1 //===- SimplifyLibCalls.cpp - Optimize specific well-known library calls --===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Reid Spencer and is distributed under the
6 // University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements a module pass that applies a variety of small
11 // optimizations for calls to specific well-known function calls (e.g. runtime
12 // library functions). For example, a call to the function "exit(3)" that
13 // occurs within the main() function can be transformed into a simple "return 3"
14 // instruction. Any optimization that takes this form (replace call to library
15 // function with simpler code that provides the same result) belongs in this
18 //===----------------------------------------------------------------------===//
20 #define DEBUG_TYPE "simplify-libcalls"
21 #include "llvm/Constants.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/Instructions.h"
24 #include "llvm/Module.h"
25 #include "llvm/Pass.h"
26 #include "llvm/ADT/hash_map"
27 #include "llvm/ADT/Statistic.h"
28 #include "llvm/Support/Debug.h"
29 #include "llvm/Target/TargetData.h"
30 #include "llvm/Transforms/IPO.h"
36 /// This statistic keeps track of the total number of library calls that have
37 /// been simplified regardless of which call it is.
38 Statistic<> SimplifiedLibCalls("simplify-libcalls",
39 "Number of library calls simplified");
41 // Forward declarations
42 class LibCallOptimization;
43 class SimplifyLibCalls;
45 /// This hash map is populated by the constructor for LibCallOptimization class.
46 /// Therefore all subclasses are registered here at static initialization time
47 /// and this list is what the SimplifyLibCalls pass uses to apply the individual
48 /// optimizations to the call sites.
49 /// @brief The list of optimizations deriving from LibCallOptimization
50 static hash_map<std::string,LibCallOptimization*> optlist;
52 /// This class is the abstract base class for the set of optimizations that
53 /// corresponds to one library call. The SimplifyLibCalls pass will call the
54 /// ValidateCalledFunction method to ask the optimization if a given Function
55 /// is the kind that the optimization can handle. If the subclass returns true,
56 /// then SImplifyLibCalls will also call the OptimizeCall method to perform,
57 /// or attempt to perform, the optimization(s) for the library call. Otherwise,
58 /// OptimizeCall won't be called. Subclasses are responsible for providing the
59 /// name of the library call (strlen, strcpy, etc.) to the LibCallOptimization
60 /// constructor. This is used to efficiently select which call instructions to
61 /// optimize. The criteria for a "lib call" is "anything with well known
62 /// semantics", typically a library function that is defined by an international
63 /// standard. Because the semantics are well known, the optimizations can
64 /// generally short-circuit actually calling the function if there's a simpler
65 /// way (e.g. strlen(X) can be reduced to a constant if X is a constant global).
66 /// @brief Base class for library call optimizations
67 class LibCallOptimization
70 /// The \p fname argument must be the name of the library function being
71 /// optimized by the subclass.
72 /// @brief Constructor that registers the optimization.
73 LibCallOptimization(const char* fname, const char* description )
76 , occurrences("simplify-libcalls",description)
79 // Register this call optimizer in the optlist (a hash_map)
80 optlist[fname] = this;
83 /// @brief Deregister from the optlist
84 virtual ~LibCallOptimization() { optlist.erase(func_name); }
86 /// The implementation of this function in subclasses should determine if
87 /// \p F is suitable for the optimization. This method is called by
88 /// SimplifyLibCalls::runOnModule to short circuit visiting all the call
89 /// sites of such a function if that function is not suitable in the first
90 /// place. If the called function is suitabe, this method should return true;
91 /// false, otherwise. This function should also perform any lazy
92 /// initialization that the LibCallOptimization needs to do, if its to return
93 /// true. This avoids doing initialization until the optimizer is actually
94 /// going to be called upon to do some optimization.
95 /// @brief Determine if the function is suitable for optimization
96 virtual bool ValidateCalledFunction(
97 const Function* F, ///< The function that is the target of call sites
98 SimplifyLibCalls& SLC ///< The pass object invoking us
101 /// The implementations of this function in subclasses is the heart of the
102 /// SimplifyLibCalls algorithm. Sublcasses of this class implement
103 /// OptimizeCall to determine if (a) the conditions are right for optimizing
104 /// the call and (b) to perform the optimization. If an action is taken
105 /// against ci, the subclass is responsible for returning true and ensuring
106 /// that ci is erased from its parent.
107 /// @brief Optimize a call, if possible.
108 virtual bool OptimizeCall(
109 CallInst* ci, ///< The call instruction that should be optimized.
110 SimplifyLibCalls& SLC ///< The pass object invoking us
113 /// @brief Get the name of the library call being optimized
114 const char * getFunctionName() const { return func_name; }
117 /// @brief Called by SimplifyLibCalls to update the occurrences statistic.
118 void succeeded() { DEBUG(++occurrences); }
122 const char* func_name; ///< Name of the library call we optimize
124 Statistic<> occurrences; ///< debug statistic (-debug-only=simplify-libcalls)
128 /// This class is an LLVM Pass that applies each of the LibCallOptimization
129 /// instances to all the call sites in a module, relatively efficiently. The
130 /// purpose of this pass is to provide optimizations for calls to well-known
131 /// functions with well-known semantics, such as those in the c library. The
132 /// class provides the basic infrastructure for handling runOnModule. Whenever
133 /// this pass finds a function call, it asks the appropriate optimizer to
134 /// validate the call (ValidateLibraryCall). If it is validated, then
135 /// the OptimizeCall method is also called.
136 /// @brief A ModulePass for optimizing well-known function calls.
137 class SimplifyLibCalls : public ModulePass
140 /// We need some target data for accurate signature details that are
141 /// target dependent. So we require target data in our AnalysisUsage.
142 /// @brief Require TargetData from AnalysisUsage.
143 virtual void getAnalysisUsage(AnalysisUsage& Info) const
145 // Ask that the TargetData analysis be performed before us so we can use
147 Info.addRequired<TargetData>();
150 /// For this pass, process all of the function calls in the module, calling
151 /// ValidateLibraryCall and OptimizeCall as appropriate.
152 /// @brief Run all the lib call optimizations on a Module.
153 virtual bool runOnModule(Module &M)
159 // The call optimizations can be recursive. That is, the optimization might
160 // generate a call to another function which can also be optimized. This way
161 // we make the LibCallOptimization instances very specific to the case they
162 // handle. It also means we need to keep running over the function calls in
163 // the module until we don't get any more optimizations possible.
164 bool found_optimization = false;
167 found_optimization = false;
168 for (Module::iterator FI = M.begin(), FE = M.end(); FI != FE; ++FI)
170 // All the "well-known" functions are external and have external linkage
171 // because they live in a runtime library somewhere and were (probably)
172 // not compiled by LLVM. So, we only act on external functions that
173 // have external linkage and non-empty uses.
174 if (!FI->isExternal() || !FI->hasExternalLinkage() || FI->use_empty())
177 // Get the optimization class that pertains to this function
178 LibCallOptimization* CO = optlist[FI->getName().c_str()];
182 // Make sure the called function is suitable for the optimization
183 if (!CO->ValidateCalledFunction(FI,*this))
186 // Loop over each of the uses of the function
187 for (Value::use_iterator UI = FI->use_begin(), UE = FI->use_end();
190 // If the use of the function is a call instruction
191 if (CallInst* CI = dyn_cast<CallInst>(*UI++))
193 // Do the optimization on the LibCallOptimization.
194 if (CO->OptimizeCall(CI,*this))
196 ++SimplifiedLibCalls;
197 found_optimization = result = true;
205 } while (found_optimization);
209 /// @brief Return the *current* module we're working on.
210 Module* getModule() const { return M; }
212 /// @brief Return the *current* target data for the module we're working on.
213 TargetData* getTargetData() const { return TD; }
215 /// @brief Return the size_t type -- syntactic shortcut
216 const Type* getIntPtrType() const { return TD->getIntPtrType(); }
218 /// @brief Return a Function* for the fputc libcall
219 Function* get_fputc(const Type* FILEptr_type)
223 std::vector<const Type*> args;
224 args.push_back(Type::IntTy);
225 args.push_back(FILEptr_type);
226 FunctionType* fputc_type =
227 FunctionType::get(Type::IntTy, args, false);
228 fputc_func = M->getOrInsertFunction("fputc",fputc_type);
233 /// @brief Return a Function* for the fwrite libcall
234 Function* get_fwrite(const Type* FILEptr_type)
238 std::vector<const Type*> args;
239 args.push_back(PointerType::get(Type::SByteTy));
240 args.push_back(TD->getIntPtrType());
241 args.push_back(TD->getIntPtrType());
242 args.push_back(FILEptr_type);
243 FunctionType* fwrite_type =
244 FunctionType::get(TD->getIntPtrType(), args, false);
245 fwrite_func = M->getOrInsertFunction("fwrite",fwrite_type);
250 /// @brief Return a Function* for the sqrt libcall
255 std::vector<const Type*> args;
256 args.push_back(Type::DoubleTy);
257 FunctionType* sqrt_type =
258 FunctionType::get(Type::DoubleTy, args, false);
259 sqrt_func = M->getOrInsertFunction("sqrt",sqrt_type);
264 /// @brief Return a Function* for the strlen libcall
265 Function* get_strcpy()
269 std::vector<const Type*> args;
270 args.push_back(PointerType::get(Type::SByteTy));
271 args.push_back(PointerType::get(Type::SByteTy));
272 FunctionType* strcpy_type =
273 FunctionType::get(PointerType::get(Type::SByteTy), args, false);
274 strcpy_func = M->getOrInsertFunction("strcpy",strcpy_type);
279 /// @brief Return a Function* for the strlen libcall
280 Function* get_strlen()
284 std::vector<const Type*> args;
285 args.push_back(PointerType::get(Type::SByteTy));
286 FunctionType* strlen_type =
287 FunctionType::get(TD->getIntPtrType(), args, false);
288 strlen_func = M->getOrInsertFunction("strlen",strlen_type);
293 /// @brief Return a Function* for the memchr libcall
294 Function* get_memchr()
298 std::vector<const Type*> args;
299 args.push_back(PointerType::get(Type::SByteTy));
300 args.push_back(Type::IntTy);
301 args.push_back(TD->getIntPtrType());
302 FunctionType* memchr_type = FunctionType::get(
303 PointerType::get(Type::SByteTy), args, false);
304 memchr_func = M->getOrInsertFunction("memchr",memchr_type);
309 /// @brief Return a Function* for the memcpy libcall
310 Function* get_memcpy() {
312 const Type *SBP = PointerType::get(Type::SByteTy);
313 memcpy_func = M->getOrInsertFunction("llvm.memcpy", Type::VoidTy,SBP, SBP,
314 Type::UIntTy, Type::UIntTy, 0);
319 Function* get_floorf() {
321 floorf_func = M->getOrInsertFunction("floorf", Type::FloatTy,
327 /// @brief Reset our cached data for a new Module
328 void reset(Module& mod)
331 TD = &getAnalysis<TargetData>();
343 Function* fputc_func; ///< Cached fputc function
344 Function* fwrite_func; ///< Cached fwrite function
345 Function* memcpy_func; ///< Cached llvm.memcpy function
346 Function* memchr_func; ///< Cached memchr function
347 Function* sqrt_func; ///< Cached sqrt function
348 Function* strcpy_func; ///< Cached strcpy function
349 Function* strlen_func; ///< Cached strlen function
350 Function* floorf_func; ///< Cached floorf function
351 Module* M; ///< Cached Module
352 TargetData* TD; ///< Cached TargetData
356 RegisterOpt<SimplifyLibCalls>
357 X("simplify-libcalls","Simplify well-known library calls");
359 } // anonymous namespace
361 // The only public symbol in this file which just instantiates the pass object
362 ModulePass *llvm::createSimplifyLibCallsPass()
364 return new SimplifyLibCalls();
367 // Classes below here, in the anonymous namespace, are all subclasses of the
368 // LibCallOptimization class, each implementing all optimizations possible for a
369 // single well-known library call. Each has a static singleton instance that
370 // auto registers it into the "optlist" global above.
373 // Forward declare utility functions.
374 bool getConstantStringLength(Value* V, uint64_t& len, ConstantArray** A = 0 );
375 Value *CastToCStr(Value *V, Instruction &IP);
377 /// This LibCallOptimization will find instances of a call to "exit" that occurs
378 /// within the "main" function and change it to a simple "ret" instruction with
379 /// the same value passed to the exit function. When this is done, it splits the
380 /// basic block at the exit(3) call and deletes the call instruction.
381 /// @brief Replace calls to exit in main with a simple return
382 struct ExitInMainOptimization : public LibCallOptimization
384 ExitInMainOptimization() : LibCallOptimization("exit",
385 "Number of 'exit' calls simplified") {}
386 virtual ~ExitInMainOptimization() {}
388 // Make sure the called function looks like exit (int argument, int return
389 // type, external linkage, not varargs).
390 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC)
392 if (f->arg_size() >= 1)
393 if (f->arg_begin()->getType()->isInteger())
398 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC)
400 // To be careful, we check that the call to exit is coming from "main", that
401 // main has external linkage, and the return type of main and the argument
402 // to exit have the same type.
403 Function *from = ci->getParent()->getParent();
404 if (from->hasExternalLinkage())
405 if (from->getReturnType() == ci->getOperand(1)->getType())
406 if (from->getName() == "main")
408 // Okay, time to actually do the optimization. First, get the basic
409 // block of the call instruction
410 BasicBlock* bb = ci->getParent();
412 // Create a return instruction that we'll replace the call with.
413 // Note that the argument of the return is the argument of the call
415 ReturnInst* ri = new ReturnInst(ci->getOperand(1), ci);
417 // Split the block at the call instruction which places it in a new
419 bb->splitBasicBlock(ci);
421 // The block split caused a branch instruction to be inserted into
422 // the end of the original block, right after the return instruction
423 // that we put there. That's not a valid block, so delete the branch
425 bb->getInstList().pop_back();
427 // Now we can finally get rid of the call instruction which now lives
428 // in the new basic block.
429 ci->eraseFromParent();
431 // Optimization succeeded, return true.
434 // We didn't pass the criteria for this optimization so return false
437 } ExitInMainOptimizer;
439 /// This LibCallOptimization will simplify a call to the strcat library
440 /// function. The simplification is possible only if the string being
441 /// concatenated is a constant array or a constant expression that results in
442 /// a constant string. In this case we can replace it with strlen + llvm.memcpy
443 /// of the constant string. Both of these calls are further reduced, if possible
444 /// on subsequent passes.
445 /// @brief Simplify the strcat library function.
446 struct StrCatOptimization : public LibCallOptimization
449 /// @brief Default constructor
450 StrCatOptimization() : LibCallOptimization("strcat",
451 "Number of 'strcat' calls simplified") {}
454 /// @breif Destructor
455 virtual ~StrCatOptimization() {}
457 /// @brief Make sure that the "strcat" function has the right prototype
458 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC)
460 if (f->getReturnType() == PointerType::get(Type::SByteTy))
461 if (f->arg_size() == 2)
463 Function::const_arg_iterator AI = f->arg_begin();
464 if (AI++->getType() == PointerType::get(Type::SByteTy))
465 if (AI->getType() == PointerType::get(Type::SByteTy))
467 // Indicate this is a suitable call type.
474 /// @brief Optimize the strcat library function
475 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC)
477 // Extract some information from the instruction
478 Module* M = ci->getParent()->getParent()->getParent();
479 Value* dest = ci->getOperand(1);
480 Value* src = ci->getOperand(2);
482 // Extract the initializer (while making numerous checks) from the
483 // source operand of the call to strcat. If we get null back, one of
484 // a variety of checks in get_GVInitializer failed
486 if (!getConstantStringLength(src,len))
489 // Handle the simple, do-nothing case
492 ci->replaceAllUsesWith(dest);
493 ci->eraseFromParent();
497 // Increment the length because we actually want to memcpy the null
498 // terminator as well.
501 // We need to find the end of the destination string. That's where the
502 // memory is to be moved to. We just generate a call to strlen (further
503 // optimized in another pass). Note that the SLC.get_strlen() call
504 // caches the Function* for us.
505 CallInst* strlen_inst =
506 new CallInst(SLC.get_strlen(), dest, dest->getName()+".len",ci);
508 // Now that we have the destination's length, we must index into the
509 // destination's pointer to get the actual memcpy destination (end of
510 // the string .. we're concatenating).
511 std::vector<Value*> idx;
512 idx.push_back(strlen_inst);
513 GetElementPtrInst* gep =
514 new GetElementPtrInst(dest,idx,dest->getName()+".indexed",ci);
516 // We have enough information to now generate the memcpy call to
517 // do the concatenation for us.
518 std::vector<Value*> vals;
519 vals.push_back(gep); // destination
520 vals.push_back(ci->getOperand(2)); // source
521 vals.push_back(ConstantUInt::get(Type::UIntTy,len)); // length
522 vals.push_back(ConstantUInt::get(Type::UIntTy,1)); // alignment
523 new CallInst(SLC.get_memcpy(), vals, "", ci);
525 // Finally, substitute the first operand of the strcat call for the
526 // strcat call itself since strcat returns its first operand; and,
527 // kill the strcat CallInst.
528 ci->replaceAllUsesWith(dest);
529 ci->eraseFromParent();
534 /// This LibCallOptimization will simplify a call to the strchr library
535 /// function. It optimizes out cases where the arguments are both constant
536 /// and the result can be determined statically.
537 /// @brief Simplify the strcmp library function.
538 struct StrChrOptimization : public LibCallOptimization
541 StrChrOptimization() : LibCallOptimization("strchr",
542 "Number of 'strchr' calls simplified") {}
543 virtual ~StrChrOptimization() {}
545 /// @brief Make sure that the "strchr" function has the right prototype
546 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC)
548 if (f->getReturnType() == PointerType::get(Type::SByteTy) &&
554 /// @brief Perform the strchr optimizations
555 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC)
557 // If there aren't three operands, bail
558 if (ci->getNumOperands() != 3)
561 // Check that the first argument to strchr is a constant array of sbyte.
562 // If it is, get the length and data, otherwise return false.
565 if (!getConstantStringLength(ci->getOperand(1),len,&CA))
568 // Check that the second argument to strchr is a constant int, return false
570 ConstantSInt* CSI = dyn_cast<ConstantSInt>(ci->getOperand(2));
573 // Just lower this to memchr since we know the length of the string as
575 Function* f = SLC.get_memchr();
576 std::vector<Value*> args;
577 args.push_back(ci->getOperand(1));
578 args.push_back(ci->getOperand(2));
579 args.push_back(ConstantUInt::get(SLC.getIntPtrType(),len));
580 ci->replaceAllUsesWith( new CallInst(f,args,ci->getName(),ci));
581 ci->eraseFromParent();
585 // Get the character we're looking for
586 int64_t chr = CSI->getValue();
588 // Compute the offset
590 bool char_found = false;
591 for (uint64_t i = 0; i < len; ++i)
593 if (ConstantSInt* CI = dyn_cast<ConstantSInt>(CA->getOperand(i)))
595 // Check for the null terminator
596 if (CI->isNullValue())
597 break; // we found end of string
598 else if (CI->getValue() == chr)
607 // strchr(s,c) -> offset_of_in(c,s)
608 // (if c is a constant integer and s is a constant string)
611 std::vector<Value*> indices;
612 indices.push_back(ConstantUInt::get(Type::ULongTy,offset));
613 GetElementPtrInst* GEP = new GetElementPtrInst(ci->getOperand(1),indices,
614 ci->getOperand(1)->getName()+".strchr",ci);
615 ci->replaceAllUsesWith(GEP);
618 ci->replaceAllUsesWith(
619 ConstantPointerNull::get(PointerType::get(Type::SByteTy)));
621 ci->eraseFromParent();
626 /// This LibCallOptimization will simplify a call to the strcmp library
627 /// function. It optimizes out cases where one or both arguments are constant
628 /// and the result can be determined statically.
629 /// @brief Simplify the strcmp library function.
630 struct StrCmpOptimization : public LibCallOptimization
633 StrCmpOptimization() : LibCallOptimization("strcmp",
634 "Number of 'strcmp' calls simplified") {}
635 virtual ~StrCmpOptimization() {}
637 /// @brief Make sure that the "strcmp" function has the right prototype
638 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC)
640 if (f->getReturnType() == Type::IntTy && f->arg_size() == 2)
645 /// @brief Perform the strcmp optimization
646 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC)
648 // First, check to see if src and destination are the same. If they are,
649 // then the optimization is to replace the CallInst with a constant 0
650 // because the call is a no-op.
651 Value* s1 = ci->getOperand(1);
652 Value* s2 = ci->getOperand(2);
656 ci->replaceAllUsesWith(ConstantInt::get(Type::IntTy,0));
657 ci->eraseFromParent();
661 bool isstr_1 = false;
664 if (getConstantStringLength(s1,len_1,&A1))
669 // strcmp("",x) -> *x
671 new LoadInst(CastToCStr(s2,*ci), ci->getName()+".load",ci);
673 new CastInst(load,Type::IntTy,ci->getName()+".int",ci);
674 ci->replaceAllUsesWith(cast);
675 ci->eraseFromParent();
680 bool isstr_2 = false;
683 if (getConstantStringLength(s2,len_2,&A2))
688 // strcmp(x,"") -> *x
690 new LoadInst(CastToCStr(s1,*ci),ci->getName()+".val",ci);
692 new CastInst(load,Type::IntTy,ci->getName()+".int",ci);
693 ci->replaceAllUsesWith(cast);
694 ci->eraseFromParent();
699 if (isstr_1 && isstr_2)
701 // strcmp(x,y) -> cnst (if both x and y are constant strings)
702 std::string str1 = A1->getAsString();
703 std::string str2 = A2->getAsString();
704 int result = strcmp(str1.c_str(), str2.c_str());
705 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy,result));
706 ci->eraseFromParent();
713 /// This LibCallOptimization will simplify a call to the strncmp library
714 /// function. It optimizes out cases where one or both arguments are constant
715 /// and the result can be determined statically.
716 /// @brief Simplify the strncmp library function.
717 struct StrNCmpOptimization : public LibCallOptimization
720 StrNCmpOptimization() : LibCallOptimization("strncmp",
721 "Number of 'strncmp' calls simplified") {}
722 virtual ~StrNCmpOptimization() {}
724 /// @brief Make sure that the "strncmp" function has the right prototype
725 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC)
727 if (f->getReturnType() == Type::IntTy && f->arg_size() == 3)
732 /// @brief Perform the strncpy optimization
733 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC)
735 // First, check to see if src and destination are the same. If they are,
736 // then the optimization is to replace the CallInst with a constant 0
737 // because the call is a no-op.
738 Value* s1 = ci->getOperand(1);
739 Value* s2 = ci->getOperand(2);
742 // strncmp(x,x,l) -> 0
743 ci->replaceAllUsesWith(ConstantInt::get(Type::IntTy,0));
744 ci->eraseFromParent();
748 // Check the length argument, if it is Constant zero then the strings are
750 uint64_t len_arg = 0;
751 bool len_arg_is_const = false;
752 if (ConstantInt* len_CI = dyn_cast<ConstantInt>(ci->getOperand(3)))
754 len_arg_is_const = true;
755 len_arg = len_CI->getRawValue();
758 // strncmp(x,y,0) -> 0
759 ci->replaceAllUsesWith(ConstantInt::get(Type::IntTy,0));
760 ci->eraseFromParent();
765 bool isstr_1 = false;
768 if (getConstantStringLength(s1,len_1,&A1))
773 // strncmp("",x) -> *x
774 LoadInst* load = new LoadInst(s1,ci->getName()+".load",ci);
776 new CastInst(load,Type::IntTy,ci->getName()+".int",ci);
777 ci->replaceAllUsesWith(cast);
778 ci->eraseFromParent();
783 bool isstr_2 = false;
786 if (getConstantStringLength(s2,len_2,&A2))
791 // strncmp(x,"") -> *x
792 LoadInst* load = new LoadInst(s2,ci->getName()+".val",ci);
794 new CastInst(load,Type::IntTy,ci->getName()+".int",ci);
795 ci->replaceAllUsesWith(cast);
796 ci->eraseFromParent();
801 if (isstr_1 && isstr_2 && len_arg_is_const)
803 // strncmp(x,y,const) -> constant
804 std::string str1 = A1->getAsString();
805 std::string str2 = A2->getAsString();
806 int result = strncmp(str1.c_str(), str2.c_str(), len_arg);
807 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy,result));
808 ci->eraseFromParent();
815 /// This LibCallOptimization will simplify a call to the strcpy library
816 /// function. Two optimizations are possible:
817 /// (1) If src and dest are the same and not volatile, just return dest
818 /// (2) If the src is a constant then we can convert to llvm.memmove
819 /// @brief Simplify the strcpy library function.
820 struct StrCpyOptimization : public LibCallOptimization
823 StrCpyOptimization() : LibCallOptimization("strcpy",
824 "Number of 'strcpy' calls simplified") {}
825 virtual ~StrCpyOptimization() {}
827 /// @brief Make sure that the "strcpy" function has the right prototype
828 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC)
830 if (f->getReturnType() == PointerType::get(Type::SByteTy))
831 if (f->arg_size() == 2)
833 Function::const_arg_iterator AI = f->arg_begin();
834 if (AI++->getType() == PointerType::get(Type::SByteTy))
835 if (AI->getType() == PointerType::get(Type::SByteTy))
837 // Indicate this is a suitable call type.
844 /// @brief Perform the strcpy optimization
845 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC)
847 // First, check to see if src and destination are the same. If they are,
848 // then the optimization is to replace the CallInst with the destination
849 // because the call is a no-op. Note that this corresponds to the
850 // degenerate strcpy(X,X) case which should have "undefined" results
851 // according to the C specification. However, it occurs sometimes and
852 // we optimize it as a no-op.
853 Value* dest = ci->getOperand(1);
854 Value* src = ci->getOperand(2);
857 ci->replaceAllUsesWith(dest);
858 ci->eraseFromParent();
862 // Get the length of the constant string referenced by the second operand,
863 // the "src" parameter. Fail the optimization if we can't get the length
864 // (note that getConstantStringLength does lots of checks to make sure this
867 if (!getConstantStringLength(ci->getOperand(2),len))
870 // If the constant string's length is zero we can optimize this by just
871 // doing a store of 0 at the first byte of the destination
874 new StoreInst(ConstantInt::get(Type::SByteTy,0),ci->getOperand(1),ci);
875 ci->replaceAllUsesWith(dest);
876 ci->eraseFromParent();
880 // Increment the length because we actually want to memcpy the null
881 // terminator as well.
884 // Extract some information from the instruction
885 Module* M = ci->getParent()->getParent()->getParent();
887 // We have enough information to now generate the memcpy call to
888 // do the concatenation for us.
889 std::vector<Value*> vals;
890 vals.push_back(dest); // destination
891 vals.push_back(src); // source
892 vals.push_back(ConstantUInt::get(Type::UIntTy,len)); // length
893 vals.push_back(ConstantUInt::get(Type::UIntTy,1)); // alignment
894 new CallInst(SLC.get_memcpy(), vals, "", ci);
896 // Finally, substitute the first operand of the strcat call for the
897 // strcat call itself since strcat returns its first operand; and,
898 // kill the strcat CallInst.
899 ci->replaceAllUsesWith(dest);
900 ci->eraseFromParent();
905 /// This LibCallOptimization will simplify a call to the strlen library
906 /// function by replacing it with a constant value if the string provided to
907 /// it is a constant array.
908 /// @brief Simplify the strlen library function.
909 struct StrLenOptimization : public LibCallOptimization
911 StrLenOptimization() : LibCallOptimization("strlen",
912 "Number of 'strlen' calls simplified") {}
913 virtual ~StrLenOptimization() {}
915 /// @brief Make sure that the "strlen" function has the right prototype
916 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC)
918 if (f->getReturnType() == SLC.getTargetData()->getIntPtrType())
919 if (f->arg_size() == 1)
920 if (Function::const_arg_iterator AI = f->arg_begin())
921 if (AI->getType() == PointerType::get(Type::SByteTy))
926 /// @brief Perform the strlen optimization
927 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC)
929 // Make sure we're dealing with an sbyte* here.
930 Value* str = ci->getOperand(1);
931 if (str->getType() != PointerType::get(Type::SByteTy))
934 // Does the call to strlen have exactly one use?
936 // Is that single use a binary operator?
937 if (BinaryOperator* bop = dyn_cast<BinaryOperator>(ci->use_back()))
938 // Is it compared against a constant integer?
939 if (ConstantInt* CI = dyn_cast<ConstantInt>(bop->getOperand(1)))
941 // Get the value the strlen result is compared to
942 uint64_t val = CI->getRawValue();
944 // If its compared against length 0 with == or !=
946 (bop->getOpcode() == Instruction::SetEQ ||
947 bop->getOpcode() == Instruction::SetNE))
949 // strlen(x) != 0 -> *x != 0
950 // strlen(x) == 0 -> *x == 0
951 LoadInst* load = new LoadInst(str,str->getName()+".first",ci);
952 BinaryOperator* rbop = BinaryOperator::create(bop->getOpcode(),
953 load, ConstantSInt::get(Type::SByteTy,0),
954 bop->getName()+".strlen", ci);
955 bop->replaceAllUsesWith(rbop);
956 bop->eraseFromParent();
957 ci->eraseFromParent();
962 // Get the length of the constant string operand
964 if (!getConstantStringLength(ci->getOperand(1),len))
967 // strlen("xyz") -> 3 (for example)
968 const Type *Ty = SLC.getTargetData()->getIntPtrType();
970 ci->replaceAllUsesWith(ConstantSInt::get(Ty, len));
972 ci->replaceAllUsesWith(ConstantUInt::get(Ty, len));
974 ci->eraseFromParent();
979 /// IsOnlyUsedInEqualsComparison - Return true if it only matters that the value
980 /// is equal or not-equal to zero.
981 static bool IsOnlyUsedInEqualsZeroComparison(Instruction *I) {
982 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
984 Instruction *User = cast<Instruction>(*UI);
985 if (User->getOpcode() == Instruction::SetNE ||
986 User->getOpcode() == Instruction::SetEQ) {
987 if (isa<Constant>(User->getOperand(1)) &&
988 cast<Constant>(User->getOperand(1))->isNullValue())
990 } else if (CastInst *CI = dyn_cast<CastInst>(User))
991 if (CI->getType() == Type::BoolTy)
993 // Unknown instruction.
999 /// This memcmpOptimization will simplify a call to the memcmp library
1001 struct memcmpOptimization : public LibCallOptimization {
1002 /// @brief Default Constructor
1003 memcmpOptimization()
1004 : LibCallOptimization("memcmp", "Number of 'memcmp' calls simplified") {}
1006 /// @brief Make sure that the "memcmp" function has the right prototype
1007 virtual bool ValidateCalledFunction(const Function *F, SimplifyLibCalls &TD) {
1008 Function::const_arg_iterator AI = F->arg_begin();
1009 if (F->arg_size() != 3 || !isa<PointerType>(AI->getType())) return false;
1010 if (!isa<PointerType>((++AI)->getType())) return false;
1011 if (!(++AI)->getType()->isInteger()) return false;
1012 if (!F->getReturnType()->isInteger()) return false;
1016 /// Because of alignment and instruction information that we don't have, we
1017 /// leave the bulk of this to the code generators.
1019 /// Note that we could do much more if we could force alignment on otherwise
1020 /// small aligned allocas, or if we could indicate that loads have a small
1022 virtual bool OptimizeCall(CallInst *CI, SimplifyLibCalls &TD) {
1023 Value *LHS = CI->getOperand(1), *RHS = CI->getOperand(2);
1025 // If the two operands are the same, return zero.
1027 // memcmp(s,s,x) -> 0
1028 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
1029 CI->eraseFromParent();
1033 // Make sure we have a constant length.
1034 ConstantInt *LenC = dyn_cast<ConstantInt>(CI->getOperand(3));
1035 if (!LenC) return false;
1036 uint64_t Len = LenC->getRawValue();
1038 // If the length is zero, this returns 0.
1041 // memcmp(s1,s2,0) -> 0
1042 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
1043 CI->eraseFromParent();
1046 // memcmp(S1,S2,1) -> *(ubyte*)S1 - *(ubyte*)S2
1047 const Type *UCharPtr = PointerType::get(Type::UByteTy);
1048 CastInst *Op1Cast = new CastInst(LHS, UCharPtr, LHS->getName(), CI);
1049 CastInst *Op2Cast = new CastInst(RHS, UCharPtr, RHS->getName(), CI);
1050 Value *S1V = new LoadInst(Op1Cast, LHS->getName()+".val", CI);
1051 Value *S2V = new LoadInst(Op2Cast, RHS->getName()+".val", CI);
1052 Value *RV = BinaryOperator::createSub(S1V, S2V, CI->getName()+".diff",CI);
1053 if (RV->getType() != CI->getType())
1054 RV = new CastInst(RV, CI->getType(), RV->getName(), CI);
1055 CI->replaceAllUsesWith(RV);
1056 CI->eraseFromParent();
1060 if (IsOnlyUsedInEqualsZeroComparison(CI)) {
1061 // TODO: IF both are aligned, use a short load/compare.
1063 // memcmp(S1,S2,2) -> S1[0]-S2[0] | S1[1]-S2[1] iff only ==/!= 0 matters
1064 const Type *UCharPtr = PointerType::get(Type::UByteTy);
1065 CastInst *Op1Cast = new CastInst(LHS, UCharPtr, LHS->getName(), CI);
1066 CastInst *Op2Cast = new CastInst(RHS, UCharPtr, RHS->getName(), CI);
1067 Value *S1V1 = new LoadInst(Op1Cast, LHS->getName()+".val1", CI);
1068 Value *S2V1 = new LoadInst(Op2Cast, RHS->getName()+".val1", CI);
1069 Value *D1 = BinaryOperator::createSub(S1V1, S2V1,
1070 CI->getName()+".d1", CI);
1071 Constant *One = ConstantInt::get(Type::IntTy, 1);
1072 Value *G1 = new GetElementPtrInst(Op1Cast, One, "next1v", CI);
1073 Value *G2 = new GetElementPtrInst(Op2Cast, One, "next2v", CI);
1074 Value *S1V2 = new LoadInst(G1, LHS->getName()+".val2", CI);
1075 Value *S2V2 = new LoadInst(G1, RHS->getName()+".val2", CI);
1076 Value *D2 = BinaryOperator::createSub(S1V2, S2V2,
1077 CI->getName()+".d1", CI);
1078 Value *Or = BinaryOperator::createOr(D1, D2, CI->getName()+".res", CI);
1079 if (Or->getType() != CI->getType())
1080 Or = new CastInst(Or, CI->getType(), Or->getName(), CI);
1081 CI->replaceAllUsesWith(Or);
1082 CI->eraseFromParent();
1100 /// This LibCallOptimization will simplify a call to the memcpy library
1101 /// function by expanding it out to a single store of size 0, 1, 2, 4, or 8
1102 /// bytes depending on the length of the string and the alignment. Additional
1103 /// optimizations are possible in code generation (sequence of immediate store)
1104 /// @brief Simplify the memcpy library function.
1105 struct LLVMMemCpyOptimization : public LibCallOptimization
1107 /// @brief Default Constructor
1108 LLVMMemCpyOptimization() : LibCallOptimization("llvm.memcpy",
1109 "Number of 'llvm.memcpy' calls simplified") {}
1112 /// @brief Subclass Constructor
1113 LLVMMemCpyOptimization(const char* fname, const char* desc)
1114 : LibCallOptimization(fname, desc) {}
1116 /// @brief Destructor
1117 virtual ~LLVMMemCpyOptimization() {}
1119 /// @brief Make sure that the "memcpy" function has the right prototype
1120 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& TD)
1122 // Just make sure this has 4 arguments per LLVM spec.
1123 return (f->arg_size() == 4);
1126 /// Because of alignment and instruction information that we don't have, we
1127 /// leave the bulk of this to the code generators. The optimization here just
1128 /// deals with a few degenerate cases where the length of the string and the
1129 /// alignment match the sizes of our intrinsic types so we can do a load and
1130 /// store instead of the memcpy call.
1131 /// @brief Perform the memcpy optimization.
1132 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& TD)
1134 // Make sure we have constant int values to work with
1135 ConstantInt* LEN = dyn_cast<ConstantInt>(ci->getOperand(3));
1138 ConstantInt* ALIGN = dyn_cast<ConstantInt>(ci->getOperand(4));
1142 // If the length is larger than the alignment, we can't optimize
1143 uint64_t len = LEN->getRawValue();
1144 uint64_t alignment = ALIGN->getRawValue();
1146 alignment = 1; // Alignment 0 is identity for alignment 1
1147 if (len > alignment)
1150 // Get the type we will cast to, based on size of the string
1151 Value* dest = ci->getOperand(1);
1152 Value* src = ci->getOperand(2);
1157 // memcpy(d,s,0,a) -> noop
1158 ci->eraseFromParent();
1160 case 1: castType = Type::SByteTy; break;
1161 case 2: castType = Type::ShortTy; break;
1162 case 4: castType = Type::IntTy; break;
1163 case 8: castType = Type::LongTy; break;
1168 // Cast source and dest to the right sized primitive and then load/store
1170 new CastInst(src,PointerType::get(castType),src->getName()+".cast",ci);
1171 CastInst* DestCast =
1172 new CastInst(dest,PointerType::get(castType),dest->getName()+".cast",ci);
1173 LoadInst* LI = new LoadInst(SrcCast,SrcCast->getName()+".val",ci);
1174 StoreInst* SI = new StoreInst(LI, DestCast, ci);
1175 ci->eraseFromParent();
1178 } LLVMMemCpyOptimizer;
1180 /// This LibCallOptimization will simplify a call to the memmove library
1181 /// function. It is identical to MemCopyOptimization except for the name of
1183 /// @brief Simplify the memmove library function.
1184 struct LLVMMemMoveOptimization : public LLVMMemCpyOptimization
1186 /// @brief Default Constructor
1187 LLVMMemMoveOptimization() : LLVMMemCpyOptimization("llvm.memmove",
1188 "Number of 'llvm.memmove' calls simplified") {}
1190 } LLVMMemMoveOptimizer;
1192 /// This LibCallOptimization will simplify a call to the memset library
1193 /// function by expanding it out to a single store of size 0, 1, 2, 4, or 8
1194 /// bytes depending on the length argument.
1195 struct LLVMMemSetOptimization : public LibCallOptimization
1197 /// @brief Default Constructor
1198 LLVMMemSetOptimization() : LibCallOptimization("llvm.memset",
1199 "Number of 'llvm.memset' calls simplified") {}
1202 /// @brief Destructor
1203 virtual ~LLVMMemSetOptimization() {}
1205 /// @brief Make sure that the "memset" function has the right prototype
1206 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& TD)
1208 // Just make sure this has 3 arguments per LLVM spec.
1209 return (f->arg_size() == 4);
1212 /// Because of alignment and instruction information that we don't have, we
1213 /// leave the bulk of this to the code generators. The optimization here just
1214 /// deals with a few degenerate cases where the length parameter is constant
1215 /// and the alignment matches the sizes of our intrinsic types so we can do
1216 /// store instead of the memcpy call. Other calls are transformed into the
1217 /// llvm.memset intrinsic.
1218 /// @brief Perform the memset optimization.
1219 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& TD)
1221 // Make sure we have constant int values to work with
1222 ConstantInt* LEN = dyn_cast<ConstantInt>(ci->getOperand(3));
1225 ConstantInt* ALIGN = dyn_cast<ConstantInt>(ci->getOperand(4));
1229 // Extract the length and alignment
1230 uint64_t len = LEN->getRawValue();
1231 uint64_t alignment = ALIGN->getRawValue();
1233 // Alignment 0 is identity for alignment 1
1237 // If the length is zero, this is a no-op
1240 // memset(d,c,0,a) -> noop
1241 ci->eraseFromParent();
1245 // If the length is larger than the alignment, we can't optimize
1246 if (len > alignment)
1249 // Make sure we have a constant ubyte to work with so we can extract
1250 // the value to be filled.
1251 ConstantUInt* FILL = dyn_cast<ConstantUInt>(ci->getOperand(2));
1254 if (FILL->getType() != Type::UByteTy)
1257 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
1259 // Extract the fill character
1260 uint64_t fill_char = FILL->getValue();
1261 uint64_t fill_value = fill_char;
1263 // Get the type we will cast to, based on size of memory area to fill, and
1264 // and the value we will store there.
1265 Value* dest = ci->getOperand(1);
1270 castType = Type::UByteTy;
1273 castType = Type::UShortTy;
1274 fill_value |= fill_char << 8;
1277 castType = Type::UIntTy;
1278 fill_value |= fill_char << 8 | fill_char << 16 | fill_char << 24;
1281 castType = Type::ULongTy;
1282 fill_value |= fill_char << 8 | fill_char << 16 | fill_char << 24;
1283 fill_value |= fill_char << 32 | fill_char << 40 | fill_char << 48;
1284 fill_value |= fill_char << 56;
1290 // Cast dest to the right sized primitive and then load/store
1291 CastInst* DestCast =
1292 new CastInst(dest,PointerType::get(castType),dest->getName()+".cast",ci);
1293 new StoreInst(ConstantUInt::get(castType,fill_value),DestCast, ci);
1294 ci->eraseFromParent();
1297 } LLVMMemSetOptimizer;
1299 /// This LibCallOptimization will simplify calls to the "pow" library
1300 /// function. It looks for cases where the result of pow is well known and
1301 /// substitutes the appropriate value.
1302 /// @brief Simplify the pow library function.
1303 struct PowOptimization : public LibCallOptimization
1306 /// @brief Default Constructor
1307 PowOptimization() : LibCallOptimization("pow",
1308 "Number of 'pow' calls simplified") {}
1310 /// @brief Destructor
1311 virtual ~PowOptimization() {}
1313 /// @brief Make sure that the "pow" function has the right prototype
1314 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC)
1316 // Just make sure this has 2 arguments
1317 return (f->arg_size() == 2);
1320 /// @brief Perform the pow optimization.
1321 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC)
1323 const Type *Ty = cast<Function>(ci->getOperand(0))->getReturnType();
1324 Value* base = ci->getOperand(1);
1325 Value* expn = ci->getOperand(2);
1326 if (ConstantFP *Op1 = dyn_cast<ConstantFP>(base)) {
1327 double Op1V = Op1->getValue();
1330 // pow(1.0,x) -> 1.0
1331 ci->replaceAllUsesWith(ConstantFP::get(Ty,1.0));
1332 ci->eraseFromParent();
1336 else if (ConstantFP* Op2 = dyn_cast<ConstantFP>(expn))
1338 double Op2V = Op2->getValue();
1341 // pow(x,0.0) -> 1.0
1342 ci->replaceAllUsesWith(ConstantFP::get(Ty,1.0));
1343 ci->eraseFromParent();
1346 else if (Op2V == 0.5)
1348 // pow(x,0.5) -> sqrt(x)
1349 CallInst* sqrt_inst = new CallInst(SLC.get_sqrt(), base,
1350 ci->getName()+".pow",ci);
1351 ci->replaceAllUsesWith(sqrt_inst);
1352 ci->eraseFromParent();
1355 else if (Op2V == 1.0)
1358 ci->replaceAllUsesWith(base);
1359 ci->eraseFromParent();
1362 else if (Op2V == -1.0)
1364 // pow(x,-1.0) -> 1.0/x
1365 BinaryOperator* div_inst= BinaryOperator::createDiv(
1366 ConstantFP::get(Ty,1.0), base, ci->getName()+".pow", ci);
1367 ci->replaceAllUsesWith(div_inst);
1368 ci->eraseFromParent();
1372 return false; // opt failed
1376 /// This LibCallOptimization will simplify calls to the "fprintf" library
1377 /// function. It looks for cases where the result of fprintf is not used and the
1378 /// operation can be reduced to something simpler.
1379 /// @brief Simplify the pow library function.
1380 struct FPrintFOptimization : public LibCallOptimization
1383 /// @brief Default Constructor
1384 FPrintFOptimization() : LibCallOptimization("fprintf",
1385 "Number of 'fprintf' calls simplified") {}
1387 /// @brief Destructor
1388 virtual ~FPrintFOptimization() {}
1390 /// @brief Make sure that the "fprintf" function has the right prototype
1391 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC)
1393 // Just make sure this has at least 2 arguments
1394 return (f->arg_size() >= 2);
1397 /// @brief Perform the fprintf optimization.
1398 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC)
1400 // If the call has more than 3 operands, we can't optimize it
1401 if (ci->getNumOperands() > 4 || ci->getNumOperands() <= 2)
1404 // If the result of the fprintf call is used, none of these optimizations
1406 if (!ci->use_empty())
1409 // All the optimizations depend on the length of the second argument and the
1410 // fact that it is a constant string array. Check that now
1412 ConstantArray* CA = 0;
1413 if (!getConstantStringLength(ci->getOperand(2), len, &CA))
1416 if (ci->getNumOperands() == 3)
1418 // Make sure there's no % in the constant array
1419 for (unsigned i = 0; i < len; ++i)
1421 if (ConstantInt* CI = dyn_cast<ConstantInt>(CA->getOperand(i)))
1423 // Check for the null terminator
1424 if (CI->getRawValue() == '%')
1425 return false; // we found end of string
1431 // fprintf(file,fmt) -> fwrite(fmt,strlen(fmt),file)
1432 const Type* FILEptr_type = ci->getOperand(1)->getType();
1433 Function* fwrite_func = SLC.get_fwrite(FILEptr_type);
1437 // Make sure that the fprintf() and fwrite() functions both take the
1438 // same type of char pointer.
1439 if (ci->getOperand(2)->getType() !=
1440 fwrite_func->getFunctionType()->getParamType(0))
1443 std::vector<Value*> args;
1444 args.push_back(ci->getOperand(2));
1445 args.push_back(ConstantUInt::get(SLC.getIntPtrType(),len));
1446 args.push_back(ConstantUInt::get(SLC.getIntPtrType(),1));
1447 args.push_back(ci->getOperand(1));
1448 new CallInst(fwrite_func,args,ci->getName(),ci);
1449 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy,len));
1450 ci->eraseFromParent();
1454 // The remaining optimizations require the format string to be length 2
1459 // The first character has to be a %
1460 if (ConstantInt* CI = dyn_cast<ConstantInt>(CA->getOperand(0)))
1461 if (CI->getRawValue() != '%')
1464 // Get the second character and switch on its value
1465 ConstantInt* CI = dyn_cast<ConstantInt>(CA->getOperand(1));
1466 switch (CI->getRawValue())
1471 ConstantArray* CA = 0;
1472 if (!getConstantStringLength(ci->getOperand(3), len, &CA))
1475 // fprintf(file,"%s",str) -> fwrite(fmt,strlen(fmt),1,file)
1476 const Type* FILEptr_type = ci->getOperand(1)->getType();
1477 Function* fwrite_func = SLC.get_fwrite(FILEptr_type);
1480 std::vector<Value*> args;
1481 args.push_back(CastToCStr(ci->getOperand(3), *ci));
1482 args.push_back(ConstantUInt::get(SLC.getIntPtrType(),len));
1483 args.push_back(ConstantUInt::get(SLC.getIntPtrType(),1));
1484 args.push_back(ci->getOperand(1));
1485 new CallInst(fwrite_func,args,ci->getName(),ci);
1486 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy,len));
1491 ConstantInt* CI = dyn_cast<ConstantInt>(ci->getOperand(3));
1495 const Type* FILEptr_type = ci->getOperand(1)->getType();
1496 Function* fputc_func = SLC.get_fputc(FILEptr_type);
1499 CastInst* cast = new CastInst(CI,Type::IntTy,CI->getName()+".int",ci);
1500 new CallInst(fputc_func,cast,ci->getOperand(1),"",ci);
1501 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy,1));
1507 ci->eraseFromParent();
1512 /// This LibCallOptimization will simplify calls to the "sprintf" library
1513 /// function. It looks for cases where the result of sprintf is not used and the
1514 /// operation can be reduced to something simpler.
1515 /// @brief Simplify the pow library function.
1516 struct SPrintFOptimization : public LibCallOptimization
1519 /// @brief Default Constructor
1520 SPrintFOptimization() : LibCallOptimization("sprintf",
1521 "Number of 'sprintf' calls simplified") {}
1523 /// @brief Destructor
1524 virtual ~SPrintFOptimization() {}
1526 /// @brief Make sure that the "fprintf" function has the right prototype
1527 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC)
1529 // Just make sure this has at least 2 arguments
1530 return (f->getReturnType() == Type::IntTy && f->arg_size() >= 2);
1533 /// @brief Perform the sprintf optimization.
1534 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC)
1536 // If the call has more than 3 operands, we can't optimize it
1537 if (ci->getNumOperands() > 4 || ci->getNumOperands() < 3)
1540 // All the optimizations depend on the length of the second argument and the
1541 // fact that it is a constant string array. Check that now
1543 ConstantArray* CA = 0;
1544 if (!getConstantStringLength(ci->getOperand(2), len, &CA))
1547 if (ci->getNumOperands() == 3)
1551 // If the length is 0, we just need to store a null byte
1552 new StoreInst(ConstantInt::get(Type::SByteTy,0),ci->getOperand(1),ci);
1553 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy,0));
1554 ci->eraseFromParent();
1558 // Make sure there's no % in the constant array
1559 for (unsigned i = 0; i < len; ++i)
1561 if (ConstantInt* CI = dyn_cast<ConstantInt>(CA->getOperand(i)))
1563 // Check for the null terminator
1564 if (CI->getRawValue() == '%')
1565 return false; // we found a %, can't optimize
1568 return false; // initializer is not constant int, can't optimize
1571 // Increment length because we want to copy the null byte too
1574 // sprintf(str,fmt) -> llvm.memcpy(str,fmt,strlen(fmt),1)
1575 Function* memcpy_func = SLC.get_memcpy();
1578 std::vector<Value*> args;
1579 args.push_back(ci->getOperand(1));
1580 args.push_back(ci->getOperand(2));
1581 args.push_back(ConstantUInt::get(Type::UIntTy,len));
1582 args.push_back(ConstantUInt::get(Type::UIntTy,1));
1583 new CallInst(memcpy_func,args,"",ci);
1584 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy,len));
1585 ci->eraseFromParent();
1589 // The remaining optimizations require the format string to be length 2
1594 // The first character has to be a %
1595 if (ConstantInt* CI = dyn_cast<ConstantInt>(CA->getOperand(0)))
1596 if (CI->getRawValue() != '%')
1599 // Get the second character and switch on its value
1600 ConstantInt* CI = dyn_cast<ConstantInt>(CA->getOperand(1));
1601 switch (CI->getRawValue()) {
1603 // sprintf(dest,"%s",str) -> llvm.memcpy(dest, str, strlen(str)+1, 1)
1604 Function* strlen_func = SLC.get_strlen();
1605 Function* memcpy_func = SLC.get_memcpy();
1606 if (!strlen_func || !memcpy_func)
1609 Value *Len = new CallInst(strlen_func, CastToCStr(ci->getOperand(3), *ci),
1610 ci->getOperand(3)->getName()+".len", ci);
1611 Value *Len1 = BinaryOperator::createAdd(Len,
1612 ConstantInt::get(Len->getType(), 1),
1613 Len->getName()+"1", ci);
1614 if (Len1->getType() != Type::UIntTy)
1615 Len1 = new CastInst(Len1, Type::UIntTy, Len1->getName(), ci);
1616 std::vector<Value*> args;
1617 args.push_back(CastToCStr(ci->getOperand(1), *ci));
1618 args.push_back(CastToCStr(ci->getOperand(3), *ci));
1619 args.push_back(Len1);
1620 args.push_back(ConstantUInt::get(Type::UIntTy,1));
1621 new CallInst(memcpy_func, args, "", ci);
1623 // The strlen result is the unincremented number of bytes in the string.
1624 if (!ci->use_empty()) {
1625 if (Len->getType() != ci->getType())
1626 Len = new CastInst(Len, ci->getType(), Len->getName(), ci);
1627 ci->replaceAllUsesWith(Len);
1629 ci->eraseFromParent();
1633 // sprintf(dest,"%c",chr) -> store chr, dest
1634 CastInst* cast = new CastInst(ci->getOperand(3),Type::SByteTy,"char",ci);
1635 new StoreInst(cast, ci->getOperand(1), ci);
1636 GetElementPtrInst* gep = new GetElementPtrInst(ci->getOperand(1),
1637 ConstantUInt::get(Type::UIntTy,1),ci->getOperand(1)->getName()+".end",
1639 new StoreInst(ConstantInt::get(Type::SByteTy,0),gep,ci);
1640 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy,1));
1641 ci->eraseFromParent();
1649 /// This LibCallOptimization will simplify calls to the "fputs" library
1650 /// function. It looks for cases where the result of fputs is not used and the
1651 /// operation can be reduced to something simpler.
1652 /// @brief Simplify the pow library function.
1653 struct PutsOptimization : public LibCallOptimization
1656 /// @brief Default Constructor
1657 PutsOptimization() : LibCallOptimization("fputs",
1658 "Number of 'fputs' calls simplified") {}
1660 /// @brief Destructor
1661 virtual ~PutsOptimization() {}
1663 /// @brief Make sure that the "fputs" function has the right prototype
1664 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC)
1666 // Just make sure this has 2 arguments
1667 return (f->arg_size() == 2);
1670 /// @brief Perform the fputs optimization.
1671 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC)
1673 // If the result is used, none of these optimizations work
1674 if (!ci->use_empty())
1677 // All the optimizations depend on the length of the first argument and the
1678 // fact that it is a constant string array. Check that now
1680 if (!getConstantStringLength(ci->getOperand(1), len))
1686 // fputs("",F) -> noop
1690 // fputs(s,F) -> fputc(s[0],F) (if s is constant and strlen(s) == 1)
1691 const Type* FILEptr_type = ci->getOperand(2)->getType();
1692 Function* fputc_func = SLC.get_fputc(FILEptr_type);
1695 LoadInst* loadi = new LoadInst(ci->getOperand(1),
1696 ci->getOperand(1)->getName()+".byte",ci);
1697 CastInst* casti = new CastInst(loadi,Type::IntTy,
1698 loadi->getName()+".int",ci);
1699 new CallInst(fputc_func,casti,ci->getOperand(2),"",ci);
1704 // fputs(s,F) -> fwrite(s,1,len,F) (if s is constant and strlen(s) > 1)
1705 const Type* FILEptr_type = ci->getOperand(2)->getType();
1706 Function* fwrite_func = SLC.get_fwrite(FILEptr_type);
1709 std::vector<Value*> parms;
1710 parms.push_back(ci->getOperand(1));
1711 parms.push_back(ConstantUInt::get(SLC.getIntPtrType(),len));
1712 parms.push_back(ConstantUInt::get(SLC.getIntPtrType(),1));
1713 parms.push_back(ci->getOperand(2));
1714 new CallInst(fwrite_func,parms,"",ci);
1718 ci->eraseFromParent();
1719 return true; // success
1723 /// This LibCallOptimization will simplify calls to the "isdigit" library
1724 /// function. It simply does range checks the parameter explicitly.
1725 /// @brief Simplify the isdigit library function.
1726 struct IsDigitOptimization : public LibCallOptimization
1729 /// @brief Default Constructor
1730 IsDigitOptimization() : LibCallOptimization("isdigit",
1731 "Number of 'isdigit' calls simplified") {}
1733 /// @brief Destructor
1734 virtual ~IsDigitOptimization() {}
1736 /// @brief Make sure that the "fputs" function has the right prototype
1737 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC)
1739 // Just make sure this has 1 argument
1740 return (f->arg_size() == 1);
1743 /// @brief Perform the toascii optimization.
1744 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC)
1746 if (ConstantInt* CI = dyn_cast<ConstantInt>(ci->getOperand(1)))
1748 // isdigit(c) -> 0 or 1, if 'c' is constant
1749 uint64_t val = CI->getRawValue();
1750 if (val >= '0' && val <='9')
1751 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy,1));
1753 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy,0));
1754 ci->eraseFromParent();
1758 // isdigit(c) -> (unsigned)c - '0' <= 9
1760 new CastInst(ci->getOperand(1),Type::UIntTy,
1761 ci->getOperand(1)->getName()+".uint",ci);
1762 BinaryOperator* sub_inst = BinaryOperator::createSub(cast,
1763 ConstantUInt::get(Type::UIntTy,0x30),
1764 ci->getOperand(1)->getName()+".sub",ci);
1765 SetCondInst* setcond_inst = new SetCondInst(Instruction::SetLE,sub_inst,
1766 ConstantUInt::get(Type::UIntTy,9),
1767 ci->getOperand(1)->getName()+".cmp",ci);
1769 new CastInst(setcond_inst,Type::IntTy,
1770 ci->getOperand(1)->getName()+".isdigit",ci);
1771 ci->replaceAllUsesWith(c2);
1772 ci->eraseFromParent();
1777 /// This LibCallOptimization will simplify calls to the "toascii" library
1778 /// function. It simply does the corresponding and operation to restrict the
1779 /// range of values to the ASCII character set (0-127).
1780 /// @brief Simplify the toascii library function.
1781 struct ToAsciiOptimization : public LibCallOptimization
1784 /// @brief Default Constructor
1785 ToAsciiOptimization() : LibCallOptimization("toascii",
1786 "Number of 'toascii' calls simplified") {}
1788 /// @brief Destructor
1789 virtual ~ToAsciiOptimization() {}
1791 /// @brief Make sure that the "fputs" function has the right prototype
1792 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC)
1794 // Just make sure this has 2 arguments
1795 return (f->arg_size() == 1);
1798 /// @brief Perform the toascii optimization.
1799 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC)
1801 // toascii(c) -> (c & 0x7f)
1802 Value* chr = ci->getOperand(1);
1803 BinaryOperator* and_inst = BinaryOperator::createAnd(chr,
1804 ConstantInt::get(chr->getType(),0x7F),ci->getName()+".toascii",ci);
1805 ci->replaceAllUsesWith(and_inst);
1806 ci->eraseFromParent();
1811 /// This LibCallOptimization will simplify calls to the "ffs" library
1812 /// calls which find the first set bit in an int, long, or long long. The
1813 /// optimization is to compute the result at compile time if the argument is
1815 /// @brief Simplify the ffs library function.
1816 struct FFSOptimization : public LibCallOptimization
1819 /// @brief Subclass Constructor
1820 FFSOptimization(const char* funcName, const char* description)
1821 : LibCallOptimization(funcName, description)
1825 /// @brief Default Constructor
1826 FFSOptimization() : LibCallOptimization("ffs",
1827 "Number of 'ffs' calls simplified") {}
1829 /// @brief Destructor
1830 virtual ~FFSOptimization() {}
1832 /// @brief Make sure that the "fputs" function has the right prototype
1833 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC)
1835 // Just make sure this has 2 arguments
1836 return (f->arg_size() == 1 && f->getReturnType() == Type::IntTy);
1839 /// @brief Perform the ffs optimization.
1840 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC)
1842 if (ConstantInt* CI = dyn_cast<ConstantInt>(ci->getOperand(1)))
1844 // ffs(cnst) -> bit#
1845 // ffsl(cnst) -> bit#
1846 // ffsll(cnst) -> bit#
1847 uint64_t val = CI->getRawValue();
1855 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy, result));
1856 ci->eraseFromParent();
1860 // ffs(x) -> ( x == 0 ? 0 : llvm.cttz(x)+1)
1861 // ffsl(x) -> ( x == 0 ? 0 : llvm.cttz(x)+1)
1862 // ffsll(x) -> ( x == 0 ? 0 : llvm.cttz(x)+1)
1863 const Type* arg_type = ci->getOperand(1)->getType();
1864 std::vector<const Type*> args;
1865 args.push_back(arg_type);
1866 FunctionType* llvm_cttz_type = FunctionType::get(arg_type,args,false);
1868 SLC.getModule()->getOrInsertFunction("llvm.cttz",llvm_cttz_type);
1869 std::string inst_name(ci->getName()+".ffs");
1871 new CallInst(F, ci->getOperand(1), inst_name, ci);
1872 if (arg_type != Type::IntTy)
1873 call = new CastInst(call, Type::IntTy, inst_name, ci);
1874 BinaryOperator* add = BinaryOperator::createAdd(call,
1875 ConstantSInt::get(Type::IntTy,1), inst_name, ci);
1876 SetCondInst* eq = new SetCondInst(Instruction::SetEQ,ci->getOperand(1),
1877 ConstantSInt::get(ci->getOperand(1)->getType(),0),inst_name,ci);
1878 SelectInst* select = new SelectInst(eq,ConstantSInt::get(Type::IntTy,0),add,
1880 ci->replaceAllUsesWith(select);
1881 ci->eraseFromParent();
1886 /// This LibCallOptimization will simplify calls to the "ffsl" library
1887 /// calls. It simply uses FFSOptimization for which the transformation is
1889 /// @brief Simplify the ffsl library function.
1890 struct FFSLOptimization : public FFSOptimization
1893 /// @brief Default Constructor
1894 FFSLOptimization() : FFSOptimization("ffsl",
1895 "Number of 'ffsl' calls simplified") {}
1899 /// This LibCallOptimization will simplify calls to the "ffsll" library
1900 /// calls. It simply uses FFSOptimization for which the transformation is
1902 /// @brief Simplify the ffsl library function.
1903 struct FFSLLOptimization : public FFSOptimization
1906 /// @brief Default Constructor
1907 FFSLLOptimization() : FFSOptimization("ffsll",
1908 "Number of 'ffsll' calls simplified") {}
1913 /// This LibCallOptimization will simplify calls to the "floor" library
1915 /// @brief Simplify the floor library function.
1916 struct FloorOptimization : public LibCallOptimization {
1918 : LibCallOptimization("floor", "Number of 'floor' calls simplified") {}
1920 /// @brief Make sure that the "floor" function has the right prototype
1921 virtual bool ValidateCalledFunction(const Function *F, SimplifyLibCalls &SLC){
1922 return F->arg_size() == 1 && F->arg_begin()->getType() == Type::DoubleTy &&
1923 F->getReturnType() == Type::DoubleTy;
1926 virtual bool OptimizeCall(CallInst *CI, SimplifyLibCalls &SLC) {
1927 // If this is a float argument passed in, convert to floorf.
1928 // e.g. floor((double)FLT) -> (double)floorf(FLT). There can be no loss of
1929 // precision due to this.
1930 if (CastInst *Cast = dyn_cast<CastInst>(CI->getOperand(1)))
1931 if (Cast->getOperand(0)->getType() == Type::FloatTy) {
1932 Value *New = new CallInst(SLC.get_floorf(), Cast->getOperand(0),
1934 New = new CastInst(New, Type::DoubleTy, CI->getName(), CI);
1935 CI->replaceAllUsesWith(New);
1936 CI->eraseFromParent();
1937 if (Cast->use_empty())
1938 Cast->eraseFromParent();
1941 return false; // opt failed
1947 /// A function to compute the length of a null-terminated constant array of
1948 /// integers. This function can't rely on the size of the constant array
1949 /// because there could be a null terminator in the middle of the array.
1950 /// We also have to bail out if we find a non-integer constant initializer
1951 /// of one of the elements or if there is no null-terminator. The logic
1952 /// below checks each of these conditions and will return true only if all
1953 /// conditions are met. In that case, the \p len parameter is set to the length
1954 /// of the null-terminated string. If false is returned, the conditions were
1955 /// not met and len is set to 0.
1956 /// @brief Get the length of a constant string (null-terminated array).
1957 bool getConstantStringLength(Value* V, uint64_t& len, ConstantArray** CA )
1959 assert(V != 0 && "Invalid args to getConstantStringLength");
1960 len = 0; // make sure we initialize this
1962 // If the value is not a GEP instruction nor a constant expression with a
1963 // GEP instruction, then return false because ConstantArray can't occur
1965 if (GetElementPtrInst* GEPI = dyn_cast<GetElementPtrInst>(V))
1967 else if (ConstantExpr* CE = dyn_cast<ConstantExpr>(V))
1968 if (CE->getOpcode() == Instruction::GetElementPtr)
1975 // Make sure the GEP has exactly three arguments.
1976 if (GEP->getNumOperands() != 3)
1979 // Check to make sure that the first operand of the GEP is an integer and
1980 // has value 0 so that we are sure we're indexing into the initializer.
1981 if (ConstantInt* op1 = dyn_cast<ConstantInt>(GEP->getOperand(1)))
1983 if (!op1->isNullValue())
1989 // Ensure that the second operand is a ConstantInt. If it isn't then this
1990 // GEP is wonky and we're not really sure what were referencing into and
1991 // better of not optimizing it. While we're at it, get the second index
1992 // value. We'll need this later for indexing the ConstantArray.
1993 uint64_t start_idx = 0;
1994 if (ConstantInt* CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
1995 start_idx = CI->getRawValue();
1999 // The GEP instruction, constant or instruction, must reference a global
2000 // variable that is a constant and is initialized. The referenced constant
2001 // initializer is the array that we'll use for optimization.
2002 GlobalVariable* GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
2003 if (!GV || !GV->isConstant() || !GV->hasInitializer())
2006 // Get the initializer.
2007 Constant* INTLZR = GV->getInitializer();
2009 // Handle the ConstantAggregateZero case
2010 if (ConstantAggregateZero* CAZ = dyn_cast<ConstantAggregateZero>(INTLZR))
2012 // This is a degenerate case. The initializer is constant zero so the
2013 // length of the string must be zero.
2018 // Must be a Constant Array
2019 ConstantArray* A = dyn_cast<ConstantArray>(INTLZR);
2023 // Get the number of elements in the array
2024 uint64_t max_elems = A->getType()->getNumElements();
2026 // Traverse the constant array from start_idx (derived above) which is
2027 // the place the GEP refers to in the array.
2028 for ( len = start_idx; len < max_elems; len++)
2030 if (ConstantInt* CI = dyn_cast<ConstantInt>(A->getOperand(len)))
2032 // Check for the null terminator
2033 if (CI->isNullValue())
2034 break; // we found end of string
2037 return false; // This array isn't suitable, non-int initializer
2039 if (len >= max_elems)
2040 return false; // This array isn't null terminated
2042 // Subtract out the initial value from the length
2046 return true; // success!
2049 /// CastToCStr - Return V if it is an sbyte*, otherwise cast it to sbyte*,
2050 /// inserting the cast before IP, and return the cast.
2051 /// @brief Cast a value to a "C" string.
2052 Value *CastToCStr(Value *V, Instruction &IP) {
2053 const Type *SBPTy = PointerType::get(Type::SByteTy);
2054 if (V->getType() != SBPTy)
2055 return new CastInst(V, SBPTy, V->getName(), &IP);
2060 // Additional cases that we need to add to this file:
2063 // * cbrt(expN(X)) -> expN(x/3)
2064 // * cbrt(sqrt(x)) -> pow(x,1/6)
2065 // * cbrt(sqrt(x)) -> pow(x,1/9)
2068 // * cos(-x) -> cos(x)
2071 // * exp(log(x)) -> x
2074 // * isascii(c) -> ((c & ~0x7f) == 0)
2077 // * isdigit(c) -> (unsigned)(c) - '0' <= 9
2080 // * log(exp(x)) -> x
2081 // * log(x**y) -> y*log(x)
2082 // * log(exp(y)) -> y*log(e)
2083 // * log(exp2(y)) -> y*log(2)
2084 // * log(exp10(y)) -> y*log(10)
2085 // * log(sqrt(x)) -> 0.5*log(x)
2086 // * log(pow(x,y)) -> y*log(x)
2088 // lround, lroundf, lroundl:
2089 // * lround(cnst) -> cnst'
2092 // * memcmp(x,y,l) -> cnst
2093 // (if all arguments are constant and strlen(x) <= l and strlen(y) <= l)
2096 // * memmove(d,s,l,a) -> memcpy(d,s,l,a)
2097 // (if s is a global constant array)
2100 // * pow(exp(x),y) -> exp(x*y)
2101 // * pow(sqrt(x),y) -> pow(x,y*0.5)
2102 // * pow(pow(x,y),z)-> pow(x,y*z)
2105 // * puts("") -> fputc("\n",stdout) (how do we get "stdout"?)
2107 // round, roundf, roundl:
2108 // * round(cnst) -> cnst'
2111 // * signbit(cnst) -> cnst'
2112 // * signbit(nncst) -> 0 (if pstv is a non-negative constant)
2114 // sqrt, sqrtf, sqrtl:
2115 // * sqrt(expN(x)) -> expN(x*0.5)
2116 // * sqrt(Nroot(x)) -> pow(x,1/(2*N))
2117 // * sqrt(pow(x,y)) -> pow(|x|,y*0.5)
2120 // * stpcpy(str, "literal") ->
2121 // llvm.memcpy(str,"literal",strlen("literal")+1,1)
2123 // * strrchr(s,c) -> reverse_offset_of_in(c,s)
2124 // (if c is a constant integer and s is a constant string)
2125 // * strrchr(s1,0) -> strchr(s1,0)
2128 // * strncat(x,y,0) -> x
2129 // * strncat(x,y,0) -> x (if strlen(y) = 0)
2130 // * strncat(x,y,l) -> strcat(x,y) (if y and l are constants an l > strlen(y))
2133 // * strncpy(d,s,0) -> d
2134 // * strncpy(d,s,l) -> memcpy(d,s,l,1)
2135 // (if s and l are constants)
2138 // * strpbrk(s,a) -> offset_in_for(s,a)
2139 // (if s and a are both constant strings)
2140 // * strpbrk(s,"") -> 0
2141 // * strpbrk(s,a) -> strchr(s,a[0]) (if a is constant string of length 1)
2144 // * strspn(s,a) -> const_int (if both args are constant)
2145 // * strspn("",a) -> 0
2146 // * strspn(s,"") -> 0
2147 // * strcspn(s,a) -> const_int (if both args are constant)
2148 // * strcspn("",a) -> 0
2149 // * strcspn(s,"") -> strlen(a)
2152 // * strstr(x,x) -> x
2153 // * strstr(s1,s2) -> offset_of_s2_in(s1)
2154 // (if s1 and s2 are constant strings)
2157 // * tan(atan(x)) -> x
2159 // trunc, truncf, truncl:
2160 // * trunc(cnst) -> cnst'