering(), AtomicOrdering::NotAtomic, Libcalls);
1423
1424 // The expansion failed: either there were no libcalls at all for
1425 // the operation (min/max), or there were only size-specialized
1426 // libcalls (add/sub/etc) and we needed a generic. So, expand to a
1427 // CAS libcall, via a CAS loop, instead.
1428 if (!Success) {
1429 expandAtomicRMWToCmpXchg(I, [this](IRBuilder<> &Builder, Value *Addr,
1430 Value *Loaded, Value *NewVal,
1431 AtomicOrdering MemOpOrder,
1432 Value *&Success, Value *&NewLoaded) {
1433 // Create the CAS instruction normally...
1434 AtomicCmpXchgInst *Pair = Builder.CreateAtomicCmpXchg(
1435 Addr, Loaded, NewVal, MemOpOrder,
1436 AtomicCmpXchgInst::getStrongestFailureOrdering(MemOpOrder));
1437 Success = Builder.CreateExtractValue(Pair, 1, "success");
1438 NewLoaded = Builder.CreateExtractValue(Pair, 0, "newloaded");
1439
1440 // ...and then expand the CAS into a libcall.
1441 expandAtomicCASToLibcall(Pair);
1442 });
1443 }
1444 }
1445
1446 // A helper routine for the above expandAtomic*ToLibcall functions.
1447 //
1448 // 'Libcalls' contains an array of enum values for the particular
1449 // ATOMIC libcalls to be emitted. All of the other arguments besides
1450 // 'I' are extracted from the Instruction subclass by the
1451 // caller. Depending on the particular call, some will be null.
expandAtomicOpToLibcall(Instruction * I,unsigned Size,unsigned Align,Value * PointerOperand,Value * ValueOperand,Value * CASExpected,AtomicOrdering Ordering,AtomicOrdering Ordering2,ArrayRef<RTLIB::Libcall> Libcalls)1452 bool AtomicExpand::expandAtomicOpToLibcall(
1453 Instruction *I, unsigned Size, unsigned Align, Value *PointerOperand,
1454 Value *ValueOperand, Value *CASExpected, AtomicOrdering Ordering,
1455 AtomicOrdering Ordering2, ArrayRef<RTLIB::Libcall> Libcalls) {
1456 assert(Libcalls.size() == 6);
1457
1458 LLVMContext &Ctx = I->getContext();
1459 Module *M = I->getModule();
1460 const DataLayout &DL = M->getDataLayout();
1461 IRBuilder<> Builder(I);
1462 IRBuilder<> AllocaBuilder(&I->getFunction()->getEntryBlock().front());
1463
1464 bool UseSizedLibcall = canUseSizedAtomicCall(Size, Align, DL);
1465 Type *SizedIntTy = Type::getIntNTy(Ctx, Size * 8);
1466
1467 unsigned AllocaAlignment = DL.getPrefTypeAlignment(SizedIntTy);
1468
1469 // TODO: the "order" argument type is "int", not int32. So
1470 // getInt32Ty may be wrong if the arch uses e.g. 16-bit ints.
1471 ConstantInt *SizeVal64 = ConstantInt::get(Type::getInt64Ty(Ctx), Size);
1472 assert(Ordering != AtomicOrdering::NotAtomic && "expect atomic MO");
1473 Constant *OrderingVal =
1474 ConstantInt::get(Type::getInt32Ty(Ctx), (int)toCABI(Ordering));
1475 Constant *Ordering2Val = nullptr;
1476 if (CASExpected) {
1477 assert(Ordering2 != AtomicOrdering::NotAtomic && "expect atomic MO");
1478 Ordering2Val =
1479 ConstantInt::get(Type::getInt32Ty(Ctx), (int)toCABI(Ordering2));
1480 }
1481 bool HasResult = I->getType() != Type::getVoidTy(Ctx);
1482
1483 RTLIB::Libcall RTLibType;
1484 if (UseSizedLibcall) {
1485 switch (Size) {
1486 case 1: RTLibType = Libcalls[1]; break;
1487 case 2: RTLibType = Libcalls[2]; break;
1488 case 4: RTLibType = Libcalls[3]; break;
1489 case 8: RTLibType = Libcalls[4]; break;
1490 case 16: RTLibType = Libcalls[5]; break;
1491 }
1492 } else if (Libcalls[0] != RTLIB::UNKNOWN_LIBCALL) {
1493 RTLibType = Libcalls[0];
1494 } else {
1495 // Can't use sized function, and there's no generic for this
1496 // operation, so give up.
1497 return false;
1498 }
1499
1500 // Build up the function call. There's two kinds. First, the sized
1501 // variants. These calls are going to be one of the following (with
1502 // N=1,2,4,8,16):
1503 // iN __atomic_load_N(iN *ptr, int ordering)
1504 // void __atomic_store_N(iN *ptr, iN val, int ordering)
1505 // iN __atomic_{exchange|fetch_*}_N(iN *ptr, iN val, int ordering)
1506 // bool __atomic_compare_exchange_N(iN *ptr, iN *expected, iN desired,
1507 // int success_order, int failure_order)
1508 //
1509 // Note that these functions can be used for non-integer atomic
1510 // operations, the values just need to be bitcast to integers on the
1511 // way in and out.
1512 //
1513 // And, then, the generic variants. They look like the following:
1514 // void __atomic_load(size_t size, void *ptr, void *ret, int ordering)
1515 // void __atomic_store(size_t size, void *ptr, void *val, int ordering)
1516 // void __atomic_exchange(size_t size, void *ptr, void *val, void *ret,
1517 // int ordering)
1518 // bool __atomic_compare_exchange(size_t size, void *ptr, void *expected,
1519 // void *desired, int success_order,
1520 // int failure_order)
1521 //
1522 // The different signatures are built up depending on the
1523 // 'UseSizedLibcall', 'CASExpected', 'ValueOperand', and 'HasResult'
1524 // variables.
1525
1526 AllocaInst *AllocaCASExpected = nullptr;
1527 Value *AllocaCASExpected_i8 = nullptr;
1528 AllocaInst *AllocaValue = nullptr;
1529 Value *AllocaValue_i8 = nullptr;
1530 AllocaInst *AllocaResult = nullptr;
1531 Value *AllocaResult_i8 = nullptr;
1532
1533 Type *ResultTy;
1534 SmallVector<Value *, 6> Args;
1535 AttributeSet Attr;
1536
1537 // 'size' argument.
1538 if (!UseSizedLibcall) {
1539 // Note, getIntPtrType is assumed equivalent to size_t.
1540 Args.push_back(ConstantInt::get(DL.getIntPtrType(Ctx), Size));
1541 }
1542
1543 // 'ptr' argument.
1544 Value *PtrVal =
1545 Builder.CreateBitCast(PointerOperand, Type::getInt8PtrTy(Ctx));
1546 Args.push_back(PtrVal);
1547
1548 // 'expected' argument, if present.
1549 if (CASExpected) {
1550 AllocaCASExpected = AllocaBuilder.CreateAlloca(CASExpected->getType());
1551 AllocaCASExpected->setAlignment(AllocaAlignment);
1552 AllocaCASExpected_i8 =
1553 Builder.CreateBitCast(AllocaCASExpected, Type::getInt8PtrTy(Ctx));
1554 Builder.CreateLifetimeStart(AllocaCASExpected_i8, SizeVal64);
1555 Builder.CreateAlignedStore(CASExpected, AllocaCASExpected, AllocaAlignment);
1556 Args.push_back(AllocaCASExpected_i8);
1557 }
1558
1559 // 'val' argument ('desired' for cas), if present.
1560 if (ValueOperand) {
1561 if (UseSizedLibcall) {
1562 Value *IntValue =
1563 Builder.CreateBitOrPointerCast(ValueOperand, SizedIntTy);
1564 Args.push_back(IntValue);
1565 } else {
1566 AllocaValue = AllocaBuilder.CreateAlloca(ValueOperand->getType());
1567 AllocaValue->setAlignment(AllocaAlignment);
1568 AllocaValue_i8 =
1569 Builder.CreateBitCast(AllocaValue, Type::getInt8PtrTy(Ctx));
1570 Builder.CreateLifetimeStart(AllocaValue_i8, SizeVal64);
1571 Builder.CreateAlignedStore(ValueOperand, AllocaValue, AllocaAlignment);
1572 Args.push_back(AllocaValue_i8);
1573 }
1574 }
1575
1576 // 'ret' argument.
1577 if (!CASExpected && HasResult && !UseSizedLibcall) {
1578 AllocaResult = AllocaBuilder.CreateAlloca(I->getType());
1579 AllocaResult->setAlignment(AllocaAlignment);
1580 AllocaResult_i8 =
1581 Builder.CreateBitCast(AllocaResult, Type::getInt8PtrTy(Ctx));
1582 Builder.CreateLifetimeStart(AllocaResult_i8, SizeVal64);
1583 Args.push_back(AllocaResult_i8);
1584 }
1585
1586 // 'ordering' ('success_order' for cas) argument.
1587 Args.push_back(OrderingVal);
1588
1589 // 'failure_order' argument, if present.
1590 if (Ordering2Val)
1591 Args.push_back(Ordering2Val);
1592
1593 // Now, the return type.
1594 if (CASExpected) {
1595 ResultTy = Type::getInt1Ty(Ctx);
1596 Attr = Attr.addAttribute(Ctx, AttributeSet::ReturnIndex, Attribute::ZExt);
1597 } else if (HasResult && UseSizedLibcall)
1598 ResultTy = SizedIntTy;
1599 else
1600 ResultTy = Type::getVoidTy(Ctx);
1601
1602 // Done with setting up arguments and return types, create the call:
1603 SmallVector<Type *, 6> ArgTys;
1604 for (Value *Arg : Args)
1605 ArgTys.push_back(Arg->getType());
1606 FunctionType *FnType = FunctionType::get(ResultTy, ArgTys, false);
1607 Constant *LibcallFn =
1608 M->getOrInsertFunction(TLI->getLibcallName(RTLibType), FnType, Attr);
1609 CallInst *Call = Builder.CreateCall(LibcallFn, Args);
1610 Call->setAttributes(Attr);
1611 Value *Result = Call;
1612
1613 // And then, extract the results...
1614 if (ValueOperand && !UseSizedLibcall)
1615 Builder.CreateLifetimeEnd(AllocaValue_i8, SizeVal64);
1616
1617 if (CASExpected) {
1618 // The final result from the CAS is {load of 'expected' alloca, bool result
1619 // from call}
1620 Type *FinalResultTy = I->getType();
1621 Value *V = UndefValue::get(FinalResultTy);
1622 Value *ExpectedOut =
1623 Builder.CreateAlignedLoad(AllocaCASExpected, AllocaAlignment);
1624 Builder.CreateLifetimeEnd(AllocaCASExpected_i8, SizeVal64);
1625 V = Builder.CreateInsertValue(V, ExpectedOut, 0);
1626 V = Builder.CreateInsertValue(V, Result, 1);
1627 I->replaceAllUsesWith(V);
1628 } else if (HasResult) {
1629 Value *V;
1630 if (UseSizedLibcall)
1631 V = Builder.CreateBitOrPointerCast(Result, I->getType());
1632 else {
1633 V = Builder.CreateAlignedLoad(AllocaResult, AllocaAlignment);
1634 Builder.CreateLifetimeEnd(AllocaResult_i8, SizeVal64);
1635 }
1636 I->replaceAllUsesWith(V);
1637 }
1638 I->eraseFromParent();
1639 return true;
1640 }
1641