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/Config/config.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Target/TargetData.h"
31 #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 list 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 LibCallOptimization *OptList = 0;
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 {
68 LibCallOptimization **Prev, *Next;
69 const char *FunctionName; ///< Name of the library call we optimize
71 Statistic<> occurrences; ///< debug statistic (-debug-only=simplify-libcalls)
74 /// The \p fname argument must be the name of the library function being
75 /// optimized by the subclass.
76 /// @brief Constructor that registers the optimization.
77 LibCallOptimization(const char *FName, const char *Description)
80 , occurrences("simplify-libcalls", Description)
83 // Register this optimizer in the list of optimizations.
87 if (Next) Next->Prev = &Next;
90 /// getNext - All libcall optimizations are chained together into a list,
91 /// return the next one in the list.
92 LibCallOptimization *getNext() { return Next; }
94 /// @brief Deregister from the optlist
95 virtual ~LibCallOptimization() {
97 if (Next) Next->Prev = Prev;
100 /// The implementation of this function in subclasses should determine if
101 /// \p F is suitable for the optimization. This method is called by
102 /// SimplifyLibCalls::runOnModule to short circuit visiting all the call
103 /// sites of such a function if that function is not suitable in the first
104 /// place. If the called function is suitabe, this method should return true;
105 /// false, otherwise. This function should also perform any lazy
106 /// initialization that the LibCallOptimization needs to do, if its to return
107 /// true. This avoids doing initialization until the optimizer is actually
108 /// going to be called upon to do some optimization.
109 /// @brief Determine if the function is suitable for optimization
110 virtual bool ValidateCalledFunction(
111 const Function* F, ///< The function that is the target of call sites
112 SimplifyLibCalls& SLC ///< The pass object invoking us
115 /// The implementations of this function in subclasses is the heart of the
116 /// SimplifyLibCalls algorithm. Sublcasses of this class implement
117 /// OptimizeCall to determine if (a) the conditions are right for optimizing
118 /// the call and (b) to perform the optimization. If an action is taken
119 /// against ci, the subclass is responsible for returning true and ensuring
120 /// that ci is erased from its parent.
121 /// @brief Optimize a call, if possible.
122 virtual bool OptimizeCall(
123 CallInst* ci, ///< The call instruction that should be optimized.
124 SimplifyLibCalls& SLC ///< The pass object invoking us
127 /// @brief Get the name of the library call being optimized
128 const char *getFunctionName() const { return FunctionName; }
130 /// @brief Called by SimplifyLibCalls to update the occurrences statistic.
133 DEBUG(++occurrences);
138 /// This class is an LLVM Pass that applies each of the LibCallOptimization
139 /// instances to all the call sites in a module, relatively efficiently. The
140 /// purpose of this pass is to provide optimizations for calls to well-known
141 /// functions with well-known semantics, such as those in the c library. The
142 /// class provides the basic infrastructure for handling runOnModule. Whenever
143 /// this pass finds a function call, it asks the appropriate optimizer to
144 /// validate the call (ValidateLibraryCall). If it is validated, then
145 /// the OptimizeCall method is also called.
146 /// @brief A ModulePass for optimizing well-known function calls.
147 class SimplifyLibCalls : public ModulePass {
149 /// We need some target data for accurate signature details that are
150 /// target dependent. So we require target data in our AnalysisUsage.
151 /// @brief Require TargetData from AnalysisUsage.
152 virtual void getAnalysisUsage(AnalysisUsage& Info) const {
153 // Ask that the TargetData analysis be performed before us so we can use
155 Info.addRequired<TargetData>();
158 /// For this pass, process all of the function calls in the module, calling
159 /// ValidateLibraryCall and OptimizeCall as appropriate.
160 /// @brief Run all the lib call optimizations on a Module.
161 virtual bool runOnModule(Module &M) {
165 hash_map<std::string, LibCallOptimization*> OptznMap;
166 for (LibCallOptimization *Optzn = OptList; Optzn; Optzn = Optzn->getNext())
167 OptznMap[Optzn->getFunctionName()] = Optzn;
169 // The call optimizations can be recursive. That is, the optimization might
170 // generate a call to another function which can also be optimized. This way
171 // we make the LibCallOptimization instances very specific to the case they
172 // handle. It also means we need to keep running over the function calls in
173 // the module until we don't get any more optimizations possible.
174 bool found_optimization = false;
176 found_optimization = false;
177 for (Module::iterator FI = M.begin(), FE = M.end(); FI != FE; ++FI) {
178 // All the "well-known" functions are external and have external linkage
179 // because they live in a runtime library somewhere and were (probably)
180 // not compiled by LLVM. So, we only act on external functions that
181 // have external linkage and non-empty uses.
182 if (!FI->isExternal() || !FI->hasExternalLinkage() || FI->use_empty())
185 // Get the optimization class that pertains to this function
186 hash_map<std::string, LibCallOptimization*>::iterator OMI =
187 OptznMap.find(FI->getName());
188 if (OMI == OptznMap.end()) continue;
190 LibCallOptimization *CO = OMI->second;
192 // Make sure the called function is suitable for the optimization
193 if (!CO->ValidateCalledFunction(FI, *this))
196 // Loop over each of the uses of the function
197 for (Value::use_iterator UI = FI->use_begin(), UE = FI->use_end();
199 // If the use of the function is a call instruction
200 if (CallInst* CI = dyn_cast<CallInst>(*UI++)) {
201 // Do the optimization on the LibCallOptimization.
202 if (CO->OptimizeCall(CI, *this)) {
203 ++SimplifiedLibCalls;
204 found_optimization = result = true;
210 } while (found_optimization);
215 /// @brief Return the *current* module we're working on.
216 Module* getModule() const { return M; }
218 /// @brief Return the *current* target data for the module we're working on.
219 TargetData* getTargetData() const { return TD; }
221 /// @brief Return the size_t type -- syntactic shortcut
222 const Type* getIntPtrType() const { return TD->getIntPtrType(); }
224 /// @brief Return a Function* for the putchar libcall
225 Function* get_putchar() {
227 putchar_func = M->getOrInsertFunction("putchar", Type::IntTy, Type::IntTy,
232 /// @brief Return a Function* for the puts libcall
233 Function* get_puts() {
235 puts_func = M->getOrInsertFunction("puts", Type::IntTy,
236 PointerType::get(Type::SByteTy),
241 /// @brief Return a Function* for the fputc libcall
242 Function* get_fputc(const Type* FILEptr_type) {
244 fputc_func = M->getOrInsertFunction("fputc", Type::IntTy, Type::IntTy,
249 /// @brief Return a Function* for the fputs libcall
250 Function* get_fputs(const Type* FILEptr_type) {
252 fputs_func = M->getOrInsertFunction("fputs", Type::IntTy,
253 PointerType::get(Type::SByteTy),
258 /// @brief Return a Function* for the fwrite libcall
259 Function* get_fwrite(const Type* FILEptr_type) {
261 fwrite_func = M->getOrInsertFunction("fwrite", TD->getIntPtrType(),
262 PointerType::get(Type::SByteTy),
269 /// @brief Return a Function* for the sqrt libcall
270 Function* get_sqrt() {
272 sqrt_func = M->getOrInsertFunction("sqrt", Type::DoubleTy,
273 Type::DoubleTy, NULL);
277 /// @brief Return a Function* for the strlen libcall
278 Function* get_strcpy() {
280 strcpy_func = M->getOrInsertFunction("strcpy",
281 PointerType::get(Type::SByteTy),
282 PointerType::get(Type::SByteTy),
283 PointerType::get(Type::SByteTy),
288 /// @brief Return a Function* for the strlen libcall
289 Function* get_strlen() {
291 strlen_func = M->getOrInsertFunction("strlen", TD->getIntPtrType(),
292 PointerType::get(Type::SByteTy),
297 /// @brief Return a Function* for the memchr libcall
298 Function* get_memchr() {
300 memchr_func = M->getOrInsertFunction("memchr",
301 PointerType::get(Type::SByteTy),
302 PointerType::get(Type::SByteTy),
303 Type::IntTy, TD->getIntPtrType(),
308 /// @brief Return a Function* for the memcpy libcall
309 Function* get_memcpy() {
311 const Type *SBP = PointerType::get(Type::SByteTy);
312 const char *N = TD->getIntPtrType() == Type::UIntTy ?
313 "llvm.memcpy.i32" : "llvm.memcpy.i64";
314 memcpy_func = M->getOrInsertFunction(N, Type::VoidTy, SBP, SBP,
315 TD->getIntPtrType(), Type::UIntTy,
321 Function *getUnaryFloatFunction(const char *Name, Function *&Cache) {
323 Cache = M->getOrInsertFunction(Name, Type::FloatTy, Type::FloatTy, NULL);
327 Function *get_floorf() { return getUnaryFloatFunction("floorf", floorf_func);}
328 Function *get_ceilf() { return getUnaryFloatFunction( "ceilf", ceilf_func);}
329 Function *get_roundf() { return getUnaryFloatFunction("roundf", roundf_func);}
330 Function *get_rintf() { return getUnaryFloatFunction( "rintf", rintf_func);}
331 Function *get_nearbyintf() { return getUnaryFloatFunction("nearbyintf",
334 /// @brief Reset our cached data for a new Module
335 void reset(Module& mod) {
337 TD = &getAnalysis<TargetData>();
356 /// Caches for function pointers.
357 Function *putchar_func, *puts_func;
358 Function *fputc_func, *fputs_func, *fwrite_func;
359 Function *memcpy_func, *memchr_func;
361 Function *strcpy_func, *strlen_func;
362 Function *floorf_func, *ceilf_func, *roundf_func;
363 Function *rintf_func, *nearbyintf_func;
364 Module *M; ///< Cached Module
365 TargetData *TD; ///< Cached TargetData
369 RegisterOpt<SimplifyLibCalls>
370 X("simplify-libcalls","Simplify well-known library calls");
372 } // anonymous namespace
374 // The only public symbol in this file which just instantiates the pass object
375 ModulePass *llvm::createSimplifyLibCallsPass() {
376 return new SimplifyLibCalls();
379 // Classes below here, in the anonymous namespace, are all subclasses of the
380 // LibCallOptimization class, each implementing all optimizations possible for a
381 // single well-known library call. Each has a static singleton instance that
382 // auto registers it into the "optlist" global above.
385 // Forward declare utility functions.
386 bool getConstantStringLength(Value* V, uint64_t& len, ConstantArray** A = 0 );
387 Value *CastToCStr(Value *V, Instruction &IP);
389 /// This LibCallOptimization will find instances of a call to "exit" that occurs
390 /// within the "main" function and change it to a simple "ret" instruction with
391 /// the same value passed to the exit function. When this is done, it splits the
392 /// basic block at the exit(3) call and deletes the call instruction.
393 /// @brief Replace calls to exit in main with a simple return
394 struct ExitInMainOptimization : public LibCallOptimization {
395 ExitInMainOptimization() : LibCallOptimization("exit",
396 "Number of 'exit' calls simplified") {}
398 // Make sure the called function looks like exit (int argument, int return
399 // type, external linkage, not varargs).
400 virtual bool ValidateCalledFunction(const Function *F, SimplifyLibCalls &SLC){
401 return F->arg_size() >= 1 && F->arg_begin()->getType()->isInteger();
404 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC) {
405 // To be careful, we check that the call to exit is coming from "main", that
406 // main has external linkage, and the return type of main and the argument
407 // to exit have the same type.
408 Function *from = ci->getParent()->getParent();
409 if (from->hasExternalLinkage())
410 if (from->getReturnType() == ci->getOperand(1)->getType())
411 if (from->getName() == "main") {
412 // Okay, time to actually do the optimization. First, get the basic
413 // block of the call instruction
414 BasicBlock* bb = ci->getParent();
416 // Create a return instruction that we'll replace the call with.
417 // Note that the argument of the return is the argument of the call
419 new ReturnInst(ci->getOperand(1), ci);
421 // Split the block at the call instruction which places it in a new
423 bb->splitBasicBlock(ci);
425 // The block split caused a branch instruction to be inserted into
426 // the end of the original block, right after the return instruction
427 // that we put there. That's not a valid block, so delete the branch
429 bb->getInstList().pop_back();
431 // Now we can finally get rid of the call instruction which now lives
432 // in the new basic block.
433 ci->eraseFromParent();
435 // Optimization succeeded, return true.
438 // We didn't pass the criteria for this optimization so return false
441 } ExitInMainOptimizer;
443 /// This LibCallOptimization will simplify a call to the strcat library
444 /// function. The simplification is possible only if the string being
445 /// concatenated is a constant array or a constant expression that results in
446 /// a constant string. In this case we can replace it with strlen + llvm.memcpy
447 /// of the constant string. Both of these calls are further reduced, if possible
448 /// on subsequent passes.
449 /// @brief Simplify the strcat library function.
450 struct StrCatOptimization : public LibCallOptimization {
452 /// @brief Default constructor
453 StrCatOptimization() : LibCallOptimization("strcat",
454 "Number of 'strcat' calls simplified") {}
458 /// @brief Make sure that the "strcat" function has the right prototype
459 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) {
476 // Extract some information from the instruction
477 Value* dest = ci->getOperand(1);
478 Value* src = ci->getOperand(2);
480 // Extract the initializer (while making numerous checks) from the
481 // source operand of the call to strcat. If we get null back, one of
482 // a variety of checks in get_GVInitializer failed
484 if (!getConstantStringLength(src,len))
487 // 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(SLC.getIntPtrType(),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 {
537 StrChrOptimization() : LibCallOptimization("strchr",
538 "Number of 'strchr' calls simplified") {}
540 /// @brief Make sure that the "strchr" function has the right prototype
541 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC){
542 if (f->getReturnType() == PointerType::get(Type::SByteTy) &&
548 /// @brief Perform the strchr optimizations
549 virtual bool OptimizeCall(CallInst *ci, SimplifyLibCalls &SLC) {
550 // If there aren't three operands, bail
551 if (ci->getNumOperands() != 3)
554 // Check that the first argument to strchr is a constant array of sbyte.
555 // If it is, get the length and data, otherwise return false.
558 if (!getConstantStringLength(ci->getOperand(1),len,&CA))
561 // Check that the second argument to strchr is a constant int, return false
563 ConstantSInt* CSI = dyn_cast<ConstantSInt>(ci->getOperand(2));
565 // Just lower this to memchr since we know the length of the string as
567 Function* f = SLC.get_memchr();
568 std::vector<Value*> args;
569 args.push_back(ci->getOperand(1));
570 args.push_back(ci->getOperand(2));
571 args.push_back(ConstantUInt::get(SLC.getIntPtrType(),len));
572 ci->replaceAllUsesWith( new CallInst(f,args,ci->getName(),ci));
573 ci->eraseFromParent();
577 // Get the character we're looking for
578 int64_t chr = CSI->getValue();
580 // Compute the offset
582 bool char_found = false;
583 for (uint64_t i = 0; i < len; ++i) {
584 if (ConstantSInt* CI = dyn_cast<ConstantSInt>(CA->getOperand(i))) {
585 // Check for the null terminator
586 if (CI->isNullValue())
587 break; // we found end of string
588 else if (CI->getValue() == chr) {
596 // strchr(s,c) -> offset_of_in(c,s)
597 // (if c is a constant integer and s is a constant string)
599 std::vector<Value*> indices;
600 indices.push_back(ConstantUInt::get(Type::ULongTy,offset));
601 GetElementPtrInst* GEP = new GetElementPtrInst(ci->getOperand(1),indices,
602 ci->getOperand(1)->getName()+".strchr",ci);
603 ci->replaceAllUsesWith(GEP);
605 ci->replaceAllUsesWith(
606 ConstantPointerNull::get(PointerType::get(Type::SByteTy)));
608 ci->eraseFromParent();
613 /// This LibCallOptimization will simplify a call to the strcmp library
614 /// function. It optimizes out cases where one or both arguments are constant
615 /// and the result can be determined statically.
616 /// @brief Simplify the strcmp library function.
617 struct StrCmpOptimization : public LibCallOptimization {
619 StrCmpOptimization() : LibCallOptimization("strcmp",
620 "Number of 'strcmp' calls simplified") {}
622 /// @brief Make sure that the "strcmp" function has the right prototype
623 virtual bool ValidateCalledFunction(const Function *F, SimplifyLibCalls &SLC){
624 return F->getReturnType() == Type::IntTy && F->arg_size() == 2;
627 /// @brief Perform the strcmp optimization
628 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC) {
629 // First, check to see if src and destination are the same. If they are,
630 // then the optimization is to replace the CallInst with a constant 0
631 // because the call is a no-op.
632 Value* s1 = ci->getOperand(1);
633 Value* s2 = ci->getOperand(2);
636 ci->replaceAllUsesWith(ConstantInt::get(Type::IntTy,0));
637 ci->eraseFromParent();
641 bool isstr_1 = false;
644 if (getConstantStringLength(s1,len_1,&A1)) {
647 // strcmp("",x) -> *x
649 new LoadInst(CastToCStr(s2,*ci), ci->getName()+".load",ci);
651 new CastInst(load,Type::IntTy,ci->getName()+".int",ci);
652 ci->replaceAllUsesWith(cast);
653 ci->eraseFromParent();
658 bool isstr_2 = false;
661 if (getConstantStringLength(s2, len_2, &A2)) {
664 // strcmp(x,"") -> *x
666 new LoadInst(CastToCStr(s1,*ci),ci->getName()+".val",ci);
668 new CastInst(load,Type::IntTy,ci->getName()+".int",ci);
669 ci->replaceAllUsesWith(cast);
670 ci->eraseFromParent();
675 if (isstr_1 && isstr_2) {
676 // strcmp(x,y) -> cnst (if both x and y are constant strings)
677 std::string str1 = A1->getAsString();
678 std::string str2 = A2->getAsString();
679 int result = strcmp(str1.c_str(), str2.c_str());
680 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy,result));
681 ci->eraseFromParent();
688 /// This LibCallOptimization will simplify a call to the strncmp library
689 /// function. It optimizes out cases where one or both arguments are constant
690 /// and the result can be determined statically.
691 /// @brief Simplify the strncmp library function.
692 struct StrNCmpOptimization : public LibCallOptimization {
694 StrNCmpOptimization() : LibCallOptimization("strncmp",
695 "Number of 'strncmp' calls simplified") {}
697 /// @brief Make sure that the "strncmp" function has the right prototype
698 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC){
699 if (f->getReturnType() == Type::IntTy && f->arg_size() == 3)
704 /// @brief Perform the strncpy optimization
705 virtual bool OptimizeCall(CallInst *ci, SimplifyLibCalls &SLC) {
706 // First, check to see if src and destination are the same. If they are,
707 // then the optimization is to replace the CallInst with a constant 0
708 // because the call is a no-op.
709 Value* s1 = ci->getOperand(1);
710 Value* s2 = ci->getOperand(2);
712 // strncmp(x,x,l) -> 0
713 ci->replaceAllUsesWith(ConstantInt::get(Type::IntTy,0));
714 ci->eraseFromParent();
718 // Check the length argument, if it is Constant zero then the strings are
720 uint64_t len_arg = 0;
721 bool len_arg_is_const = false;
722 if (ConstantInt* len_CI = dyn_cast<ConstantInt>(ci->getOperand(3))) {
723 len_arg_is_const = true;
724 len_arg = len_CI->getRawValue();
726 // strncmp(x,y,0) -> 0
727 ci->replaceAllUsesWith(ConstantInt::get(Type::IntTy,0));
728 ci->eraseFromParent();
733 bool isstr_1 = false;
736 if (getConstantStringLength(s1, len_1, &A1)) {
739 // strncmp("",x) -> *x
740 LoadInst* load = new LoadInst(s1,ci->getName()+".load",ci);
742 new CastInst(load,Type::IntTy,ci->getName()+".int",ci);
743 ci->replaceAllUsesWith(cast);
744 ci->eraseFromParent();
749 bool isstr_2 = false;
752 if (getConstantStringLength(s2,len_2,&A2)) {
755 // strncmp(x,"") -> *x
756 LoadInst* load = new LoadInst(s2,ci->getName()+".val",ci);
758 new CastInst(load,Type::IntTy,ci->getName()+".int",ci);
759 ci->replaceAllUsesWith(cast);
760 ci->eraseFromParent();
765 if (isstr_1 && isstr_2 && len_arg_is_const) {
766 // strncmp(x,y,const) -> constant
767 std::string str1 = A1->getAsString();
768 std::string str2 = A2->getAsString();
769 int result = strncmp(str1.c_str(), str2.c_str(), len_arg);
770 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy,result));
771 ci->eraseFromParent();
778 /// This LibCallOptimization will simplify a call to the strcpy library
779 /// function. Two optimizations are possible:
780 /// (1) If src and dest are the same and not volatile, just return dest
781 /// (2) If the src is a constant then we can convert to llvm.memmove
782 /// @brief Simplify the strcpy library function.
783 struct StrCpyOptimization : public LibCallOptimization {
785 StrCpyOptimization() : LibCallOptimization("strcpy",
786 "Number of 'strcpy' calls simplified") {}
788 /// @brief Make sure that the "strcpy" function has the right prototype
789 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC){
790 if (f->getReturnType() == PointerType::get(Type::SByteTy))
791 if (f->arg_size() == 2) {
792 Function::const_arg_iterator AI = f->arg_begin();
793 if (AI++->getType() == PointerType::get(Type::SByteTy))
794 if (AI->getType() == PointerType::get(Type::SByteTy)) {
795 // Indicate this is a suitable call type.
802 /// @brief Perform the strcpy optimization
803 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC) {
804 // First, check to see if src and destination are the same. If they are,
805 // then the optimization is to replace the CallInst with the destination
806 // because the call is a no-op. Note that this corresponds to the
807 // degenerate strcpy(X,X) case which should have "undefined" results
808 // according to the C specification. However, it occurs sometimes and
809 // we optimize it as a no-op.
810 Value* dest = ci->getOperand(1);
811 Value* src = ci->getOperand(2);
813 ci->replaceAllUsesWith(dest);
814 ci->eraseFromParent();
818 // Get the length of the constant string referenced by the second operand,
819 // the "src" parameter. Fail the optimization if we can't get the length
820 // (note that getConstantStringLength does lots of checks to make sure this
823 if (!getConstantStringLength(ci->getOperand(2),len))
826 // If the constant string's length is zero we can optimize this by just
827 // doing a store of 0 at the first byte of the destination
829 new StoreInst(ConstantInt::get(Type::SByteTy,0),ci->getOperand(1),ci);
830 ci->replaceAllUsesWith(dest);
831 ci->eraseFromParent();
835 // Increment the length because we actually want to memcpy the null
836 // terminator as well.
839 // We have enough information to now generate the memcpy call to
840 // do the concatenation for us.
841 std::vector<Value*> vals;
842 vals.push_back(dest); // destination
843 vals.push_back(src); // source
844 vals.push_back(ConstantUInt::get(SLC.getIntPtrType(),len)); // length
845 vals.push_back(ConstantUInt::get(Type::UIntTy,1)); // alignment
846 new CallInst(SLC.get_memcpy(), vals, "", ci);
848 // Finally, substitute the first operand of the strcat call for the
849 // strcat call itself since strcat returns its first operand; and,
850 // kill the strcat CallInst.
851 ci->replaceAllUsesWith(dest);
852 ci->eraseFromParent();
857 /// This LibCallOptimization will simplify a call to the strlen library
858 /// function by replacing it with a constant value if the string provided to
859 /// it is a constant array.
860 /// @brief Simplify the strlen library function.
861 struct StrLenOptimization : public LibCallOptimization {
862 StrLenOptimization() : LibCallOptimization("strlen",
863 "Number of 'strlen' calls simplified") {}
865 /// @brief Make sure that the "strlen" function has the right prototype
866 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC)
868 if (f->getReturnType() == SLC.getTargetData()->getIntPtrType())
869 if (f->arg_size() == 1)
870 if (Function::const_arg_iterator AI = f->arg_begin())
871 if (AI->getType() == PointerType::get(Type::SByteTy))
876 /// @brief Perform the strlen optimization
877 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC)
879 // Make sure we're dealing with an sbyte* here.
880 Value* str = ci->getOperand(1);
881 if (str->getType() != PointerType::get(Type::SByteTy))
884 // Does the call to strlen have exactly one use?
886 // Is that single use a binary operator?
887 if (BinaryOperator* bop = dyn_cast<BinaryOperator>(ci->use_back()))
888 // Is it compared against a constant integer?
889 if (ConstantInt* CI = dyn_cast<ConstantInt>(bop->getOperand(1)))
891 // Get the value the strlen result is compared to
892 uint64_t val = CI->getRawValue();
894 // If its compared against length 0 with == or !=
896 (bop->getOpcode() == Instruction::SetEQ ||
897 bop->getOpcode() == Instruction::SetNE))
899 // strlen(x) != 0 -> *x != 0
900 // strlen(x) == 0 -> *x == 0
901 LoadInst* load = new LoadInst(str,str->getName()+".first",ci);
902 BinaryOperator* rbop = BinaryOperator::create(bop->getOpcode(),
903 load, ConstantSInt::get(Type::SByteTy,0),
904 bop->getName()+".strlen", ci);
905 bop->replaceAllUsesWith(rbop);
906 bop->eraseFromParent();
907 ci->eraseFromParent();
912 // Get the length of the constant string operand
914 if (!getConstantStringLength(ci->getOperand(1),len))
917 // strlen("xyz") -> 3 (for example)
918 const Type *Ty = SLC.getTargetData()->getIntPtrType();
920 ci->replaceAllUsesWith(ConstantSInt::get(Ty, len));
922 ci->replaceAllUsesWith(ConstantUInt::get(Ty, len));
924 ci->eraseFromParent();
929 /// IsOnlyUsedInEqualsComparison - Return true if it only matters that the value
930 /// is equal or not-equal to zero.
931 static bool IsOnlyUsedInEqualsZeroComparison(Instruction *I) {
932 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
934 Instruction *User = cast<Instruction>(*UI);
935 if (User->getOpcode() == Instruction::SetNE ||
936 User->getOpcode() == Instruction::SetEQ) {
937 if (isa<Constant>(User->getOperand(1)) &&
938 cast<Constant>(User->getOperand(1))->isNullValue())
940 } else if (CastInst *CI = dyn_cast<CastInst>(User))
941 if (CI->getType() == Type::BoolTy)
943 // Unknown instruction.
949 /// This memcmpOptimization will simplify a call to the memcmp library
951 struct memcmpOptimization : public LibCallOptimization {
952 /// @brief Default Constructor
954 : LibCallOptimization("memcmp", "Number of 'memcmp' calls simplified") {}
956 /// @brief Make sure that the "memcmp" function has the right prototype
957 virtual bool ValidateCalledFunction(const Function *F, SimplifyLibCalls &TD) {
958 Function::const_arg_iterator AI = F->arg_begin();
959 if (F->arg_size() != 3 || !isa<PointerType>(AI->getType())) return false;
960 if (!isa<PointerType>((++AI)->getType())) return false;
961 if (!(++AI)->getType()->isInteger()) return false;
962 if (!F->getReturnType()->isInteger()) return false;
966 /// Because of alignment and instruction information that we don't have, we
967 /// leave the bulk of this to the code generators.
969 /// Note that we could do much more if we could force alignment on otherwise
970 /// small aligned allocas, or if we could indicate that loads have a small
972 virtual bool OptimizeCall(CallInst *CI, SimplifyLibCalls &TD) {
973 Value *LHS = CI->getOperand(1), *RHS = CI->getOperand(2);
975 // If the two operands are the same, return zero.
977 // memcmp(s,s,x) -> 0
978 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
979 CI->eraseFromParent();
983 // Make sure we have a constant length.
984 ConstantInt *LenC = dyn_cast<ConstantInt>(CI->getOperand(3));
985 if (!LenC) return false;
986 uint64_t Len = LenC->getRawValue();
988 // If the length is zero, this returns 0.
991 // memcmp(s1,s2,0) -> 0
992 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
993 CI->eraseFromParent();
996 // memcmp(S1,S2,1) -> *(ubyte*)S1 - *(ubyte*)S2
997 const Type *UCharPtr = PointerType::get(Type::UByteTy);
998 CastInst *Op1Cast = new CastInst(LHS, UCharPtr, LHS->getName(), CI);
999 CastInst *Op2Cast = new CastInst(RHS, UCharPtr, RHS->getName(), CI);
1000 Value *S1V = new LoadInst(Op1Cast, LHS->getName()+".val", CI);
1001 Value *S2V = new LoadInst(Op2Cast, RHS->getName()+".val", CI);
1002 Value *RV = BinaryOperator::createSub(S1V, S2V, CI->getName()+".diff",CI);
1003 if (RV->getType() != CI->getType())
1004 RV = new CastInst(RV, CI->getType(), RV->getName(), CI);
1005 CI->replaceAllUsesWith(RV);
1006 CI->eraseFromParent();
1010 if (IsOnlyUsedInEqualsZeroComparison(CI)) {
1011 // TODO: IF both are aligned, use a short load/compare.
1013 // memcmp(S1,S2,2) -> S1[0]-S2[0] | S1[1]-S2[1] iff only ==/!= 0 matters
1014 const Type *UCharPtr = PointerType::get(Type::UByteTy);
1015 CastInst *Op1Cast = new CastInst(LHS, UCharPtr, LHS->getName(), CI);
1016 CastInst *Op2Cast = new CastInst(RHS, UCharPtr, RHS->getName(), CI);
1017 Value *S1V1 = new LoadInst(Op1Cast, LHS->getName()+".val1", CI);
1018 Value *S2V1 = new LoadInst(Op2Cast, RHS->getName()+".val1", CI);
1019 Value *D1 = BinaryOperator::createSub(S1V1, S2V1,
1020 CI->getName()+".d1", CI);
1021 Constant *One = ConstantInt::get(Type::IntTy, 1);
1022 Value *G1 = new GetElementPtrInst(Op1Cast, One, "next1v", CI);
1023 Value *G2 = new GetElementPtrInst(Op2Cast, One, "next2v", CI);
1024 Value *S1V2 = new LoadInst(G1, LHS->getName()+".val2", CI);
1025 Value *S2V2 = new LoadInst(G2, RHS->getName()+".val2", CI);
1026 Value *D2 = BinaryOperator::createSub(S1V2, S2V2,
1027 CI->getName()+".d1", CI);
1028 Value *Or = BinaryOperator::createOr(D1, D2, CI->getName()+".res", CI);
1029 if (Or->getType() != CI->getType())
1030 Or = new CastInst(Or, CI->getType(), Or->getName(), CI);
1031 CI->replaceAllUsesWith(Or);
1032 CI->eraseFromParent();
1045 /// This LibCallOptimization will simplify a call to the memcpy library
1046 /// function by expanding it out to a single store of size 0, 1, 2, 4, or 8
1047 /// bytes depending on the length of the string and the alignment. Additional
1048 /// optimizations are possible in code generation (sequence of immediate store)
1049 /// @brief Simplify the memcpy library function.
1050 struct LLVMMemCpyMoveOptzn : public LibCallOptimization {
1051 LLVMMemCpyMoveOptzn(const char* fname, const char* desc)
1052 : LibCallOptimization(fname, desc) {}
1054 /// @brief Make sure that the "memcpy" function has the right prototype
1055 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& TD) {
1056 // Just make sure this has 4 arguments per LLVM spec.
1057 return (f->arg_size() == 4);
1060 /// Because of alignment and instruction information that we don't have, we
1061 /// leave the bulk of this to the code generators. The optimization here just
1062 /// deals with a few degenerate cases where the length of the string and the
1063 /// alignment match the sizes of our intrinsic types so we can do a load and
1064 /// store instead of the memcpy call.
1065 /// @brief Perform the memcpy optimization.
1066 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& TD) {
1067 // Make sure we have constant int values to work with
1068 ConstantInt* LEN = dyn_cast<ConstantInt>(ci->getOperand(3));
1071 ConstantInt* ALIGN = dyn_cast<ConstantInt>(ci->getOperand(4));
1075 // If the length is larger than the alignment, we can't optimize
1076 uint64_t len = LEN->getRawValue();
1077 uint64_t alignment = ALIGN->getRawValue();
1079 alignment = 1; // Alignment 0 is identity for alignment 1
1080 if (len > alignment)
1083 // Get the type we will cast to, based on size of the string
1084 Value* dest = ci->getOperand(1);
1085 Value* src = ci->getOperand(2);
1090 // memcpy(d,s,0,a) -> noop
1091 ci->eraseFromParent();
1093 case 1: castType = Type::SByteTy; break;
1094 case 2: castType = Type::ShortTy; break;
1095 case 4: castType = Type::IntTy; break;
1096 case 8: castType = Type::LongTy; break;
1101 // Cast source and dest to the right sized primitive and then load/store
1103 new CastInst(src,PointerType::get(castType),src->getName()+".cast",ci);
1104 CastInst* DestCast =
1105 new CastInst(dest,PointerType::get(castType),dest->getName()+".cast",ci);
1106 LoadInst* LI = new LoadInst(SrcCast,SrcCast->getName()+".val",ci);
1107 StoreInst* SI = new StoreInst(LI, DestCast, ci);
1108 ci->eraseFromParent();
1113 /// This LibCallOptimization will simplify a call to the memcpy/memmove library
1115 LLVMMemCpyMoveOptzn LLVMMemCpyOptimizer32("llvm.memcpy.i32",
1116 "Number of 'llvm.memcpy' calls simplified");
1117 LLVMMemCpyMoveOptzn LLVMMemCpyOptimizer64("llvm.memcpy.i64",
1118 "Number of 'llvm.memcpy' calls simplified");
1119 LLVMMemCpyMoveOptzn LLVMMemMoveOptimizer32("llvm.memmove.i32",
1120 "Number of 'llvm.memmove' calls simplified");
1121 LLVMMemCpyMoveOptzn LLVMMemMoveOptimizer64("llvm.memmove.i64",
1122 "Number of 'llvm.memmove' calls simplified");
1124 /// This LibCallOptimization will simplify a call to the memset library
1125 /// function by expanding it out to a single store of size 0, 1, 2, 4, or 8
1126 /// bytes depending on the length argument.
1127 struct LLVMMemSetOptimization : public LibCallOptimization {
1128 /// @brief Default Constructor
1129 LLVMMemSetOptimization(const char *Name) : LibCallOptimization(Name,
1130 "Number of 'llvm.memset' calls simplified") {}
1132 /// @brief Make sure that the "memset" function has the right prototype
1133 virtual bool ValidateCalledFunction(const Function *F, SimplifyLibCalls &TD) {
1134 // Just make sure this has 3 arguments per LLVM spec.
1135 return F->arg_size() == 4;
1138 /// Because of alignment and instruction information that we don't have, we
1139 /// leave the bulk of this to the code generators. The optimization here just
1140 /// deals with a few degenerate cases where the length parameter is constant
1141 /// and the alignment matches the sizes of our intrinsic types so we can do
1142 /// store instead of the memcpy call. Other calls are transformed into the
1143 /// llvm.memset intrinsic.
1144 /// @brief Perform the memset optimization.
1145 virtual bool OptimizeCall(CallInst *ci, SimplifyLibCalls &TD) {
1146 // Make sure we have constant int values to work with
1147 ConstantInt* LEN = dyn_cast<ConstantInt>(ci->getOperand(3));
1150 ConstantInt* ALIGN = dyn_cast<ConstantInt>(ci->getOperand(4));
1154 // Extract the length and alignment
1155 uint64_t len = LEN->getRawValue();
1156 uint64_t alignment = ALIGN->getRawValue();
1158 // Alignment 0 is identity for alignment 1
1162 // If the length is zero, this is a no-op
1164 // memset(d,c,0,a) -> noop
1165 ci->eraseFromParent();
1169 // If the length is larger than the alignment, we can't optimize
1170 if (len > alignment)
1173 // Make sure we have a constant ubyte to work with so we can extract
1174 // the value to be filled.
1175 ConstantUInt* FILL = dyn_cast<ConstantUInt>(ci->getOperand(2));
1178 if (FILL->getType() != Type::UByteTy)
1181 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
1183 // Extract the fill character
1184 uint64_t fill_char = FILL->getValue();
1185 uint64_t fill_value = fill_char;
1187 // Get the type we will cast to, based on size of memory area to fill, and
1188 // and the value we will store there.
1189 Value* dest = ci->getOperand(1);
1193 castType = Type::UByteTy;
1196 castType = Type::UShortTy;
1197 fill_value |= fill_char << 8;
1200 castType = Type::UIntTy;
1201 fill_value |= fill_char << 8 | fill_char << 16 | fill_char << 24;
1204 castType = Type::ULongTy;
1205 fill_value |= fill_char << 8 | fill_char << 16 | fill_char << 24;
1206 fill_value |= fill_char << 32 | fill_char << 40 | fill_char << 48;
1207 fill_value |= fill_char << 56;
1213 // Cast dest to the right sized primitive and then load/store
1214 CastInst* DestCast =
1215 new CastInst(dest,PointerType::get(castType),dest->getName()+".cast",ci);
1216 new StoreInst(ConstantUInt::get(castType,fill_value),DestCast, ci);
1217 ci->eraseFromParent();
1222 LLVMMemSetOptimization MemSet32Optimizer("llvm.memset.i32");
1223 LLVMMemSetOptimization MemSet64Optimizer("llvm.memset.i64");
1226 /// This LibCallOptimization will simplify calls to the "pow" library
1227 /// function. It looks for cases where the result of pow is well known and
1228 /// substitutes the appropriate value.
1229 /// @brief Simplify the pow library function.
1230 struct PowOptimization : public LibCallOptimization {
1232 /// @brief Default Constructor
1233 PowOptimization() : LibCallOptimization("pow",
1234 "Number of 'pow' calls simplified") {}
1236 /// @brief Make sure that the "pow" function has the right prototype
1237 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC){
1238 // Just make sure this has 2 arguments
1239 return (f->arg_size() == 2);
1242 /// @brief Perform the pow optimization.
1243 virtual bool OptimizeCall(CallInst *ci, SimplifyLibCalls &SLC) {
1244 const Type *Ty = cast<Function>(ci->getOperand(0))->getReturnType();
1245 Value* base = ci->getOperand(1);
1246 Value* expn = ci->getOperand(2);
1247 if (ConstantFP *Op1 = dyn_cast<ConstantFP>(base)) {
1248 double Op1V = Op1->getValue();
1250 // pow(1.0,x) -> 1.0
1251 ci->replaceAllUsesWith(ConstantFP::get(Ty,1.0));
1252 ci->eraseFromParent();
1255 } else if (ConstantFP* Op2 = dyn_cast<ConstantFP>(expn)) {
1256 double Op2V = Op2->getValue();
1258 // pow(x,0.0) -> 1.0
1259 ci->replaceAllUsesWith(ConstantFP::get(Ty,1.0));
1260 ci->eraseFromParent();
1262 } else if (Op2V == 0.5) {
1263 // pow(x,0.5) -> sqrt(x)
1264 CallInst* sqrt_inst = new CallInst(SLC.get_sqrt(), base,
1265 ci->getName()+".pow",ci);
1266 ci->replaceAllUsesWith(sqrt_inst);
1267 ci->eraseFromParent();
1269 } else if (Op2V == 1.0) {
1271 ci->replaceAllUsesWith(base);
1272 ci->eraseFromParent();
1274 } else if (Op2V == -1.0) {
1275 // pow(x,-1.0) -> 1.0/x
1276 BinaryOperator* div_inst= BinaryOperator::createDiv(
1277 ConstantFP::get(Ty,1.0), base, ci->getName()+".pow", ci);
1278 ci->replaceAllUsesWith(div_inst);
1279 ci->eraseFromParent();
1283 return false; // opt failed
1287 /// This LibCallOptimization will simplify calls to the "printf" library
1288 /// function. It looks for cases where the result of printf is not used and the
1289 /// operation can be reduced to something simpler.
1290 /// @brief Simplify the printf library function.
1291 struct PrintfOptimization : public LibCallOptimization {
1293 /// @brief Default Constructor
1294 PrintfOptimization() : LibCallOptimization("printf",
1295 "Number of 'printf' calls simplified") {}
1297 /// @brief Make sure that the "printf" function has the right prototype
1298 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC){
1299 // Just make sure this has at least 1 arguments
1300 return (f->arg_size() >= 1);
1303 /// @brief Perform the printf optimization.
1304 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC) {
1305 // If the call has more than 2 operands, we can't optimize it
1306 if (ci->getNumOperands() > 3 || ci->getNumOperands() <= 2)
1309 // If the result of the printf call is used, none of these optimizations
1311 if (!ci->use_empty())
1314 // All the optimizations depend on the length of the first argument and the
1315 // fact that it is a constant string array. Check that now
1317 ConstantArray* CA = 0;
1318 if (!getConstantStringLength(ci->getOperand(1), len, &CA))
1321 if (len != 2 && len != 3)
1324 // The first character has to be a %
1325 if (ConstantInt* CI = dyn_cast<ConstantInt>(CA->getOperand(0)))
1326 if (CI->getRawValue() != '%')
1329 // Get the second character and switch on its value
1330 ConstantInt* CI = dyn_cast<ConstantInt>(CA->getOperand(1));
1331 switch (CI->getRawValue()) {
1335 dyn_cast<ConstantInt>(CA->getOperand(2))->getRawValue() != '\n')
1338 // printf("%s\n",str) -> puts(str)
1339 Function* puts_func = SLC.get_puts();
1342 std::vector<Value*> args;
1343 args.push_back(ci->getOperand(2));
1344 new CallInst(puts_func,args,ci->getName(),ci);
1345 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy,len));
1350 // printf("%c",c) -> putchar(c)
1354 Function* putchar_func = SLC.get_putchar();
1357 CastInst* cast = new CastInst(ci->getOperand(2), Type::IntTy,
1358 CI->getName()+".int", ci);
1359 new CallInst(putchar_func, cast, "", ci);
1360 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy, 1));
1366 ci->eraseFromParent();
1371 /// This LibCallOptimization will simplify calls to the "fprintf" library
1372 /// function. It looks for cases where the result of fprintf is not used and the
1373 /// operation can be reduced to something simpler.
1374 /// @brief Simplify the fprintf library function.
1375 struct FPrintFOptimization : public LibCallOptimization {
1377 /// @brief Default Constructor
1378 FPrintFOptimization() : LibCallOptimization("fprintf",
1379 "Number of 'fprintf' calls simplified") {}
1381 /// @brief Make sure that the "fprintf" function has the right prototype
1382 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC){
1383 // Just make sure this has at least 2 arguments
1384 return (f->arg_size() >= 2);
1387 /// @brief Perform the fprintf optimization.
1388 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC) {
1389 // If the call has more than 3 operands, we can't optimize it
1390 if (ci->getNumOperands() > 4 || ci->getNumOperands() <= 2)
1393 // If the result of the fprintf call is used, none of these optimizations
1395 if (!ci->use_empty())
1398 // All the optimizations depend on the length of the second argument and the
1399 // fact that it is a constant string array. Check that now
1401 ConstantArray* CA = 0;
1402 if (!getConstantStringLength(ci->getOperand(2), len, &CA))
1405 if (ci->getNumOperands() == 3) {
1406 // Make sure there's no % in the constant array
1407 for (unsigned i = 0; i < len; ++i) {
1408 if (ConstantInt* CI = dyn_cast<ConstantInt>(CA->getOperand(i))) {
1409 // Check for the null terminator
1410 if (CI->getRawValue() == '%')
1411 return false; // we found end of string
1417 // fprintf(file,fmt) -> fwrite(fmt,strlen(fmt),file)
1418 const Type* FILEptr_type = ci->getOperand(1)->getType();
1419 Function* fwrite_func = SLC.get_fwrite(FILEptr_type);
1423 // Make sure that the fprintf() and fwrite() functions both take the
1424 // same type of char pointer.
1425 if (ci->getOperand(2)->getType() !=
1426 fwrite_func->getFunctionType()->getParamType(0))
1429 std::vector<Value*> args;
1430 args.push_back(ci->getOperand(2));
1431 args.push_back(ConstantUInt::get(SLC.getIntPtrType(),len));
1432 args.push_back(ConstantUInt::get(SLC.getIntPtrType(),1));
1433 args.push_back(ci->getOperand(1));
1434 new CallInst(fwrite_func,args,ci->getName(),ci);
1435 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy,len));
1436 ci->eraseFromParent();
1440 // The remaining optimizations require the format string to be length 2
1445 // The first character has to be a %
1446 if (ConstantInt* CI = dyn_cast<ConstantInt>(CA->getOperand(0)))
1447 if (CI->getRawValue() != '%')
1450 // Get the second character and switch on its value
1451 ConstantInt* CI = dyn_cast<ConstantInt>(CA->getOperand(1));
1452 switch (CI->getRawValue()) {
1456 ConstantArray* CA = 0;
1457 if (getConstantStringLength(ci->getOperand(3), len, &CA)) {
1458 // fprintf(file,"%s",str) -> fwrite(str,strlen(str),1,file)
1459 const Type* FILEptr_type = ci->getOperand(1)->getType();
1460 Function* fwrite_func = SLC.get_fwrite(FILEptr_type);
1463 std::vector<Value*> args;
1464 args.push_back(CastToCStr(ci->getOperand(3), *ci));
1465 args.push_back(ConstantUInt::get(SLC.getIntPtrType(),len));
1466 args.push_back(ConstantUInt::get(SLC.getIntPtrType(),1));
1467 args.push_back(ci->getOperand(1));
1468 new CallInst(fwrite_func,args,ci->getName(),ci);
1469 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy,len));
1471 // fprintf(file,"%s",str) -> fputs(str,file)
1472 const Type* FILEptr_type = ci->getOperand(1)->getType();
1473 Function* fputs_func = SLC.get_fputs(FILEptr_type);
1476 std::vector<Value*> args;
1477 args.push_back(ci->getOperand(3));
1478 args.push_back(ci->getOperand(1));
1479 new CallInst(fputs_func,args,ci->getName(),ci);
1480 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy,len));
1486 // fprintf(file,"%c",c) -> fputc(c,file)
1487 const Type* FILEptr_type = ci->getOperand(1)->getType();
1488 Function* fputc_func = SLC.get_fputc(FILEptr_type);
1491 CastInst* cast = new CastInst(ci->getOperand(3), Type::IntTy,
1492 CI->getName()+".int", ci);
1493 new CallInst(fputc_func,cast,ci->getOperand(1),"",ci);
1494 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy,1));
1500 ci->eraseFromParent();
1505 /// This LibCallOptimization will simplify calls to the "sprintf" library
1506 /// function. It looks for cases where the result of sprintf is not used and the
1507 /// operation can be reduced to something simpler.
1508 /// @brief Simplify the sprintf library function.
1509 struct SPrintFOptimization : public LibCallOptimization {
1511 /// @brief Default Constructor
1512 SPrintFOptimization() : LibCallOptimization("sprintf",
1513 "Number of 'sprintf' calls simplified") {}
1515 /// @brief Make sure that the "fprintf" function has the right prototype
1516 virtual bool ValidateCalledFunction(const Function *f, SimplifyLibCalls &SLC){
1517 // Just make sure this has at least 2 arguments
1518 return (f->getReturnType() == Type::IntTy && f->arg_size() >= 2);
1521 /// @brief Perform the sprintf optimization.
1522 virtual bool OptimizeCall(CallInst *ci, SimplifyLibCalls &SLC) {
1523 // If the call has more than 3 operands, we can't optimize it
1524 if (ci->getNumOperands() > 4 || ci->getNumOperands() < 3)
1527 // All the optimizations depend on the length of the second argument and the
1528 // fact that it is a constant string array. Check that now
1530 ConstantArray* CA = 0;
1531 if (!getConstantStringLength(ci->getOperand(2), len, &CA))
1534 if (ci->getNumOperands() == 3) {
1536 // If the length is 0, we just need to store a null byte
1537 new StoreInst(ConstantInt::get(Type::SByteTy,0),ci->getOperand(1),ci);
1538 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy,0));
1539 ci->eraseFromParent();
1543 // Make sure there's no % in the constant array
1544 for (unsigned i = 0; i < len; ++i) {
1545 if (ConstantInt* CI = dyn_cast<ConstantInt>(CA->getOperand(i))) {
1546 // Check for the null terminator
1547 if (CI->getRawValue() == '%')
1548 return false; // we found a %, can't optimize
1550 return false; // initializer is not constant int, can't optimize
1554 // Increment length because we want to copy the null byte too
1557 // sprintf(str,fmt) -> llvm.memcpy(str,fmt,strlen(fmt),1)
1558 Function* memcpy_func = SLC.get_memcpy();
1561 std::vector<Value*> args;
1562 args.push_back(ci->getOperand(1));
1563 args.push_back(ci->getOperand(2));
1564 args.push_back(ConstantUInt::get(SLC.getIntPtrType(),len));
1565 args.push_back(ConstantUInt::get(Type::UIntTy,1));
1566 new CallInst(memcpy_func,args,"",ci);
1567 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy,len));
1568 ci->eraseFromParent();
1572 // The remaining optimizations require the format string to be length 2
1577 // The first character has to be a %
1578 if (ConstantInt* CI = dyn_cast<ConstantInt>(CA->getOperand(0)))
1579 if (CI->getRawValue() != '%')
1582 // Get the second character and switch on its value
1583 ConstantInt* CI = dyn_cast<ConstantInt>(CA->getOperand(1));
1584 switch (CI->getRawValue()) {
1586 // sprintf(dest,"%s",str) -> llvm.memcpy(dest, str, strlen(str)+1, 1)
1587 Function* strlen_func = SLC.get_strlen();
1588 Function* memcpy_func = SLC.get_memcpy();
1589 if (!strlen_func || !memcpy_func)
1592 Value *Len = new CallInst(strlen_func, CastToCStr(ci->getOperand(3), *ci),
1593 ci->getOperand(3)->getName()+".len", ci);
1594 Value *Len1 = BinaryOperator::createAdd(Len,
1595 ConstantInt::get(Len->getType(), 1),
1596 Len->getName()+"1", ci);
1597 if (Len1->getType() != SLC.getIntPtrType())
1598 Len1 = new CastInst(Len1, SLC.getIntPtrType(), Len1->getName(), ci);
1599 std::vector<Value*> args;
1600 args.push_back(CastToCStr(ci->getOperand(1), *ci));
1601 args.push_back(CastToCStr(ci->getOperand(3), *ci));
1602 args.push_back(Len1);
1603 args.push_back(ConstantUInt::get(Type::UIntTy,1));
1604 new CallInst(memcpy_func, args, "", ci);
1606 // The strlen result is the unincremented number of bytes in the string.
1607 if (!ci->use_empty()) {
1608 if (Len->getType() != ci->getType())
1609 Len = new CastInst(Len, ci->getType(), Len->getName(), ci);
1610 ci->replaceAllUsesWith(Len);
1612 ci->eraseFromParent();
1616 // sprintf(dest,"%c",chr) -> store chr, dest
1617 CastInst* cast = new CastInst(ci->getOperand(3),Type::SByteTy,"char",ci);
1618 new StoreInst(cast, ci->getOperand(1), ci);
1619 GetElementPtrInst* gep = new GetElementPtrInst(ci->getOperand(1),
1620 ConstantUInt::get(Type::UIntTy,1),ci->getOperand(1)->getName()+".end",
1622 new StoreInst(ConstantInt::get(Type::SByteTy,0),gep,ci);
1623 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy,1));
1624 ci->eraseFromParent();
1632 /// This LibCallOptimization will simplify calls to the "fputs" library
1633 /// function. It looks for cases where the result of fputs is not used and the
1634 /// operation can be reduced to something simpler.
1635 /// @brief Simplify the puts library function.
1636 struct PutsOptimization : public LibCallOptimization {
1638 /// @brief Default Constructor
1639 PutsOptimization() : LibCallOptimization("fputs",
1640 "Number of 'fputs' calls simplified") {}
1642 /// @brief Make sure that the "fputs" function has the right prototype
1643 virtual bool ValidateCalledFunction(const Function *F, SimplifyLibCalls &SLC){
1644 // Just make sure this has 2 arguments
1645 return F->arg_size() == 2;
1648 /// @brief Perform the fputs optimization.
1649 virtual bool OptimizeCall(CallInst* ci, SimplifyLibCalls& SLC) {
1650 // If the result is used, none of these optimizations work
1651 if (!ci->use_empty())
1654 // All the optimizations depend on the length of the first argument and the
1655 // fact that it is a constant string array. Check that now
1657 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){
1709 // Just make sure this has 1 argument
1710 return (f->arg_size() == 1);
1713 /// @brief Perform the toascii optimization.
1714 virtual bool OptimizeCall(CallInst *ci, SimplifyLibCalls &SLC) {
1715 if (ConstantInt* CI = dyn_cast<ConstantInt>(ci->getOperand(1))) {
1716 // isdigit(c) -> 0 or 1, if 'c' is constant
1717 uint64_t val = CI->getRawValue();
1718 if (val >= '0' && val <='9')
1719 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy,1));
1721 ci->replaceAllUsesWith(ConstantSInt::get(Type::IntTy,0));
1722 ci->eraseFromParent();
1726 // isdigit(c) -> (unsigned)c - '0' <= 9
1728 new CastInst(ci->getOperand(1),Type::UIntTy,
1729 ci->getOperand(1)->getName()+".uint",ci);
1730 BinaryOperator* sub_inst = BinaryOperator::createSub(cast,
1731 ConstantUInt::get(Type::UIntTy,0x30),
1732 ci->getOperand(1)->getName()+".sub",ci);
1733 SetCondInst* setcond_inst = new SetCondInst(Instruction::SetLE,sub_inst,
1734 ConstantUInt::get(Type::UIntTy,9),
1735 ci->getOperand(1)->getName()+".cmp",ci);
1737 new CastInst(setcond_inst,Type::IntTy,
1738 ci->getOperand(1)->getName()+".isdigit",ci);
1739 ci->replaceAllUsesWith(c2);
1740 ci->eraseFromParent();
1745 struct isasciiOptimization : public LibCallOptimization {
1747 isasciiOptimization()
1748 : LibCallOptimization("isascii", "Number of 'isascii' calls simplified") {}
1750 virtual bool ValidateCalledFunction(const Function *F, SimplifyLibCalls &SLC){
1751 return F->arg_size() == 1 && F->arg_begin()->getType()->isInteger() &&
1752 F->getReturnType()->isInteger();
1755 /// @brief Perform the isascii optimization.
1756 virtual bool OptimizeCall(CallInst *CI, SimplifyLibCalls &SLC) {
1757 // isascii(c) -> (unsigned)c < 128
1758 Value *V = CI->getOperand(1);
1759 if (V->getType()->isSigned())
1760 V = new CastInst(V, V->getType()->getUnsignedVersion(), V->getName(), CI);
1761 Value *Cmp = BinaryOperator::createSetLT(V, ConstantUInt::get(V->getType(),
1763 V->getName()+".isascii", CI);
1764 if (Cmp->getType() != CI->getType())
1765 Cmp = new CastInst(Cmp, CI->getType(), Cmp->getName(), CI);
1766 CI->replaceAllUsesWith(Cmp);
1767 CI->eraseFromParent();
1773 /// This LibCallOptimization will simplify calls to the "toascii" library
1774 /// function. It simply does the corresponding and operation to restrict the
1775 /// range of values to the ASCII character set (0-127).
1776 /// @brief Simplify the toascii library function.
1777 struct ToAsciiOptimization : public LibCallOptimization {
1779 /// @brief Default Constructor
1780 ToAsciiOptimization() : LibCallOptimization("toascii",
1781 "Number of 'toascii' calls simplified") {}
1783 /// @brief Make sure that the "fputs" function has the right prototype
1784 virtual bool ValidateCalledFunction(const Function* f, SimplifyLibCalls& SLC){
1785 // Just make sure this has 2 arguments
1786 return (f->arg_size() == 1);
1789 /// @brief Perform the toascii optimization.
1790 virtual bool OptimizeCall(CallInst *ci, SimplifyLibCalls &SLC) {
1791 // toascii(c) -> (c & 0x7f)
1792 Value* chr = ci->getOperand(1);
1793 BinaryOperator* and_inst = BinaryOperator::createAnd(chr,
1794 ConstantInt::get(chr->getType(),0x7F),ci->getName()+".toascii",ci);
1795 ci->replaceAllUsesWith(and_inst);
1796 ci->eraseFromParent();
1801 /// This LibCallOptimization will simplify calls to the "ffs" library
1802 /// calls which find the first set bit in an int, long, or long long. The
1803 /// optimization is to compute the result at compile time if the argument is
1805 /// @brief Simplify the ffs library function.
1806 struct FFSOptimization : public LibCallOptimization {
1808 /// @brief Subclass Constructor
1809 FFSOptimization(const char* funcName, const char* description)
1810 : LibCallOptimization(funcName, description) {}
1813 /// @brief Default Constructor
1814 FFSOptimization() : LibCallOptimization("ffs",
1815 "Number of 'ffs' calls simplified") {}
1817 /// @brief Make sure that the "ffs" function has the right prototype
1818 virtual bool ValidateCalledFunction(const Function *F, SimplifyLibCalls &SLC){
1819 // Just make sure this has 2 arguments
1820 return F->arg_size() == 1 && F->getReturnType() == Type::IntTy;
1823 /// @brief Perform the ffs optimization.
1824 virtual bool OptimizeCall(CallInst *TheCall, SimplifyLibCalls &SLC) {
1825 if (ConstantInt *CI = dyn_cast<ConstantInt>(TheCall->getOperand(1))) {
1826 // ffs(cnst) -> bit#
1827 // ffsl(cnst) -> bit#
1828 // ffsll(cnst) -> bit#
1829 uint64_t val = CI->getRawValue();
1833 while ((val & 1) == 0) {
1838 TheCall->replaceAllUsesWith(ConstantSInt::get(Type::IntTy, result));
1839 TheCall->eraseFromParent();
1843 // ffs(x) -> x == 0 ? 0 : llvm.cttz(x)+1
1844 // ffsl(x) -> x == 0 ? 0 : llvm.cttz(x)+1
1845 // ffsll(x) -> x == 0 ? 0 : llvm.cttz(x)+1
1846 const Type *ArgType = TheCall->getOperand(1)->getType();
1847 ArgType = ArgType->getUnsignedVersion();
1848 const char *CTTZName;
1849 switch (ArgType->getTypeID()) {
1850 default: assert(0 && "Unknown unsigned type!");
1851 case Type::UByteTyID : CTTZName = "llvm.cttz.i8" ; break;
1852 case Type::UShortTyID: CTTZName = "llvm.cttz.i16"; break;
1853 case Type::UIntTyID : CTTZName = "llvm.cttz.i32"; break;
1854 case Type::ULongTyID : CTTZName = "llvm.cttz.i64"; break;
1857 Function *F = SLC.getModule()->getOrInsertFunction(CTTZName, ArgType,
1859 Value *V = new CastInst(TheCall->getOperand(1), ArgType, "tmp", TheCall);
1860 Value *V2 = new CallInst(F, V, "tmp", TheCall);
1861 V2 = new CastInst(V2, Type::IntTy, "tmp", TheCall);
1862 V2 = BinaryOperator::createAdd(V2, ConstantSInt::get(Type::IntTy, 1),
1865 BinaryOperator::createSetEQ(V, Constant::getNullValue(V->getType()),
1867 V2 = new SelectInst(Cond, ConstantInt::get(Type::IntTy, 0), V2,
1868 TheCall->getName(), TheCall);
1869 TheCall->replaceAllUsesWith(V2);
1870 TheCall->eraseFromParent();
1875 /// This LibCallOptimization will simplify calls to the "ffsl" library
1876 /// calls. It simply uses FFSOptimization for which the transformation is
1878 /// @brief Simplify the ffsl library function.
1879 struct FFSLOptimization : public FFSOptimization {
1881 /// @brief Default Constructor
1882 FFSLOptimization() : FFSOptimization("ffsl",
1883 "Number of 'ffsl' calls simplified") {}
1887 /// This LibCallOptimization will simplify calls to the "ffsll" library
1888 /// calls. It simply uses FFSOptimization for which the transformation is
1890 /// @brief Simplify the ffsl library function.
1891 struct FFSLLOptimization : public FFSOptimization {
1893 /// @brief Default Constructor
1894 FFSLLOptimization() : FFSOptimization("ffsll",
1895 "Number of 'ffsll' calls simplified") {}
1899 /// This optimizes unary functions that take and return doubles.
1900 struct UnaryDoubleFPOptimizer : public LibCallOptimization {
1901 UnaryDoubleFPOptimizer(const char *Fn, const char *Desc)
1902 : LibCallOptimization(Fn, Desc) {}
1904 // Make sure that this function has the right prototype
1905 virtual bool ValidateCalledFunction(const Function *F, SimplifyLibCalls &SLC){
1906 return F->arg_size() == 1 && F->arg_begin()->getType() == Type::DoubleTy &&
1907 F->getReturnType() == Type::DoubleTy;
1910 /// ShrinkFunctionToFloatVersion - If the input to this function is really a
1911 /// float, strength reduce this to a float version of the function,
1912 /// e.g. floor((double)FLT) -> (double)floorf(FLT). This can only be called
1913 /// when the target supports the destination function and where there can be
1914 /// no precision loss.
1915 static bool ShrinkFunctionToFloatVersion(CallInst *CI, SimplifyLibCalls &SLC,
1916 Function *(SimplifyLibCalls::*FP)()){
1917 if (CastInst *Cast = dyn_cast<CastInst>(CI->getOperand(1)))
1918 if (Cast->getOperand(0)->getType() == Type::FloatTy) {
1919 Value *New = new CallInst((SLC.*FP)(), Cast->getOperand(0),
1921 New = new CastInst(New, Type::DoubleTy, CI->getName(), CI);
1922 CI->replaceAllUsesWith(New);
1923 CI->eraseFromParent();
1924 if (Cast->use_empty())
1925 Cast->eraseFromParent();
1933 struct FloorOptimization : public UnaryDoubleFPOptimizer {
1935 : UnaryDoubleFPOptimizer("floor", "Number of 'floor' calls simplified") {}
1937 virtual bool OptimizeCall(CallInst *CI, SimplifyLibCalls &SLC) {
1939 // If this is a float argument passed in, convert to floorf.
1940 if (ShrinkFunctionToFloatVersion(CI, SLC, &SimplifyLibCalls::get_floorf))
1943 return false; // opt failed
1947 struct CeilOptimization : public UnaryDoubleFPOptimizer {
1949 : UnaryDoubleFPOptimizer("ceil", "Number of 'ceil' calls simplified") {}
1951 virtual bool OptimizeCall(CallInst *CI, SimplifyLibCalls &SLC) {
1953 // If this is a float argument passed in, convert to ceilf.
1954 if (ShrinkFunctionToFloatVersion(CI, SLC, &SimplifyLibCalls::get_ceilf))
1957 return false; // opt failed
1961 struct RoundOptimization : public UnaryDoubleFPOptimizer {
1963 : UnaryDoubleFPOptimizer("round", "Number of 'round' calls simplified") {}
1965 virtual bool OptimizeCall(CallInst *CI, SimplifyLibCalls &SLC) {
1967 // If this is a float argument passed in, convert to roundf.
1968 if (ShrinkFunctionToFloatVersion(CI, SLC, &SimplifyLibCalls::get_roundf))
1971 return false; // opt failed
1975 struct RintOptimization : public UnaryDoubleFPOptimizer {
1977 : UnaryDoubleFPOptimizer("rint", "Number of 'rint' calls simplified") {}
1979 virtual bool OptimizeCall(CallInst *CI, SimplifyLibCalls &SLC) {
1981 // If this is a float argument passed in, convert to rintf.
1982 if (ShrinkFunctionToFloatVersion(CI, SLC, &SimplifyLibCalls::get_rintf))
1985 return false; // opt failed
1989 struct NearByIntOptimization : public UnaryDoubleFPOptimizer {
1990 NearByIntOptimization()
1991 : UnaryDoubleFPOptimizer("nearbyint",
1992 "Number of 'nearbyint' calls simplified") {}
1994 virtual bool OptimizeCall(CallInst *CI, SimplifyLibCalls &SLC) {
1995 #ifdef HAVE_NEARBYINTF
1996 // If this is a float argument passed in, convert to nearbyintf.
1997 if (ShrinkFunctionToFloatVersion(CI, SLC,&SimplifyLibCalls::get_nearbyintf))
2000 return false; // opt failed
2002 } NearByIntOptimizer;
2004 /// A function to compute the length of a null-terminated constant array of
2005 /// integers. This function can't rely on the size of the constant array
2006 /// because there could be a null terminator in the middle of the array.
2007 /// We also have to bail out if we find a non-integer constant initializer
2008 /// of one of the elements or if there is no null-terminator. The logic
2009 /// below checks each of these conditions and will return true only if all
2010 /// conditions are met. In that case, the \p len parameter is set to the length
2011 /// of the null-terminated string. If false is returned, the conditions were
2012 /// not met and len is set to 0.
2013 /// @brief Get the length of a constant string (null-terminated array).
2014 bool getConstantStringLength(Value *V, uint64_t &len, ConstantArray **CA) {
2015 assert(V != 0 && "Invalid args to getConstantStringLength");
2016 len = 0; // make sure we initialize this
2018 // If the value is not a GEP instruction nor a constant expression with a
2019 // GEP instruction, then return false because ConstantArray can't occur
2021 if (GetElementPtrInst* GEPI = dyn_cast<GetElementPtrInst>(V))
2023 else if (ConstantExpr* CE = dyn_cast<ConstantExpr>(V))
2024 if (CE->getOpcode() == Instruction::GetElementPtr)
2031 // Make sure the GEP has exactly three arguments.
2032 if (GEP->getNumOperands() != 3)
2035 // Check to make sure that the first operand of the GEP is an integer and
2036 // has value 0 so that we are sure we're indexing into the initializer.
2037 if (ConstantInt* op1 = dyn_cast<ConstantInt>(GEP->getOperand(1))) {
2038 if (!op1->isNullValue())
2043 // Ensure that the second operand is a ConstantInt. If it isn't then this
2044 // GEP is wonky and we're not really sure what were referencing into and
2045 // better of not optimizing it. While we're at it, get the second index
2046 // value. We'll need this later for indexing the ConstantArray.
2047 uint64_t start_idx = 0;
2048 if (ConstantInt* CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
2049 start_idx = CI->getRawValue();
2053 // The GEP instruction, constant or instruction, must reference a global
2054 // variable that is a constant and is initialized. The referenced constant
2055 // initializer is the array that we'll use for optimization.
2056 GlobalVariable* GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
2057 if (!GV || !GV->isConstant() || !GV->hasInitializer())
2060 // Get the initializer.
2061 Constant* INTLZR = GV->getInitializer();
2063 // Handle the ConstantAggregateZero case
2064 if (ConstantAggregateZero *CAZ = dyn_cast<ConstantAggregateZero>(INTLZR)) {
2065 // This is a degenerate case. The initializer is constant zero so the
2066 // length of the string must be zero.
2071 // Must be a Constant Array
2072 ConstantArray* A = dyn_cast<ConstantArray>(INTLZR);
2076 // Get the number of elements in the array
2077 uint64_t max_elems = A->getType()->getNumElements();
2079 // Traverse the constant array from start_idx (derived above) which is
2080 // the place the GEP refers to in the array.
2081 for (len = start_idx; len < max_elems; len++) {
2082 if (ConstantInt *CI = dyn_cast<ConstantInt>(A->getOperand(len))) {
2083 // Check for the null terminator
2084 if (CI->isNullValue())
2085 break; // we found end of string
2087 return false; // This array isn't suitable, non-int initializer
2090 if (len >= max_elems)
2091 return false; // This array isn't null terminated
2093 // Subtract out the initial value from the length
2097 return true; // success!
2100 /// CastToCStr - Return V if it is an sbyte*, otherwise cast it to sbyte*,
2101 /// inserting the cast before IP, and return the cast.
2102 /// @brief Cast a value to a "C" string.
2103 Value *CastToCStr(Value *V, Instruction &IP) {
2104 const Type *SBPTy = PointerType::get(Type::SByteTy);
2105 if (V->getType() != SBPTy)
2106 return new CastInst(V, SBPTy, V->getName(), &IP);
2111 // Additional cases that we need to add to this file:
2114 // * cbrt(expN(X)) -> expN(x/3)
2115 // * cbrt(sqrt(x)) -> pow(x,1/6)
2116 // * cbrt(sqrt(x)) -> pow(x,1/9)
2119 // * cos(-x) -> cos(x)
2122 // * exp(log(x)) -> x
2125 // * log(exp(x)) -> x
2126 // * log(x**y) -> y*log(x)
2127 // * log(exp(y)) -> y*log(e)
2128 // * log(exp2(y)) -> y*log(2)
2129 // * log(exp10(y)) -> y*log(10)
2130 // * log(sqrt(x)) -> 0.5*log(x)
2131 // * log(pow(x,y)) -> y*log(x)
2133 // lround, lroundf, lroundl:
2134 // * lround(cnst) -> cnst'
2137 // * memcmp(x,y,l) -> cnst
2138 // (if all arguments are constant and strlen(x) <= l and strlen(y) <= l)
2141 // * memmove(d,s,l,a) -> memcpy(d,s,l,a)
2142 // (if s is a global constant array)
2145 // * pow(exp(x),y) -> exp(x*y)
2146 // * pow(sqrt(x),y) -> pow(x,y*0.5)
2147 // * pow(pow(x,y),z)-> pow(x,y*z)
2150 // * puts("") -> fputc("\n",stdout) (how do we get "stdout"?)
2152 // round, roundf, roundl:
2153 // * round(cnst) -> cnst'
2156 // * signbit(cnst) -> cnst'
2157 // * signbit(nncst) -> 0 (if pstv is a non-negative constant)
2159 // sqrt, sqrtf, sqrtl:
2160 // * sqrt(expN(x)) -> expN(x*0.5)
2161 // * sqrt(Nroot(x)) -> pow(x,1/(2*N))
2162 // * sqrt(pow(x,y)) -> pow(|x|,y*0.5)
2165 // * stpcpy(str, "literal") ->
2166 // llvm.memcpy(str,"literal",strlen("literal")+1,1)
2168 // * strrchr(s,c) -> reverse_offset_of_in(c,s)
2169 // (if c is a constant integer and s is a constant string)
2170 // * strrchr(s1,0) -> strchr(s1,0)
2173 // * strncat(x,y,0) -> x
2174 // * strncat(x,y,0) -> x (if strlen(y) = 0)
2175 // * strncat(x,y,l) -> strcat(x,y) (if y and l are constants an l > strlen(y))
2178 // * strncpy(d,s,0) -> d
2179 // * strncpy(d,s,l) -> memcpy(d,s,l,1)
2180 // (if s and l are constants)
2183 // * strpbrk(s,a) -> offset_in_for(s,a)
2184 // (if s and a are both constant strings)
2185 // * strpbrk(s,"") -> 0
2186 // * strpbrk(s,a) -> strchr(s,a[0]) (if a is constant string of length 1)
2189 // * strspn(s,a) -> const_int (if both args are constant)
2190 // * strspn("",a) -> 0
2191 // * strspn(s,"") -> 0
2192 // * strcspn(s,a) -> const_int (if both args are constant)
2193 // * strcspn("",a) -> 0
2194 // * strcspn(s,"") -> strlen(a)
2197 // * strstr(x,x) -> x
2198 // * strstr(s1,s2) -> offset_of_s2_in(s1)
2199 // (if s1 and s2 are constant strings)
2202 // * tan(atan(x)) -> x
2204 // trunc, truncf, truncl:
2205 // * trunc(cnst) -> cnst'