1 //===-- ArgumentPromotion.cpp - Promote by-reference arguments ------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass promotes "by reference" arguments to be "by value" arguments. In
11 // practice, this means looking for internal functions that have pointer
12 // arguments. If we can prove, through the use of alias analysis, that an
13 // argument is *only* loaded, then we can pass the value into the function
14 // instead of the address of the value. This can cause recursive simplification
15 // of code and lead to the elimination of allocas (especially in C++ template
16 // code like the STL).
18 // This pass also handles aggregate arguments that are passed into a function,
19 // scalarizing them if the elements of the aggregate are only loaded. Note that
20 // we refuse to scalarize aggregates which would require passing in more than
21 // three operands to the function, because we don't want to pass thousands of
22 // operands for a large array or structure!
24 // Note that this transformation could also be done for arguments that are only
25 // stored to (returning the value instead), but we do not currently handle that
26 // case. This case would be best handled when and if we start supporting
27 // multiple return values from functions.
29 //===----------------------------------------------------------------------===//
31 #define DEBUG_TYPE "argpromotion"
32 #include "llvm/Transforms/IPO.h"
33 #include "llvm/Constants.h"
34 #include "llvm/DerivedTypes.h"
35 #include "llvm/Module.h"
36 #include "llvm/CallGraphSCCPass.h"
37 #include "llvm/Instructions.h"
38 #include "llvm/Analysis/AliasAnalysis.h"
39 #include "llvm/Analysis/CallGraph.h"
40 #include "llvm/Target/TargetData.h"
41 #include "llvm/Support/CallSite.h"
42 #include "llvm/Support/CFG.h"
43 #include "llvm/Support/Debug.h"
44 #include "llvm/ADT/DepthFirstIterator.h"
45 #include "llvm/ADT/Statistic.h"
46 #include "llvm/ADT/StringExtras.h"
51 Statistic<> NumArgumentsPromoted("argpromotion",
52 "Number of pointer arguments promoted");
53 Statistic<> NumAggregatesPromoted("argpromotion",
54 "Number of aggregate arguments promoted");
55 Statistic<> NumArgumentsDead("argpromotion",
56 "Number of dead pointer args eliminated");
58 /// ArgPromotion - The 'by reference' to 'by value' argument promotion pass.
60 struct ArgPromotion : public CallGraphSCCPass {
61 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
62 AU.addRequired<AliasAnalysis>();
63 AU.addRequired<TargetData>();
64 CallGraphSCCPass::getAnalysisUsage(AU);
67 virtual bool runOnSCC(const std::vector<CallGraphNode *> &SCC);
69 bool PromoteArguments(CallGraphNode *CGN);
70 bool isSafeToPromoteArgument(Argument *Arg) const;
71 Function *DoPromotion(Function *F, std::vector<Argument*> &ArgsToPromote);
74 RegisterOpt<ArgPromotion> X("argpromotion",
75 "Promote 'by reference' arguments to scalars");
78 ModulePass *llvm::createArgumentPromotionPass() {
79 return new ArgPromotion();
82 bool ArgPromotion::runOnSCC(const std::vector<CallGraphNode *> &SCC) {
83 bool Changed = false, LocalChange;
85 do { // Iterate until we stop promoting from this SCC.
87 // Attempt to promote arguments from all functions in this SCC.
88 for (unsigned i = 0, e = SCC.size(); i != e; ++i)
89 LocalChange |= PromoteArguments(SCC[i]);
90 Changed |= LocalChange; // Remember that we changed something.
91 } while (LocalChange);
96 /// PromoteArguments - This method checks the specified function to see if there
97 /// are any promotable arguments and if it is safe to promote the function (for
98 /// example, all callers are direct). If safe to promote some arguments, it
99 /// calls the DoPromotion method.
101 bool ArgPromotion::PromoteArguments(CallGraphNode *CGN) {
102 Function *F = CGN->getFunction();
104 // Make sure that it is local to this module.
105 if (!F || !F->hasInternalLinkage()) return false;
107 // First check: see if there are any pointer arguments! If not, quick exit.
108 std::vector<Argument*> PointerArgs;
109 for (Function::aiterator I = F->abegin(), E = F->aend(); I != E; ++I)
110 if (isa<PointerType>(I->getType()))
111 PointerArgs.push_back(I);
112 if (PointerArgs.empty()) return false;
114 // Second check: make sure that all callers are direct callers. We can't
115 // transform functions that have indirect callers.
116 for (Value::use_iterator UI = F->use_begin(), E = F->use_end();
118 CallSite CS = CallSite::get(*UI);
119 if (!CS.getInstruction()) // "Taking the address" of the function
122 // Ensure that this call site is CALLING the function, not passing it as
124 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
126 if (*AI == F) return false; // Passing the function address in!
129 // Check to see which arguments are promotable. If an argument is not
130 // promotable, remove it from the PointerArgs vector.
131 for (unsigned i = 0; i != PointerArgs.size(); ++i)
132 if (!isSafeToPromoteArgument(PointerArgs[i])) {
133 std::swap(PointerArgs[i--], PointerArgs.back());
134 PointerArgs.pop_back();
137 // No promotable pointer arguments.
138 if (PointerArgs.empty()) return false;
140 // Okay, promote all of the arguments are rewrite the callees!
141 Function *NewF = DoPromotion(F, PointerArgs);
143 // Update the call graph to know that the old function is gone.
144 getAnalysis<CallGraph>().changeFunction(F, NewF);
149 /// isSafeToPromoteArgument - As you might guess from the name of this method,
150 /// it checks to see if it is both safe and useful to promote the argument.
151 /// This method limits promotion of aggregates to only promote up to three
152 /// elements of the aggregate in order to avoid exploding the number of
153 /// arguments passed in.
154 bool ArgPromotion::isSafeToPromoteArgument(Argument *Arg) const {
155 // We can only promote this argument if all of the uses are loads, or are GEP
156 // instructions (with constant indices) that are subsequently loaded.
157 std::vector<LoadInst*> Loads;
158 std::vector<std::vector<ConstantInt*> > GEPIndices;
159 for (Value::use_iterator UI = Arg->use_begin(), E = Arg->use_end();
161 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
162 if (LI->isVolatile()) return false; // Don't hack volatile loads
164 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) {
165 if (GEP->use_empty()) {
166 // Dead GEP's cause trouble later. Just remove them if we run into
168 getAnalysis<AliasAnalysis>().deleteValue(GEP);
169 GEP->getParent()->getInstList().erase(GEP);
170 return isSafeToPromoteArgument(Arg);
172 // Ensure that all of the indices are constants.
173 std::vector<ConstantInt*> Operands;
174 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
175 if (ConstantInt *C = dyn_cast<ConstantInt>(GEP->getOperand(i)))
176 Operands.push_back(C);
178 return false; // Not a constant operand GEP!
180 // Ensure that the only users of the GEP are load instructions.
181 for (Value::use_iterator UI = GEP->use_begin(), E = GEP->use_end();
183 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
184 if (LI->isVolatile()) return false; // Don't hack volatile loads
190 // See if there is already a GEP with these indices. If not, check to
191 // make sure that we aren't promoting too many elements. If so, nothing
193 if (std::find(GEPIndices.begin(), GEPIndices.end(), Operands) ==
195 if (GEPIndices.size() == 3) {
196 DEBUG(std::cerr << "argpromotion disable promoting argument '"
197 << Arg->getName() << "' because it would require adding more "
198 << "than 3 arguments to the function.\n");
199 // We limit aggregate promotion to only promoting up to three elements
203 GEPIndices.push_back(Operands);
206 return false; // Not a load or a GEP.
209 if (Loads.empty()) return true; // No users, this is a dead argument.
211 // Okay, now we know that the argument is only used by load instructions. Use
212 // alias analysis to check to see if the pointer is guaranteed to not be
213 // modified from entry of the function to each of the load instructions.
214 Function &F = *Arg->getParent();
216 // Because there could be several/many load instructions, remember which
217 // blocks we know to be transparent to the load.
218 std::set<BasicBlock*> TranspBlocks;
220 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
221 TargetData &TD = getAnalysis<TargetData>();
223 for (unsigned i = 0, e = Loads.size(); i != e; ++i) {
224 // Check to see if the load is invalidated from the start of the block to
226 LoadInst *Load = Loads[i];
227 BasicBlock *BB = Load->getParent();
229 const PointerType *LoadTy =
230 cast<PointerType>(Load->getOperand(0)->getType());
231 unsigned LoadSize = TD.getTypeSize(LoadTy->getElementType());
233 if (AA.canInstructionRangeModify(BB->front(), *Load, Arg, LoadSize))
234 return false; // Pointer is invalidated!
236 // Now check every path from the entry block to the load for transparency.
237 // To do this, we perform a depth first search on the inverse CFG from the
239 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
240 for (idf_ext_iterator<BasicBlock*> I = idf_ext_begin(*PI, TranspBlocks),
241 E = idf_ext_end(*PI, TranspBlocks); I != E; ++I)
242 if (AA.canBasicBlockModify(**I, Arg, LoadSize))
246 // If the path from the entry of the function to each load is free of
247 // instructions that potentially invalidate the load, we can make the
253 /// GEPIdxComparator - Provide a strong ordering for GEP indices. All Value*
254 /// elements are instances of ConstantInt.
256 struct GEPIdxComparator {
257 bool operator()(const std::vector<Value*> &LHS,
258 const std::vector<Value*> &RHS) const {
260 for (; idx < LHS.size() && idx < RHS.size(); ++idx) {
261 if (LHS[idx] != RHS[idx]) {
262 return cast<ConstantInt>(LHS[idx])->getRawValue() <
263 cast<ConstantInt>(RHS[idx])->getRawValue();
267 // Return less than if we ran out of stuff in LHS and we didn't run out of
269 return idx == LHS.size() && idx != RHS.size();
275 /// DoPromotion - This method actually performs the promotion of the specified
276 /// arguments, and returns the new function. At this point, we know that it's
278 Function *ArgPromotion::DoPromotion(Function *F,
279 std::vector<Argument*> &Args2Prom) {
280 std::set<Argument*> ArgsToPromote(Args2Prom.begin(), Args2Prom.end());
282 // Start by computing a new prototype for the function, which is the same as
283 // the old function, but has modified arguments.
284 const FunctionType *FTy = F->getFunctionType();
285 std::vector<const Type*> Params;
287 typedef std::set<std::vector<Value*>, GEPIdxComparator> ScalarizeTable;
289 // ScalarizedElements - If we are promoting a pointer that has elements
290 // accessed out of it, keep track of which elements are accessed so that we
291 // can add one argument for each.
293 // Arguments that are directly loaded will have a zero element value here, to
294 // handle cases where there are both a direct load and GEP accesses.
296 std::map<Argument*, ScalarizeTable> ScalarizedElements;
298 // OriginalLoads - Keep track of a representative load instruction from the
299 // original function so that we can tell the alias analysis implementation
300 // what the new GEP/Load instructions we are inserting look like.
301 std::map<std::vector<Value*>, LoadInst*> OriginalLoads;
303 for (Function::aiterator I = F->abegin(), E = F->aend(); I != E; ++I)
304 if (!ArgsToPromote.count(I)) {
305 Params.push_back(I->getType());
306 } else if (I->use_empty()) {
309 // Okay, this is being promoted. Check to see if there are any GEP uses
311 ScalarizeTable &ArgIndices = ScalarizedElements[I];
312 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
314 Instruction *User = cast<Instruction>(*UI);
315 assert(isa<LoadInst>(User) || isa<GetElementPtrInst>(User));
316 std::vector<Value*> Indices(User->op_begin()+1, User->op_end());
317 ArgIndices.insert(Indices);
319 if (LoadInst *L = dyn_cast<LoadInst>(User))
322 OrigLoad = cast<LoadInst>(User->use_back());
323 OriginalLoads[Indices] = OrigLoad;
326 // Add a parameter to the function for each element passed in.
327 for (ScalarizeTable::iterator SI = ArgIndices.begin(),
328 E = ArgIndices.end(); SI != E; ++SI)
329 Params.push_back(GetElementPtrInst::getIndexedType(I->getType(), *SI));
331 if (ArgIndices.size() == 1 && ArgIndices.begin()->empty())
332 ++NumArgumentsPromoted;
334 ++NumAggregatesPromoted;
337 const Type *RetTy = FTy->getReturnType();
339 // Work around LLVM bug PR56: the CWriter cannot emit varargs functions which
340 // have zero fixed arguments.
341 bool ExtraArgHack = false;
342 if (Params.empty() && FTy->isVarArg()) {
344 Params.push_back(Type::IntTy);
346 FunctionType *NFTy = FunctionType::get(RetTy, Params, FTy->isVarArg());
348 // Create the new function body and insert it into the module...
349 Function *NF = new Function(NFTy, F->getLinkage(), F->getName());
350 F->getParent()->getFunctionList().insert(F, NF);
352 // Get the alias analysis information that we need to update to reflect our
354 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
356 // Loop over all of the callers of the function, transforming the call sites
357 // to pass in the loaded pointers.
359 std::vector<Value*> Args;
360 while (!F->use_empty()) {
361 CallSite CS = CallSite::get(F->use_back());
362 Instruction *Call = CS.getInstruction();
364 // Loop over the operands, inserting GEP and loads in the caller as
366 CallSite::arg_iterator AI = CS.arg_begin();
367 for (Function::aiterator I = F->abegin(), E = F->aend(); I != E; ++I, ++AI)
368 if (!ArgsToPromote.count(I))
369 Args.push_back(*AI); // Unmodified argument
370 else if (!I->use_empty()) {
371 // Non-dead argument: insert GEPs and loads as appropriate.
372 ScalarizeTable &ArgIndices = ScalarizedElements[I];
373 for (ScalarizeTable::iterator SI = ArgIndices.begin(),
374 E = ArgIndices.end(); SI != E; ++SI) {
376 LoadInst *OrigLoad = OriginalLoads[*SI];
378 V = new GetElementPtrInst(V, *SI, V->getName()+".idx", Call);
379 AA.copyValue(OrigLoad->getOperand(0), V);
381 Args.push_back(new LoadInst(V, V->getName()+".val", Call));
382 AA.copyValue(OrigLoad, Args.back());
387 Args.push_back(Constant::getNullValue(Type::IntTy));
389 // Push any varargs arguments on the list
390 for (; AI != CS.arg_end(); ++AI)
394 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) {
395 New = new InvokeInst(NF, II->getNormalDest(), II->getUnwindDest(),
398 New = new CallInst(NF, Args, "", Call);
402 // Update the alias analysis implementation to know that we are replacing
403 // the old call with a new one.
404 AA.replaceWithNewValue(Call, New);
406 if (!Call->use_empty()) {
407 Call->replaceAllUsesWith(New);
408 std::string Name = Call->getName();
413 // Finally, remove the old call from the program, reducing the use-count of
415 Call->getParent()->getInstList().erase(Call);
418 // Since we have now created the new function, splice the body of the old
419 // function right into the new function, leaving the old rotting hulk of the
421 NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList());
423 // Loop over the argument list, transfering uses of the old arguments over to
424 // the new arguments, also transfering over the names as well.
426 for (Function::aiterator I = F->abegin(), E = F->aend(), I2 = NF->abegin();
428 if (!ArgsToPromote.count(I)) {
429 // If this is an unmodified argument, move the name and users over to the
431 I->replaceAllUsesWith(I2);
432 I2->setName(I->getName());
433 AA.replaceWithNewValue(I, I2);
435 } else if (I->use_empty()) {
438 // Otherwise, if we promoted this argument, then all users are load
439 // instructions, and all loads should be using the new argument that we
441 ScalarizeTable &ArgIndices = ScalarizedElements[I];
443 while (!I->use_empty()) {
444 if (LoadInst *LI = dyn_cast<LoadInst>(I->use_back())) {
445 assert(ArgIndices.begin()->empty() &&
446 "Load element should sort to front!");
447 I2->setName(I->getName()+".val");
448 LI->replaceAllUsesWith(I2);
449 AA.replaceWithNewValue(LI, I2);
450 LI->getParent()->getInstList().erase(LI);
451 DEBUG(std::cerr << "*** Promoted load of argument '" << I->getName()
452 << "' in function '" << F->getName() << "'\n");
454 GetElementPtrInst *GEP = cast<GetElementPtrInst>(I->use_back());
455 std::vector<Value*> Operands(GEP->op_begin()+1, GEP->op_end());
458 Function::aiterator TheArg = I2;
459 for (ScalarizeTable::iterator It = ArgIndices.begin();
460 *It != Operands; ++It, ++TheArg) {
461 assert(It != ArgIndices.end() && "GEP not handled??");
464 std::string NewName = I->getName();
465 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
466 if (ConstantInt *CI = dyn_cast<ConstantInt>(Operands[i]))
467 NewName += "."+itostr((int64_t)CI->getRawValue());
470 TheArg->setName(NewName+".val");
472 DEBUG(std::cerr << "*** Promoted agg argument '" << TheArg->getName()
473 << "' of function '" << F->getName() << "'\n");
475 // All of the uses must be load instructions. Replace them all with
476 // the argument specified by ArgNo.
477 while (!GEP->use_empty()) {
478 LoadInst *L = cast<LoadInst>(GEP->use_back());
479 L->replaceAllUsesWith(TheArg);
480 AA.replaceWithNewValue(L, TheArg);
481 L->getParent()->getInstList().erase(L);
484 GEP->getParent()->getInstList().erase(GEP);
488 // Increment I2 past all of the arguments added for this promoted pointer.
489 for (unsigned i = 0, e = ArgIndices.size(); i != e; ++i)
493 // Notify the alias analysis implementation that we inserted a new argument.
495 AA.copyValue(Constant::getNullValue(Type::IntTy), NF->abegin());
498 // Tell the alias analysis that the old function is about to disappear.
499 AA.replaceWithNewValue(F, NF);
501 // Now that the old function is dead, delete it.
502 F->getParent()->getFunctionList().erase(F);