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") {}
387 // Make sure the called function looks like exit (int argument, int return
388 // type, external linkage, not varargs).
389 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC)
391 if (f->arg_size() >= 1)
392 if (f->arg_begin()->getType()->isInteger())
397 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC)
399 // To be careful, we check that the call to exit is coming from "main", that
400 // main has external linkage, and the return type of main and the argument
401 // to exit have the same type.
402 Function *from = ci->getParent()->getParent();
403 if (from->hasExternalLinkage())
404 if (from->getReturnType() == ci->getOperand(1)->getType())
405 if (from->getName() == "main")
407 // Okay, time to actually do the optimization. First, get the basic
408 // block of the call instruction
409 BasicBlock* bb = ci->getParent();
411 // Create a return instruction that we'll replace the call with.
412 // Note that the argument of the return is the argument of the call
414 ReturnInst* ri = new ReturnInst(ci->getOperand(1), ci);
416 // Split the block at the call instruction which places it in a new
418 bb->splitBasicBlock(ci);
420 // The block split caused a branch instruction to be inserted into
421 // the end of the original block, right after the return instruction
422 // that we put there. That's not a valid block, so delete the branch
424 bb->getInstList().pop_back();
426 // Now we can finally get rid of the call instruction which now lives
427 // in the new basic block.
428 ci->eraseFromParent();
430 // Optimization succeeded, return true.
433 // We didn't pass the criteria for this optimization so return false
436 } ExitInMainOptimizer;
438 /// This LibCallOptimization will simplify a call to the strcat library
439 /// function. The simplification is possible only if the string being
440 /// concatenated is a constant array or a constant expression that results in
441 /// a constant string. In this case we can replace it with strlen + llvm.memcpy
442 /// of the constant string. Both of these calls are further reduced, if possible
443 /// on subsequent passes.
444 /// @brief Simplify the strcat library function.
445 struct StrCatOptimization : public LibCallOptimization
448 /// @brief Default constructor
449 StrCatOptimization() : LibCallOptimization("strcat",
450 "Number of 'strcat' calls simplified") {}
454 /// @brief Make sure that the "strcat" function has the right prototype
455 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC)
457 if (f->getReturnType() == PointerType::get(Type::SByteTy))
458 if (f->arg_size() == 2)
460 Function::const_arg_iterator AI = f->arg_begin();
461 if (AI++->getType() == PointerType::get(Type::SByteTy))
462 if (AI->getType() == PointerType::get(Type::SByteTy))
464 // Indicate this is a suitable call type.
471 /// @brief Optimize the strcat library function
472 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC)
474 // Extract some information from the instruction
475 Module* M = ci->getParent()->getParent()->getParent();
476 Value* dest = ci->getOperand(1);
477 Value* src = ci->getOperand(2);
479 // Extract the initializer (while making numerous checks) from the
480 // source operand of the call to strcat. If we get null back, one of
481 // a variety of checks in get_GVInitializer failed
483 if (!getConstantStringLength(src,len))
486 // Handle the simple, do-nothing case
489 ci->replaceAllUsesWith(dest);
490 ci->eraseFromParent();
494 // Increment the length because we actually want to memcpy the null
495 // terminator as well.
498 // We need to find the end of the destination string. That's where the
499 // memory is to be moved to. We just generate a call to strlen (further
500 // optimized in another pass). Note that the SLC.get_strlen() call
501 // caches the Function* for us.
502 CallInst* strlen_inst =
503 new CallInst(SLC.get_strlen(), dest, dest->getName()+".len",ci);
505 // Now that we have the destination's length, we must index into the
506 // destination's pointer to get the actual memcpy destination (end of
507 // the string .. we're concatenating).
508 std::vector<Value*> idx;
509 idx.push_back(strlen_inst);
510 GetElementPtrInst* gep =
511 new GetElementPtrInst(dest,idx,dest->getName()+".indexed",ci);
513 // We have enough information to now generate the memcpy call to
514 // do the concatenation for us.
515 std::vector<Value*> vals;
516 vals.push_back(gep); // destination
517 vals.push_back(ci->getOperand(2)); // source
518 vals.push_back(ConstantUInt::get(Type::UIntTy,len)); // length
519 vals.push_back(ConstantUInt::get(Type::UIntTy,1)); // alignment
520 new CallInst(SLC.get_memcpy(), vals, "", ci);
522 // Finally, substitute the first operand of the strcat call for the
523 // strcat call itself since strcat returns its first operand; and,
524 // kill the strcat CallInst.
525 ci->replaceAllUsesWith(dest);
526 ci->eraseFromParent();
531 /// This LibCallOptimization will simplify a call to the strchr library
532 /// function. It optimizes out cases where the arguments are both constant
533 /// and the result can be determined statically.
534 /// @brief Simplify the strcmp library function.
535 struct StrChrOptimization : public LibCallOptimization
538 StrChrOptimization() : LibCallOptimization("strchr",
539 "Number of 'strchr' calls simplified") {}
541 /// @brief Make sure that the "strchr" function has the right prototype
542 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC)
544 if (f->getReturnType() == PointerType::get(Type::SByteTy) &&
550 /// @brief Perform the strchr optimizations
551 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC)
553 // If there aren't three operands, bail
554 if (ci->getNumOperands() != 3)
557 // Check that the first argument to strchr is a constant array of sbyte.
558 // If it is, get the length and data, otherwise return false.
561 if (!getConstantStringLength(ci->getOperand(1),len,&CA))
564 // Check that the second argument to strchr is a constant int, return false
566 ConstantSInt* CSI = dyn_cast<ConstantSInt>(ci->getOperand(2));
569 // Just lower this to memchr since we know the length of the string as
571 Function* f = SLC.get_memchr();
572 std::vector<Value*> args;
573 args.push_back(ci->getOperand(1));
574 args.push_back(ci->getOperand(2));
575 args.push_back(ConstantUInt::get(SLC.getIntPtrType(),len));
576 ci->replaceAllUsesWith( new CallInst(f,args,ci->getName(),ci));
577 ci->eraseFromParent();
581 // Get the character we're looking for
582 int64_t chr = CSI->getValue();
584 // Compute the offset
586 bool char_found = false;
587 for (uint64_t i = 0; i < len; ++i)
589 if (ConstantSInt* CI = dyn_cast<ConstantSInt>(CA->getOperand(i)))
591 // Check for the null terminator
592 if (CI->isNullValue())
593 break; // we found end of string
594 else if (CI->getValue() == chr)
603 // strchr(s,c) -> offset_of_in(c,s)
604 // (if c is a constant integer and s is a constant string)
607 std::vector<Value*> indices;
608 indices.push_back(ConstantUInt::get(Type::ULongTy,offset));
609 GetElementPtrInst* GEP = new GetElementPtrInst(ci->getOperand(1),indices,
610 ci->getOperand(1)->getName()+".strchr",ci);
611 ci->replaceAllUsesWith(GEP);
614 ci->replaceAllUsesWith(
615 ConstantPointerNull::get(PointerType::get(Type::SByteTy)));
617 ci->eraseFromParent();
622 /// This LibCallOptimization will simplify a call to the strcmp library
623 /// function. It optimizes out cases where one or both arguments are constant
624 /// and the result can be determined statically.
625 /// @brief Simplify the strcmp library function.
626 struct StrCmpOptimization : public LibCallOptimization
629 StrCmpOptimization() : LibCallOptimization("strcmp",
630 "Number of 'strcmp' calls simplified") {}
632 /// @brief Make sure that the "strcmp" function has the right prototype
633 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC)
635 if (f->getReturnType() == Type::IntTy && f->arg_size() == 2)
640 /// @brief Perform the strcmp optimization
641 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC)
643 // First, check to see if src and destination are the same. If they are,
644 // then the optimization is to replace the CallInst with a constant 0
645 // because the call is a no-op.
646 Value* s1 = ci->getOperand(1);
647 Value* s2 = ci->getOperand(2);
651 ci->replaceAllUsesWith(ConstantInt::get(Type::IntTy,0));
652 ci->eraseFromParent();
656 bool isstr_1 = false;
659 if (getConstantStringLength(s1,len_1,&A1))
664 // strcmp("",x) -> *x
666 new LoadInst(CastToCStr(s2,*ci), ci->getName()+".load",ci);
668 new CastInst(load,Type::IntTy,ci->getName()+".int",ci);
669 ci->replaceAllUsesWith(cast);
670 ci->eraseFromParent();
675 bool isstr_2 = false;
678 if (getConstantStringLength(s2,len_2,&A2))
683 // strcmp(x,"") -> *x
685 new LoadInst(CastToCStr(s1,*ci),ci->getName()+".val",ci);
687 new CastInst(load,Type::IntTy,ci->getName()+".int",ci);
688 ci->replaceAllUsesWith(cast);
689 ci->eraseFromParent();
694 if (isstr_1 && isstr_2)
696 // strcmp(x,y) -> cnst (if both x and y are constant strings)
697 std::string str1 = A1->getAsString();
698 std::string str2 = A2->getAsString();
699 int result = strcmp(str1.c_str(), str2.c_str());
700 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy,result));
701 ci->eraseFromParent();
708 /// This LibCallOptimization will simplify a call to the strncmp library
709 /// function. It optimizes out cases where one or both arguments are constant
710 /// and the result can be determined statically.
711 /// @brief Simplify the strncmp library function.
712 struct StrNCmpOptimization : public LibCallOptimization
715 StrNCmpOptimization() : LibCallOptimization("strncmp",
716 "Number of 'strncmp' calls simplified") {}
718 /// @brief Make sure that the "strncmp" function has the right prototype
719 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC)
721 if (f->getReturnType() == Type::IntTy && f->arg_size() == 3)
726 /// @brief Perform the strncpy optimization
727 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC)
729 // First, check to see if src and destination are the same. If they are,
730 // then the optimization is to replace the CallInst with a constant 0
731 // because the call is a no-op.
732 Value* s1 = ci->getOperand(1);
733 Value* s2 = ci->getOperand(2);
736 // strncmp(x,x,l) -> 0
737 ci->replaceAllUsesWith(ConstantInt::get(Type::IntTy,0));
738 ci->eraseFromParent();
742 // Check the length argument, if it is Constant zero then the strings are
744 uint64_t len_arg = 0;
745 bool len_arg_is_const = false;
746 if (ConstantInt* len_CI = dyn_cast<ConstantInt>(ci->getOperand(3)))
748 len_arg_is_const = true;
749 len_arg = len_CI->getRawValue();
752 // strncmp(x,y,0) -> 0
753 ci->replaceAllUsesWith(ConstantInt::get(Type::IntTy,0));
754 ci->eraseFromParent();
759 bool isstr_1 = false;
762 if (getConstantStringLength(s1,len_1,&A1))
767 // strncmp("",x) -> *x
768 LoadInst* load = new LoadInst(s1,ci->getName()+".load",ci);
770 new CastInst(load,Type::IntTy,ci->getName()+".int",ci);
771 ci->replaceAllUsesWith(cast);
772 ci->eraseFromParent();
777 bool isstr_2 = false;
780 if (getConstantStringLength(s2,len_2,&A2))
785 // strncmp(x,"") -> *x
786 LoadInst* load = new LoadInst(s2,ci->getName()+".val",ci);
788 new CastInst(load,Type::IntTy,ci->getName()+".int",ci);
789 ci->replaceAllUsesWith(cast);
790 ci->eraseFromParent();
795 if (isstr_1 && isstr_2 && len_arg_is_const)
797 // strncmp(x,y,const) -> constant
798 std::string str1 = A1->getAsString();
799 std::string str2 = A2->getAsString();
800 int result = strncmp(str1.c_str(), str2.c_str(), len_arg);
801 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy,result));
802 ci->eraseFromParent();
809 /// This LibCallOptimization will simplify a call to the strcpy library
810 /// function. Two optimizations are possible:
811 /// (1) If src and dest are the same and not volatile, just return dest
812 /// (2) If the src is a constant then we can convert to llvm.memmove
813 /// @brief Simplify the strcpy library function.
814 struct StrCpyOptimization : public LibCallOptimization
817 StrCpyOptimization() : LibCallOptimization("strcpy",
818 "Number of 'strcpy' calls simplified") {}
820 /// @brief Make sure that the "strcpy" function has the right prototype
821 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC)
823 if (f->getReturnType() == PointerType::get(Type::SByteTy))
824 if (f->arg_size() == 2)
826 Function::const_arg_iterator AI = f->arg_begin();
827 if (AI++->getType() == PointerType::get(Type::SByteTy))
828 if (AI->getType() == PointerType::get(Type::SByteTy))
830 // Indicate this is a suitable call type.
837 /// @brief Perform the strcpy optimization
838 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC)
840 // First, check to see if src and destination are the same. If they are,
841 // then the optimization is to replace the CallInst with the destination
842 // because the call is a no-op. Note that this corresponds to the
843 // degenerate strcpy(X,X) case which should have "undefined" results
844 // according to the C specification. However, it occurs sometimes and
845 // we optimize it as a no-op.
846 Value* dest = ci->getOperand(1);
847 Value* src = ci->getOperand(2);
850 ci->replaceAllUsesWith(dest);
851 ci->eraseFromParent();
855 // Get the length of the constant string referenced by the second operand,
856 // the "src" parameter. Fail the optimization if we can't get the length
857 // (note that getConstantStringLength does lots of checks to make sure this
860 if (!getConstantStringLength(ci->getOperand(2),len))
863 // If the constant string's length is zero we can optimize this by just
864 // doing a store of 0 at the first byte of the destination
867 new StoreInst(ConstantInt::get(Type::SByteTy,0),ci->getOperand(1),ci);
868 ci->replaceAllUsesWith(dest);
869 ci->eraseFromParent();
873 // Increment the length because we actually want to memcpy the null
874 // terminator as well.
877 // Extract some information from the instruction
878 Module* M = ci->getParent()->getParent()->getParent();
880 // We have enough information to now generate the memcpy call to
881 // do the concatenation for us.
882 std::vector<Value*> vals;
883 vals.push_back(dest); // destination
884 vals.push_back(src); // source
885 vals.push_back(ConstantUInt::get(Type::UIntTy,len)); // length
886 vals.push_back(ConstantUInt::get(Type::UIntTy,1)); // alignment
887 new CallInst(SLC.get_memcpy(), vals, "", ci);
889 // Finally, substitute the first operand of the strcat call for the
890 // strcat call itself since strcat returns its first operand; and,
891 // kill the strcat CallInst.
892 ci->replaceAllUsesWith(dest);
893 ci->eraseFromParent();
898 /// This LibCallOptimization will simplify a call to the strlen library
899 /// function by replacing it with a constant value if the string provided to
900 /// it is a constant array.
901 /// @brief Simplify the strlen library function.
902 struct StrLenOptimization : public LibCallOptimization
904 StrLenOptimization() : LibCallOptimization("strlen",
905 "Number of 'strlen' calls simplified") {}
907 /// @brief Make sure that the "strlen" function has the right prototype
908 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC)
910 if (f->getReturnType() == SLC.getTargetData()->getIntPtrType())
911 if (f->arg_size() == 1)
912 if (Function::const_arg_iterator AI = f->arg_begin())
913 if (AI->getType() == PointerType::get(Type::SByteTy))
918 /// @brief Perform the strlen optimization
919 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC)
921 // Make sure we're dealing with an sbyte* here.
922 Value* str = ci->getOperand(1);
923 if (str->getType() != PointerType::get(Type::SByteTy))
926 // Does the call to strlen have exactly one use?
928 // Is that single use a binary operator?
929 if (BinaryOperator* bop = dyn_cast<BinaryOperator>(ci->use_back()))
930 // Is it compared against a constant integer?
931 if (ConstantInt* CI = dyn_cast<ConstantInt>(bop->getOperand(1)))
933 // Get the value the strlen result is compared to
934 uint64_t val = CI->getRawValue();
936 // If its compared against length 0 with == or !=
938 (bop->getOpcode() == Instruction::SetEQ ||
939 bop->getOpcode() == Instruction::SetNE))
941 // strlen(x) != 0 -> *x != 0
942 // strlen(x) == 0 -> *x == 0
943 LoadInst* load = new LoadInst(str,str->getName()+".first",ci);
944 BinaryOperator* rbop = BinaryOperator::create(bop->getOpcode(),
945 load, ConstantSInt::get(Type::SByteTy,0),
946 bop->getName()+".strlen", ci);
947 bop->replaceAllUsesWith(rbop);
948 bop->eraseFromParent();
949 ci->eraseFromParent();
954 // Get the length of the constant string operand
956 if (!getConstantStringLength(ci->getOperand(1),len))
959 // strlen("xyz") -> 3 (for example)
960 const Type *Ty = SLC.getTargetData()->getIntPtrType();
962 ci->replaceAllUsesWith(ConstantSInt::get(Ty, len));
964 ci->replaceAllUsesWith(ConstantUInt::get(Ty, len));
966 ci->eraseFromParent();
971 /// IsOnlyUsedInEqualsComparison - Return true if it only matters that the value
972 /// is equal or not-equal to zero.
973 static bool IsOnlyUsedInEqualsZeroComparison(Instruction *I) {
974 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
976 Instruction *User = cast<Instruction>(*UI);
977 if (User->getOpcode() == Instruction::SetNE ||
978 User->getOpcode() == Instruction::SetEQ) {
979 if (isa<Constant>(User->getOperand(1)) &&
980 cast<Constant>(User->getOperand(1))->isNullValue())
982 } else if (CastInst *CI = dyn_cast<CastInst>(User))
983 if (CI->getType() == Type::BoolTy)
985 // Unknown instruction.
991 /// This memcmpOptimization will simplify a call to the memcmp library
993 struct memcmpOptimization : public LibCallOptimization {
994 /// @brief Default Constructor
996 : LibCallOptimization("memcmp", "Number of 'memcmp' calls simplified") {}
998 /// @brief Make sure that the "memcmp" function has the right prototype
999 virtual bool ValidateCalledFunction(const Function *F, SimplifyLibCalls &TD) {
1000 Function::const_arg_iterator AI = F->arg_begin();
1001 if (F->arg_size() != 3 || !isa<PointerType>(AI->getType())) return false;
1002 if (!isa<PointerType>((++AI)->getType())) return false;
1003 if (!(++AI)->getType()->isInteger()) return false;
1004 if (!F->getReturnType()->isInteger()) return false;
1008 /// Because of alignment and instruction information that we don't have, we
1009 /// leave the bulk of this to the code generators.
1011 /// Note that we could do much more if we could force alignment on otherwise
1012 /// small aligned allocas, or if we could indicate that loads have a small
1014 virtual bool OptimizeCall(CallInst *CI, SimplifyLibCalls &TD) {
1015 Value *LHS = CI->getOperand(1), *RHS = CI->getOperand(2);
1017 // If the two operands are the same, return zero.
1019 // memcmp(s,s,x) -> 0
1020 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
1021 CI->eraseFromParent();
1025 // Make sure we have a constant length.
1026 ConstantInt *LenC = dyn_cast<ConstantInt>(CI->getOperand(3));
1027 if (!LenC) return false;
1028 uint64_t Len = LenC->getRawValue();
1030 // If the length is zero, this returns 0.
1033 // memcmp(s1,s2,0) -> 0
1034 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
1035 CI->eraseFromParent();
1038 // memcmp(S1,S2,1) -> *(ubyte*)S1 - *(ubyte*)S2
1039 const Type *UCharPtr = PointerType::get(Type::UByteTy);
1040 CastInst *Op1Cast = new CastInst(LHS, UCharPtr, LHS->getName(), CI);
1041 CastInst *Op2Cast = new CastInst(RHS, UCharPtr, RHS->getName(), CI);
1042 Value *S1V = new LoadInst(Op1Cast, LHS->getName()+".val", CI);
1043 Value *S2V = new LoadInst(Op2Cast, RHS->getName()+".val", CI);
1044 Value *RV = BinaryOperator::createSub(S1V, S2V, CI->getName()+".diff",CI);
1045 if (RV->getType() != CI->getType())
1046 RV = new CastInst(RV, CI->getType(), RV->getName(), CI);
1047 CI->replaceAllUsesWith(RV);
1048 CI->eraseFromParent();
1052 if (IsOnlyUsedInEqualsZeroComparison(CI)) {
1053 // TODO: IF both are aligned, use a short load/compare.
1055 // memcmp(S1,S2,2) -> S1[0]-S2[0] | S1[1]-S2[1] iff only ==/!= 0 matters
1056 const Type *UCharPtr = PointerType::get(Type::UByteTy);
1057 CastInst *Op1Cast = new CastInst(LHS, UCharPtr, LHS->getName(), CI);
1058 CastInst *Op2Cast = new CastInst(RHS, UCharPtr, RHS->getName(), CI);
1059 Value *S1V1 = new LoadInst(Op1Cast, LHS->getName()+".val1", CI);
1060 Value *S2V1 = new LoadInst(Op2Cast, RHS->getName()+".val1", CI);
1061 Value *D1 = BinaryOperator::createSub(S1V1, S2V1,
1062 CI->getName()+".d1", CI);
1063 Constant *One = ConstantInt::get(Type::IntTy, 1);
1064 Value *G1 = new GetElementPtrInst(Op1Cast, One, "next1v", CI);
1065 Value *G2 = new GetElementPtrInst(Op2Cast, One, "next2v", CI);
1066 Value *S1V2 = new LoadInst(G1, LHS->getName()+".val2", CI);
1067 Value *S2V2 = new LoadInst(G1, RHS->getName()+".val2", CI);
1068 Value *D2 = BinaryOperator::createSub(S1V2, S2V2,
1069 CI->getName()+".d1", CI);
1070 Value *Or = BinaryOperator::createOr(D1, D2, CI->getName()+".res", CI);
1071 if (Or->getType() != CI->getType())
1072 Or = new CastInst(Or, CI->getType(), Or->getName(), CI);
1073 CI->replaceAllUsesWith(Or);
1074 CI->eraseFromParent();
1092 /// This LibCallOptimization will simplify a call to the memcpy library
1093 /// function by expanding it out to a single store of size 0, 1, 2, 4, or 8
1094 /// bytes depending on the length of the string and the alignment. Additional
1095 /// optimizations are possible in code generation (sequence of immediate store)
1096 /// @brief Simplify the memcpy library function.
1097 struct LLVMMemCpyOptimization : public LibCallOptimization
1099 /// @brief Default Constructor
1100 LLVMMemCpyOptimization() : LibCallOptimization("llvm.memcpy",
1101 "Number of 'llvm.memcpy' calls simplified") {}
1104 /// @brief Subclass Constructor
1105 LLVMMemCpyOptimization(const char* fname, const char* desc)
1106 : LibCallOptimization(fname, desc) {}
1109 /// @brief Make sure that the "memcpy" function has the right prototype
1110 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& TD)
1112 // Just make sure this has 4 arguments per LLVM spec.
1113 return (f->arg_size() == 4);
1116 /// Because of alignment and instruction information that we don't have, we
1117 /// leave the bulk of this to the code generators. The optimization here just
1118 /// deals with a few degenerate cases where the length of the string and the
1119 /// alignment match the sizes of our intrinsic types so we can do a load and
1120 /// store instead of the memcpy call.
1121 /// @brief Perform the memcpy optimization.
1122 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& TD)
1124 // Make sure we have constant int values to work with
1125 ConstantInt* LEN = dyn_cast<ConstantInt>(ci->getOperand(3));
1128 ConstantInt* ALIGN = dyn_cast<ConstantInt>(ci->getOperand(4));
1132 // If the length is larger than the alignment, we can't optimize
1133 uint64_t len = LEN->getRawValue();
1134 uint64_t alignment = ALIGN->getRawValue();
1136 alignment = 1; // Alignment 0 is identity for alignment 1
1137 if (len > alignment)
1140 // Get the type we will cast to, based on size of the string
1141 Value* dest = ci->getOperand(1);
1142 Value* src = ci->getOperand(2);
1147 // memcpy(d,s,0,a) -> noop
1148 ci->eraseFromParent();
1150 case 1: castType = Type::SByteTy; break;
1151 case 2: castType = Type::ShortTy; break;
1152 case 4: castType = Type::IntTy; break;
1153 case 8: castType = Type::LongTy; break;
1158 // Cast source and dest to the right sized primitive and then load/store
1160 new CastInst(src,PointerType::get(castType),src->getName()+".cast",ci);
1161 CastInst* DestCast =
1162 new CastInst(dest,PointerType::get(castType),dest->getName()+".cast",ci);
1163 LoadInst* LI = new LoadInst(SrcCast,SrcCast->getName()+".val",ci);
1164 StoreInst* SI = new StoreInst(LI, DestCast, ci);
1165 ci->eraseFromParent();
1168 } LLVMMemCpyOptimizer;
1170 /// This LibCallOptimization will simplify a call to the memmove library
1171 /// function. It is identical to MemCopyOptimization except for the name of
1173 /// @brief Simplify the memmove library function.
1174 struct LLVMMemMoveOptimization : public LLVMMemCpyOptimization
1176 /// @brief Default Constructor
1177 LLVMMemMoveOptimization() : LLVMMemCpyOptimization("llvm.memmove",
1178 "Number of 'llvm.memmove' calls simplified") {}
1180 } LLVMMemMoveOptimizer;
1182 /// This LibCallOptimization will simplify a call to the memset library
1183 /// function by expanding it out to a single store of size 0, 1, 2, 4, or 8
1184 /// bytes depending on the length argument.
1185 struct LLVMMemSetOptimization : public LibCallOptimization
1187 /// @brief Default Constructor
1188 LLVMMemSetOptimization() : LibCallOptimization("llvm.memset",
1189 "Number of 'llvm.memset' calls simplified") {}
1193 /// @brief Make sure that the "memset" function has the right prototype
1194 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& TD)
1196 // Just make sure this has 3 arguments per LLVM spec.
1197 return (f->arg_size() == 4);
1200 /// Because of alignment and instruction information that we don't have, we
1201 /// leave the bulk of this to the code generators. The optimization here just
1202 /// deals with a few degenerate cases where the length parameter is constant
1203 /// and the alignment matches the sizes of our intrinsic types so we can do
1204 /// store instead of the memcpy call. Other calls are transformed into the
1205 /// llvm.memset intrinsic.
1206 /// @brief Perform the memset optimization.
1207 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& TD)
1209 // Make sure we have constant int values to work with
1210 ConstantInt* LEN = dyn_cast<ConstantInt>(ci->getOperand(3));
1213 ConstantInt* ALIGN = dyn_cast<ConstantInt>(ci->getOperand(4));
1217 // Extract the length and alignment
1218 uint64_t len = LEN->getRawValue();
1219 uint64_t alignment = ALIGN->getRawValue();
1221 // Alignment 0 is identity for alignment 1
1225 // If the length is zero, this is a no-op
1228 // memset(d,c,0,a) -> noop
1229 ci->eraseFromParent();
1233 // If the length is larger than the alignment, we can't optimize
1234 if (len > alignment)
1237 // Make sure we have a constant ubyte to work with so we can extract
1238 // the value to be filled.
1239 ConstantUInt* FILL = dyn_cast<ConstantUInt>(ci->getOperand(2));
1242 if (FILL->getType() != Type::UByteTy)
1245 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
1247 // Extract the fill character
1248 uint64_t fill_char = FILL->getValue();
1249 uint64_t fill_value = fill_char;
1251 // Get the type we will cast to, based on size of memory area to fill, and
1252 // and the value we will store there.
1253 Value* dest = ci->getOperand(1);
1258 castType = Type::UByteTy;
1261 castType = Type::UShortTy;
1262 fill_value |= fill_char << 8;
1265 castType = Type::UIntTy;
1266 fill_value |= fill_char << 8 | fill_char << 16 | fill_char << 24;
1269 castType = Type::ULongTy;
1270 fill_value |= fill_char << 8 | fill_char << 16 | fill_char << 24;
1271 fill_value |= fill_char << 32 | fill_char << 40 | fill_char << 48;
1272 fill_value |= fill_char << 56;
1278 // Cast dest to the right sized primitive and then load/store
1279 CastInst* DestCast =
1280 new CastInst(dest,PointerType::get(castType),dest->getName()+".cast",ci);
1281 new StoreInst(ConstantUInt::get(castType,fill_value),DestCast, ci);
1282 ci->eraseFromParent();
1285 } LLVMMemSetOptimizer;
1287 /// This LibCallOptimization will simplify calls to the "pow" library
1288 /// function. It looks for cases where the result of pow is well known and
1289 /// substitutes the appropriate value.
1290 /// @brief Simplify the pow library function.
1291 struct PowOptimization : public LibCallOptimization
1294 /// @brief Default Constructor
1295 PowOptimization() : LibCallOptimization("pow",
1296 "Number of 'pow' calls simplified") {}
1298 /// @brief Make sure that the "pow" function has the right prototype
1299 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC)
1301 // Just make sure this has 2 arguments
1302 return (f->arg_size() == 2);
1305 /// @brief Perform the pow optimization.
1306 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC)
1308 const Type *Ty = cast<Function>(ci->getOperand(0))->getReturnType();
1309 Value* base = ci->getOperand(1);
1310 Value* expn = ci->getOperand(2);
1311 if (ConstantFP *Op1 = dyn_cast<ConstantFP>(base)) {
1312 double Op1V = Op1->getValue();
1315 // pow(1.0,x) -> 1.0
1316 ci->replaceAllUsesWith(ConstantFP::get(Ty,1.0));
1317 ci->eraseFromParent();
1321 else if (ConstantFP* Op2 = dyn_cast<ConstantFP>(expn))
1323 double Op2V = Op2->getValue();
1326 // pow(x,0.0) -> 1.0
1327 ci->replaceAllUsesWith(ConstantFP::get(Ty,1.0));
1328 ci->eraseFromParent();
1331 else if (Op2V == 0.5)
1333 // pow(x,0.5) -> sqrt(x)
1334 CallInst* sqrt_inst = new CallInst(SLC.get_sqrt(), base,
1335 ci->getName()+".pow",ci);
1336 ci->replaceAllUsesWith(sqrt_inst);
1337 ci->eraseFromParent();
1340 else if (Op2V == 1.0)
1343 ci->replaceAllUsesWith(base);
1344 ci->eraseFromParent();
1347 else if (Op2V == -1.0)
1349 // pow(x,-1.0) -> 1.0/x
1350 BinaryOperator* div_inst= BinaryOperator::createDiv(
1351 ConstantFP::get(Ty,1.0), base, ci->getName()+".pow", ci);
1352 ci->replaceAllUsesWith(div_inst);
1353 ci->eraseFromParent();
1357 return false; // opt failed
1361 /// This LibCallOptimization will simplify calls to the "fprintf" library
1362 /// function. It looks for cases where the result of fprintf is not used and the
1363 /// operation can be reduced to something simpler.
1364 /// @brief Simplify the pow library function.
1365 struct FPrintFOptimization : public LibCallOptimization
1368 /// @brief Default Constructor
1369 FPrintFOptimization() : LibCallOptimization("fprintf",
1370 "Number of 'fprintf' calls simplified") {}
1372 /// @brief Make sure that the "fprintf" function has the right prototype
1373 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC)
1375 // Just make sure this has at least 2 arguments
1376 return (f->arg_size() >= 2);
1379 /// @brief Perform the fprintf optimization.
1380 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC)
1382 // If the call has more than 3 operands, we can't optimize it
1383 if (ci->getNumOperands() > 4 || ci->getNumOperands() <= 2)
1386 // If the result of the fprintf call is used, none of these optimizations
1388 if (!ci->use_empty())
1391 // All the optimizations depend on the length of the second argument and the
1392 // fact that it is a constant string array. Check that now
1394 ConstantArray* CA = 0;
1395 if (!getConstantStringLength(ci->getOperand(2), len, &CA))
1398 if (ci->getNumOperands() == 3)
1400 // Make sure there's no % in the constant array
1401 for (unsigned i = 0; i < len; ++i)
1403 if (ConstantInt* CI = dyn_cast<ConstantInt>(CA->getOperand(i)))
1405 // Check for the null terminator
1406 if (CI->getRawValue() == '%')
1407 return false; // we found end of string
1413 // fprintf(file,fmt) -> fwrite(fmt,strlen(fmt),file)
1414 const Type* FILEptr_type = ci->getOperand(1)->getType();
1415 Function* fwrite_func = SLC.get_fwrite(FILEptr_type);
1419 // Make sure that the fprintf() and fwrite() functions both take the
1420 // same type of char pointer.
1421 if (ci->getOperand(2)->getType() !=
1422 fwrite_func->getFunctionType()->getParamType(0))
1425 std::vector<Value*> args;
1426 args.push_back(ci->getOperand(2));
1427 args.push_back(ConstantUInt::get(SLC.getIntPtrType(),len));
1428 args.push_back(ConstantUInt::get(SLC.getIntPtrType(),1));
1429 args.push_back(ci->getOperand(1));
1430 new CallInst(fwrite_func,args,ci->getName(),ci);
1431 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy,len));
1432 ci->eraseFromParent();
1436 // The remaining optimizations require the format string to be length 2
1441 // The first character has to be a %
1442 if (ConstantInt* CI = dyn_cast<ConstantInt>(CA->getOperand(0)))
1443 if (CI->getRawValue() != '%')
1446 // Get the second character and switch on its value
1447 ConstantInt* CI = dyn_cast<ConstantInt>(CA->getOperand(1));
1448 switch (CI->getRawValue())
1453 ConstantArray* CA = 0;
1454 if (!getConstantStringLength(ci->getOperand(3), len, &CA))
1457 // fprintf(file,"%s",str) -> fwrite(fmt,strlen(fmt),1,file)
1458 const Type* FILEptr_type = ci->getOperand(1)->getType();
1459 Function* fwrite_func = SLC.get_fwrite(FILEptr_type);
1462 std::vector<Value*> args;
1463 args.push_back(CastToCStr(ci->getOperand(3), *ci));
1464 args.push_back(ConstantUInt::get(SLC.getIntPtrType(),len));
1465 args.push_back(ConstantUInt::get(SLC.getIntPtrType(),1));
1466 args.push_back(ci->getOperand(1));
1467 new CallInst(fwrite_func,args,ci->getName(),ci);
1468 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy,len));
1473 ConstantInt* CI = dyn_cast<ConstantInt>(ci->getOperand(3));
1477 const Type* FILEptr_type = ci->getOperand(1)->getType();
1478 Function* fputc_func = SLC.get_fputc(FILEptr_type);
1481 CastInst* cast = new CastInst(CI,Type::IntTy,CI->getName()+".int",ci);
1482 new CallInst(fputc_func,cast,ci->getOperand(1),"",ci);
1483 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy,1));
1489 ci->eraseFromParent();
1494 /// This LibCallOptimization will simplify calls to the "sprintf" library
1495 /// function. It looks for cases where the result of sprintf is not used and the
1496 /// operation can be reduced to something simpler.
1497 /// @brief Simplify the pow library function.
1498 struct SPrintFOptimization : public LibCallOptimization
1501 /// @brief Default Constructor
1502 SPrintFOptimization() : LibCallOptimization("sprintf",
1503 "Number of 'sprintf' calls simplified") {}
1505 /// @brief Make sure that the "fprintf" function has the right prototype
1506 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC)
1508 // Just make sure this has at least 2 arguments
1509 return (f->getReturnType() == Type::IntTy && f->arg_size() >= 2);
1512 /// @brief Perform the sprintf optimization.
1513 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC)
1515 // If the call has more than 3 operands, we can't optimize it
1516 if (ci->getNumOperands() > 4 || ci->getNumOperands() < 3)
1519 // All the optimizations depend on the length of the second argument and the
1520 // fact that it is a constant string array. Check that now
1522 ConstantArray* CA = 0;
1523 if (!getConstantStringLength(ci->getOperand(2), len, &CA))
1526 if (ci->getNumOperands() == 3)
1530 // If the length is 0, we just need to store a null byte
1531 new StoreInst(ConstantInt::get(Type::SByteTy,0),ci->getOperand(1),ci);
1532 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy,0));
1533 ci->eraseFromParent();
1537 // Make sure there's no % in the constant array
1538 for (unsigned i = 0; i < len; ++i)
1540 if (ConstantInt* CI = dyn_cast<ConstantInt>(CA->getOperand(i)))
1542 // Check for the null terminator
1543 if (CI->getRawValue() == '%')
1544 return false; // we found a %, can't optimize
1547 return false; // initializer is not constant int, can't optimize
1550 // Increment length because we want to copy the null byte too
1553 // sprintf(str,fmt) -> llvm.memcpy(str,fmt,strlen(fmt),1)
1554 Function* memcpy_func = SLC.get_memcpy();
1557 std::vector<Value*> args;
1558 args.push_back(ci->getOperand(1));
1559 args.push_back(ci->getOperand(2));
1560 args.push_back(ConstantUInt::get(Type::UIntTy,len));
1561 args.push_back(ConstantUInt::get(Type::UIntTy,1));
1562 new CallInst(memcpy_func,args,"",ci);
1563 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy,len));
1564 ci->eraseFromParent();
1568 // The remaining optimizations require the format string to be length 2
1573 // The first character has to be a %
1574 if (ConstantInt* CI = dyn_cast<ConstantInt>(CA->getOperand(0)))
1575 if (CI->getRawValue() != '%')
1578 // Get the second character and switch on its value
1579 ConstantInt* CI = dyn_cast<ConstantInt>(CA->getOperand(1));
1580 switch (CI->getRawValue()) {
1582 // sprintf(dest,"%s",str) -> llvm.memcpy(dest, str, strlen(str)+1, 1)
1583 Function* strlen_func = SLC.get_strlen();
1584 Function* memcpy_func = SLC.get_memcpy();
1585 if (!strlen_func || !memcpy_func)
1588 Value *Len = new CallInst(strlen_func, CastToCStr(ci->getOperand(3), *ci),
1589 ci->getOperand(3)->getName()+".len", ci);
1590 Value *Len1 = BinaryOperator::createAdd(Len,
1591 ConstantInt::get(Len->getType(), 1),
1592 Len->getName()+"1", ci);
1593 if (Len1->getType() != Type::UIntTy)
1594 Len1 = new CastInst(Len1, Type::UIntTy, Len1->getName(), ci);
1595 std::vector<Value*> args;
1596 args.push_back(CastToCStr(ci->getOperand(1), *ci));
1597 args.push_back(CastToCStr(ci->getOperand(3), *ci));
1598 args.push_back(Len1);
1599 args.push_back(ConstantUInt::get(Type::UIntTy,1));
1600 new CallInst(memcpy_func, args, "", ci);
1602 // The strlen result is the unincremented number of bytes in the string.
1603 if (!ci->use_empty()) {
1604 if (Len->getType() != ci->getType())
1605 Len = new CastInst(Len, ci->getType(), Len->getName(), ci);
1606 ci->replaceAllUsesWith(Len);
1608 ci->eraseFromParent();
1612 // sprintf(dest,"%c",chr) -> store chr, dest
1613 CastInst* cast = new CastInst(ci->getOperand(3),Type::SByteTy,"char",ci);
1614 new StoreInst(cast, ci->getOperand(1), ci);
1615 GetElementPtrInst* gep = new GetElementPtrInst(ci->getOperand(1),
1616 ConstantUInt::get(Type::UIntTy,1),ci->getOperand(1)->getName()+".end",
1618 new StoreInst(ConstantInt::get(Type::SByteTy,0),gep,ci);
1619 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy,1));
1620 ci->eraseFromParent();
1628 /// This LibCallOptimization will simplify calls to the "fputs" library
1629 /// function. It looks for cases where the result of fputs is not used and the
1630 /// operation can be reduced to something simpler.
1631 /// @brief Simplify the pow library function.
1632 struct PutsOptimization : public LibCallOptimization
1635 /// @brief Default Constructor
1636 PutsOptimization() : LibCallOptimization("fputs",
1637 "Number of 'fputs' calls simplified") {}
1639 /// @brief Make sure that the "fputs" function has the right prototype
1640 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC)
1642 // Just make sure this has 2 arguments
1643 return (f->arg_size() == 2);
1646 /// @brief Perform the fputs optimization.
1647 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC)
1649 // If the result is used, none of these optimizations work
1650 if (!ci->use_empty())
1653 // All the optimizations depend on the length of the first argument and the
1654 // fact that it is a constant string array. Check that now
1656 if (!getConstantStringLength(ci->getOperand(1), len))
1662 // fputs("",F) -> noop
1666 // fputs(s,F) -> fputc(s[0],F) (if s is constant and strlen(s) == 1)
1667 const Type* FILEptr_type = ci->getOperand(2)->getType();
1668 Function* fputc_func = SLC.get_fputc(FILEptr_type);
1671 LoadInst* loadi = new LoadInst(ci->getOperand(1),
1672 ci->getOperand(1)->getName()+".byte",ci);
1673 CastInst* casti = new CastInst(loadi,Type::IntTy,
1674 loadi->getName()+".int",ci);
1675 new CallInst(fputc_func,casti,ci->getOperand(2),"",ci);
1680 // fputs(s,F) -> fwrite(s,1,len,F) (if s is constant and strlen(s) > 1)
1681 const Type* FILEptr_type = ci->getOperand(2)->getType();
1682 Function* fwrite_func = SLC.get_fwrite(FILEptr_type);
1685 std::vector<Value*> parms;
1686 parms.push_back(ci->getOperand(1));
1687 parms.push_back(ConstantUInt::get(SLC.getIntPtrType(),len));
1688 parms.push_back(ConstantUInt::get(SLC.getIntPtrType(),1));
1689 parms.push_back(ci->getOperand(2));
1690 new CallInst(fwrite_func,parms,"",ci);
1694 ci->eraseFromParent();
1695 return true; // success
1699 /// This LibCallOptimization will simplify calls to the "isdigit" library
1700 /// function. It simply does range checks the parameter explicitly.
1701 /// @brief Simplify the isdigit library function.
1702 struct isdigitOptimization : public LibCallOptimization {
1704 isdigitOptimization() : LibCallOptimization("isdigit",
1705 "Number of 'isdigit' calls simplified") {}
1707 /// @brief Make sure that the "isdigit" function has the right prototype
1708 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC)
1710 // Just make sure this has 1 argument
1711 return (f->arg_size() == 1);
1714 /// @brief Perform the toascii optimization.
1715 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC)
1717 if (ConstantInt* CI = dyn_cast<ConstantInt>(ci->getOperand(1)))
1719 // isdigit(c) -> 0 or 1, if 'c' is constant
1720 uint64_t val = CI->getRawValue();
1721 if (val >= '0' && val <='9')
1722 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy,1));
1724 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy,0));
1725 ci->eraseFromParent();
1729 // isdigit(c) -> (unsigned)c - '0' <= 9
1731 new CastInst(ci->getOperand(1),Type::UIntTy,
1732 ci->getOperand(1)->getName()+".uint",ci);
1733 BinaryOperator* sub_inst = BinaryOperator::createSub(cast,
1734 ConstantUInt::get(Type::UIntTy,0x30),
1735 ci->getOperand(1)->getName()+".sub",ci);
1736 SetCondInst* setcond_inst = new SetCondInst(Instruction::SetLE,sub_inst,
1737 ConstantUInt::get(Type::UIntTy,9),
1738 ci->getOperand(1)->getName()+".cmp",ci);
1740 new CastInst(setcond_inst,Type::IntTy,
1741 ci->getOperand(1)->getName()+".isdigit",ci);
1742 ci->replaceAllUsesWith(c2);
1743 ci->eraseFromParent();
1748 struct isasciiOptimization : public LibCallOptimization {
1750 isasciiOptimization()
1751 : LibCallOptimization("isascii", "Number of 'isascii' calls simplified") {}
1753 virtual bool ValidateCalledFunction(const Function *F, SimplifyLibCalls &SLC){
1754 return F->arg_size() == 1 && F->arg_begin()->getType()->isInteger() &&
1755 F->getReturnType()->isInteger();
1758 /// @brief Perform the isascii optimization.
1759 virtual bool OptimizeCall(CallInst *CI, SimplifyLibCalls &SLC) {
1760 // isascii(c) -> (unsigned)c < 128
1761 Value *V = CI->getOperand(1);
1762 if (V->getType()->isSigned())
1763 V = new CastInst(V, V->getType()->getUnsignedVersion(), V->getName(), CI);
1764 Value *Cmp = BinaryOperator::createSetLT(V, ConstantUInt::get(V->getType(),
1766 V->getName()+".isascii", CI);
1767 if (Cmp->getType() != CI->getType())
1768 Cmp = new CastInst(Cmp, CI->getType(), Cmp->getName(), CI);
1769 CI->replaceAllUsesWith(Cmp);
1770 CI->eraseFromParent();
1776 /// This LibCallOptimization will simplify calls to the "toascii" library
1777 /// function. It simply does the corresponding and operation to restrict the
1778 /// range of values to the ASCII character set (0-127).
1779 /// @brief Simplify the toascii library function.
1780 struct ToAsciiOptimization : public LibCallOptimization
1783 /// @brief Default Constructor
1784 ToAsciiOptimization() : LibCallOptimization("toascii",
1785 "Number of 'toascii' calls simplified") {}
1787 /// @brief Make sure that the "fputs" function has the right prototype
1788 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC)
1790 // Just make sure this has 2 arguments
1791 return (f->arg_size() == 1);
1794 /// @brief Perform the toascii optimization.
1795 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC)
1797 // toascii(c) -> (c & 0x7f)
1798 Value* chr = ci->getOperand(1);
1799 BinaryOperator* and_inst = BinaryOperator::createAnd(chr,
1800 ConstantInt::get(chr->getType(),0x7F),ci->getName()+".toascii",ci);
1801 ci->replaceAllUsesWith(and_inst);
1802 ci->eraseFromParent();
1807 /// This LibCallOptimization will simplify calls to the "ffs" library
1808 /// calls which find the first set bit in an int, long, or long long. The
1809 /// optimization is to compute the result at compile time if the argument is
1811 /// @brief Simplify the ffs library function.
1812 struct FFSOptimization : public LibCallOptimization
1815 /// @brief Subclass Constructor
1816 FFSOptimization(const char* funcName, const char* description)
1817 : LibCallOptimization(funcName, description)
1821 /// @brief Default Constructor
1822 FFSOptimization() : LibCallOptimization("ffs",
1823 "Number of 'ffs' calls simplified") {}
1825 /// @brief Make sure that the "fputs" function has the right prototype
1826 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC)
1828 // Just make sure this has 2 arguments
1829 return (f->arg_size() == 1 && f->getReturnType() == Type::IntTy);
1832 /// @brief Perform the ffs optimization.
1833 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC)
1835 if (ConstantInt* CI = dyn_cast<ConstantInt>(ci->getOperand(1)))
1837 // ffs(cnst) -> bit#
1838 // ffsl(cnst) -> bit#
1839 // ffsll(cnst) -> bit#
1840 uint64_t val = CI->getRawValue();
1848 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy, result));
1849 ci->eraseFromParent();
1853 // ffs(x) -> ( x == 0 ? 0 : llvm.cttz(x)+1)
1854 // ffsl(x) -> ( x == 0 ? 0 : llvm.cttz(x)+1)
1855 // ffsll(x) -> ( x == 0 ? 0 : llvm.cttz(x)+1)
1856 const Type* arg_type = ci->getOperand(1)->getType();
1857 std::vector<const Type*> args;
1858 args.push_back(arg_type);
1859 FunctionType* llvm_cttz_type = FunctionType::get(arg_type,args,false);
1861 SLC.getModule()->getOrInsertFunction("llvm.cttz",llvm_cttz_type);
1862 std::string inst_name(ci->getName()+".ffs");
1864 new CallInst(F, ci->getOperand(1), inst_name, ci);
1865 if (arg_type != Type::IntTy)
1866 call = new CastInst(call, Type::IntTy, inst_name, ci);
1867 BinaryOperator* add = BinaryOperator::createAdd(call,
1868 ConstantSInt::get(Type::IntTy,1), inst_name, ci);
1869 SetCondInst* eq = new SetCondInst(Instruction::SetEQ,ci->getOperand(1),
1870 ConstantSInt::get(ci->getOperand(1)->getType(),0),inst_name,ci);
1871 SelectInst* select = new SelectInst(eq,ConstantSInt::get(Type::IntTy,0),add,
1873 ci->replaceAllUsesWith(select);
1874 ci->eraseFromParent();
1879 /// This LibCallOptimization will simplify calls to the "ffsl" library
1880 /// calls. It simply uses FFSOptimization for which the transformation is
1882 /// @brief Simplify the ffsl library function.
1883 struct FFSLOptimization : public FFSOptimization
1886 /// @brief Default Constructor
1887 FFSLOptimization() : FFSOptimization("ffsl",
1888 "Number of 'ffsl' calls simplified") {}
1892 /// This LibCallOptimization will simplify calls to the "ffsll" library
1893 /// calls. It simply uses FFSOptimization for which the transformation is
1895 /// @brief Simplify the ffsl library function.
1896 struct FFSLLOptimization : public FFSOptimization
1899 /// @brief Default Constructor
1900 FFSLLOptimization() : FFSOptimization("ffsll",
1901 "Number of 'ffsll' calls simplified") {}
1906 /// This LibCallOptimization will simplify calls to the "floor" library
1908 /// @brief Simplify the floor library function.
1909 struct FloorOptimization : public LibCallOptimization {
1911 : LibCallOptimization("floor", "Number of 'floor' calls simplified") {}
1913 /// @brief Make sure that the "floor" function has the right prototype
1914 virtual bool ValidateCalledFunction(const Function *F, SimplifyLibCalls &SLC){
1915 return F->arg_size() == 1 && F->arg_begin()->getType() == Type::DoubleTy &&
1916 F->getReturnType() == Type::DoubleTy;
1919 virtual bool OptimizeCall(CallInst *CI, SimplifyLibCalls &SLC) {
1920 // If this is a float argument passed in, convert to floorf.
1921 // e.g. floor((double)FLT) -> (double)floorf(FLT). There can be no loss of
1922 // precision due to this.
1923 if (CastInst *Cast = dyn_cast<CastInst>(CI->getOperand(1)))
1924 if (Cast->getOperand(0)->getType() == Type::FloatTy) {
1925 Value *New = new CallInst(SLC.get_floorf(), Cast->getOperand(0),
1927 New = new CastInst(New, Type::DoubleTy, CI->getName(), CI);
1928 CI->replaceAllUsesWith(New);
1929 CI->eraseFromParent();
1930 if (Cast->use_empty())
1931 Cast->eraseFromParent();
1934 return false; // opt failed
1940 /// A function to compute the length of a null-terminated constant array of
1941 /// integers. This function can't rely on the size of the constant array
1942 /// because there could be a null terminator in the middle of the array.
1943 /// We also have to bail out if we find a non-integer constant initializer
1944 /// of one of the elements or if there is no null-terminator. The logic
1945 /// below checks each of these conditions and will return true only if all
1946 /// conditions are met. In that case, the \p len parameter is set to the length
1947 /// of the null-terminated string. If false is returned, the conditions were
1948 /// not met and len is set to 0.
1949 /// @brief Get the length of a constant string (null-terminated array).
1950 bool getConstantStringLength(Value* V, uint64_t& len, ConstantArray** CA )
1952 assert(V != 0 && "Invalid args to getConstantStringLength");
1953 len = 0; // make sure we initialize this
1955 // If the value is not a GEP instruction nor a constant expression with a
1956 // GEP instruction, then return false because ConstantArray can't occur
1958 if (GetElementPtrInst* GEPI = dyn_cast<GetElementPtrInst>(V))
1960 else if (ConstantExpr* CE = dyn_cast<ConstantExpr>(V))
1961 if (CE->getOpcode() == Instruction::GetElementPtr)
1968 // Make sure the GEP has exactly three arguments.
1969 if (GEP->getNumOperands() != 3)
1972 // Check to make sure that the first operand of the GEP is an integer and
1973 // has value 0 so that we are sure we're indexing into the initializer.
1974 if (ConstantInt* op1 = dyn_cast<ConstantInt>(GEP->getOperand(1)))
1976 if (!op1->isNullValue())
1982 // Ensure that the second operand is a ConstantInt. If it isn't then this
1983 // GEP is wonky and we're not really sure what were referencing into and
1984 // better of not optimizing it. While we're at it, get the second index
1985 // value. We'll need this later for indexing the ConstantArray.
1986 uint64_t start_idx = 0;
1987 if (ConstantInt* CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
1988 start_idx = CI->getRawValue();
1992 // The GEP instruction, constant or instruction, must reference a global
1993 // variable that is a constant and is initialized. The referenced constant
1994 // initializer is the array that we'll use for optimization.
1995 GlobalVariable* GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
1996 if (!GV || !GV->isConstant() || !GV->hasInitializer())
1999 // Get the initializer.
2000 Constant* INTLZR = GV->getInitializer();
2002 // Handle the ConstantAggregateZero case
2003 if (ConstantAggregateZero* CAZ = dyn_cast<ConstantAggregateZero>(INTLZR))
2005 // This is a degenerate case. The initializer is constant zero so the
2006 // length of the string must be zero.
2011 // Must be a Constant Array
2012 ConstantArray* A = dyn_cast<ConstantArray>(INTLZR);
2016 // Get the number of elements in the array
2017 uint64_t max_elems = A->getType()->getNumElements();
2019 // Traverse the constant array from start_idx (derived above) which is
2020 // the place the GEP refers to in the array.
2021 for ( len = start_idx; len < max_elems; len++)
2023 if (ConstantInt* CI = dyn_cast<ConstantInt>(A->getOperand(len)))
2025 // Check for the null terminator
2026 if (CI->isNullValue())
2027 break; // we found end of string
2030 return false; // This array isn't suitable, non-int initializer
2032 if (len >= max_elems)
2033 return false; // This array isn't null terminated
2035 // Subtract out the initial value from the length
2039 return true; // success!
2042 /// CastToCStr - Return V if it is an sbyte*, otherwise cast it to sbyte*,
2043 /// inserting the cast before IP, and return the cast.
2044 /// @brief Cast a value to a "C" string.
2045 Value *CastToCStr(Value *V, Instruction &IP) {
2046 const Type *SBPTy = PointerType::get(Type::SByteTy);
2047 if (V->getType() != SBPTy)
2048 return new CastInst(V, SBPTy, V->getName(), &IP);
2053 // Additional cases that we need to add to this file:
2056 // * cbrt(expN(X)) -> expN(x/3)
2057 // * cbrt(sqrt(x)) -> pow(x,1/6)
2058 // * cbrt(sqrt(x)) -> pow(x,1/9)
2061 // * cos(-x) -> cos(x)
2064 // * exp(log(x)) -> x
2067 // * log(exp(x)) -> x
2068 // * log(x**y) -> y*log(x)
2069 // * log(exp(y)) -> y*log(e)
2070 // * log(exp2(y)) -> y*log(2)
2071 // * log(exp10(y)) -> y*log(10)
2072 // * log(sqrt(x)) -> 0.5*log(x)
2073 // * log(pow(x,y)) -> y*log(x)
2075 // lround, lroundf, lroundl:
2076 // * lround(cnst) -> cnst'
2079 // * memcmp(x,y,l) -> cnst
2080 // (if all arguments are constant and strlen(x) <= l and strlen(y) <= l)
2083 // * memmove(d,s,l,a) -> memcpy(d,s,l,a)
2084 // (if s is a global constant array)
2087 // * pow(exp(x),y) -> exp(x*y)
2088 // * pow(sqrt(x),y) -> pow(x,y*0.5)
2089 // * pow(pow(x,y),z)-> pow(x,y*z)
2092 // * puts("") -> fputc("\n",stdout) (how do we get "stdout"?)
2094 // round, roundf, roundl:
2095 // * round(cnst) -> cnst'
2098 // * signbit(cnst) -> cnst'
2099 // * signbit(nncst) -> 0 (if pstv is a non-negative constant)
2101 // sqrt, sqrtf, sqrtl:
2102 // * sqrt(expN(x)) -> expN(x*0.5)
2103 // * sqrt(Nroot(x)) -> pow(x,1/(2*N))
2104 // * sqrt(pow(x,y)) -> pow(|x|,y*0.5)
2107 // * stpcpy(str, "literal") ->
2108 // llvm.memcpy(str,"literal",strlen("literal")+1,1)
2110 // * strrchr(s,c) -> reverse_offset_of_in(c,s)
2111 // (if c is a constant integer and s is a constant string)
2112 // * strrchr(s1,0) -> strchr(s1,0)
2115 // * strncat(x,y,0) -> x
2116 // * strncat(x,y,0) -> x (if strlen(y) = 0)
2117 // * strncat(x,y,l) -> strcat(x,y) (if y and l are constants an l > strlen(y))
2120 // * strncpy(d,s,0) -> d
2121 // * strncpy(d,s,l) -> memcpy(d,s,l,1)
2122 // (if s and l are constants)
2125 // * strpbrk(s,a) -> offset_in_for(s,a)
2126 // (if s and a are both constant strings)
2127 // * strpbrk(s,"") -> 0
2128 // * strpbrk(s,a) -> strchr(s,a[0]) (if a is constant string of length 1)
2131 // * strspn(s,a) -> const_int (if both args are constant)
2132 // * strspn("",a) -> 0
2133 // * strspn(s,"") -> 0
2134 // * strcspn(s,a) -> const_int (if both args are constant)
2135 // * strcspn("",a) -> 0
2136 // * strcspn(s,"") -> strlen(a)
2139 // * strstr(x,x) -> x
2140 // * strstr(s1,s2) -> offset_of_s2_in(s1)
2141 // (if s1 and s2 are constant strings)
2144 // * tan(atan(x)) -> x
2146 // trunc, truncf, truncl:
2147 // * trunc(cnst) -> cnst'