xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/IPO/DeadArgumentElimination.cpp (revision 77013d11e6483b970af25e13c9b892075742f7e5)
1 //===- DeadArgumentElimination.cpp - Eliminate dead arguments -------------===//
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 pass deletes dead arguments from internal functions.  Dead argument
10 // elimination removes arguments which are directly dead, as well as arguments
11 // only passed into function calls as dead arguments of other functions.  This
12 // pass also deletes dead return values in a similar way.
13 //
14 // This pass is often useful as a cleanup pass to run after aggressive
15 // interprocedural passes, which add possibly-dead arguments or return values.
16 //
17 //===----------------------------------------------------------------------===//
18 
19 #include "llvm/Transforms/IPO/DeadArgumentElimination.h"
20 #include "llvm/ADT/SmallVector.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/IR/Argument.h"
23 #include "llvm/IR/Attributes.h"
24 #include "llvm/IR/BasicBlock.h"
25 #include "llvm/IR/Constants.h"
26 #include "llvm/IR/DerivedTypes.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/IRBuilder.h"
29 #include "llvm/IR/InstrTypes.h"
30 #include "llvm/IR/Instruction.h"
31 #include "llvm/IR/Instructions.h"
32 #include "llvm/IR/IntrinsicInst.h"
33 #include "llvm/IR/Intrinsics.h"
34 #include "llvm/IR/Module.h"
35 #include "llvm/IR/NoFolder.h"
36 #include "llvm/IR/PassManager.h"
37 #include "llvm/IR/Type.h"
38 #include "llvm/IR/Use.h"
39 #include "llvm/IR/User.h"
40 #include "llvm/IR/Value.h"
41 #include "llvm/InitializePasses.h"
42 #include "llvm/Pass.h"
43 #include "llvm/Support/Casting.h"
44 #include "llvm/Support/Debug.h"
45 #include "llvm/Support/raw_ostream.h"
46 #include "llvm/Transforms/IPO.h"
47 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
48 #include <cassert>
49 #include <cstdint>
50 #include <utility>
51 #include <vector>
52 
53 using namespace llvm;
54 
55 #define DEBUG_TYPE "deadargelim"
56 
57 STATISTIC(NumArgumentsEliminated, "Number of unread args removed");
58 STATISTIC(NumRetValsEliminated  , "Number of unused return values removed");
59 STATISTIC(NumArgumentsReplacedWithUndef,
60           "Number of unread args replaced with undef");
61 
62 namespace {
63 
64   /// DAE - The dead argument elimination pass.
65   class DAE : public ModulePass {
66   protected:
67     // DAH uses this to specify a different ID.
68     explicit DAE(char &ID) : ModulePass(ID) {}
69 
70   public:
71     static char ID; // Pass identification, replacement for typeid
72 
73     DAE() : ModulePass(ID) {
74       initializeDAEPass(*PassRegistry::getPassRegistry());
75     }
76 
77     bool runOnModule(Module &M) override {
78       if (skipModule(M))
79         return false;
80       DeadArgumentEliminationPass DAEP(ShouldHackArguments());
81       ModuleAnalysisManager DummyMAM;
82       PreservedAnalyses PA = DAEP.run(M, DummyMAM);
83       return !PA.areAllPreserved();
84     }
85 
86     virtual bool ShouldHackArguments() const { return false; }
87   };
88 
89 } // end anonymous namespace
90 
91 char DAE::ID = 0;
92 
93 INITIALIZE_PASS(DAE, "deadargelim", "Dead Argument Elimination", false, false)
94 
95 namespace {
96 
97   /// DAH - DeadArgumentHacking pass - Same as dead argument elimination, but
98   /// deletes arguments to functions which are external.  This is only for use
99   /// by bugpoint.
100   struct DAH : public DAE {
101     static char ID;
102 
103     DAH() : DAE(ID) {}
104 
105     bool ShouldHackArguments() const override { return true; }
106   };
107 
108 } // end anonymous namespace
109 
110 char DAH::ID = 0;
111 
112 INITIALIZE_PASS(DAH, "deadarghaX0r",
113                 "Dead Argument Hacking (BUGPOINT USE ONLY; DO NOT USE)",
114                 false, false)
115 
116 /// createDeadArgEliminationPass - This pass removes arguments from functions
117 /// which are not used by the body of the function.
118 ModulePass *llvm::createDeadArgEliminationPass() { return new DAE(); }
119 
120 ModulePass *llvm::createDeadArgHackingPass() { return new DAH(); }
121 
122 /// DeleteDeadVarargs - If this is an function that takes a ... list, and if
123 /// llvm.vastart is never called, the varargs list is dead for the function.
124 bool DeadArgumentEliminationPass::DeleteDeadVarargs(Function &Fn) {
125   assert(Fn.getFunctionType()->isVarArg() && "Function isn't varargs!");
126   if (Fn.isDeclaration() || !Fn.hasLocalLinkage()) return false;
127 
128   // Ensure that the function is only directly called.
129   if (Fn.hasAddressTaken())
130     return false;
131 
132   // Don't touch naked functions. The assembly might be using an argument, or
133   // otherwise rely on the frame layout in a way that this analysis will not
134   // see.
135   if (Fn.hasFnAttribute(Attribute::Naked)) {
136     return false;
137   }
138 
139   // Okay, we know we can transform this function if safe.  Scan its body
140   // looking for calls marked musttail or calls to llvm.vastart.
141   for (BasicBlock &BB : Fn) {
142     for (Instruction &I : BB) {
143       CallInst *CI = dyn_cast<CallInst>(&I);
144       if (!CI)
145         continue;
146       if (CI->isMustTailCall())
147         return false;
148       if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
149         if (II->getIntrinsicID() == Intrinsic::vastart)
150           return false;
151       }
152     }
153   }
154 
155   // If we get here, there are no calls to llvm.vastart in the function body,
156   // remove the "..." and adjust all the calls.
157 
158   // Start by computing a new prototype for the function, which is the same as
159   // the old function, but doesn't have isVarArg set.
160   FunctionType *FTy = Fn.getFunctionType();
161 
162   std::vector<Type *> Params(FTy->param_begin(), FTy->param_end());
163   FunctionType *NFTy = FunctionType::get(FTy->getReturnType(),
164                                                 Params, false);
165   unsigned NumArgs = Params.size();
166 
167   // Create the new function body and insert it into the module...
168   Function *NF = Function::Create(NFTy, Fn.getLinkage(), Fn.getAddressSpace());
169   NF->copyAttributesFrom(&Fn);
170   NF->setComdat(Fn.getComdat());
171   Fn.getParent()->getFunctionList().insert(Fn.getIterator(), NF);
172   NF->takeName(&Fn);
173 
174   // Loop over all of the callers of the function, transforming the call sites
175   // to pass in a smaller number of arguments into the new function.
176   //
177   std::vector<Value *> Args;
178   for (Value::user_iterator I = Fn.user_begin(), E = Fn.user_end(); I != E; ) {
179     CallBase *CB = dyn_cast<CallBase>(*I++);
180     if (!CB)
181       continue;
182 
183     // Pass all the same arguments.
184     Args.assign(CB->arg_begin(), CB->arg_begin() + NumArgs);
185 
186     // Drop any attributes that were on the vararg arguments.
187     AttributeList PAL = CB->getAttributes();
188     if (!PAL.isEmpty()) {
189       SmallVector<AttributeSet, 8> ArgAttrs;
190       for (unsigned ArgNo = 0; ArgNo < NumArgs; ++ArgNo)
191         ArgAttrs.push_back(PAL.getParamAttributes(ArgNo));
192       PAL = AttributeList::get(Fn.getContext(), PAL.getFnAttributes(),
193                                PAL.getRetAttributes(), ArgAttrs);
194     }
195 
196     SmallVector<OperandBundleDef, 1> OpBundles;
197     CB->getOperandBundlesAsDefs(OpBundles);
198 
199     CallBase *NewCB = nullptr;
200     if (InvokeInst *II = dyn_cast<InvokeInst>(CB)) {
201       NewCB = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(),
202                                  Args, OpBundles, "", CB);
203     } else {
204       NewCB = CallInst::Create(NF, Args, OpBundles, "", CB);
205       cast<CallInst>(NewCB)->setTailCallKind(
206           cast<CallInst>(CB)->getTailCallKind());
207     }
208     NewCB->setCallingConv(CB->getCallingConv());
209     NewCB->setAttributes(PAL);
210     NewCB->copyMetadata(*CB, {LLVMContext::MD_prof, LLVMContext::MD_dbg});
211 
212     Args.clear();
213 
214     if (!CB->use_empty())
215       CB->replaceAllUsesWith(NewCB);
216 
217     NewCB->takeName(CB);
218 
219     // Finally, remove the old call from the program, reducing the use-count of
220     // F.
221     CB->eraseFromParent();
222   }
223 
224   // Since we have now created the new function, splice the body of the old
225   // function right into the new function, leaving the old rotting hulk of the
226   // function empty.
227   NF->getBasicBlockList().splice(NF->begin(), Fn.getBasicBlockList());
228 
229   // Loop over the argument list, transferring uses of the old arguments over to
230   // the new arguments, also transferring over the names as well.  While we're at
231   // it, remove the dead arguments from the DeadArguments list.
232   for (Function::arg_iterator I = Fn.arg_begin(), E = Fn.arg_end(),
233        I2 = NF->arg_begin(); I != E; ++I, ++I2) {
234     // Move the name and users over to the new version.
235     I->replaceAllUsesWith(&*I2);
236     I2->takeName(&*I);
237   }
238 
239   // Clone metadatas from the old function, including debug info descriptor.
240   SmallVector<std::pair<unsigned, MDNode *>, 1> MDs;
241   Fn.getAllMetadata(MDs);
242   for (auto MD : MDs)
243     NF->addMetadata(MD.first, *MD.second);
244 
245   // Fix up any BlockAddresses that refer to the function.
246   Fn.replaceAllUsesWith(ConstantExpr::getBitCast(NF, Fn.getType()));
247   // Delete the bitcast that we just created, so that NF does not
248   // appear to be address-taken.
249   NF->removeDeadConstantUsers();
250   // Finally, nuke the old function.
251   Fn.eraseFromParent();
252   return true;
253 }
254 
255 /// RemoveDeadArgumentsFromCallers - Checks if the given function has any
256 /// arguments that are unused, and changes the caller parameters to be undefined
257 /// instead.
258 bool DeadArgumentEliminationPass::RemoveDeadArgumentsFromCallers(Function &Fn) {
259   // We cannot change the arguments if this TU does not define the function or
260   // if the linker may choose a function body from another TU, even if the
261   // nominal linkage indicates that other copies of the function have the same
262   // semantics. In the below example, the dead load from %p may not have been
263   // eliminated from the linker-chosen copy of f, so replacing %p with undef
264   // in callers may introduce undefined behavior.
265   //
266   // define linkonce_odr void @f(i32* %p) {
267   //   %v = load i32 %p
268   //   ret void
269   // }
270   if (!Fn.hasExactDefinition())
271     return false;
272 
273   // Functions with local linkage should already have been handled, except the
274   // fragile (variadic) ones which we can improve here.
275   if (Fn.hasLocalLinkage() && !Fn.getFunctionType()->isVarArg())
276     return false;
277 
278   // Don't touch naked functions. The assembly might be using an argument, or
279   // otherwise rely on the frame layout in a way that this analysis will not
280   // see.
281   if (Fn.hasFnAttribute(Attribute::Naked))
282     return false;
283 
284   if (Fn.use_empty())
285     return false;
286 
287   SmallVector<unsigned, 8> UnusedArgs;
288   bool Changed = false;
289 
290   for (Argument &Arg : Fn.args()) {
291     if (!Arg.hasSwiftErrorAttr() && Arg.use_empty() &&
292         !Arg.hasPassPointeeByValueCopyAttr()) {
293       if (Arg.isUsedByMetadata()) {
294         Arg.replaceAllUsesWith(UndefValue::get(Arg.getType()));
295         Changed = true;
296       }
297       UnusedArgs.push_back(Arg.getArgNo());
298     }
299   }
300 
301   if (UnusedArgs.empty())
302     return false;
303 
304   for (Use &U : Fn.uses()) {
305     CallBase *CB = dyn_cast<CallBase>(U.getUser());
306     if (!CB || !CB->isCallee(&U))
307       continue;
308 
309     // Now go through all unused args and replace them with "undef".
310     for (unsigned I = 0, E = UnusedArgs.size(); I != E; ++I) {
311       unsigned ArgNo = UnusedArgs[I];
312 
313       Value *Arg = CB->getArgOperand(ArgNo);
314       CB->setArgOperand(ArgNo, UndefValue::get(Arg->getType()));
315       ++NumArgumentsReplacedWithUndef;
316       Changed = true;
317     }
318   }
319 
320   return Changed;
321 }
322 
323 /// Convenience function that returns the number of return values. It returns 0
324 /// for void functions and 1 for functions not returning a struct. It returns
325 /// the number of struct elements for functions returning a struct.
326 static unsigned NumRetVals(const Function *F) {
327   Type *RetTy = F->getReturnType();
328   if (RetTy->isVoidTy())
329     return 0;
330   else if (StructType *STy = dyn_cast<StructType>(RetTy))
331     return STy->getNumElements();
332   else if (ArrayType *ATy = dyn_cast<ArrayType>(RetTy))
333     return ATy->getNumElements();
334   else
335     return 1;
336 }
337 
338 /// Returns the sub-type a function will return at a given Idx. Should
339 /// correspond to the result type of an ExtractValue instruction executed with
340 /// just that one Idx (i.e. only top-level structure is considered).
341 static Type *getRetComponentType(const Function *F, unsigned Idx) {
342   Type *RetTy = F->getReturnType();
343   assert(!RetTy->isVoidTy() && "void type has no subtype");
344 
345   if (StructType *STy = dyn_cast<StructType>(RetTy))
346     return STy->getElementType(Idx);
347   else if (ArrayType *ATy = dyn_cast<ArrayType>(RetTy))
348     return ATy->getElementType();
349   else
350     return RetTy;
351 }
352 
353 /// MarkIfNotLive - This checks Use for liveness in LiveValues. If Use is not
354 /// live, it adds Use to the MaybeLiveUses argument. Returns the determined
355 /// liveness of Use.
356 DeadArgumentEliminationPass::Liveness
357 DeadArgumentEliminationPass::MarkIfNotLive(RetOrArg Use,
358                                            UseVector &MaybeLiveUses) {
359   // We're live if our use or its Function is already marked as live.
360   if (IsLive(Use))
361     return Live;
362 
363   // We're maybe live otherwise, but remember that we must become live if
364   // Use becomes live.
365   MaybeLiveUses.push_back(Use);
366   return MaybeLive;
367 }
368 
369 /// SurveyUse - This looks at a single use of an argument or return value
370 /// and determines if it should be alive or not. Adds this use to MaybeLiveUses
371 /// if it causes the used value to become MaybeLive.
372 ///
373 /// RetValNum is the return value number to use when this use is used in a
374 /// return instruction. This is used in the recursion, you should always leave
375 /// it at 0.
376 DeadArgumentEliminationPass::Liveness
377 DeadArgumentEliminationPass::SurveyUse(const Use *U, UseVector &MaybeLiveUses,
378                                        unsigned RetValNum) {
379     const User *V = U->getUser();
380     if (const ReturnInst *RI = dyn_cast<ReturnInst>(V)) {
381       // The value is returned from a function. It's only live when the
382       // function's return value is live. We use RetValNum here, for the case
383       // that U is really a use of an insertvalue instruction that uses the
384       // original Use.
385       const Function *F = RI->getParent()->getParent();
386       if (RetValNum != -1U) {
387         RetOrArg Use = CreateRet(F, RetValNum);
388         // We might be live, depending on the liveness of Use.
389         return MarkIfNotLive(Use, MaybeLiveUses);
390       } else {
391         DeadArgumentEliminationPass::Liveness Result = MaybeLive;
392         for (unsigned Ri = 0; Ri < NumRetVals(F); ++Ri) {
393           RetOrArg Use = CreateRet(F, Ri);
394           // We might be live, depending on the liveness of Use. If any
395           // sub-value is live, then the entire value is considered live. This
396           // is a conservative choice, and better tracking is possible.
397           DeadArgumentEliminationPass::Liveness SubResult =
398               MarkIfNotLive(Use, MaybeLiveUses);
399           if (Result != Live)
400             Result = SubResult;
401         }
402         return Result;
403       }
404     }
405     if (const InsertValueInst *IV = dyn_cast<InsertValueInst>(V)) {
406       if (U->getOperandNo() != InsertValueInst::getAggregateOperandIndex()
407           && IV->hasIndices())
408         // The use we are examining is inserted into an aggregate. Our liveness
409         // depends on all uses of that aggregate, but if it is used as a return
410         // value, only index at which we were inserted counts.
411         RetValNum = *IV->idx_begin();
412 
413       // Note that if we are used as the aggregate operand to the insertvalue,
414       // we don't change RetValNum, but do survey all our uses.
415 
416       Liveness Result = MaybeLive;
417       for (const Use &UU : IV->uses()) {
418         Result = SurveyUse(&UU, MaybeLiveUses, RetValNum);
419         if (Result == Live)
420           break;
421       }
422       return Result;
423     }
424 
425     if (const auto *CB = dyn_cast<CallBase>(V)) {
426       const Function *F = CB->getCalledFunction();
427       if (F) {
428         // Used in a direct call.
429 
430         // The function argument is live if it is used as a bundle operand.
431         if (CB->isBundleOperand(U))
432           return Live;
433 
434         // Find the argument number. We know for sure that this use is an
435         // argument, since if it was the function argument this would be an
436         // indirect call and the we know can't be looking at a value of the
437         // label type (for the invoke instruction).
438         unsigned ArgNo = CB->getArgOperandNo(U);
439 
440         if (ArgNo >= F->getFunctionType()->getNumParams())
441           // The value is passed in through a vararg! Must be live.
442           return Live;
443 
444         assert(CB->getArgOperand(ArgNo) == CB->getOperand(U->getOperandNo()) &&
445                "Argument is not where we expected it");
446 
447         // Value passed to a normal call. It's only live when the corresponding
448         // argument to the called function turns out live.
449         RetOrArg Use = CreateArg(F, ArgNo);
450         return MarkIfNotLive(Use, MaybeLiveUses);
451       }
452     }
453     // Used in any other way? Value must be live.
454     return Live;
455 }
456 
457 /// SurveyUses - This looks at all the uses of the given value
458 /// Returns the Liveness deduced from the uses of this value.
459 ///
460 /// Adds all uses that cause the result to be MaybeLive to MaybeLiveRetUses. If
461 /// the result is Live, MaybeLiveUses might be modified but its content should
462 /// be ignored (since it might not be complete).
463 DeadArgumentEliminationPass::Liveness
464 DeadArgumentEliminationPass::SurveyUses(const Value *V,
465                                         UseVector &MaybeLiveUses) {
466   // Assume it's dead (which will only hold if there are no uses at all..).
467   Liveness Result = MaybeLive;
468   // Check each use.
469   for (const Use &U : V->uses()) {
470     Result = SurveyUse(&U, MaybeLiveUses);
471     if (Result == Live)
472       break;
473   }
474   return Result;
475 }
476 
477 // SurveyFunction - This performs the initial survey of the specified function,
478 // checking out whether or not it uses any of its incoming arguments or whether
479 // any callers use the return value.  This fills in the LiveValues set and Uses
480 // map.
481 //
482 // We consider arguments of non-internal functions to be intrinsically alive as
483 // well as arguments to functions which have their "address taken".
484 void DeadArgumentEliminationPass::SurveyFunction(const Function &F) {
485   // Functions with inalloca/preallocated parameters are expecting args in a
486   // particular register and memory layout.
487   if (F.getAttributes().hasAttrSomewhere(Attribute::InAlloca) ||
488       F.getAttributes().hasAttrSomewhere(Attribute::Preallocated)) {
489     MarkLive(F);
490     return;
491   }
492 
493   // Don't touch naked functions. The assembly might be using an argument, or
494   // otherwise rely on the frame layout in a way that this analysis will not
495   // see.
496   if (F.hasFnAttribute(Attribute::Naked)) {
497     MarkLive(F);
498     return;
499   }
500 
501   unsigned RetCount = NumRetVals(&F);
502 
503   // Assume all return values are dead
504   using RetVals = SmallVector<Liveness, 5>;
505 
506   RetVals RetValLiveness(RetCount, MaybeLive);
507 
508   using RetUses = SmallVector<UseVector, 5>;
509 
510   // These vectors map each return value to the uses that make it MaybeLive, so
511   // we can add those to the Uses map if the return value really turns out to be
512   // MaybeLive. Initialized to a list of RetCount empty lists.
513   RetUses MaybeLiveRetUses(RetCount);
514 
515   bool HasMustTailCalls = false;
516 
517   for (Function::const_iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
518     if (const ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
519       if (RI->getNumOperands() != 0 && RI->getOperand(0)->getType()
520           != F.getFunctionType()->getReturnType()) {
521         // We don't support old style multiple return values.
522         MarkLive(F);
523         return;
524       }
525     }
526 
527     // If we have any returns of `musttail` results - the signature can't
528     // change
529     if (BB->getTerminatingMustTailCall() != nullptr)
530       HasMustTailCalls = true;
531   }
532 
533   if (HasMustTailCalls) {
534     LLVM_DEBUG(dbgs() << "DeadArgumentEliminationPass - " << F.getName()
535                       << " has musttail calls\n");
536   }
537 
538   if (!F.hasLocalLinkage() && (!ShouldHackArguments || F.isIntrinsic())) {
539     MarkLive(F);
540     return;
541   }
542 
543   LLVM_DEBUG(
544       dbgs() << "DeadArgumentEliminationPass - Inspecting callers for fn: "
545              << F.getName() << "\n");
546   // Keep track of the number of live retvals, so we can skip checks once all
547   // of them turn out to be live.
548   unsigned NumLiveRetVals = 0;
549 
550   bool HasMustTailCallers = false;
551 
552   // Loop all uses of the function.
553   for (const Use &U : F.uses()) {
554     // If the function is PASSED IN as an argument, its address has been
555     // taken.
556     const auto *CB = dyn_cast<CallBase>(U.getUser());
557     if (!CB || !CB->isCallee(&U)) {
558       MarkLive(F);
559       return;
560     }
561 
562     // The number of arguments for `musttail` call must match the number of
563     // arguments of the caller
564     if (CB->isMustTailCall())
565       HasMustTailCallers = true;
566 
567     // If we end up here, we are looking at a direct call to our function.
568 
569     // Now, check how our return value(s) is/are used in this caller. Don't
570     // bother checking return values if all of them are live already.
571     if (NumLiveRetVals == RetCount)
572       continue;
573 
574     // Check all uses of the return value.
575     for (const Use &U : CB->uses()) {
576       if (ExtractValueInst *Ext = dyn_cast<ExtractValueInst>(U.getUser())) {
577         // This use uses a part of our return value, survey the uses of
578         // that part and store the results for this index only.
579         unsigned Idx = *Ext->idx_begin();
580         if (RetValLiveness[Idx] != Live) {
581           RetValLiveness[Idx] = SurveyUses(Ext, MaybeLiveRetUses[Idx]);
582           if (RetValLiveness[Idx] == Live)
583             NumLiveRetVals++;
584         }
585       } else {
586         // Used by something else than extractvalue. Survey, but assume that the
587         // result applies to all sub-values.
588         UseVector MaybeLiveAggregateUses;
589         if (SurveyUse(&U, MaybeLiveAggregateUses) == Live) {
590           NumLiveRetVals = RetCount;
591           RetValLiveness.assign(RetCount, Live);
592           break;
593         } else {
594           for (unsigned Ri = 0; Ri != RetCount; ++Ri) {
595             if (RetValLiveness[Ri] != Live)
596               MaybeLiveRetUses[Ri].append(MaybeLiveAggregateUses.begin(),
597                                           MaybeLiveAggregateUses.end());
598           }
599         }
600       }
601     }
602   }
603 
604   if (HasMustTailCallers) {
605     LLVM_DEBUG(dbgs() << "DeadArgumentEliminationPass - " << F.getName()
606                       << " has musttail callers\n");
607   }
608 
609   // Now we've inspected all callers, record the liveness of our return values.
610   for (unsigned Ri = 0; Ri != RetCount; ++Ri)
611     MarkValue(CreateRet(&F, Ri), RetValLiveness[Ri], MaybeLiveRetUses[Ri]);
612 
613   LLVM_DEBUG(dbgs() << "DeadArgumentEliminationPass - Inspecting args for fn: "
614                     << F.getName() << "\n");
615 
616   // Now, check all of our arguments.
617   unsigned ArgI = 0;
618   UseVector MaybeLiveArgUses;
619   for (Function::const_arg_iterator AI = F.arg_begin(), E = F.arg_end();
620        AI != E; ++AI, ++ArgI) {
621     Liveness Result;
622     if (F.getFunctionType()->isVarArg() || HasMustTailCallers ||
623         HasMustTailCalls) {
624       // Variadic functions will already have a va_arg function expanded inside
625       // them, making them potentially very sensitive to ABI changes resulting
626       // from removing arguments entirely, so don't. For example AArch64 handles
627       // register and stack HFAs very differently, and this is reflected in the
628       // IR which has already been generated.
629       //
630       // `musttail` calls to this function restrict argument removal attempts.
631       // The signature of the caller must match the signature of the function.
632       //
633       // `musttail` calls in this function prevents us from changing its
634       // signature
635       Result = Live;
636     } else {
637       // See what the effect of this use is (recording any uses that cause
638       // MaybeLive in MaybeLiveArgUses).
639       Result = SurveyUses(&*AI, MaybeLiveArgUses);
640     }
641 
642     // Mark the result.
643     MarkValue(CreateArg(&F, ArgI), Result, MaybeLiveArgUses);
644     // Clear the vector again for the next iteration.
645     MaybeLiveArgUses.clear();
646   }
647 }
648 
649 /// MarkValue - This function marks the liveness of RA depending on L. If L is
650 /// MaybeLive, it also takes all uses in MaybeLiveUses and records them in Uses,
651 /// such that RA will be marked live if any use in MaybeLiveUses gets marked
652 /// live later on.
653 void DeadArgumentEliminationPass::MarkValue(const RetOrArg &RA, Liveness L,
654                                             const UseVector &MaybeLiveUses) {
655   switch (L) {
656     case Live:
657       MarkLive(RA);
658       break;
659     case MaybeLive:
660       assert(!IsLive(RA) && "Use is already live!");
661       for (const auto &MaybeLiveUse : MaybeLiveUses) {
662         if (IsLive(MaybeLiveUse)) {
663           // A use is live, so this value is live.
664           MarkLive(RA);
665           break;
666         } else {
667           // Note any uses of this value, so this value can be
668           // marked live whenever one of the uses becomes live.
669           Uses.insert(std::make_pair(MaybeLiveUse, RA));
670         }
671       }
672       break;
673   }
674 }
675 
676 /// MarkLive - Mark the given Function as alive, meaning that it cannot be
677 /// changed in any way. Additionally,
678 /// mark any values that are used as this function's parameters or by its return
679 /// values (according to Uses) live as well.
680 void DeadArgumentEliminationPass::MarkLive(const Function &F) {
681   LLVM_DEBUG(dbgs() << "DeadArgumentEliminationPass - Intrinsically live fn: "
682                     << F.getName() << "\n");
683   // Mark the function as live.
684   LiveFunctions.insert(&F);
685   // Mark all arguments as live.
686   for (unsigned ArgI = 0, E = F.arg_size(); ArgI != E; ++ArgI)
687     PropagateLiveness(CreateArg(&F, ArgI));
688   // Mark all return values as live.
689   for (unsigned Ri = 0, E = NumRetVals(&F); Ri != E; ++Ri)
690     PropagateLiveness(CreateRet(&F, Ri));
691 }
692 
693 /// MarkLive - Mark the given return value or argument as live. Additionally,
694 /// mark any values that are used by this value (according to Uses) live as
695 /// well.
696 void DeadArgumentEliminationPass::MarkLive(const RetOrArg &RA) {
697   if (IsLive(RA))
698     return; // Already marked Live.
699 
700   LiveValues.insert(RA);
701 
702   LLVM_DEBUG(dbgs() << "DeadArgumentEliminationPass - Marking "
703                     << RA.getDescription() << " live\n");
704   PropagateLiveness(RA);
705 }
706 
707 bool DeadArgumentEliminationPass::IsLive(const RetOrArg &RA) {
708   return LiveFunctions.count(RA.F) || LiveValues.count(RA);
709 }
710 
711 /// PropagateLiveness - Given that RA is a live value, propagate it's liveness
712 /// to any other values it uses (according to Uses).
713 void DeadArgumentEliminationPass::PropagateLiveness(const RetOrArg &RA) {
714   // We don't use upper_bound (or equal_range) here, because our recursive call
715   // to ourselves is likely to cause the upper_bound (which is the first value
716   // not belonging to RA) to become erased and the iterator invalidated.
717   UseMap::iterator Begin = Uses.lower_bound(RA);
718   UseMap::iterator E = Uses.end();
719   UseMap::iterator I;
720   for (I = Begin; I != E && I->first == RA; ++I)
721     MarkLive(I->second);
722 
723   // Erase RA from the Uses map (from the lower bound to wherever we ended up
724   // after the loop).
725   Uses.erase(Begin, I);
726 }
727 
728 // RemoveDeadStuffFromFunction - Remove any arguments and return values from F
729 // that are not in LiveValues. Transform the function and all of the callees of
730 // the function to not have these arguments and return values.
731 //
732 bool DeadArgumentEliminationPass::RemoveDeadStuffFromFunction(Function *F) {
733   // Don't modify fully live functions
734   if (LiveFunctions.count(F))
735     return false;
736 
737   // Start by computing a new prototype for the function, which is the same as
738   // the old function, but has fewer arguments and a different return type.
739   FunctionType *FTy = F->getFunctionType();
740   std::vector<Type*> Params;
741 
742   // Keep track of if we have a live 'returned' argument
743   bool HasLiveReturnedArg = false;
744 
745   // Set up to build a new list of parameter attributes.
746   SmallVector<AttributeSet, 8> ArgAttrVec;
747   const AttributeList &PAL = F->getAttributes();
748 
749   // Remember which arguments are still alive.
750   SmallVector<bool, 10> ArgAlive(FTy->getNumParams(), false);
751   // Construct the new parameter list from non-dead arguments. Also construct
752   // a new set of parameter attributes to correspond. Skip the first parameter
753   // attribute, since that belongs to the return value.
754   unsigned ArgI = 0;
755   for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E;
756        ++I, ++ArgI) {
757     RetOrArg Arg = CreateArg(F, ArgI);
758     if (LiveValues.erase(Arg)) {
759       Params.push_back(I->getType());
760       ArgAlive[ArgI] = true;
761       ArgAttrVec.push_back(PAL.getParamAttributes(ArgI));
762       HasLiveReturnedArg |= PAL.hasParamAttribute(ArgI, Attribute::Returned);
763     } else {
764       ++NumArgumentsEliminated;
765       LLVM_DEBUG(dbgs() << "DeadArgumentEliminationPass - Removing argument "
766                         << ArgI << " (" << I->getName() << ") from "
767                         << F->getName() << "\n");
768     }
769   }
770 
771   // Find out the new return value.
772   Type *RetTy = FTy->getReturnType();
773   Type *NRetTy = nullptr;
774   unsigned RetCount = NumRetVals(F);
775 
776   // -1 means unused, other numbers are the new index
777   SmallVector<int, 5> NewRetIdxs(RetCount, -1);
778   std::vector<Type*> RetTypes;
779 
780   // If there is a function with a live 'returned' argument but a dead return
781   // value, then there are two possible actions:
782   // 1) Eliminate the return value and take off the 'returned' attribute on the
783   //    argument.
784   // 2) Retain the 'returned' attribute and treat the return value (but not the
785   //    entire function) as live so that it is not eliminated.
786   //
787   // It's not clear in the general case which option is more profitable because,
788   // even in the absence of explicit uses of the return value, code generation
789   // is free to use the 'returned' attribute to do things like eliding
790   // save/restores of registers across calls. Whether or not this happens is
791   // target and ABI-specific as well as depending on the amount of register
792   // pressure, so there's no good way for an IR-level pass to figure this out.
793   //
794   // Fortunately, the only places where 'returned' is currently generated by
795   // the FE are places where 'returned' is basically free and almost always a
796   // performance win, so the second option can just be used always for now.
797   //
798   // This should be revisited if 'returned' is ever applied more liberally.
799   if (RetTy->isVoidTy() || HasLiveReturnedArg) {
800     NRetTy = RetTy;
801   } else {
802     // Look at each of the original return values individually.
803     for (unsigned Ri = 0; Ri != RetCount; ++Ri) {
804       RetOrArg Ret = CreateRet(F, Ri);
805       if (LiveValues.erase(Ret)) {
806         RetTypes.push_back(getRetComponentType(F, Ri));
807         NewRetIdxs[Ri] = RetTypes.size() - 1;
808       } else {
809         ++NumRetValsEliminated;
810         LLVM_DEBUG(
811             dbgs() << "DeadArgumentEliminationPass - Removing return value "
812                    << Ri << " from " << F->getName() << "\n");
813       }
814     }
815     if (RetTypes.size() > 1) {
816       // More than one return type? Reduce it down to size.
817       if (StructType *STy = dyn_cast<StructType>(RetTy)) {
818         // Make the new struct packed if we used to return a packed struct
819         // already.
820         NRetTy = StructType::get(STy->getContext(), RetTypes, STy->isPacked());
821       } else {
822         assert(isa<ArrayType>(RetTy) && "unexpected multi-value return");
823         NRetTy = ArrayType::get(RetTypes[0], RetTypes.size());
824       }
825     } else if (RetTypes.size() == 1)
826       // One return type? Just a simple value then, but only if we didn't use to
827       // return a struct with that simple value before.
828       NRetTy = RetTypes.front();
829     else if (RetTypes.empty())
830       // No return types? Make it void, but only if we didn't use to return {}.
831       NRetTy = Type::getVoidTy(F->getContext());
832   }
833 
834   assert(NRetTy && "No new return type found?");
835 
836   // The existing function return attributes.
837   AttrBuilder RAttrs(PAL.getRetAttributes());
838 
839   // Remove any incompatible attributes, but only if we removed all return
840   // values. Otherwise, ensure that we don't have any conflicting attributes
841   // here. Currently, this should not be possible, but special handling might be
842   // required when new return value attributes are added.
843   if (NRetTy->isVoidTy())
844     RAttrs.remove(AttributeFuncs::typeIncompatible(NRetTy));
845   else
846     assert(!RAttrs.overlaps(AttributeFuncs::typeIncompatible(NRetTy)) &&
847            "Return attributes no longer compatible?");
848 
849   AttributeSet RetAttrs = AttributeSet::get(F->getContext(), RAttrs);
850 
851   // Strip allocsize attributes. They might refer to the deleted arguments.
852   AttributeSet FnAttrs = PAL.getFnAttributes().removeAttribute(
853       F->getContext(), Attribute::AllocSize);
854 
855   // Reconstruct the AttributesList based on the vector we constructed.
856   assert(ArgAttrVec.size() == Params.size());
857   AttributeList NewPAL =
858       AttributeList::get(F->getContext(), FnAttrs, RetAttrs, ArgAttrVec);
859 
860   // Create the new function type based on the recomputed parameters.
861   FunctionType *NFTy = FunctionType::get(NRetTy, Params, FTy->isVarArg());
862 
863   // No change?
864   if (NFTy == FTy)
865     return false;
866 
867   // Create the new function body and insert it into the module...
868   Function *NF = Function::Create(NFTy, F->getLinkage(), F->getAddressSpace());
869   NF->copyAttributesFrom(F);
870   NF->setComdat(F->getComdat());
871   NF->setAttributes(NewPAL);
872   // Insert the new function before the old function, so we won't be processing
873   // it again.
874   F->getParent()->getFunctionList().insert(F->getIterator(), NF);
875   NF->takeName(F);
876 
877   // Loop over all of the callers of the function, transforming the call sites
878   // to pass in a smaller number of arguments into the new function.
879   std::vector<Value*> Args;
880   while (!F->use_empty()) {
881     CallBase &CB = cast<CallBase>(*F->user_back());
882 
883     ArgAttrVec.clear();
884     const AttributeList &CallPAL = CB.getAttributes();
885 
886     // Adjust the call return attributes in case the function was changed to
887     // return void.
888     AttrBuilder RAttrs(CallPAL.getRetAttributes());
889     RAttrs.remove(AttributeFuncs::typeIncompatible(NRetTy));
890     AttributeSet RetAttrs = AttributeSet::get(F->getContext(), RAttrs);
891 
892     // Declare these outside of the loops, so we can reuse them for the second
893     // loop, which loops the varargs.
894     auto I = CB.arg_begin();
895     unsigned Pi = 0;
896     // Loop over those operands, corresponding to the normal arguments to the
897     // original function, and add those that are still alive.
898     for (unsigned E = FTy->getNumParams(); Pi != E; ++I, ++Pi)
899       if (ArgAlive[Pi]) {
900         Args.push_back(*I);
901         // Get original parameter attributes, but skip return attributes.
902         AttributeSet Attrs = CallPAL.getParamAttributes(Pi);
903         if (NRetTy != RetTy && Attrs.hasAttribute(Attribute::Returned)) {
904           // If the return type has changed, then get rid of 'returned' on the
905           // call site. The alternative is to make all 'returned' attributes on
906           // call sites keep the return value alive just like 'returned'
907           // attributes on function declaration but it's less clearly a win and
908           // this is not an expected case anyway
909           ArgAttrVec.push_back(AttributeSet::get(
910               F->getContext(),
911               AttrBuilder(Attrs).removeAttribute(Attribute::Returned)));
912         } else {
913           // Otherwise, use the original attributes.
914           ArgAttrVec.push_back(Attrs);
915         }
916       }
917 
918     // Push any varargs arguments on the list. Don't forget their attributes.
919     for (auto E = CB.arg_end(); I != E; ++I, ++Pi) {
920       Args.push_back(*I);
921       ArgAttrVec.push_back(CallPAL.getParamAttributes(Pi));
922     }
923 
924     // Reconstruct the AttributesList based on the vector we constructed.
925     assert(ArgAttrVec.size() == Args.size());
926 
927     // Again, be sure to remove any allocsize attributes, since their indices
928     // may now be incorrect.
929     AttributeSet FnAttrs = CallPAL.getFnAttributes().removeAttribute(
930         F->getContext(), Attribute::AllocSize);
931 
932     AttributeList NewCallPAL = AttributeList::get(
933         F->getContext(), FnAttrs, RetAttrs, ArgAttrVec);
934 
935     SmallVector<OperandBundleDef, 1> OpBundles;
936     CB.getOperandBundlesAsDefs(OpBundles);
937 
938     CallBase *NewCB = nullptr;
939     if (InvokeInst *II = dyn_cast<InvokeInst>(&CB)) {
940       NewCB = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(),
941                                  Args, OpBundles, "", CB.getParent());
942     } else {
943       NewCB = CallInst::Create(NFTy, NF, Args, OpBundles, "", &CB);
944       cast<CallInst>(NewCB)->setTailCallKind(
945           cast<CallInst>(&CB)->getTailCallKind());
946     }
947     NewCB->setCallingConv(CB.getCallingConv());
948     NewCB->setAttributes(NewCallPAL);
949     NewCB->copyMetadata(CB, {LLVMContext::MD_prof, LLVMContext::MD_dbg});
950     Args.clear();
951     ArgAttrVec.clear();
952 
953     if (!CB.use_empty() || CB.isUsedByMetadata()) {
954       if (NewCB->getType() == CB.getType()) {
955         // Return type not changed? Just replace users then.
956         CB.replaceAllUsesWith(NewCB);
957         NewCB->takeName(&CB);
958       } else if (NewCB->getType()->isVoidTy()) {
959         // If the return value is dead, replace any uses of it with undef
960         // (any non-debug value uses will get removed later on).
961         if (!CB.getType()->isX86_MMXTy())
962           CB.replaceAllUsesWith(UndefValue::get(CB.getType()));
963       } else {
964         assert((RetTy->isStructTy() || RetTy->isArrayTy()) &&
965                "Return type changed, but not into a void. The old return type"
966                " must have been a struct or an array!");
967         Instruction *InsertPt = &CB;
968         if (InvokeInst *II = dyn_cast<InvokeInst>(&CB)) {
969           BasicBlock *NewEdge =
970               SplitEdge(NewCB->getParent(), II->getNormalDest());
971           InsertPt = &*NewEdge->getFirstInsertionPt();
972         }
973 
974         // We used to return a struct or array. Instead of doing smart stuff
975         // with all the uses, we will just rebuild it using extract/insertvalue
976         // chaining and let instcombine clean that up.
977         //
978         // Start out building up our return value from undef
979         Value *RetVal = UndefValue::get(RetTy);
980         for (unsigned Ri = 0; Ri != RetCount; ++Ri)
981           if (NewRetIdxs[Ri] != -1) {
982             Value *V;
983             IRBuilder<NoFolder> IRB(InsertPt);
984             if (RetTypes.size() > 1)
985               // We are still returning a struct, so extract the value from our
986               // return value
987               V = IRB.CreateExtractValue(NewCB, NewRetIdxs[Ri], "newret");
988             else
989               // We are now returning a single element, so just insert that
990               V = NewCB;
991             // Insert the value at the old position
992             RetVal = IRB.CreateInsertValue(RetVal, V, Ri, "oldret");
993           }
994         // Now, replace all uses of the old call instruction with the return
995         // struct we built
996         CB.replaceAllUsesWith(RetVal);
997         NewCB->takeName(&CB);
998       }
999     }
1000 
1001     // Finally, remove the old call from the program, reducing the use-count of
1002     // F.
1003     CB.eraseFromParent();
1004   }
1005 
1006   // Since we have now created the new function, splice the body of the old
1007   // function right into the new function, leaving the old rotting hulk of the
1008   // function empty.
1009   NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList());
1010 
1011   // Loop over the argument list, transferring uses of the old arguments over to
1012   // the new arguments, also transferring over the names as well.
1013   ArgI = 0;
1014   for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(),
1015                               I2 = NF->arg_begin();
1016        I != E; ++I, ++ArgI)
1017     if (ArgAlive[ArgI]) {
1018       // If this is a live argument, move the name and users over to the new
1019       // version.
1020       I->replaceAllUsesWith(&*I2);
1021       I2->takeName(&*I);
1022       ++I2;
1023     } else {
1024       // If this argument is dead, replace any uses of it with undef
1025       // (any non-debug value uses will get removed later on).
1026       if (!I->getType()->isX86_MMXTy())
1027         I->replaceAllUsesWith(UndefValue::get(I->getType()));
1028     }
1029 
1030   // If we change the return value of the function we must rewrite any return
1031   // instructions.  Check this now.
1032   if (F->getReturnType() != NF->getReturnType())
1033     for (BasicBlock &BB : *NF)
1034       if (ReturnInst *RI = dyn_cast<ReturnInst>(BB.getTerminator())) {
1035         IRBuilder<NoFolder> IRB(RI);
1036         Value *RetVal = nullptr;
1037 
1038         if (!NFTy->getReturnType()->isVoidTy()) {
1039           assert(RetTy->isStructTy() || RetTy->isArrayTy());
1040           // The original return value was a struct or array, insert
1041           // extractvalue/insertvalue chains to extract only the values we need
1042           // to return and insert them into our new result.
1043           // This does generate messy code, but we'll let it to instcombine to
1044           // clean that up.
1045           Value *OldRet = RI->getOperand(0);
1046           // Start out building up our return value from undef
1047           RetVal = UndefValue::get(NRetTy);
1048           for (unsigned RetI = 0; RetI != RetCount; ++RetI)
1049             if (NewRetIdxs[RetI] != -1) {
1050               Value *EV = IRB.CreateExtractValue(OldRet, RetI, "oldret");
1051 
1052               if (RetTypes.size() > 1) {
1053                 // We're still returning a struct, so reinsert the value into
1054                 // our new return value at the new index
1055 
1056                 RetVal = IRB.CreateInsertValue(RetVal, EV, NewRetIdxs[RetI],
1057                                                "newret");
1058               } else {
1059                 // We are now only returning a simple value, so just return the
1060                 // extracted value.
1061                 RetVal = EV;
1062               }
1063             }
1064         }
1065         // Replace the return instruction with one returning the new return
1066         // value (possibly 0 if we became void).
1067         auto *NewRet = ReturnInst::Create(F->getContext(), RetVal, RI);
1068         NewRet->setDebugLoc(RI->getDebugLoc());
1069         BB.getInstList().erase(RI);
1070       }
1071 
1072   // Clone metadatas from the old function, including debug info descriptor.
1073   SmallVector<std::pair<unsigned, MDNode *>, 1> MDs;
1074   F->getAllMetadata(MDs);
1075   for (auto MD : MDs)
1076     NF->addMetadata(MD.first, *MD.second);
1077 
1078   // Now that the old function is dead, delete it.
1079   F->eraseFromParent();
1080 
1081   return true;
1082 }
1083 
1084 PreservedAnalyses DeadArgumentEliminationPass::run(Module &M,
1085                                                    ModuleAnalysisManager &) {
1086   bool Changed = false;
1087 
1088   // First pass: Do a simple check to see if any functions can have their "..."
1089   // removed.  We can do this if they never call va_start.  This loop cannot be
1090   // fused with the next loop, because deleting a function invalidates
1091   // information computed while surveying other functions.
1092   LLVM_DEBUG(dbgs() << "DeadArgumentEliminationPass - Deleting dead varargs\n");
1093   for (Module::iterator I = M.begin(), E = M.end(); I != E; ) {
1094     Function &F = *I++;
1095     if (F.getFunctionType()->isVarArg())
1096       Changed |= DeleteDeadVarargs(F);
1097   }
1098 
1099   // Second phase:loop through the module, determining which arguments are live.
1100   // We assume all arguments are dead unless proven otherwise (allowing us to
1101   // determine that dead arguments passed into recursive functions are dead).
1102   //
1103   LLVM_DEBUG(dbgs() << "DeadArgumentEliminationPass - Determining liveness\n");
1104   for (auto &F : M)
1105     SurveyFunction(F);
1106 
1107   // Now, remove all dead arguments and return values from each function in
1108   // turn.
1109   for (Module::iterator I = M.begin(), E = M.end(); I != E; ) {
1110     // Increment now, because the function will probably get removed (ie.
1111     // replaced by a new one).
1112     Function *F = &*I++;
1113     Changed |= RemoveDeadStuffFromFunction(F);
1114   }
1115 
1116   // Finally, look for any unused parameters in functions with non-local
1117   // linkage and replace the passed in parameters with undef.
1118   for (auto &F : M)
1119     Changed |= RemoveDeadArgumentsFromCallers(F);
1120 
1121   if (!Changed)
1122     return PreservedAnalyses::all();
1123   return PreservedAnalyses::none();
1124 }
1125