Implement count leading zeros (ctlz), count trailing zeros (cttz), and count
[oota-llvm.git] / lib / CodeGen / IntrinsicLowering.cpp
1 //===-- IntrinsicLowering.cpp - Intrinsic Lowering default implementation -===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements the default intrinsic lowering implementation.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "llvm/CodeGen/IntrinsicLowering.h"
15 #include "llvm/Constants.h"
16 #include "llvm/DerivedTypes.h"
17 #include "llvm/Module.h"
18 #include "llvm/Instructions.h"
19 #include "llvm/Type.h"
20 #include <iostream>
21
22 using namespace llvm;
23
24 template <class ArgIt>
25 static Function *EnsureFunctionExists(Module &M, const char *Name,
26                                       ArgIt ArgBegin, ArgIt ArgEnd,
27                                       const Type *RetTy) {
28   if (Function *F = M.getNamedFunction(Name)) return F;
29   // It doesn't already exist in the program, insert a new definition now.
30   std::vector<const Type *> ParamTys;
31   for (ArgIt I = ArgBegin; I != ArgEnd; ++I)
32     ParamTys.push_back(I->getType());
33   return M.getOrInsertFunction(Name, FunctionType::get(RetTy, ParamTys, false));
34 }
35
36 /// ReplaceCallWith - This function is used when we want to lower an intrinsic
37 /// call to a call of an external function.  This handles hard cases such as
38 /// when there was already a prototype for the external function, and if that
39 /// prototype doesn't match the arguments we expect to pass in.
40 template <class ArgIt>
41 static CallInst *ReplaceCallWith(const char *NewFn, CallInst *CI,
42                                  ArgIt ArgBegin, ArgIt ArgEnd,
43                                  const Type *RetTy, Function *&FCache) {
44   if (!FCache) {
45     // If we haven't already looked up this function, check to see if the
46     // program already contains a function with this name.
47     Module *M = CI->getParent()->getParent()->getParent();
48     FCache = M->getNamedFunction(NewFn);
49     if (!FCache) {
50       // It doesn't already exist in the program, insert a new definition now.
51       std::vector<const Type *> ParamTys;
52       for (ArgIt I = ArgBegin; I != ArgEnd; ++I)
53         ParamTys.push_back((*I)->getType());
54       FCache = M->getOrInsertFunction(NewFn,
55                                      FunctionType::get(RetTy, ParamTys, false));
56     }
57    }
58
59   const FunctionType *FT = FCache->getFunctionType();
60   std::vector<Value*> Operands;
61   unsigned ArgNo = 0;
62   for (ArgIt I = ArgBegin; I != ArgEnd && ArgNo != FT->getNumParams();
63        ++I, ++ArgNo) {
64     Value *Arg = *I;
65     if (Arg->getType() != FT->getParamType(ArgNo))
66       Arg = new CastInst(Arg, FT->getParamType(ArgNo), Arg->getName(), CI);
67     Operands.push_back(Arg);
68   }
69   // Pass nulls into any additional arguments...
70   for (; ArgNo != FT->getNumParams(); ++ArgNo)
71     Operands.push_back(Constant::getNullValue(FT->getParamType(ArgNo)));
72
73   std::string Name = CI->getName(); CI->setName("");
74   if (FT->getReturnType() == Type::VoidTy) Name.clear();
75   CallInst *NewCI = new CallInst(FCache, Operands, Name, CI);
76   if (!CI->use_empty()) {
77     Value *V = NewCI;
78     if (CI->getType() != NewCI->getType())
79       V = new CastInst(NewCI, CI->getType(), Name, CI);
80     CI->replaceAllUsesWith(V);
81   }
82   return NewCI;
83 }
84
85 void DefaultIntrinsicLowering::AddPrototypes(Module &M) {
86   for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
87     if (I->isExternal() && !I->use_empty())
88       switch (I->getIntrinsicID()) {
89       default: break;
90       case Intrinsic::setjmp:
91         EnsureFunctionExists(M, "setjmp", I->arg_begin(), I->arg_end(), Type::IntTy);
92         break;
93       case Intrinsic::longjmp:
94         EnsureFunctionExists(M, "longjmp", I->arg_begin(), I->arg_end(),Type::VoidTy);
95         break;
96       case Intrinsic::siglongjmp:
97         EnsureFunctionExists(M, "abort", I->arg_end(), I->arg_end(), Type::VoidTy);
98         break;
99       case Intrinsic::memcpy:
100         EnsureFunctionExists(M, "memcpy", I->arg_begin(), --I->arg_end(),
101                              I->arg_begin()->getType());
102         break;
103       case Intrinsic::memmove:
104         EnsureFunctionExists(M, "memmove", I->arg_begin(), --I->arg_end(),
105                              I->arg_begin()->getType());
106         break;
107       case Intrinsic::memset:
108         EnsureFunctionExists(M, "memset", I->arg_begin(), --I->arg_end(),
109                              I->arg_begin()->getType());
110         break;
111       case Intrinsic::isunordered:
112         EnsureFunctionExists(M, "isunordered", I->arg_begin(), I->arg_end(), Type::BoolTy);
113         break;
114       case Intrinsic::sqrt:
115         if(I->arg_begin()->getType() == Type::FloatTy)
116           EnsureFunctionExists(M, "sqrtf", I->arg_begin(), I->arg_end(), Type::FloatTy);
117         else
118           EnsureFunctionExists(M, "sqrt", I->arg_begin(), I->arg_end(), Type::DoubleTy);
119         break;
120       }
121 }
122
123 void DefaultIntrinsicLowering::LowerIntrinsicCall(CallInst *CI) {
124   Function *Callee = CI->getCalledFunction();
125   assert(Callee && "Cannot lower an indirect call!");
126
127   switch (Callee->getIntrinsicID()) {
128   case Intrinsic::not_intrinsic:
129     std::cerr << "Cannot lower a call to a non-intrinsic function '"
130               << Callee->getName() << "'!\n";
131     abort();
132   default:
133     std::cerr << "Error: Code generator does not support intrinsic function '"
134               << Callee->getName() << "'!\n";
135     abort();
136
137     // The setjmp/longjmp intrinsics should only exist in the code if it was
138     // never optimized (ie, right out of the CFE), or if it has been hacked on
139     // by the lowerinvoke pass.  In both cases, the right thing to do is to
140     // convert the call to an explicit setjmp or longjmp call.
141   case Intrinsic::setjmp: {
142     static Function *SetjmpFCache = 0;
143     Value *V = ReplaceCallWith("setjmp", CI, CI->op_begin()+1, CI->op_end(),
144                                Type::IntTy, SetjmpFCache);
145     if (CI->getType() != Type::VoidTy)
146       CI->replaceAllUsesWith(V);
147     break;
148   }
149   case Intrinsic::sigsetjmp:
150      if (CI->getType() != Type::VoidTy)
151        CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
152      break;
153
154   case Intrinsic::longjmp: {
155     static Function *LongjmpFCache = 0;
156     ReplaceCallWith("longjmp", CI, CI->op_begin()+1, CI->op_end(),
157                     Type::VoidTy, LongjmpFCache);
158     break;
159   }
160
161   case Intrinsic::siglongjmp: {
162     // Insert the call to abort
163     static Function *AbortFCache = 0;
164     ReplaceCallWith("abort", CI, CI->op_end(), CI->op_end(), Type::VoidTy,
165                     AbortFCache);
166     break;
167   }
168   case Intrinsic::ctpop: {
169     Value *Src = CI->getOperand(1);
170     switch (CI->getOperand(0)->getType()->getTypeID())
171     {
172     case Type::SByteTyID:
173     case Type::UByteTyID:
174       {
175         Value* SA = ConstantUInt::get(Type::UByteTy, 1);
176         Value* MA = ConstantUInt::get(Type::UIntTy, 0x55);
177         Src = BinaryOperator::create(Instruction::Add,
178                 BinaryOperator::create(Instruction::And, Src, MA),
179                 BinaryOperator::create(Instruction::And, 
180                                        new ShiftInst(Instruction::Shr, Src, SA),
181                                        MA));
182         SA = ConstantUInt::get(Type::UByteTy, 2);
183         MA = ConstantUInt::get(Type::UIntTy, 0x33);
184         Src = BinaryOperator::create(Instruction::Add,
185                 BinaryOperator::create(Instruction::And, Src, MA),
186                 BinaryOperator::create(Instruction::And,
187                                        new ShiftInst(Instruction::Shr, Src, SA),
188                                        MA));
189         SA = ConstantUInt::get(Type::UByteTy, 4);
190         MA = ConstantUInt::get(Type::UIntTy, 0x0F);
191         Src = BinaryOperator::create(Instruction::Add,
192                 BinaryOperator::create(Instruction::And, Src, MA),
193                 BinaryOperator::create(Instruction::And, 
194                                        new ShiftInst(Instruction::Shr, Src, SA),
195                                        MA), "ctpop");
196       }
197       break;
198     case Type::ShortTyID:
199     case Type::UShortTyID:
200       {
201         Value* SA = ConstantUInt::get(Type::UByteTy, 1);
202         Value* MA = ConstantUInt::get(Type::UIntTy, 0x5555);
203         Src = BinaryOperator::create(Instruction::Add,
204                 BinaryOperator::create(Instruction::And, Src, MA),
205                 BinaryOperator::create(Instruction::And, 
206                                        new ShiftInst(Instruction::Shr, Src, SA),
207                                        MA));
208         SA = ConstantUInt::get(Type::UByteTy, 2);
209         MA = ConstantUInt::get(Type::UIntTy, 0x3333);
210         Src = BinaryOperator::create(Instruction::Add,
211                 BinaryOperator::create(Instruction::And, Src, MA),
212                 BinaryOperator::create(Instruction::And, 
213                                        new ShiftInst(Instruction::Shr, Src, SA),
214                                        MA));
215         SA = ConstantUInt::get(Type::UByteTy, 4);
216         MA = ConstantUInt::get(Type::UIntTy, 0x0F0F);
217         Src = BinaryOperator::create(Instruction::Add,
218                 BinaryOperator::create(Instruction::And, Src, MA),
219                 BinaryOperator::create(Instruction::And, 
220                                        new ShiftInst(Instruction::Shr, Src, SA),
221                                        MA));
222         SA = ConstantUInt::get(Type::UByteTy, 8);
223         MA = ConstantUInt::get(Type::UIntTy, 0x00FF);
224         Src = BinaryOperator::create(Instruction::Add,
225                 BinaryOperator::create(Instruction::And, Src, MA),
226                 BinaryOperator::create(Instruction::And, 
227                                        new ShiftInst(Instruction::Shr, Src, SA),
228                                        MA), "ctpop");
229
230       }
231       break;
232     case Type::IntTyID:
233     case Type::UIntTyID:
234       {
235         Value* SA = ConstantUInt::get(Type::UByteTy, 1);
236         Value* MA = ConstantUInt::get(Type::UIntTy, 0x55555555);
237         Src = BinaryOperator::create(Instruction::Add,
238                 BinaryOperator::create(Instruction::And, Src, MA),
239                 BinaryOperator::create(Instruction::And, 
240                                        new ShiftInst(Instruction::Shr, Src, SA),
241                                        MA));
242         SA = ConstantUInt::get(Type::UByteTy, 2);
243         MA = ConstantUInt::get(Type::UIntTy, 0x33333333);
244         Src = BinaryOperator::create(Instruction::Add,
245                 BinaryOperator::create(Instruction::And, Src, MA),
246                 BinaryOperator::create(Instruction::And, 
247                                        new ShiftInst(Instruction::Shr, Src, SA),
248                                        MA));
249         SA = ConstantUInt::get(Type::UByteTy, 4);
250         MA = ConstantUInt::get(Type::UIntTy, 0x0F0F0F0F);
251         Src = BinaryOperator::create(Instruction::Add,
252                 BinaryOperator::create(Instruction::And, Src, MA),
253                 BinaryOperator::create(Instruction::And, 
254                                        new ShiftInst(Instruction::Shr, Src, SA),
255                                        MA));
256         SA = ConstantUInt::get(Type::UByteTy, 8);
257         MA = ConstantUInt::get(Type::UIntTy, 0x00FF00FF);
258         Src = BinaryOperator::create(Instruction::Add,
259                 BinaryOperator::create(Instruction::And, Src, MA),
260                 BinaryOperator::create(Instruction::And, 
261                                        new ShiftInst(Instruction::Shr, Src, SA),
262                                        MA));
263         SA = ConstantUInt::get(Type::UByteTy, 8);
264         MA = ConstantUInt::get(Type::UIntTy, 0x0000FFFF);
265         Src = BinaryOperator::create(Instruction::Add,
266                 BinaryOperator::create(Instruction::And, Src, MA),
267                 BinaryOperator::create(Instruction::And, 
268                                        new ShiftInst(Instruction::Shr, Src, SA),
269                                        MA), "ctpop");
270       }
271       break;
272     case Type::LongTyID:
273     case Type::ULongTyID:
274       {
275         Value* SA = ConstantUInt::get(Type::UByteTy, 1);
276         Value* MA = ConstantUInt::get(Type::ULongTy, 0x5555555555555555ULL);
277         Src = BinaryOperator::create(Instruction::Add,
278                 BinaryOperator::create(Instruction::And, Src, MA),
279                 BinaryOperator::create(Instruction::And, 
280                                        new ShiftInst(Instruction::Shr, Src, SA),
281                                        MA));
282         SA = ConstantUInt::get(Type::UByteTy, 2);
283         MA = ConstantUInt::get(Type::ULongTy, 0x3333333333333333ULL);
284         Src = BinaryOperator::create(Instruction::Add,
285                 BinaryOperator::create(Instruction::And, Src, MA),
286                 BinaryOperator::create(Instruction::And, 
287                                        new ShiftInst(Instruction::Shr, Src, SA),
288                                        MA));
289         SA = ConstantUInt::get(Type::UByteTy, 4);
290         MA = ConstantUInt::get(Type::ULongTy, 0x0F0F0F0F0F0F0F0FULL);
291         Src = BinaryOperator::create(Instruction::Add,
292                 BinaryOperator::create(Instruction::And, Src, MA),
293                 BinaryOperator::create(Instruction::And, 
294                                        new ShiftInst(Instruction::Shr, Src, SA),
295                                        MA));
296         SA = ConstantUInt::get(Type::UByteTy, 8);
297         MA = ConstantUInt::get(Type::ULongTy, 0x00FF00FF00FF00FFULL);
298         Src = BinaryOperator::create(Instruction::Add,
299                 BinaryOperator::create(Instruction::And, Src, MA),
300                 BinaryOperator::create(Instruction::And, 
301                                        new ShiftInst(Instruction::Shr, Src, SA),
302                                        MA));
303         SA = ConstantUInt::get(Type::UByteTy, 16);
304         MA = ConstantUInt::get(Type::ULongTy, 0x00000000FFFFFFFFULL);
305         Src = BinaryOperator::create(Instruction::Add,
306                 BinaryOperator::create(Instruction::And, Src, MA),
307                 BinaryOperator::create(Instruction::And, 
308                                        new ShiftInst(Instruction::Shr, Src, SA),
309                                        MA), "ctpop");
310       }
311       break;
312     default:
313       abort();
314     }
315         
316     CI->replaceAllUsesWith(Src);
317     break;
318   }
319   case Intrinsic::ctlz: {
320     Value *Src = CI->getOperand(1);
321     Value* SA;
322     switch (CI->getOperand(0)->getType()->getTypeID())
323     {
324     case Type::LongTyID:
325     case Type::ULongTyID:
326       SA = ConstantUInt::get(Type::UByteTy, 32);
327       Src = BinaryOperator::create(Instruction::Or, Src, new ShiftInst(Instruction::Shr, Src, SA));
328     case Type::IntTyID:
329     case Type::UIntTyID:
330       SA = ConstantUInt::get(Type::UByteTy, 16);
331       Src = BinaryOperator::create(Instruction::Or, Src, new ShiftInst(Instruction::Shr, Src, SA));
332     case Type::ShortTyID:
333     case Type::UShortTyID:
334       SA = ConstantUInt::get(Type::UByteTy, 8);
335       Src = BinaryOperator::create(Instruction::Or, Src, new ShiftInst(Instruction::Shr, Src, SA));
336     default:
337       SA = ConstantUInt::get(Type::UByteTy, 1);
338       Src = BinaryOperator::create(Instruction::Or, Src, new ShiftInst(Instruction::Shr, Src, SA));
339       SA = ConstantUInt::get(Type::UByteTy, 2);
340       Src = BinaryOperator::create(Instruction::Or, Src, new ShiftInst(Instruction::Shr, Src, SA));
341       SA = ConstantUInt::get(Type::UByteTy, 4);
342       Src = BinaryOperator::create(Instruction::Or, Src, new ShiftInst(Instruction::Shr, Src, SA));
343     };
344     Src = BinaryOperator::createNot(Src);
345
346     Src = new CallInst(new Function(CI->getCalledFunction()->getFunctionType(),
347                                     CI->getCalledFunction()->getLinkage(),
348                                     "llvm.ctpop"), Src);
349     CI->replaceAllUsesWith(Src);
350     break;
351   }
352   case Intrinsic::cttz: {
353     Value *Src = CI->getOperand(1);
354     Src = BinaryOperator::create(Instruction::And, BinaryOperator::createNot(Src),
355             BinaryOperator::create(Instruction::Sub, Src, 
356                             ConstantUInt::get(CI->getOperand(0)->getType(), 1)));
357     Src = new CallInst(new Function(CI->getCalledFunction()->getFunctionType(),
358                                     CI->getCalledFunction()->getLinkage(),
359                                     "llvm.ctpop"), Src);
360     CI->replaceAllUsesWith(Src);
361     break;
362   }
363
364   case Intrinsic::returnaddress:
365   case Intrinsic::frameaddress:
366     std::cerr << "WARNING: this target does not support the llvm."
367               << (Callee->getIntrinsicID() == Intrinsic::returnaddress ?
368                   "return" : "frame") << "address intrinsic.\n";
369     CI->replaceAllUsesWith(ConstantPointerNull::get(
370                                             cast<PointerType>(CI->getType())));
371     break;
372
373   case Intrinsic::prefetch:
374     break;    // Simply strip out prefetches on unsupported architectures
375
376   case Intrinsic::pcmarker:
377     break;    // Simply strip out pcmarker on unsupported architectures
378
379   case Intrinsic::dbg_stoppoint:
380   case Intrinsic::dbg_region_start:
381   case Intrinsic::dbg_region_end:
382   case Intrinsic::dbg_declare:
383   case Intrinsic::dbg_func_start:
384     if (CI->getType() != Type::VoidTy)
385       CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
386     break;    // Simply strip out debugging intrinsics
387
388   case Intrinsic::memcpy: {
389     // The memcpy intrinsic take an extra alignment argument that the memcpy
390     // libc function does not.
391     static Function *MemcpyFCache = 0;
392     ReplaceCallWith("memcpy", CI, CI->op_begin()+1, CI->op_end()-1,
393                     (*(CI->op_begin()+1))->getType(), MemcpyFCache);
394     break;
395   }
396   case Intrinsic::memmove: {
397     // The memmove intrinsic take an extra alignment argument that the memmove
398     // libc function does not.
399     static Function *MemmoveFCache = 0;
400     ReplaceCallWith("memmove", CI, CI->op_begin()+1, CI->op_end()-1,
401                     (*(CI->op_begin()+1))->getType(), MemmoveFCache);
402     break;
403   }
404   case Intrinsic::memset: {
405     // The memset intrinsic take an extra alignment argument that the memset
406     // libc function does not.
407     static Function *MemsetFCache = 0;
408     ReplaceCallWith("memset", CI, CI->op_begin()+1, CI->op_end()-1,
409                     (*(CI->op_begin()+1))->getType(), MemsetFCache);
410     break;
411   }
412   case Intrinsic::isunordered: {
413     Value *L = CI->getOperand(1);
414     Value *R = CI->getOperand(2);
415
416     Value *LIsNan = new SetCondInst(Instruction::SetNE, L, L, "LIsNan", CI);
417     Value *RIsNan = new SetCondInst(Instruction::SetNE, R, R, "RIsNan", CI);
418     CI->replaceAllUsesWith(
419       BinaryOperator::create(Instruction::Or, LIsNan, RIsNan,
420                              "isunordered", CI));
421     break;
422   }
423   case Intrinsic::sqrt: {
424     static Function *sqrtFCache = 0;
425     static Function *sqrtfFCache = 0;
426     if(CI->getType() == Type::FloatTy)
427       ReplaceCallWith("sqrtf", CI, CI->op_begin()+1, CI->op_end(),
428                       Type::FloatTy, sqrtfFCache);
429     else
430       ReplaceCallWith("sqrt", CI, CI->op_begin()+1, CI->op_end(),
431                       Type::DoubleTy, sqrtFCache);
432     break;
433   }
434   }
435
436   assert(CI->use_empty() &&
437          "Lowering should have eliminated any uses of the intrinsic call!");
438   CI->getParent()->getInstList().erase(CI);
439 }