1 //===- MemoryLocation.cpp - Memory location descriptions -------------------==//
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/Analysis/MemoryLocation.h"
10 #include "llvm/Analysis/TargetLibraryInfo.h"
11 #include "llvm/IR/BasicBlock.h"
12 #include "llvm/IR/DataLayout.h"
13 #include "llvm/IR/Instructions.h"
14 #include "llvm/IR/IntrinsicInst.h"
15 #include "llvm/IR/IntrinsicsARM.h"
16 #include "llvm/IR/LLVMContext.h"
17 #include "llvm/IR/Module.h"
18 #include "llvm/IR/Type.h"
19 using namespace llvm;
20 
21 void LocationSize::print(raw_ostream &OS) const {
22   OS << "LocationSize::";
23   if (*this == beforeOrAfterPointer())
24     OS << "beforeOrAfterPointer";
25   else if (*this == afterPointer())
26     OS << "afterPointer";
27   else if (*this == mapEmpty())
28     OS << "mapEmpty";
29   else if (*this == mapTombstone())
30     OS << "mapTombstone";
31   else if (isPrecise())
32     OS << "precise(" << getValue() << ')';
33   else
34     OS << "upperBound(" << getValue() << ')';
35 }
36 
37 MemoryLocation MemoryLocation::get(const LoadInst *LI) {
38   const auto &DL = LI->getModule()->getDataLayout();
39 
40   return MemoryLocation(
41       LI->getPointerOperand(),
42       LocationSize::precise(DL.getTypeStoreSize(LI->getType())),
43       LI->getAAMetadata());
44 }
45 
46 MemoryLocation MemoryLocation::get(const StoreInst *SI) {
47   const auto &DL = SI->getModule()->getDataLayout();
48 
49   return MemoryLocation(SI->getPointerOperand(),
50                         LocationSize::precise(DL.getTypeStoreSize(
51                             SI->getValueOperand()->getType())),
52                         SI->getAAMetadata());
53 }
54 
55 MemoryLocation MemoryLocation::get(const VAArgInst *VI) {
56   return MemoryLocation(VI->getPointerOperand(),
57                         LocationSize::afterPointer(), VI->getAAMetadata());
58 }
59 
60 MemoryLocation MemoryLocation::get(const AtomicCmpXchgInst *CXI) {
61   const auto &DL = CXI->getModule()->getDataLayout();
62 
63   return MemoryLocation(CXI->getPointerOperand(),
64                         LocationSize::precise(DL.getTypeStoreSize(
65                             CXI->getCompareOperand()->getType())),
66                         CXI->getAAMetadata());
67 }
68 
69 MemoryLocation MemoryLocation::get(const AtomicRMWInst *RMWI) {
70   const auto &DL = RMWI->getModule()->getDataLayout();
71 
72   return MemoryLocation(RMWI->getPointerOperand(),
73                         LocationSize::precise(DL.getTypeStoreSize(
74                             RMWI->getValOperand()->getType())),
75                         RMWI->getAAMetadata());
76 }
77 
78 Optional<MemoryLocation> MemoryLocation::getOrNone(const Instruction *Inst) {
79   switch (Inst->getOpcode()) {
80   case Instruction::Load:
81     return get(cast<LoadInst>(Inst));
82   case Instruction::Store:
83     return get(cast<StoreInst>(Inst));
84   case Instruction::VAArg:
85     return get(cast<VAArgInst>(Inst));
86   case Instruction::AtomicCmpXchg:
87     return get(cast<AtomicCmpXchgInst>(Inst));
88   case Instruction::AtomicRMW:
89     return get(cast<AtomicRMWInst>(Inst));
90   default:
91     return None;
92   }
93 }
94 
95 MemoryLocation MemoryLocation::getForSource(const MemTransferInst *MTI) {
96   return getForSource(cast<AnyMemTransferInst>(MTI));
97 }
98 
99 MemoryLocation MemoryLocation::getForSource(const AtomicMemTransferInst *MTI) {
100   return getForSource(cast<AnyMemTransferInst>(MTI));
101 }
102 
103 MemoryLocation MemoryLocation::getForSource(const AnyMemTransferInst *MTI) {
104   assert(MTI->getRawSource() == MTI->getArgOperand(1));
105   return getForArgument(MTI, 1, nullptr);
106 }
107 
108 MemoryLocation MemoryLocation::getForDest(const MemIntrinsic *MI) {
109   return getForDest(cast<AnyMemIntrinsic>(MI));
110 }
111 
112 MemoryLocation MemoryLocation::getForDest(const AtomicMemIntrinsic *MI) {
113   return getForDest(cast<AnyMemIntrinsic>(MI));
114 }
115 
116 MemoryLocation MemoryLocation::getForDest(const AnyMemIntrinsic *MI) {
117   assert(MI->getRawDest() == MI->getArgOperand(0));
118   return getForArgument(MI, 0, nullptr);
119 }
120 
121 Optional<MemoryLocation>
122 MemoryLocation::getForDest(const CallBase *CB, const TargetLibraryInfo &TLI) {
123   if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(CB)) {
124     if (auto *MemInst = dyn_cast<AnyMemIntrinsic>(CB))
125       return getForDest(MemInst);
126 
127     switch (II->getIntrinsicID()) {
128     default:
129       return None;
130     case Intrinsic::init_trampoline:
131       return MemoryLocation::getForArgument(CB, 0, TLI);
132     case Intrinsic::masked_store:
133       return MemoryLocation::getForArgument(CB, 1, TLI);
134     }
135   }
136 
137   if (!CB->onlyAccessesArgMemory())
138     return None;
139 
140   if (CB->hasOperandBundles())
141     // TODO: remove implementation restriction
142     return None;
143 
144   Value *UsedV = nullptr;
145   Optional<unsigned> UsedIdx;
146   for (unsigned i = 0; i < CB->arg_size(); i++) {
147     if (!CB->getArgOperand(i)->getType()->isPointerTy())
148       continue;
149      if (CB->onlyReadsMemory(i))
150        continue;
151     if (!UsedV) {
152       // First potentially writing parameter
153       UsedV = CB->getArgOperand(i);
154       UsedIdx = i;
155       continue;
156     }
157     UsedIdx = None;
158     if (UsedV != CB->getArgOperand(i))
159       // Can't describe writing to two distinct locations.
160       // TODO: This results in an inprecision when two values derived from the
161       // same object are passed as arguments to the same function.
162       return None;
163   }
164   if (!UsedV)
165     // We don't currently have a way to represent a "does not write" result
166     // and thus have to be conservative and return unknown.
167     return None;
168 
169   if (UsedIdx)
170     return getForArgument(CB, *UsedIdx, &TLI);
171   return MemoryLocation::getBeforeOrAfter(UsedV, CB->getAAMetadata());
172 }
173 
174 MemoryLocation MemoryLocation::getForArgument(const CallBase *Call,
175                                               unsigned ArgIdx,
176                                               const TargetLibraryInfo *TLI) {
177   AAMDNodes AATags = Call->getAAMetadata();
178   const Value *Arg = Call->getArgOperand(ArgIdx);
179 
180   // We may be able to produce an exact size for known intrinsics.
181   if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Call)) {
182     const DataLayout &DL = II->getModule()->getDataLayout();
183 
184     switch (II->getIntrinsicID()) {
185     default:
186       break;
187     case Intrinsic::memset:
188     case Intrinsic::memcpy:
189     case Intrinsic::memcpy_inline:
190     case Intrinsic::memmove:
191     case Intrinsic::memcpy_element_unordered_atomic:
192     case Intrinsic::memmove_element_unordered_atomic:
193     case Intrinsic::memset_element_unordered_atomic:
194       assert((ArgIdx == 0 || ArgIdx == 1) &&
195              "Invalid argument index for memory intrinsic");
196       if (ConstantInt *LenCI = dyn_cast<ConstantInt>(II->getArgOperand(2)))
197         return MemoryLocation(Arg, LocationSize::precise(LenCI->getZExtValue()),
198                               AATags);
199       return MemoryLocation::getAfter(Arg, AATags);
200 
201     case Intrinsic::lifetime_start:
202     case Intrinsic::lifetime_end:
203     case Intrinsic::invariant_start:
204       assert(ArgIdx == 1 && "Invalid argument index");
205       return MemoryLocation(
206           Arg,
207           LocationSize::precise(
208               cast<ConstantInt>(II->getArgOperand(0))->getZExtValue()),
209           AATags);
210 
211     case Intrinsic::masked_load:
212       assert(ArgIdx == 0 && "Invalid argument index");
213       return MemoryLocation(
214           Arg,
215           LocationSize::upperBound(DL.getTypeStoreSize(II->getType())),
216           AATags);
217 
218     case Intrinsic::masked_store:
219       assert(ArgIdx == 1 && "Invalid argument index");
220       return MemoryLocation(
221           Arg,
222           LocationSize::upperBound(
223               DL.getTypeStoreSize(II->getArgOperand(0)->getType())),
224           AATags);
225 
226     case Intrinsic::invariant_end:
227       // The first argument to an invariant.end is a "descriptor" type (e.g. a
228       // pointer to a empty struct) which is never actually dereferenced.
229       if (ArgIdx == 0)
230         return MemoryLocation(Arg, LocationSize::precise(0), AATags);
231       assert(ArgIdx == 2 && "Invalid argument index");
232       return MemoryLocation(
233           Arg,
234           LocationSize::precise(
235               cast<ConstantInt>(II->getArgOperand(1))->getZExtValue()),
236           AATags);
237 
238     case Intrinsic::arm_neon_vld1:
239       assert(ArgIdx == 0 && "Invalid argument index");
240       // LLVM's vld1 and vst1 intrinsics currently only support a single
241       // vector register.
242       return MemoryLocation(
243           Arg, LocationSize::precise(DL.getTypeStoreSize(II->getType())),
244           AATags);
245 
246     case Intrinsic::arm_neon_vst1:
247       assert(ArgIdx == 0 && "Invalid argument index");
248       return MemoryLocation(Arg,
249                             LocationSize::precise(DL.getTypeStoreSize(
250                                 II->getArgOperand(1)->getType())),
251                             AATags);
252     }
253 
254     assert(
255         !isa<AnyMemTransferInst>(II) &&
256         "all memory transfer intrinsics should be handled by the switch above");
257   }
258 
259   // We can bound the aliasing properties of memset_pattern16 just as we can
260   // for memcpy/memset.  This is particularly important because the
261   // LoopIdiomRecognizer likes to turn loops into calls to memset_pattern16
262   // whenever possible.
263   LibFunc F;
264   if (TLI && TLI->getLibFunc(*Call, F) && TLI->has(F)) {
265     switch (F) {
266     case LibFunc_strcpy:
267     case LibFunc_strcat:
268     case LibFunc_strncat:
269       assert((ArgIdx == 0 || ArgIdx == 1) && "Invalid argument index for str function");
270       return MemoryLocation::getAfter(Arg, AATags);
271 
272     case LibFunc_memset_chk: {
273       assert(ArgIdx == 0 && "Invalid argument index for memset_chk");
274       LocationSize Size = LocationSize::afterPointer();
275       if (const auto *Len = dyn_cast<ConstantInt>(Call->getArgOperand(2))) {
276         // memset_chk writes at most Len bytes. It may write less, if Len
277         // exceeds the specified max size and aborts.
278         Size = LocationSize::upperBound(Len->getZExtValue());
279       }
280       return MemoryLocation(Arg, Size, AATags);
281     }
282     case LibFunc_strncpy: {
283       assert((ArgIdx == 0 || ArgIdx == 1) &&
284              "Invalid argument index for strncpy");
285       LocationSize Size = LocationSize::afterPointer();
286       if (const auto *Len = dyn_cast<ConstantInt>(Call->getArgOperand(2))) {
287         // strncpy is guaranteed to write Len bytes, but only reads up to Len
288         // bytes.
289         Size = ArgIdx == 0 ? LocationSize::precise(Len->getZExtValue())
290                            : LocationSize::upperBound(Len->getZExtValue());
291       }
292       return MemoryLocation(Arg, Size, AATags);
293     }
294     case LibFunc_memset_pattern16:
295     case LibFunc_memset_pattern4:
296     case LibFunc_memset_pattern8:
297       assert((ArgIdx == 0 || ArgIdx == 1) &&
298              "Invalid argument index for memset_pattern16");
299       if (ArgIdx == 1) {
300         unsigned Size = 16;
301         if (F == LibFunc_memset_pattern4)
302           Size = 4;
303         else if (F == LibFunc_memset_pattern8)
304           Size = 8;
305         return MemoryLocation(Arg, LocationSize::precise(Size), AATags);
306       }
307       if (const ConstantInt *LenCI =
308               dyn_cast<ConstantInt>(Call->getArgOperand(2)))
309         return MemoryLocation(Arg, LocationSize::precise(LenCI->getZExtValue()),
310                               AATags);
311       return MemoryLocation::getAfter(Arg, AATags);
312     case LibFunc_bcmp:
313     case LibFunc_memcmp:
314       assert((ArgIdx == 0 || ArgIdx == 1) &&
315              "Invalid argument index for memcmp/bcmp");
316       if (const ConstantInt *LenCI =
317               dyn_cast<ConstantInt>(Call->getArgOperand(2)))
318         return MemoryLocation(Arg, LocationSize::precise(LenCI->getZExtValue()),
319                               AATags);
320       return MemoryLocation::getAfter(Arg, AATags);
321     case LibFunc_memchr:
322       assert((ArgIdx == 0) && "Invalid argument index for memchr");
323       if (const ConstantInt *LenCI =
324               dyn_cast<ConstantInt>(Call->getArgOperand(2)))
325         return MemoryLocation(Arg, LocationSize::precise(LenCI->getZExtValue()),
326                               AATags);
327       return MemoryLocation::getAfter(Arg, AATags);
328     case LibFunc_memccpy:
329       assert((ArgIdx == 0 || ArgIdx == 1) &&
330              "Invalid argument index for memccpy");
331       // We only know an upper bound on the number of bytes read/written.
332       if (const ConstantInt *LenCI =
333               dyn_cast<ConstantInt>(Call->getArgOperand(3)))
334         return MemoryLocation(
335             Arg, LocationSize::upperBound(LenCI->getZExtValue()), AATags);
336       return MemoryLocation::getAfter(Arg, AATags);
337     default:
338       break;
339     };
340   }
341 
342   return MemoryLocation::getBeforeOrAfter(Call->getArgOperand(ArgIdx), AATags);
343 }
344