File: | src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Transforms/IPO/OpenMPOpt.cpp |
Warning: | line 3402, column 7 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===-- IPO/OpenMPOpt.cpp - Collection of OpenMP specific optimizations ---===// | |||
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 | // OpenMP specific optimizations: | |||
10 | // | |||
11 | // - Deduplication of runtime calls, e.g., omp_get_thread_num. | |||
12 | // - Replacing globalized device memory with stack memory. | |||
13 | // - Replacing globalized device memory with shared memory. | |||
14 | // - Parallel region merging. | |||
15 | // - Transforming generic-mode device kernels to SPMD mode. | |||
16 | // - Specializing the state machine for generic-mode device kernels. | |||
17 | // | |||
18 | //===----------------------------------------------------------------------===// | |||
19 | ||||
20 | #include "llvm/Transforms/IPO/OpenMPOpt.h" | |||
21 | ||||
22 | #include "llvm/ADT/EnumeratedArray.h" | |||
23 | #include "llvm/ADT/PostOrderIterator.h" | |||
24 | #include "llvm/ADT/Statistic.h" | |||
25 | #include "llvm/Analysis/CallGraph.h" | |||
26 | #include "llvm/Analysis/CallGraphSCCPass.h" | |||
27 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" | |||
28 | #include "llvm/Analysis/ValueTracking.h" | |||
29 | #include "llvm/Frontend/OpenMP/OMPConstants.h" | |||
30 | #include "llvm/Frontend/OpenMP/OMPIRBuilder.h" | |||
31 | #include "llvm/IR/Assumptions.h" | |||
32 | #include "llvm/IR/DiagnosticInfo.h" | |||
33 | #include "llvm/IR/GlobalValue.h" | |||
34 | #include "llvm/IR/Instruction.h" | |||
35 | #include "llvm/IR/IntrinsicInst.h" | |||
36 | #include "llvm/InitializePasses.h" | |||
37 | #include "llvm/Support/CommandLine.h" | |||
38 | #include "llvm/Transforms/IPO.h" | |||
39 | #include "llvm/Transforms/IPO/Attributor.h" | |||
40 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" | |||
41 | #include "llvm/Transforms/Utils/CallGraphUpdater.h" | |||
42 | #include "llvm/Transforms/Utils/CodeExtractor.h" | |||
43 | ||||
44 | using namespace llvm; | |||
45 | using namespace omp; | |||
46 | ||||
47 | #define DEBUG_TYPE"openmp-opt" "openmp-opt" | |||
48 | ||||
49 | static cl::opt<bool> DisableOpenMPOptimizations( | |||
50 | "openmp-opt-disable", cl::ZeroOrMore, | |||
51 | cl::desc("Disable OpenMP specific optimizations."), cl::Hidden, | |||
52 | cl::init(false)); | |||
53 | ||||
54 | static cl::opt<bool> EnableParallelRegionMerging( | |||
55 | "openmp-opt-enable-merging", cl::ZeroOrMore, | |||
56 | cl::desc("Enable the OpenMP region merging optimization."), cl::Hidden, | |||
57 | cl::init(false)); | |||
58 | ||||
59 | static cl::opt<bool> | |||
60 | DisableInternalization("openmp-opt-disable-internalization", cl::ZeroOrMore, | |||
61 | cl::desc("Disable function internalization."), | |||
62 | cl::Hidden, cl::init(false)); | |||
63 | ||||
64 | static cl::opt<bool> PrintICVValues("openmp-print-icv-values", cl::init(false), | |||
65 | cl::Hidden); | |||
66 | static cl::opt<bool> PrintOpenMPKernels("openmp-print-gpu-kernels", | |||
67 | cl::init(false), cl::Hidden); | |||
68 | ||||
69 | static cl::opt<bool> HideMemoryTransferLatency( | |||
70 | "openmp-hide-memory-transfer-latency", | |||
71 | cl::desc("[WIP] Tries to hide the latency of host to device memory" | |||
72 | " transfers"), | |||
73 | cl::Hidden, cl::init(false)); | |||
74 | ||||
75 | STATISTIC(NumOpenMPRuntimeCallsDeduplicated,static llvm::Statistic NumOpenMPRuntimeCallsDeduplicated = {"openmp-opt" , "NumOpenMPRuntimeCallsDeduplicated", "Number of OpenMP runtime calls deduplicated" } | |||
76 | "Number of OpenMP runtime calls deduplicated")static llvm::Statistic NumOpenMPRuntimeCallsDeduplicated = {"openmp-opt" , "NumOpenMPRuntimeCallsDeduplicated", "Number of OpenMP runtime calls deduplicated" }; | |||
77 | STATISTIC(NumOpenMPParallelRegionsDeleted,static llvm::Statistic NumOpenMPParallelRegionsDeleted = {"openmp-opt" , "NumOpenMPParallelRegionsDeleted", "Number of OpenMP parallel regions deleted" } | |||
78 | "Number of OpenMP parallel regions deleted")static llvm::Statistic NumOpenMPParallelRegionsDeleted = {"openmp-opt" , "NumOpenMPParallelRegionsDeleted", "Number of OpenMP parallel regions deleted" }; | |||
79 | STATISTIC(NumOpenMPRuntimeFunctionsIdentified,static llvm::Statistic NumOpenMPRuntimeFunctionsIdentified = { "openmp-opt", "NumOpenMPRuntimeFunctionsIdentified", "Number of OpenMP runtime functions identified" } | |||
80 | "Number of OpenMP runtime functions identified")static llvm::Statistic NumOpenMPRuntimeFunctionsIdentified = { "openmp-opt", "NumOpenMPRuntimeFunctionsIdentified", "Number of OpenMP runtime functions identified" }; | |||
81 | STATISTIC(NumOpenMPRuntimeFunctionUsesIdentified,static llvm::Statistic NumOpenMPRuntimeFunctionUsesIdentified = {"openmp-opt", "NumOpenMPRuntimeFunctionUsesIdentified", "Number of OpenMP runtime function uses identified" } | |||
82 | "Number of OpenMP runtime function uses identified")static llvm::Statistic NumOpenMPRuntimeFunctionUsesIdentified = {"openmp-opt", "NumOpenMPRuntimeFunctionUsesIdentified", "Number of OpenMP runtime function uses identified" }; | |||
83 | STATISTIC(NumOpenMPTargetRegionKernels,static llvm::Statistic NumOpenMPTargetRegionKernels = {"openmp-opt" , "NumOpenMPTargetRegionKernels", "Number of OpenMP target region entry points (=kernels) identified" } | |||
84 | "Number of OpenMP target region entry points (=kernels) identified")static llvm::Statistic NumOpenMPTargetRegionKernels = {"openmp-opt" , "NumOpenMPTargetRegionKernels", "Number of OpenMP target region entry points (=kernels) identified" }; | |||
85 | STATISTIC(NumOpenMPTargetRegionKernelsSPMD,static llvm::Statistic NumOpenMPTargetRegionKernelsSPMD = {"openmp-opt" , "NumOpenMPTargetRegionKernelsSPMD", "Number of OpenMP target region entry points (=kernels) executed in " "SPMD-mode instead of generic-mode"} | |||
86 | "Number of OpenMP target region entry points (=kernels) executed in "static llvm::Statistic NumOpenMPTargetRegionKernelsSPMD = {"openmp-opt" , "NumOpenMPTargetRegionKernelsSPMD", "Number of OpenMP target region entry points (=kernels) executed in " "SPMD-mode instead of generic-mode"} | |||
87 | "SPMD-mode instead of generic-mode")static llvm::Statistic NumOpenMPTargetRegionKernelsSPMD = {"openmp-opt" , "NumOpenMPTargetRegionKernelsSPMD", "Number of OpenMP target region entry points (=kernels) executed in " "SPMD-mode instead of generic-mode"}; | |||
88 | STATISTIC(NumOpenMPTargetRegionKernelsWithoutStateMachine,static llvm::Statistic NumOpenMPTargetRegionKernelsWithoutStateMachine = {"openmp-opt", "NumOpenMPTargetRegionKernelsWithoutStateMachine" , "Number of OpenMP target region entry points (=kernels) executed in " "generic-mode without a state machines"} | |||
89 | "Number of OpenMP target region entry points (=kernels) executed in "static llvm::Statistic NumOpenMPTargetRegionKernelsWithoutStateMachine = {"openmp-opt", "NumOpenMPTargetRegionKernelsWithoutStateMachine" , "Number of OpenMP target region entry points (=kernels) executed in " "generic-mode without a state machines"} | |||
90 | "generic-mode without a state machines")static llvm::Statistic NumOpenMPTargetRegionKernelsWithoutStateMachine = {"openmp-opt", "NumOpenMPTargetRegionKernelsWithoutStateMachine" , "Number of OpenMP target region entry points (=kernels) executed in " "generic-mode without a state machines"}; | |||
91 | STATISTIC(NumOpenMPTargetRegionKernelsCustomStateMachineWithFallback,static llvm::Statistic NumOpenMPTargetRegionKernelsCustomStateMachineWithFallback = {"openmp-opt", "NumOpenMPTargetRegionKernelsCustomStateMachineWithFallback" , "Number of OpenMP target region entry points (=kernels) executed in " "generic-mode with customized state machines with fallback"} | |||
92 | "Number of OpenMP target region entry points (=kernels) executed in "static llvm::Statistic NumOpenMPTargetRegionKernelsCustomStateMachineWithFallback = {"openmp-opt", "NumOpenMPTargetRegionKernelsCustomStateMachineWithFallback" , "Number of OpenMP target region entry points (=kernels) executed in " "generic-mode with customized state machines with fallback"} | |||
93 | "generic-mode with customized state machines with fallback")static llvm::Statistic NumOpenMPTargetRegionKernelsCustomStateMachineWithFallback = {"openmp-opt", "NumOpenMPTargetRegionKernelsCustomStateMachineWithFallback" , "Number of OpenMP target region entry points (=kernels) executed in " "generic-mode with customized state machines with fallback"}; | |||
94 | STATISTIC(NumOpenMPTargetRegionKernelsCustomStateMachineWithoutFallback,static llvm::Statistic NumOpenMPTargetRegionKernelsCustomStateMachineWithoutFallback = {"openmp-opt", "NumOpenMPTargetRegionKernelsCustomStateMachineWithoutFallback" , "Number of OpenMP target region entry points (=kernels) executed in " "generic-mode with customized state machines without fallback" } | |||
95 | "Number of OpenMP target region entry points (=kernels) executed in "static llvm::Statistic NumOpenMPTargetRegionKernelsCustomStateMachineWithoutFallback = {"openmp-opt", "NumOpenMPTargetRegionKernelsCustomStateMachineWithoutFallback" , "Number of OpenMP target region entry points (=kernels) executed in " "generic-mode with customized state machines without fallback" } | |||
96 | "generic-mode with customized state machines without fallback")static llvm::Statistic NumOpenMPTargetRegionKernelsCustomStateMachineWithoutFallback = {"openmp-opt", "NumOpenMPTargetRegionKernelsCustomStateMachineWithoutFallback" , "Number of OpenMP target region entry points (=kernels) executed in " "generic-mode with customized state machines without fallback" }; | |||
97 | STATISTIC(static llvm::Statistic NumOpenMPParallelRegionsReplacedInGPUStateMachine = {"openmp-opt", "NumOpenMPParallelRegionsReplacedInGPUStateMachine" , "Number of OpenMP parallel regions replaced with ID in GPU state machines" } | |||
98 | NumOpenMPParallelRegionsReplacedInGPUStateMachine,static llvm::Statistic NumOpenMPParallelRegionsReplacedInGPUStateMachine = {"openmp-opt", "NumOpenMPParallelRegionsReplacedInGPUStateMachine" , "Number of OpenMP parallel regions replaced with ID in GPU state machines" } | |||
99 | "Number of OpenMP parallel regions replaced with ID in GPU state machines")static llvm::Statistic NumOpenMPParallelRegionsReplacedInGPUStateMachine = {"openmp-opt", "NumOpenMPParallelRegionsReplacedInGPUStateMachine" , "Number of OpenMP parallel regions replaced with ID in GPU state machines" }; | |||
100 | STATISTIC(NumOpenMPParallelRegionsMerged,static llvm::Statistic NumOpenMPParallelRegionsMerged = {"openmp-opt" , "NumOpenMPParallelRegionsMerged", "Number of OpenMP parallel regions merged" } | |||
101 | "Number of OpenMP parallel regions merged")static llvm::Statistic NumOpenMPParallelRegionsMerged = {"openmp-opt" , "NumOpenMPParallelRegionsMerged", "Number of OpenMP parallel regions merged" }; | |||
102 | STATISTIC(NumBytesMovedToSharedMemory,static llvm::Statistic NumBytesMovedToSharedMemory = {"openmp-opt" , "NumBytesMovedToSharedMemory", "Amount of memory pushed to shared memory" } | |||
103 | "Amount of memory pushed to shared memory")static llvm::Statistic NumBytesMovedToSharedMemory = {"openmp-opt" , "NumBytesMovedToSharedMemory", "Amount of memory pushed to shared memory" }; | |||
104 | ||||
105 | #if !defined(NDEBUG1) | |||
106 | static constexpr auto TAG = "[" DEBUG_TYPE"openmp-opt" "]"; | |||
107 | #endif | |||
108 | ||||
109 | namespace { | |||
110 | ||||
111 | enum class AddressSpace : unsigned { | |||
112 | Generic = 0, | |||
113 | Global = 1, | |||
114 | Shared = 3, | |||
115 | Constant = 4, | |||
116 | Local = 5, | |||
117 | }; | |||
118 | ||||
119 | struct AAHeapToShared; | |||
120 | ||||
121 | struct AAICVTracker; | |||
122 | ||||
123 | /// OpenMP specific information. For now, stores RFIs and ICVs also needed for | |||
124 | /// Attributor runs. | |||
125 | struct OMPInformationCache : public InformationCache { | |||
126 | OMPInformationCache(Module &M, AnalysisGetter &AG, | |||
127 | BumpPtrAllocator &Allocator, SetVector<Function *> &CGSCC, | |||
128 | SmallPtrSetImpl<Kernel> &Kernels) | |||
129 | : InformationCache(M, AG, Allocator, &CGSCC), OMPBuilder(M), | |||
130 | Kernels(Kernels) { | |||
131 | ||||
132 | OMPBuilder.initialize(); | |||
133 | initializeRuntimeFunctions(); | |||
134 | initializeInternalControlVars(); | |||
135 | } | |||
136 | ||||
137 | /// Generic information that describes an internal control variable. | |||
138 | struct InternalControlVarInfo { | |||
139 | /// The kind, as described by InternalControlVar enum. | |||
140 | InternalControlVar Kind; | |||
141 | ||||
142 | /// The name of the ICV. | |||
143 | StringRef Name; | |||
144 | ||||
145 | /// Environment variable associated with this ICV. | |||
146 | StringRef EnvVarName; | |||
147 | ||||
148 | /// Initial value kind. | |||
149 | ICVInitValue InitKind; | |||
150 | ||||
151 | /// Initial value. | |||
152 | ConstantInt *InitValue; | |||
153 | ||||
154 | /// Setter RTL function associated with this ICV. | |||
155 | RuntimeFunction Setter; | |||
156 | ||||
157 | /// Getter RTL function associated with this ICV. | |||
158 | RuntimeFunction Getter; | |||
159 | ||||
160 | /// RTL Function corresponding to the override clause of this ICV | |||
161 | RuntimeFunction Clause; | |||
162 | }; | |||
163 | ||||
164 | /// Generic information that describes a runtime function | |||
165 | struct RuntimeFunctionInfo { | |||
166 | ||||
167 | /// The kind, as described by the RuntimeFunction enum. | |||
168 | RuntimeFunction Kind; | |||
169 | ||||
170 | /// The name of the function. | |||
171 | StringRef Name; | |||
172 | ||||
173 | /// Flag to indicate a variadic function. | |||
174 | bool IsVarArg; | |||
175 | ||||
176 | /// The return type of the function. | |||
177 | Type *ReturnType; | |||
178 | ||||
179 | /// The argument types of the function. | |||
180 | SmallVector<Type *, 8> ArgumentTypes; | |||
181 | ||||
182 | /// The declaration if available. | |||
183 | Function *Declaration = nullptr; | |||
184 | ||||
185 | /// Uses of this runtime function per function containing the use. | |||
186 | using UseVector = SmallVector<Use *, 16>; | |||
187 | ||||
188 | /// Clear UsesMap for runtime function. | |||
189 | void clearUsesMap() { UsesMap.clear(); } | |||
190 | ||||
191 | /// Boolean conversion that is true if the runtime function was found. | |||
192 | operator bool() const { return Declaration; } | |||
193 | ||||
194 | /// Return the vector of uses in function \p F. | |||
195 | UseVector &getOrCreateUseVector(Function *F) { | |||
196 | std::shared_ptr<UseVector> &UV = UsesMap[F]; | |||
197 | if (!UV) | |||
198 | UV = std::make_shared<UseVector>(); | |||
199 | return *UV; | |||
200 | } | |||
201 | ||||
202 | /// Return the vector of uses in function \p F or `nullptr` if there are | |||
203 | /// none. | |||
204 | const UseVector *getUseVector(Function &F) const { | |||
205 | auto I = UsesMap.find(&F); | |||
206 | if (I != UsesMap.end()) | |||
207 | return I->second.get(); | |||
208 | return nullptr; | |||
209 | } | |||
210 | ||||
211 | /// Return how many functions contain uses of this runtime function. | |||
212 | size_t getNumFunctionsWithUses() const { return UsesMap.size(); } | |||
213 | ||||
214 | /// Return the number of arguments (or the minimal number for variadic | |||
215 | /// functions). | |||
216 | size_t getNumArgs() const { return ArgumentTypes.size(); } | |||
217 | ||||
218 | /// Run the callback \p CB on each use and forget the use if the result is | |||
219 | /// true. The callback will be fed the function in which the use was | |||
220 | /// encountered as second argument. | |||
221 | void foreachUse(SmallVectorImpl<Function *> &SCC, | |||
222 | function_ref<bool(Use &, Function &)> CB) { | |||
223 | for (Function *F : SCC) | |||
224 | foreachUse(CB, F); | |||
225 | } | |||
226 | ||||
227 | /// Run the callback \p CB on each use within the function \p F and forget | |||
228 | /// the use if the result is true. | |||
229 | void foreachUse(function_ref<bool(Use &, Function &)> CB, Function *F) { | |||
230 | SmallVector<unsigned, 8> ToBeDeleted; | |||
231 | ToBeDeleted.clear(); | |||
232 | ||||
233 | unsigned Idx = 0; | |||
234 | UseVector &UV = getOrCreateUseVector(F); | |||
235 | ||||
236 | for (Use *U : UV) { | |||
237 | if (CB(*U, *F)) | |||
238 | ToBeDeleted.push_back(Idx); | |||
239 | ++Idx; | |||
240 | } | |||
241 | ||||
242 | // Remove the to-be-deleted indices in reverse order as prior | |||
243 | // modifications will not modify the smaller indices. | |||
244 | while (!ToBeDeleted.empty()) { | |||
245 | unsigned Idx = ToBeDeleted.pop_back_val(); | |||
246 | UV[Idx] = UV.back(); | |||
247 | UV.pop_back(); | |||
248 | } | |||
249 | } | |||
250 | ||||
251 | private: | |||
252 | /// Map from functions to all uses of this runtime function contained in | |||
253 | /// them. | |||
254 | DenseMap<Function *, std::shared_ptr<UseVector>> UsesMap; | |||
255 | ||||
256 | public: | |||
257 | /// Iterators for the uses of this runtime function. | |||
258 | decltype(UsesMap)::iterator begin() { return UsesMap.begin(); } | |||
259 | decltype(UsesMap)::iterator end() { return UsesMap.end(); } | |||
260 | }; | |||
261 | ||||
262 | /// An OpenMP-IR-Builder instance | |||
263 | OpenMPIRBuilder OMPBuilder; | |||
264 | ||||
265 | /// Map from runtime function kind to the runtime function description. | |||
266 | EnumeratedArray<RuntimeFunctionInfo, RuntimeFunction, | |||
267 | RuntimeFunction::OMPRTL___last> | |||
268 | RFIs; | |||
269 | ||||
270 | /// Map from function declarations/definitions to their runtime enum type. | |||
271 | DenseMap<Function *, RuntimeFunction> RuntimeFunctionIDMap; | |||
272 | ||||
273 | /// Map from ICV kind to the ICV description. | |||
274 | EnumeratedArray<InternalControlVarInfo, InternalControlVar, | |||
275 | InternalControlVar::ICV___last> | |||
276 | ICVs; | |||
277 | ||||
278 | /// Helper to initialize all internal control variable information for those | |||
279 | /// defined in OMPKinds.def. | |||
280 | void initializeInternalControlVars() { | |||
281 | #define ICV_RT_SET(_Name, RTL) \ | |||
282 | { \ | |||
283 | auto &ICV = ICVs[_Name]; \ | |||
284 | ICV.Setter = RTL; \ | |||
285 | } | |||
286 | #define ICV_RT_GET(Name, RTL) \ | |||
287 | { \ | |||
288 | auto &ICV = ICVs[Name]; \ | |||
289 | ICV.Getter = RTL; \ | |||
290 | } | |||
291 | #define ICV_DATA_ENV(Enum, _Name, _EnvVarName, Init) \ | |||
292 | { \ | |||
293 | auto &ICV = ICVs[Enum]; \ | |||
294 | ICV.Name = _Name; \ | |||
295 | ICV.Kind = Enum; \ | |||
296 | ICV.InitKind = Init; \ | |||
297 | ICV.EnvVarName = _EnvVarName; \ | |||
298 | switch (ICV.InitKind) { \ | |||
299 | case ICV_IMPLEMENTATION_DEFINED: \ | |||
300 | ICV.InitValue = nullptr; \ | |||
301 | break; \ | |||
302 | case ICV_ZERO: \ | |||
303 | ICV.InitValue = ConstantInt::get( \ | |||
304 | Type::getInt32Ty(OMPBuilder.Int32->getContext()), 0); \ | |||
305 | break; \ | |||
306 | case ICV_FALSE: \ | |||
307 | ICV.InitValue = ConstantInt::getFalse(OMPBuilder.Int1->getContext()); \ | |||
308 | break; \ | |||
309 | case ICV_LAST: \ | |||
310 | break; \ | |||
311 | } \ | |||
312 | } | |||
313 | #include "llvm/Frontend/OpenMP/OMPKinds.def" | |||
314 | } | |||
315 | ||||
316 | /// Returns true if the function declaration \p F matches the runtime | |||
317 | /// function types, that is, return type \p RTFRetType, and argument types | |||
318 | /// \p RTFArgTypes. | |||
319 | static bool declMatchesRTFTypes(Function *F, Type *RTFRetType, | |||
320 | SmallVector<Type *, 8> &RTFArgTypes) { | |||
321 | // TODO: We should output information to the user (under debug output | |||
322 | // and via remarks). | |||
323 | ||||
324 | if (!F) | |||
325 | return false; | |||
326 | if (F->getReturnType() != RTFRetType) | |||
327 | return false; | |||
328 | if (F->arg_size() != RTFArgTypes.size()) | |||
329 | return false; | |||
330 | ||||
331 | auto RTFTyIt = RTFArgTypes.begin(); | |||
332 | for (Argument &Arg : F->args()) { | |||
333 | if (Arg.getType() != *RTFTyIt) | |||
334 | return false; | |||
335 | ||||
336 | ++RTFTyIt; | |||
337 | } | |||
338 | ||||
339 | return true; | |||
340 | } | |||
341 | ||||
342 | // Helper to collect all uses of the declaration in the UsesMap. | |||
343 | unsigned collectUses(RuntimeFunctionInfo &RFI, bool CollectStats = true) { | |||
344 | unsigned NumUses = 0; | |||
345 | if (!RFI.Declaration) | |||
346 | return NumUses; | |||
347 | OMPBuilder.addAttributes(RFI.Kind, *RFI.Declaration); | |||
348 | ||||
349 | if (CollectStats) { | |||
350 | NumOpenMPRuntimeFunctionsIdentified += 1; | |||
351 | NumOpenMPRuntimeFunctionUsesIdentified += RFI.Declaration->getNumUses(); | |||
352 | } | |||
353 | ||||
354 | // TODO: We directly convert uses into proper calls and unknown uses. | |||
355 | for (Use &U : RFI.Declaration->uses()) { | |||
356 | if (Instruction *UserI = dyn_cast<Instruction>(U.getUser())) { | |||
357 | if (ModuleSlice.count(UserI->getFunction())) { | |||
358 | RFI.getOrCreateUseVector(UserI->getFunction()).push_back(&U); | |||
359 | ++NumUses; | |||
360 | } | |||
361 | } else { | |||
362 | RFI.getOrCreateUseVector(nullptr).push_back(&U); | |||
363 | ++NumUses; | |||
364 | } | |||
365 | } | |||
366 | return NumUses; | |||
367 | } | |||
368 | ||||
369 | // Helper function to recollect uses of a runtime function. | |||
370 | void recollectUsesForFunction(RuntimeFunction RTF) { | |||
371 | auto &RFI = RFIs[RTF]; | |||
372 | RFI.clearUsesMap(); | |||
373 | collectUses(RFI, /*CollectStats*/ false); | |||
374 | } | |||
375 | ||||
376 | // Helper function to recollect uses of all runtime functions. | |||
377 | void recollectUses() { | |||
378 | for (int Idx = 0; Idx < RFIs.size(); ++Idx) | |||
379 | recollectUsesForFunction(static_cast<RuntimeFunction>(Idx)); | |||
380 | } | |||
381 | ||||
382 | /// Helper to initialize all runtime function information for those defined | |||
383 | /// in OpenMPKinds.def. | |||
384 | void initializeRuntimeFunctions() { | |||
385 | Module &M = *((*ModuleSlice.begin())->getParent()); | |||
386 | ||||
387 | // Helper macros for handling __VA_ARGS__ in OMP_RTL | |||
388 | #define OMP_TYPE(VarName, ...) \ | |||
389 | Type *VarName = OMPBuilder.VarName; \ | |||
390 | (void)VarName; | |||
391 | ||||
392 | #define OMP_ARRAY_TYPE(VarName, ...) \ | |||
393 | ArrayType *VarName##Ty = OMPBuilder.VarName##Ty; \ | |||
394 | (void)VarName##Ty; \ | |||
395 | PointerType *VarName##PtrTy = OMPBuilder.VarName##PtrTy; \ | |||
396 | (void)VarName##PtrTy; | |||
397 | ||||
398 | #define OMP_FUNCTION_TYPE(VarName, ...) \ | |||
399 | FunctionType *VarName = OMPBuilder.VarName; \ | |||
400 | (void)VarName; \ | |||
401 | PointerType *VarName##Ptr = OMPBuilder.VarName##Ptr; \ | |||
402 | (void)VarName##Ptr; | |||
403 | ||||
404 | #define OMP_STRUCT_TYPE(VarName, ...) \ | |||
405 | StructType *VarName = OMPBuilder.VarName; \ | |||
406 | (void)VarName; \ | |||
407 | PointerType *VarName##Ptr = OMPBuilder.VarName##Ptr; \ | |||
408 | (void)VarName##Ptr; | |||
409 | ||||
410 | #define OMP_RTL(_Enum, _Name, _IsVarArg, _ReturnType, ...) \ | |||
411 | { \ | |||
412 | SmallVector<Type *, 8> ArgsTypes({__VA_ARGS__}); \ | |||
413 | Function *F = M.getFunction(_Name); \ | |||
414 | RTLFunctions.insert(F); \ | |||
415 | if (declMatchesRTFTypes(F, OMPBuilder._ReturnType, ArgsTypes)) { \ | |||
416 | RuntimeFunctionIDMap[F] = _Enum; \ | |||
417 | F->removeFnAttr(Attribute::NoInline); \ | |||
418 | auto &RFI = RFIs[_Enum]; \ | |||
419 | RFI.Kind = _Enum; \ | |||
420 | RFI.Name = _Name; \ | |||
421 | RFI.IsVarArg = _IsVarArg; \ | |||
422 | RFI.ReturnType = OMPBuilder._ReturnType; \ | |||
423 | RFI.ArgumentTypes = std::move(ArgsTypes); \ | |||
424 | RFI.Declaration = F; \ | |||
425 | unsigned NumUses = collectUses(RFI); \ | |||
426 | (void)NumUses; \ | |||
427 | LLVM_DEBUG({ \do { } while (false) | |||
428 | dbgs() << TAG << RFI.Name << (RFI.Declaration ? "" : " not") \do { } while (false) | |||
429 | << " found\n"; \do { } while (false) | |||
430 | if (RFI.Declaration) \do { } while (false) | |||
431 | dbgs() << TAG << "-> got " << NumUses << " uses in " \do { } while (false) | |||
432 | << RFI.getNumFunctionsWithUses() \do { } while (false) | |||
433 | << " different functions.\n"; \do { } while (false) | |||
434 | })do { } while (false); \ | |||
435 | } \ | |||
436 | } | |||
437 | #include "llvm/Frontend/OpenMP/OMPKinds.def" | |||
438 | ||||
439 | // TODO: We should attach the attributes defined in OMPKinds.def. | |||
440 | } | |||
441 | ||||
442 | /// Collection of known kernels (\see Kernel) in the module. | |||
443 | SmallPtrSetImpl<Kernel> &Kernels; | |||
444 | ||||
445 | /// Collection of known OpenMP runtime functions.. | |||
446 | DenseSet<const Function *> RTLFunctions; | |||
447 | }; | |||
448 | ||||
449 | template <typename Ty, bool InsertInvalidates = true> | |||
450 | struct BooleanStateWithSetVector : public BooleanState { | |||
451 | bool contains(const Ty &Elem) const { return Set.contains(Elem); } | |||
452 | bool insert(const Ty &Elem) { | |||
453 | if (InsertInvalidates) | |||
454 | BooleanState::indicatePessimisticFixpoint(); | |||
455 | return Set.insert(Elem); | |||
456 | } | |||
457 | ||||
458 | const Ty &operator[](int Idx) const { return Set[Idx]; } | |||
459 | bool operator==(const BooleanStateWithSetVector &RHS) const { | |||
460 | return BooleanState::operator==(RHS) && Set == RHS.Set; | |||
461 | } | |||
462 | bool operator!=(const BooleanStateWithSetVector &RHS) const { | |||
463 | return !(*this == RHS); | |||
464 | } | |||
465 | ||||
466 | bool empty() const { return Set.empty(); } | |||
467 | size_t size() const { return Set.size(); } | |||
468 | ||||
469 | /// "Clamp" this state with \p RHS. | |||
470 | BooleanStateWithSetVector &operator^=(const BooleanStateWithSetVector &RHS) { | |||
471 | BooleanState::operator^=(RHS); | |||
472 | Set.insert(RHS.Set.begin(), RHS.Set.end()); | |||
473 | return *this; | |||
474 | } | |||
475 | ||||
476 | private: | |||
477 | /// A set to keep track of elements. | |||
478 | SetVector<Ty> Set; | |||
479 | ||||
480 | public: | |||
481 | typename decltype(Set)::iterator begin() { return Set.begin(); } | |||
482 | typename decltype(Set)::iterator end() { return Set.end(); } | |||
483 | typename decltype(Set)::const_iterator begin() const { return Set.begin(); } | |||
484 | typename decltype(Set)::const_iterator end() const { return Set.end(); } | |||
485 | }; | |||
486 | ||||
487 | template <typename Ty, bool InsertInvalidates = true> | |||
488 | using BooleanStateWithPtrSetVector = | |||
489 | BooleanStateWithSetVector<Ty *, InsertInvalidates>; | |||
490 | ||||
491 | struct KernelInfoState : AbstractState { | |||
492 | /// Flag to track if we reached a fixpoint. | |||
493 | bool IsAtFixpoint = false; | |||
494 | ||||
495 | /// The parallel regions (identified by the outlined parallel functions) that | |||
496 | /// can be reached from the associated function. | |||
497 | BooleanStateWithPtrSetVector<Function, /* InsertInvalidates */ false> | |||
498 | ReachedKnownParallelRegions; | |||
499 | ||||
500 | /// State to track what parallel region we might reach. | |||
501 | BooleanStateWithPtrSetVector<CallBase> ReachedUnknownParallelRegions; | |||
502 | ||||
503 | /// State to track if we are in SPMD-mode, assumed or know, and why we decided | |||
504 | /// we cannot be. If it is assumed, then RequiresFullRuntime should also be | |||
505 | /// false. | |||
506 | BooleanStateWithPtrSetVector<Instruction> SPMDCompatibilityTracker; | |||
507 | ||||
508 | /// The __kmpc_target_init call in this kernel, if any. If we find more than | |||
509 | /// one we abort as the kernel is malformed. | |||
510 | CallBase *KernelInitCB = nullptr; | |||
511 | ||||
512 | /// The __kmpc_target_deinit call in this kernel, if any. If we find more than | |||
513 | /// one we abort as the kernel is malformed. | |||
514 | CallBase *KernelDeinitCB = nullptr; | |||
515 | ||||
516 | /// Flag to indicate if the associated function is a kernel entry. | |||
517 | bool IsKernelEntry = false; | |||
518 | ||||
519 | /// State to track what kernel entries can reach the associated function. | |||
520 | BooleanStateWithPtrSetVector<Function, false> ReachingKernelEntries; | |||
521 | ||||
522 | /// State to indicate if we can track parallel level of the associated | |||
523 | /// function. We will give up tracking if we encounter unknown caller or the | |||
524 | /// caller is __kmpc_parallel_51. | |||
525 | BooleanStateWithSetVector<uint8_t> ParallelLevels; | |||
526 | ||||
527 | /// Abstract State interface | |||
528 | ///{ | |||
529 | ||||
530 | KernelInfoState() {} | |||
531 | KernelInfoState(bool BestState) { | |||
532 | if (!BestState) | |||
533 | indicatePessimisticFixpoint(); | |||
534 | } | |||
535 | ||||
536 | /// See AbstractState::isValidState(...) | |||
537 | bool isValidState() const override { return true; } | |||
538 | ||||
539 | /// See AbstractState::isAtFixpoint(...) | |||
540 | bool isAtFixpoint() const override { return IsAtFixpoint; } | |||
541 | ||||
542 | /// See AbstractState::indicatePessimisticFixpoint(...) | |||
543 | ChangeStatus indicatePessimisticFixpoint() override { | |||
544 | IsAtFixpoint = true; | |||
545 | SPMDCompatibilityTracker.indicatePessimisticFixpoint(); | |||
546 | ReachedUnknownParallelRegions.indicatePessimisticFixpoint(); | |||
547 | return ChangeStatus::CHANGED; | |||
548 | } | |||
549 | ||||
550 | /// See AbstractState::indicateOptimisticFixpoint(...) | |||
551 | ChangeStatus indicateOptimisticFixpoint() override { | |||
552 | IsAtFixpoint = true; | |||
553 | return ChangeStatus::UNCHANGED; | |||
554 | } | |||
555 | ||||
556 | /// Return the assumed state | |||
557 | KernelInfoState &getAssumed() { return *this; } | |||
558 | const KernelInfoState &getAssumed() const { return *this; } | |||
559 | ||||
560 | bool operator==(const KernelInfoState &RHS) const { | |||
561 | if (SPMDCompatibilityTracker != RHS.SPMDCompatibilityTracker) | |||
562 | return false; | |||
563 | if (ReachedKnownParallelRegions != RHS.ReachedKnownParallelRegions) | |||
564 | return false; | |||
565 | if (ReachedUnknownParallelRegions != RHS.ReachedUnknownParallelRegions) | |||
566 | return false; | |||
567 | if (ReachingKernelEntries != RHS.ReachingKernelEntries) | |||
568 | return false; | |||
569 | return true; | |||
570 | } | |||
571 | ||||
572 | /// Return empty set as the best state of potential values. | |||
573 | static KernelInfoState getBestState() { return KernelInfoState(true); } | |||
574 | ||||
575 | static KernelInfoState getBestState(KernelInfoState &KIS) { | |||
576 | return getBestState(); | |||
577 | } | |||
578 | ||||
579 | /// Return full set as the worst state of potential values. | |||
580 | static KernelInfoState getWorstState() { return KernelInfoState(false); } | |||
581 | ||||
582 | /// "Clamp" this state with \p KIS. | |||
583 | KernelInfoState operator^=(const KernelInfoState &KIS) { | |||
584 | // Do not merge two different _init and _deinit call sites. | |||
585 | if (KIS.KernelInitCB) { | |||
586 | if (KernelInitCB && KernelInitCB != KIS.KernelInitCB) | |||
587 | indicatePessimisticFixpoint(); | |||
588 | KernelInitCB = KIS.KernelInitCB; | |||
589 | } | |||
590 | if (KIS.KernelDeinitCB) { | |||
591 | if (KernelDeinitCB && KernelDeinitCB != KIS.KernelDeinitCB) | |||
592 | indicatePessimisticFixpoint(); | |||
593 | KernelDeinitCB = KIS.KernelDeinitCB; | |||
594 | } | |||
595 | SPMDCompatibilityTracker ^= KIS.SPMDCompatibilityTracker; | |||
596 | ReachedKnownParallelRegions ^= KIS.ReachedKnownParallelRegions; | |||
597 | ReachedUnknownParallelRegions ^= KIS.ReachedUnknownParallelRegions; | |||
598 | return *this; | |||
599 | } | |||
600 | ||||
601 | KernelInfoState operator&=(const KernelInfoState &KIS) { | |||
602 | return (*this ^= KIS); | |||
603 | } | |||
604 | ||||
605 | ///} | |||
606 | }; | |||
607 | ||||
608 | /// Used to map the values physically (in the IR) stored in an offload | |||
609 | /// array, to a vector in memory. | |||
610 | struct OffloadArray { | |||
611 | /// Physical array (in the IR). | |||
612 | AllocaInst *Array = nullptr; | |||
613 | /// Mapped values. | |||
614 | SmallVector<Value *, 8> StoredValues; | |||
615 | /// Last stores made in the offload array. | |||
616 | SmallVector<StoreInst *, 8> LastAccesses; | |||
617 | ||||
618 | OffloadArray() = default; | |||
619 | ||||
620 | /// Initializes the OffloadArray with the values stored in \p Array before | |||
621 | /// instruction \p Before is reached. Returns false if the initialization | |||
622 | /// fails. | |||
623 | /// This MUST be used immediately after the construction of the object. | |||
624 | bool initialize(AllocaInst &Array, Instruction &Before) { | |||
625 | if (!Array.getAllocatedType()->isArrayTy()) | |||
626 | return false; | |||
627 | ||||
628 | if (!getValues(Array, Before)) | |||
629 | return false; | |||
630 | ||||
631 | this->Array = &Array; | |||
632 | return true; | |||
633 | } | |||
634 | ||||
635 | static const unsigned DeviceIDArgNum = 1; | |||
636 | static const unsigned BasePtrsArgNum = 3; | |||
637 | static const unsigned PtrsArgNum = 4; | |||
638 | static const unsigned SizesArgNum = 5; | |||
639 | ||||
640 | private: | |||
641 | /// Traverses the BasicBlock where \p Array is, collecting the stores made to | |||
642 | /// \p Array, leaving StoredValues with the values stored before the | |||
643 | /// instruction \p Before is reached. | |||
644 | bool getValues(AllocaInst &Array, Instruction &Before) { | |||
645 | // Initialize container. | |||
646 | const uint64_t NumValues = Array.getAllocatedType()->getArrayNumElements(); | |||
647 | StoredValues.assign(NumValues, nullptr); | |||
648 | LastAccesses.assign(NumValues, nullptr); | |||
649 | ||||
650 | // TODO: This assumes the instruction \p Before is in the same | |||
651 | // BasicBlock as Array. Make it general, for any control flow graph. | |||
652 | BasicBlock *BB = Array.getParent(); | |||
653 | if (BB != Before.getParent()) | |||
654 | return false; | |||
655 | ||||
656 | const DataLayout &DL = Array.getModule()->getDataLayout(); | |||
657 | const unsigned int PointerSize = DL.getPointerSize(); | |||
658 | ||||
659 | for (Instruction &I : *BB) { | |||
660 | if (&I == &Before) | |||
661 | break; | |||
662 | ||||
663 | if (!isa<StoreInst>(&I)) | |||
664 | continue; | |||
665 | ||||
666 | auto *S = cast<StoreInst>(&I); | |||
667 | int64_t Offset = -1; | |||
668 | auto *Dst = | |||
669 | GetPointerBaseWithConstantOffset(S->getPointerOperand(), Offset, DL); | |||
670 | if (Dst == &Array) { | |||
671 | int64_t Idx = Offset / PointerSize; | |||
672 | StoredValues[Idx] = getUnderlyingObject(S->getValueOperand()); | |||
673 | LastAccesses[Idx] = S; | |||
674 | } | |||
675 | } | |||
676 | ||||
677 | return isFilled(); | |||
678 | } | |||
679 | ||||
680 | /// Returns true if all values in StoredValues and | |||
681 | /// LastAccesses are not nullptrs. | |||
682 | bool isFilled() { | |||
683 | const unsigned NumValues = StoredValues.size(); | |||
684 | for (unsigned I = 0; I < NumValues; ++I) { | |||
685 | if (!StoredValues[I] || !LastAccesses[I]) | |||
686 | return false; | |||
687 | } | |||
688 | ||||
689 | return true; | |||
690 | } | |||
691 | }; | |||
692 | ||||
693 | struct OpenMPOpt { | |||
694 | ||||
695 | using OptimizationRemarkGetter = | |||
696 | function_ref<OptimizationRemarkEmitter &(Function *)>; | |||
697 | ||||
698 | OpenMPOpt(SmallVectorImpl<Function *> &SCC, CallGraphUpdater &CGUpdater, | |||
699 | OptimizationRemarkGetter OREGetter, | |||
700 | OMPInformationCache &OMPInfoCache, Attributor &A) | |||
701 | : M(*(*SCC.begin())->getParent()), SCC(SCC), CGUpdater(CGUpdater), | |||
702 | OREGetter(OREGetter), OMPInfoCache(OMPInfoCache), A(A) {} | |||
703 | ||||
704 | /// Check if any remarks are enabled for openmp-opt | |||
705 | bool remarksEnabled() { | |||
706 | auto &Ctx = M.getContext(); | |||
707 | return Ctx.getDiagHandlerPtr()->isAnyRemarkEnabled(DEBUG_TYPE"openmp-opt"); | |||
708 | } | |||
709 | ||||
710 | /// Run all OpenMP optimizations on the underlying SCC/ModuleSlice. | |||
711 | bool run(bool IsModulePass) { | |||
712 | if (SCC.empty()) | |||
713 | return false; | |||
714 | ||||
715 | bool Changed = false; | |||
716 | ||||
717 | LLVM_DEBUG(dbgs() << TAG << "Run on SCC with " << SCC.size()do { } while (false) | |||
718 | << " functions in a slice with "do { } while (false) | |||
719 | << OMPInfoCache.ModuleSlice.size() << " functions\n")do { } while (false); | |||
720 | ||||
721 | if (IsModulePass) { | |||
722 | Changed |= runAttributor(IsModulePass); | |||
723 | ||||
724 | // Recollect uses, in case Attributor deleted any. | |||
725 | OMPInfoCache.recollectUses(); | |||
726 | ||||
727 | // TODO: This should be folded into buildCustomStateMachine. | |||
728 | Changed |= rewriteDeviceCodeStateMachine(); | |||
729 | ||||
730 | if (remarksEnabled()) | |||
731 | analysisGlobalization(); | |||
732 | } else { | |||
733 | if (PrintICVValues) | |||
734 | printICVs(); | |||
735 | if (PrintOpenMPKernels) | |||
736 | printKernels(); | |||
737 | ||||
738 | Changed |= runAttributor(IsModulePass); | |||
739 | ||||
740 | // Recollect uses, in case Attributor deleted any. | |||
741 | OMPInfoCache.recollectUses(); | |||
742 | ||||
743 | Changed |= deleteParallelRegions(); | |||
744 | ||||
745 | if (HideMemoryTransferLatency) | |||
746 | Changed |= hideMemTransfersLatency(); | |||
747 | Changed |= deduplicateRuntimeCalls(); | |||
748 | if (EnableParallelRegionMerging) { | |||
749 | if (mergeParallelRegions()) { | |||
750 | deduplicateRuntimeCalls(); | |||
751 | Changed = true; | |||
752 | } | |||
753 | } | |||
754 | } | |||
755 | ||||
756 | return Changed; | |||
757 | } | |||
758 | ||||
759 | /// Print initial ICV values for testing. | |||
760 | /// FIXME: This should be done from the Attributor once it is added. | |||
761 | void printICVs() const { | |||
762 | InternalControlVar ICVs[] = {ICV_nthreads, ICV_active_levels, ICV_cancel, | |||
763 | ICV_proc_bind}; | |||
764 | ||||
765 | for (Function *F : OMPInfoCache.ModuleSlice) { | |||
766 | for (auto ICV : ICVs) { | |||
767 | auto ICVInfo = OMPInfoCache.ICVs[ICV]; | |||
768 | auto Remark = [&](OptimizationRemarkAnalysis ORA) { | |||
769 | return ORA << "OpenMP ICV " << ore::NV("OpenMPICV", ICVInfo.Name) | |||
770 | << " Value: " | |||
771 | << (ICVInfo.InitValue | |||
772 | ? toString(ICVInfo.InitValue->getValue(), 10, true) | |||
773 | : "IMPLEMENTATION_DEFINED"); | |||
774 | }; | |||
775 | ||||
776 | emitRemark<OptimizationRemarkAnalysis>(F, "OpenMPICVTracker", Remark); | |||
777 | } | |||
778 | } | |||
779 | } | |||
780 | ||||
781 | /// Print OpenMP GPU kernels for testing. | |||
782 | void printKernels() const { | |||
783 | for (Function *F : SCC) { | |||
784 | if (!OMPInfoCache.Kernels.count(F)) | |||
785 | continue; | |||
786 | ||||
787 | auto Remark = [&](OptimizationRemarkAnalysis ORA) { | |||
788 | return ORA << "OpenMP GPU kernel " | |||
789 | << ore::NV("OpenMPGPUKernel", F->getName()) << "\n"; | |||
790 | }; | |||
791 | ||||
792 | emitRemark<OptimizationRemarkAnalysis>(F, "OpenMPGPU", Remark); | |||
793 | } | |||
794 | } | |||
795 | ||||
796 | /// Return the call if \p U is a callee use in a regular call. If \p RFI is | |||
797 | /// given it has to be the callee or a nullptr is returned. | |||
798 | static CallInst *getCallIfRegularCall( | |||
799 | Use &U, OMPInformationCache::RuntimeFunctionInfo *RFI = nullptr) { | |||
800 | CallInst *CI = dyn_cast<CallInst>(U.getUser()); | |||
801 | if (CI && CI->isCallee(&U) && !CI->hasOperandBundles() && | |||
802 | (!RFI || | |||
803 | (RFI->Declaration && CI->getCalledFunction() == RFI->Declaration))) | |||
804 | return CI; | |||
805 | return nullptr; | |||
806 | } | |||
807 | ||||
808 | /// Return the call if \p V is a regular call. If \p RFI is given it has to be | |||
809 | /// the callee or a nullptr is returned. | |||
810 | static CallInst *getCallIfRegularCall( | |||
811 | Value &V, OMPInformationCache::RuntimeFunctionInfo *RFI = nullptr) { | |||
812 | CallInst *CI = dyn_cast<CallInst>(&V); | |||
813 | if (CI && !CI->hasOperandBundles() && | |||
814 | (!RFI || | |||
815 | (RFI->Declaration && CI->getCalledFunction() == RFI->Declaration))) | |||
816 | return CI; | |||
817 | return nullptr; | |||
818 | } | |||
819 | ||||
820 | private: | |||
821 | /// Merge parallel regions when it is safe. | |||
822 | bool mergeParallelRegions() { | |||
823 | const unsigned CallbackCalleeOperand = 2; | |||
824 | const unsigned CallbackFirstArgOperand = 3; | |||
825 | using InsertPointTy = OpenMPIRBuilder::InsertPointTy; | |||
826 | ||||
827 | // Check if there are any __kmpc_fork_call calls to merge. | |||
828 | OMPInformationCache::RuntimeFunctionInfo &RFI = | |||
829 | OMPInfoCache.RFIs[OMPRTL___kmpc_fork_call]; | |||
830 | ||||
831 | if (!RFI.Declaration) | |||
832 | return false; | |||
833 | ||||
834 | // Unmergable calls that prevent merging a parallel region. | |||
835 | OMPInformationCache::RuntimeFunctionInfo UnmergableCallsInfo[] = { | |||
836 | OMPInfoCache.RFIs[OMPRTL___kmpc_push_proc_bind], | |||
837 | OMPInfoCache.RFIs[OMPRTL___kmpc_push_num_threads], | |||
838 | }; | |||
839 | ||||
840 | bool Changed = false; | |||
841 | LoopInfo *LI = nullptr; | |||
842 | DominatorTree *DT = nullptr; | |||
843 | ||||
844 | SmallDenseMap<BasicBlock *, SmallPtrSet<Instruction *, 4>> BB2PRMap; | |||
845 | ||||
846 | BasicBlock *StartBB = nullptr, *EndBB = nullptr; | |||
847 | auto BodyGenCB = [&](InsertPointTy AllocaIP, InsertPointTy CodeGenIP, | |||
848 | BasicBlock &ContinuationIP) { | |||
849 | BasicBlock *CGStartBB = CodeGenIP.getBlock(); | |||
850 | BasicBlock *CGEndBB = | |||
851 | SplitBlock(CGStartBB, &*CodeGenIP.getPoint(), DT, LI); | |||
852 | assert(StartBB != nullptr && "StartBB should not be null")((void)0); | |||
853 | CGStartBB->getTerminator()->setSuccessor(0, StartBB); | |||
854 | assert(EndBB != nullptr && "EndBB should not be null")((void)0); | |||
855 | EndBB->getTerminator()->setSuccessor(0, CGEndBB); | |||
856 | }; | |||
857 | ||||
858 | auto PrivCB = [&](InsertPointTy AllocaIP, InsertPointTy CodeGenIP, Value &, | |||
859 | Value &Inner, Value *&ReplacementValue) -> InsertPointTy { | |||
860 | ReplacementValue = &Inner; | |||
861 | return CodeGenIP; | |||
862 | }; | |||
863 | ||||
864 | auto FiniCB = [&](InsertPointTy CodeGenIP) {}; | |||
865 | ||||
866 | /// Create a sequential execution region within a merged parallel region, | |||
867 | /// encapsulated in a master construct with a barrier for synchronization. | |||
868 | auto CreateSequentialRegion = [&](Function *OuterFn, | |||
869 | BasicBlock *OuterPredBB, | |||
870 | Instruction *SeqStartI, | |||
871 | Instruction *SeqEndI) { | |||
872 | // Isolate the instructions of the sequential region to a separate | |||
873 | // block. | |||
874 | BasicBlock *ParentBB = SeqStartI->getParent(); | |||
875 | BasicBlock *SeqEndBB = | |||
876 | SplitBlock(ParentBB, SeqEndI->getNextNode(), DT, LI); | |||
877 | BasicBlock *SeqAfterBB = | |||
878 | SplitBlock(SeqEndBB, &*SeqEndBB->getFirstInsertionPt(), DT, LI); | |||
879 | BasicBlock *SeqStartBB = | |||
880 | SplitBlock(ParentBB, SeqStartI, DT, LI, nullptr, "seq.par.merged"); | |||
881 | ||||
882 | assert(ParentBB->getUniqueSuccessor() == SeqStartBB &&((void)0) | |||
883 | "Expected a different CFG")((void)0); | |||
884 | const DebugLoc DL = ParentBB->getTerminator()->getDebugLoc(); | |||
885 | ParentBB->getTerminator()->eraseFromParent(); | |||
886 | ||||
887 | auto BodyGenCB = [&](InsertPointTy AllocaIP, InsertPointTy CodeGenIP, | |||
888 | BasicBlock &ContinuationIP) { | |||
889 | BasicBlock *CGStartBB = CodeGenIP.getBlock(); | |||
890 | BasicBlock *CGEndBB = | |||
891 | SplitBlock(CGStartBB, &*CodeGenIP.getPoint(), DT, LI); | |||
892 | assert(SeqStartBB != nullptr && "SeqStartBB should not be null")((void)0); | |||
893 | CGStartBB->getTerminator()->setSuccessor(0, SeqStartBB); | |||
894 | assert(SeqEndBB != nullptr && "SeqEndBB should not be null")((void)0); | |||
895 | SeqEndBB->getTerminator()->setSuccessor(0, CGEndBB); | |||
896 | }; | |||
897 | auto FiniCB = [&](InsertPointTy CodeGenIP) {}; | |||
898 | ||||
899 | // Find outputs from the sequential region to outside users and | |||
900 | // broadcast their values to them. | |||
901 | for (Instruction &I : *SeqStartBB) { | |||
902 | SmallPtrSet<Instruction *, 4> OutsideUsers; | |||
903 | for (User *Usr : I.users()) { | |||
904 | Instruction &UsrI = *cast<Instruction>(Usr); | |||
905 | // Ignore outputs to LT intrinsics, code extraction for the merged | |||
906 | // parallel region will fix them. | |||
907 | if (UsrI.isLifetimeStartOrEnd()) | |||
908 | continue; | |||
909 | ||||
910 | if (UsrI.getParent() != SeqStartBB) | |||
911 | OutsideUsers.insert(&UsrI); | |||
912 | } | |||
913 | ||||
914 | if (OutsideUsers.empty()) | |||
915 | continue; | |||
916 | ||||
917 | // Emit an alloca in the outer region to store the broadcasted | |||
918 | // value. | |||
919 | const DataLayout &DL = M.getDataLayout(); | |||
920 | AllocaInst *AllocaI = new AllocaInst( | |||
921 | I.getType(), DL.getAllocaAddrSpace(), nullptr, | |||
922 | I.getName() + ".seq.output.alloc", &OuterFn->front().front()); | |||
923 | ||||
924 | // Emit a store instruction in the sequential BB to update the | |||
925 | // value. | |||
926 | new StoreInst(&I, AllocaI, SeqStartBB->getTerminator()); | |||
927 | ||||
928 | // Emit a load instruction and replace the use of the output value | |||
929 | // with it. | |||
930 | for (Instruction *UsrI : OutsideUsers) { | |||
931 | LoadInst *LoadI = new LoadInst( | |||
932 | I.getType(), AllocaI, I.getName() + ".seq.output.load", UsrI); | |||
933 | UsrI->replaceUsesOfWith(&I, LoadI); | |||
934 | } | |||
935 | } | |||
936 | ||||
937 | OpenMPIRBuilder::LocationDescription Loc( | |||
938 | InsertPointTy(ParentBB, ParentBB->end()), DL); | |||
939 | InsertPointTy SeqAfterIP = | |||
940 | OMPInfoCache.OMPBuilder.createMaster(Loc, BodyGenCB, FiniCB); | |||
941 | ||||
942 | OMPInfoCache.OMPBuilder.createBarrier(SeqAfterIP, OMPD_parallel); | |||
943 | ||||
944 | BranchInst::Create(SeqAfterBB, SeqAfterIP.getBlock()); | |||
945 | ||||
946 | LLVM_DEBUG(dbgs() << TAG << "After sequential inlining " << *OuterFndo { } while (false) | |||
947 | << "\n")do { } while (false); | |||
948 | }; | |||
949 | ||||
950 | // Helper to merge the __kmpc_fork_call calls in MergableCIs. They are all | |||
951 | // contained in BB and only separated by instructions that can be | |||
952 | // redundantly executed in parallel. The block BB is split before the first | |||
953 | // call (in MergableCIs) and after the last so the entire region we merge | |||
954 | // into a single parallel region is contained in a single basic block | |||
955 | // without any other instructions. We use the OpenMPIRBuilder to outline | |||
956 | // that block and call the resulting function via __kmpc_fork_call. | |||
957 | auto Merge = [&](SmallVectorImpl<CallInst *> &MergableCIs, BasicBlock *BB) { | |||
958 | // TODO: Change the interface to allow single CIs expanded, e.g, to | |||
959 | // include an outer loop. | |||
960 | assert(MergableCIs.size() > 1 && "Assumed multiple mergable CIs")((void)0); | |||
961 | ||||
962 | auto Remark = [&](OptimizationRemark OR) { | |||
963 | OR << "Parallel region merged with parallel region" | |||
964 | << (MergableCIs.size() > 2 ? "s" : "") << " at "; | |||
965 | for (auto *CI : llvm::drop_begin(MergableCIs)) { | |||
966 | OR << ore::NV("OpenMPParallelMerge", CI->getDebugLoc()); | |||
967 | if (CI != MergableCIs.back()) | |||
968 | OR << ", "; | |||
969 | } | |||
970 | return OR << "."; | |||
971 | }; | |||
972 | ||||
973 | emitRemark<OptimizationRemark>(MergableCIs.front(), "OMP150", Remark); | |||
974 | ||||
975 | Function *OriginalFn = BB->getParent(); | |||
976 | LLVM_DEBUG(dbgs() << TAG << "Merge " << MergableCIs.size()do { } while (false) | |||
977 | << " parallel regions in " << OriginalFn->getName()do { } while (false) | |||
978 | << "\n")do { } while (false); | |||
979 | ||||
980 | // Isolate the calls to merge in a separate block. | |||
981 | EndBB = SplitBlock(BB, MergableCIs.back()->getNextNode(), DT, LI); | |||
982 | BasicBlock *AfterBB = | |||
983 | SplitBlock(EndBB, &*EndBB->getFirstInsertionPt(), DT, LI); | |||
984 | StartBB = SplitBlock(BB, MergableCIs.front(), DT, LI, nullptr, | |||
985 | "omp.par.merged"); | |||
986 | ||||
987 | assert(BB->getUniqueSuccessor() == StartBB && "Expected a different CFG")((void)0); | |||
988 | const DebugLoc DL = BB->getTerminator()->getDebugLoc(); | |||
989 | BB->getTerminator()->eraseFromParent(); | |||
990 | ||||
991 | // Create sequential regions for sequential instructions that are | |||
992 | // in-between mergable parallel regions. | |||
993 | for (auto *It = MergableCIs.begin(), *End = MergableCIs.end() - 1; | |||
994 | It != End; ++It) { | |||
995 | Instruction *ForkCI = *It; | |||
996 | Instruction *NextForkCI = *(It + 1); | |||
997 | ||||
998 | // Continue if there are not in-between instructions. | |||
999 | if (ForkCI->getNextNode() == NextForkCI) | |||
1000 | continue; | |||
1001 | ||||
1002 | CreateSequentialRegion(OriginalFn, BB, ForkCI->getNextNode(), | |||
1003 | NextForkCI->getPrevNode()); | |||
1004 | } | |||
1005 | ||||
1006 | OpenMPIRBuilder::LocationDescription Loc(InsertPointTy(BB, BB->end()), | |||
1007 | DL); | |||
1008 | IRBuilder<>::InsertPoint AllocaIP( | |||
1009 | &OriginalFn->getEntryBlock(), | |||
1010 | OriginalFn->getEntryBlock().getFirstInsertionPt()); | |||
1011 | // Create the merged parallel region with default proc binding, to | |||
1012 | // avoid overriding binding settings, and without explicit cancellation. | |||
1013 | InsertPointTy AfterIP = OMPInfoCache.OMPBuilder.createParallel( | |||
1014 | Loc, AllocaIP, BodyGenCB, PrivCB, FiniCB, nullptr, nullptr, | |||
1015 | OMP_PROC_BIND_default, /* IsCancellable */ false); | |||
1016 | BranchInst::Create(AfterBB, AfterIP.getBlock()); | |||
1017 | ||||
1018 | // Perform the actual outlining. | |||
1019 | OMPInfoCache.OMPBuilder.finalize(OriginalFn, | |||
1020 | /* AllowExtractorSinking */ true); | |||
1021 | ||||
1022 | Function *OutlinedFn = MergableCIs.front()->getCaller(); | |||
1023 | ||||
1024 | // Replace the __kmpc_fork_call calls with direct calls to the outlined | |||
1025 | // callbacks. | |||
1026 | SmallVector<Value *, 8> Args; | |||
1027 | for (auto *CI : MergableCIs) { | |||
1028 | Value *Callee = | |||
1029 | CI->getArgOperand(CallbackCalleeOperand)->stripPointerCasts(); | |||
1030 | FunctionType *FT = | |||
1031 | cast<FunctionType>(Callee->getType()->getPointerElementType()); | |||
1032 | Args.clear(); | |||
1033 | Args.push_back(OutlinedFn->getArg(0)); | |||
1034 | Args.push_back(OutlinedFn->getArg(1)); | |||
1035 | for (unsigned U = CallbackFirstArgOperand, E = CI->getNumArgOperands(); | |||
1036 | U < E; ++U) | |||
1037 | Args.push_back(CI->getArgOperand(U)); | |||
1038 | ||||
1039 | CallInst *NewCI = CallInst::Create(FT, Callee, Args, "", CI); | |||
1040 | if (CI->getDebugLoc()) | |||
1041 | NewCI->setDebugLoc(CI->getDebugLoc()); | |||
1042 | ||||
1043 | // Forward parameter attributes from the callback to the callee. | |||
1044 | for (unsigned U = CallbackFirstArgOperand, E = CI->getNumArgOperands(); | |||
1045 | U < E; ++U) | |||
1046 | for (const Attribute &A : CI->getAttributes().getParamAttributes(U)) | |||
1047 | NewCI->addParamAttr( | |||
1048 | U - (CallbackFirstArgOperand - CallbackCalleeOperand), A); | |||
1049 | ||||
1050 | // Emit an explicit barrier to replace the implicit fork-join barrier. | |||
1051 | if (CI != MergableCIs.back()) { | |||
1052 | // TODO: Remove barrier if the merged parallel region includes the | |||
1053 | // 'nowait' clause. | |||
1054 | OMPInfoCache.OMPBuilder.createBarrier( | |||
1055 | InsertPointTy(NewCI->getParent(), | |||
1056 | NewCI->getNextNode()->getIterator()), | |||
1057 | OMPD_parallel); | |||
1058 | } | |||
1059 | ||||
1060 | CI->eraseFromParent(); | |||
1061 | } | |||
1062 | ||||
1063 | assert(OutlinedFn != OriginalFn && "Outlining failed")((void)0); | |||
1064 | CGUpdater.registerOutlinedFunction(*OriginalFn, *OutlinedFn); | |||
1065 | CGUpdater.reanalyzeFunction(*OriginalFn); | |||
1066 | ||||
1067 | NumOpenMPParallelRegionsMerged += MergableCIs.size(); | |||
1068 | ||||
1069 | return true; | |||
1070 | }; | |||
1071 | ||||
1072 | // Helper function that identifes sequences of | |||
1073 | // __kmpc_fork_call uses in a basic block. | |||
1074 | auto DetectPRsCB = [&](Use &U, Function &F) { | |||
1075 | CallInst *CI = getCallIfRegularCall(U, &RFI); | |||
1076 | BB2PRMap[CI->getParent()].insert(CI); | |||
1077 | ||||
1078 | return false; | |||
1079 | }; | |||
1080 | ||||
1081 | BB2PRMap.clear(); | |||
1082 | RFI.foreachUse(SCC, DetectPRsCB); | |||
1083 | SmallVector<SmallVector<CallInst *, 4>, 4> MergableCIsVector; | |||
1084 | // Find mergable parallel regions within a basic block that are | |||
1085 | // safe to merge, that is any in-between instructions can safely | |||
1086 | // execute in parallel after merging. | |||
1087 | // TODO: support merging across basic-blocks. | |||
1088 | for (auto &It : BB2PRMap) { | |||
1089 | auto &CIs = It.getSecond(); | |||
1090 | if (CIs.size() < 2) | |||
1091 | continue; | |||
1092 | ||||
1093 | BasicBlock *BB = It.getFirst(); | |||
1094 | SmallVector<CallInst *, 4> MergableCIs; | |||
1095 | ||||
1096 | /// Returns true if the instruction is mergable, false otherwise. | |||
1097 | /// A terminator instruction is unmergable by definition since merging | |||
1098 | /// works within a BB. Instructions before the mergable region are | |||
1099 | /// mergable if they are not calls to OpenMP runtime functions that may | |||
1100 | /// set different execution parameters for subsequent parallel regions. | |||
1101 | /// Instructions in-between parallel regions are mergable if they are not | |||
1102 | /// calls to any non-intrinsic function since that may call a non-mergable | |||
1103 | /// OpenMP runtime function. | |||
1104 | auto IsMergable = [&](Instruction &I, bool IsBeforeMergableRegion) { | |||
1105 | // We do not merge across BBs, hence return false (unmergable) if the | |||
1106 | // instruction is a terminator. | |||
1107 | if (I.isTerminator()) | |||
1108 | return false; | |||
1109 | ||||
1110 | if (!isa<CallInst>(&I)) | |||
1111 | return true; | |||
1112 | ||||
1113 | CallInst *CI = cast<CallInst>(&I); | |||
1114 | if (IsBeforeMergableRegion) { | |||
1115 | Function *CalledFunction = CI->getCalledFunction(); | |||
1116 | if (!CalledFunction) | |||
1117 | return false; | |||
1118 | // Return false (unmergable) if the call before the parallel | |||
1119 | // region calls an explicit affinity (proc_bind) or number of | |||
1120 | // threads (num_threads) compiler-generated function. Those settings | |||
1121 | // may be incompatible with following parallel regions. | |||
1122 | // TODO: ICV tracking to detect compatibility. | |||
1123 | for (const auto &RFI : UnmergableCallsInfo) { | |||
1124 | if (CalledFunction == RFI.Declaration) | |||
1125 | return false; | |||
1126 | } | |||
1127 | } else { | |||
1128 | // Return false (unmergable) if there is a call instruction | |||
1129 | // in-between parallel regions when it is not an intrinsic. It | |||
1130 | // may call an unmergable OpenMP runtime function in its callpath. | |||
1131 | // TODO: Keep track of possible OpenMP calls in the callpath. | |||
1132 | if (!isa<IntrinsicInst>(CI)) | |||
1133 | return false; | |||
1134 | } | |||
1135 | ||||
1136 | return true; | |||
1137 | }; | |||
1138 | // Find maximal number of parallel region CIs that are safe to merge. | |||
1139 | for (auto It = BB->begin(), End = BB->end(); It != End;) { | |||
1140 | Instruction &I = *It; | |||
1141 | ++It; | |||
1142 | ||||
1143 | if (CIs.count(&I)) { | |||
1144 | MergableCIs.push_back(cast<CallInst>(&I)); | |||
1145 | continue; | |||
1146 | } | |||
1147 | ||||
1148 | // Continue expanding if the instruction is mergable. | |||
1149 | if (IsMergable(I, MergableCIs.empty())) | |||
1150 | continue; | |||
1151 | ||||
1152 | // Forward the instruction iterator to skip the next parallel region | |||
1153 | // since there is an unmergable instruction which can affect it. | |||
1154 | for (; It != End; ++It) { | |||
1155 | Instruction &SkipI = *It; | |||
1156 | if (CIs.count(&SkipI)) { | |||
1157 | LLVM_DEBUG(dbgs() << TAG << "Skip parallel region " << SkipIdo { } while (false) | |||
1158 | << " due to " << I << "\n")do { } while (false); | |||
1159 | ++It; | |||
1160 | break; | |||
1161 | } | |||
1162 | } | |||
1163 | ||||
1164 | // Store mergable regions found. | |||
1165 | if (MergableCIs.size() > 1) { | |||
1166 | MergableCIsVector.push_back(MergableCIs); | |||
1167 | LLVM_DEBUG(dbgs() << TAG << "Found " << MergableCIs.size()do { } while (false) | |||
1168 | << " parallel regions in block " << BB->getName()do { } while (false) | |||
1169 | << " of function " << BB->getParent()->getName()do { } while (false) | |||
1170 | << "\n";)do { } while (false); | |||
1171 | } | |||
1172 | ||||
1173 | MergableCIs.clear(); | |||
1174 | } | |||
1175 | ||||
1176 | if (!MergableCIsVector.empty()) { | |||
1177 | Changed = true; | |||
1178 | ||||
1179 | for (auto &MergableCIs : MergableCIsVector) | |||
1180 | Merge(MergableCIs, BB); | |||
1181 | MergableCIsVector.clear(); | |||
1182 | } | |||
1183 | } | |||
1184 | ||||
1185 | if (Changed) { | |||
1186 | /// Re-collect use for fork calls, emitted barrier calls, and | |||
1187 | /// any emitted master/end_master calls. | |||
1188 | OMPInfoCache.recollectUsesForFunction(OMPRTL___kmpc_fork_call); | |||
1189 | OMPInfoCache.recollectUsesForFunction(OMPRTL___kmpc_barrier); | |||
1190 | OMPInfoCache.recollectUsesForFunction(OMPRTL___kmpc_master); | |||
1191 | OMPInfoCache.recollectUsesForFunction(OMPRTL___kmpc_end_master); | |||
1192 | } | |||
1193 | ||||
1194 | return Changed; | |||
1195 | } | |||
1196 | ||||
1197 | /// Try to delete parallel regions if possible. | |||
1198 | bool deleteParallelRegions() { | |||
1199 | const unsigned CallbackCalleeOperand = 2; | |||
1200 | ||||
1201 | OMPInformationCache::RuntimeFunctionInfo &RFI = | |||
1202 | OMPInfoCache.RFIs[OMPRTL___kmpc_fork_call]; | |||
1203 | ||||
1204 | if (!RFI.Declaration) | |||
1205 | return false; | |||
1206 | ||||
1207 | bool Changed = false; | |||
1208 | auto DeleteCallCB = [&](Use &U, Function &) { | |||
1209 | CallInst *CI = getCallIfRegularCall(U); | |||
1210 | if (!CI) | |||
1211 | return false; | |||
1212 | auto *Fn = dyn_cast<Function>( | |||
1213 | CI->getArgOperand(CallbackCalleeOperand)->stripPointerCasts()); | |||
1214 | if (!Fn) | |||
1215 | return false; | |||
1216 | if (!Fn->onlyReadsMemory()) | |||
1217 | return false; | |||
1218 | if (!Fn->hasFnAttribute(Attribute::WillReturn)) | |||
1219 | return false; | |||
1220 | ||||
1221 | LLVM_DEBUG(dbgs() << TAG << "Delete read-only parallel region in "do { } while (false) | |||
1222 | << CI->getCaller()->getName() << "\n")do { } while (false); | |||
1223 | ||||
1224 | auto Remark = [&](OptimizationRemark OR) { | |||
1225 | return OR << "Removing parallel region with no side-effects."; | |||
1226 | }; | |||
1227 | emitRemark<OptimizationRemark>(CI, "OMP160", Remark); | |||
1228 | ||||
1229 | CGUpdater.removeCallSite(*CI); | |||
1230 | CI->eraseFromParent(); | |||
1231 | Changed = true; | |||
1232 | ++NumOpenMPParallelRegionsDeleted; | |||
1233 | return true; | |||
1234 | }; | |||
1235 | ||||
1236 | RFI.foreachUse(SCC, DeleteCallCB); | |||
1237 | ||||
1238 | return Changed; | |||
1239 | } | |||
1240 | ||||
1241 | /// Try to eliminate runtime calls by reusing existing ones. | |||
1242 | bool deduplicateRuntimeCalls() { | |||
1243 | bool Changed = false; | |||
1244 | ||||
1245 | RuntimeFunction DeduplicableRuntimeCallIDs[] = { | |||
1246 | OMPRTL_omp_get_num_threads, | |||
1247 | OMPRTL_omp_in_parallel, | |||
1248 | OMPRTL_omp_get_cancellation, | |||
1249 | OMPRTL_omp_get_thread_limit, | |||
1250 | OMPRTL_omp_get_supported_active_levels, | |||
1251 | OMPRTL_omp_get_level, | |||
1252 | OMPRTL_omp_get_ancestor_thread_num, | |||
1253 | OMPRTL_omp_get_team_size, | |||
1254 | OMPRTL_omp_get_active_level, | |||
1255 | OMPRTL_omp_in_final, | |||
1256 | OMPRTL_omp_get_proc_bind, | |||
1257 | OMPRTL_omp_get_num_places, | |||
1258 | OMPRTL_omp_get_num_procs, | |||
1259 | OMPRTL_omp_get_place_num, | |||
1260 | OMPRTL_omp_get_partition_num_places, | |||
1261 | OMPRTL_omp_get_partition_place_nums}; | |||
1262 | ||||
1263 | // Global-tid is handled separately. | |||
1264 | SmallSetVector<Value *, 16> GTIdArgs; | |||
1265 | collectGlobalThreadIdArguments(GTIdArgs); | |||
1266 | LLVM_DEBUG(dbgs() << TAG << "Found " << GTIdArgs.size()do { } while (false) | |||
1267 | << " global thread ID arguments\n")do { } while (false); | |||
1268 | ||||
1269 | for (Function *F : SCC) { | |||
1270 | for (auto DeduplicableRuntimeCallID : DeduplicableRuntimeCallIDs) | |||
1271 | Changed |= deduplicateRuntimeCalls( | |||
1272 | *F, OMPInfoCache.RFIs[DeduplicableRuntimeCallID]); | |||
1273 | ||||
1274 | // __kmpc_global_thread_num is special as we can replace it with an | |||
1275 | // argument in enough cases to make it worth trying. | |||
1276 | Value *GTIdArg = nullptr; | |||
1277 | for (Argument &Arg : F->args()) | |||
1278 | if (GTIdArgs.count(&Arg)) { | |||
1279 | GTIdArg = &Arg; | |||
1280 | break; | |||
1281 | } | |||
1282 | Changed |= deduplicateRuntimeCalls( | |||
1283 | *F, OMPInfoCache.RFIs[OMPRTL___kmpc_global_thread_num], GTIdArg); | |||
1284 | } | |||
1285 | ||||
1286 | return Changed; | |||
1287 | } | |||
1288 | ||||
1289 | /// Tries to hide the latency of runtime calls that involve host to | |||
1290 | /// device memory transfers by splitting them into their "issue" and "wait" | |||
1291 | /// versions. The "issue" is moved upwards as much as possible. The "wait" is | |||
1292 | /// moved downards as much as possible. The "issue" issues the memory transfer | |||
1293 | /// asynchronously, returning a handle. The "wait" waits in the returned | |||
1294 | /// handle for the memory transfer to finish. | |||
1295 | bool hideMemTransfersLatency() { | |||
1296 | auto &RFI = OMPInfoCache.RFIs[OMPRTL___tgt_target_data_begin_mapper]; | |||
1297 | bool Changed = false; | |||
1298 | auto SplitMemTransfers = [&](Use &U, Function &Decl) { | |||
1299 | auto *RTCall = getCallIfRegularCall(U, &RFI); | |||
1300 | if (!RTCall) | |||
1301 | return false; | |||
1302 | ||||
1303 | OffloadArray OffloadArrays[3]; | |||
1304 | if (!getValuesInOffloadArrays(*RTCall, OffloadArrays)) | |||
1305 | return false; | |||
1306 | ||||
1307 | LLVM_DEBUG(dumpValuesInOffloadArrays(OffloadArrays))do { } while (false); | |||
1308 | ||||
1309 | // TODO: Check if can be moved upwards. | |||
1310 | bool WasSplit = false; | |||
1311 | Instruction *WaitMovementPoint = canBeMovedDownwards(*RTCall); | |||
1312 | if (WaitMovementPoint) | |||
1313 | WasSplit = splitTargetDataBeginRTC(*RTCall, *WaitMovementPoint); | |||
1314 | ||||
1315 | Changed |= WasSplit; | |||
1316 | return WasSplit; | |||
1317 | }; | |||
1318 | RFI.foreachUse(SCC, SplitMemTransfers); | |||
1319 | ||||
1320 | return Changed; | |||
1321 | } | |||
1322 | ||||
1323 | void analysisGlobalization() { | |||
1324 | auto &RFI = OMPInfoCache.RFIs[OMPRTL___kmpc_alloc_shared]; | |||
1325 | ||||
1326 | auto CheckGlobalization = [&](Use &U, Function &Decl) { | |||
1327 | if (CallInst *CI = getCallIfRegularCall(U, &RFI)) { | |||
1328 | auto Remark = [&](OptimizationRemarkMissed ORM) { | |||
1329 | return ORM | |||
1330 | << "Found thread data sharing on the GPU. " | |||
1331 | << "Expect degraded performance due to data globalization."; | |||
1332 | }; | |||
1333 | emitRemark<OptimizationRemarkMissed>(CI, "OMP112", Remark); | |||
1334 | } | |||
1335 | ||||
1336 | return false; | |||
1337 | }; | |||
1338 | ||||
1339 | RFI.foreachUse(SCC, CheckGlobalization); | |||
1340 | } | |||
1341 | ||||
1342 | /// Maps the values stored in the offload arrays passed as arguments to | |||
1343 | /// \p RuntimeCall into the offload arrays in \p OAs. | |||
1344 | bool getValuesInOffloadArrays(CallInst &RuntimeCall, | |||
1345 | MutableArrayRef<OffloadArray> OAs) { | |||
1346 | assert(OAs.size() == 3 && "Need space for three offload arrays!")((void)0); | |||
1347 | ||||
1348 | // A runtime call that involves memory offloading looks something like: | |||
1349 | // call void @__tgt_target_data_begin_mapper(arg0, arg1, | |||
1350 | // i8** %offload_baseptrs, i8** %offload_ptrs, i64* %offload_sizes, | |||
1351 | // ...) | |||
1352 | // So, the idea is to access the allocas that allocate space for these | |||
1353 | // offload arrays, offload_baseptrs, offload_ptrs, offload_sizes. | |||
1354 | // Therefore: | |||
1355 | // i8** %offload_baseptrs. | |||
1356 | Value *BasePtrsArg = | |||
1357 | RuntimeCall.getArgOperand(OffloadArray::BasePtrsArgNum); | |||
1358 | // i8** %offload_ptrs. | |||
1359 | Value *PtrsArg = RuntimeCall.getArgOperand(OffloadArray::PtrsArgNum); | |||
1360 | // i8** %offload_sizes. | |||
1361 | Value *SizesArg = RuntimeCall.getArgOperand(OffloadArray::SizesArgNum); | |||
1362 | ||||
1363 | // Get values stored in **offload_baseptrs. | |||
1364 | auto *V = getUnderlyingObject(BasePtrsArg); | |||
1365 | if (!isa<AllocaInst>(V)) | |||
1366 | return false; | |||
1367 | auto *BasePtrsArray = cast<AllocaInst>(V); | |||
1368 | if (!OAs[0].initialize(*BasePtrsArray, RuntimeCall)) | |||
1369 | return false; | |||
1370 | ||||
1371 | // Get values stored in **offload_baseptrs. | |||
1372 | V = getUnderlyingObject(PtrsArg); | |||
1373 | if (!isa<AllocaInst>(V)) | |||
1374 | return false; | |||
1375 | auto *PtrsArray = cast<AllocaInst>(V); | |||
1376 | if (!OAs[1].initialize(*PtrsArray, RuntimeCall)) | |||
1377 | return false; | |||
1378 | ||||
1379 | // Get values stored in **offload_sizes. | |||
1380 | V = getUnderlyingObject(SizesArg); | |||
1381 | // If it's a [constant] global array don't analyze it. | |||
1382 | if (isa<GlobalValue>(V)) | |||
1383 | return isa<Constant>(V); | |||
1384 | if (!isa<AllocaInst>(V)) | |||
1385 | return false; | |||
1386 | ||||
1387 | auto *SizesArray = cast<AllocaInst>(V); | |||
1388 | if (!OAs[2].initialize(*SizesArray, RuntimeCall)) | |||
1389 | return false; | |||
1390 | ||||
1391 | return true; | |||
1392 | } | |||
1393 | ||||
1394 | /// Prints the values in the OffloadArrays \p OAs using LLVM_DEBUG. | |||
1395 | /// For now this is a way to test that the function getValuesInOffloadArrays | |||
1396 | /// is working properly. | |||
1397 | /// TODO: Move this to a unittest when unittests are available for OpenMPOpt. | |||
1398 | void dumpValuesInOffloadArrays(ArrayRef<OffloadArray> OAs) { | |||
1399 | assert(OAs.size() == 3 && "There are three offload arrays to debug!")((void)0); | |||
1400 | ||||
1401 | LLVM_DEBUG(dbgs() << TAG << " Successfully got offload values:\n")do { } while (false); | |||
1402 | std::string ValuesStr; | |||
1403 | raw_string_ostream Printer(ValuesStr); | |||
1404 | std::string Separator = " --- "; | |||
1405 | ||||
1406 | for (auto *BP : OAs[0].StoredValues) { | |||
1407 | BP->print(Printer); | |||
1408 | Printer << Separator; | |||
1409 | } | |||
1410 | LLVM_DEBUG(dbgs() << "\t\toffload_baseptrs: " << Printer.str() << "\n")do { } while (false); | |||
1411 | ValuesStr.clear(); | |||
1412 | ||||
1413 | for (auto *P : OAs[1].StoredValues) { | |||
1414 | P->print(Printer); | |||
1415 | Printer << Separator; | |||
1416 | } | |||
1417 | LLVM_DEBUG(dbgs() << "\t\toffload_ptrs: " << Printer.str() << "\n")do { } while (false); | |||
1418 | ValuesStr.clear(); | |||
1419 | ||||
1420 | for (auto *S : OAs[2].StoredValues) { | |||
1421 | S->print(Printer); | |||
1422 | Printer << Separator; | |||
1423 | } | |||
1424 | LLVM_DEBUG(dbgs() << "\t\toffload_sizes: " << Printer.str() << "\n")do { } while (false); | |||
1425 | } | |||
1426 | ||||
1427 | /// Returns the instruction where the "wait" counterpart \p RuntimeCall can be | |||
1428 | /// moved. Returns nullptr if the movement is not possible, or not worth it. | |||
1429 | Instruction *canBeMovedDownwards(CallInst &RuntimeCall) { | |||
1430 | // FIXME: This traverses only the BasicBlock where RuntimeCall is. | |||
1431 | // Make it traverse the CFG. | |||
1432 | ||||
1433 | Instruction *CurrentI = &RuntimeCall; | |||
1434 | bool IsWorthIt = false; | |||
1435 | while ((CurrentI = CurrentI->getNextNode())) { | |||
1436 | ||||
1437 | // TODO: Once we detect the regions to be offloaded we should use the | |||
1438 | // alias analysis manager to check if CurrentI may modify one of | |||
1439 | // the offloaded regions. | |||
1440 | if (CurrentI->mayHaveSideEffects() || CurrentI->mayReadFromMemory()) { | |||
1441 | if (IsWorthIt) | |||
1442 | return CurrentI; | |||
1443 | ||||
1444 | return nullptr; | |||
1445 | } | |||
1446 | ||||
1447 | // FIXME: For now if we move it over anything without side effect | |||
1448 | // is worth it. | |||
1449 | IsWorthIt = true; | |||
1450 | } | |||
1451 | ||||
1452 | // Return end of BasicBlock. | |||
1453 | return RuntimeCall.getParent()->getTerminator(); | |||
1454 | } | |||
1455 | ||||
1456 | /// Splits \p RuntimeCall into its "issue" and "wait" counterparts. | |||
1457 | bool splitTargetDataBeginRTC(CallInst &RuntimeCall, | |||
1458 | Instruction &WaitMovementPoint) { | |||
1459 | // Create stack allocated handle (__tgt_async_info) at the beginning of the | |||
1460 | // function. Used for storing information of the async transfer, allowing to | |||
1461 | // wait on it later. | |||
1462 | auto &IRBuilder = OMPInfoCache.OMPBuilder; | |||
1463 | auto *F = RuntimeCall.getCaller(); | |||
1464 | Instruction *FirstInst = &(F->getEntryBlock().front()); | |||
1465 | AllocaInst *Handle = new AllocaInst( | |||
1466 | IRBuilder.AsyncInfo, F->getAddressSpace(), "handle", FirstInst); | |||
1467 | ||||
1468 | // Add "issue" runtime call declaration: | |||
1469 | // declare %struct.tgt_async_info @__tgt_target_data_begin_issue(i64, i32, | |||
1470 | // i8**, i8**, i64*, i64*) | |||
1471 | FunctionCallee IssueDecl = IRBuilder.getOrCreateRuntimeFunction( | |||
1472 | M, OMPRTL___tgt_target_data_begin_mapper_issue); | |||
1473 | ||||
1474 | // Change RuntimeCall call site for its asynchronous version. | |||
1475 | SmallVector<Value *, 16> Args; | |||
1476 | for (auto &Arg : RuntimeCall.args()) | |||
1477 | Args.push_back(Arg.get()); | |||
1478 | Args.push_back(Handle); | |||
1479 | ||||
1480 | CallInst *IssueCallsite = | |||
1481 | CallInst::Create(IssueDecl, Args, /*NameStr=*/"", &RuntimeCall); | |||
1482 | RuntimeCall.eraseFromParent(); | |||
1483 | ||||
1484 | // Add "wait" runtime call declaration: | |||
1485 | // declare void @__tgt_target_data_begin_wait(i64, %struct.__tgt_async_info) | |||
1486 | FunctionCallee WaitDecl = IRBuilder.getOrCreateRuntimeFunction( | |||
1487 | M, OMPRTL___tgt_target_data_begin_mapper_wait); | |||
1488 | ||||
1489 | Value *WaitParams[2] = { | |||
1490 | IssueCallsite->getArgOperand( | |||
1491 | OffloadArray::DeviceIDArgNum), // device_id. | |||
1492 | Handle // handle to wait on. | |||
1493 | }; | |||
1494 | CallInst::Create(WaitDecl, WaitParams, /*NameStr=*/"", &WaitMovementPoint); | |||
1495 | ||||
1496 | return true; | |||
1497 | } | |||
1498 | ||||
1499 | static Value *combinedIdentStruct(Value *CurrentIdent, Value *NextIdent, | |||
1500 | bool GlobalOnly, bool &SingleChoice) { | |||
1501 | if (CurrentIdent == NextIdent) | |||
1502 | return CurrentIdent; | |||
1503 | ||||
1504 | // TODO: Figure out how to actually combine multiple debug locations. For | |||
1505 | // now we just keep an existing one if there is a single choice. | |||
1506 | if (!GlobalOnly || isa<GlobalValue>(NextIdent)) { | |||
1507 | SingleChoice = !CurrentIdent; | |||
1508 | return NextIdent; | |||
1509 | } | |||
1510 | return nullptr; | |||
1511 | } | |||
1512 | ||||
1513 | /// Return an `struct ident_t*` value that represents the ones used in the | |||
1514 | /// calls of \p RFI inside of \p F. If \p GlobalOnly is true, we will not | |||
1515 | /// return a local `struct ident_t*`. For now, if we cannot find a suitable | |||
1516 | /// return value we create one from scratch. We also do not yet combine | |||
1517 | /// information, e.g., the source locations, see combinedIdentStruct. | |||
1518 | Value * | |||
1519 | getCombinedIdentFromCallUsesIn(OMPInformationCache::RuntimeFunctionInfo &RFI, | |||
1520 | Function &F, bool GlobalOnly) { | |||
1521 | bool SingleChoice = true; | |||
1522 | Value *Ident = nullptr; | |||
1523 | auto CombineIdentStruct = [&](Use &U, Function &Caller) { | |||
1524 | CallInst *CI = getCallIfRegularCall(U, &RFI); | |||
1525 | if (!CI || &F != &Caller) | |||
1526 | return false; | |||
1527 | Ident = combinedIdentStruct(Ident, CI->getArgOperand(0), | |||
1528 | /* GlobalOnly */ true, SingleChoice); | |||
1529 | return false; | |||
1530 | }; | |||
1531 | RFI.foreachUse(SCC, CombineIdentStruct); | |||
1532 | ||||
1533 | if (!Ident || !SingleChoice) { | |||
1534 | // The IRBuilder uses the insertion block to get to the module, this is | |||
1535 | // unfortunate but we work around it for now. | |||
1536 | if (!OMPInfoCache.OMPBuilder.getInsertionPoint().getBlock()) | |||
1537 | OMPInfoCache.OMPBuilder.updateToLocation(OpenMPIRBuilder::InsertPointTy( | |||
1538 | &F.getEntryBlock(), F.getEntryBlock().begin())); | |||
1539 | // Create a fallback location if non was found. | |||
1540 | // TODO: Use the debug locations of the calls instead. | |||
1541 | Constant *Loc = OMPInfoCache.OMPBuilder.getOrCreateDefaultSrcLocStr(); | |||
1542 | Ident = OMPInfoCache.OMPBuilder.getOrCreateIdent(Loc); | |||
1543 | } | |||
1544 | return Ident; | |||
1545 | } | |||
1546 | ||||
1547 | /// Try to eliminate calls of \p RFI in \p F by reusing an existing one or | |||
1548 | /// \p ReplVal if given. | |||
1549 | bool deduplicateRuntimeCalls(Function &F, | |||
1550 | OMPInformationCache::RuntimeFunctionInfo &RFI, | |||
1551 | Value *ReplVal = nullptr) { | |||
1552 | auto *UV = RFI.getUseVector(F); | |||
1553 | if (!UV || UV->size() + (ReplVal != nullptr) < 2) | |||
1554 | return false; | |||
1555 | ||||
1556 | LLVM_DEBUG(do { } while (false) | |||
1557 | dbgs() << TAG << "Deduplicate " << UV->size() << " uses of " << RFI.Namedo { } while (false) | |||
1558 | << (ReplVal ? " with an existing value\n" : "\n") << "\n")do { } while (false); | |||
1559 | ||||
1560 | assert((!ReplVal || (isa<Argument>(ReplVal) &&((void)0) | |||
1561 | cast<Argument>(ReplVal)->getParent() == &F)) &&((void)0) | |||
1562 | "Unexpected replacement value!")((void)0); | |||
1563 | ||||
1564 | // TODO: Use dominance to find a good position instead. | |||
1565 | auto CanBeMoved = [this](CallBase &CB) { | |||
1566 | unsigned NumArgs = CB.getNumArgOperands(); | |||
1567 | if (NumArgs == 0) | |||
1568 | return true; | |||
1569 | if (CB.getArgOperand(0)->getType() != OMPInfoCache.OMPBuilder.IdentPtr) | |||
1570 | return false; | |||
1571 | for (unsigned u = 1; u < NumArgs; ++u) | |||
1572 | if (isa<Instruction>(CB.getArgOperand(u))) | |||
1573 | return false; | |||
1574 | return true; | |||
1575 | }; | |||
1576 | ||||
1577 | if (!ReplVal) { | |||
1578 | for (Use *U : *UV) | |||
1579 | if (CallInst *CI = getCallIfRegularCall(*U, &RFI)) { | |||
1580 | if (!CanBeMoved(*CI)) | |||
1581 | continue; | |||
1582 | ||||
1583 | // If the function is a kernel, dedup will move | |||
1584 | // the runtime call right after the kernel init callsite. Otherwise, | |||
1585 | // it will move it to the beginning of the caller function. | |||
1586 | if (isKernel(F)) { | |||
1587 | auto &KernelInitRFI = OMPInfoCache.RFIs[OMPRTL___kmpc_target_init]; | |||
1588 | auto *KernelInitUV = KernelInitRFI.getUseVector(F); | |||
1589 | ||||
1590 | if (KernelInitUV->empty()) | |||
1591 | continue; | |||
1592 | ||||
1593 | assert(KernelInitUV->size() == 1 &&((void)0) | |||
1594 | "Expected a single __kmpc_target_init in kernel\n")((void)0); | |||
1595 | ||||
1596 | CallInst *KernelInitCI = | |||
1597 | getCallIfRegularCall(*KernelInitUV->front(), &KernelInitRFI); | |||
1598 | assert(KernelInitCI &&((void)0) | |||
1599 | "Expected a call to __kmpc_target_init in kernel\n")((void)0); | |||
1600 | ||||
1601 | CI->moveAfter(KernelInitCI); | |||
1602 | } else | |||
1603 | CI->moveBefore(&*F.getEntryBlock().getFirstInsertionPt()); | |||
1604 | ReplVal = CI; | |||
1605 | break; | |||
1606 | } | |||
1607 | if (!ReplVal) | |||
1608 | return false; | |||
1609 | } | |||
1610 | ||||
1611 | // If we use a call as a replacement value we need to make sure the ident is | |||
1612 | // valid at the new location. For now we just pick a global one, either | |||
1613 | // existing and used by one of the calls, or created from scratch. | |||
1614 | if (CallBase *CI = dyn_cast<CallBase>(ReplVal)) { | |||
1615 | if (CI->getNumArgOperands() > 0 && | |||
1616 | CI->getArgOperand(0)->getType() == OMPInfoCache.OMPBuilder.IdentPtr) { | |||
1617 | Value *Ident = getCombinedIdentFromCallUsesIn(RFI, F, | |||
1618 | /* GlobalOnly */ true); | |||
1619 | CI->setArgOperand(0, Ident); | |||
1620 | } | |||
1621 | } | |||
1622 | ||||
1623 | bool Changed = false; | |||
1624 | auto ReplaceAndDeleteCB = [&](Use &U, Function &Caller) { | |||
1625 | CallInst *CI = getCallIfRegularCall(U, &RFI); | |||
1626 | if (!CI || CI == ReplVal || &F != &Caller) | |||
1627 | return false; | |||
1628 | assert(CI->getCaller() == &F && "Unexpected call!")((void)0); | |||
1629 | ||||
1630 | auto Remark = [&](OptimizationRemark OR) { | |||
1631 | return OR << "OpenMP runtime call " | |||
1632 | << ore::NV("OpenMPOptRuntime", RFI.Name) << " deduplicated."; | |||
1633 | }; | |||
1634 | if (CI->getDebugLoc()) | |||
1635 | emitRemark<OptimizationRemark>(CI, "OMP170", Remark); | |||
1636 | else | |||
1637 | emitRemark<OptimizationRemark>(&F, "OMP170", Remark); | |||
1638 | ||||
1639 | CGUpdater.removeCallSite(*CI); | |||
1640 | CI->replaceAllUsesWith(ReplVal); | |||
1641 | CI->eraseFromParent(); | |||
1642 | ++NumOpenMPRuntimeCallsDeduplicated; | |||
1643 | Changed = true; | |||
1644 | return true; | |||
1645 | }; | |||
1646 | RFI.foreachUse(SCC, ReplaceAndDeleteCB); | |||
1647 | ||||
1648 | return Changed; | |||
1649 | } | |||
1650 | ||||
1651 | /// Collect arguments that represent the global thread id in \p GTIdArgs. | |||
1652 | void collectGlobalThreadIdArguments(SmallSetVector<Value *, 16> >IdArgs) { | |||
1653 | // TODO: Below we basically perform a fixpoint iteration with a pessimistic | |||
1654 | // initialization. We could define an AbstractAttribute instead and | |||
1655 | // run the Attributor here once it can be run as an SCC pass. | |||
1656 | ||||
1657 | // Helper to check the argument \p ArgNo at all call sites of \p F for | |||
1658 | // a GTId. | |||
1659 | auto CallArgOpIsGTId = [&](Function &F, unsigned ArgNo, CallInst &RefCI) { | |||
1660 | if (!F.hasLocalLinkage()) | |||
1661 | return false; | |||
1662 | for (Use &U : F.uses()) { | |||
1663 | if (CallInst *CI = getCallIfRegularCall(U)) { | |||
1664 | Value *ArgOp = CI->getArgOperand(ArgNo); | |||
1665 | if (CI == &RefCI || GTIdArgs.count(ArgOp) || | |||
1666 | getCallIfRegularCall( | |||
1667 | *ArgOp, &OMPInfoCache.RFIs[OMPRTL___kmpc_global_thread_num])) | |||
1668 | continue; | |||
1669 | } | |||
1670 | return false; | |||
1671 | } | |||
1672 | return true; | |||
1673 | }; | |||
1674 | ||||
1675 | // Helper to identify uses of a GTId as GTId arguments. | |||
1676 | auto AddUserArgs = [&](Value >Id) { | |||
1677 | for (Use &U : GTId.uses()) | |||
1678 | if (CallInst *CI = dyn_cast<CallInst>(U.getUser())) | |||
1679 | if (CI->isArgOperand(&U)) | |||
1680 | if (Function *Callee = CI->getCalledFunction()) | |||
1681 | if (CallArgOpIsGTId(*Callee, U.getOperandNo(), *CI)) | |||
1682 | GTIdArgs.insert(Callee->getArg(U.getOperandNo())); | |||
1683 | }; | |||
1684 | ||||
1685 | // The argument users of __kmpc_global_thread_num calls are GTIds. | |||
1686 | OMPInformationCache::RuntimeFunctionInfo &GlobThreadNumRFI = | |||
1687 | OMPInfoCache.RFIs[OMPRTL___kmpc_global_thread_num]; | |||
1688 | ||||
1689 | GlobThreadNumRFI.foreachUse(SCC, [&](Use &U, Function &F) { | |||
1690 | if (CallInst *CI = getCallIfRegularCall(U, &GlobThreadNumRFI)) | |||
1691 | AddUserArgs(*CI); | |||
1692 | return false; | |||
1693 | }); | |||
1694 | ||||
1695 | // Transitively search for more arguments by looking at the users of the | |||
1696 | // ones we know already. During the search the GTIdArgs vector is extended | |||
1697 | // so we cannot cache the size nor can we use a range based for. | |||
1698 | for (unsigned u = 0; u < GTIdArgs.size(); ++u) | |||
1699 | AddUserArgs(*GTIdArgs[u]); | |||
1700 | } | |||
1701 | ||||
1702 | /// Kernel (=GPU) optimizations and utility functions | |||
1703 | /// | |||
1704 | ///{{ | |||
1705 | ||||
1706 | /// Check if \p F is a kernel, hence entry point for target offloading. | |||
1707 | bool isKernel(Function &F) { return OMPInfoCache.Kernels.count(&F); } | |||
1708 | ||||
1709 | /// Cache to remember the unique kernel for a function. | |||
1710 | DenseMap<Function *, Optional<Kernel>> UniqueKernelMap; | |||
1711 | ||||
1712 | /// Find the unique kernel that will execute \p F, if any. | |||
1713 | Kernel getUniqueKernelFor(Function &F); | |||
1714 | ||||
1715 | /// Find the unique kernel that will execute \p I, if any. | |||
1716 | Kernel getUniqueKernelFor(Instruction &I) { | |||
1717 | return getUniqueKernelFor(*I.getFunction()); | |||
1718 | } | |||
1719 | ||||
1720 | /// Rewrite the device (=GPU) code state machine create in non-SPMD mode in | |||
1721 | /// the cases we can avoid taking the address of a function. | |||
1722 | bool rewriteDeviceCodeStateMachine(); | |||
1723 | ||||
1724 | /// | |||
1725 | ///}} | |||
1726 | ||||
1727 | /// Emit a remark generically | |||
1728 | /// | |||
1729 | /// This template function can be used to generically emit a remark. The | |||
1730 | /// RemarkKind should be one of the following: | |||
1731 | /// - OptimizationRemark to indicate a successful optimization attempt | |||
1732 | /// - OptimizationRemarkMissed to report a failed optimization attempt | |||
1733 | /// - OptimizationRemarkAnalysis to provide additional information about an | |||
1734 | /// optimization attempt | |||
1735 | /// | |||
1736 | /// The remark is built using a callback function provided by the caller that | |||
1737 | /// takes a RemarkKind as input and returns a RemarkKind. | |||
1738 | template <typename RemarkKind, typename RemarkCallBack> | |||
1739 | void emitRemark(Instruction *I, StringRef RemarkName, | |||
1740 | RemarkCallBack &&RemarkCB) const { | |||
1741 | Function *F = I->getParent()->getParent(); | |||
1742 | auto &ORE = OREGetter(F); | |||
1743 | ||||
1744 | if (RemarkName.startswith("OMP")) | |||
1745 | ORE.emit([&]() { | |||
1746 | return RemarkCB(RemarkKind(DEBUG_TYPE"openmp-opt", RemarkName, I)) | |||
1747 | << " [" << RemarkName << "]"; | |||
1748 | }); | |||
1749 | else | |||
1750 | ORE.emit( | |||
1751 | [&]() { return RemarkCB(RemarkKind(DEBUG_TYPE"openmp-opt", RemarkName, I)); }); | |||
1752 | } | |||
1753 | ||||
1754 | /// Emit a remark on a function. | |||
1755 | template <typename RemarkKind, typename RemarkCallBack> | |||
1756 | void emitRemark(Function *F, StringRef RemarkName, | |||
1757 | RemarkCallBack &&RemarkCB) const { | |||
1758 | auto &ORE = OREGetter(F); | |||
1759 | ||||
1760 | if (RemarkName.startswith("OMP")) | |||
1761 | ORE.emit([&]() { | |||
1762 | return RemarkCB(RemarkKind(DEBUG_TYPE"openmp-opt", RemarkName, F)) | |||
1763 | << " [" << RemarkName << "]"; | |||
1764 | }); | |||
1765 | else | |||
1766 | ORE.emit( | |||
1767 | [&]() { return RemarkCB(RemarkKind(DEBUG_TYPE"openmp-opt", RemarkName, F)); }); | |||
1768 | } | |||
1769 | ||||
1770 | /// RAII struct to temporarily change an RTL function's linkage to external. | |||
1771 | /// This prevents it from being mistakenly removed by other optimizations. | |||
1772 | struct ExternalizationRAII { | |||
1773 | ExternalizationRAII(OMPInformationCache &OMPInfoCache, | |||
1774 | RuntimeFunction RFKind) | |||
1775 | : Declaration(OMPInfoCache.RFIs[RFKind].Declaration) { | |||
1776 | if (!Declaration) | |||
1777 | return; | |||
1778 | ||||
1779 | LinkageType = Declaration->getLinkage(); | |||
1780 | Declaration->setLinkage(GlobalValue::ExternalLinkage); | |||
1781 | } | |||
1782 | ||||
1783 | ~ExternalizationRAII() { | |||
1784 | if (!Declaration) | |||
1785 | return; | |||
1786 | ||||
1787 | Declaration->setLinkage(LinkageType); | |||
1788 | } | |||
1789 | ||||
1790 | Function *Declaration; | |||
1791 | GlobalValue::LinkageTypes LinkageType; | |||
1792 | }; | |||
1793 | ||||
1794 | /// The underlying module. | |||
1795 | Module &M; | |||
1796 | ||||
1797 | /// The SCC we are operating on. | |||
1798 | SmallVectorImpl<Function *> &SCC; | |||
1799 | ||||
1800 | /// Callback to update the call graph, the first argument is a removed call, | |||
1801 | /// the second an optional replacement call. | |||
1802 | CallGraphUpdater &CGUpdater; | |||
1803 | ||||
1804 | /// Callback to get an OptimizationRemarkEmitter from a Function * | |||
1805 | OptimizationRemarkGetter OREGetter; | |||
1806 | ||||
1807 | /// OpenMP-specific information cache. Also Used for Attributor runs. | |||
1808 | OMPInformationCache &OMPInfoCache; | |||
1809 | ||||
1810 | /// Attributor instance. | |||
1811 | Attributor &A; | |||
1812 | ||||
1813 | /// Helper function to run Attributor on SCC. | |||
1814 | bool runAttributor(bool IsModulePass) { | |||
1815 | if (SCC.empty()) | |||
1816 | return false; | |||
1817 | ||||
1818 | // Temporarily make these function have external linkage so the Attributor | |||
1819 | // doesn't remove them when we try to look them up later. | |||
1820 | ExternalizationRAII Parallel(OMPInfoCache, OMPRTL___kmpc_kernel_parallel); | |||
1821 | ExternalizationRAII EndParallel(OMPInfoCache, | |||
1822 | OMPRTL___kmpc_kernel_end_parallel); | |||
1823 | ExternalizationRAII BarrierSPMD(OMPInfoCache, | |||
1824 | OMPRTL___kmpc_barrier_simple_spmd); | |||
1825 | ||||
1826 | registerAAs(IsModulePass); | |||
1827 | ||||
1828 | ChangeStatus Changed = A.run(); | |||
1829 | ||||
1830 | LLVM_DEBUG(dbgs() << "[Attributor] Done with " << SCC.size()do { } while (false) | |||
1831 | << " functions, result: " << Changed << ".\n")do { } while (false); | |||
1832 | ||||
1833 | return Changed == ChangeStatus::CHANGED; | |||
1834 | } | |||
1835 | ||||
1836 | void registerFoldRuntimeCall(RuntimeFunction RF); | |||
1837 | ||||
1838 | /// Populate the Attributor with abstract attribute opportunities in the | |||
1839 | /// function. | |||
1840 | void registerAAs(bool IsModulePass); | |||
1841 | }; | |||
1842 | ||||
1843 | Kernel OpenMPOpt::getUniqueKernelFor(Function &F) { | |||
1844 | if (!OMPInfoCache.ModuleSlice.count(&F)) | |||
1845 | return nullptr; | |||
1846 | ||||
1847 | // Use a scope to keep the lifetime of the CachedKernel short. | |||
1848 | { | |||
1849 | Optional<Kernel> &CachedKernel = UniqueKernelMap[&F]; | |||
1850 | if (CachedKernel) | |||
1851 | return *CachedKernel; | |||
1852 | ||||
1853 | // TODO: We should use an AA to create an (optimistic and callback | |||
1854 | // call-aware) call graph. For now we stick to simple patterns that | |||
1855 | // are less powerful, basically the worst fixpoint. | |||
1856 | if (isKernel(F)) { | |||
1857 | CachedKernel = Kernel(&F); | |||
1858 | return *CachedKernel; | |||
1859 | } | |||
1860 | ||||
1861 | CachedKernel = nullptr; | |||
1862 | if (!F.hasLocalLinkage()) { | |||
1863 | ||||
1864 | // See https://openmp.llvm.org/remarks/OptimizationRemarks.html | |||
1865 | auto Remark = [&](OptimizationRemarkAnalysis ORA) { | |||
1866 | return ORA << "Potentially unknown OpenMP target region caller."; | |||
1867 | }; | |||
1868 | emitRemark<OptimizationRemarkAnalysis>(&F, "OMP100", Remark); | |||
1869 | ||||
1870 | return nullptr; | |||
1871 | } | |||
1872 | } | |||
1873 | ||||
1874 | auto GetUniqueKernelForUse = [&](const Use &U) -> Kernel { | |||
1875 | if (auto *Cmp = dyn_cast<ICmpInst>(U.getUser())) { | |||
1876 | // Allow use in equality comparisons. | |||
1877 | if (Cmp->isEquality()) | |||
1878 | return getUniqueKernelFor(*Cmp); | |||
1879 | return nullptr; | |||
1880 | } | |||
1881 | if (auto *CB = dyn_cast<CallBase>(U.getUser())) { | |||
1882 | // Allow direct calls. | |||
1883 | if (CB->isCallee(&U)) | |||
1884 | return getUniqueKernelFor(*CB); | |||
1885 | ||||
1886 | OMPInformationCache::RuntimeFunctionInfo &KernelParallelRFI = | |||
1887 | OMPInfoCache.RFIs[OMPRTL___kmpc_parallel_51]; | |||
1888 | // Allow the use in __kmpc_parallel_51 calls. | |||
1889 | if (OpenMPOpt::getCallIfRegularCall(*U.getUser(), &KernelParallelRFI)) | |||
1890 | return getUniqueKernelFor(*CB); | |||
1891 | return nullptr; | |||
1892 | } | |||
1893 | // Disallow every other use. | |||
1894 | return nullptr; | |||
1895 | }; | |||
1896 | ||||
1897 | // TODO: In the future we want to track more than just a unique kernel. | |||
1898 | SmallPtrSet<Kernel, 2> PotentialKernels; | |||
1899 | OMPInformationCache::foreachUse(F, [&](const Use &U) { | |||
1900 | PotentialKernels.insert(GetUniqueKernelForUse(U)); | |||
1901 | }); | |||
1902 | ||||
1903 | Kernel K = nullptr; | |||
1904 | if (PotentialKernels.size() == 1) | |||
1905 | K = *PotentialKernels.begin(); | |||
1906 | ||||
1907 | // Cache the result. | |||
1908 | UniqueKernelMap[&F] = K; | |||
1909 | ||||
1910 | return K; | |||
1911 | } | |||
1912 | ||||
1913 | bool OpenMPOpt::rewriteDeviceCodeStateMachine() { | |||
1914 | OMPInformationCache::RuntimeFunctionInfo &KernelParallelRFI = | |||
1915 | OMPInfoCache.RFIs[OMPRTL___kmpc_parallel_51]; | |||
1916 | ||||
1917 | bool Changed = false; | |||
1918 | if (!KernelParallelRFI) | |||
1919 | return Changed; | |||
1920 | ||||
1921 | for (Function *F : SCC) { | |||
1922 | ||||
1923 | // Check if the function is a use in a __kmpc_parallel_51 call at | |||
1924 | // all. | |||
1925 | bool UnknownUse = false; | |||
1926 | bool KernelParallelUse = false; | |||
1927 | unsigned NumDirectCalls = 0; | |||
1928 | ||||
1929 | SmallVector<Use *, 2> ToBeReplacedStateMachineUses; | |||
1930 | OMPInformationCache::foreachUse(*F, [&](Use &U) { | |||
1931 | if (auto *CB = dyn_cast<CallBase>(U.getUser())) | |||
1932 | if (CB->isCallee(&U)) { | |||
1933 | ++NumDirectCalls; | |||
1934 | return; | |||
1935 | } | |||
1936 | ||||
1937 | if (isa<ICmpInst>(U.getUser())) { | |||
1938 | ToBeReplacedStateMachineUses.push_back(&U); | |||
1939 | return; | |||
1940 | } | |||
1941 | ||||
1942 | // Find wrapper functions that represent parallel kernels. | |||
1943 | CallInst *CI = | |||
1944 | OpenMPOpt::getCallIfRegularCall(*U.getUser(), &KernelParallelRFI); | |||
1945 | const unsigned int WrapperFunctionArgNo = 6; | |||
1946 | if (!KernelParallelUse && CI && | |||
1947 | CI->getArgOperandNo(&U) == WrapperFunctionArgNo) { | |||
1948 | KernelParallelUse = true; | |||
1949 | ToBeReplacedStateMachineUses.push_back(&U); | |||
1950 | return; | |||
1951 | } | |||
1952 | UnknownUse = true; | |||
1953 | }); | |||
1954 | ||||
1955 | // Do not emit a remark if we haven't seen a __kmpc_parallel_51 | |||
1956 | // use. | |||
1957 | if (!KernelParallelUse) | |||
1958 | continue; | |||
1959 | ||||
1960 | // If this ever hits, we should investigate. | |||
1961 | // TODO: Checking the number of uses is not a necessary restriction and | |||
1962 | // should be lifted. | |||
1963 | if (UnknownUse || NumDirectCalls != 1 || | |||
1964 | ToBeReplacedStateMachineUses.size() > 2) { | |||
1965 | auto Remark = [&](OptimizationRemarkAnalysis ORA) { | |||
1966 | return ORA << "Parallel region is used in " | |||
1967 | << (UnknownUse ? "unknown" : "unexpected") | |||
1968 | << " ways. Will not attempt to rewrite the state machine."; | |||
1969 | }; | |||
1970 | emitRemark<OptimizationRemarkAnalysis>(F, "OMP101", Remark); | |||
1971 | continue; | |||
1972 | } | |||
1973 | ||||
1974 | // Even if we have __kmpc_parallel_51 calls, we (for now) give | |||
1975 | // up if the function is not called from a unique kernel. | |||
1976 | Kernel K = getUniqueKernelFor(*F); | |||
1977 | if (!K) { | |||
1978 | auto Remark = [&](OptimizationRemarkAnalysis ORA) { | |||
1979 | return ORA << "Parallel region is not called from a unique kernel. " | |||
1980 | "Will not attempt to rewrite the state machine."; | |||
1981 | }; | |||
1982 | emitRemark<OptimizationRemarkAnalysis>(F, "OMP102", Remark); | |||
1983 | continue; | |||
1984 | } | |||
1985 | ||||
1986 | // We now know F is a parallel body function called only from the kernel K. | |||
1987 | // We also identified the state machine uses in which we replace the | |||
1988 | // function pointer by a new global symbol for identification purposes. This | |||
1989 | // ensures only direct calls to the function are left. | |||
1990 | ||||
1991 | Module &M = *F->getParent(); | |||
1992 | Type *Int8Ty = Type::getInt8Ty(M.getContext()); | |||
1993 | ||||
1994 | auto *ID = new GlobalVariable( | |||
1995 | M, Int8Ty, /* isConstant */ true, GlobalValue::PrivateLinkage, | |||
1996 | UndefValue::get(Int8Ty), F->getName() + ".ID"); | |||
1997 | ||||
1998 | for (Use *U : ToBeReplacedStateMachineUses) | |||
1999 | U->set(ConstantExpr::getPointerBitCastOrAddrSpaceCast( | |||
2000 | ID, U->get()->getType())); | |||
2001 | ||||
2002 | ++NumOpenMPParallelRegionsReplacedInGPUStateMachine; | |||
2003 | ||||
2004 | Changed = true; | |||
2005 | } | |||
2006 | ||||
2007 | return Changed; | |||
2008 | } | |||
2009 | ||||
2010 | /// Abstract Attribute for tracking ICV values. | |||
2011 | struct AAICVTracker : public StateWrapper<BooleanState, AbstractAttribute> { | |||
2012 | using Base = StateWrapper<BooleanState, AbstractAttribute>; | |||
2013 | AAICVTracker(const IRPosition &IRP, Attributor &A) : Base(IRP) {} | |||
2014 | ||||
2015 | void initialize(Attributor &A) override { | |||
2016 | Function *F = getAnchorScope(); | |||
2017 | if (!F || !A.isFunctionIPOAmendable(*F)) | |||
2018 | indicatePessimisticFixpoint(); | |||
2019 | } | |||
2020 | ||||
2021 | /// Returns true if value is assumed to be tracked. | |||
2022 | bool isAssumedTracked() const { return getAssumed(); } | |||
2023 | ||||
2024 | /// Returns true if value is known to be tracked. | |||
2025 | bool isKnownTracked() const { return getAssumed(); } | |||
2026 | ||||
2027 | /// Create an abstract attribute biew for the position \p IRP. | |||
2028 | static AAICVTracker &createForPosition(const IRPosition &IRP, Attributor &A); | |||
2029 | ||||
2030 | /// Return the value with which \p I can be replaced for specific \p ICV. | |||
2031 | virtual Optional<Value *> getReplacementValue(InternalControlVar ICV, | |||
2032 | const Instruction *I, | |||
2033 | Attributor &A) const { | |||
2034 | return None; | |||
2035 | } | |||
2036 | ||||
2037 | /// Return an assumed unique ICV value if a single candidate is found. If | |||
2038 | /// there cannot be one, return a nullptr. If it is not clear yet, return the | |||
2039 | /// Optional::NoneType. | |||
2040 | virtual Optional<Value *> | |||
2041 | getUniqueReplacementValue(InternalControlVar ICV) const = 0; | |||
2042 | ||||
2043 | // Currently only nthreads is being tracked. | |||
2044 | // this array will only grow with time. | |||
2045 | InternalControlVar TrackableICVs[1] = {ICV_nthreads}; | |||
2046 | ||||
2047 | /// See AbstractAttribute::getName() | |||
2048 | const std::string getName() const override { return "AAICVTracker"; } | |||
2049 | ||||
2050 | /// See AbstractAttribute::getIdAddr() | |||
2051 | const char *getIdAddr() const override { return &ID; } | |||
2052 | ||||
2053 | /// This function should return true if the type of the \p AA is AAICVTracker | |||
2054 | static bool classof(const AbstractAttribute *AA) { | |||
2055 | return (AA->getIdAddr() == &ID); | |||
2056 | } | |||
2057 | ||||
2058 | static const char ID; | |||
2059 | }; | |||
2060 | ||||
2061 | struct AAICVTrackerFunction : public AAICVTracker { | |||
2062 | AAICVTrackerFunction(const IRPosition &IRP, Attributor &A) | |||
2063 | : AAICVTracker(IRP, A) {} | |||
2064 | ||||
2065 | // FIXME: come up with better string. | |||
2066 | const std::string getAsStr() const override { return "ICVTrackerFunction"; } | |||
2067 | ||||
2068 | // FIXME: come up with some stats. | |||
2069 | void trackStatistics() const override {} | |||
2070 | ||||
2071 | /// We don't manifest anything for this AA. | |||
2072 | ChangeStatus manifest(Attributor &A) override { | |||
2073 | return ChangeStatus::UNCHANGED; | |||
2074 | } | |||
2075 | ||||
2076 | // Map of ICV to their values at specific program point. | |||
2077 | EnumeratedArray<DenseMap<Instruction *, Value *>, InternalControlVar, | |||
2078 | InternalControlVar::ICV___last> | |||
2079 | ICVReplacementValuesMap; | |||
2080 | ||||
2081 | ChangeStatus updateImpl(Attributor &A) override { | |||
2082 | ChangeStatus HasChanged = ChangeStatus::UNCHANGED; | |||
2083 | ||||
2084 | Function *F = getAnchorScope(); | |||
2085 | ||||
2086 | auto &OMPInfoCache = static_cast<OMPInformationCache &>(A.getInfoCache()); | |||
2087 | ||||
2088 | for (InternalControlVar ICV : TrackableICVs) { | |||
2089 | auto &SetterRFI = OMPInfoCache.RFIs[OMPInfoCache.ICVs[ICV].Setter]; | |||
2090 | ||||
2091 | auto &ValuesMap = ICVReplacementValuesMap[ICV]; | |||
2092 | auto TrackValues = [&](Use &U, Function &) { | |||
2093 | CallInst *CI = OpenMPOpt::getCallIfRegularCall(U); | |||
2094 | if (!CI) | |||
2095 | return false; | |||
2096 | ||||
2097 | // FIXME: handle setters with more that 1 arguments. | |||
2098 | /// Track new value. | |||
2099 | if (ValuesMap.insert(std::make_pair(CI, CI->getArgOperand(0))).second) | |||
2100 | HasChanged = ChangeStatus::CHANGED; | |||
2101 | ||||
2102 | return false; | |||
2103 | }; | |||
2104 | ||||
2105 | auto CallCheck = [&](Instruction &I) { | |||
2106 | Optional<Value *> ReplVal = getValueForCall(A, &I, ICV); | |||
2107 | if (ReplVal.hasValue() && | |||
2108 | ValuesMap.insert(std::make_pair(&I, *ReplVal)).second) | |||
2109 | HasChanged = ChangeStatus::CHANGED; | |||
2110 | ||||
2111 | return true; | |||
2112 | }; | |||
2113 | ||||
2114 | // Track all changes of an ICV. | |||
2115 | SetterRFI.foreachUse(TrackValues, F); | |||
2116 | ||||
2117 | bool UsedAssumedInformation = false; | |||
2118 | A.checkForAllInstructions(CallCheck, *this, {Instruction::Call}, | |||
2119 | UsedAssumedInformation, | |||
2120 | /* CheckBBLivenessOnly */ true); | |||
2121 | ||||
2122 | /// TODO: Figure out a way to avoid adding entry in | |||
2123 | /// ICVReplacementValuesMap | |||
2124 | Instruction *Entry = &F->getEntryBlock().front(); | |||
2125 | if (HasChanged == ChangeStatus::CHANGED && !ValuesMap.count(Entry)) | |||
2126 | ValuesMap.insert(std::make_pair(Entry, nullptr)); | |||
2127 | } | |||
2128 | ||||
2129 | return HasChanged; | |||
2130 | } | |||
2131 | ||||
2132 | /// Hepler to check if \p I is a call and get the value for it if it is | |||
2133 | /// unique. | |||
2134 | Optional<Value *> getValueForCall(Attributor &A, const Instruction *I, | |||
2135 | InternalControlVar &ICV) const { | |||
2136 | ||||
2137 | const auto *CB = dyn_cast<CallBase>(I); | |||
2138 | if (!CB || CB->hasFnAttr("no_openmp") || | |||
2139 | CB->hasFnAttr("no_openmp_routines")) | |||
2140 | return None; | |||
2141 | ||||
2142 | auto &OMPInfoCache = static_cast<OMPInformationCache &>(A.getInfoCache()); | |||
2143 | auto &GetterRFI = OMPInfoCache.RFIs[OMPInfoCache.ICVs[ICV].Getter]; | |||
2144 | auto &SetterRFI = OMPInfoCache.RFIs[OMPInfoCache.ICVs[ICV].Setter]; | |||
2145 | Function *CalledFunction = CB->getCalledFunction(); | |||
2146 | ||||
2147 | // Indirect call, assume ICV changes. | |||
2148 | if (CalledFunction == nullptr) | |||
2149 | return nullptr; | |||
2150 | if (CalledFunction == GetterRFI.Declaration) | |||
2151 | return None; | |||
2152 | if (CalledFunction == SetterRFI.Declaration) { | |||
2153 | if (ICVReplacementValuesMap[ICV].count(I)) | |||
2154 | return ICVReplacementValuesMap[ICV].lookup(I); | |||
2155 | ||||
2156 | return nullptr; | |||
2157 | } | |||
2158 | ||||
2159 | // Since we don't know, assume it changes the ICV. | |||
2160 | if (CalledFunction->isDeclaration()) | |||
2161 | return nullptr; | |||
2162 | ||||
2163 | const auto &ICVTrackingAA = A.getAAFor<AAICVTracker>( | |||
2164 | *this, IRPosition::callsite_returned(*CB), DepClassTy::REQUIRED); | |||
2165 | ||||
2166 | if (ICVTrackingAA.isAssumedTracked()) | |||
2167 | return ICVTrackingAA.getUniqueReplacementValue(ICV); | |||
2168 | ||||
2169 | // If we don't know, assume it changes. | |||
2170 | return nullptr; | |||
2171 | } | |||
2172 | ||||
2173 | // We don't check unique value for a function, so return None. | |||
2174 | Optional<Value *> | |||
2175 | getUniqueReplacementValue(InternalControlVar ICV) const override { | |||
2176 | return None; | |||
2177 | } | |||
2178 | ||||
2179 | /// Return the value with which \p I can be replaced for specific \p ICV. | |||
2180 | Optional<Value *> getReplacementValue(InternalControlVar ICV, | |||
2181 | const Instruction *I, | |||
2182 | Attributor &A) const override { | |||
2183 | const auto &ValuesMap = ICVReplacementValuesMap[ICV]; | |||
2184 | if (ValuesMap.count(I)) | |||
2185 | return ValuesMap.lookup(I); | |||
2186 | ||||
2187 | SmallVector<const Instruction *, 16> Worklist; | |||
2188 | SmallPtrSet<const Instruction *, 16> Visited; | |||
2189 | Worklist.push_back(I); | |||
2190 | ||||
2191 | Optional<Value *> ReplVal; | |||
2192 | ||||
2193 | while (!Worklist.empty()) { | |||
2194 | const Instruction *CurrInst = Worklist.pop_back_val(); | |||
2195 | if (!Visited.insert(CurrInst).second) | |||
2196 | continue; | |||
2197 | ||||
2198 | const BasicBlock *CurrBB = CurrInst->getParent(); | |||
2199 | ||||
2200 | // Go up and look for all potential setters/calls that might change the | |||
2201 | // ICV. | |||
2202 | while ((CurrInst = CurrInst->getPrevNode())) { | |||
2203 | if (ValuesMap.count(CurrInst)) { | |||
2204 | Optional<Value *> NewReplVal = ValuesMap.lookup(CurrInst); | |||
2205 | // Unknown value, track new. | |||
2206 | if (!ReplVal.hasValue()) { | |||
2207 | ReplVal = NewReplVal; | |||
2208 | break; | |||
2209 | } | |||
2210 | ||||
2211 | // If we found a new value, we can't know the icv value anymore. | |||
2212 | if (NewReplVal.hasValue()) | |||
2213 | if (ReplVal != NewReplVal) | |||
2214 | return nullptr; | |||
2215 | ||||
2216 | break; | |||
2217 | } | |||
2218 | ||||
2219 | Optional<Value *> NewReplVal = getValueForCall(A, CurrInst, ICV); | |||
2220 | if (!NewReplVal.hasValue()) | |||
2221 | continue; | |||
2222 | ||||
2223 | // Unknown value, track new. | |||
2224 | if (!ReplVal.hasValue()) { | |||
2225 | ReplVal = NewReplVal; | |||
2226 | break; | |||
2227 | } | |||
2228 | ||||
2229 | // if (NewReplVal.hasValue()) | |||
2230 | // We found a new value, we can't know the icv value anymore. | |||
2231 | if (ReplVal != NewReplVal) | |||
2232 | return nullptr; | |||
2233 | } | |||
2234 | ||||
2235 | // If we are in the same BB and we have a value, we are done. | |||
2236 | if (CurrBB == I->getParent() && ReplVal.hasValue()) | |||
2237 | return ReplVal; | |||
2238 | ||||
2239 | // Go through all predecessors and add terminators for analysis. | |||
2240 | for (const BasicBlock *Pred : predecessors(CurrBB)) | |||
2241 | if (const Instruction *Terminator = Pred->getTerminator()) | |||
2242 | Worklist.push_back(Terminator); | |||
2243 | } | |||
2244 | ||||
2245 | return ReplVal; | |||
2246 | } | |||
2247 | }; | |||
2248 | ||||
2249 | struct AAICVTrackerFunctionReturned : AAICVTracker { | |||
2250 | AAICVTrackerFunctionReturned(const IRPosition &IRP, Attributor &A) | |||
2251 | : AAICVTracker(IRP, A) {} | |||
2252 | ||||
2253 | // FIXME: come up with better string. | |||
2254 | const std::string getAsStr() const override { | |||
2255 | return "ICVTrackerFunctionReturned"; | |||
2256 | } | |||
2257 | ||||
2258 | // FIXME: come up with some stats. | |||
2259 | void trackStatistics() const override {} | |||
2260 | ||||
2261 | /// We don't manifest anything for this AA. | |||
2262 | ChangeStatus manifest(Attributor &A) override { | |||
2263 | return ChangeStatus::UNCHANGED; | |||
2264 | } | |||
2265 | ||||
2266 | // Map of ICV to their values at specific program point. | |||
2267 | EnumeratedArray<Optional<Value *>, InternalControlVar, | |||
2268 | InternalControlVar::ICV___last> | |||
2269 | ICVReplacementValuesMap; | |||
2270 | ||||
2271 | /// Return the value with which \p I can be replaced for specific \p ICV. | |||
2272 | Optional<Value *> | |||
2273 | getUniqueReplacementValue(InternalControlVar ICV) const override { | |||
2274 | return ICVReplacementValuesMap[ICV]; | |||
2275 | } | |||
2276 | ||||
2277 | ChangeStatus updateImpl(Attributor &A) override { | |||
2278 | ChangeStatus Changed = ChangeStatus::UNCHANGED; | |||
2279 | const auto &ICVTrackingAA = A.getAAFor<AAICVTracker>( | |||
2280 | *this, IRPosition::function(*getAnchorScope()), DepClassTy::REQUIRED); | |||
2281 | ||||
2282 | if (!ICVTrackingAA.isAssumedTracked()) | |||
2283 | return indicatePessimisticFixpoint(); | |||
2284 | ||||
2285 | for (InternalControlVar ICV : TrackableICVs) { | |||
2286 | Optional<Value *> &ReplVal = ICVReplacementValuesMap[ICV]; | |||
2287 | Optional<Value *> UniqueICVValue; | |||
2288 | ||||
2289 | auto CheckReturnInst = [&](Instruction &I) { | |||
2290 | Optional<Value *> NewReplVal = | |||
2291 | ICVTrackingAA.getReplacementValue(ICV, &I, A); | |||
2292 | ||||
2293 | // If we found a second ICV value there is no unique returned value. | |||
2294 | if (UniqueICVValue.hasValue() && UniqueICVValue != NewReplVal) | |||
2295 | return false; | |||
2296 | ||||
2297 | UniqueICVValue = NewReplVal; | |||
2298 | ||||
2299 | return true; | |||
2300 | }; | |||
2301 | ||||
2302 | bool UsedAssumedInformation = false; | |||
2303 | if (!A.checkForAllInstructions(CheckReturnInst, *this, {Instruction::Ret}, | |||
2304 | UsedAssumedInformation, | |||
2305 | /* CheckBBLivenessOnly */ true)) | |||
2306 | UniqueICVValue = nullptr; | |||
2307 | ||||
2308 | if (UniqueICVValue == ReplVal) | |||
2309 | continue; | |||
2310 | ||||
2311 | ReplVal = UniqueICVValue; | |||
2312 | Changed = ChangeStatus::CHANGED; | |||
2313 | } | |||
2314 | ||||
2315 | return Changed; | |||
2316 | } | |||
2317 | }; | |||
2318 | ||||
2319 | struct AAICVTrackerCallSite : AAICVTracker { | |||
2320 | AAICVTrackerCallSite(const IRPosition &IRP, Attributor &A) | |||
2321 | : AAICVTracker(IRP, A) {} | |||
2322 | ||||
2323 | void initialize(Attributor &A) override { | |||
2324 | Function *F = getAnchorScope(); | |||
2325 | if (!F || !A.isFunctionIPOAmendable(*F)) | |||
2326 | indicatePessimisticFixpoint(); | |||
2327 | ||||
2328 | // We only initialize this AA for getters, so we need to know which ICV it | |||
2329 | // gets. | |||
2330 | auto &OMPInfoCache = static_cast<OMPInformationCache &>(A.getInfoCache()); | |||
2331 | for (InternalControlVar ICV : TrackableICVs) { | |||
2332 | auto ICVInfo = OMPInfoCache.ICVs[ICV]; | |||
2333 | auto &Getter = OMPInfoCache.RFIs[ICVInfo.Getter]; | |||
2334 | if (Getter.Declaration == getAssociatedFunction()) { | |||
2335 | AssociatedICV = ICVInfo.Kind; | |||
2336 | return; | |||
2337 | } | |||
2338 | } | |||
2339 | ||||
2340 | /// Unknown ICV. | |||
2341 | indicatePessimisticFixpoint(); | |||
2342 | } | |||
2343 | ||||
2344 | ChangeStatus manifest(Attributor &A) override { | |||
2345 | if (!ReplVal.hasValue() || !ReplVal.getValue()) | |||
2346 | return ChangeStatus::UNCHANGED; | |||
2347 | ||||
2348 | A.changeValueAfterManifest(*getCtxI(), **ReplVal); | |||
2349 | A.deleteAfterManifest(*getCtxI()); | |||
2350 | ||||
2351 | return ChangeStatus::CHANGED; | |||
2352 | } | |||
2353 | ||||
2354 | // FIXME: come up with better string. | |||
2355 | const std::string getAsStr() const override { return "ICVTrackerCallSite"; } | |||
2356 | ||||
2357 | // FIXME: come up with some stats. | |||
2358 | void trackStatistics() const override {} | |||
2359 | ||||
2360 | InternalControlVar AssociatedICV; | |||
2361 | Optional<Value *> ReplVal; | |||
2362 | ||||
2363 | ChangeStatus updateImpl(Attributor &A) override { | |||
2364 | const auto &ICVTrackingAA = A.getAAFor<AAICVTracker>( | |||
2365 | *this, IRPosition::function(*getAnchorScope()), DepClassTy::REQUIRED); | |||
2366 | ||||
2367 | // We don't have any information, so we assume it changes the ICV. | |||
2368 | if (!ICVTrackingAA.isAssumedTracked()) | |||
2369 | return indicatePessimisticFixpoint(); | |||
2370 | ||||
2371 | Optional<Value *> NewReplVal = | |||
2372 | ICVTrackingAA.getReplacementValue(AssociatedICV, getCtxI(), A); | |||
2373 | ||||
2374 | if (ReplVal == NewReplVal) | |||
2375 | return ChangeStatus::UNCHANGED; | |||
2376 | ||||
2377 | ReplVal = NewReplVal; | |||
2378 | return ChangeStatus::CHANGED; | |||
2379 | } | |||
2380 | ||||
2381 | // Return the value with which associated value can be replaced for specific | |||
2382 | // \p ICV. | |||
2383 | Optional<Value *> | |||
2384 | getUniqueReplacementValue(InternalControlVar ICV) const override { | |||
2385 | return ReplVal; | |||
2386 | } | |||
2387 | }; | |||
2388 | ||||
2389 | struct AAICVTrackerCallSiteReturned : AAICVTracker { | |||
2390 | AAICVTrackerCallSiteReturned(const IRPosition &IRP, Attributor &A) | |||
2391 | : AAICVTracker(IRP, A) {} | |||
2392 | ||||
2393 | // FIXME: come up with better string. | |||
2394 | const std::string getAsStr() const override { | |||
2395 | return "ICVTrackerCallSiteReturned"; | |||
2396 | } | |||
2397 | ||||
2398 | // FIXME: come up with some stats. | |||
2399 | void trackStatistics() const override {} | |||
2400 | ||||
2401 | /// We don't manifest anything for this AA. | |||
2402 | ChangeStatus manifest(Attributor &A) override { | |||
2403 | return ChangeStatus::UNCHANGED; | |||
2404 | } | |||
2405 | ||||
2406 | // Map of ICV to their values at specific program point. | |||
2407 | EnumeratedArray<Optional<Value *>, InternalControlVar, | |||
2408 | InternalControlVar::ICV___last> | |||
2409 | ICVReplacementValuesMap; | |||
2410 | ||||
2411 | /// Return the value with which associated value can be replaced for specific | |||
2412 | /// \p ICV. | |||
2413 | Optional<Value *> | |||
2414 | getUniqueReplacementValue(InternalControlVar ICV) const override { | |||
2415 | return ICVReplacementValuesMap[ICV]; | |||
2416 | } | |||
2417 | ||||
2418 | ChangeStatus updateImpl(Attributor &A) override { | |||
2419 | ChangeStatus Changed = ChangeStatus::UNCHANGED; | |||
2420 | const auto &ICVTrackingAA = A.getAAFor<AAICVTracker>( | |||
2421 | *this, IRPosition::returned(*getAssociatedFunction()), | |||
2422 | DepClassTy::REQUIRED); | |||
2423 | ||||
2424 | // We don't have any information, so we assume it changes the ICV. | |||
2425 | if (!ICVTrackingAA.isAssumedTracked()) | |||
2426 | return indicatePessimisticFixpoint(); | |||
2427 | ||||
2428 | for (InternalControlVar ICV : TrackableICVs) { | |||
2429 | Optional<Value *> &ReplVal = ICVReplacementValuesMap[ICV]; | |||
2430 | Optional<Value *> NewReplVal = | |||
2431 | ICVTrackingAA.getUniqueReplacementValue(ICV); | |||
2432 | ||||
2433 | if (ReplVal == NewReplVal) | |||
2434 | continue; | |||
2435 | ||||
2436 | ReplVal = NewReplVal; | |||
2437 | Changed = ChangeStatus::CHANGED; | |||
2438 | } | |||
2439 | return Changed; | |||
2440 | } | |||
2441 | }; | |||
2442 | ||||
2443 | struct AAExecutionDomainFunction : public AAExecutionDomain { | |||
2444 | AAExecutionDomainFunction(const IRPosition &IRP, Attributor &A) | |||
2445 | : AAExecutionDomain(IRP, A) {} | |||
2446 | ||||
2447 | const std::string getAsStr() const override { | |||
2448 | return "[AAExecutionDomain] " + std::to_string(SingleThreadedBBs.size()) + | |||
2449 | "/" + std::to_string(NumBBs) + " BBs thread 0 only."; | |||
2450 | } | |||
2451 | ||||
2452 | /// See AbstractAttribute::trackStatistics(). | |||
2453 | void trackStatistics() const override {} | |||
2454 | ||||
2455 | void initialize(Attributor &A) override { | |||
2456 | Function *F = getAnchorScope(); | |||
2457 | for (const auto &BB : *F) | |||
2458 | SingleThreadedBBs.insert(&BB); | |||
2459 | NumBBs = SingleThreadedBBs.size(); | |||
2460 | } | |||
2461 | ||||
2462 | ChangeStatus manifest(Attributor &A) override { | |||
2463 | LLVM_DEBUG({do { } while (false) | |||
2464 | for (const BasicBlock *BB : SingleThreadedBBs)do { } while (false) | |||
2465 | dbgs() << TAG << " Basic block @" << getAnchorScope()->getName() << " "do { } while (false) | |||
2466 | << BB->getName() << " is executed by a single thread.\n";do { } while (false) | |||
2467 | })do { } while (false); | |||
2468 | return ChangeStatus::UNCHANGED; | |||
2469 | } | |||
2470 | ||||
2471 | ChangeStatus updateImpl(Attributor &A) override; | |||
2472 | ||||
2473 | /// Check if an instruction is executed by a single thread. | |||
2474 | bool isExecutedByInitialThreadOnly(const Instruction &I) const override { | |||
2475 | return isExecutedByInitialThreadOnly(*I.getParent()); | |||
2476 | } | |||
2477 | ||||
2478 | bool isExecutedByInitialThreadOnly(const BasicBlock &BB) const override { | |||
2479 | return isValidState() && SingleThreadedBBs.contains(&BB); | |||
2480 | } | |||
2481 | ||||
2482 | /// Set of basic blocks that are executed by a single thread. | |||
2483 | DenseSet<const BasicBlock *> SingleThreadedBBs; | |||
2484 | ||||
2485 | /// Total number of basic blocks in this function. | |||
2486 | long unsigned NumBBs; | |||
2487 | }; | |||
2488 | ||||
2489 | ChangeStatus AAExecutionDomainFunction::updateImpl(Attributor &A) { | |||
2490 | Function *F = getAnchorScope(); | |||
2491 | ReversePostOrderTraversal<Function *> RPOT(F); | |||
2492 | auto NumSingleThreadedBBs = SingleThreadedBBs.size(); | |||
2493 | ||||
2494 | bool AllCallSitesKnown; | |||
2495 | auto PredForCallSite = [&](AbstractCallSite ACS) { | |||
2496 | const auto &ExecutionDomainAA = A.getAAFor<AAExecutionDomain>( | |||
2497 | *this, IRPosition::function(*ACS.getInstruction()->getFunction()), | |||
2498 | DepClassTy::REQUIRED); | |||
2499 | return ACS.isDirectCall() && | |||
2500 | ExecutionDomainAA.isExecutedByInitialThreadOnly( | |||
2501 | *ACS.getInstruction()); | |||
2502 | }; | |||
2503 | ||||
2504 | if (!A.checkForAllCallSites(PredForCallSite, *this, | |||
2505 | /* RequiresAllCallSites */ true, | |||
2506 | AllCallSitesKnown)) | |||
2507 | SingleThreadedBBs.erase(&F->getEntryBlock()); | |||
2508 | ||||
2509 | auto &OMPInfoCache = static_cast<OMPInformationCache &>(A.getInfoCache()); | |||
2510 | auto &RFI = OMPInfoCache.RFIs[OMPRTL___kmpc_target_init]; | |||
2511 | ||||
2512 | // Check if the edge into the successor block compares the __kmpc_target_init | |||
2513 | // result with -1. If we are in non-SPMD-mode that signals only the main | |||
2514 | // thread will execute the edge. | |||
2515 | auto IsInitialThreadOnly = [&](BranchInst *Edge, BasicBlock *SuccessorBB) { | |||
2516 | if (!Edge || !Edge->isConditional()) | |||
2517 | return false; | |||
2518 | if (Edge->getSuccessor(0) != SuccessorBB) | |||
2519 | return false; | |||
2520 | ||||
2521 | auto *Cmp = dyn_cast<CmpInst>(Edge->getCondition()); | |||
2522 | if (!Cmp || !Cmp->isTrueWhenEqual() || !Cmp->isEquality()) | |||
2523 | return false; | |||
2524 | ||||
2525 | ConstantInt *C = dyn_cast<ConstantInt>(Cmp->getOperand(1)); | |||
2526 | if (!C) | |||
2527 | return false; | |||
2528 | ||||
2529 | // Match: -1 == __kmpc_target_init (for non-SPMD kernels only!) | |||
2530 | if (C->isAllOnesValue()) { | |||
2531 | auto *CB = dyn_cast<CallBase>(Cmp->getOperand(0)); | |||
2532 | CB = CB ? OpenMPOpt::getCallIfRegularCall(*CB, &RFI) : nullptr; | |||
2533 | if (!CB) | |||
2534 | return false; | |||
2535 | const int InitIsSPMDArgNo = 1; | |||
2536 | auto *IsSPMDModeCI = | |||
2537 | dyn_cast<ConstantInt>(CB->getOperand(InitIsSPMDArgNo)); | |||
2538 | return IsSPMDModeCI && IsSPMDModeCI->isZero(); | |||
2539 | } | |||
2540 | ||||
2541 | return false; | |||
2542 | }; | |||
2543 | ||||
2544 | // Merge all the predecessor states into the current basic block. A basic | |||
2545 | // block is executed by a single thread if all of its predecessors are. | |||
2546 | auto MergePredecessorStates = [&](BasicBlock *BB) { | |||
2547 | if (pred_begin(BB) == pred_end(BB)) | |||
2548 | return SingleThreadedBBs.contains(BB); | |||
2549 | ||||
2550 | bool IsInitialThread = true; | |||
2551 | for (auto PredBB = pred_begin(BB), PredEndBB = pred_end(BB); | |||
2552 | PredBB != PredEndBB; ++PredBB) { | |||
2553 | if (!IsInitialThreadOnly(dyn_cast<BranchInst>((*PredBB)->getTerminator()), | |||
2554 | BB)) | |||
2555 | IsInitialThread &= SingleThreadedBBs.contains(*PredBB); | |||
2556 | } | |||
2557 | ||||
2558 | return IsInitialThread; | |||
2559 | }; | |||
2560 | ||||
2561 | for (auto *BB : RPOT) { | |||
2562 | if (!MergePredecessorStates(BB)) | |||
2563 | SingleThreadedBBs.erase(BB); | |||
2564 | } | |||
2565 | ||||
2566 | return (NumSingleThreadedBBs == SingleThreadedBBs.size()) | |||
2567 | ? ChangeStatus::UNCHANGED | |||
2568 | : ChangeStatus::CHANGED; | |||
2569 | } | |||
2570 | ||||
2571 | /// Try to replace memory allocation calls called by a single thread with a | |||
2572 | /// static buffer of shared memory. | |||
2573 | struct AAHeapToShared : public StateWrapper<BooleanState, AbstractAttribute> { | |||
2574 | using Base = StateWrapper<BooleanState, AbstractAttribute>; | |||
2575 | AAHeapToShared(const IRPosition &IRP, Attributor &A) : Base(IRP) {} | |||
2576 | ||||
2577 | /// Create an abstract attribute view for the position \p IRP. | |||
2578 | static AAHeapToShared &createForPosition(const IRPosition &IRP, | |||
2579 | Attributor &A); | |||
2580 | ||||
2581 | /// Returns true if HeapToShared conversion is assumed to be possible. | |||
2582 | virtual bool isAssumedHeapToShared(CallBase &CB) const = 0; | |||
2583 | ||||
2584 | /// Returns true if HeapToShared conversion is assumed and the CB is a | |||
2585 | /// callsite to a free operation to be removed. | |||
2586 | virtual bool isAssumedHeapToSharedRemovedFree(CallBase &CB) const = 0; | |||
2587 | ||||
2588 | /// See AbstractAttribute::getName(). | |||
2589 | const std::string getName() const override { return "AAHeapToShared"; } | |||
2590 | ||||
2591 | /// See AbstractAttribute::getIdAddr(). | |||
2592 | const char *getIdAddr() const override { return &ID; } | |||
2593 | ||||
2594 | /// This function should return true if the type of the \p AA is | |||
2595 | /// AAHeapToShared. | |||
2596 | static bool classof(const AbstractAttribute *AA) { | |||
2597 | return (AA->getIdAddr() == &ID); | |||
2598 | } | |||
2599 | ||||
2600 | /// Unique ID (due to the unique address) | |||
2601 | static const char ID; | |||
2602 | }; | |||
2603 | ||||
2604 | struct AAHeapToSharedFunction : public AAHeapToShared { | |||
2605 | AAHeapToSharedFunction(const IRPosition &IRP, Attributor &A) | |||
2606 | : AAHeapToShared(IRP, A) {} | |||
2607 | ||||
2608 | const std::string getAsStr() const override { | |||
2609 | return "[AAHeapToShared] " + std::to_string(MallocCalls.size()) + | |||
2610 | " malloc calls eligible."; | |||
2611 | } | |||
2612 | ||||
2613 | /// See AbstractAttribute::trackStatistics(). | |||
2614 | void trackStatistics() const override {} | |||
2615 | ||||
2616 | /// This functions finds free calls that will be removed by the | |||
2617 | /// HeapToShared transformation. | |||
2618 | void findPotentialRemovedFreeCalls(Attributor &A) { | |||
2619 | auto &OMPInfoCache = static_cast<OMPInformationCache &>(A.getInfoCache()); | |||
2620 | auto &FreeRFI = OMPInfoCache.RFIs[OMPRTL___kmpc_free_shared]; | |||
2621 | ||||
2622 | PotentialRemovedFreeCalls.clear(); | |||
2623 | // Update free call users of found malloc calls. | |||
2624 | for (CallBase *CB : MallocCalls) { | |||
2625 | SmallVector<CallBase *, 4> FreeCalls; | |||
2626 | for (auto *U : CB->users()) { | |||
2627 | CallBase *C = dyn_cast<CallBase>(U); | |||
2628 | if (C && C->getCalledFunction() == FreeRFI.Declaration) | |||
2629 | FreeCalls.push_back(C); | |||
2630 | } | |||
2631 | ||||
2632 | if (FreeCalls.size() != 1) | |||
2633 | continue; | |||
2634 | ||||
2635 | PotentialRemovedFreeCalls.insert(FreeCalls.front()); | |||
2636 | } | |||
2637 | } | |||
2638 | ||||
2639 | void initialize(Attributor &A) override { | |||
2640 | auto &OMPInfoCache = static_cast<OMPInformationCache &>(A.getInfoCache()); | |||
2641 | auto &RFI = OMPInfoCache.RFIs[OMPRTL___kmpc_alloc_shared]; | |||
2642 | ||||
2643 | for (User *U : RFI.Declaration->users()) | |||
2644 | if (CallBase *CB = dyn_cast<CallBase>(U)) | |||
2645 | MallocCalls.insert(CB); | |||
2646 | ||||
2647 | findPotentialRemovedFreeCalls(A); | |||
2648 | } | |||
2649 | ||||
2650 | bool isAssumedHeapToShared(CallBase &CB) const override { | |||
2651 | return isValidState() && MallocCalls.count(&CB); | |||
2652 | } | |||
2653 | ||||
2654 | bool isAssumedHeapToSharedRemovedFree(CallBase &CB) const override { | |||
2655 | return isValidState() && PotentialRemovedFreeCalls.count(&CB); | |||
2656 | } | |||
2657 | ||||
2658 | ChangeStatus manifest(Attributor &A) override { | |||
2659 | if (MallocCalls.empty()) | |||
2660 | return ChangeStatus::UNCHANGED; | |||
2661 | ||||
2662 | auto &OMPInfoCache = static_cast<OMPInformationCache &>(A.getInfoCache()); | |||
2663 | auto &FreeCall = OMPInfoCache.RFIs[OMPRTL___kmpc_free_shared]; | |||
2664 | ||||
2665 | Function *F = getAnchorScope(); | |||
2666 | auto *HS = A.lookupAAFor<AAHeapToStack>(IRPosition::function(*F), this, | |||
2667 | DepClassTy::OPTIONAL); | |||
2668 | ||||
2669 | ChangeStatus Changed = ChangeStatus::UNCHANGED; | |||
2670 | for (CallBase *CB : MallocCalls) { | |||
2671 | // Skip replacing this if HeapToStack has already claimed it. | |||
2672 | if (HS && HS->isAssumedHeapToStack(*CB)) | |||
2673 | continue; | |||
2674 | ||||
2675 | // Find the unique free call to remove it. | |||
2676 | SmallVector<CallBase *, 4> FreeCalls; | |||
2677 | for (auto *U : CB->users()) { | |||
2678 | CallBase *C = dyn_cast<CallBase>(U); | |||
2679 | if (C && C->getCalledFunction() == FreeCall.Declaration) | |||
2680 | FreeCalls.push_back(C); | |||
2681 | } | |||
2682 | if (FreeCalls.size() != 1) | |||
2683 | continue; | |||
2684 | ||||
2685 | ConstantInt *AllocSize = dyn_cast<ConstantInt>(CB->getArgOperand(0)); | |||
2686 | ||||
2687 | LLVM_DEBUG(dbgs() << TAG << "Replace globalization call in "do { } while (false) | |||
2688 | << CB->getCaller()->getName() << " with "do { } while (false) | |||
2689 | << AllocSize->getZExtValue()do { } while (false) | |||
2690 | << " bytes of shared memory\n")do { } while (false); | |||
2691 | ||||
2692 | // Create a new shared memory buffer of the same size as the allocation | |||
2693 | // and replace all the uses of the original allocation with it. | |||
2694 | Module *M = CB->getModule(); | |||
2695 | Type *Int8Ty = Type::getInt8Ty(M->getContext()); | |||
2696 | Type *Int8ArrTy = ArrayType::get(Int8Ty, AllocSize->getZExtValue()); | |||
2697 | auto *SharedMem = new GlobalVariable( | |||
2698 | *M, Int8ArrTy, /* IsConstant */ false, GlobalValue::InternalLinkage, | |||
2699 | UndefValue::get(Int8ArrTy), CB->getName(), nullptr, | |||
2700 | GlobalValue::NotThreadLocal, | |||
2701 | static_cast<unsigned>(AddressSpace::Shared)); | |||
2702 | auto *NewBuffer = | |||
2703 | ConstantExpr::getPointerCast(SharedMem, Int8Ty->getPointerTo()); | |||
2704 | ||||
2705 | auto Remark = [&](OptimizationRemark OR) { | |||
2706 | return OR << "Replaced globalized variable with " | |||
2707 | << ore::NV("SharedMemory", AllocSize->getZExtValue()) | |||
2708 | << ((AllocSize->getZExtValue() != 1) ? " bytes " : " byte ") | |||
2709 | << "of shared memory."; | |||
2710 | }; | |||
2711 | A.emitRemark<OptimizationRemark>(CB, "OMP111", Remark); | |||
2712 | ||||
2713 | SharedMem->setAlignment(MaybeAlign(32)); | |||
2714 | ||||
2715 | A.changeValueAfterManifest(*CB, *NewBuffer); | |||
2716 | A.deleteAfterManifest(*CB); | |||
2717 | A.deleteAfterManifest(*FreeCalls.front()); | |||
2718 | ||||
2719 | NumBytesMovedToSharedMemory += AllocSize->getZExtValue(); | |||
2720 | Changed = ChangeStatus::CHANGED; | |||
2721 | } | |||
2722 | ||||
2723 | return Changed; | |||
2724 | } | |||
2725 | ||||
2726 | ChangeStatus updateImpl(Attributor &A) override { | |||
2727 | auto &OMPInfoCache = static_cast<OMPInformationCache &>(A.getInfoCache()); | |||
2728 | auto &RFI = OMPInfoCache.RFIs[OMPRTL___kmpc_alloc_shared]; | |||
2729 | Function *F = getAnchorScope(); | |||
2730 | ||||
2731 | auto NumMallocCalls = MallocCalls.size(); | |||
2732 | ||||
2733 | // Only consider malloc calls executed by a single thread with a constant. | |||
2734 | for (User *U : RFI.Declaration->users()) { | |||
2735 | const auto &ED = A.getAAFor<AAExecutionDomain>( | |||
2736 | *this, IRPosition::function(*F), DepClassTy::REQUIRED); | |||
2737 | if (CallBase *CB = dyn_cast<CallBase>(U)) | |||
2738 | if (!dyn_cast<ConstantInt>(CB->getArgOperand(0)) || | |||
2739 | !ED.isExecutedByInitialThreadOnly(*CB)) | |||
2740 | MallocCalls.erase(CB); | |||
2741 | } | |||
2742 | ||||
2743 | findPotentialRemovedFreeCalls(A); | |||
2744 | ||||
2745 | if (NumMallocCalls != MallocCalls.size()) | |||
2746 | return ChangeStatus::CHANGED; | |||
2747 | ||||
2748 | return ChangeStatus::UNCHANGED; | |||
2749 | } | |||
2750 | ||||
2751 | /// Collection of all malloc calls in a function. | |||
2752 | SmallPtrSet<CallBase *, 4> MallocCalls; | |||
2753 | /// Collection of potentially removed free calls in a function. | |||
2754 | SmallPtrSet<CallBase *, 4> PotentialRemovedFreeCalls; | |||
2755 | }; | |||
2756 | ||||
2757 | struct AAKernelInfo : public StateWrapper<KernelInfoState, AbstractAttribute> { | |||
2758 | using Base = StateWrapper<KernelInfoState, AbstractAttribute>; | |||
2759 | AAKernelInfo(const IRPosition &IRP, Attributor &A) : Base(IRP) {} | |||
2760 | ||||
2761 | /// Statistics are tracked as part of manifest for now. | |||
2762 | void trackStatistics() const override {} | |||
2763 | ||||
2764 | /// See AbstractAttribute::getAsStr() | |||
2765 | const std::string getAsStr() const override { | |||
2766 | if (!isValidState()) | |||
2767 | return "<invalid>"; | |||
2768 | return std::string(SPMDCompatibilityTracker.isAssumed() ? "SPMD" | |||
2769 | : "generic") + | |||
2770 | std::string(SPMDCompatibilityTracker.isAtFixpoint() ? " [FIX]" | |||
2771 | : "") + | |||
2772 | std::string(" #PRs: ") + | |||
2773 | std::to_string(ReachedKnownParallelRegions.size()) + | |||
2774 | ", #Unknown PRs: " + | |||
2775 | std::to_string(ReachedUnknownParallelRegions.size()); | |||
2776 | } | |||
2777 | ||||
2778 | /// Create an abstract attribute biew for the position \p IRP. | |||
2779 | static AAKernelInfo &createForPosition(const IRPosition &IRP, Attributor &A); | |||
2780 | ||||
2781 | /// See AbstractAttribute::getName() | |||
2782 | const std::string getName() const override { return "AAKernelInfo"; } | |||
2783 | ||||
2784 | /// See AbstractAttribute::getIdAddr() | |||
2785 | const char *getIdAddr() const override { return &ID; } | |||
2786 | ||||
2787 | /// This function should return true if the type of the \p AA is AAKernelInfo | |||
2788 | static bool classof(const AbstractAttribute *AA) { | |||
2789 | return (AA->getIdAddr() == &ID); | |||
2790 | } | |||
2791 | ||||
2792 | static const char ID; | |||
2793 | }; | |||
2794 | ||||
2795 | /// The function kernel info abstract attribute, basically, what can we say | |||
2796 | /// about a function with regards to the KernelInfoState. | |||
2797 | struct AAKernelInfoFunction : AAKernelInfo { | |||
2798 | AAKernelInfoFunction(const IRPosition &IRP, Attributor &A) | |||
2799 | : AAKernelInfo(IRP, A) {} | |||
2800 | ||||
2801 | /// See AbstractAttribute::initialize(...). | |||
2802 | void initialize(Attributor &A) override { | |||
2803 | // This is a high-level transform that might change the constant arguments | |||
2804 | // of the init and dinit calls. We need to tell the Attributor about this | |||
2805 | // to avoid other parts using the current constant value for simpliication. | |||
2806 | auto &OMPInfoCache = static_cast<OMPInformationCache &>(A.getInfoCache()); | |||
2807 | ||||
2808 | Function *Fn = getAnchorScope(); | |||
2809 | if (!OMPInfoCache.Kernels.count(Fn)) | |||
2810 | return; | |||
2811 | ||||
2812 | // Add itself to the reaching kernel and set IsKernelEntry. | |||
2813 | ReachingKernelEntries.insert(Fn); | |||
2814 | IsKernelEntry = true; | |||
2815 | ||||
2816 | OMPInformationCache::RuntimeFunctionInfo &InitRFI = | |||
2817 | OMPInfoCache.RFIs[OMPRTL___kmpc_target_init]; | |||
2818 | OMPInformationCache::RuntimeFunctionInfo &DeinitRFI = | |||
2819 | OMPInfoCache.RFIs[OMPRTL___kmpc_target_deinit]; | |||
2820 | ||||
2821 | // For kernels we perform more initialization work, first we find the init | |||
2822 | // and deinit calls. | |||
2823 | auto StoreCallBase = [](Use &U, | |||
2824 | OMPInformationCache::RuntimeFunctionInfo &RFI, | |||
2825 | CallBase *&Storage) { | |||
2826 | CallBase *CB = OpenMPOpt::getCallIfRegularCall(U, &RFI); | |||
2827 | assert(CB &&((void)0) | |||
2828 | "Unexpected use of __kmpc_target_init or __kmpc_target_deinit!")((void)0); | |||
2829 | assert(!Storage &&((void)0) | |||
2830 | "Multiple uses of __kmpc_target_init or __kmpc_target_deinit!")((void)0); | |||
2831 | Storage = CB; | |||
2832 | return false; | |||
2833 | }; | |||
2834 | InitRFI.foreachUse( | |||
2835 | [&](Use &U, Function &) { | |||
2836 | StoreCallBase(U, InitRFI, KernelInitCB); | |||
2837 | return false; | |||
2838 | }, | |||
2839 | Fn); | |||
2840 | DeinitRFI.foreachUse( | |||
2841 | [&](Use &U, Function &) { | |||
2842 | StoreCallBase(U, DeinitRFI, KernelDeinitCB); | |||
2843 | return false; | |||
2844 | }, | |||
2845 | Fn); | |||
2846 | ||||
2847 | assert((KernelInitCB && KernelDeinitCB) &&((void)0) | |||
2848 | "Kernel without __kmpc_target_init or __kmpc_target_deinit!")((void)0); | |||
2849 | ||||
2850 | // For kernels we might need to initialize/finalize the IsSPMD state and | |||
2851 | // we need to register a simplification callback so that the Attributor | |||
2852 | // knows the constant arguments to __kmpc_target_init and | |||
2853 | // __kmpc_target_deinit might actually change. | |||
2854 | ||||
2855 | Attributor::SimplifictionCallbackTy StateMachineSimplifyCB = | |||
2856 | [&](const IRPosition &IRP, const AbstractAttribute *AA, | |||
2857 | bool &UsedAssumedInformation) -> Optional<Value *> { | |||
2858 | // IRP represents the "use generic state machine" argument of an | |||
2859 | // __kmpc_target_init call. We will answer this one with the internal | |||
2860 | // state. As long as we are not in an invalid state, we will create a | |||
2861 | // custom state machine so the value should be a `i1 false`. If we are | |||
2862 | // in an invalid state, we won't change the value that is in the IR. | |||
2863 | if (!isValidState()) | |||
2864 | return nullptr; | |||
2865 | if (AA) | |||
2866 | A.recordDependence(*this, *AA, DepClassTy::OPTIONAL); | |||
2867 | UsedAssumedInformation = !isAtFixpoint(); | |||
2868 | auto *FalseVal = | |||
2869 | ConstantInt::getBool(IRP.getAnchorValue().getContext(), 0); | |||
2870 | return FalseVal; | |||
2871 | }; | |||
2872 | ||||
2873 | Attributor::SimplifictionCallbackTy IsSPMDModeSimplifyCB = | |||
2874 | [&](const IRPosition &IRP, const AbstractAttribute *AA, | |||
2875 | bool &UsedAssumedInformation) -> Optional<Value *> { | |||
2876 | // IRP represents the "SPMDCompatibilityTracker" argument of an | |||
2877 | // __kmpc_target_init or | |||
2878 | // __kmpc_target_deinit call. We will answer this one with the internal | |||
2879 | // state. | |||
2880 | if (!SPMDCompatibilityTracker.isValidState()) | |||
2881 | return nullptr; | |||
2882 | if (!SPMDCompatibilityTracker.isAtFixpoint()) { | |||
2883 | if (AA) | |||
2884 | A.recordDependence(*this, *AA, DepClassTy::OPTIONAL); | |||
2885 | UsedAssumedInformation = true; | |||
2886 | } else { | |||
2887 | UsedAssumedInformation = false; | |||
2888 | } | |||
2889 | auto *Val = ConstantInt::getBool(IRP.getAnchorValue().getContext(), | |||
2890 | SPMDCompatibilityTracker.isAssumed()); | |||
2891 | return Val; | |||
2892 | }; | |||
2893 | ||||
2894 | Attributor::SimplifictionCallbackTy IsGenericModeSimplifyCB = | |||
2895 | [&](const IRPosition &IRP, const AbstractAttribute *AA, | |||
2896 | bool &UsedAssumedInformation) -> Optional<Value *> { | |||
2897 | // IRP represents the "RequiresFullRuntime" argument of an | |||
2898 | // __kmpc_target_init or __kmpc_target_deinit call. We will answer this | |||
2899 | // one with the internal state of the SPMDCompatibilityTracker, so if | |||
2900 | // generic then true, if SPMD then false. | |||
2901 | if (!SPMDCompatibilityTracker.isValidState()) | |||
2902 | return nullptr; | |||
2903 | if (!SPMDCompatibilityTracker.isAtFixpoint()) { | |||
2904 | if (AA) | |||
2905 | A.recordDependence(*this, *AA, DepClassTy::OPTIONAL); | |||
2906 | UsedAssumedInformation = true; | |||
2907 | } else { | |||
2908 | UsedAssumedInformation = false; | |||
2909 | } | |||
2910 | auto *Val = ConstantInt::getBool(IRP.getAnchorValue().getContext(), | |||
2911 | !SPMDCompatibilityTracker.isAssumed()); | |||
2912 | return Val; | |||
2913 | }; | |||
2914 | ||||
2915 | constexpr const int InitIsSPMDArgNo = 1; | |||
2916 | constexpr const int DeinitIsSPMDArgNo = 1; | |||
2917 | constexpr const int InitUseStateMachineArgNo = 2; | |||
2918 | constexpr const int InitRequiresFullRuntimeArgNo = 3; | |||
2919 | constexpr const int DeinitRequiresFullRuntimeArgNo = 2; | |||
2920 | A.registerSimplificationCallback( | |||
2921 | IRPosition::callsite_argument(*KernelInitCB, InitUseStateMachineArgNo), | |||
2922 | StateMachineSimplifyCB); | |||
2923 | A.registerSimplificationCallback( | |||
2924 | IRPosition::callsite_argument(*KernelInitCB, InitIsSPMDArgNo), | |||
2925 | IsSPMDModeSimplifyCB); | |||
2926 | A.registerSimplificationCallback( | |||
2927 | IRPosition::callsite_argument(*KernelDeinitCB, DeinitIsSPMDArgNo), | |||
2928 | IsSPMDModeSimplifyCB); | |||
2929 | A.registerSimplificationCallback( | |||
2930 | IRPosition::callsite_argument(*KernelInitCB, | |||
2931 | InitRequiresFullRuntimeArgNo), | |||
2932 | IsGenericModeSimplifyCB); | |||
2933 | A.registerSimplificationCallback( | |||
2934 | IRPosition::callsite_argument(*KernelDeinitCB, | |||
2935 | DeinitRequiresFullRuntimeArgNo), | |||
2936 | IsGenericModeSimplifyCB); | |||
2937 | ||||
2938 | // Check if we know we are in SPMD-mode already. | |||
2939 | ConstantInt *IsSPMDArg = | |||
2940 | dyn_cast<ConstantInt>(KernelInitCB->getArgOperand(InitIsSPMDArgNo)); | |||
2941 | if (IsSPMDArg && !IsSPMDArg->isZero()) | |||
2942 | SPMDCompatibilityTracker.indicateOptimisticFixpoint(); | |||
2943 | } | |||
2944 | ||||
2945 | /// Modify the IR based on the KernelInfoState as the fixpoint iteration is | |||
2946 | /// finished now. | |||
2947 | ChangeStatus manifest(Attributor &A) override { | |||
2948 | // If we are not looking at a kernel with __kmpc_target_init and | |||
2949 | // __kmpc_target_deinit call we cannot actually manifest the information. | |||
2950 | if (!KernelInitCB || !KernelDeinitCB) | |||
2951 | return ChangeStatus::UNCHANGED; | |||
2952 | ||||
2953 | // Known SPMD-mode kernels need no manifest changes. | |||
2954 | if (SPMDCompatibilityTracker.isKnown()) | |||
2955 | return ChangeStatus::UNCHANGED; | |||
2956 | ||||
2957 | // If we can we change the execution mode to SPMD-mode otherwise we build a | |||
2958 | // custom state machine. | |||
2959 | if (!changeToSPMDMode(A)) | |||
2960 | buildCustomStateMachine(A); | |||
2961 | ||||
2962 | return ChangeStatus::CHANGED; | |||
2963 | } | |||
2964 | ||||
2965 | bool changeToSPMDMode(Attributor &A) { | |||
2966 | auto &OMPInfoCache = static_cast<OMPInformationCache &>(A.getInfoCache()); | |||
2967 | ||||
2968 | if (!SPMDCompatibilityTracker.isAssumed()) { | |||
2969 | for (Instruction *NonCompatibleI : SPMDCompatibilityTracker) { | |||
2970 | if (!NonCompatibleI) | |||
2971 | continue; | |||
2972 | ||||
2973 | // Skip diagnostics on calls to known OpenMP runtime functions for now. | |||
2974 | if (auto *CB = dyn_cast<CallBase>(NonCompatibleI)) | |||
2975 | if (OMPInfoCache.RTLFunctions.contains(CB->getCalledFunction())) | |||
2976 | continue; | |||
2977 | ||||
2978 | auto Remark = [&](OptimizationRemarkAnalysis ORA) { | |||
2979 | ORA << "Value has potential side effects preventing SPMD-mode " | |||
2980 | "execution"; | |||
2981 | if (isa<CallBase>(NonCompatibleI)) { | |||
2982 | ORA << ". Add `__attribute__((assume(\"ompx_spmd_amenable\")))` to " | |||
2983 | "the called function to override"; | |||
2984 | } | |||
2985 | return ORA << "."; | |||
2986 | }; | |||
2987 | A.emitRemark<OptimizationRemarkAnalysis>(NonCompatibleI, "OMP121", | |||
2988 | Remark); | |||
2989 | ||||
2990 | LLVM_DEBUG(dbgs() << TAG << "SPMD-incompatible side-effect: "do { } while (false) | |||
2991 | << *NonCompatibleI << "\n")do { } while (false); | |||
2992 | } | |||
2993 | ||||
2994 | return false; | |||
2995 | } | |||
2996 | ||||
2997 | // Adjust the global exec mode flag that tells the runtime what mode this | |||
2998 | // kernel is executed in. | |||
2999 | Function *Kernel = getAnchorScope(); | |||
3000 | GlobalVariable *ExecMode = Kernel->getParent()->getGlobalVariable( | |||
3001 | (Kernel->getName() + "_exec_mode").str()); | |||
3002 | assert(ExecMode && "Kernel without exec mode?")((void)0); | |||
3003 | assert(ExecMode->getInitializer() &&((void)0) | |||
3004 | ExecMode->getInitializer()->isOneValue() &&((void)0) | |||
3005 | "Initially non-SPMD kernel has SPMD exec mode!")((void)0); | |||
3006 | ||||
3007 | // Set the global exec mode flag to indicate SPMD-Generic mode. | |||
3008 | constexpr int SPMDGeneric = 2; | |||
3009 | if (!ExecMode->getInitializer()->isZeroValue()) | |||
3010 | ExecMode->setInitializer( | |||
3011 | ConstantInt::get(ExecMode->getInitializer()->getType(), SPMDGeneric)); | |||
3012 | ||||
3013 | // Next rewrite the init and deinit calls to indicate we use SPMD-mode now. | |||
3014 | const int InitIsSPMDArgNo = 1; | |||
3015 | const int DeinitIsSPMDArgNo = 1; | |||
3016 | const int InitUseStateMachineArgNo = 2; | |||
3017 | const int InitRequiresFullRuntimeArgNo = 3; | |||
3018 | const int DeinitRequiresFullRuntimeArgNo = 2; | |||
3019 | ||||
3020 | auto &Ctx = getAnchorValue().getContext(); | |||
3021 | A.changeUseAfterManifest(KernelInitCB->getArgOperandUse(InitIsSPMDArgNo), | |||
3022 | *ConstantInt::getBool(Ctx, 1)); | |||
3023 | A.changeUseAfterManifest( | |||
3024 | KernelInitCB->getArgOperandUse(InitUseStateMachineArgNo), | |||
3025 | *ConstantInt::getBool(Ctx, 0)); | |||
3026 | A.changeUseAfterManifest( | |||
3027 | KernelDeinitCB->getArgOperandUse(DeinitIsSPMDArgNo), | |||
3028 | *ConstantInt::getBool(Ctx, 1)); | |||
3029 | A.changeUseAfterManifest( | |||
3030 | KernelInitCB->getArgOperandUse(InitRequiresFullRuntimeArgNo), | |||
3031 | *ConstantInt::getBool(Ctx, 0)); | |||
3032 | A.changeUseAfterManifest( | |||
3033 | KernelDeinitCB->getArgOperandUse(DeinitRequiresFullRuntimeArgNo), | |||
3034 | *ConstantInt::getBool(Ctx, 0)); | |||
3035 | ||||
3036 | ++NumOpenMPTargetRegionKernelsSPMD; | |||
3037 | ||||
3038 | auto Remark = [&](OptimizationRemark OR) { | |||
3039 | return OR << "Transformed generic-mode kernel to SPMD-mode."; | |||
3040 | }; | |||
3041 | A.emitRemark<OptimizationRemark>(KernelInitCB, "OMP120", Remark); | |||
3042 | return true; | |||
3043 | }; | |||
3044 | ||||
3045 | ChangeStatus buildCustomStateMachine(Attributor &A) { | |||
3046 | assert(ReachedKnownParallelRegions.isValidState() &&((void)0) | |||
3047 | "Custom state machine with invalid parallel region states?")((void)0); | |||
3048 | ||||
3049 | const int InitIsSPMDArgNo = 1; | |||
3050 | const int InitUseStateMachineArgNo = 2; | |||
3051 | ||||
3052 | // Check if the current configuration is non-SPMD and generic state machine. | |||
3053 | // If we already have SPMD mode or a custom state machine we do not need to | |||
3054 | // go any further. If it is anything but a constant something is weird and | |||
3055 | // we give up. | |||
3056 | ConstantInt *UseStateMachine = dyn_cast<ConstantInt>( | |||
3057 | KernelInitCB->getArgOperand(InitUseStateMachineArgNo)); | |||
3058 | ConstantInt *IsSPMD = | |||
3059 | dyn_cast<ConstantInt>(KernelInitCB->getArgOperand(InitIsSPMDArgNo)); | |||
3060 | ||||
3061 | // If we are stuck with generic mode, try to create a custom device (=GPU) | |||
3062 | // state machine which is specialized for the parallel regions that are | |||
3063 | // reachable by the kernel. | |||
3064 | if (!UseStateMachine || UseStateMachine->isZero() || !IsSPMD || | |||
3065 | !IsSPMD->isZero()) | |||
3066 | return ChangeStatus::UNCHANGED; | |||
3067 | ||||
3068 | // If not SPMD mode, indicate we use a custom state machine now. | |||
3069 | auto &Ctx = getAnchorValue().getContext(); | |||
3070 | auto *FalseVal = ConstantInt::getBool(Ctx, 0); | |||
3071 | A.changeUseAfterManifest( | |||
3072 | KernelInitCB->getArgOperandUse(InitUseStateMachineArgNo), *FalseVal); | |||
3073 | ||||
3074 | // If we don't actually need a state machine we are done here. This can | |||
3075 | // happen if there simply are no parallel regions. In the resulting kernel | |||
3076 | // all worker threads will simply exit right away, leaving the main thread | |||
3077 | // to do the work alone. | |||
3078 | if (ReachedKnownParallelRegions.empty() && | |||
3079 | ReachedUnknownParallelRegions.empty()) { | |||
3080 | ++NumOpenMPTargetRegionKernelsWithoutStateMachine; | |||
3081 | ||||
3082 | auto Remark = [&](OptimizationRemark OR) { | |||
3083 | return OR << "Removing unused state machine from generic-mode kernel."; | |||
3084 | }; | |||
3085 | A.emitRemark<OptimizationRemark>(KernelInitCB, "OMP130", Remark); | |||
3086 | ||||
3087 | return ChangeStatus::CHANGED; | |||
3088 | } | |||
3089 | ||||
3090 | // Keep track in the statistics of our new shiny custom state machine. | |||
3091 | if (ReachedUnknownParallelRegions.empty()) { | |||
3092 | ++NumOpenMPTargetRegionKernelsCustomStateMachineWithoutFallback; | |||
3093 | ||||
3094 | auto Remark = [&](OptimizationRemark OR) { | |||
3095 | return OR << "Rewriting generic-mode kernel with a customized state " | |||
3096 | "machine."; | |||
3097 | }; | |||
3098 | A.emitRemark<OptimizationRemark>(KernelInitCB, "OMP131", Remark); | |||
3099 | } else { | |||
3100 | ++NumOpenMPTargetRegionKernelsCustomStateMachineWithFallback; | |||
3101 | ||||
3102 | auto Remark = [&](OptimizationRemarkAnalysis OR) { | |||
3103 | return OR << "Generic-mode kernel is executed with a customized state " | |||
3104 | "machine that requires a fallback."; | |||
3105 | }; | |||
3106 | A.emitRemark<OptimizationRemarkAnalysis>(KernelInitCB, "OMP132", Remark); | |||
3107 | ||||
3108 | // Tell the user why we ended up with a fallback. | |||
3109 | for (CallBase *UnknownParallelRegionCB : ReachedUnknownParallelRegions) { | |||
3110 | if (!UnknownParallelRegionCB) | |||
3111 | continue; | |||
3112 | auto Remark = [&](OptimizationRemarkAnalysis ORA) { | |||
3113 | return ORA << "Call may contain unknown parallel regions. Use " | |||
3114 | << "`__attribute__((assume(\"omp_no_parallelism\")))` to " | |||
3115 | "override."; | |||
3116 | }; | |||
3117 | A.emitRemark<OptimizationRemarkAnalysis>(UnknownParallelRegionCB, | |||
3118 | "OMP133", Remark); | |||
3119 | } | |||
3120 | } | |||
3121 | ||||
3122 | // Create all the blocks: | |||
3123 | // | |||
3124 | // InitCB = __kmpc_target_init(...) | |||
3125 | // bool IsWorker = InitCB >= 0; | |||
3126 | // if (IsWorker) { | |||
3127 | // SMBeginBB: __kmpc_barrier_simple_spmd(...); | |||
3128 | // void *WorkFn; | |||
3129 | // bool Active = __kmpc_kernel_parallel(&WorkFn); | |||
3130 | // if (!WorkFn) return; | |||
3131 | // SMIsActiveCheckBB: if (Active) { | |||
3132 | // SMIfCascadeCurrentBB: if (WorkFn == <ParFn0>) | |||
3133 | // ParFn0(...); | |||
3134 | // SMIfCascadeCurrentBB: else if (WorkFn == <ParFn1>) | |||
3135 | // ParFn1(...); | |||
3136 | // ... | |||
3137 | // SMIfCascadeCurrentBB: else | |||
3138 | // ((WorkFnTy*)WorkFn)(...); | |||
3139 | // SMEndParallelBB: __kmpc_kernel_end_parallel(...); | |||
3140 | // } | |||
3141 | // SMDoneBB: __kmpc_barrier_simple_spmd(...); | |||
3142 | // goto SMBeginBB; | |||
3143 | // } | |||
3144 | // UserCodeEntryBB: // user code | |||
3145 | // __kmpc_target_deinit(...) | |||
3146 | // | |||
3147 | Function *Kernel = getAssociatedFunction(); | |||
3148 | assert(Kernel && "Expected an associated function!")((void)0); | |||
3149 | ||||
3150 | BasicBlock *InitBB = KernelInitCB->getParent(); | |||
3151 | BasicBlock *UserCodeEntryBB = InitBB->splitBasicBlock( | |||
3152 | KernelInitCB->getNextNode(), "thread.user_code.check"); | |||
3153 | BasicBlock *StateMachineBeginBB = BasicBlock::Create( | |||
3154 | Ctx, "worker_state_machine.begin", Kernel, UserCodeEntryBB); | |||
3155 | BasicBlock *StateMachineFinishedBB = BasicBlock::Create( | |||
3156 | Ctx, "worker_state_machine.finished", Kernel, UserCodeEntryBB); | |||
3157 | BasicBlock *StateMachineIsActiveCheckBB = BasicBlock::Create( | |||
3158 | Ctx, "worker_state_machine.is_active.check", Kernel, UserCodeEntryBB); | |||
3159 | BasicBlock *StateMachineIfCascadeCurrentBB = | |||
3160 | BasicBlock::Create(Ctx, "worker_state_machine.parallel_region.check", | |||
3161 | Kernel, UserCodeEntryBB); | |||
3162 | BasicBlock *StateMachineEndParallelBB = | |||
3163 | BasicBlock::Create(Ctx, "worker_state_machine.parallel_region.end", | |||
3164 | Kernel, UserCodeEntryBB); | |||
3165 | BasicBlock *StateMachineDoneBarrierBB = BasicBlock::Create( | |||
3166 | Ctx, "worker_state_machine.done.barrier", Kernel, UserCodeEntryBB); | |||
3167 | A.registerManifestAddedBasicBlock(*InitBB); | |||
3168 | A.registerManifestAddedBasicBlock(*UserCodeEntryBB); | |||
3169 | A.registerManifestAddedBasicBlock(*StateMachineBeginBB); | |||
3170 | A.registerManifestAddedBasicBlock(*StateMachineFinishedBB); | |||
3171 | A.registerManifestAddedBasicBlock(*StateMachineIsActiveCheckBB); | |||
3172 | A.registerManifestAddedBasicBlock(*StateMachineIfCascadeCurrentBB); | |||
3173 | A.registerManifestAddedBasicBlock(*StateMachineEndParallelBB); | |||
3174 | A.registerManifestAddedBasicBlock(*StateMachineDoneBarrierBB); | |||
3175 | ||||
3176 | const DebugLoc &DLoc = KernelInitCB->getDebugLoc(); | |||
3177 | ReturnInst::Create(Ctx, StateMachineFinishedBB)->setDebugLoc(DLoc); | |||
3178 | ||||
3179 | InitBB->getTerminator()->eraseFromParent(); | |||
3180 | Instruction *IsWorker = | |||
3181 | ICmpInst::Create(ICmpInst::ICmp, llvm::CmpInst::ICMP_NE, KernelInitCB, | |||
3182 | ConstantInt::get(KernelInitCB->getType(), -1), | |||
3183 | "thread.is_worker", InitBB); | |||
3184 | IsWorker->setDebugLoc(DLoc); | |||
3185 | BranchInst::Create(StateMachineBeginBB, UserCodeEntryBB, IsWorker, InitBB); | |||
3186 | ||||
3187 | Module &M = *Kernel->getParent(); | |||
3188 | ||||
3189 | // Create local storage for the work function pointer. | |||
3190 | const DataLayout &DL = M.getDataLayout(); | |||
3191 | Type *VoidPtrTy = Type::getInt8PtrTy(Ctx); | |||
3192 | Instruction *WorkFnAI = | |||
3193 | new AllocaInst(VoidPtrTy, DL.getAllocaAddrSpace(), nullptr, | |||
3194 | "worker.work_fn.addr", &Kernel->getEntryBlock().front()); | |||
3195 | WorkFnAI->setDebugLoc(DLoc); | |||
3196 | ||||
3197 | auto &OMPInfoCache = static_cast<OMPInformationCache &>(A.getInfoCache()); | |||
3198 | OMPInfoCache.OMPBuilder.updateToLocation( | |||
3199 | OpenMPIRBuilder::LocationDescription( | |||
3200 | IRBuilder<>::InsertPoint(StateMachineBeginBB, | |||
3201 | StateMachineBeginBB->end()), | |||
3202 | DLoc)); | |||
3203 | ||||
3204 | Value *Ident = KernelInitCB->getArgOperand(0); | |||
3205 | Value *GTid = KernelInitCB; | |||
3206 | ||||
3207 | FunctionCallee BarrierFn = | |||
3208 | OMPInfoCache.OMPBuilder.getOrCreateRuntimeFunction( | |||
3209 | M, OMPRTL___kmpc_barrier_simple_spmd); | |||
3210 | CallInst::Create(BarrierFn, {Ident, GTid}, "", StateMachineBeginBB) | |||
3211 | ->setDebugLoc(DLoc); | |||
3212 | ||||
3213 | if (WorkFnAI->getType()->getPointerAddressSpace() != | |||
3214 | (unsigned int)AddressSpace::Generic) { | |||
3215 | WorkFnAI = new AddrSpaceCastInst( | |||
3216 | WorkFnAI, | |||
3217 | PointerType::getWithSamePointeeType( | |||
3218 | cast<PointerType>(WorkFnAI->getType()), | |||
3219 | (unsigned int)AddressSpace::Generic), | |||
3220 | WorkFnAI->getName() + ".generic", StateMachineBeginBB); | |||
3221 | WorkFnAI->setDebugLoc(DLoc); | |||
3222 | } | |||
3223 | ||||
3224 | FunctionCallee KernelParallelFn = | |||
3225 | OMPInfoCache.OMPBuilder.getOrCreateRuntimeFunction( | |||
3226 | M, OMPRTL___kmpc_kernel_parallel); | |||
3227 | Instruction *IsActiveWorker = CallInst::Create( | |||
3228 | KernelParallelFn, {WorkFnAI}, "worker.is_active", StateMachineBeginBB); | |||
3229 | IsActiveWorker->setDebugLoc(DLoc); | |||
3230 | Instruction *WorkFn = new LoadInst(VoidPtrTy, WorkFnAI, "worker.work_fn", | |||
3231 | StateMachineBeginBB); | |||
3232 | WorkFn->setDebugLoc(DLoc); | |||
3233 | ||||
3234 | FunctionType *ParallelRegionFnTy = FunctionType::get( | |||
3235 | Type::getVoidTy(Ctx), {Type::getInt16Ty(Ctx), Type::getInt32Ty(Ctx)}, | |||
3236 | false); | |||
3237 | Value *WorkFnCast = BitCastInst::CreatePointerBitCastOrAddrSpaceCast( | |||
3238 | WorkFn, ParallelRegionFnTy->getPointerTo(), "worker.work_fn.addr_cast", | |||
3239 | StateMachineBeginBB); | |||
3240 | ||||
3241 | Instruction *IsDone = | |||
3242 | ICmpInst::Create(ICmpInst::ICmp, llvm::CmpInst::ICMP_EQ, WorkFn, | |||
3243 | Constant::getNullValue(VoidPtrTy), "worker.is_done", | |||
3244 | StateMachineBeginBB); | |||
3245 | IsDone->setDebugLoc(DLoc); | |||
3246 | BranchInst::Create(StateMachineFinishedBB, StateMachineIsActiveCheckBB, | |||
3247 | IsDone, StateMachineBeginBB) | |||
3248 | ->setDebugLoc(DLoc); | |||
3249 | ||||
3250 | BranchInst::Create(StateMachineIfCascadeCurrentBB, | |||
3251 | StateMachineDoneBarrierBB, IsActiveWorker, | |||
3252 | StateMachineIsActiveCheckBB) | |||
3253 | ->setDebugLoc(DLoc); | |||
3254 | ||||
3255 | Value *ZeroArg = | |||
3256 | Constant::getNullValue(ParallelRegionFnTy->getParamType(0)); | |||
3257 | ||||
3258 | // Now that we have most of the CFG skeleton it is time for the if-cascade | |||
3259 | // that checks the function pointer we got from the runtime against the | |||
3260 | // parallel regions we expect, if there are any. | |||
3261 | for (int i = 0, e = ReachedKnownParallelRegions.size(); i < e; ++i) { | |||
3262 | auto *ParallelRegion = ReachedKnownParallelRegions[i]; | |||
3263 | BasicBlock *PRExecuteBB = BasicBlock::Create( | |||
3264 | Ctx, "worker_state_machine.parallel_region.execute", Kernel, | |||
3265 | StateMachineEndParallelBB); | |||
3266 | CallInst::Create(ParallelRegion, {ZeroArg, GTid}, "", PRExecuteBB) | |||
3267 | ->setDebugLoc(DLoc); | |||
3268 | BranchInst::Create(StateMachineEndParallelBB, PRExecuteBB) | |||
3269 | ->setDebugLoc(DLoc); | |||
3270 | ||||
3271 | BasicBlock *PRNextBB = | |||
3272 | BasicBlock::Create(Ctx, "worker_state_machine.parallel_region.check", | |||
3273 | Kernel, StateMachineEndParallelBB); | |||
3274 | ||||
3275 | // Check if we need to compare the pointer at all or if we can just | |||
3276 | // call the parallel region function. | |||
3277 | Value *IsPR; | |||
3278 | if (i + 1 < e || !ReachedUnknownParallelRegions.empty()) { | |||
3279 | Instruction *CmpI = ICmpInst::Create( | |||
3280 | ICmpInst::ICmp, llvm::CmpInst::ICMP_EQ, WorkFnCast, ParallelRegion, | |||
3281 | "worker.check_parallel_region", StateMachineIfCascadeCurrentBB); | |||
3282 | CmpI->setDebugLoc(DLoc); | |||
3283 | IsPR = CmpI; | |||
3284 | } else { | |||
3285 | IsPR = ConstantInt::getTrue(Ctx); | |||
3286 | } | |||
3287 | ||||
3288 | BranchInst::Create(PRExecuteBB, PRNextBB, IsPR, | |||
3289 | StateMachineIfCascadeCurrentBB) | |||
3290 | ->setDebugLoc(DLoc); | |||
3291 | StateMachineIfCascadeCurrentBB = PRNextBB; | |||
3292 | } | |||
3293 | ||||
3294 | // At the end of the if-cascade we place the indirect function pointer call | |||
3295 | // in case we might need it, that is if there can be parallel regions we | |||
3296 | // have not handled in the if-cascade above. | |||
3297 | if (!ReachedUnknownParallelRegions.empty()) { | |||
3298 | StateMachineIfCascadeCurrentBB->setName( | |||
3299 | "worker_state_machine.parallel_region.fallback.execute"); | |||
3300 | CallInst::Create(ParallelRegionFnTy, WorkFnCast, {ZeroArg, GTid}, "", | |||
3301 | StateMachineIfCascadeCurrentBB) | |||
3302 | ->setDebugLoc(DLoc); | |||
3303 | } | |||
3304 | BranchInst::Create(StateMachineEndParallelBB, | |||
3305 | StateMachineIfCascadeCurrentBB) | |||
3306 | ->setDebugLoc(DLoc); | |||
3307 | ||||
3308 | CallInst::Create(OMPInfoCache.OMPBuilder.getOrCreateRuntimeFunction( | |||
3309 | M, OMPRTL___kmpc_kernel_end_parallel), | |||
3310 | {}, "", StateMachineEndParallelBB) | |||
3311 | ->setDebugLoc(DLoc); | |||
3312 | BranchInst::Create(StateMachineDoneBarrierBB, StateMachineEndParallelBB) | |||
3313 | ->setDebugLoc(DLoc); | |||
3314 | ||||
3315 | CallInst::Create(BarrierFn, {Ident, GTid}, "", StateMachineDoneBarrierBB) | |||
3316 | ->setDebugLoc(DLoc); | |||
3317 | BranchInst::Create(StateMachineBeginBB, StateMachineDoneBarrierBB) | |||
3318 | ->setDebugLoc(DLoc); | |||
3319 | ||||
3320 | return ChangeStatus::CHANGED; | |||
3321 | } | |||
3322 | ||||
3323 | /// Fixpoint iteration update function. Will be called every time a dependence | |||
3324 | /// changed its state (and in the beginning). | |||
3325 | ChangeStatus updateImpl(Attributor &A) override { | |||
3326 | KernelInfoState StateBefore = getState(); | |||
3327 | ||||
3328 | // Callback to check a read/write instruction. | |||
3329 | auto CheckRWInst = [&](Instruction &I) { | |||
3330 | // We handle calls later. | |||
3331 | if (isa<CallBase>(I)) | |||
3332 | return true; | |||
3333 | // We only care about write effects. | |||
3334 | if (!I.mayWriteToMemory()) | |||
3335 | return true; | |||
3336 | if (auto *SI = dyn_cast<StoreInst>(&I)) { | |||
3337 | SmallVector<const Value *> Objects; | |||
3338 | getUnderlyingObjects(SI->getPointerOperand(), Objects); | |||
3339 | if (llvm::all_of(Objects, | |||
3340 | [](const Value *Obj) { return isa<AllocaInst>(Obj); })) | |||
3341 | return true; | |||
3342 | } | |||
3343 | // For now we give up on everything but stores. | |||
3344 | SPMDCompatibilityTracker.insert(&I); | |||
3345 | return true; | |||
3346 | }; | |||
3347 | ||||
3348 | bool UsedAssumedInformationInCheckRWInst = false; | |||
3349 | if (!SPMDCompatibilityTracker.isAtFixpoint()) | |||
3350 | if (!A.checkForAllReadWriteInstructions( | |||
3351 | CheckRWInst, *this, UsedAssumedInformationInCheckRWInst)) | |||
3352 | SPMDCompatibilityTracker.indicatePessimisticFixpoint(); | |||
3353 | ||||
3354 | if (!IsKernelEntry) { | |||
3355 | updateReachingKernelEntries(A); | |||
3356 | updateParallelLevels(A); | |||
3357 | } | |||
3358 | ||||
3359 | // Callback to check a call instruction. | |||
3360 | bool AllSPMDStatesWereFixed = true; | |||
3361 | auto CheckCallInst = [&](Instruction &I) { | |||
3362 | auto &CB = cast<CallBase>(I); | |||
3363 | auto &CBAA = A.getAAFor<AAKernelInfo>( | |||
3364 | *this, IRPosition::callsite_function(CB), DepClassTy::OPTIONAL); | |||
3365 | getState() ^= CBAA.getState(); | |||
3366 | AllSPMDStatesWereFixed &= CBAA.SPMDCompatibilityTracker.isAtFixpoint(); | |||
3367 | return true; | |||
3368 | }; | |||
3369 | ||||
3370 | bool UsedAssumedInformationInCheckCallInst = false; | |||
3371 | if (!A.checkForAllCallLikeInstructions( | |||
3372 | CheckCallInst, *this, UsedAssumedInformationInCheckCallInst)) | |||
3373 | return indicatePessimisticFixpoint(); | |||
3374 | ||||
3375 | // If we haven't used any assumed information for the SPMD state we can fix | |||
3376 | // it. | |||
3377 | if (!UsedAssumedInformationInCheckRWInst && | |||
3378 | !UsedAssumedInformationInCheckCallInst && AllSPMDStatesWereFixed) | |||
3379 | SPMDCompatibilityTracker.indicateOptimisticFixpoint(); | |||
3380 | ||||
3381 | return StateBefore == getState() ? ChangeStatus::UNCHANGED | |||
3382 | : ChangeStatus::CHANGED; | |||
3383 | } | |||
3384 | ||||
3385 | private: | |||
3386 | /// Update info regarding reaching kernels. | |||
3387 | void updateReachingKernelEntries(Attributor &A) { | |||
3388 | auto PredCallSite = [&](AbstractCallSite ACS) { | |||
3389 | Function *Caller = ACS.getInstruction()->getFunction(); | |||
3390 | ||||
3391 | assert(Caller && "Caller is nullptr")((void)0); | |||
3392 | ||||
3393 | auto &CAA = A.getOrCreateAAFor<AAKernelInfo>( | |||
3394 | IRPosition::function(*Caller), this, DepClassTy::REQUIRED); | |||
3395 | if (CAA.ReachingKernelEntries.isValidState()) { | |||
| ||||
3396 | ReachingKernelEntries ^= CAA.ReachingKernelEntries; | |||
3397 | return true; | |||
3398 | } | |||
3399 | ||||
3400 | // We lost track of the caller of the associated function, any kernel | |||
3401 | // could reach now. | |||
3402 | ReachingKernelEntries.indicatePessimisticFixpoint(); | |||
| ||||
3403 | ||||
3404 | return true; | |||
3405 | }; | |||
3406 | ||||
3407 | bool AllCallSitesKnown; | |||
3408 | if (!A.checkForAllCallSites(PredCallSite, *this, | |||
3409 | true /* RequireAllCallSites */, | |||
3410 | AllCallSitesKnown)) | |||
3411 | ReachingKernelEntries.indicatePessimisticFixpoint(); | |||
3412 | } | |||
3413 | ||||
3414 | /// Update info regarding parallel levels. | |||
3415 | void updateParallelLevels(Attributor &A) { | |||
3416 | auto &OMPInfoCache = static_cast<OMPInformationCache &>(A.getInfoCache()); | |||
3417 | OMPInformationCache::RuntimeFunctionInfo &Parallel51RFI = | |||
3418 | OMPInfoCache.RFIs[OMPRTL___kmpc_parallel_51]; | |||
3419 | ||||
3420 | auto PredCallSite = [&](AbstractCallSite ACS) { | |||
3421 | Function *Caller = ACS.getInstruction()->getFunction(); | |||
3422 | ||||
3423 | assert(Caller && "Caller is nullptr")((void)0); | |||
3424 | ||||
3425 | auto &CAA = | |||
3426 | A.getOrCreateAAFor<AAKernelInfo>(IRPosition::function(*Caller)); | |||
3427 | if (CAA.ParallelLevels.isValidState()) { | |||
3428 | // Any function that is called by `__kmpc_parallel_51` will not be | |||
3429 | // folded as the parallel level in the function is updated. In order to | |||
3430 | // get it right, all the analysis would depend on the implentation. That | |||
3431 | // said, if in the future any change to the implementation, the analysis | |||
3432 | // could be wrong. As a consequence, we are just conservative here. | |||
3433 | if (Caller == Parallel51RFI.Declaration) { | |||
3434 | ParallelLevels.indicatePessimisticFixpoint(); | |||
3435 | return true; | |||
3436 | } | |||
3437 | ||||
3438 | ParallelLevels ^= CAA.ParallelLevels; | |||
3439 | ||||
3440 | return true; | |||
3441 | } | |||
3442 | ||||
3443 | // We lost track of the caller of the associated function, any kernel | |||
3444 | // could reach now. | |||
3445 | ParallelLevels.indicatePessimisticFixpoint(); | |||
3446 | ||||
3447 | return true; | |||
3448 | }; | |||
3449 | ||||
3450 | bool AllCallSitesKnown = true; | |||
3451 | if (!A.checkForAllCallSites(PredCallSite, *this, | |||
3452 | true /* RequireAllCallSites */, | |||
3453 | AllCallSitesKnown)) | |||
3454 | ParallelLevels.indicatePessimisticFixpoint(); | |||
3455 | } | |||
3456 | }; | |||
3457 | ||||
3458 | /// The call site kernel info abstract attribute, basically, what can we say | |||
3459 | /// about a call site with regards to the KernelInfoState. For now this simply | |||
3460 | /// forwards the information from the callee. | |||
3461 | struct AAKernelInfoCallSite : AAKernelInfo { | |||
3462 | AAKernelInfoCallSite(const IRPosition &IRP, Attributor &A) | |||
3463 | : AAKernelInfo(IRP, A) {} | |||
3464 | ||||
3465 | /// See AbstractAttribute::initialize(...). | |||
3466 | void initialize(Attributor &A) override { | |||
3467 | AAKernelInfo::initialize(A); | |||
3468 | ||||
3469 | CallBase &CB = cast<CallBase>(getAssociatedValue()); | |||
3470 | Function *Callee = getAssociatedFunction(); | |||
3471 | ||||
3472 | // Helper to lookup an assumption string. | |||
3473 | auto HasAssumption = [](Function *Fn, StringRef AssumptionStr) { | |||
3474 | return Fn && hasAssumption(*Fn, AssumptionStr); | |||
3475 | }; | |||
3476 | ||||
3477 | // Check for SPMD-mode assumptions. | |||
3478 | if (HasAssumption(Callee, "ompx_spmd_amenable")) | |||
3479 | SPMDCompatibilityTracker.indicateOptimisticFixpoint(); | |||
3480 | ||||
3481 | // First weed out calls we do not care about, that is readonly/readnone | |||
3482 | // calls, intrinsics, and "no_openmp" calls. Neither of these can reach a | |||
3483 | // parallel region or anything else we are looking for. | |||
3484 | if (!CB.mayWriteToMemory() || isa<IntrinsicInst>(CB)) { | |||
3485 | indicateOptimisticFixpoint(); | |||
3486 | return; | |||
3487 | } | |||
3488 | ||||
3489 | // Next we check if we know the callee. If it is a known OpenMP function | |||
3490 | // we will handle them explicitly in the switch below. If it is not, we | |||
3491 | // will use an AAKernelInfo object on the callee to gather information and | |||
3492 | // merge that into the current state. The latter happens in the updateImpl. | |||
3493 | auto &OMPInfoCache = static_cast<OMPInformationCache &>(A.getInfoCache()); | |||
3494 | const auto &It = OMPInfoCache.RuntimeFunctionIDMap.find(Callee); | |||
3495 | if (It == OMPInfoCache.RuntimeFunctionIDMap.end()) { | |||
3496 | // Unknown caller or declarations are not analyzable, we give up. | |||
3497 | if (!Callee || !A.isFunctionIPOAmendable(*Callee)) { | |||
3498 | ||||
3499 | // Unknown callees might contain parallel regions, except if they have | |||
3500 | // an appropriate assumption attached. | |||
3501 | if (!(HasAssumption(Callee, "omp_no_openmp") || | |||
3502 | HasAssumption(Callee, "omp_no_parallelism"))) | |||
3503 | ReachedUnknownParallelRegions.insert(&CB); | |||
3504 | ||||
3505 | // If SPMDCompatibilityTracker is not fixed, we need to give up on the | |||
3506 | // idea we can run something unknown in SPMD-mode. | |||
3507 | if (!SPMDCompatibilityTracker.isAtFixpoint()) | |||
3508 | SPMDCompatibilityTracker.insert(&CB); | |||
3509 | ||||
3510 | // We have updated the state for this unknown call properly, there won't | |||
3511 | // be any change so we indicate a fixpoint. | |||
3512 | indicateOptimisticFixpoint(); | |||
3513 | } | |||
3514 | // If the callee is known and can be used in IPO, we will update the state | |||
3515 | // based on the callee state in updateImpl. | |||
3516 | return; | |||
3517 | } | |||
3518 | ||||
3519 | const unsigned int WrapperFunctionArgNo = 6; | |||
3520 | RuntimeFunction RF = It->getSecond(); | |||
3521 | switch (RF) { | |||
3522 | // All the functions we know are compatible with SPMD mode. | |||
3523 | case OMPRTL___kmpc_is_spmd_exec_mode: | |||
3524 | case OMPRTL___kmpc_for_static_fini: | |||
3525 | case OMPRTL___kmpc_global_thread_num: | |||
3526 | case OMPRTL___kmpc_get_hardware_num_threads_in_block: | |||
3527 | case OMPRTL___kmpc_get_hardware_num_blocks: | |||
3528 | case OMPRTL___kmpc_single: | |||
3529 | case OMPRTL___kmpc_end_single: | |||
3530 | case OMPRTL___kmpc_master: | |||
3531 | case OMPRTL___kmpc_end_master: | |||
3532 | case OMPRTL___kmpc_barrier: | |||
3533 | break; | |||
3534 | case OMPRTL___kmpc_for_static_init_4: | |||
3535 | case OMPRTL___kmpc_for_static_init_4u: | |||
3536 | case OMPRTL___kmpc_for_static_init_8: | |||
3537 | case OMPRTL___kmpc_for_static_init_8u: { | |||
3538 | // Check the schedule and allow static schedule in SPMD mode. | |||
3539 | unsigned ScheduleArgOpNo = 2; | |||
3540 | auto *ScheduleTypeCI = | |||
3541 | dyn_cast<ConstantInt>(CB.getArgOperand(ScheduleArgOpNo)); | |||
3542 | unsigned ScheduleTypeVal = | |||
3543 | ScheduleTypeCI ? ScheduleTypeCI->getZExtValue() : 0; | |||
3544 | switch (OMPScheduleType(ScheduleTypeVal)) { | |||
3545 | case OMPScheduleType::Static: | |||
3546 | case OMPScheduleType::StaticChunked: | |||
3547 | case OMPScheduleType::Distribute: | |||
3548 | case OMPScheduleType::DistributeChunked: | |||
3549 | break; | |||
3550 | default: | |||
3551 | SPMDCompatibilityTracker.insert(&CB); | |||
3552 | break; | |||
3553 | }; | |||
3554 | } break; | |||
3555 | case OMPRTL___kmpc_target_init: | |||
3556 | KernelInitCB = &CB; | |||
3557 | break; | |||
3558 | case OMPRTL___kmpc_target_deinit: | |||
3559 | KernelDeinitCB = &CB; | |||
3560 | break; | |||
3561 | case OMPRTL___kmpc_parallel_51: | |||
3562 | if (auto *ParallelRegion = dyn_cast<Function>( | |||
3563 | CB.getArgOperand(WrapperFunctionArgNo)->stripPointerCasts())) { | |||
3564 | ReachedKnownParallelRegions.insert(ParallelRegion); | |||
3565 | break; | |||
3566 | } | |||
3567 | // The condition above should usually get the parallel region function | |||
3568 | // pointer and record it. In the off chance it doesn't we assume the | |||
3569 | // worst. | |||
3570 | ReachedUnknownParallelRegions.insert(&CB); | |||
3571 | break; | |||
3572 | case OMPRTL___kmpc_omp_task: | |||
3573 | // We do not look into tasks right now, just give up. | |||
3574 | SPMDCompatibilityTracker.insert(&CB); | |||
3575 | ReachedUnknownParallelRegions.insert(&CB); | |||
3576 | break; | |||
3577 | case OMPRTL___kmpc_alloc_shared: | |||
3578 | case OMPRTL___kmpc_free_shared: | |||
3579 | // Return without setting a fixpoint, to be resolved in updateImpl. | |||
3580 | return; | |||
3581 | default: | |||
3582 | // Unknown OpenMP runtime calls cannot be executed in SPMD-mode, | |||
3583 | // generally. | |||
3584 | SPMDCompatibilityTracker.insert(&CB); | |||
3585 | break; | |||
3586 | } | |||
3587 | // All other OpenMP runtime calls will not reach parallel regions so they | |||
3588 | // can be safely ignored for now. Since it is a known OpenMP runtime call we | |||
3589 | // have now modeled all effects and there is no need for any update. | |||
3590 | indicateOptimisticFixpoint(); | |||
3591 | } | |||
3592 | ||||
3593 | ChangeStatus updateImpl(Attributor &A) override { | |||
3594 | // TODO: Once we have call site specific value information we can provide | |||
3595 | // call site specific liveness information and then it makes | |||
3596 | // sense to specialize attributes for call sites arguments instead of | |||
3597 | // redirecting requests to the callee argument. | |||
3598 | Function *F = getAssociatedFunction(); | |||
3599 | ||||
3600 | auto &OMPInfoCache = static_cast<OMPInformationCache &>(A.getInfoCache()); | |||
3601 | const auto &It = OMPInfoCache.RuntimeFunctionIDMap.find(F); | |||
3602 | ||||
3603 | // If F is not a runtime function, propagate the AAKernelInfo of the callee. | |||
3604 | if (It == OMPInfoCache.RuntimeFunctionIDMap.end()) { | |||
3605 | const IRPosition &FnPos = IRPosition::function(*F); | |||
3606 | auto &FnAA = A.getAAFor<AAKernelInfo>(*this, FnPos, DepClassTy::REQUIRED); | |||
3607 | if (getState() == FnAA.getState()) | |||
3608 | return ChangeStatus::UNCHANGED; | |||
3609 | getState() = FnAA.getState(); | |||
3610 | return ChangeStatus::CHANGED; | |||
3611 | } | |||
3612 | ||||
3613 | // F is a runtime function that allocates or frees memory, check | |||
3614 | // AAHeapToStack and AAHeapToShared. | |||
3615 | KernelInfoState StateBefore = getState(); | |||
3616 | assert((It->getSecond() == OMPRTL___kmpc_alloc_shared ||((void)0) | |||
3617 | It->getSecond() == OMPRTL___kmpc_free_shared) &&((void)0) | |||
3618 | "Expected a __kmpc_alloc_shared or __kmpc_free_shared runtime call")((void)0); | |||
3619 | ||||
3620 | CallBase &CB = cast<CallBase>(getAssociatedValue()); | |||
3621 | ||||
3622 | auto &HeapToStackAA = A.getAAFor<AAHeapToStack>( | |||
3623 | *this, IRPosition::function(*CB.getCaller()), DepClassTy::OPTIONAL); | |||
3624 | auto &HeapToSharedAA = A.getAAFor<AAHeapToShared>( | |||
3625 | *this, IRPosition::function(*CB.getCaller()), DepClassTy::OPTIONAL); | |||
3626 | ||||
3627 | RuntimeFunction RF = It->getSecond(); | |||
3628 | ||||
3629 | switch (RF) { | |||
3630 | // If neither HeapToStack nor HeapToShared assume the call is removed, | |||
3631 | // assume SPMD incompatibility. | |||
3632 | case OMPRTL___kmpc_alloc_shared: | |||
3633 | if (!HeapToStackAA.isAssumedHeapToStack(CB) && | |||
3634 | !HeapToSharedAA.isAssumedHeapToShared(CB)) | |||
3635 | SPMDCompatibilityTracker.insert(&CB); | |||
3636 | break; | |||
3637 | case OMPRTL___kmpc_free_shared: | |||
3638 | if (!HeapToStackAA.isAssumedHeapToStackRemovedFree(CB) && | |||
3639 | !HeapToSharedAA.isAssumedHeapToSharedRemovedFree(CB)) | |||
3640 | SPMDCompatibilityTracker.insert(&CB); | |||
3641 | break; | |||
3642 | default: | |||
3643 | SPMDCompatibilityTracker.insert(&CB); | |||
3644 | } | |||
3645 | ||||
3646 | return StateBefore == getState() ? ChangeStatus::UNCHANGED | |||
3647 | : ChangeStatus::CHANGED; | |||
3648 | } | |||
3649 | }; | |||
3650 | ||||
3651 | struct AAFoldRuntimeCall | |||
3652 | : public StateWrapper<BooleanState, AbstractAttribute> { | |||
3653 | using Base = StateWrapper<BooleanState, AbstractAttribute>; | |||
3654 | ||||
3655 | AAFoldRuntimeCall(const IRPosition &IRP, Attributor &A) : Base(IRP) {} | |||
3656 | ||||
3657 | /// Statistics are tracked as part of manifest for now. | |||
3658 | void trackStatistics() const override {} | |||
3659 | ||||
3660 | /// Create an abstract attribute biew for the position \p IRP. | |||
3661 | static AAFoldRuntimeCall &createForPosition(const IRPosition &IRP, | |||
3662 | Attributor &A); | |||
3663 | ||||
3664 | /// See AbstractAttribute::getName() | |||
3665 | const std::string getName() const override { return "AAFoldRuntimeCall"; } | |||
3666 | ||||
3667 | /// See AbstractAttribute::getIdAddr() | |||
3668 | const char *getIdAddr() const override { return &ID; } | |||
3669 | ||||
3670 | /// This function should return true if the type of the \p AA is | |||
3671 | /// AAFoldRuntimeCall | |||
3672 | static bool classof(const AbstractAttribute *AA) { | |||
3673 | return (AA->getIdAddr() == &ID); | |||
3674 | } | |||
3675 | ||||
3676 | static const char ID; | |||
3677 | }; | |||
3678 | ||||
3679 | struct AAFoldRuntimeCallCallSiteReturned : AAFoldRuntimeCall { | |||
3680 | AAFoldRuntimeCallCallSiteReturned(const IRPosition &IRP, Attributor &A) | |||
3681 | : AAFoldRuntimeCall(IRP, A) {} | |||
3682 | ||||
3683 | /// See AbstractAttribute::getAsStr() | |||
3684 | const std::string getAsStr() const override { | |||
3685 | if (!isValidState()) | |||
3686 | return "<invalid>"; | |||
3687 | ||||
3688 | std::string Str("simplified value: "); | |||
3689 | ||||
3690 | if (!SimplifiedValue.hasValue()) | |||
3691 | return Str + std::string("none"); | |||
3692 | ||||
3693 | if (!SimplifiedValue.getValue()) | |||
3694 | return Str + std::string("nullptr"); | |||
3695 | ||||
3696 | if (ConstantInt *CI = dyn_cast<ConstantInt>(SimplifiedValue.getValue())) | |||
3697 | return Str + std::to_string(CI->getSExtValue()); | |||
3698 | ||||
3699 | return Str + std::string("unknown"); | |||
3700 | } | |||
3701 | ||||
3702 | void initialize(Attributor &A) override { | |||
3703 | Function *Callee = getAssociatedFunction(); | |||
3704 | ||||
3705 | auto &OMPInfoCache = static_cast<OMPInformationCache &>(A.getInfoCache()); | |||
3706 | const auto &It = OMPInfoCache.RuntimeFunctionIDMap.find(Callee); | |||
3707 | assert(It != OMPInfoCache.RuntimeFunctionIDMap.end() &&((void)0) | |||
3708 | "Expected a known OpenMP runtime function")((void)0); | |||
3709 | ||||
3710 | RFKind = It->getSecond(); | |||
3711 | ||||
3712 | CallBase &CB = cast<CallBase>(getAssociatedValue()); | |||
3713 | A.registerSimplificationCallback( | |||
3714 | IRPosition::callsite_returned(CB), | |||
3715 | [&](const IRPosition &IRP, const AbstractAttribute *AA, | |||
3716 | bool &UsedAssumedInformation) -> Optional<Value *> { | |||
3717 | assert((isValidState() || (SimplifiedValue.hasValue() &&((void)0) | |||
3718 | SimplifiedValue.getValue() == nullptr)) &&((void)0) | |||
3719 | "Unexpected invalid state!")((void)0); | |||
3720 | ||||
3721 | if (!isAtFixpoint()) { | |||
3722 | UsedAssumedInformation = true; | |||
3723 | if (AA) | |||
3724 | A.recordDependence(*this, *AA, DepClassTy::OPTIONAL); | |||
3725 | } | |||
3726 | return SimplifiedValue; | |||
3727 | }); | |||
3728 | } | |||
3729 | ||||
3730 | ChangeStatus updateImpl(Attributor &A) override { | |||
3731 | ChangeStatus Changed = ChangeStatus::UNCHANGED; | |||
3732 | switch (RFKind) { | |||
3733 | case OMPRTL___kmpc_is_spmd_exec_mode: | |||
3734 | Changed |= foldIsSPMDExecMode(A); | |||
3735 | break; | |||
3736 | case OMPRTL___kmpc_is_generic_main_thread_id: | |||
3737 | Changed |= foldIsGenericMainThread(A); | |||
3738 | break; | |||
3739 | case OMPRTL___kmpc_parallel_level: | |||
3740 | Changed |= foldParallelLevel(A); | |||
3741 | break; | |||
3742 | case OMPRTL___kmpc_get_hardware_num_threads_in_block: | |||
3743 | Changed = Changed | foldKernelFnAttribute(A, "omp_target_thread_limit"); | |||
3744 | break; | |||
3745 | case OMPRTL___kmpc_get_hardware_num_blocks: | |||
3746 | Changed = Changed | foldKernelFnAttribute(A, "omp_target_num_teams"); | |||
3747 | break; | |||
3748 | default: | |||
3749 | llvm_unreachable("Unhandled OpenMP runtime function!")__builtin_unreachable(); | |||
3750 | } | |||
3751 | ||||
3752 | return Changed; | |||
3753 | } | |||
3754 | ||||
3755 | ChangeStatus manifest(Attributor &A) override { | |||
3756 | ChangeStatus Changed = ChangeStatus::UNCHANGED; | |||
3757 | ||||
3758 | if (SimplifiedValue.hasValue() && SimplifiedValue.getValue()) { | |||
3759 | Instruction &CB = *getCtxI(); | |||
3760 | A.changeValueAfterManifest(CB, **SimplifiedValue); | |||
3761 | A.deleteAfterManifest(CB); | |||
3762 | ||||
3763 | LLVM_DEBUG(dbgs() << TAG << "Folding runtime call: " << CB << " with "do { } while (false) | |||
3764 | << **SimplifiedValue << "\n")do { } while (false); | |||
3765 | ||||
3766 | Changed = ChangeStatus::CHANGED; | |||
3767 | } | |||
3768 | ||||
3769 | return Changed; | |||
3770 | } | |||
3771 | ||||
3772 | ChangeStatus indicatePessimisticFixpoint() override { | |||
3773 | SimplifiedValue = nullptr; | |||
3774 | return AAFoldRuntimeCall::indicatePessimisticFixpoint(); | |||
3775 | } | |||
3776 | ||||
3777 | private: | |||
3778 | /// Fold __kmpc_is_spmd_exec_mode into a constant if possible. | |||
3779 | ChangeStatus foldIsSPMDExecMode(Attributor &A) { | |||
3780 | Optional<Value *> SimplifiedValueBefore = SimplifiedValue; | |||
3781 | ||||
3782 | unsigned AssumedSPMDCount = 0, KnownSPMDCount = 0; | |||
3783 | unsigned AssumedNonSPMDCount = 0, KnownNonSPMDCount = 0; | |||
3784 | auto &CallerKernelInfoAA = A.getAAFor<AAKernelInfo>( | |||
3785 | *this, IRPosition::function(*getAnchorScope()), DepClassTy::REQUIRED); | |||
3786 | ||||
3787 | if (!CallerKernelInfoAA.ReachingKernelEntries.isValidState()) | |||
3788 | return indicatePessimisticFixpoint(); | |||
3789 | ||||
3790 | for (Kernel K : CallerKernelInfoAA.ReachingKernelEntries) { | |||
3791 | auto &AA = A.getAAFor<AAKernelInfo>(*this, IRPosition::function(*K), | |||
3792 | DepClassTy::REQUIRED); | |||
3793 | ||||
3794 | if (!AA.isValidState()) { | |||
3795 | SimplifiedValue = nullptr; | |||
3796 | return indicatePessimisticFixpoint(); | |||
3797 | } | |||
3798 | ||||
3799 | if (AA.SPMDCompatibilityTracker.isAssumed()) { | |||
3800 | if (AA.SPMDCompatibilityTracker.isAtFixpoint()) | |||
3801 | ++KnownSPMDCount; | |||
3802 | else | |||
3803 | ++AssumedSPMDCount; | |||
3804 | } else { | |||
3805 | if (AA.SPMDCompatibilityTracker.isAtFixpoint()) | |||
3806 | ++KnownNonSPMDCount; | |||
3807 | else | |||
3808 | ++AssumedNonSPMDCount; | |||
3809 | } | |||
3810 | } | |||
3811 | ||||
3812 | if ((AssumedSPMDCount + KnownSPMDCount) && | |||
3813 | (AssumedNonSPMDCount + KnownNonSPMDCount)) | |||
3814 | return indicatePessimisticFixpoint(); | |||
3815 | ||||
3816 | auto &Ctx = getAnchorValue().getContext(); | |||
3817 | if (KnownSPMDCount || AssumedSPMDCount) { | |||
3818 | assert(KnownNonSPMDCount == 0 && AssumedNonSPMDCount == 0 &&((void)0) | |||
3819 | "Expected only SPMD kernels!")((void)0); | |||
3820 | // All reaching kernels are in SPMD mode. Update all function calls to | |||
3821 | // __kmpc_is_spmd_exec_mode to 1. | |||
3822 | SimplifiedValue = ConstantInt::get(Type::getInt8Ty(Ctx), true); | |||
3823 | } else if (KnownNonSPMDCount || AssumedNonSPMDCount) { | |||
3824 | assert(KnownSPMDCount == 0 && AssumedSPMDCount == 0 &&((void)0) | |||
3825 | "Expected only non-SPMD kernels!")((void)0); | |||
3826 | // All reaching kernels are in non-SPMD mode. Update all function | |||
3827 | // calls to __kmpc_is_spmd_exec_mode to 0. | |||
3828 | SimplifiedValue = ConstantInt::get(Type::getInt8Ty(Ctx), false); | |||
3829 | } else { | |||
3830 | // We have empty reaching kernels, therefore we cannot tell if the | |||
3831 | // associated call site can be folded. At this moment, SimplifiedValue | |||
3832 | // must be none. | |||
3833 | assert(!SimplifiedValue.hasValue() && "SimplifiedValue should be none")((void)0); | |||
3834 | } | |||
3835 | ||||
3836 | return SimplifiedValue == SimplifiedValueBefore ? ChangeStatus::UNCHANGED | |||
3837 | : ChangeStatus::CHANGED; | |||
3838 | } | |||
3839 | ||||
3840 | /// Fold __kmpc_is_generic_main_thread_id into a constant if possible. | |||
3841 | ChangeStatus foldIsGenericMainThread(Attributor &A) { | |||
3842 | Optional<Value *> SimplifiedValueBefore = SimplifiedValue; | |||
3843 | ||||
3844 | CallBase &CB = cast<CallBase>(getAssociatedValue()); | |||
3845 | Function *F = CB.getFunction(); | |||
3846 | const auto &ExecutionDomainAA = A.getAAFor<AAExecutionDomain>( | |||
3847 | *this, IRPosition::function(*F), DepClassTy::REQUIRED); | |||
3848 | ||||
3849 | if (!ExecutionDomainAA.isValidState()) | |||
3850 | return indicatePessimisticFixpoint(); | |||
3851 | ||||
3852 | auto &Ctx = getAnchorValue().getContext(); | |||
3853 | if (ExecutionDomainAA.isExecutedByInitialThreadOnly(CB)) | |||
3854 | SimplifiedValue = ConstantInt::get(Type::getInt8Ty(Ctx), true); | |||
3855 | else | |||
3856 | return indicatePessimisticFixpoint(); | |||
3857 | ||||
3858 | return SimplifiedValue == SimplifiedValueBefore ? ChangeStatus::UNCHANGED | |||
3859 | : ChangeStatus::CHANGED; | |||
3860 | } | |||
3861 | ||||
3862 | /// Fold __kmpc_parallel_level into a constant if possible. | |||
3863 | ChangeStatus foldParallelLevel(Attributor &A) { | |||
3864 | Optional<Value *> SimplifiedValueBefore = SimplifiedValue; | |||
3865 | ||||
3866 | auto &CallerKernelInfoAA = A.getAAFor<AAKernelInfo>( | |||
3867 | *this, IRPosition::function(*getAnchorScope()), DepClassTy::REQUIRED); | |||
3868 | ||||
3869 | if (!CallerKernelInfoAA.ParallelLevels.isValidState()) | |||
3870 | return indicatePessimisticFixpoint(); | |||
3871 | ||||
3872 | if (!CallerKernelInfoAA.ReachingKernelEntries.isValidState()) | |||
3873 | return indicatePessimisticFixpoint(); | |||
3874 | ||||
3875 | if (CallerKernelInfoAA.ReachingKernelEntries.empty()) { | |||
3876 | assert(!SimplifiedValue.hasValue() &&((void)0) | |||
3877 | "SimplifiedValue should keep none at this point")((void)0); | |||
3878 | return ChangeStatus::UNCHANGED; | |||
3879 | } | |||
3880 | ||||
3881 | unsigned AssumedSPMDCount = 0, KnownSPMDCount = 0; | |||
3882 | unsigned AssumedNonSPMDCount = 0, KnownNonSPMDCount = 0; | |||
3883 | for (Kernel K : CallerKernelInfoAA.ReachingKernelEntries) { | |||
3884 | auto &AA = A.getAAFor<AAKernelInfo>(*this, IRPosition::function(*K), | |||
3885 | DepClassTy::REQUIRED); | |||
3886 | if (!AA.SPMDCompatibilityTracker.isValidState()) | |||
3887 | return indicatePessimisticFixpoint(); | |||
3888 | ||||
3889 | if (AA.SPMDCompatibilityTracker.isAssumed()) { | |||
3890 | if (AA.SPMDCompatibilityTracker.isAtFixpoint()) | |||
3891 | ++KnownSPMDCount; | |||
3892 | else | |||
3893 | ++AssumedSPMDCount; | |||
3894 | } else { | |||
3895 | if (AA.SPMDCompatibilityTracker.isAtFixpoint()) | |||
3896 | ++KnownNonSPMDCount; | |||
3897 | else | |||
3898 | ++AssumedNonSPMDCount; | |||
3899 | } | |||
3900 | } | |||
3901 | ||||
3902 | if ((AssumedSPMDCount + KnownSPMDCount) && | |||
3903 | (AssumedNonSPMDCount + KnownNonSPMDCount)) | |||
3904 | return indicatePessimisticFixpoint(); | |||
3905 | ||||
3906 | auto &Ctx = getAnchorValue().getContext(); | |||
3907 | // If the caller can only be reached by SPMD kernel entries, the parallel | |||
3908 | // level is 1. Similarly, if the caller can only be reached by non-SPMD | |||
3909 | // kernel entries, it is 0. | |||
3910 | if (AssumedSPMDCount || KnownSPMDCount) { | |||
3911 | assert(KnownNonSPMDCount == 0 && AssumedNonSPMDCount == 0 &&((void)0) | |||
3912 | "Expected only SPMD kernels!")((void)0); | |||
3913 | SimplifiedValue = ConstantInt::get(Type::getInt8Ty(Ctx), 1); | |||
3914 | } else { | |||
3915 | assert(KnownSPMDCount == 0 && AssumedSPMDCount == 0 &&((void)0) | |||
3916 | "Expected only non-SPMD kernels!")((void)0); | |||
3917 | SimplifiedValue = ConstantInt::get(Type::getInt8Ty(Ctx), 0); | |||
3918 | } | |||
3919 | return SimplifiedValue == SimplifiedValueBefore ? ChangeStatus::UNCHANGED | |||
3920 | : ChangeStatus::CHANGED; | |||
3921 | } | |||
3922 | ||||
3923 | ChangeStatus foldKernelFnAttribute(Attributor &A, llvm::StringRef Attr) { | |||
3924 | // Specialize only if all the calls agree with the attribute constant value | |||
3925 | int32_t CurrentAttrValue = -1; | |||
3926 | Optional<Value *> SimplifiedValueBefore = SimplifiedValue; | |||
3927 | ||||
3928 | auto &CallerKernelInfoAA = A.getAAFor<AAKernelInfo>( | |||
3929 | *this, IRPosition::function(*getAnchorScope()), DepClassTy::REQUIRED); | |||
3930 | ||||
3931 | if (!CallerKernelInfoAA.ReachingKernelEntries.isValidState()) | |||
3932 | return indicatePessimisticFixpoint(); | |||
3933 | ||||
3934 | // Iterate over the kernels that reach this function | |||
3935 | for (Kernel K : CallerKernelInfoAA.ReachingKernelEntries) { | |||
3936 | int32_t NextAttrVal = -1; | |||
3937 | if (K->hasFnAttribute(Attr)) | |||
3938 | NextAttrVal = | |||
3939 | std::stoi(K->getFnAttribute(Attr).getValueAsString().str()); | |||
3940 | ||||
3941 | if (NextAttrVal == -1 || | |||
3942 | (CurrentAttrValue != -1 && CurrentAttrValue != NextAttrVal)) | |||
3943 | return indicatePessimisticFixpoint(); | |||
3944 | CurrentAttrValue = NextAttrVal; | |||
3945 | } | |||
3946 | ||||
3947 | if (CurrentAttrValue != -1) { | |||
3948 | auto &Ctx = getAnchorValue().getContext(); | |||
3949 | SimplifiedValue = | |||
3950 | ConstantInt::get(Type::getInt32Ty(Ctx), CurrentAttrValue); | |||
3951 | } | |||
3952 | return SimplifiedValue == SimplifiedValueBefore ? ChangeStatus::UNCHANGED | |||
3953 | : ChangeStatus::CHANGED; | |||
3954 | } | |||
3955 | ||||
3956 | /// An optional value the associated value is assumed to fold to. That is, we | |||
3957 | /// assume the associated value (which is a call) can be replaced by this | |||
3958 | /// simplified value. | |||
3959 | Optional<Value *> SimplifiedValue; | |||
3960 | ||||
3961 | /// The runtime function kind of the callee of the associated call site. | |||
3962 | RuntimeFunction RFKind; | |||
3963 | }; | |||
3964 | ||||
3965 | } // namespace | |||
3966 | ||||
3967 | /// Register folding callsite | |||
3968 | void OpenMPOpt::registerFoldRuntimeCall(RuntimeFunction RF) { | |||
3969 | auto &RFI = OMPInfoCache.RFIs[RF]; | |||
3970 | RFI.foreachUse(SCC, [&](Use &U, Function &F) { | |||
3971 | CallInst *CI = OpenMPOpt::getCallIfRegularCall(U, &RFI); | |||
3972 | if (!CI) | |||
3973 | return false; | |||
3974 | A.getOrCreateAAFor<AAFoldRuntimeCall>( | |||
3975 | IRPosition::callsite_returned(*CI), /* QueryingAA */ nullptr, | |||
3976 | DepClassTy::NONE, /* ForceUpdate */ false, | |||
3977 | /* UpdateAfterInit */ false); | |||
3978 | return false; | |||
3979 | }); | |||
3980 | } | |||
3981 | ||||
3982 | void OpenMPOpt::registerAAs(bool IsModulePass) { | |||
3983 | if (SCC.empty()) | |||
3984 | ||||
3985 | return; | |||
3986 | if (IsModulePass) { | |||
3987 | // Ensure we create the AAKernelInfo AAs first and without triggering an | |||
3988 | // update. This will make sure we register all value simplification | |||
3989 | // callbacks before any other AA has the chance to create an AAValueSimplify | |||
3990 | // or similar. | |||
3991 | for (Function *Kernel : OMPInfoCache.Kernels) | |||
3992 | A.getOrCreateAAFor<AAKernelInfo>( | |||
3993 | IRPosition::function(*Kernel), /* QueryingAA */ nullptr, | |||
3994 | DepClassTy::NONE, /* ForceUpdate */ false, | |||
3995 | /* UpdateAfterInit */ false); | |||
3996 | ||||
3997 | ||||
3998 | registerFoldRuntimeCall(OMPRTL___kmpc_is_generic_main_thread_id); | |||
3999 | registerFoldRuntimeCall(OMPRTL___kmpc_is_spmd_exec_mode); | |||
4000 | registerFoldRuntimeCall(OMPRTL___kmpc_parallel_level); | |||
4001 | registerFoldRuntimeCall(OMPRTL___kmpc_get_hardware_num_threads_in_block); | |||
4002 | registerFoldRuntimeCall(OMPRTL___kmpc_get_hardware_num_blocks); | |||
4003 | } | |||
4004 | ||||
4005 | // Create CallSite AA for all Getters. | |||
4006 | for (int Idx = 0; Idx < OMPInfoCache.ICVs.size() - 1; ++Idx) { | |||
4007 | auto ICVInfo = OMPInfoCache.ICVs[static_cast<InternalControlVar>(Idx)]; | |||
4008 | ||||
4009 | auto &GetterRFI = OMPInfoCache.RFIs[ICVInfo.Getter]; | |||
4010 | ||||
4011 | auto CreateAA = [&](Use &U, Function &Caller) { | |||
4012 | CallInst *CI = OpenMPOpt::getCallIfRegularCall(U, &GetterRFI); | |||
4013 | if (!CI) | |||
4014 | return false; | |||
4015 | ||||
4016 | auto &CB = cast<CallBase>(*CI); | |||
4017 | ||||
4018 | IRPosition CBPos = IRPosition::callsite_function(CB); | |||
4019 | A.getOrCreateAAFor<AAICVTracker>(CBPos); | |||
4020 | return false; | |||
4021 | }; | |||
4022 | ||||
4023 | GetterRFI.foreachUse(SCC, CreateAA); | |||
4024 | } | |||
4025 | auto &GlobalizationRFI = OMPInfoCache.RFIs[OMPRTL___kmpc_alloc_shared]; | |||
4026 | auto CreateAA = [&](Use &U, Function &F) { | |||
4027 | A.getOrCreateAAFor<AAHeapToShared>(IRPosition::function(F)); | |||
4028 | return false; | |||
4029 | }; | |||
4030 | GlobalizationRFI.foreachUse(SCC, CreateAA); | |||
4031 | ||||
4032 | // Create an ExecutionDomain AA for every function and a HeapToStack AA for | |||
4033 | // every function if there is a device kernel. | |||
4034 | if (!isOpenMPDevice(M)) | |||
4035 | return; | |||
4036 | ||||
4037 | for (auto *F : SCC) { | |||
4038 | if (F->isDeclaration()) | |||
4039 | continue; | |||
4040 | ||||
4041 | A.getOrCreateAAFor<AAExecutionDomain>(IRPosition::function(*F)); | |||
4042 | A.getOrCreateAAFor<AAHeapToStack>(IRPosition::function(*F)); | |||
4043 | ||||
4044 | for (auto &I : instructions(*F)) { | |||
4045 | if (auto *LI = dyn_cast<LoadInst>(&I)) { | |||
4046 | bool UsedAssumedInformation = false; | |||
4047 | A.getAssumedSimplified(IRPosition::value(*LI), /* AA */ nullptr, | |||
4048 | UsedAssumedInformation); | |||
4049 | } | |||
4050 | } | |||
4051 | } | |||
4052 | } | |||
4053 | ||||
4054 | const char AAICVTracker::ID = 0; | |||
4055 | const char AAKernelInfo::ID = 0; | |||
4056 | const char AAExecutionDomain::ID = 0; | |||
4057 | const char AAHeapToShared::ID = 0; | |||
4058 | const char AAFoldRuntimeCall::ID = 0; | |||
4059 | ||||
4060 | AAICVTracker &AAICVTracker::createForPosition(const IRPosition &IRP, | |||
4061 | Attributor &A) { | |||
4062 | AAICVTracker *AA = nullptr; | |||
4063 | switch (IRP.getPositionKind()) { | |||
4064 | case IRPosition::IRP_INVALID: | |||
4065 | case IRPosition::IRP_FLOAT: | |||
4066 | case IRPosition::IRP_ARGUMENT: | |||
4067 | case IRPosition::IRP_CALL_SITE_ARGUMENT: | |||
4068 | llvm_unreachable("ICVTracker can only be created for function position!")__builtin_unreachable(); | |||
4069 | case IRPosition::IRP_RETURNED: | |||
4070 | AA = new (A.Allocator) AAICVTrackerFunctionReturned(IRP, A); | |||
4071 | break; | |||
4072 | case IRPosition::IRP_CALL_SITE_RETURNED: | |||
4073 | AA = new (A.Allocator) AAICVTrackerCallSiteReturned(IRP, A); | |||
4074 | break; | |||
4075 | case IRPosition::IRP_CALL_SITE: | |||
4076 | AA = new (A.Allocator) AAICVTrackerCallSite(IRP, A); | |||
4077 | break; | |||
4078 | case IRPosition::IRP_FUNCTION: | |||
4079 | AA = new (A.Allocator) AAICVTrackerFunction(IRP, A); | |||
4080 | break; | |||
4081 | } | |||
4082 | ||||
4083 | return *AA; | |||
4084 | } | |||
4085 | ||||
4086 | AAExecutionDomain &AAExecutionDomain::createForPosition(const IRPosition &IRP, | |||
4087 | Attributor &A) { | |||
4088 | AAExecutionDomainFunction *AA = nullptr; | |||
4089 | switch (IRP.getPositionKind()) { | |||
4090 | case IRPosition::IRP_INVALID: | |||
4091 | case IRPosition::IRP_FLOAT: | |||
4092 | case IRPosition::IRP_ARGUMENT: | |||
4093 | case IRPosition::IRP_CALL_SITE_ARGUMENT: | |||
4094 | case IRPosition::IRP_RETURNED: | |||
4095 | case IRPosition::IRP_CALL_SITE_RETURNED: | |||
4096 | case IRPosition::IRP_CALL_SITE: | |||
4097 | llvm_unreachable(__builtin_unreachable() | |||
4098 | "AAExecutionDomain can only be created for function position!")__builtin_unreachable(); | |||
4099 | case IRPosition::IRP_FUNCTION: | |||
4100 | AA = new (A.Allocator) AAExecutionDomainFunction(IRP, A); | |||
4101 | break; | |||
4102 | } | |||
4103 | ||||
4104 | return *AA; | |||
4105 | } | |||
4106 | ||||
4107 | AAHeapToShared &AAHeapToShared::createForPosition(const IRPosition &IRP, | |||
4108 | Attributor &A) { | |||
4109 | AAHeapToSharedFunction *AA = nullptr; | |||
4110 | switch (IRP.getPositionKind()) { | |||
4111 | case IRPosition::IRP_INVALID: | |||
4112 | case IRPosition::IRP_FLOAT: | |||
4113 | case IRPosition::IRP_ARGUMENT: | |||
4114 | case IRPosition::IRP_CALL_SITE_ARGUMENT: | |||
4115 | case IRPosition::IRP_RETURNED: | |||
4116 | case IRPosition::IRP_CALL_SITE_RETURNED: | |||
4117 | case IRPosition::IRP_CALL_SITE: | |||
4118 | llvm_unreachable(__builtin_unreachable() | |||
4119 | "AAHeapToShared can only be created for function position!")__builtin_unreachable(); | |||
4120 | case IRPosition::IRP_FUNCTION: | |||
4121 | AA = new (A.Allocator) AAHeapToSharedFunction(IRP, A); | |||
4122 | break; | |||
4123 | } | |||
4124 | ||||
4125 | return *AA; | |||
4126 | } | |||
4127 | ||||
4128 | AAKernelInfo &AAKernelInfo::createForPosition(const IRPosition &IRP, | |||
4129 | Attributor &A) { | |||
4130 | AAKernelInfo *AA = nullptr; | |||
4131 | switch (IRP.getPositionKind()) { | |||
4132 | case IRPosition::IRP_INVALID: | |||
4133 | case IRPosition::IRP_FLOAT: | |||
4134 | case IRPosition::IRP_ARGUMENT: | |||
4135 | case IRPosition::IRP_RETURNED: | |||
4136 | case IRPosition::IRP_CALL_SITE_RETURNED: | |||
4137 | case IRPosition::IRP_CALL_SITE_ARGUMENT: | |||
4138 | llvm_unreachable("KernelInfo can only be created for function position!")__builtin_unreachable(); | |||
4139 | case IRPosition::IRP_CALL_SITE: | |||
4140 | AA = new (A.Allocator) AAKernelInfoCallSite(IRP, A); | |||
4141 | break; | |||
4142 | case IRPosition::IRP_FUNCTION: | |||
4143 | AA = new (A.Allocator) AAKernelInfoFunction(IRP, A); | |||
4144 | break; | |||
4145 | } | |||
4146 | ||||
4147 | return *AA; | |||
4148 | } | |||
4149 | ||||
4150 | AAFoldRuntimeCall &AAFoldRuntimeCall::createForPosition(const IRPosition &IRP, | |||
4151 | Attributor &A) { | |||
4152 | AAFoldRuntimeCall *AA = nullptr; | |||
4153 | switch (IRP.getPositionKind()) { | |||
4154 | case IRPosition::IRP_INVALID: | |||
4155 | case IRPosition::IRP_FLOAT: | |||
4156 | case IRPosition::IRP_ARGUMENT: | |||
4157 | case IRPosition::IRP_RETURNED: | |||
4158 | case IRPosition::IRP_FUNCTION: | |||
4159 | case IRPosition::IRP_CALL_SITE: | |||
4160 | case IRPosition::IRP_CALL_SITE_ARGUMENT: | |||
4161 | llvm_unreachable("KernelInfo can only be created for call site position!")__builtin_unreachable(); | |||
4162 | case IRPosition::IRP_CALL_SITE_RETURNED: | |||
4163 | AA = new (A.Allocator) AAFoldRuntimeCallCallSiteReturned(IRP, A); | |||
4164 | break; | |||
4165 | } | |||
4166 | ||||
4167 | return *AA; | |||
4168 | } | |||
4169 | ||||
4170 | PreservedAnalyses OpenMPOptPass::run(Module &M, ModuleAnalysisManager &AM) { | |||
4171 | if (!containsOpenMP(M)) | |||
4172 | return PreservedAnalyses::all(); | |||
4173 | if (DisableOpenMPOptimizations) | |||
4174 | return PreservedAnalyses::all(); | |||
4175 | ||||
4176 | FunctionAnalysisManager &FAM = | |||
4177 | AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager(); | |||
4178 | KernelSet Kernels = getDeviceKernels(M); | |||
4179 | ||||
4180 | auto IsCalled = [&](Function &F) { | |||
4181 | if (Kernels.contains(&F)) | |||
4182 | return true; | |||
4183 | for (const User *U : F.users()) | |||
4184 | if (!isa<BlockAddress>(U)) | |||
4185 | return true; | |||
4186 | return false; | |||
4187 | }; | |||
4188 | ||||
4189 | auto EmitRemark = [&](Function &F) { | |||
4190 | auto &ORE = FAM.getResult<OptimizationRemarkEmitterAnalysis>(F); | |||
4191 | ORE.emit([&]() { | |||
4192 | OptimizationRemarkAnalysis ORA(DEBUG_TYPE"openmp-opt", "OMP140", &F); | |||
4193 | return ORA << "Could not internalize function. " | |||
4194 | << "Some optimizations may not be possible. [OMP140]"; | |||
4195 | }); | |||
4196 | }; | |||
4197 | ||||
4198 | // Create internal copies of each function if this is a kernel Module. This | |||
4199 | // allows iterprocedural passes to see every call edge. | |||
4200 | DenseMap<Function *, Function *> InternalizedMap; | |||
4201 | if (isOpenMPDevice(M)) { | |||
4202 | SmallPtrSet<Function *, 16> InternalizeFns; | |||
4203 | for (Function &F : M) | |||
4204 | if (!F.isDeclaration() && !Kernels.contains(&F) && IsCalled(F) && | |||
4205 | !DisableInternalization) { | |||
4206 | if (Attributor::isInternalizable(F)) { | |||
4207 | InternalizeFns.insert(&F); | |||
4208 | } else if (!F.hasLocalLinkage() && !F.hasFnAttribute(Attribute::Cold)) { | |||
4209 | EmitRemark(F); | |||
4210 | } | |||
4211 | } | |||
4212 | ||||
4213 | Attributor::internalizeFunctions(InternalizeFns, InternalizedMap); | |||
4214 | } | |||
4215 | ||||
4216 | // Look at every function in the Module unless it was internalized. | |||
4217 | SmallVector<Function *, 16> SCC; | |||
4218 | for (Function &F : M) | |||
4219 | if (!F.isDeclaration() && !InternalizedMap.lookup(&F)) | |||
4220 | SCC.push_back(&F); | |||
4221 | ||||
4222 | if (SCC.empty()) | |||
4223 | return PreservedAnalyses::all(); | |||
4224 | ||||
4225 | AnalysisGetter AG(FAM); | |||
4226 | ||||
4227 | auto OREGetter = [&FAM](Function *F) -> OptimizationRemarkEmitter & { | |||
4228 | return FAM.getResult<OptimizationRemarkEmitterAnalysis>(*F); | |||
4229 | }; | |||
4230 | ||||
4231 | BumpPtrAllocator Allocator; | |||
4232 | CallGraphUpdater CGUpdater; | |||
4233 | ||||
4234 | SetVector<Function *> Functions(SCC.begin(), SCC.end()); | |||
4235 | OMPInformationCache InfoCache(M, AG, Allocator, /*CGSCC*/ Functions, Kernels); | |||
4236 | ||||
4237 | unsigned MaxFixpointIterations = (isOpenMPDevice(M)) ? 128 : 32; | |||
4238 | Attributor A(Functions, InfoCache, CGUpdater, nullptr, true, false, | |||
4239 | MaxFixpointIterations, OREGetter, DEBUG_TYPE"openmp-opt"); | |||
4240 | ||||
4241 | OpenMPOpt OMPOpt(SCC, CGUpdater, OREGetter, InfoCache, A); | |||
4242 | bool Changed = OMPOpt.run(true); | |||
4243 | if (Changed) | |||
4244 | return PreservedAnalyses::none(); | |||
4245 | ||||
4246 | return PreservedAnalyses::all(); | |||
4247 | } | |||
4248 | ||||
4249 | PreservedAnalyses OpenMPOptCGSCCPass::run(LazyCallGraph::SCC &C, | |||
4250 | CGSCCAnalysisManager &AM, | |||
4251 | LazyCallGraph &CG, | |||
4252 | CGSCCUpdateResult &UR) { | |||
4253 | if (!containsOpenMP(*C.begin()->getFunction().getParent())) | |||
4254 | return PreservedAnalyses::all(); | |||
4255 | if (DisableOpenMPOptimizations) | |||
4256 | return PreservedAnalyses::all(); | |||
4257 | ||||
4258 | SmallVector<Function *, 16> SCC; | |||
4259 | // If there are kernels in the module, we have to run on all SCC's. | |||
4260 | for (LazyCallGraph::Node &N : C) { | |||
4261 | Function *Fn = &N.getFunction(); | |||
4262 | SCC.push_back(Fn); | |||
4263 | } | |||
4264 | ||||
4265 | if (SCC.empty()) | |||
4266 | return PreservedAnalyses::all(); | |||
4267 | ||||
4268 | Module &M = *C.begin()->getFunction().getParent(); | |||
4269 | ||||
4270 | KernelSet Kernels = getDeviceKernels(M); | |||
4271 | ||||
4272 | FunctionAnalysisManager &FAM = | |||
4273 | AM.getResult<FunctionAnalysisManagerCGSCCProxy>(C, CG).getManager(); | |||
4274 | ||||
4275 | AnalysisGetter AG(FAM); | |||
4276 | ||||
4277 | auto OREGetter = [&FAM](Function *F) -> OptimizationRemarkEmitter & { | |||
4278 | return FAM.getResult<OptimizationRemarkEmitterAnalysis>(*F); | |||
4279 | }; | |||
4280 | ||||
4281 | BumpPtrAllocator Allocator; | |||
4282 | CallGraphUpdater CGUpdater; | |||
4283 | CGUpdater.initialize(CG, C, AM, UR); | |||
4284 | ||||
4285 | SetVector<Function *> Functions(SCC.begin(), SCC.end()); | |||
4286 | OMPInformationCache InfoCache(*(Functions.back()->getParent()), AG, Allocator, | |||
4287 | /*CGSCC*/ Functions, Kernels); | |||
4288 | ||||
4289 | unsigned MaxFixpointIterations = (isOpenMPDevice(M)) ? 128 : 32; | |||
4290 | Attributor A(Functions, InfoCache, CGUpdater, nullptr, false, true, | |||
4291 | MaxFixpointIterations, OREGetter, DEBUG_TYPE"openmp-opt"); | |||
4292 | ||||
4293 | OpenMPOpt OMPOpt(SCC, CGUpdater, OREGetter, InfoCache, A); | |||
4294 | bool Changed = OMPOpt.run(false); | |||
4295 | if (Changed) | |||
4296 | return PreservedAnalyses::none(); | |||
4297 | ||||
4298 | return PreservedAnalyses::all(); | |||
4299 | } | |||
4300 | ||||
4301 | namespace { | |||
4302 | ||||
4303 | struct OpenMPOptCGSCCLegacyPass : public CallGraphSCCPass { | |||
4304 | CallGraphUpdater CGUpdater; | |||
4305 | static char ID; | |||
4306 | ||||
4307 | OpenMPOptCGSCCLegacyPass() : CallGraphSCCPass(ID) { | |||
4308 | initializeOpenMPOptCGSCCLegacyPassPass(*PassRegistry::getPassRegistry()); | |||
4309 | } | |||
4310 | ||||
4311 | void getAnalysisUsage(AnalysisUsage &AU) const override { | |||
4312 | CallGraphSCCPass::getAnalysisUsage(AU); | |||
4313 | } | |||
4314 | ||||
4315 | bool runOnSCC(CallGraphSCC &CGSCC) override { | |||
4316 | if (!containsOpenMP(CGSCC.getCallGraph().getModule())) | |||
4317 | return false; | |||
4318 | if (DisableOpenMPOptimizations || skipSCC(CGSCC)) | |||
4319 | return false; | |||
4320 | ||||
4321 | SmallVector<Function *, 16> SCC; | |||
4322 | // If there are kernels in the module, we have to run on all SCC's. | |||
4323 | for (CallGraphNode *CGN : CGSCC) { | |||
4324 | Function *Fn = CGN->getFunction(); | |||
4325 | if (!Fn || Fn->isDeclaration()) | |||
4326 | continue; | |||
4327 | SCC.push_back(Fn); | |||
4328 | } | |||
4329 | ||||
4330 | if (SCC.empty()) | |||
4331 | return false; | |||
4332 | ||||
4333 | Module &M = CGSCC.getCallGraph().getModule(); | |||
4334 | KernelSet Kernels = getDeviceKernels(M); | |||
4335 | ||||
4336 | CallGraph &CG = getAnalysis<CallGraphWrapperPass>().getCallGraph(); | |||
4337 | CGUpdater.initialize(CG, CGSCC); | |||
4338 | ||||
4339 | // Maintain a map of functions to avoid rebuilding the ORE | |||
4340 | DenseMap<Function *, std::unique_ptr<OptimizationRemarkEmitter>> OREMap; | |||
4341 | auto OREGetter = [&OREMap](Function *F) -> OptimizationRemarkEmitter & { | |||
4342 | std::unique_ptr<OptimizationRemarkEmitter> &ORE = OREMap[F]; | |||
4343 | if (!ORE) | |||
4344 | ORE = std::make_unique<OptimizationRemarkEmitter>(F); | |||
4345 | return *ORE; | |||
4346 | }; | |||
4347 | ||||
4348 | AnalysisGetter AG; | |||
4349 | SetVector<Function *> Functions(SCC.begin(), SCC.end()); | |||
4350 | BumpPtrAllocator Allocator; | |||
4351 | OMPInformationCache InfoCache(*(Functions.back()->getParent()), AG, | |||
4352 | Allocator, | |||
4353 | /*CGSCC*/ Functions, Kernels); | |||
4354 | ||||
4355 | unsigned MaxFixpointIterations = (isOpenMPDevice(M)) ? 128 : 32; | |||
4356 | Attributor A(Functions, InfoCache, CGUpdater, nullptr, false, true, | |||
4357 | MaxFixpointIterations, OREGetter, DEBUG_TYPE"openmp-opt"); | |||
4358 | ||||
4359 | OpenMPOpt OMPOpt(SCC, CGUpdater, OREGetter, InfoCache, A); | |||
4360 | return OMPOpt.run(false); | |||
4361 | } | |||
4362 | ||||
4363 | bool doFinalization(CallGraph &CG) override { return CGUpdater.finalize(); } | |||
4364 | }; | |||
4365 | ||||
4366 | } // end anonymous namespace | |||
4367 | ||||
4368 | KernelSet llvm::omp::getDeviceKernels(Module &M) { | |||
4369 | // TODO: Create a more cross-platform way of determining device kernels. | |||
4370 | NamedMDNode *MD = M.getOrInsertNamedMetadata("nvvm.annotations"); | |||
4371 | KernelSet Kernels; | |||
4372 | ||||
4373 | if (!MD) | |||
4374 | return Kernels; | |||
4375 | ||||
4376 | for (auto *Op : MD->operands()) { | |||
4377 | if (Op->getNumOperands() < 2) | |||
4378 | continue; | |||
4379 | MDString *KindID = dyn_cast<MDString>(Op->getOperand(1)); | |||
4380 | if (!KindID || KindID->getString() != "kernel") | |||
4381 | continue; | |||
4382 | ||||
4383 | Function *KernelFn = | |||
4384 | mdconst::dyn_extract_or_null<Function>(Op->getOperand(0)); | |||
4385 | if (!KernelFn) | |||
4386 | continue; | |||
4387 | ||||
4388 | ++NumOpenMPTargetRegionKernels; | |||
4389 | ||||
4390 | Kernels.insert(KernelFn); | |||
4391 | } | |||
4392 | ||||
4393 | return Kernels; | |||
4394 | } | |||
4395 | ||||
4396 | bool llvm::omp::containsOpenMP(Module &M) { | |||
4397 | Metadata *MD = M.getModuleFlag("openmp"); | |||
4398 | if (!MD) | |||
4399 | return false; | |||
4400 | ||||
4401 | return true; | |||
4402 | } | |||
4403 | ||||
4404 | bool llvm::omp::isOpenMPDevice(Module &M) { | |||
4405 | Metadata *MD = M.getModuleFlag("openmp-device"); | |||
4406 | if (!MD) | |||
4407 | return false; | |||
4408 | ||||
4409 | return true; | |||
4410 | } | |||
4411 | ||||
4412 | char OpenMPOptCGSCCLegacyPass::ID = 0; | |||
4413 | ||||
4414 | INITIALIZE_PASS_BEGIN(OpenMPOptCGSCCLegacyPass, "openmp-opt-cgscc",static void *initializeOpenMPOptCGSCCLegacyPassPassOnce(PassRegistry &Registry) { | |||
4415 | "OpenMP specific optimizations", false, false)static void *initializeOpenMPOptCGSCCLegacyPassPassOnce(PassRegistry &Registry) { | |||
4416 | INITIALIZE_PASS_DEPENDENCY(CallGraphWrapperPass)initializeCallGraphWrapperPassPass(Registry); | |||
4417 | INITIALIZE_PASS_END(OpenMPOptCGSCCLegacyPass, "openmp-opt-cgscc",PassInfo *PI = new PassInfo( "OpenMP specific optimizations", "openmp-opt-cgscc", &OpenMPOptCGSCCLegacyPass::ID, PassInfo ::NormalCtor_t(callDefaultCtor<OpenMPOptCGSCCLegacyPass> ), false, false); Registry.registerPass(*PI, true); return PI ; } static llvm::once_flag InitializeOpenMPOptCGSCCLegacyPassPassFlag ; void llvm::initializeOpenMPOptCGSCCLegacyPassPass(PassRegistry &Registry) { llvm::call_once(InitializeOpenMPOptCGSCCLegacyPassPassFlag , initializeOpenMPOptCGSCCLegacyPassPassOnce, std::ref(Registry )); } | |||
4418 | "OpenMP specific optimizations", false, false)PassInfo *PI = new PassInfo( "OpenMP specific optimizations", "openmp-opt-cgscc", &OpenMPOptCGSCCLegacyPass::ID, PassInfo ::NormalCtor_t(callDefaultCtor<OpenMPOptCGSCCLegacyPass> ), false, false); Registry.registerPass(*PI, true); return PI ; } static llvm::once_flag InitializeOpenMPOptCGSCCLegacyPassPassFlag ; void llvm::initializeOpenMPOptCGSCCLegacyPassPass(PassRegistry &Registry) { llvm::call_once(InitializeOpenMPOptCGSCCLegacyPassPassFlag , initializeOpenMPOptCGSCCLegacyPassPassOnce, std::ref(Registry )); } | |||
4419 | ||||
4420 | Pass *llvm::createOpenMPOptCGSCCLegacyPass() { | |||
4421 | return new OpenMPOptCGSCCLegacyPass(); | |||
4422 | } |
1 | //===- Attributor.h --- Module-wide attribute deduction ---------*- 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 | // Attributor: An inter procedural (abstract) "attribute" deduction framework. |
10 | // |
11 | // The Attributor framework is an inter procedural abstract analysis (fixpoint |
12 | // iteration analysis). The goal is to allow easy deduction of new attributes as |
13 | // well as information exchange between abstract attributes in-flight. |
14 | // |
15 | // The Attributor class is the driver and the link between the various abstract |
16 | // attributes. The Attributor will iterate until a fixpoint state is reached by |
17 | // all abstract attributes in-flight, or until it will enforce a pessimistic fix |
18 | // point because an iteration limit is reached. |
19 | // |
20 | // Abstract attributes, derived from the AbstractAttribute class, actually |
21 | // describe properties of the code. They can correspond to actual LLVM-IR |
22 | // attributes, or they can be more general, ultimately unrelated to LLVM-IR |
23 | // attributes. The latter is useful when an abstract attributes provides |
24 | // information to other abstract attributes in-flight but we might not want to |
25 | // manifest the information. The Attributor allows to query in-flight abstract |
26 | // attributes through the `Attributor::getAAFor` method (see the method |
27 | // description for an example). If the method is used by an abstract attribute |
28 | // P, and it results in an abstract attribute Q, the Attributor will |
29 | // automatically capture a potential dependence from Q to P. This dependence |
30 | // will cause P to be reevaluated whenever Q changes in the future. |
31 | // |
32 | // The Attributor will only reevaluate abstract attributes that might have |
33 | // changed since the last iteration. That means that the Attribute will not |
34 | // revisit all instructions/blocks/functions in the module but only query |
35 | // an update from a subset of the abstract attributes. |
36 | // |
37 | // The update method `AbstractAttribute::updateImpl` is implemented by the |
38 | // specific "abstract attribute" subclasses. The method is invoked whenever the |
39 | // currently assumed state (see the AbstractState class) might not be valid |
40 | // anymore. This can, for example, happen if the state was dependent on another |
41 | // abstract attribute that changed. In every invocation, the update method has |
42 | // to adjust the internal state of an abstract attribute to a point that is |
43 | // justifiable by the underlying IR and the current state of abstract attributes |
44 | // in-flight. Since the IR is given and assumed to be valid, the information |
45 | // derived from it can be assumed to hold. However, information derived from |
46 | // other abstract attributes is conditional on various things. If the justifying |
47 | // state changed, the `updateImpl` has to revisit the situation and potentially |
48 | // find another justification or limit the optimistic assumes made. |
49 | // |
50 | // Change is the key in this framework. Until a state of no-change, thus a |
51 | // fixpoint, is reached, the Attributor will query the abstract attributes |
52 | // in-flight to re-evaluate their state. If the (current) state is too |
53 | // optimistic, hence it cannot be justified anymore through other abstract |
54 | // attributes or the state of the IR, the state of the abstract attribute will |
55 | // have to change. Generally, we assume abstract attribute state to be a finite |
56 | // height lattice and the update function to be monotone. However, these |
57 | // conditions are not enforced because the iteration limit will guarantee |
58 | // termination. If an optimistic fixpoint is reached, or a pessimistic fix |
59 | // point is enforced after a timeout, the abstract attributes are tasked to |
60 | // manifest their result in the IR for passes to come. |
61 | // |
62 | // Attribute manifestation is not mandatory. If desired, there is support to |
63 | // generate a single or multiple LLVM-IR attributes already in the helper struct |
64 | // IRAttribute. In the simplest case, a subclass inherits from IRAttribute with |
65 | // a proper Attribute::AttrKind as template parameter. The Attributor |
66 | // manifestation framework will then create and place a new attribute if it is |
67 | // allowed to do so (based on the abstract state). Other use cases can be |
68 | // achieved by overloading AbstractAttribute or IRAttribute methods. |
69 | // |
70 | // |
71 | // The "mechanics" of adding a new "abstract attribute": |
72 | // - Define a class (transitively) inheriting from AbstractAttribute and one |
73 | // (which could be the same) that (transitively) inherits from AbstractState. |
74 | // For the latter, consider the already available BooleanState and |
75 | // {Inc,Dec,Bit}IntegerState if they fit your needs, e.g., you require only a |
76 | // number tracking or bit-encoding. |
77 | // - Implement all pure methods. Also use overloading if the attribute is not |
78 | // conforming with the "default" behavior: A (set of) LLVM-IR attribute(s) for |
79 | // an argument, call site argument, function return value, or function. See |
80 | // the class and method descriptions for more information on the two |
81 | // "Abstract" classes and their respective methods. |
82 | // - Register opportunities for the new abstract attribute in the |
83 | // `Attributor::identifyDefaultAbstractAttributes` method if it should be |
84 | // counted as a 'default' attribute. |
85 | // - Add sufficient tests. |
86 | // - Add a Statistics object for bookkeeping. If it is a simple (set of) |
87 | // attribute(s) manifested through the Attributor manifestation framework, see |
88 | // the bookkeeping function in Attributor.cpp. |
89 | // - If instructions with a certain opcode are interesting to the attribute, add |
90 | // that opcode to the switch in `Attributor::identifyAbstractAttributes`. This |
91 | // will make it possible to query all those instructions through the |
92 | // `InformationCache::getOpcodeInstMapForFunction` interface and eliminate the |
93 | // need to traverse the IR repeatedly. |
94 | // |
95 | //===----------------------------------------------------------------------===// |
96 | |
97 | #ifndef LLVM_TRANSFORMS_IPO_ATTRIBUTOR_H |
98 | #define LLVM_TRANSFORMS_IPO_ATTRIBUTOR_H |
99 | |
100 | #include "llvm/ADT/DenseSet.h" |
101 | #include "llvm/ADT/GraphTraits.h" |
102 | #include "llvm/ADT/MapVector.h" |
103 | #include "llvm/ADT/STLExtras.h" |
104 | #include "llvm/ADT/SetVector.h" |
105 | #include "llvm/ADT/Triple.h" |
106 | #include "llvm/ADT/iterator.h" |
107 | #include "llvm/Analysis/AssumeBundleQueries.h" |
108 | #include "llvm/Analysis/CFG.h" |
109 | #include "llvm/Analysis/CGSCCPassManager.h" |
110 | #include "llvm/Analysis/LazyCallGraph.h" |
111 | #include "llvm/Analysis/LoopInfo.h" |
112 | #include "llvm/Analysis/MustExecute.h" |
113 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
114 | #include "llvm/Analysis/PostDominators.h" |
115 | #include "llvm/Analysis/TargetLibraryInfo.h" |
116 | #include "llvm/IR/AbstractCallSite.h" |
117 | #include "llvm/IR/ConstantRange.h" |
118 | #include "llvm/IR/PassManager.h" |
119 | #include "llvm/Support/Allocator.h" |
120 | #include "llvm/Support/Casting.h" |
121 | #include "llvm/Support/GraphWriter.h" |
122 | #include "llvm/Support/TimeProfiler.h" |
123 | #include "llvm/Transforms/Utils/CallGraphUpdater.h" |
124 | |
125 | namespace llvm { |
126 | |
127 | struct AADepGraphNode; |
128 | struct AADepGraph; |
129 | struct Attributor; |
130 | struct AbstractAttribute; |
131 | struct InformationCache; |
132 | struct AAIsDead; |
133 | struct AttributorCallGraph; |
134 | |
135 | class AAManager; |
136 | class AAResults; |
137 | class Function; |
138 | |
139 | /// Abstract Attribute helper functions. |
140 | namespace AA { |
141 | |
142 | /// Return true if \p V is dynamically unique, that is, there are no two |
143 | /// "instances" of \p V at runtime with different values. |
144 | bool isDynamicallyUnique(Attributor &A, const AbstractAttribute &QueryingAA, |
145 | const Value &V); |
146 | |
147 | /// Return true if \p V is a valid value in \p Scope, that is a constant or an |
148 | /// instruction/argument of \p Scope. |
149 | bool isValidInScope(const Value &V, const Function *Scope); |
150 | |
151 | /// Return true if \p V is a valid value at position \p CtxI, that is a |
152 | /// constant, an argument of the same function as \p CtxI, or an instruction in |
153 | /// that function that dominates \p CtxI. |
154 | bool isValidAtPosition(const Value &V, const Instruction &CtxI, |
155 | InformationCache &InfoCache); |
156 | |
157 | /// Try to convert \p V to type \p Ty without introducing new instructions. If |
158 | /// this is not possible return `nullptr`. Note: this function basically knows |
159 | /// how to cast various constants. |
160 | Value *getWithType(Value &V, Type &Ty); |
161 | |
162 | /// Return the combination of \p A and \p B such that the result is a possible |
163 | /// value of both. \p B is potentially casted to match the type \p Ty or the |
164 | /// type of \p A if \p Ty is null. |
165 | /// |
166 | /// Examples: |
167 | /// X + none => X |
168 | /// not_none + undef => not_none |
169 | /// V1 + V2 => nullptr |
170 | Optional<Value *> |
171 | combineOptionalValuesInAAValueLatice(const Optional<Value *> &A, |
172 | const Optional<Value *> &B, Type *Ty); |
173 | |
174 | /// Return the initial value of \p Obj with type \p Ty if that is a constant. |
175 | Constant *getInitialValueForObj(Value &Obj, Type &Ty); |
176 | |
177 | /// Collect all potential underlying objects of \p Ptr at position \p CtxI in |
178 | /// \p Objects. Assumed information is used and dependences onto \p QueryingAA |
179 | /// are added appropriately. |
180 | /// |
181 | /// \returns True if \p Objects contains all assumed underlying objects, and |
182 | /// false if something went wrong and the objects could not be |
183 | /// determined. |
184 | bool getAssumedUnderlyingObjects(Attributor &A, const Value &Ptr, |
185 | SmallVectorImpl<Value *> &Objects, |
186 | const AbstractAttribute &QueryingAA, |
187 | const Instruction *CtxI); |
188 | |
189 | /// Collect all potential values of the one stored by \p SI into |
190 | /// \p PotentialCopies. That is, the only copies that were made via the |
191 | /// store are assumed to be known and all in \p PotentialCopies. Dependences |
192 | /// onto \p QueryingAA are properly tracked, \p UsedAssumedInformation will |
193 | /// inform the caller if assumed information was used. |
194 | /// |
195 | /// \returns True if the assumed potential copies are all in \p PotentialCopies, |
196 | /// false if something went wrong and the copies could not be |
197 | /// determined. |
198 | bool getPotentialCopiesOfStoredValue( |
199 | Attributor &A, StoreInst &SI, SmallSetVector<Value *, 4> &PotentialCopies, |
200 | const AbstractAttribute &QueryingAA, bool &UsedAssumedInformation); |
201 | |
202 | } // namespace AA |
203 | |
204 | /// The value passed to the line option that defines the maximal initialization |
205 | /// chain length. |
206 | extern unsigned MaxInitializationChainLength; |
207 | |
208 | ///{ |
209 | enum class ChangeStatus { |
210 | CHANGED, |
211 | UNCHANGED, |
212 | }; |
213 | |
214 | ChangeStatus operator|(ChangeStatus l, ChangeStatus r); |
215 | ChangeStatus &operator|=(ChangeStatus &l, ChangeStatus r); |
216 | ChangeStatus operator&(ChangeStatus l, ChangeStatus r); |
217 | ChangeStatus &operator&=(ChangeStatus &l, ChangeStatus r); |
218 | |
219 | enum class DepClassTy { |
220 | REQUIRED, ///< The target cannot be valid if the source is not. |
221 | OPTIONAL, ///< The target may be valid if the source is not. |
222 | NONE, ///< Do not track a dependence between source and target. |
223 | }; |
224 | ///} |
225 | |
226 | /// The data structure for the nodes of a dependency graph |
227 | struct AADepGraphNode { |
228 | public: |
229 | virtual ~AADepGraphNode(){}; |
230 | using DepTy = PointerIntPair<AADepGraphNode *, 1>; |
231 | |
232 | protected: |
233 | /// Set of dependency graph nodes which should be updated if this one |
234 | /// is updated. The bit encodes if it is optional. |
235 | TinyPtrVector<DepTy> Deps; |
236 | |
237 | static AADepGraphNode *DepGetVal(DepTy &DT) { return DT.getPointer(); } |
238 | static AbstractAttribute *DepGetValAA(DepTy &DT) { |
239 | return cast<AbstractAttribute>(DT.getPointer()); |
240 | } |
241 | |
242 | operator AbstractAttribute *() { return cast<AbstractAttribute>(this); } |
243 | |
244 | public: |
245 | using iterator = |
246 | mapped_iterator<TinyPtrVector<DepTy>::iterator, decltype(&DepGetVal)>; |
247 | using aaiterator = |
248 | mapped_iterator<TinyPtrVector<DepTy>::iterator, decltype(&DepGetValAA)>; |
249 | |
250 | aaiterator begin() { return aaiterator(Deps.begin(), &DepGetValAA); } |
251 | aaiterator end() { return aaiterator(Deps.end(), &DepGetValAA); } |
252 | iterator child_begin() { return iterator(Deps.begin(), &DepGetVal); } |
253 | iterator child_end() { return iterator(Deps.end(), &DepGetVal); } |
254 | |
255 | virtual void print(raw_ostream &OS) const { OS << "AADepNode Impl\n"; } |
256 | TinyPtrVector<DepTy> &getDeps() { return Deps; } |
257 | |
258 | friend struct Attributor; |
259 | friend struct AADepGraph; |
260 | }; |
261 | |
262 | /// The data structure for the dependency graph |
263 | /// |
264 | /// Note that in this graph if there is an edge from A to B (A -> B), |
265 | /// then it means that B depends on A, and when the state of A is |
266 | /// updated, node B should also be updated |
267 | struct AADepGraph { |
268 | AADepGraph() {} |
269 | ~AADepGraph() {} |
270 | |
271 | using DepTy = AADepGraphNode::DepTy; |
272 | static AADepGraphNode *DepGetVal(DepTy &DT) { return DT.getPointer(); } |
273 | using iterator = |
274 | mapped_iterator<TinyPtrVector<DepTy>::iterator, decltype(&DepGetVal)>; |
275 | |
276 | /// There is no root node for the dependency graph. But the SCCIterator |
277 | /// requires a single entry point, so we maintain a fake("synthetic") root |
278 | /// node that depends on every node. |
279 | AADepGraphNode SyntheticRoot; |
280 | AADepGraphNode *GetEntryNode() { return &SyntheticRoot; } |
281 | |
282 | iterator begin() { return SyntheticRoot.child_begin(); } |
283 | iterator end() { return SyntheticRoot.child_end(); } |
284 | |
285 | void viewGraph(); |
286 | |
287 | /// Dump graph to file |
288 | void dumpGraph(); |
289 | |
290 | /// Print dependency graph |
291 | void print(); |
292 | }; |
293 | |
294 | /// Helper to describe and deal with positions in the LLVM-IR. |
295 | /// |
296 | /// A position in the IR is described by an anchor value and an "offset" that |
297 | /// could be the argument number, for call sites and arguments, or an indicator |
298 | /// of the "position kind". The kinds, specified in the Kind enum below, include |
299 | /// the locations in the attribute list, i.a., function scope and return value, |
300 | /// as well as a distinction between call sites and functions. Finally, there |
301 | /// are floating values that do not have a corresponding attribute list |
302 | /// position. |
303 | struct IRPosition { |
304 | // NOTE: In the future this definition can be changed to support recursive |
305 | // functions. |
306 | using CallBaseContext = CallBase; |
307 | |
308 | /// The positions we distinguish in the IR. |
309 | enum Kind : char { |
310 | IRP_INVALID, ///< An invalid position. |
311 | IRP_FLOAT, ///< A position that is not associated with a spot suitable |
312 | ///< for attributes. This could be any value or instruction. |
313 | IRP_RETURNED, ///< An attribute for the function return value. |
314 | IRP_CALL_SITE_RETURNED, ///< An attribute for a call site return value. |
315 | IRP_FUNCTION, ///< An attribute for a function (scope). |
316 | IRP_CALL_SITE, ///< An attribute for a call site (function scope). |
317 | IRP_ARGUMENT, ///< An attribute for a function argument. |
318 | IRP_CALL_SITE_ARGUMENT, ///< An attribute for a call site argument. |
319 | }; |
320 | |
321 | /// Default constructor available to create invalid positions implicitly. All |
322 | /// other positions need to be created explicitly through the appropriate |
323 | /// static member function. |
324 | IRPosition() : Enc(nullptr, ENC_VALUE) { verify(); } |
325 | |
326 | /// Create a position describing the value of \p V. |
327 | static const IRPosition value(const Value &V, |
328 | const CallBaseContext *CBContext = nullptr) { |
329 | if (auto *Arg = dyn_cast<Argument>(&V)) |
330 | return IRPosition::argument(*Arg, CBContext); |
331 | if (auto *CB = dyn_cast<CallBase>(&V)) |
332 | return IRPosition::callsite_returned(*CB); |
333 | return IRPosition(const_cast<Value &>(V), IRP_FLOAT, CBContext); |
334 | } |
335 | |
336 | /// Create a position describing the function scope of \p F. |
337 | /// \p CBContext is used for call base specific analysis. |
338 | static const IRPosition function(const Function &F, |
339 | const CallBaseContext *CBContext = nullptr) { |
340 | return IRPosition(const_cast<Function &>(F), IRP_FUNCTION, CBContext); |
341 | } |
342 | |
343 | /// Create a position describing the returned value of \p F. |
344 | /// \p CBContext is used for call base specific analysis. |
345 | static const IRPosition returned(const Function &F, |
346 | const CallBaseContext *CBContext = nullptr) { |
347 | return IRPosition(const_cast<Function &>(F), IRP_RETURNED, CBContext); |
348 | } |
349 | |
350 | /// Create a position describing the argument \p Arg. |
351 | /// \p CBContext is used for call base specific analysis. |
352 | static const IRPosition argument(const Argument &Arg, |
353 | const CallBaseContext *CBContext = nullptr) { |
354 | return IRPosition(const_cast<Argument &>(Arg), IRP_ARGUMENT, CBContext); |
355 | } |
356 | |
357 | /// Create a position describing the function scope of \p CB. |
358 | static const IRPosition callsite_function(const CallBase &CB) { |
359 | return IRPosition(const_cast<CallBase &>(CB), IRP_CALL_SITE); |
360 | } |
361 | |
362 | /// Create a position describing the returned value of \p CB. |
363 | static const IRPosition callsite_returned(const CallBase &CB) { |
364 | return IRPosition(const_cast<CallBase &>(CB), IRP_CALL_SITE_RETURNED); |
365 | } |
366 | |
367 | /// Create a position describing the argument of \p CB at position \p ArgNo. |
368 | static const IRPosition callsite_argument(const CallBase &CB, |
369 | unsigned ArgNo) { |
370 | return IRPosition(const_cast<Use &>(CB.getArgOperandUse(ArgNo)), |
371 | IRP_CALL_SITE_ARGUMENT); |
372 | } |
373 | |
374 | /// Create a position describing the argument of \p ACS at position \p ArgNo. |
375 | static const IRPosition callsite_argument(AbstractCallSite ACS, |
376 | unsigned ArgNo) { |
377 | if (ACS.getNumArgOperands() <= ArgNo) |
378 | return IRPosition(); |
379 | int CSArgNo = ACS.getCallArgOperandNo(ArgNo); |
380 | if (CSArgNo >= 0) |
381 | return IRPosition::callsite_argument( |
382 | cast<CallBase>(*ACS.getInstruction()), CSArgNo); |
383 | return IRPosition(); |
384 | } |
385 | |
386 | /// Create a position with function scope matching the "context" of \p IRP. |
387 | /// If \p IRP is a call site (see isAnyCallSitePosition()) then the result |
388 | /// will be a call site position, otherwise the function position of the |
389 | /// associated function. |
390 | static const IRPosition |
391 | function_scope(const IRPosition &IRP, |
392 | const CallBaseContext *CBContext = nullptr) { |
393 | if (IRP.isAnyCallSitePosition()) { |
394 | return IRPosition::callsite_function( |
395 | cast<CallBase>(IRP.getAnchorValue())); |
396 | } |
397 | assert(IRP.getAssociatedFunction())((void)0); |
398 | return IRPosition::function(*IRP.getAssociatedFunction(), CBContext); |
399 | } |
400 | |
401 | bool operator==(const IRPosition &RHS) const { |
402 | return Enc == RHS.Enc && RHS.CBContext == CBContext; |
403 | } |
404 | bool operator!=(const IRPosition &RHS) const { return !(*this == RHS); } |
405 | |
406 | /// Return the value this abstract attribute is anchored with. |
407 | /// |
408 | /// The anchor value might not be the associated value if the latter is not |
409 | /// sufficient to determine where arguments will be manifested. This is, so |
410 | /// far, only the case for call site arguments as the value is not sufficient |
411 | /// to pinpoint them. Instead, we can use the call site as an anchor. |
412 | Value &getAnchorValue() const { |
413 | switch (getEncodingBits()) { |
414 | case ENC_VALUE: |
415 | case ENC_RETURNED_VALUE: |
416 | case ENC_FLOATING_FUNCTION: |
417 | return *getAsValuePtr(); |
418 | case ENC_CALL_SITE_ARGUMENT_USE: |
419 | return *(getAsUsePtr()->getUser()); |
420 | default: |
421 | llvm_unreachable("Unkown encoding!")__builtin_unreachable(); |
422 | }; |
423 | } |
424 | |
425 | /// Return the associated function, if any. |
426 | Function *getAssociatedFunction() const { |
427 | if (auto *CB = dyn_cast<CallBase>(&getAnchorValue())) { |
428 | // We reuse the logic that associates callback calles to arguments of a |
429 | // call site here to identify the callback callee as the associated |
430 | // function. |
431 | if (Argument *Arg = getAssociatedArgument()) |
432 | return Arg->getParent(); |
433 | return CB->getCalledFunction(); |
434 | } |
435 | return getAnchorScope(); |
436 | } |
437 | |
438 | /// Return the associated argument, if any. |
439 | Argument *getAssociatedArgument() const; |
440 | |
441 | /// Return true if the position refers to a function interface, that is the |
442 | /// function scope, the function return, or an argument. |
443 | bool isFnInterfaceKind() const { |
444 | switch (getPositionKind()) { |
445 | case IRPosition::IRP_FUNCTION: |
446 | case IRPosition::IRP_RETURNED: |
447 | case IRPosition::IRP_ARGUMENT: |
448 | return true; |
449 | default: |
450 | return false; |
451 | } |
452 | } |
453 | |
454 | /// Return the Function surrounding the anchor value. |
455 | Function *getAnchorScope() const { |
456 | Value &V = getAnchorValue(); |
457 | if (isa<Function>(V)) |
458 | return &cast<Function>(V); |
459 | if (isa<Argument>(V)) |
460 | return cast<Argument>(V).getParent(); |
461 | if (isa<Instruction>(V)) |
462 | return cast<Instruction>(V).getFunction(); |
463 | return nullptr; |
464 | } |
465 | |
466 | /// Return the context instruction, if any. |
467 | Instruction *getCtxI() const { |
468 | Value &V = getAnchorValue(); |
469 | if (auto *I = dyn_cast<Instruction>(&V)) |
470 | return I; |
471 | if (auto *Arg = dyn_cast<Argument>(&V)) |
472 | if (!Arg->getParent()->isDeclaration()) |
473 | return &Arg->getParent()->getEntryBlock().front(); |
474 | if (auto *F = dyn_cast<Function>(&V)) |
475 | if (!F->isDeclaration()) |
476 | return &(F->getEntryBlock().front()); |
477 | return nullptr; |
478 | } |
479 | |
480 | /// Return the value this abstract attribute is associated with. |
481 | Value &getAssociatedValue() const { |
482 | if (getCallSiteArgNo() < 0 || isa<Argument>(&getAnchorValue())) |
483 | return getAnchorValue(); |
484 | assert(isa<CallBase>(&getAnchorValue()) && "Expected a call base!")((void)0); |
485 | return *cast<CallBase>(&getAnchorValue()) |
486 | ->getArgOperand(getCallSiteArgNo()); |
487 | } |
488 | |
489 | /// Return the type this abstract attribute is associated with. |
490 | Type *getAssociatedType() const { |
491 | if (getPositionKind() == IRPosition::IRP_RETURNED) |
492 | return getAssociatedFunction()->getReturnType(); |
493 | return getAssociatedValue().getType(); |
494 | } |
495 | |
496 | /// Return the callee argument number of the associated value if it is an |
497 | /// argument or call site argument, otherwise a negative value. In contrast to |
498 | /// `getCallSiteArgNo` this method will always return the "argument number" |
499 | /// from the perspective of the callee. This may not the same as the call site |
500 | /// if this is a callback call. |
501 | int getCalleeArgNo() const { |
502 | return getArgNo(/* CallbackCalleeArgIfApplicable */ true); |
503 | } |
504 | |
505 | /// Return the call site argument number of the associated value if it is an |
506 | /// argument or call site argument, otherwise a negative value. In contrast to |
507 | /// `getCalleArgNo` this method will always return the "operand number" from |
508 | /// the perspective of the call site. This may not the same as the callee |
509 | /// perspective if this is a callback call. |
510 | int getCallSiteArgNo() const { |
511 | return getArgNo(/* CallbackCalleeArgIfApplicable */ false); |
512 | } |
513 | |
514 | /// Return the index in the attribute list for this position. |
515 | unsigned getAttrIdx() const { |
516 | switch (getPositionKind()) { |
517 | case IRPosition::IRP_INVALID: |
518 | case IRPosition::IRP_FLOAT: |
519 | break; |
520 | case IRPosition::IRP_FUNCTION: |
521 | case IRPosition::IRP_CALL_SITE: |
522 | return AttributeList::FunctionIndex; |
523 | case IRPosition::IRP_RETURNED: |
524 | case IRPosition::IRP_CALL_SITE_RETURNED: |
525 | return AttributeList::ReturnIndex; |
526 | case IRPosition::IRP_ARGUMENT: |
527 | case IRPosition::IRP_CALL_SITE_ARGUMENT: |
528 | return getCallSiteArgNo() + AttributeList::FirstArgIndex; |
529 | } |
530 | llvm_unreachable(__builtin_unreachable() |
531 | "There is no attribute index for a floating or invalid position!")__builtin_unreachable(); |
532 | } |
533 | |
534 | /// Return the associated position kind. |
535 | Kind getPositionKind() const { |
536 | char EncodingBits = getEncodingBits(); |
537 | if (EncodingBits == ENC_CALL_SITE_ARGUMENT_USE) |
538 | return IRP_CALL_SITE_ARGUMENT; |
539 | if (EncodingBits == ENC_FLOATING_FUNCTION) |
540 | return IRP_FLOAT; |
541 | |
542 | Value *V = getAsValuePtr(); |
543 | if (!V) |
544 | return IRP_INVALID; |
545 | if (isa<Argument>(V)) |
546 | return IRP_ARGUMENT; |
547 | if (isa<Function>(V)) |
548 | return isReturnPosition(EncodingBits) ? IRP_RETURNED : IRP_FUNCTION; |
549 | if (isa<CallBase>(V)) |
550 | return isReturnPosition(EncodingBits) ? IRP_CALL_SITE_RETURNED |
551 | : IRP_CALL_SITE; |
552 | return IRP_FLOAT; |
553 | } |
554 | |
555 | /// TODO: Figure out if the attribute related helper functions should live |
556 | /// here or somewhere else. |
557 | |
558 | /// Return true if any kind in \p AKs existing in the IR at a position that |
559 | /// will affect this one. See also getAttrs(...). |
560 | /// \param IgnoreSubsumingPositions Flag to determine if subsuming positions, |
561 | /// e.g., the function position if this is an |
562 | /// argument position, should be ignored. |
563 | bool hasAttr(ArrayRef<Attribute::AttrKind> AKs, |
564 | bool IgnoreSubsumingPositions = false, |
565 | Attributor *A = nullptr) const; |
566 | |
567 | /// Return the attributes of any kind in \p AKs existing in the IR at a |
568 | /// position that will affect this one. While each position can only have a |
569 | /// single attribute of any kind in \p AKs, there are "subsuming" positions |
570 | /// that could have an attribute as well. This method returns all attributes |
571 | /// found in \p Attrs. |
572 | /// \param IgnoreSubsumingPositions Flag to determine if subsuming positions, |
573 | /// e.g., the function position if this is an |
574 | /// argument position, should be ignored. |
575 | void getAttrs(ArrayRef<Attribute::AttrKind> AKs, |
576 | SmallVectorImpl<Attribute> &Attrs, |
577 | bool IgnoreSubsumingPositions = false, |
578 | Attributor *A = nullptr) const; |
579 | |
580 | /// Remove the attribute of kind \p AKs existing in the IR at this position. |
581 | void removeAttrs(ArrayRef<Attribute::AttrKind> AKs) const { |
582 | if (getPositionKind() == IRP_INVALID || getPositionKind() == IRP_FLOAT) |
583 | return; |
584 | |
585 | AttributeList AttrList; |
586 | auto *CB = dyn_cast<CallBase>(&getAnchorValue()); |
587 | if (CB) |
588 | AttrList = CB->getAttributes(); |
589 | else |
590 | AttrList = getAssociatedFunction()->getAttributes(); |
591 | |
592 | LLVMContext &Ctx = getAnchorValue().getContext(); |
593 | for (Attribute::AttrKind AK : AKs) |
594 | AttrList = AttrList.removeAttribute(Ctx, getAttrIdx(), AK); |
595 | |
596 | if (CB) |
597 | CB->setAttributes(AttrList); |
598 | else |
599 | getAssociatedFunction()->setAttributes(AttrList); |
600 | } |
601 | |
602 | bool isAnyCallSitePosition() const { |
603 | switch (getPositionKind()) { |
604 | case IRPosition::IRP_CALL_SITE: |
605 | case IRPosition::IRP_CALL_SITE_RETURNED: |
606 | case IRPosition::IRP_CALL_SITE_ARGUMENT: |
607 | return true; |
608 | default: |
609 | return false; |
610 | } |
611 | } |
612 | |
613 | /// Return true if the position is an argument or call site argument. |
614 | bool isArgumentPosition() const { |
615 | switch (getPositionKind()) { |
616 | case IRPosition::IRP_ARGUMENT: |
617 | case IRPosition::IRP_CALL_SITE_ARGUMENT: |
618 | return true; |
619 | default: |
620 | return false; |
621 | } |
622 | } |
623 | |
624 | /// Return the same position without the call base context. |
625 | IRPosition stripCallBaseContext() const { |
626 | IRPosition Result = *this; |
627 | Result.CBContext = nullptr; |
628 | return Result; |
629 | } |
630 | |
631 | /// Get the call base context from the position. |
632 | const CallBaseContext *getCallBaseContext() const { return CBContext; } |
633 | |
634 | /// Check if the position has any call base context. |
635 | bool hasCallBaseContext() const { return CBContext != nullptr; } |
636 | |
637 | /// Special DenseMap key values. |
638 | /// |
639 | ///{ |
640 | static const IRPosition EmptyKey; |
641 | static const IRPosition TombstoneKey; |
642 | ///} |
643 | |
644 | /// Conversion into a void * to allow reuse of pointer hashing. |
645 | operator void *() const { return Enc.getOpaqueValue(); } |
646 | |
647 | private: |
648 | /// Private constructor for special values only! |
649 | explicit IRPosition(void *Ptr, const CallBaseContext *CBContext = nullptr) |
650 | : CBContext(CBContext) { |
651 | Enc.setFromOpaqueValue(Ptr); |
652 | } |
653 | |
654 | /// IRPosition anchored at \p AnchorVal with kind/argument numbet \p PK. |
655 | explicit IRPosition(Value &AnchorVal, Kind PK, |
656 | const CallBaseContext *CBContext = nullptr) |
657 | : CBContext(CBContext) { |
658 | switch (PK) { |
659 | case IRPosition::IRP_INVALID: |
660 | llvm_unreachable("Cannot create invalid IRP with an anchor value!")__builtin_unreachable(); |
661 | break; |
662 | case IRPosition::IRP_FLOAT: |
663 | // Special case for floating functions. |
664 | if (isa<Function>(AnchorVal)) |
665 | Enc = {&AnchorVal, ENC_FLOATING_FUNCTION}; |
666 | else |
667 | Enc = {&AnchorVal, ENC_VALUE}; |
668 | break; |
669 | case IRPosition::IRP_FUNCTION: |
670 | case IRPosition::IRP_CALL_SITE: |
671 | Enc = {&AnchorVal, ENC_VALUE}; |
672 | break; |
673 | case IRPosition::IRP_RETURNED: |
674 | case IRPosition::IRP_CALL_SITE_RETURNED: |
675 | Enc = {&AnchorVal, ENC_RETURNED_VALUE}; |
676 | break; |
677 | case IRPosition::IRP_ARGUMENT: |
678 | Enc = {&AnchorVal, ENC_VALUE}; |
679 | break; |
680 | case IRPosition::IRP_CALL_SITE_ARGUMENT: |
681 | llvm_unreachable(__builtin_unreachable() |
682 | "Cannot create call site argument IRP with an anchor value!")__builtin_unreachable(); |
683 | break; |
684 | } |
685 | verify(); |
686 | } |
687 | |
688 | /// Return the callee argument number of the associated value if it is an |
689 | /// argument or call site argument. See also `getCalleeArgNo` and |
690 | /// `getCallSiteArgNo`. |
691 | int getArgNo(bool CallbackCalleeArgIfApplicable) const { |
692 | if (CallbackCalleeArgIfApplicable) |
693 | if (Argument *Arg = getAssociatedArgument()) |
694 | return Arg->getArgNo(); |
695 | switch (getPositionKind()) { |
696 | case IRPosition::IRP_ARGUMENT: |
697 | return cast<Argument>(getAsValuePtr())->getArgNo(); |
698 | case IRPosition::IRP_CALL_SITE_ARGUMENT: { |
699 | Use &U = *getAsUsePtr(); |
700 | return cast<CallBase>(U.getUser())->getArgOperandNo(&U); |
701 | } |
702 | default: |
703 | return -1; |
704 | } |
705 | } |
706 | |
707 | /// IRPosition for the use \p U. The position kind \p PK needs to be |
708 | /// IRP_CALL_SITE_ARGUMENT, the anchor value is the user, the associated value |
709 | /// the used value. |
710 | explicit IRPosition(Use &U, Kind PK) { |
711 | assert(PK == IRP_CALL_SITE_ARGUMENT &&((void)0) |
712 | "Use constructor is for call site arguments only!")((void)0); |
713 | Enc = {&U, ENC_CALL_SITE_ARGUMENT_USE}; |
714 | verify(); |
715 | } |
716 | |
717 | /// Verify internal invariants. |
718 | void verify(); |
719 | |
720 | /// Return the attributes of kind \p AK existing in the IR as attribute. |
721 | bool getAttrsFromIRAttr(Attribute::AttrKind AK, |
722 | SmallVectorImpl<Attribute> &Attrs) const; |
723 | |
724 | /// Return the attributes of kind \p AK existing in the IR as operand bundles |
725 | /// of an llvm.assume. |
726 | bool getAttrsFromAssumes(Attribute::AttrKind AK, |
727 | SmallVectorImpl<Attribute> &Attrs, |
728 | Attributor &A) const; |
729 | |
730 | /// Return the underlying pointer as Value *, valid for all positions but |
731 | /// IRP_CALL_SITE_ARGUMENT. |
732 | Value *getAsValuePtr() const { |
733 | assert(getEncodingBits() != ENC_CALL_SITE_ARGUMENT_USE &&((void)0) |
734 | "Not a value pointer!")((void)0); |
735 | return reinterpret_cast<Value *>(Enc.getPointer()); |
736 | } |
737 | |
738 | /// Return the underlying pointer as Use *, valid only for |
739 | /// IRP_CALL_SITE_ARGUMENT positions. |
740 | Use *getAsUsePtr() const { |
741 | assert(getEncodingBits() == ENC_CALL_SITE_ARGUMENT_USE &&((void)0) |
742 | "Not a value pointer!")((void)0); |
743 | return reinterpret_cast<Use *>(Enc.getPointer()); |
744 | } |
745 | |
746 | /// Return true if \p EncodingBits describe a returned or call site returned |
747 | /// position. |
748 | static bool isReturnPosition(char EncodingBits) { |
749 | return EncodingBits == ENC_RETURNED_VALUE; |
750 | } |
751 | |
752 | /// Return true if the encoding bits describe a returned or call site returned |
753 | /// position. |
754 | bool isReturnPosition() const { return isReturnPosition(getEncodingBits()); } |
755 | |
756 | /// The encoding of the IRPosition is a combination of a pointer and two |
757 | /// encoding bits. The values of the encoding bits are defined in the enum |
758 | /// below. The pointer is either a Value* (for the first three encoding bit |
759 | /// combinations) or Use* (for ENC_CALL_SITE_ARGUMENT_USE). |
760 | /// |
761 | ///{ |
762 | enum { |
763 | ENC_VALUE = 0b00, |
764 | ENC_RETURNED_VALUE = 0b01, |
765 | ENC_FLOATING_FUNCTION = 0b10, |
766 | ENC_CALL_SITE_ARGUMENT_USE = 0b11, |
767 | }; |
768 | |
769 | // Reserve the maximal amount of bits so there is no need to mask out the |
770 | // remaining ones. We will not encode anything else in the pointer anyway. |
771 | static constexpr int NumEncodingBits = |
772 | PointerLikeTypeTraits<void *>::NumLowBitsAvailable; |
773 | static_assert(NumEncodingBits >= 2, "At least two bits are required!"); |
774 | |
775 | /// The pointer with the encoding bits. |
776 | PointerIntPair<void *, NumEncodingBits, char> Enc; |
777 | ///} |
778 | |
779 | /// Call base context. Used for callsite specific analysis. |
780 | const CallBaseContext *CBContext = nullptr; |
781 | |
782 | /// Return the encoding bits. |
783 | char getEncodingBits() const { return Enc.getInt(); } |
784 | }; |
785 | |
786 | /// Helper that allows IRPosition as a key in a DenseMap. |
787 | template <> struct DenseMapInfo<IRPosition> { |
788 | static inline IRPosition getEmptyKey() { return IRPosition::EmptyKey; } |
789 | static inline IRPosition getTombstoneKey() { |
790 | return IRPosition::TombstoneKey; |
791 | } |
792 | static unsigned getHashValue(const IRPosition &IRP) { |
793 | return (DenseMapInfo<void *>::getHashValue(IRP) << 4) ^ |
794 | (DenseMapInfo<Value *>::getHashValue(IRP.getCallBaseContext())); |
795 | } |
796 | |
797 | static bool isEqual(const IRPosition &a, const IRPosition &b) { |
798 | return a == b; |
799 | } |
800 | }; |
801 | |
802 | /// A visitor class for IR positions. |
803 | /// |
804 | /// Given a position P, the SubsumingPositionIterator allows to visit "subsuming |
805 | /// positions" wrt. attributes/information. Thus, if a piece of information |
806 | /// holds for a subsuming position, it also holds for the position P. |
807 | /// |
808 | /// The subsuming positions always include the initial position and then, |
809 | /// depending on the position kind, additionally the following ones: |
810 | /// - for IRP_RETURNED: |
811 | /// - the function (IRP_FUNCTION) |
812 | /// - for IRP_ARGUMENT: |
813 | /// - the function (IRP_FUNCTION) |
814 | /// - for IRP_CALL_SITE: |
815 | /// - the callee (IRP_FUNCTION), if known |
816 | /// - for IRP_CALL_SITE_RETURNED: |
817 | /// - the callee (IRP_RETURNED), if known |
818 | /// - the call site (IRP_FUNCTION) |
819 | /// - the callee (IRP_FUNCTION), if known |
820 | /// - for IRP_CALL_SITE_ARGUMENT: |
821 | /// - the argument of the callee (IRP_ARGUMENT), if known |
822 | /// - the callee (IRP_FUNCTION), if known |
823 | /// - the position the call site argument is associated with if it is not |
824 | /// anchored to the call site, e.g., if it is an argument then the argument |
825 | /// (IRP_ARGUMENT) |
826 | class SubsumingPositionIterator { |
827 | SmallVector<IRPosition, 4> IRPositions; |
828 | using iterator = decltype(IRPositions)::iterator; |
829 | |
830 | public: |
831 | SubsumingPositionIterator(const IRPosition &IRP); |
832 | iterator begin() { return IRPositions.begin(); } |
833 | iterator end() { return IRPositions.end(); } |
834 | }; |
835 | |
836 | /// Wrapper for FunctoinAnalysisManager. |
837 | struct AnalysisGetter { |
838 | template <typename Analysis> |
839 | typename Analysis::Result *getAnalysis(const Function &F) { |
840 | if (!FAM || !F.getParent()) |
841 | return nullptr; |
842 | return &FAM->getResult<Analysis>(const_cast<Function &>(F)); |
843 | } |
844 | |
845 | AnalysisGetter(FunctionAnalysisManager &FAM) : FAM(&FAM) {} |
846 | AnalysisGetter() {} |
847 | |
848 | private: |
849 | FunctionAnalysisManager *FAM = nullptr; |
850 | }; |
851 | |
852 | /// Data structure to hold cached (LLVM-IR) information. |
853 | /// |
854 | /// All attributes are given an InformationCache object at creation time to |
855 | /// avoid inspection of the IR by all of them individually. This default |
856 | /// InformationCache will hold information required by 'default' attributes, |
857 | /// thus the ones deduced when Attributor::identifyDefaultAbstractAttributes(..) |
858 | /// is called. |
859 | /// |
860 | /// If custom abstract attributes, registered manually through |
861 | /// Attributor::registerAA(...), need more information, especially if it is not |
862 | /// reusable, it is advised to inherit from the InformationCache and cast the |
863 | /// instance down in the abstract attributes. |
864 | struct InformationCache { |
865 | InformationCache(const Module &M, AnalysisGetter &AG, |
866 | BumpPtrAllocator &Allocator, SetVector<Function *> *CGSCC) |
867 | : DL(M.getDataLayout()), Allocator(Allocator), |
868 | Explorer( |
869 | /* ExploreInterBlock */ true, /* ExploreCFGForward */ true, |
870 | /* ExploreCFGBackward */ true, |
871 | /* LIGetter */ |
872 | [&](const Function &F) { return AG.getAnalysis<LoopAnalysis>(F); }, |
873 | /* DTGetter */ |
874 | [&](const Function &F) { |
875 | return AG.getAnalysis<DominatorTreeAnalysis>(F); |
876 | }, |
877 | /* PDTGetter */ |
878 | [&](const Function &F) { |
879 | return AG.getAnalysis<PostDominatorTreeAnalysis>(F); |
880 | }), |
881 | AG(AG), CGSCC(CGSCC), TargetTriple(M.getTargetTriple()) { |
882 | if (CGSCC) |
883 | initializeModuleSlice(*CGSCC); |
884 | } |
885 | |
886 | ~InformationCache() { |
887 | // The FunctionInfo objects are allocated via a BumpPtrAllocator, we call |
888 | // the destructor manually. |
889 | for (auto &It : FuncInfoMap) |
890 | It.getSecond()->~FunctionInfo(); |
891 | } |
892 | |
893 | /// Apply \p CB to all uses of \p F. If \p LookThroughConstantExprUses is |
894 | /// true, constant expression users are not given to \p CB but their uses are |
895 | /// traversed transitively. |
896 | template <typename CBTy> |
897 | static void foreachUse(Function &F, CBTy CB, |
898 | bool LookThroughConstantExprUses = true) { |
899 | SmallVector<Use *, 8> Worklist(make_pointer_range(F.uses())); |
900 | |
901 | for (unsigned Idx = 0; Idx < Worklist.size(); ++Idx) { |
902 | Use &U = *Worklist[Idx]; |
903 | |
904 | // Allow use in constant bitcasts and simply look through them. |
905 | if (LookThroughConstantExprUses && isa<ConstantExpr>(U.getUser())) { |
906 | for (Use &CEU : cast<ConstantExpr>(U.getUser())->uses()) |
907 | Worklist.push_back(&CEU); |
908 | continue; |
909 | } |
910 | |
911 | CB(U); |
912 | } |
913 | } |
914 | |
915 | /// Initialize the ModuleSlice member based on \p SCC. ModuleSlices contains |
916 | /// (a subset of) all functions that we can look at during this SCC traversal. |
917 | /// This includes functions (transitively) called from the SCC and the |
918 | /// (transitive) callers of SCC functions. We also can look at a function if |
919 | /// there is a "reference edge", i.a., if the function somehow uses (!=calls) |
920 | /// a function in the SCC or a caller of a function in the SCC. |
921 | void initializeModuleSlice(SetVector<Function *> &SCC) { |
922 | ModuleSlice.insert(SCC.begin(), SCC.end()); |
923 | |
924 | SmallPtrSet<Function *, 16> Seen; |
925 | SmallVector<Function *, 16> Worklist(SCC.begin(), SCC.end()); |
926 | while (!Worklist.empty()) { |
927 | Function *F = Worklist.pop_back_val(); |
928 | ModuleSlice.insert(F); |
929 | |
930 | for (Instruction &I : instructions(*F)) |
931 | if (auto *CB = dyn_cast<CallBase>(&I)) |
932 | if (Function *Callee = CB->getCalledFunction()) |
933 | if (Seen.insert(Callee).second) |
934 | Worklist.push_back(Callee); |
935 | } |
936 | |
937 | Seen.clear(); |
938 | Worklist.append(SCC.begin(), SCC.end()); |
939 | while (!Worklist.empty()) { |
940 | Function *F = Worklist.pop_back_val(); |
941 | ModuleSlice.insert(F); |
942 | |
943 | // Traverse all transitive uses. |
944 | foreachUse(*F, [&](Use &U) { |
945 | if (auto *UsrI = dyn_cast<Instruction>(U.getUser())) |
946 | if (Seen.insert(UsrI->getFunction()).second) |
947 | Worklist.push_back(UsrI->getFunction()); |
948 | }); |
949 | } |
950 | } |
951 | |
952 | /// The slice of the module we are allowed to look at. |
953 | SmallPtrSet<Function *, 8> ModuleSlice; |
954 | |
955 | /// A vector type to hold instructions. |
956 | using InstructionVectorTy = SmallVector<Instruction *, 8>; |
957 | |
958 | /// A map type from opcodes to instructions with this opcode. |
959 | using OpcodeInstMapTy = DenseMap<unsigned, InstructionVectorTy *>; |
960 | |
961 | /// Return the map that relates "interesting" opcodes with all instructions |
962 | /// with that opcode in \p F. |
963 | OpcodeInstMapTy &getOpcodeInstMapForFunction(const Function &F) { |
964 | return getFunctionInfo(F).OpcodeInstMap; |
965 | } |
966 | |
967 | /// Return the instructions in \p F that may read or write memory. |
968 | InstructionVectorTy &getReadOrWriteInstsForFunction(const Function &F) { |
969 | return getFunctionInfo(F).RWInsts; |
970 | } |
971 | |
972 | /// Return MustBeExecutedContextExplorer |
973 | MustBeExecutedContextExplorer &getMustBeExecutedContextExplorer() { |
974 | return Explorer; |
975 | } |
976 | |
977 | /// Return TargetLibraryInfo for function \p F. |
978 | TargetLibraryInfo *getTargetLibraryInfoForFunction(const Function &F) { |
979 | return AG.getAnalysis<TargetLibraryAnalysis>(F); |
980 | } |
981 | |
982 | /// Return AliasAnalysis Result for function \p F. |
983 | AAResults *getAAResultsForFunction(const Function &F); |
984 | |
985 | /// Return true if \p Arg is involved in a must-tail call, thus the argument |
986 | /// of the caller or callee. |
987 | bool isInvolvedInMustTailCall(const Argument &Arg) { |
988 | FunctionInfo &FI = getFunctionInfo(*Arg.getParent()); |
989 | return FI.CalledViaMustTail || FI.ContainsMustTailCall; |
990 | } |
991 | |
992 | /// Return the analysis result from a pass \p AP for function \p F. |
993 | template <typename AP> |
994 | typename AP::Result *getAnalysisResultForFunction(const Function &F) { |
995 | return AG.getAnalysis<AP>(F); |
996 | } |
997 | |
998 | /// Return SCC size on call graph for function \p F or 0 if unknown. |
999 | unsigned getSccSize(const Function &F) { |
1000 | if (CGSCC && CGSCC->count(const_cast<Function *>(&F))) |
1001 | return CGSCC->size(); |
1002 | return 0; |
1003 | } |
1004 | |
1005 | /// Return datalayout used in the module. |
1006 | const DataLayout &getDL() { return DL; } |
1007 | |
1008 | /// Return the map conaining all the knowledge we have from `llvm.assume`s. |
1009 | const RetainedKnowledgeMap &getKnowledgeMap() const { return KnowledgeMap; } |
1010 | |
1011 | /// Return if \p To is potentially reachable form \p From or not |
1012 | /// If the same query was answered, return cached result |
1013 | bool getPotentiallyReachable(const Instruction &From, const Instruction &To) { |
1014 | auto KeyPair = std::make_pair(&From, &To); |
1015 | auto Iter = PotentiallyReachableMap.find(KeyPair); |
1016 | if (Iter != PotentiallyReachableMap.end()) |
1017 | return Iter->second; |
1018 | const Function &F = *From.getFunction(); |
1019 | bool Result = true; |
1020 | if (From.getFunction() == To.getFunction()) |
1021 | Result = isPotentiallyReachable(&From, &To, nullptr, |
1022 | AG.getAnalysis<DominatorTreeAnalysis>(F), |
1023 | AG.getAnalysis<LoopAnalysis>(F)); |
1024 | PotentiallyReachableMap.insert(std::make_pair(KeyPair, Result)); |
1025 | return Result; |
1026 | } |
1027 | |
1028 | /// Check whether \p F is part of module slice. |
1029 | bool isInModuleSlice(const Function &F) { |
1030 | return ModuleSlice.count(const_cast<Function *>(&F)); |
1031 | } |
1032 | |
1033 | /// Return true if the stack (llvm::Alloca) can be accessed by other threads. |
1034 | bool stackIsAccessibleByOtherThreads() { return !targetIsGPU(); } |
1035 | |
1036 | /// Return true if the target is a GPU. |
1037 | bool targetIsGPU() { |
1038 | return TargetTriple.isAMDGPU() || TargetTriple.isNVPTX(); |
1039 | } |
1040 | |
1041 | private: |
1042 | struct FunctionInfo { |
1043 | ~FunctionInfo(); |
1044 | |
1045 | /// A nested map that remembers all instructions in a function with a |
1046 | /// certain instruction opcode (Instruction::getOpcode()). |
1047 | OpcodeInstMapTy OpcodeInstMap; |
1048 | |
1049 | /// A map from functions to their instructions that may read or write |
1050 | /// memory. |
1051 | InstructionVectorTy RWInsts; |
1052 | |
1053 | /// Function is called by a `musttail` call. |
1054 | bool CalledViaMustTail; |
1055 | |
1056 | /// Function contains a `musttail` call. |
1057 | bool ContainsMustTailCall; |
1058 | }; |
1059 | |
1060 | /// A map type from functions to informatio about it. |
1061 | DenseMap<const Function *, FunctionInfo *> FuncInfoMap; |
1062 | |
1063 | /// Return information about the function \p F, potentially by creating it. |
1064 | FunctionInfo &getFunctionInfo(const Function &F) { |
1065 | FunctionInfo *&FI = FuncInfoMap[&F]; |
1066 | if (!FI) { |
1067 | FI = new (Allocator) FunctionInfo(); |
1068 | initializeInformationCache(F, *FI); |
1069 | } |
1070 | return *FI; |
1071 | } |
1072 | |
1073 | /// Initialize the function information cache \p FI for the function \p F. |
1074 | /// |
1075 | /// This method needs to be called for all function that might be looked at |
1076 | /// through the information cache interface *prior* to looking at them. |
1077 | void initializeInformationCache(const Function &F, FunctionInfo &FI); |
1078 | |
1079 | /// The datalayout used in the module. |
1080 | const DataLayout &DL; |
1081 | |
1082 | /// The allocator used to allocate memory, e.g. for `FunctionInfo`s. |
1083 | BumpPtrAllocator &Allocator; |
1084 | |
1085 | /// MustBeExecutedContextExplorer |
1086 | MustBeExecutedContextExplorer Explorer; |
1087 | |
1088 | /// A map with knowledge retained in `llvm.assume` instructions. |
1089 | RetainedKnowledgeMap KnowledgeMap; |
1090 | |
1091 | /// Getters for analysis. |
1092 | AnalysisGetter &AG; |
1093 | |
1094 | /// The underlying CGSCC, or null if not available. |
1095 | SetVector<Function *> *CGSCC; |
1096 | |
1097 | /// Set of inlineable functions |
1098 | SmallPtrSet<const Function *, 8> InlineableFunctions; |
1099 | |
1100 | /// A map for caching results of queries for isPotentiallyReachable |
1101 | DenseMap<std::pair<const Instruction *, const Instruction *>, bool> |
1102 | PotentiallyReachableMap; |
1103 | |
1104 | /// The triple describing the target machine. |
1105 | Triple TargetTriple; |
1106 | |
1107 | /// Give the Attributor access to the members so |
1108 | /// Attributor::identifyDefaultAbstractAttributes(...) can initialize them. |
1109 | friend struct Attributor; |
1110 | }; |
1111 | |
1112 | /// The fixpoint analysis framework that orchestrates the attribute deduction. |
1113 | /// |
1114 | /// The Attributor provides a general abstract analysis framework (guided |
1115 | /// fixpoint iteration) as well as helper functions for the deduction of |
1116 | /// (LLVM-IR) attributes. However, also other code properties can be deduced, |
1117 | /// propagated, and ultimately manifested through the Attributor framework. This |
1118 | /// is particularly useful if these properties interact with attributes and a |
1119 | /// co-scheduled deduction allows to improve the solution. Even if not, thus if |
1120 | /// attributes/properties are completely isolated, they should use the |
1121 | /// Attributor framework to reduce the number of fixpoint iteration frameworks |
1122 | /// in the code base. Note that the Attributor design makes sure that isolated |
1123 | /// attributes are not impacted, in any way, by others derived at the same time |
1124 | /// if there is no cross-reasoning performed. |
1125 | /// |
1126 | /// The public facing interface of the Attributor is kept simple and basically |
1127 | /// allows abstract attributes to one thing, query abstract attributes |
1128 | /// in-flight. There are two reasons to do this: |
1129 | /// a) The optimistic state of one abstract attribute can justify an |
1130 | /// optimistic state of another, allowing to framework to end up with an |
1131 | /// optimistic (=best possible) fixpoint instead of one based solely on |
1132 | /// information in the IR. |
1133 | /// b) This avoids reimplementing various kinds of lookups, e.g., to check |
1134 | /// for existing IR attributes, in favor of a single lookups interface |
1135 | /// provided by an abstract attribute subclass. |
1136 | /// |
1137 | /// NOTE: The mechanics of adding a new "concrete" abstract attribute are |
1138 | /// described in the file comment. |
1139 | struct Attributor { |
1140 | |
1141 | using OptimizationRemarkGetter = |
1142 | function_ref<OptimizationRemarkEmitter &(Function *)>; |
1143 | |
1144 | /// Constructor |
1145 | /// |
1146 | /// \param Functions The set of functions we are deriving attributes for. |
1147 | /// \param InfoCache Cache to hold various information accessible for |
1148 | /// the abstract attributes. |
1149 | /// \param CGUpdater Helper to update an underlying call graph. |
1150 | /// \param Allowed If not null, a set limiting the attribute opportunities. |
1151 | /// \param DeleteFns Whether to delete functions. |
1152 | /// \param RewriteSignatures Whether to rewrite function signatures. |
1153 | /// \param MaxFixedPointIterations Maximum number of iterations to run until |
1154 | /// fixpoint. |
1155 | Attributor(SetVector<Function *> &Functions, InformationCache &InfoCache, |
1156 | CallGraphUpdater &CGUpdater, |
1157 | DenseSet<const char *> *Allowed = nullptr, bool DeleteFns = true, |
1158 | bool RewriteSignatures = true) |
1159 | : Allocator(InfoCache.Allocator), Functions(Functions), |
1160 | InfoCache(InfoCache), CGUpdater(CGUpdater), Allowed(Allowed), |
1161 | DeleteFns(DeleteFns), RewriteSignatures(RewriteSignatures), |
1162 | MaxFixpointIterations(None), OREGetter(None), PassName("") {} |
1163 | |
1164 | /// Constructor |
1165 | /// |
1166 | /// \param Functions The set of functions we are deriving attributes for. |
1167 | /// \param InfoCache Cache to hold various information accessible for |
1168 | /// the abstract attributes. |
1169 | /// \param CGUpdater Helper to update an underlying call graph. |
1170 | /// \param Allowed If not null, a set limiting the attribute opportunities. |
1171 | /// \param DeleteFns Whether to delete functions |
1172 | /// \param MaxFixedPointIterations Maximum number of iterations to run until |
1173 | /// fixpoint. |
1174 | /// \param OREGetter A callback function that returns an ORE object from a |
1175 | /// Function pointer. |
1176 | /// \param PassName The name of the pass emitting remarks. |
1177 | Attributor(SetVector<Function *> &Functions, InformationCache &InfoCache, |
1178 | CallGraphUpdater &CGUpdater, DenseSet<const char *> *Allowed, |
1179 | bool DeleteFns, bool RewriteSignatures, |
1180 | Optional<unsigned> MaxFixpointIterations, |
1181 | OptimizationRemarkGetter OREGetter, const char *PassName) |
1182 | : Allocator(InfoCache.Allocator), Functions(Functions), |
1183 | InfoCache(InfoCache), CGUpdater(CGUpdater), Allowed(Allowed), |
1184 | DeleteFns(DeleteFns), RewriteSignatures(RewriteSignatures), |
1185 | MaxFixpointIterations(MaxFixpointIterations), |
1186 | OREGetter(Optional<OptimizationRemarkGetter>(OREGetter)), |
1187 | PassName(PassName) {} |
1188 | |
1189 | ~Attributor(); |
1190 | |
1191 | /// Run the analyses until a fixpoint is reached or enforced (timeout). |
1192 | /// |
1193 | /// The attributes registered with this Attributor can be used after as long |
1194 | /// as the Attributor is not destroyed (it owns the attributes now). |
1195 | /// |
1196 | /// \Returns CHANGED if the IR was changed, otherwise UNCHANGED. |
1197 | ChangeStatus run(); |
1198 | |
1199 | /// Lookup an abstract attribute of type \p AAType at position \p IRP. While |
1200 | /// no abstract attribute is found equivalent positions are checked, see |
1201 | /// SubsumingPositionIterator. Thus, the returned abstract attribute |
1202 | /// might be anchored at a different position, e.g., the callee if \p IRP is a |
1203 | /// call base. |
1204 | /// |
1205 | /// This method is the only (supported) way an abstract attribute can retrieve |
1206 | /// information from another abstract attribute. As an example, take an |
1207 | /// abstract attribute that determines the memory access behavior for a |
1208 | /// argument (readnone, readonly, ...). It should use `getAAFor` to get the |
1209 | /// most optimistic information for other abstract attributes in-flight, e.g. |
1210 | /// the one reasoning about the "captured" state for the argument or the one |
1211 | /// reasoning on the memory access behavior of the function as a whole. |
1212 | /// |
1213 | /// If the DepClass enum is set to `DepClassTy::None` the dependence from |
1214 | /// \p QueryingAA to the return abstract attribute is not automatically |
1215 | /// recorded. This should only be used if the caller will record the |
1216 | /// dependence explicitly if necessary, thus if it the returned abstract |
1217 | /// attribute is used for reasoning. To record the dependences explicitly use |
1218 | /// the `Attributor::recordDependence` method. |
1219 | template <typename AAType> |
1220 | const AAType &getAAFor(const AbstractAttribute &QueryingAA, |
1221 | const IRPosition &IRP, DepClassTy DepClass) { |
1222 | return getOrCreateAAFor<AAType>(IRP, &QueryingAA, DepClass, |
1223 | /* ForceUpdate */ false); |
1224 | } |
1225 | |
1226 | /// Similar to getAAFor but the return abstract attribute will be updated (via |
1227 | /// `AbstractAttribute::update`) even if it is found in the cache. This is |
1228 | /// especially useful for AAIsDead as changes in liveness can make updates |
1229 | /// possible/useful that were not happening before as the abstract attribute |
1230 | /// was assumed dead. |
1231 | template <typename AAType> |
1232 | const AAType &getAndUpdateAAFor(const AbstractAttribute &QueryingAA, |
1233 | const IRPosition &IRP, DepClassTy DepClass) { |
1234 | return getOrCreateAAFor<AAType>(IRP, &QueryingAA, DepClass, |
1235 | /* ForceUpdate */ true); |
1236 | } |
1237 | |
1238 | /// The version of getAAFor that allows to omit a querying abstract |
1239 | /// attribute. Using this after Attributor started running is restricted to |
1240 | /// only the Attributor itself. Initial seeding of AAs can be done via this |
1241 | /// function. |
1242 | /// NOTE: ForceUpdate is ignored in any stage other than the update stage. |
1243 | template <typename AAType> |
1244 | const AAType &getOrCreateAAFor(IRPosition IRP, |
1245 | const AbstractAttribute *QueryingAA, |
1246 | DepClassTy DepClass, bool ForceUpdate = false, |
1247 | bool UpdateAfterInit = true) { |
1248 | if (!shouldPropagateCallBaseContext(IRP)) |
1249 | IRP = IRP.stripCallBaseContext(); |
1250 | |
1251 | if (AAType *AAPtr = lookupAAFor<AAType>(IRP, QueryingAA, DepClass, |
1252 | /* AllowInvalidState */ true)) { |
1253 | if (ForceUpdate && Phase == AttributorPhase::UPDATE) |
1254 | updateAA(*AAPtr); |
1255 | return *AAPtr; |
1256 | } |
1257 | |
1258 | // No matching attribute found, create one. |
1259 | // Use the static create method. |
1260 | auto &AA = AAType::createForPosition(IRP, *this); |
1261 | |
1262 | // If we are currenty seeding attributes, enforce seeding rules. |
1263 | if (Phase == AttributorPhase::SEEDING && !shouldSeedAttribute(AA)) { |
1264 | AA.getState().indicatePessimisticFixpoint(); |
1265 | return AA; |
1266 | } |
1267 | |
1268 | registerAA(AA); |
1269 | |
1270 | // For now we ignore naked and optnone functions. |
1271 | bool Invalidate = Allowed && !Allowed->count(&AAType::ID); |
1272 | const Function *FnScope = IRP.getAnchorScope(); |
1273 | if (FnScope) |
1274 | Invalidate |= FnScope->hasFnAttribute(Attribute::Naked) || |
1275 | FnScope->hasFnAttribute(Attribute::OptimizeNone); |
1276 | |
1277 | // Avoid too many nested initializations to prevent a stack overflow. |
1278 | Invalidate |= InitializationChainLength > MaxInitializationChainLength; |
1279 | |
1280 | // Bootstrap the new attribute with an initial update to propagate |
1281 | // information, e.g., function -> call site. If it is not on a given |
1282 | // Allowed we will not perform updates at all. |
1283 | if (Invalidate) { |
1284 | AA.getState().indicatePessimisticFixpoint(); |
1285 | return AA; |
1286 | } |
1287 | |
1288 | { |
1289 | TimeTraceScope TimeScope(AA.getName() + "::initialize"); |
1290 | ++InitializationChainLength; |
1291 | AA.initialize(*this); |
1292 | --InitializationChainLength; |
1293 | } |
1294 | |
1295 | // Initialize and update is allowed for code outside of the current function |
1296 | // set, but only if it is part of module slice we are allowed to look at. |
1297 | // Only exception is AAIsDeadFunction whose initialization is prevented |
1298 | // directly, since we don't to compute it twice. |
1299 | if (FnScope && !Functions.count(const_cast<Function *>(FnScope))) { |
1300 | if (!getInfoCache().isInModuleSlice(*FnScope)) { |
1301 | AA.getState().indicatePessimisticFixpoint(); |
1302 | return AA; |
1303 | } |
1304 | } |
1305 | |
1306 | // If this is queried in the manifest stage, we force the AA to indicate |
1307 | // pessimistic fixpoint immediately. |
1308 | if (Phase == AttributorPhase::MANIFEST) { |
1309 | AA.getState().indicatePessimisticFixpoint(); |
1310 | return AA; |
1311 | } |
1312 | |
1313 | // Allow seeded attributes to declare dependencies. |
1314 | // Remember the seeding state. |
1315 | if (UpdateAfterInit) { |
1316 | AttributorPhase OldPhase = Phase; |
1317 | Phase = AttributorPhase::UPDATE; |
1318 | |
1319 | updateAA(AA); |
1320 | |
1321 | Phase = OldPhase; |
1322 | } |
1323 | |
1324 | if (QueryingAA && AA.getState().isValidState()) |
1325 | recordDependence(AA, const_cast<AbstractAttribute &>(*QueryingAA), |
1326 | DepClass); |
1327 | return AA; |
1328 | } |
1329 | template <typename AAType> |
1330 | const AAType &getOrCreateAAFor(const IRPosition &IRP) { |
1331 | return getOrCreateAAFor<AAType>(IRP, /* QueryingAA */ nullptr, |
1332 | DepClassTy::NONE); |
1333 | } |
1334 | |
1335 | /// Return the attribute of \p AAType for \p IRP if existing and valid. This |
1336 | /// also allows non-AA users lookup. |
1337 | template <typename AAType> |
1338 | AAType *lookupAAFor(const IRPosition &IRP, |
1339 | const AbstractAttribute *QueryingAA = nullptr, |
1340 | DepClassTy DepClass = DepClassTy::OPTIONAL, |
1341 | bool AllowInvalidState = false) { |
1342 | static_assert(std::is_base_of<AbstractAttribute, AAType>::value, |
1343 | "Cannot query an attribute with a type not derived from " |
1344 | "'AbstractAttribute'!"); |
1345 | // Lookup the abstract attribute of type AAType. If found, return it after |
1346 | // registering a dependence of QueryingAA on the one returned attribute. |
1347 | AbstractAttribute *AAPtr = AAMap.lookup({&AAType::ID, IRP}); |
1348 | if (!AAPtr) |
1349 | return nullptr; |
1350 | |
1351 | AAType *AA = static_cast<AAType *>(AAPtr); |
1352 | |
1353 | // Do not register a dependence on an attribute with an invalid state. |
1354 | if (DepClass != DepClassTy::NONE && QueryingAA && |
1355 | AA->getState().isValidState()) |
1356 | recordDependence(*AA, const_cast<AbstractAttribute &>(*QueryingAA), |
1357 | DepClass); |
1358 | |
1359 | // Return nullptr if this attribute has an invalid state. |
1360 | if (!AllowInvalidState && !AA->getState().isValidState()) |
1361 | return nullptr; |
1362 | return AA; |
1363 | } |
1364 | |
1365 | /// Explicitly record a dependence from \p FromAA to \p ToAA, that is if |
1366 | /// \p FromAA changes \p ToAA should be updated as well. |
1367 | /// |
1368 | /// This method should be used in conjunction with the `getAAFor` method and |
1369 | /// with the DepClass enum passed to the method set to None. This can |
1370 | /// be beneficial to avoid false dependences but it requires the users of |
1371 | /// `getAAFor` to explicitly record true dependences through this method. |
1372 | /// The \p DepClass flag indicates if the dependence is striclty necessary. |
1373 | /// That means for required dependences, if \p FromAA changes to an invalid |
1374 | /// state, \p ToAA can be moved to a pessimistic fixpoint because it required |
1375 | /// information from \p FromAA but none are available anymore. |
1376 | void recordDependence(const AbstractAttribute &FromAA, |
1377 | const AbstractAttribute &ToAA, DepClassTy DepClass); |
1378 | |
1379 | /// Introduce a new abstract attribute into the fixpoint analysis. |
1380 | /// |
1381 | /// Note that ownership of the attribute is given to the Attributor. It will |
1382 | /// invoke delete for the Attributor on destruction of the Attributor. |
1383 | /// |
1384 | /// Attributes are identified by their IR position (AAType::getIRPosition()) |
1385 | /// and the address of their static member (see AAType::ID). |
1386 | template <typename AAType> AAType ®isterAA(AAType &AA) { |
1387 | static_assert(std::is_base_of<AbstractAttribute, AAType>::value, |
1388 | "Cannot register an attribute with a type not derived from " |
1389 | "'AbstractAttribute'!"); |
1390 | // Put the attribute in the lookup map structure and the container we use to |
1391 | // keep track of all attributes. |
1392 | const IRPosition &IRP = AA.getIRPosition(); |
1393 | AbstractAttribute *&AAPtr = AAMap[{&AAType::ID, IRP}]; |
1394 | |
1395 | assert(!AAPtr && "Attribute already in map!")((void)0); |
1396 | AAPtr = &AA; |
1397 | |
1398 | // Register AA with the synthetic root only before the manifest stage. |
1399 | if (Phase == AttributorPhase::SEEDING || Phase == AttributorPhase::UPDATE) |
1400 | DG.SyntheticRoot.Deps.push_back( |
1401 | AADepGraphNode::DepTy(&AA, unsigned(DepClassTy::REQUIRED))); |
1402 | |
1403 | return AA; |
1404 | } |
1405 | |
1406 | /// Return the internal information cache. |
1407 | InformationCache &getInfoCache() { return InfoCache; } |
1408 | |
1409 | /// Return true if this is a module pass, false otherwise. |
1410 | bool isModulePass() const { |
1411 | return !Functions.empty() && |
1412 | Functions.size() == Functions.front()->getParent()->size(); |
1413 | } |
1414 | |
1415 | /// Return true if we derive attributes for \p Fn |
1416 | bool isRunOn(Function &Fn) const { |
1417 | return Functions.empty() || Functions.count(&Fn); |
1418 | } |
1419 | |
1420 | /// Determine opportunities to derive 'default' attributes in \p F and create |
1421 | /// abstract attribute objects for them. |
1422 | /// |
1423 | /// \param F The function that is checked for attribute opportunities. |
1424 | /// |
1425 | /// Note that abstract attribute instances are generally created even if the |
1426 | /// IR already contains the information they would deduce. The most important |
1427 | /// reason for this is the single interface, the one of the abstract attribute |
1428 | /// instance, which can be queried without the need to look at the IR in |
1429 | /// various places. |
1430 | void identifyDefaultAbstractAttributes(Function &F); |
1431 | |
1432 | /// Determine whether the function \p F is IPO amendable |
1433 | /// |
1434 | /// If a function is exactly defined or it has alwaysinline attribute |
1435 | /// and is viable to be inlined, we say it is IPO amendable |
1436 | bool isFunctionIPOAmendable(const Function &F) { |
1437 | return F.hasExactDefinition() || InfoCache.InlineableFunctions.count(&F); |
1438 | } |
1439 | |
1440 | /// Mark the internal function \p F as live. |
1441 | /// |
1442 | /// This will trigger the identification and initialization of attributes for |
1443 | /// \p F. |
1444 | void markLiveInternalFunction(const Function &F) { |
1445 | assert(F.hasLocalLinkage() &&((void)0) |
1446 | "Only local linkage is assumed dead initially.")((void)0); |
1447 | |
1448 | identifyDefaultAbstractAttributes(const_cast<Function &>(F)); |
1449 | } |
1450 | |
1451 | /// Helper function to remove callsite. |
1452 | void removeCallSite(CallInst *CI) { |
1453 | if (!CI) |
1454 | return; |
1455 | |
1456 | CGUpdater.removeCallSite(*CI); |
1457 | } |
1458 | |
1459 | /// Record that \p U is to be replaces with \p NV after information was |
1460 | /// manifested. This also triggers deletion of trivially dead istructions. |
1461 | bool changeUseAfterManifest(Use &U, Value &NV) { |
1462 | Value *&V = ToBeChangedUses[&U]; |
1463 | if (V && (V->stripPointerCasts() == NV.stripPointerCasts() || |
1464 | isa_and_nonnull<UndefValue>(V))) |
1465 | return false; |
1466 | assert((!V || V == &NV || isa<UndefValue>(NV)) &&((void)0) |
1467 | "Use was registered twice for replacement with different values!")((void)0); |
1468 | V = &NV; |
1469 | return true; |
1470 | } |
1471 | |
1472 | /// Helper function to replace all uses of \p V with \p NV. Return true if |
1473 | /// there is any change. The flag \p ChangeDroppable indicates if dropppable |
1474 | /// uses should be changed too. |
1475 | bool changeValueAfterManifest(Value &V, Value &NV, |
1476 | bool ChangeDroppable = true) { |
1477 | auto &Entry = ToBeChangedValues[&V]; |
1478 | Value *&CurNV = Entry.first; |
1479 | if (CurNV && (CurNV->stripPointerCasts() == NV.stripPointerCasts() || |
1480 | isa<UndefValue>(CurNV))) |
1481 | return false; |
1482 | assert((!CurNV || CurNV == &NV || isa<UndefValue>(NV)) &&((void)0) |
1483 | "Value replacement was registered twice with different values!")((void)0); |
1484 | CurNV = &NV; |
1485 | Entry.second = ChangeDroppable; |
1486 | return true; |
1487 | } |
1488 | |
1489 | /// Record that \p I is to be replaced with `unreachable` after information |
1490 | /// was manifested. |
1491 | void changeToUnreachableAfterManifest(Instruction *I) { |
1492 | ToBeChangedToUnreachableInsts.insert(I); |
1493 | } |
1494 | |
1495 | /// Record that \p II has at least one dead successor block. This information |
1496 | /// is used, e.g., to replace \p II with a call, after information was |
1497 | /// manifested. |
1498 | void registerInvokeWithDeadSuccessor(InvokeInst &II) { |
1499 | InvokeWithDeadSuccessor.push_back(&II); |
1500 | } |
1501 | |
1502 | /// Record that \p I is deleted after information was manifested. This also |
1503 | /// triggers deletion of trivially dead istructions. |
1504 | void deleteAfterManifest(Instruction &I) { ToBeDeletedInsts.insert(&I); } |
1505 | |
1506 | /// Record that \p BB is deleted after information was manifested. This also |
1507 | /// triggers deletion of trivially dead istructions. |
1508 | void deleteAfterManifest(BasicBlock &BB) { ToBeDeletedBlocks.insert(&BB); } |
1509 | |
1510 | // Record that \p BB is added during the manifest of an AA. Added basic blocks |
1511 | // are preserved in the IR. |
1512 | void registerManifestAddedBasicBlock(BasicBlock &BB) { |
1513 | ManifestAddedBlocks.insert(&BB); |
1514 | } |
1515 | |
1516 | /// Record that \p F is deleted after information was manifested. |
1517 | void deleteAfterManifest(Function &F) { |
1518 | if (DeleteFns) |
1519 | ToBeDeletedFunctions.insert(&F); |
1520 | } |
1521 | |
1522 | /// If \p IRP is assumed to be a constant, return it, if it is unclear yet, |
1523 | /// return None, otherwise return `nullptr`. |
1524 | Optional<Constant *> getAssumedConstant(const IRPosition &IRP, |
1525 | const AbstractAttribute &AA, |
1526 | bool &UsedAssumedInformation); |
1527 | Optional<Constant *> getAssumedConstant(const Value &V, |
1528 | const AbstractAttribute &AA, |
1529 | bool &UsedAssumedInformation) { |
1530 | return getAssumedConstant(IRPosition::value(V), AA, UsedAssumedInformation); |
1531 | } |
1532 | |
1533 | /// If \p V is assumed simplified, return it, if it is unclear yet, |
1534 | /// return None, otherwise return `nullptr`. |
1535 | Optional<Value *> getAssumedSimplified(const IRPosition &IRP, |
1536 | const AbstractAttribute &AA, |
1537 | bool &UsedAssumedInformation) { |
1538 | return getAssumedSimplified(IRP, &AA, UsedAssumedInformation); |
1539 | } |
1540 | Optional<Value *> getAssumedSimplified(const Value &V, |
1541 | const AbstractAttribute &AA, |
1542 | bool &UsedAssumedInformation) { |
1543 | return getAssumedSimplified(IRPosition::value(V), AA, |
1544 | UsedAssumedInformation); |
1545 | } |
1546 | |
1547 | /// If \p V is assumed simplified, return it, if it is unclear yet, |
1548 | /// return None, otherwise return `nullptr`. Same as the public version |
1549 | /// except that it can be used without recording dependences on any \p AA. |
1550 | Optional<Value *> getAssumedSimplified(const IRPosition &V, |
1551 | const AbstractAttribute *AA, |
1552 | bool &UsedAssumedInformation); |
1553 | |
1554 | /// Register \p CB as a simplification callback. |
1555 | /// `Attributor::getAssumedSimplified` will use these callbacks before |
1556 | /// we it will ask `AAValueSimplify`. It is important to ensure this |
1557 | /// is called before `identifyDefaultAbstractAttributes`, assuming the |
1558 | /// latter is called at all. |
1559 | using SimplifictionCallbackTy = std::function<Optional<Value *>( |
1560 | const IRPosition &, const AbstractAttribute *, bool &)>; |
1561 | void registerSimplificationCallback(const IRPosition &IRP, |
1562 | const SimplifictionCallbackTy &CB) { |
1563 | SimplificationCallbacks[IRP].emplace_back(CB); |
1564 | } |
1565 | |
1566 | /// Return true if there is a simplification callback for \p IRP. |
1567 | bool hasSimplificationCallback(const IRPosition &IRP) { |
1568 | return SimplificationCallbacks.count(IRP); |
1569 | } |
1570 | |
1571 | private: |
1572 | /// The vector with all simplification callbacks registered by outside AAs. |
1573 | DenseMap<IRPosition, SmallVector<SimplifictionCallbackTy, 1>> |
1574 | SimplificationCallbacks; |
1575 | |
1576 | public: |
1577 | /// Translate \p V from the callee context into the call site context. |
1578 | Optional<Value *> |
1579 | translateArgumentToCallSiteContent(Optional<Value *> V, CallBase &CB, |
1580 | const AbstractAttribute &AA, |
1581 | bool &UsedAssumedInformation); |