1 //===- LowerMemIntrinsics.cpp ----------------------------------*- C++ -*--===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 
9 #include "llvm/Transforms/Utils/LowerMemIntrinsics.h"
10 #include "llvm/Analysis/ScalarEvolution.h"
11 #include "llvm/Analysis/TargetTransformInfo.h"
12 #include "llvm/IR/IRBuilder.h"
13 #include "llvm/IR/IntrinsicInst.h"
14 #include "llvm/IR/MDBuilder.h"
15 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
16 
17 using namespace llvm;
18 
19 void llvm::createMemCpyLoopKnownSize(Instruction *InsertBefore, Value *SrcAddr,
20                                      Value *DstAddr, ConstantInt *CopyLen,
21                                      Align SrcAlign, Align DstAlign,
22                                      bool SrcIsVolatile, bool DstIsVolatile,
23                                      bool CanOverlap,
24                                      const TargetTransformInfo &TTI,
25                                      Optional<uint32_t> AtomicElementSize) {
26   // No need to expand zero length copies.
27   if (CopyLen->isZero())
28     return;
29 
30   BasicBlock *PreLoopBB = InsertBefore->getParent();
31   BasicBlock *PostLoopBB = nullptr;
32   Function *ParentFunc = PreLoopBB->getParent();
33   LLVMContext &Ctx = PreLoopBB->getContext();
34   const DataLayout &DL = ParentFunc->getParent()->getDataLayout();
35   MDBuilder MDB(Ctx);
36   MDNode *NewDomain = MDB.createAnonymousAliasScopeDomain("MemCopyDomain");
37   StringRef Name = "MemCopyAliasScope";
38   MDNode *NewScope = MDB.createAnonymousAliasScope(NewDomain, Name);
39 
40   unsigned SrcAS = cast<PointerType>(SrcAddr->getType())->getAddressSpace();
41   unsigned DstAS = cast<PointerType>(DstAddr->getType())->getAddressSpace();
42 
43   Type *TypeOfCopyLen = CopyLen->getType();
44   Type *LoopOpType = TTI.getMemcpyLoopLoweringType(
45       Ctx, CopyLen, SrcAS, DstAS, SrcAlign.value(), DstAlign.value(),
46       AtomicElementSize);
47   assert((!AtomicElementSize || !LoopOpType->isVectorTy()) &&
48          "Atomic memcpy lowering is not supported for vector operand type");
49 
50   unsigned LoopOpSize = DL.getTypeStoreSize(LoopOpType);
51   assert((!AtomicElementSize || LoopOpSize % *AtomicElementSize == 0) &&
52       "Atomic memcpy lowering is not supported for selected operand size");
53 
54   uint64_t LoopEndCount = CopyLen->getZExtValue() / LoopOpSize;
55 
56   if (LoopEndCount != 0) {
57     // Split
58     PostLoopBB = PreLoopBB->splitBasicBlock(InsertBefore, "memcpy-split");
59     BasicBlock *LoopBB =
60         BasicBlock::Create(Ctx, "load-store-loop", ParentFunc, PostLoopBB);
61     PreLoopBB->getTerminator()->setSuccessor(0, LoopBB);
62 
63     IRBuilder<> PLBuilder(PreLoopBB->getTerminator());
64 
65     // Cast the Src and Dst pointers to pointers to the loop operand type (if
66     // needed).
67     PointerType *SrcOpType = PointerType::get(LoopOpType, SrcAS);
68     PointerType *DstOpType = PointerType::get(LoopOpType, DstAS);
69     if (SrcAddr->getType() != SrcOpType) {
70       SrcAddr = PLBuilder.CreateBitCast(SrcAddr, SrcOpType);
71     }
72     if (DstAddr->getType() != DstOpType) {
73       DstAddr = PLBuilder.CreateBitCast(DstAddr, DstOpType);
74     }
75 
76     Align PartDstAlign(commonAlignment(DstAlign, LoopOpSize));
77     Align PartSrcAlign(commonAlignment(SrcAlign, LoopOpSize));
78 
79     IRBuilder<> LoopBuilder(LoopBB);
80     PHINode *LoopIndex = LoopBuilder.CreatePHI(TypeOfCopyLen, 2, "loop-index");
81     LoopIndex->addIncoming(ConstantInt::get(TypeOfCopyLen, 0U), PreLoopBB);
82     // Loop Body
83     Value *SrcGEP =
84         LoopBuilder.CreateInBoundsGEP(LoopOpType, SrcAddr, LoopIndex);
85     LoadInst *Load = LoopBuilder.CreateAlignedLoad(LoopOpType, SrcGEP,
86                                                    PartSrcAlign, SrcIsVolatile);
87     if (!CanOverlap) {
88       // Set alias scope for loads.
89       Load->setMetadata(LLVMContext::MD_alias_scope,
90                         MDNode::get(Ctx, NewScope));
91     }
92     Value *DstGEP =
93         LoopBuilder.CreateInBoundsGEP(LoopOpType, DstAddr, LoopIndex);
94     StoreInst *Store = LoopBuilder.CreateAlignedStore(
95         Load, DstGEP, PartDstAlign, DstIsVolatile);
96     if (!CanOverlap) {
97       // Indicate that stores don't overlap loads.
98       Store->setMetadata(LLVMContext::MD_noalias, MDNode::get(Ctx, NewScope));
99     }
100     if (AtomicElementSize) {
101       Load->setAtomic(AtomicOrdering::Unordered);
102       Store->setAtomic(AtomicOrdering::Unordered);
103     }
104     Value *NewIndex =
105         LoopBuilder.CreateAdd(LoopIndex, ConstantInt::get(TypeOfCopyLen, 1U));
106     LoopIndex->addIncoming(NewIndex, LoopBB);
107 
108     // Create the loop branch condition.
109     Constant *LoopEndCI = ConstantInt::get(TypeOfCopyLen, LoopEndCount);
110     LoopBuilder.CreateCondBr(LoopBuilder.CreateICmpULT(NewIndex, LoopEndCI),
111                              LoopBB, PostLoopBB);
112   }
113 
114   uint64_t BytesCopied = LoopEndCount * LoopOpSize;
115   uint64_t RemainingBytes = CopyLen->getZExtValue() - BytesCopied;
116   if (RemainingBytes) {
117     IRBuilder<> RBuilder(PostLoopBB ? PostLoopBB->getFirstNonPHI()
118                                     : InsertBefore);
119 
120     SmallVector<Type *, 5> RemainingOps;
121     TTI.getMemcpyLoopResidualLoweringType(RemainingOps, Ctx, RemainingBytes,
122                                           SrcAS, DstAS, SrcAlign.value(),
123                                           DstAlign.value(), AtomicElementSize);
124 
125     for (auto OpTy : RemainingOps) {
126       Align PartSrcAlign(commonAlignment(SrcAlign, BytesCopied));
127       Align PartDstAlign(commonAlignment(DstAlign, BytesCopied));
128 
129       // Calaculate the new index
130       unsigned OperandSize = DL.getTypeStoreSize(OpTy);
131       assert(
132           (!AtomicElementSize || OperandSize % *AtomicElementSize == 0) &&
133           "Atomic memcpy lowering is not supported for selected operand size");
134 
135       uint64_t GepIndex = BytesCopied / OperandSize;
136       assert(GepIndex * OperandSize == BytesCopied &&
137              "Division should have no Remainder!");
138       // Cast source to operand type and load
139       PointerType *SrcPtrType = PointerType::get(OpTy, SrcAS);
140       Value *CastedSrc = SrcAddr->getType() == SrcPtrType
141                              ? SrcAddr
142                              : RBuilder.CreateBitCast(SrcAddr, SrcPtrType);
143       Value *SrcGEP = RBuilder.CreateInBoundsGEP(
144           OpTy, CastedSrc, ConstantInt::get(TypeOfCopyLen, GepIndex));
145       LoadInst *Load =
146           RBuilder.CreateAlignedLoad(OpTy, SrcGEP, PartSrcAlign, SrcIsVolatile);
147       if (!CanOverlap) {
148         // Set alias scope for loads.
149         Load->setMetadata(LLVMContext::MD_alias_scope,
150                           MDNode::get(Ctx, NewScope));
151       }
152       // Cast destination to operand type and store.
153       PointerType *DstPtrType = PointerType::get(OpTy, DstAS);
154       Value *CastedDst = DstAddr->getType() == DstPtrType
155                              ? DstAddr
156                              : RBuilder.CreateBitCast(DstAddr, DstPtrType);
157       Value *DstGEP = RBuilder.CreateInBoundsGEP(
158           OpTy, CastedDst, ConstantInt::get(TypeOfCopyLen, GepIndex));
159       StoreInst *Store = RBuilder.CreateAlignedStore(Load, DstGEP, PartDstAlign,
160                                                      DstIsVolatile);
161       if (!CanOverlap) {
162         // Indicate that stores don't overlap loads.
163         Store->setMetadata(LLVMContext::MD_noalias, MDNode::get(Ctx, NewScope));
164       }
165       if (AtomicElementSize) {
166         Load->setAtomic(AtomicOrdering::Unordered);
167         Store->setAtomic(AtomicOrdering::Unordered);
168       }
169       BytesCopied += OperandSize;
170     }
171   }
172   assert(BytesCopied == CopyLen->getZExtValue() &&
173          "Bytes copied should match size in the call!");
174 }
175 
176 void llvm::createMemCpyLoopUnknownSize(Instruction *InsertBefore,
177                                        Value *SrcAddr, Value *DstAddr,
178                                        Value *CopyLen, Align SrcAlign,
179                                        Align DstAlign, bool SrcIsVolatile,
180                                        bool DstIsVolatile, bool CanOverlap,
181                                        const TargetTransformInfo &TTI,
182                                        Optional<uint32_t> AtomicElementSize) {
183   BasicBlock *PreLoopBB = InsertBefore->getParent();
184   BasicBlock *PostLoopBB =
185       PreLoopBB->splitBasicBlock(InsertBefore, "post-loop-memcpy-expansion");
186 
187   Function *ParentFunc = PreLoopBB->getParent();
188   const DataLayout &DL = ParentFunc->getParent()->getDataLayout();
189   LLVMContext &Ctx = PreLoopBB->getContext();
190   MDBuilder MDB(Ctx);
191   MDNode *NewDomain = MDB.createAnonymousAliasScopeDomain("MemCopyDomain");
192   StringRef Name = "MemCopyAliasScope";
193   MDNode *NewScope = MDB.createAnonymousAliasScope(NewDomain, Name);
194 
195   unsigned SrcAS = cast<PointerType>(SrcAddr->getType())->getAddressSpace();
196   unsigned DstAS = cast<PointerType>(DstAddr->getType())->getAddressSpace();
197 
198   Type *LoopOpType = TTI.getMemcpyLoopLoweringType(
199       Ctx, CopyLen, SrcAS, DstAS, SrcAlign.value(), DstAlign.value(),
200       AtomicElementSize);
201   assert((!AtomicElementSize || !LoopOpType->isVectorTy()) &&
202          "Atomic memcpy lowering is not supported for vector operand type");
203   unsigned LoopOpSize = DL.getTypeStoreSize(LoopOpType);
204   assert((!AtomicElementSize || LoopOpSize % *AtomicElementSize == 0) &&
205          "Atomic memcpy lowering is not supported for selected operand size");
206 
207   IRBuilder<> PLBuilder(PreLoopBB->getTerminator());
208 
209   PointerType *SrcOpType = PointerType::get(LoopOpType, SrcAS);
210   PointerType *DstOpType = PointerType::get(LoopOpType, DstAS);
211   if (SrcAddr->getType() != SrcOpType) {
212     SrcAddr = PLBuilder.CreateBitCast(SrcAddr, SrcOpType);
213   }
214   if (DstAddr->getType() != DstOpType) {
215     DstAddr = PLBuilder.CreateBitCast(DstAddr, DstOpType);
216   }
217 
218   // Calculate the loop trip count, and remaining bytes to copy after the loop.
219   Type *CopyLenType = CopyLen->getType();
220   IntegerType *ILengthType = dyn_cast<IntegerType>(CopyLenType);
221   assert(ILengthType &&
222          "expected size argument to memcpy to be an integer type!");
223   Type *Int8Type = Type::getInt8Ty(Ctx);
224   bool LoopOpIsInt8 = LoopOpType == Int8Type;
225   ConstantInt *CILoopOpSize = ConstantInt::get(ILengthType, LoopOpSize);
226   Value *RuntimeLoopCount = LoopOpIsInt8 ?
227                             CopyLen :
228                             PLBuilder.CreateUDiv(CopyLen, CILoopOpSize);
229   BasicBlock *LoopBB =
230       BasicBlock::Create(Ctx, "loop-memcpy-expansion", ParentFunc, PostLoopBB);
231   IRBuilder<> LoopBuilder(LoopBB);
232 
233   Align PartSrcAlign(commonAlignment(SrcAlign, LoopOpSize));
234   Align PartDstAlign(commonAlignment(DstAlign, LoopOpSize));
235 
236   PHINode *LoopIndex = LoopBuilder.CreatePHI(CopyLenType, 2, "loop-index");
237   LoopIndex->addIncoming(ConstantInt::get(CopyLenType, 0U), PreLoopBB);
238 
239   Value *SrcGEP = LoopBuilder.CreateInBoundsGEP(LoopOpType, SrcAddr, LoopIndex);
240   LoadInst *Load = LoopBuilder.CreateAlignedLoad(LoopOpType, SrcGEP,
241                                                  PartSrcAlign, SrcIsVolatile);
242   if (!CanOverlap) {
243     // Set alias scope for loads.
244     Load->setMetadata(LLVMContext::MD_alias_scope, MDNode::get(Ctx, NewScope));
245   }
246   Value *DstGEP = LoopBuilder.CreateInBoundsGEP(LoopOpType, DstAddr, LoopIndex);
247   StoreInst *Store =
248       LoopBuilder.CreateAlignedStore(Load, DstGEP, PartDstAlign, DstIsVolatile);
249   if (!CanOverlap) {
250     // Indicate that stores don't overlap loads.
251     Store->setMetadata(LLVMContext::MD_noalias, MDNode::get(Ctx, NewScope));
252   }
253   if (AtomicElementSize) {
254     Load->setAtomic(AtomicOrdering::Unordered);
255     Store->setAtomic(AtomicOrdering::Unordered);
256   }
257   Value *NewIndex =
258       LoopBuilder.CreateAdd(LoopIndex, ConstantInt::get(CopyLenType, 1U));
259   LoopIndex->addIncoming(NewIndex, LoopBB);
260 
261   bool requiresResidual =
262       !LoopOpIsInt8 && !(AtomicElementSize && LoopOpSize == AtomicElementSize);
263   if (requiresResidual) {
264     Type *ResLoopOpType = AtomicElementSize
265                               ? Type::getIntNTy(Ctx, *AtomicElementSize * 8)
266                               : Int8Type;
267     unsigned ResLoopOpSize = DL.getTypeStoreSize(ResLoopOpType);
268     assert((ResLoopOpSize == AtomicElementSize ? *AtomicElementSize : 1) &&
269            "Store size is expected to match type size");
270 
271     // Add in the
272     Value *RuntimeResidual = PLBuilder.CreateURem(CopyLen, CILoopOpSize);
273     Value *RuntimeBytesCopied = PLBuilder.CreateSub(CopyLen, RuntimeResidual);
274 
275     // Loop body for the residual copy.
276     BasicBlock *ResLoopBB = BasicBlock::Create(Ctx, "loop-memcpy-residual",
277                                                PreLoopBB->getParent(),
278                                                PostLoopBB);
279     // Residual loop header.
280     BasicBlock *ResHeaderBB = BasicBlock::Create(
281         Ctx, "loop-memcpy-residual-header", PreLoopBB->getParent(), nullptr);
282 
283     // Need to update the pre-loop basic block to branch to the correct place.
284     // branch to the main loop if the count is non-zero, branch to the residual
285     // loop if the copy size is smaller then 1 iteration of the main loop but
286     // non-zero and finally branch to after the residual loop if the memcpy
287     //  size is zero.
288     ConstantInt *Zero = ConstantInt::get(ILengthType, 0U);
289     PLBuilder.CreateCondBr(PLBuilder.CreateICmpNE(RuntimeLoopCount, Zero),
290                            LoopBB, ResHeaderBB);
291     PreLoopBB->getTerminator()->eraseFromParent();
292 
293     LoopBuilder.CreateCondBr(
294         LoopBuilder.CreateICmpULT(NewIndex, RuntimeLoopCount), LoopBB,
295         ResHeaderBB);
296 
297     // Determine if we need to branch to the residual loop or bypass it.
298     IRBuilder<> RHBuilder(ResHeaderBB);
299     RHBuilder.CreateCondBr(RHBuilder.CreateICmpNE(RuntimeResidual, Zero),
300                            ResLoopBB, PostLoopBB);
301 
302     // Copy the residual with single byte load/store loop.
303     IRBuilder<> ResBuilder(ResLoopBB);
304     PHINode *ResidualIndex =
305         ResBuilder.CreatePHI(CopyLenType, 2, "residual-loop-index");
306     ResidualIndex->addIncoming(Zero, ResHeaderBB);
307 
308     Value *SrcAsResLoopOpType = ResBuilder.CreateBitCast(
309         SrcAddr, PointerType::get(ResLoopOpType, SrcAS));
310     Value *DstAsResLoopOpType = ResBuilder.CreateBitCast(
311         DstAddr, PointerType::get(ResLoopOpType, DstAS));
312     Value *FullOffset = ResBuilder.CreateAdd(RuntimeBytesCopied, ResidualIndex);
313     Value *SrcGEP = ResBuilder.CreateInBoundsGEP(
314         ResLoopOpType, SrcAsResLoopOpType, FullOffset);
315     LoadInst *Load = ResBuilder.CreateAlignedLoad(ResLoopOpType, SrcGEP,
316                                                   PartSrcAlign, SrcIsVolatile);
317     if (!CanOverlap) {
318       // Set alias scope for loads.
319       Load->setMetadata(LLVMContext::MD_alias_scope,
320                         MDNode::get(Ctx, NewScope));
321     }
322     Value *DstGEP = ResBuilder.CreateInBoundsGEP(
323         ResLoopOpType, DstAsResLoopOpType, FullOffset);
324     StoreInst *Store = ResBuilder.CreateAlignedStore(Load, DstGEP, PartDstAlign,
325                                                      DstIsVolatile);
326     if (!CanOverlap) {
327       // Indicate that stores don't overlap loads.
328       Store->setMetadata(LLVMContext::MD_noalias, MDNode::get(Ctx, NewScope));
329     }
330     if (AtomicElementSize) {
331       Load->setAtomic(AtomicOrdering::Unordered);
332       Store->setAtomic(AtomicOrdering::Unordered);
333     }
334     Value *ResNewIndex = ResBuilder.CreateAdd(
335         ResidualIndex, ConstantInt::get(CopyLenType, ResLoopOpSize));
336     ResidualIndex->addIncoming(ResNewIndex, ResLoopBB);
337 
338     // Create the loop branch condition.
339     ResBuilder.CreateCondBr(
340         ResBuilder.CreateICmpULT(ResNewIndex, RuntimeResidual), ResLoopBB,
341         PostLoopBB);
342   } else {
343     // In this case the loop operand type was a byte, and there is no need for a
344     // residual loop to copy the remaining memory after the main loop.
345     // We do however need to patch up the control flow by creating the
346     // terminators for the preloop block and the memcpy loop.
347     ConstantInt *Zero = ConstantInt::get(ILengthType, 0U);
348     PLBuilder.CreateCondBr(PLBuilder.CreateICmpNE(RuntimeLoopCount, Zero),
349                            LoopBB, PostLoopBB);
350     PreLoopBB->getTerminator()->eraseFromParent();
351     LoopBuilder.CreateCondBr(
352         LoopBuilder.CreateICmpULT(NewIndex, RuntimeLoopCount), LoopBB,
353         PostLoopBB);
354   }
355 }
356 
357 // Lower memmove to IR. memmove is required to correctly copy overlapping memory
358 // regions; therefore, it has to check the relative positions of the source and
359 // destination pointers and choose the copy direction accordingly.
360 //
361 // The code below is an IR rendition of this C function:
362 //
363 // void* memmove(void* dst, const void* src, size_t n) {
364 //   unsigned char* d = dst;
365 //   const unsigned char* s = src;
366 //   if (s < d) {
367 //     // copy backwards
368 //     while (n--) {
369 //       d[n] = s[n];
370 //     }
371 //   } else {
372 //     // copy forward
373 //     for (size_t i = 0; i < n; ++i) {
374 //       d[i] = s[i];
375 //     }
376 //   }
377 //   return dst;
378 // }
379 static void createMemMoveLoop(Instruction *InsertBefore, Value *SrcAddr,
380                               Value *DstAddr, Value *CopyLen, Align SrcAlign,
381                               Align DstAlign, bool SrcIsVolatile,
382                               bool DstIsVolatile) {
383   Type *TypeOfCopyLen = CopyLen->getType();
384   BasicBlock *OrigBB = InsertBefore->getParent();
385   Function *F = OrigBB->getParent();
386   const DataLayout &DL = F->getParent()->getDataLayout();
387 
388   // TODO: Use different element type if possible?
389   IRBuilder<> CastBuilder(InsertBefore);
390   Type *EltTy = CastBuilder.getInt8Ty();
391   Type *PtrTy =
392       CastBuilder.getInt8PtrTy(SrcAddr->getType()->getPointerAddressSpace());
393   SrcAddr = CastBuilder.CreateBitCast(SrcAddr, PtrTy);
394   DstAddr = CastBuilder.CreateBitCast(DstAddr, PtrTy);
395 
396   // Create the a comparison of src and dst, based on which we jump to either
397   // the forward-copy part of the function (if src >= dst) or the backwards-copy
398   // part (if src < dst).
399   // SplitBlockAndInsertIfThenElse conveniently creates the basic if-then-else
400   // structure. Its block terminators (unconditional branches) are replaced by
401   // the appropriate conditional branches when the loop is built.
402   ICmpInst *PtrCompare = new ICmpInst(InsertBefore, ICmpInst::ICMP_ULT,
403                                       SrcAddr, DstAddr, "compare_src_dst");
404   Instruction *ThenTerm, *ElseTerm;
405   SplitBlockAndInsertIfThenElse(PtrCompare, InsertBefore, &ThenTerm,
406                                 &ElseTerm);
407 
408   // Each part of the function consists of two blocks:
409   //   copy_backwards:        used to skip the loop when n == 0
410   //   copy_backwards_loop:   the actual backwards loop BB
411   //   copy_forward:          used to skip the loop when n == 0
412   //   copy_forward_loop:     the actual forward loop BB
413   BasicBlock *CopyBackwardsBB = ThenTerm->getParent();
414   CopyBackwardsBB->setName("copy_backwards");
415   BasicBlock *CopyForwardBB = ElseTerm->getParent();
416   CopyForwardBB->setName("copy_forward");
417   BasicBlock *ExitBB = InsertBefore->getParent();
418   ExitBB->setName("memmove_done");
419 
420   unsigned PartSize = DL.getTypeStoreSize(EltTy);
421   Align PartSrcAlign(commonAlignment(SrcAlign, PartSize));
422   Align PartDstAlign(commonAlignment(DstAlign, PartSize));
423 
424   // Initial comparison of n == 0 that lets us skip the loops altogether. Shared
425   // between both backwards and forward copy clauses.
426   ICmpInst *CompareN =
427       new ICmpInst(OrigBB->getTerminator(), ICmpInst::ICMP_EQ, CopyLen,
428                    ConstantInt::get(TypeOfCopyLen, 0), "compare_n_to_0");
429 
430   // Copying backwards.
431   BasicBlock *LoopBB =
432     BasicBlock::Create(F->getContext(), "copy_backwards_loop", F, CopyForwardBB);
433   IRBuilder<> LoopBuilder(LoopBB);
434   PHINode *LoopPhi = LoopBuilder.CreatePHI(TypeOfCopyLen, 0);
435   Value *IndexPtr = LoopBuilder.CreateSub(
436       LoopPhi, ConstantInt::get(TypeOfCopyLen, 1), "index_ptr");
437   Value *Element = LoopBuilder.CreateAlignedLoad(
438       EltTy, LoopBuilder.CreateInBoundsGEP(EltTy, SrcAddr, IndexPtr),
439       PartSrcAlign, "element");
440   LoopBuilder.CreateAlignedStore(
441       Element, LoopBuilder.CreateInBoundsGEP(EltTy, DstAddr, IndexPtr),
442       PartDstAlign);
443   LoopBuilder.CreateCondBr(
444       LoopBuilder.CreateICmpEQ(IndexPtr, ConstantInt::get(TypeOfCopyLen, 0)),
445       ExitBB, LoopBB);
446   LoopPhi->addIncoming(IndexPtr, LoopBB);
447   LoopPhi->addIncoming(CopyLen, CopyBackwardsBB);
448   BranchInst::Create(ExitBB, LoopBB, CompareN, ThenTerm);
449   ThenTerm->eraseFromParent();
450 
451   // Copying forward.
452   BasicBlock *FwdLoopBB =
453     BasicBlock::Create(F->getContext(), "copy_forward_loop", F, ExitBB);
454   IRBuilder<> FwdLoopBuilder(FwdLoopBB);
455   PHINode *FwdCopyPhi = FwdLoopBuilder.CreatePHI(TypeOfCopyLen, 0, "index_ptr");
456   Value *SrcGEP = FwdLoopBuilder.CreateInBoundsGEP(EltTy, SrcAddr, FwdCopyPhi);
457   Value *FwdElement =
458       FwdLoopBuilder.CreateAlignedLoad(EltTy, SrcGEP, PartSrcAlign, "element");
459   Value *DstGEP = FwdLoopBuilder.CreateInBoundsGEP(EltTy, DstAddr, FwdCopyPhi);
460   FwdLoopBuilder.CreateAlignedStore(FwdElement, DstGEP, PartDstAlign);
461   Value *FwdIndexPtr = FwdLoopBuilder.CreateAdd(
462       FwdCopyPhi, ConstantInt::get(TypeOfCopyLen, 1), "index_increment");
463   FwdLoopBuilder.CreateCondBr(FwdLoopBuilder.CreateICmpEQ(FwdIndexPtr, CopyLen),
464                               ExitBB, FwdLoopBB);
465   FwdCopyPhi->addIncoming(FwdIndexPtr, FwdLoopBB);
466   FwdCopyPhi->addIncoming(ConstantInt::get(TypeOfCopyLen, 0), CopyForwardBB);
467 
468   BranchInst::Create(ExitBB, FwdLoopBB, CompareN, ElseTerm);
469   ElseTerm->eraseFromParent();
470 }
471 
472 static void createMemSetLoop(Instruction *InsertBefore, Value *DstAddr,
473                              Value *CopyLen, Value *SetValue, Align DstAlign,
474                              bool IsVolatile) {
475   Type *TypeOfCopyLen = CopyLen->getType();
476   BasicBlock *OrigBB = InsertBefore->getParent();
477   Function *F = OrigBB->getParent();
478   const DataLayout &DL = F->getParent()->getDataLayout();
479   BasicBlock *NewBB =
480       OrigBB->splitBasicBlock(InsertBefore, "split");
481   BasicBlock *LoopBB
482     = BasicBlock::Create(F->getContext(), "loadstoreloop", F, NewBB);
483 
484   IRBuilder<> Builder(OrigBB->getTerminator());
485 
486   // Cast pointer to the type of value getting stored
487   unsigned dstAS = cast<PointerType>(DstAddr->getType())->getAddressSpace();
488   DstAddr = Builder.CreateBitCast(DstAddr,
489                                   PointerType::get(SetValue->getType(), dstAS));
490 
491   Builder.CreateCondBr(
492       Builder.CreateICmpEQ(ConstantInt::get(TypeOfCopyLen, 0), CopyLen), NewBB,
493       LoopBB);
494   OrigBB->getTerminator()->eraseFromParent();
495 
496   unsigned PartSize = DL.getTypeStoreSize(SetValue->getType());
497   Align PartAlign(commonAlignment(DstAlign, PartSize));
498 
499   IRBuilder<> LoopBuilder(LoopBB);
500   PHINode *LoopIndex = LoopBuilder.CreatePHI(TypeOfCopyLen, 0);
501   LoopIndex->addIncoming(ConstantInt::get(TypeOfCopyLen, 0), OrigBB);
502 
503   LoopBuilder.CreateAlignedStore(
504       SetValue,
505       LoopBuilder.CreateInBoundsGEP(SetValue->getType(), DstAddr, LoopIndex),
506       PartAlign, IsVolatile);
507 
508   Value *NewIndex =
509       LoopBuilder.CreateAdd(LoopIndex, ConstantInt::get(TypeOfCopyLen, 1));
510   LoopIndex->addIncoming(NewIndex, LoopBB);
511 
512   LoopBuilder.CreateCondBr(LoopBuilder.CreateICmpULT(NewIndex, CopyLen), LoopBB,
513                            NewBB);
514 }
515 
516 template <typename T>
517 static bool canOverlap(MemTransferBase<T> *Memcpy, ScalarEvolution *SE) {
518   if (SE) {
519     auto *SrcSCEV = SE->getSCEV(Memcpy->getRawSource());
520     auto *DestSCEV = SE->getSCEV(Memcpy->getRawDest());
521     if (SE->isKnownPredicateAt(CmpInst::ICMP_NE, SrcSCEV, DestSCEV, Memcpy))
522       return false;
523   }
524   return true;
525 }
526 
527 void llvm::expandMemCpyAsLoop(MemCpyInst *Memcpy,
528                               const TargetTransformInfo &TTI,
529                               ScalarEvolution *SE) {
530   bool CanOverlap = canOverlap(Memcpy, SE);
531   if (ConstantInt *CI = dyn_cast<ConstantInt>(Memcpy->getLength())) {
532     createMemCpyLoopKnownSize(
533         /* InsertBefore */ Memcpy,
534         /* SrcAddr */ Memcpy->getRawSource(),
535         /* DstAddr */ Memcpy->getRawDest(),
536         /* CopyLen */ CI,
537         /* SrcAlign */ Memcpy->getSourceAlign().valueOrOne(),
538         /* DestAlign */ Memcpy->getDestAlign().valueOrOne(),
539         /* SrcIsVolatile */ Memcpy->isVolatile(),
540         /* DstIsVolatile */ Memcpy->isVolatile(),
541         /* CanOverlap */ CanOverlap,
542         /* TargetTransformInfo */ TTI);
543   } else {
544     createMemCpyLoopUnknownSize(
545         /* InsertBefore */ Memcpy,
546         /* SrcAddr */ Memcpy->getRawSource(),
547         /* DstAddr */ Memcpy->getRawDest(),
548         /* CopyLen */ Memcpy->getLength(),
549         /* SrcAlign */ Memcpy->getSourceAlign().valueOrOne(),
550         /* DestAlign */ Memcpy->getDestAlign().valueOrOne(),
551         /* SrcIsVolatile */ Memcpy->isVolatile(),
552         /* DstIsVolatile */ Memcpy->isVolatile(),
553         /* CanOverlap */ CanOverlap,
554         /* TargetTransformInfo */ TTI);
555   }
556 }
557 
558 void llvm::expandMemMoveAsLoop(MemMoveInst *Memmove) {
559   createMemMoveLoop(/* InsertBefore */ Memmove,
560                     /* SrcAddr */ Memmove->getRawSource(),
561                     /* DstAddr */ Memmove->getRawDest(),
562                     /* CopyLen */ Memmove->getLength(),
563                     /* SrcAlign */ Memmove->getSourceAlign().valueOrOne(),
564                     /* DestAlign */ Memmove->getDestAlign().valueOrOne(),
565                     /* SrcIsVolatile */ Memmove->isVolatile(),
566                     /* DstIsVolatile */ Memmove->isVolatile());
567 }
568 
569 void llvm::expandMemSetAsLoop(MemSetInst *Memset) {
570   createMemSetLoop(/* InsertBefore */ Memset,
571                    /* DstAddr */ Memset->getRawDest(),
572                    /* CopyLen */ Memset->getLength(),
573                    /* SetValue */ Memset->getValue(),
574                    /* Alignment */ Memset->getDestAlign().valueOrOne(),
575                    Memset->isVolatile());
576 }
577 
578 void llvm::expandAtomicMemCpyAsLoop(AtomicMemCpyInst *AtomicMemcpy,
579                                     const TargetTransformInfo &TTI,
580                                     ScalarEvolution *SE) {
581   if (ConstantInt *CI = dyn_cast<ConstantInt>(AtomicMemcpy->getLength())) {
582     createMemCpyLoopKnownSize(
583         /* InsertBefore */ AtomicMemcpy,
584         /* SrcAddr */ AtomicMemcpy->getRawSource(),
585         /* DstAddr */ AtomicMemcpy->getRawDest(),
586         /* CopyLen */ CI,
587         /* SrcAlign */ AtomicMemcpy->getSourceAlign().valueOrOne(),
588         /* DestAlign */ AtomicMemcpy->getDestAlign().valueOrOne(),
589         /* SrcIsVolatile */ AtomicMemcpy->isVolatile(),
590         /* DstIsVolatile */ AtomicMemcpy->isVolatile(),
591         /* CanOverlap */ false, // SrcAddr & DstAddr may not overlap by spec.
592         /* TargetTransformInfo */ TTI,
593         /* AtomicCpySize */ AtomicMemcpy->getElementSizeInBytes());
594   } else {
595     createMemCpyLoopUnknownSize(
596         /* InsertBefore */ AtomicMemcpy,
597         /* SrcAddr */ AtomicMemcpy->getRawSource(),
598         /* DstAddr */ AtomicMemcpy->getRawDest(),
599         /* CopyLen */ AtomicMemcpy->getLength(),
600         /* SrcAlign */ AtomicMemcpy->getSourceAlign().valueOrOne(),
601         /* DestAlign */ AtomicMemcpy->getDestAlign().valueOrOne(),
602         /* SrcIsVolatile */ AtomicMemcpy->isVolatile(),
603         /* DstIsVolatile */ AtomicMemcpy->isVolatile(),
604         /* CanOverlap */ false, // SrcAddr & DstAddr may not overlap by spec.
605         /* TargetTransformInfo */ TTI,
606         /* AtomicCpySize */ AtomicMemcpy->getElementSizeInBytes());
607   }
608 }
609