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