1 Inlining 2 ======== 3 4 There are several options that control which calls the analyzer will consider for 5 inlining. The major one is -analyzer-config ipa: 6 7 -analyzer-config ipa=none - All inlining is disabled. This is the only mode 8 available in LLVM 3.1 and earlier and in Xcode 4.3 and earlier. 9 10 -analyzer-config ipa=basic-inlining - Turns on inlining for C functions, C++ 11 static member functions, and blocks -- essentially, the calls that behave 12 like simple C function calls. This is essentially the mode used in 13 Xcode 4.4. 14 15 -analyzer-config ipa=inlining - Turns on inlining when we can confidently find 16 the function/method body corresponding to the call. (C functions, static 17 functions, devirtualized C++ methods, Objective-C class methods, Objective-C 18 instance methods when ExprEngine is confident about the dynamic type of the 19 instance). 20 21 -analyzer-config ipa=dynamic - Inline instance methods for which the type is 22 determined at runtime and we are not 100% sure that our type info is 23 correct. For virtual calls, inline the most plausible definition. 24 25 -analyzer-config ipa=dynamic-bifurcate - Same as -analyzer-config ipa=dynamic, 26 but the path is split. We inline on one branch and do not inline on the 27 other. This mode does not drop the coverage in cases when the parent class 28 has code that is only exercised when some of its methods are overridden. 29 30 Currently, -analyzer-config ipa=dynamic-bifurcate is the default mode. 31 32 While -analyzer-config ipa determines in general how aggressively the analyzer 33 will try to inline functions, several additional options control which types of 34 functions can inlined, in an all-or-nothing way. These options use the 35 analyzer's configuration table, so they are all specified as follows: 36 37 -analyzer-config OPTION=VALUE 38 39 ### c++-inlining ### 40 41 This option controls which C++ member functions may be inlined. 42 43 -analyzer-config c++-inlining=[none | methods | constructors | destructors] 44 45 Each of these modes implies that all the previous member function kinds will be 46 inlined as well; it doesn't make sense to inline destructors without inlining 47 constructors, for example. 48 49 The default c++-inlining mode is 'destructors', meaning that all member 50 functions with visible definitions will be considered for inlining. In some 51 cases the analyzer may still choose not to inline the function. 52 53 Note that under 'constructors', constructors for types with non-trivial 54 destructors will not be inlined. Additionally, no C++ member functions will be 55 inlined under -analyzer-config ipa=none or -analyzer-config ipa=basic-inlining, 56 regardless of the setting of the c++-inlining mode. 57 58 ### c++-template-inlining ### 59 60 This option controls whether C++ templated functions may be inlined. 61 62 -analyzer-config c++-template-inlining=[true | false] 63 64 Currently, template functions are considered for inlining by default. 65 66 The motivation behind this option is that very generic code can be a source 67 of false positives, either by considering paths that the caller considers 68 impossible (by some unstated precondition), or by inlining some but not all 69 of a deep implementation of a function. 70 71 ### c++-stdlib-inlining ### 72 73 This option controls whether functions from the C++ standard library, including 74 methods of the container classes in the Standard Template Library, should be 75 considered for inlining. 76 77 -analyzer-config c++-stdlib-inlining=[true | false] 78 79 Currently, C++ standard library functions are considered for inlining by 80 default. 81 82 The standard library functions and the STL in particular are used ubiquitously 83 enough that our tolerance for false positives is even lower here. A false 84 positive due to poor modeling of the STL leads to a poor user experience, since 85 most users would not be comfortable adding assertions to system headers in order 86 to silence analyzer warnings. 87 88 ### c++-container-inlining ### 89 90 This option controls whether constructors and destructors of "container" types 91 should be considered for inlining. 92 93 -analyzer-config c++-container-inlining=[true | false] 94 95 Currently, these constructors and destructors are NOT considered for inlining 96 by default. 97 98 The current implementation of this setting checks whether a type has a member 99 named 'iterator' or a member named 'begin'; these names are idiomatic in C++, 100 with the latter specified in the C++11 standard. The analyzer currently does a 101 fairly poor job of modeling certain data structure invariants of container-like 102 objects. For example, these three expressions should be equivalent: 103 104 std::distance(c.begin(), c.end()) == 0 105 c.begin() == c.end() 106 c.empty()) 107 108 Many of these issues are avoided if containers always have unknown, symbolic 109 state, which is what happens when their constructors are treated as opaque. 110 In the future, we may decide specific containers are "safe" to model through 111 inlining, or choose to model them directly using checkers instead. 112 113 114 Basics of Implementation 115 ----------------------- 116 117 The low-level mechanism of inlining a function is handled in 118 ExprEngine::inlineCall and ExprEngine::processCallExit. 119 120 If the conditions are right for inlining, a CallEnter node is created and added 121 to the analysis work list. The CallEnter node marks the change to a new 122 LocationContext representing the called function, and its state includes the 123 contents of the new stack frame. When the CallEnter node is actually processed, 124 its single successor will be a edge to the first CFG block in the function. 125 126 Exiting an inlined function is a bit more work, fortunately broken up into 127 reasonable steps: 128 129 1. The CoreEngine realizes we're at the end of an inlined call and generates a 130 CallExitBegin node. 131 132 2. ExprEngine takes over (in processCallExit) and finds the return value of the 133 function, if it has one. This is bound to the expression that triggered the 134 call. (In the case of calls without origin expressions, such as destructors, 135 this step is skipped.) 136 137 3. Dead symbols and bindings are cleaned out from the state, including any local 138 bindings. 139 140 4. A CallExitEnd node is generated, which marks the transition back to the 141 caller's LocationContext. 142 143 5. Custom post-call checks are processed and the final nodes are pushed back 144 onto the work list, so that evaluation of the caller can continue. 145 146 Retry Without Inlining 147 ---------------------- 148 149 In some cases, we would like to retry analysis without inlining a particular 150 call. 151 152 Currently, we use this technique to recover coverage in case we stop 153 analyzing a path due to exceeding the maximum block count inside an inlined 154 function. 155 156 When this situation is detected, we walk up the path to find the first node 157 before inlining was started and enqueue it on the WorkList with a special 158 ReplayWithoutInlining bit added to it (ExprEngine::replayWithoutInlining). The 159 path is then re-analyzed from that point without inlining that particular call. 160 161 Deciding When to Inline 162 ----------------------- 163 164 In general, the analyzer attempts to inline as much as possible, since it 165 provides a better summary of what actually happens in the program. There are 166 some cases, however, where the analyzer chooses not to inline: 167 168 - If there is no definition available for the called function or method. In 169 this case, there is no opportunity to inline. 170 171 - If the CFG cannot be constructed for a called function, or the liveness 172 cannot be computed. These are prerequisites for analyzing a function body, 173 with or without inlining. 174 175 - If the LocationContext chain for a given ExplodedNode reaches a maximum cutoff 176 depth. This prevents unbounded analysis due to infinite recursion, but also 177 serves as a useful cutoff for performance reasons. 178 179 - If the function is variadic. This is not a hard limitation, but an engineering 180 limitation. 181 182 Tracked by: <rdar://problem/12147064> Support inlining of variadic functions 183 184 - In C++, constructors are not inlined unless the destructor call will be 185 processed by the ExprEngine. Thus, if the CFG was built without nodes for 186 implicit destructors, or if the destructors for the given object are not 187 represented in the CFG, the constructor will not be inlined. (As an exception, 188 constructors for objects with trivial constructors can still be inlined.) 189 See "C++ Caveats" below. 190 191 - In C++, ExprEngine does not inline custom implementations of operator 'new' 192 or operator 'delete', nor does it inline the constructors and destructors 193 associated with these. See "C++ Caveats" below. 194 195 - Calls resulting in "dynamic dispatch" are specially handled. See more below. 196 197 - The FunctionSummaries map stores additional information about declarations, 198 some of which is collected at runtime based on previous analyses. 199 We do not inline functions which were not profitable to inline in a different 200 context (for example, if the maximum block count was exceeded; see 201 "Retry Without Inlining"). 202 203 204 Dynamic Calls and Devirtualization 205 ---------------------------------- 206 207 "Dynamic" calls are those that are resolved at runtime, such as C++ virtual 208 method calls and Objective-C message sends. Due to the path-sensitive nature of 209 the analysis, the analyzer may be able to reason about the dynamic type of the 210 object whose method is being called and thus "devirtualize" the call. 211 212 This path-sensitive devirtualization occurs when the analyzer can determine what 213 method would actually be called at runtime. This is possible when the type 214 information is constrained enough for a simulated C++/Objective-C object that 215 the analyzer can make such a decision. 216 217 == DynamicTypeInfo == 218 219 As the analyzer analyzes a path, it may accrue information to refine the 220 knowledge about the type of an object. This can then be used to make better 221 decisions about the target method of a call. 222 223 Such type information is tracked as DynamicTypeInfo. This is path-sensitive 224 data that is stored in ProgramState, which defines a mapping from MemRegions to 225 an (optional) DynamicTypeInfo. 226 227 If no DynamicTypeInfo has been explicitly set for a MemRegion, it will be lazily 228 inferred from the region's type or associated symbol. Information from symbolic 229 regions is weaker than from true typed regions. 230 231 EXAMPLE: A C++ object declared "A obj" is known to have the class 'A', but a 232 reference "A &ref" may dynamically be a subclass of 'A'. 233 234 The DynamicTypePropagation checker gathers and propagates DynamicTypeInfo, 235 updating it as information is observed along a path that can refine that type 236 information for a region. 237 238 WARNING: Not all of the existing analyzer code has been retrofitted to use 239 DynamicTypeInfo, nor is it universally appropriate. In particular, 240 DynamicTypeInfo always applies to a region with all casts stripped 241 off, but sometimes the information provided by casts can be useful. 242 243 244 == RuntimeDefinition == 245 246 The basis of devirtualization is CallEvent's getRuntimeDefinition() method, 247 which returns a RuntimeDefinition object. When asked to provide a definition, 248 the CallEvents for dynamic calls will use the DynamicTypeInfo in their 249 ProgramState to attempt to devirtualize the call. In the case of no dynamic 250 dispatch, or perfectly constrained devirtualization, the resulting 251 RuntimeDefinition contains a Decl corresponding to the definition of the called 252 function, and RuntimeDefinition::mayHaveOtherDefinitions will return FALSE. 253 254 In the case of dynamic dispatch where our information is not perfect, CallEvent 255 can make a guess, but RuntimeDefinition::mayHaveOtherDefinitions will return 256 TRUE. The RuntimeDefinition object will then also include a MemRegion 257 corresponding to the object being called (i.e., the "receiver" in Objective-C 258 parlance), which ExprEngine uses to decide whether or not the call should be 259 inlined. 260 261 == Inlining Dynamic Calls == 262 263 The -analyzer-config ipa option has five different modes: none, basic-inlining, 264 inlining, dynamic, and dynamic-bifurcate. Under -analyzer-config ipa=dynamic, 265 all dynamic calls are inlined, whether we are certain or not that this will 266 actually be the definition used at runtime. Under -analyzer-config ipa=inlining, 267 only "near-perfect" devirtualized calls are inlined*, and other dynamic calls 268 are evaluated conservatively (as if no definition were available). 269 270 * Currently, no Objective-C messages are not inlined under 271 -analyzer-config ipa=inlining, even if we are reasonably confident of the type 272 of the receiver. We plan to enable this once we have tested our heuristics 273 more thoroughly. 274 275 The last option, -analyzer-config ipa=dynamic-bifurcate, behaves similarly to 276 "dynamic", but performs a conservative invalidation in the general virtual case 277 in *addition* to inlining. The details of this are discussed below. 278 279 As stated above, -analyzer-config ipa=basic-inlining does not inline any C++ 280 member functions or Objective-C method calls, even if they are non-virtual or 281 can be safely devirtualized. 282 283 284 Bifurcation 285 ----------- 286 287 ExprEngine::BifurcateCall implements the -analyzer-config ipa=dynamic-bifurcate 288 mode. 289 290 When a call is made on an object with imprecise dynamic type information 291 (RuntimeDefinition::mayHaveOtherDefinitions() evaluates to TRUE), ExprEngine 292 bifurcates the path and marks the object's region (retrieved from the 293 RuntimeDefinition object) with a path-sensitive "mode" in the ProgramState. 294 295 Currently, there are 2 modes: 296 297 DynamicDispatchModeInlined - Models the case where the dynamic type information 298 of the receiver (MemoryRegion) is assumed to be perfectly constrained so 299 that a given definition of a method is expected to be the code actually 300 called. When this mode is set, ExprEngine uses the Decl from 301 RuntimeDefinition to inline any dynamically dispatched call sent to this 302 receiver because the function definition is considered to be fully resolved. 303 304 DynamicDispatchModeConservative - Models the case where the dynamic type 305 information is assumed to be incorrect, for example, implies that the method 306 definition is overriden in a subclass. In such cases, ExprEngine does not 307 inline the methods sent to the receiver (MemoryRegion), even if a candidate 308 definition is available. This mode is conservative about simulating the 309 effects of a call. 310 311 Going forward along the symbolic execution path, ExprEngine consults the mode 312 of the receiver's MemRegion to make decisions on whether the calls should be 313 inlined or not, which ensures that there is at most one split per region. 314 315 At a high level, "bifurcation mode" allows for increased semantic coverage in 316 cases where the parent method contains code which is only executed when the 317 class is subclassed. The disadvantages of this mode are a (considerable?) 318 performance hit and the possibility of false positives on the path where the 319 conservative mode is used. 320 321 Objective-C Message Heuristics 322 ------------------------------ 323 324 ExprEngine relies on a set of heuristics to partition the set of Objective-C 325 method calls into those that require bifurcation and those that do not. Below 326 are the cases when the DynamicTypeInfo of the object is considered precise 327 (cannot be a subclass): 328 329 - If the object was created with +alloc or +new and initialized with an -init 330 method. 331 332 - If the calls are property accesses using dot syntax. This is based on the 333 assumption that children rarely override properties, or do so in an 334 essentially compatible way. 335 336 - If the class interface is declared inside the main source file. In this case 337 it is unlikely that it will be subclassed. 338 339 - If the method is not declared outside of main source file, either by the 340 receiver's class or by any superclasses. 341 342 C++ Caveats 343 -------------------- 344 345 C++11 [class.cdtor]p4 describes how the vtable of an object is modified as it is 346 being constructed or destructed; that is, the type of the object depends on 347 which base constructors have been completed. This is tracked using 348 DynamicTypeInfo in the DynamicTypePropagation checker. 349 350 There are several limitations in the current implementation: 351 352 - Temporaries are poorly modeled right now because we're not confident in the 353 placement of their destructors in the CFG. We currently won't inline their 354 constructors unless the destructor is trivial, and don't process their 355 destructors at all, not even to invalidate the region. 356 357 - 'new' is poorly modeled due to some nasty CFG/design issues. This is tracked 358 in PR12014. 'delete' is not modeled at all. 359 360 - Arrays of objects are modeled very poorly right now. ExprEngine currently 361 only simulates the first constructor and first destructor. Because of this, 362 ExprEngine does not inline any constructors or destructors for arrays. 363 364 365 CallEvent 366 ========= 367 368 A CallEvent represents a specific call to a function, method, or other body of 369 code. It is path-sensitive, containing both the current state (ProgramStateRef) 370 and stack space (LocationContext), and provides uniform access to the argument 371 values and return type of a call, no matter how the call is written in the 372 source or what sort of code body is being invoked. 373 374 NOTE: For those familiar with Cocoa, CallEvent is roughly equivalent to 375 NSInvocation. 376 377 CallEvent should be used whenever there is logic dealing with function calls 378 that does not care how the call occurred. 379 380 Examples include checking that arguments satisfy preconditions (such as 381 __attribute__((nonnull))), and attempting to inline a call. 382 383 CallEvents are reference-counted objects managed by a CallEventManager. While 384 there is no inherent issue with persisting them (say, in a ProgramState's GDM), 385 they are intended for short-lived use, and can be recreated from CFGElements or 386 non-top-level StackFrameContexts fairly easily. 387
