1 | //===-- SCCP.cpp ----------------------------------------------------------===// |
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 | // This file implements Interprocedural Sparse Conditional Constant Propagation. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "llvm/Transforms/IPO/SCCP.h" |
14 | #include "llvm/ADT/SetVector.h" |
15 | #include "llvm/Analysis/AssumptionCache.h" |
16 | #include "llvm/Analysis/BlockFrequencyInfo.h" |
17 | #include "llvm/Analysis/PostDominators.h" |
18 | #include "llvm/Analysis/TargetLibraryInfo.h" |
19 | #include "llvm/Analysis/TargetTransformInfo.h" |
20 | #include "llvm/Analysis/ValueLattice.h" |
21 | #include "llvm/Analysis/ValueLatticeUtils.h" |
22 | #include "llvm/Analysis/ValueTracking.h" |
23 | #include "llvm/IR/AttributeMask.h" |
24 | #include "llvm/IR/Constants.h" |
25 | #include "llvm/IR/DIBuilder.h" |
26 | #include "llvm/IR/IntrinsicInst.h" |
27 | #include "llvm/Support/CommandLine.h" |
28 | #include "llvm/Support/ModRef.h" |
29 | #include "llvm/Transforms/IPO.h" |
30 | #include "llvm/Transforms/IPO/FunctionSpecialization.h" |
31 | #include "llvm/Transforms/Scalar/SCCP.h" |
32 | #include "llvm/Transforms/Utils/Local.h" |
33 | #include "llvm/Transforms/Utils/SCCPSolver.h" |
34 | |
35 | using namespace llvm; |
36 | |
37 | #define DEBUG_TYPE "sccp" |
38 | |
39 | STATISTIC(NumInstRemoved, "Number of instructions removed" ); |
40 | STATISTIC(NumArgsElimed ,"Number of arguments constant propagated" ); |
41 | STATISTIC(NumGlobalConst, "Number of globals found to be constant" ); |
42 | STATISTIC(NumDeadBlocks , "Number of basic blocks unreachable" ); |
43 | STATISTIC(NumInstReplaced, |
44 | "Number of instructions replaced with (simpler) instruction" ); |
45 | |
46 | static cl::opt<unsigned> FuncSpecMaxIters( |
47 | "funcspec-max-iters" , cl::init(Val: 10), cl::Hidden, cl::desc( |
48 | "The maximum number of iterations function specialization is run" )); |
49 | |
50 | static void findReturnsToZap(Function &F, |
51 | SmallVector<ReturnInst *, 8> &ReturnsToZap, |
52 | SCCPSolver &Solver) { |
53 | // We can only do this if we know that nothing else can call the function. |
54 | if (!Solver.isArgumentTrackedFunction(F: &F)) |
55 | return; |
56 | |
57 | if (Solver.mustPreserveReturn(F: &F)) { |
58 | LLVM_DEBUG( |
59 | dbgs() |
60 | << "Can't zap returns of the function : " << F.getName() |
61 | << " due to present musttail or \"clang.arc.attachedcall\" call of " |
62 | "it\n" ); |
63 | return; |
64 | } |
65 | |
66 | assert( |
67 | all_of(F.users(), |
68 | [&Solver](User *U) { |
69 | if (isa<Instruction>(U) && |
70 | !Solver.isBlockExecutable(cast<Instruction>(U)->getParent())) |
71 | return true; |
72 | // Non-callsite uses are not impacted by zapping. Also, constant |
73 | // uses (like blockaddresses) could stuck around, without being |
74 | // used in the underlying IR, meaning we do not have lattice |
75 | // values for them. |
76 | if (!isa<CallBase>(U)) |
77 | return true; |
78 | if (U->getType()->isStructTy()) { |
79 | return none_of(Solver.getStructLatticeValueFor(U), |
80 | SCCPSolver::isOverdefined); |
81 | } |
82 | |
83 | // We don't consider assume-like intrinsics to be actual address |
84 | // captures. |
85 | if (auto *II = dyn_cast<IntrinsicInst>(U)) { |
86 | if (II->isAssumeLikeIntrinsic()) |
87 | return true; |
88 | } |
89 | |
90 | return !SCCPSolver::isOverdefined(Solver.getLatticeValueFor(U)); |
91 | }) && |
92 | "We can only zap functions where all live users have a concrete value" ); |
93 | |
94 | for (BasicBlock &BB : F) { |
95 | if (CallInst *CI = BB.getTerminatingMustTailCall()) { |
96 | LLVM_DEBUG(dbgs() << "Can't zap return of the block due to present " |
97 | << "musttail call : " << *CI << "\n" ); |
98 | (void)CI; |
99 | return; |
100 | } |
101 | |
102 | if (auto *RI = dyn_cast<ReturnInst>(Val: BB.getTerminator())) |
103 | if (!isa<UndefValue>(Val: RI->getOperand(i_nocapture: 0))) |
104 | ReturnsToZap.push_back(Elt: RI); |
105 | } |
106 | } |
107 | |
108 | static bool runIPSCCP( |
109 | Module &M, const DataLayout &DL, FunctionAnalysisManager *FAM, |
110 | std::function<const TargetLibraryInfo &(Function &)> GetTLI, |
111 | std::function<TargetTransformInfo &(Function &)> GetTTI, |
112 | std::function<AssumptionCache &(Function &)> GetAC, |
113 | std::function<DominatorTree &(Function &)> GetDT, |
114 | std::function<BlockFrequencyInfo &(Function &)> GetBFI, |
115 | bool IsFuncSpecEnabled) { |
116 | SCCPSolver Solver(DL, GetTLI, M.getContext()); |
117 | FunctionSpecializer Specializer(Solver, M, FAM, GetBFI, GetTLI, GetTTI, |
118 | GetAC); |
119 | |
120 | // Loop over all functions, marking arguments to those with their addresses |
121 | // taken or that are external as overdefined. |
122 | for (Function &F : M) { |
123 | if (F.isDeclaration()) |
124 | continue; |
125 | |
126 | DominatorTree &DT = GetDT(F); |
127 | AssumptionCache &AC = GetAC(F); |
128 | Solver.addPredicateInfo(F, DT, AC); |
129 | |
130 | // Determine if we can track the function's return values. If so, add the |
131 | // function to the solver's set of return-tracked functions. |
132 | if (canTrackReturnsInterprocedurally(F: &F)) |
133 | Solver.addTrackedFunction(F: &F); |
134 | |
135 | // Determine if we can track the function's arguments. If so, add the |
136 | // function to the solver's set of argument-tracked functions. |
137 | if (canTrackArgumentsInterprocedurally(F: &F)) { |
138 | Solver.addArgumentTrackedFunction(F: &F); |
139 | continue; |
140 | } |
141 | |
142 | // Assume the function is called. |
143 | Solver.markBlockExecutable(BB: &F.front()); |
144 | |
145 | for (Argument &AI : F.args()) |
146 | Solver.trackValueOfArgument(V: &AI); |
147 | } |
148 | |
149 | // Determine if we can track any of the module's global variables. If so, add |
150 | // the global variables we can track to the solver's set of tracked global |
151 | // variables. |
152 | for (GlobalVariable &G : M.globals()) { |
153 | G.removeDeadConstantUsers(); |
154 | if (canTrackGlobalVariableInterprocedurally(GV: &G)) |
155 | Solver.trackValueOfGlobalVariable(GV: &G); |
156 | } |
157 | |
158 | // Solve for constants. |
159 | Solver.solveWhileResolvedUndefsIn(M); |
160 | |
161 | if (IsFuncSpecEnabled) { |
162 | unsigned Iters = 0; |
163 | while (Iters++ < FuncSpecMaxIters && Specializer.run()); |
164 | } |
165 | |
166 | // Iterate over all of the instructions in the module, replacing them with |
167 | // constants if we have found them to be of constant values. |
168 | bool MadeChanges = false; |
169 | for (Function &F : M) { |
170 | if (F.isDeclaration()) |
171 | continue; |
172 | |
173 | SmallVector<BasicBlock *, 512> BlocksToErase; |
174 | |
175 | if (Solver.isBlockExecutable(BB: &F.front())) { |
176 | bool ReplacedPointerArg = false; |
177 | for (Argument &Arg : F.args()) { |
178 | if (!Arg.use_empty() && Solver.tryToReplaceWithConstant(V: &Arg)) { |
179 | ReplacedPointerArg |= Arg.getType()->isPointerTy(); |
180 | ++NumArgsElimed; |
181 | } |
182 | } |
183 | |
184 | // If we replaced an argument, we may now also access a global (currently |
185 | // classified as "other" memory). Update memory attribute to reflect this. |
186 | if (ReplacedPointerArg) { |
187 | auto UpdateAttrs = [&](AttributeList AL) { |
188 | MemoryEffects ME = AL.getMemoryEffects(); |
189 | if (ME == MemoryEffects::unknown()) |
190 | return AL; |
191 | |
192 | ModRefInfo ArgMemMR = ME.getModRef(Loc: IRMemLocation::ArgMem); |
193 | ME |= MemoryEffects(IRMemLocation::ErrnoMem, ArgMemMR); |
194 | ME |= MemoryEffects(IRMemLocation::Other, ArgMemMR); |
195 | |
196 | return AL.addFnAttribute( |
197 | C&: F.getContext(), |
198 | Attr: Attribute::getWithMemoryEffects(Context&: F.getContext(), ME)); |
199 | }; |
200 | |
201 | F.setAttributes(UpdateAttrs(F.getAttributes())); |
202 | for (User *U : F.users()) { |
203 | auto *CB = dyn_cast<CallBase>(Val: U); |
204 | if (!CB || CB->getCalledFunction() != &F) |
205 | continue; |
206 | |
207 | CB->setAttributes(UpdateAttrs(CB->getAttributes())); |
208 | } |
209 | } |
210 | MadeChanges |= ReplacedPointerArg; |
211 | } |
212 | |
213 | SmallPtrSet<Value *, 32> InsertedValues; |
214 | for (BasicBlock &BB : F) { |
215 | if (!Solver.isBlockExecutable(BB: &BB)) { |
216 | LLVM_DEBUG(dbgs() << " BasicBlock Dead:" << BB); |
217 | ++NumDeadBlocks; |
218 | |
219 | MadeChanges = true; |
220 | |
221 | if (&BB != &F.front()) |
222 | BlocksToErase.push_back(Elt: &BB); |
223 | continue; |
224 | } |
225 | |
226 | MadeChanges |= Solver.simplifyInstsInBlock( |
227 | BB, InsertedValues, InstRemovedStat&: NumInstRemoved, InstReplacedStat&: NumInstReplaced); |
228 | } |
229 | |
230 | DominatorTree *DT = FAM->getCachedResult<DominatorTreeAnalysis>(IR&: F); |
231 | PostDominatorTree *PDT = FAM->getCachedResult<PostDominatorTreeAnalysis>(IR&: F); |
232 | DomTreeUpdater DTU(DT, PDT, DomTreeUpdater::UpdateStrategy::Lazy); |
233 | // Change dead blocks to unreachable. We do it after replacing constants |
234 | // in all executable blocks, because changeToUnreachable may remove PHI |
235 | // nodes in executable blocks we found values for. The function's entry |
236 | // block is not part of BlocksToErase, so we have to handle it separately. |
237 | for (BasicBlock *BB : BlocksToErase) { |
238 | NumInstRemoved += changeToUnreachable(I: &*BB->getFirstNonPHIOrDbg(), |
239 | /*PreserveLCSSA=*/false, DTU: &DTU); |
240 | } |
241 | if (!Solver.isBlockExecutable(BB: &F.front())) |
242 | NumInstRemoved += changeToUnreachable(I: &*F.front().getFirstNonPHIOrDbg(), |
243 | /*PreserveLCSSA=*/false, DTU: &DTU); |
244 | |
245 | BasicBlock *NewUnreachableBB = nullptr; |
246 | for (BasicBlock &BB : F) |
247 | MadeChanges |= Solver.removeNonFeasibleEdges(BB: &BB, DTU, NewUnreachableBB); |
248 | |
249 | for (BasicBlock *DeadBB : BlocksToErase) |
250 | if (!DeadBB->hasAddressTaken()) |
251 | DTU.deleteBB(DelBB: DeadBB); |
252 | |
253 | Solver.removeSSACopies(F); |
254 | } |
255 | |
256 | // If we inferred constant or undef return values for a function, we replaced |
257 | // all call uses with the inferred value. This means we don't need to bother |
258 | // actually returning anything from the function. Replace all return |
259 | // instructions with return undef. |
260 | // |
261 | // Do this in two stages: first identify the functions we should process, then |
262 | // actually zap their returns. This is important because we can only do this |
263 | // if the address of the function isn't taken. In cases where a return is the |
264 | // last use of a function, the order of processing functions would affect |
265 | // whether other functions are optimizable. |
266 | SmallVector<ReturnInst*, 8> ReturnsToZap; |
267 | |
268 | Solver.inferReturnAttributes(); |
269 | Solver.inferArgAttributes(); |
270 | for (const auto &[F, ReturnValue] : Solver.getTrackedRetVals()) { |
271 | assert(!F->getReturnType()->isVoidTy() && |
272 | "should not track void functions" ); |
273 | if (SCCPSolver::isConstant(LV: ReturnValue) || ReturnValue.isUnknownOrUndef()) |
274 | findReturnsToZap(F&: *F, ReturnsToZap, Solver); |
275 | } |
276 | |
277 | for (auto *F : Solver.getMRVFunctionsTracked()) { |
278 | assert(F->getReturnType()->isStructTy() && |
279 | "The return type should be a struct" ); |
280 | StructType *STy = cast<StructType>(Val: F->getReturnType()); |
281 | if (Solver.isStructLatticeConstant(F, STy)) |
282 | findReturnsToZap(F&: *F, ReturnsToZap, Solver); |
283 | } |
284 | |
285 | // Zap all returns which we've identified as zap to change. |
286 | SmallSetVector<Function *, 8> FuncZappedReturn; |
287 | for (ReturnInst *RI : ReturnsToZap) { |
288 | Function *F = RI->getParent()->getParent(); |
289 | RI->setOperand(i_nocapture: 0, Val_nocapture: PoisonValue::get(T: F->getReturnType())); |
290 | // Record all functions that are zapped. |
291 | FuncZappedReturn.insert(X: F); |
292 | } |
293 | |
294 | // Remove the returned attribute for zapped functions and the |
295 | // corresponding call sites. |
296 | // Also remove any attributes that convert an undef return value into |
297 | // immediate undefined behavior |
298 | AttributeMask UBImplyingAttributes = |
299 | AttributeFuncs::getUBImplyingAttributes(); |
300 | for (Function *F : FuncZappedReturn) { |
301 | for (Argument &A : F->args()) |
302 | F->removeParamAttr(ArgNo: A.getArgNo(), Kind: Attribute::Returned); |
303 | F->removeRetAttrs(Attrs: UBImplyingAttributes); |
304 | for (Use &U : F->uses()) { |
305 | CallBase *CB = dyn_cast<CallBase>(Val: U.getUser()); |
306 | if (!CB) { |
307 | assert(isa<Constant>(U.getUser()) && |
308 | all_of(U.getUser()->users(), [](const User *UserUser) { |
309 | return cast<IntrinsicInst>(UserUser)->isAssumeLikeIntrinsic(); |
310 | })); |
311 | continue; |
312 | } |
313 | |
314 | for (Use &Arg : CB->args()) |
315 | CB->removeParamAttr(ArgNo: CB->getArgOperandNo(U: &Arg), Kind: Attribute::Returned); |
316 | CB->removeRetAttrs(AttrsToRemove: UBImplyingAttributes); |
317 | } |
318 | } |
319 | |
320 | // If we inferred constant or undef values for globals variables, we can |
321 | // delete the global and any stores that remain to it. |
322 | for (const auto &I : make_early_inc_range(Range: Solver.getTrackedGlobals())) { |
323 | GlobalVariable *GV = I.first; |
324 | if (SCCPSolver::isOverdefined(LV: I.second)) |
325 | continue; |
326 | LLVM_DEBUG(dbgs() << "Found that GV '" << GV->getName() |
327 | << "' is constant!\n" ); |
328 | for (User *U : make_early_inc_range(Range: GV->users())) { |
329 | // We can remove LoadInst here, because we already replaced its users |
330 | // with a constant. |
331 | assert((isa<StoreInst>(U) || isa<LoadInst>(U)) && |
332 | "Only Store|Load Instruction can be user of GlobalVariable at " |
333 | "reaching here." ); |
334 | cast<Instruction>(Val: U)->eraseFromParent(); |
335 | } |
336 | |
337 | // Try to create a debug constant expression for the global variable |
338 | // initializer value. |
339 | SmallVector<DIGlobalVariableExpression *, 1> GVEs; |
340 | GV->getDebugInfo(GVs&: GVEs); |
341 | if (GVEs.size() == 1) { |
342 | DIBuilder DIB(M); |
343 | if (DIExpression *InitExpr = getExpressionForConstant( |
344 | DIB, C: *GV->getInitializer(), Ty&: *GV->getValueType())) |
345 | GVEs[0]->replaceOperandWith(I: 1, New: InitExpr); |
346 | } |
347 | |
348 | MadeChanges = true; |
349 | M.eraseGlobalVariable(GV); |
350 | ++NumGlobalConst; |
351 | } |
352 | |
353 | return MadeChanges; |
354 | } |
355 | |
356 | PreservedAnalyses IPSCCPPass::run(Module &M, ModuleAnalysisManager &AM) { |
357 | const DataLayout &DL = M.getDataLayout(); |
358 | auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(IR&: M).getManager(); |
359 | auto GetTLI = [&FAM](Function &F) -> const TargetLibraryInfo & { |
360 | return FAM.getResult<TargetLibraryAnalysis>(IR&: F); |
361 | }; |
362 | auto GetTTI = [&FAM](Function &F) -> TargetTransformInfo & { |
363 | return FAM.getResult<TargetIRAnalysis>(IR&: F); |
364 | }; |
365 | auto GetAC = [&FAM](Function &F) -> AssumptionCache & { |
366 | return FAM.getResult<AssumptionAnalysis>(IR&: F); |
367 | }; |
368 | auto GetDT = [&FAM](Function &F) -> DominatorTree & { |
369 | return FAM.getResult<DominatorTreeAnalysis>(IR&: F); |
370 | }; |
371 | auto GetBFI = [&FAM](Function &F) -> BlockFrequencyInfo & { |
372 | return FAM.getResult<BlockFrequencyAnalysis>(IR&: F); |
373 | }; |
374 | |
375 | |
376 | if (!runIPSCCP(M, DL, FAM: &FAM, GetTLI, GetTTI, GetAC, GetDT, GetBFI, |
377 | IsFuncSpecEnabled: isFuncSpecEnabled())) |
378 | return PreservedAnalyses::all(); |
379 | |
380 | PreservedAnalyses PA; |
381 | PA.preserve<DominatorTreeAnalysis>(); |
382 | PA.preserve<PostDominatorTreeAnalysis>(); |
383 | PA.preserve<FunctionAnalysisManagerModuleProxy>(); |
384 | return PA; |
385 | } |
386 | |