63bb5f21f8f158ac64f865ce9bc66269daf8f245
[oota-llvm.git] / lib / CodeGen / IntrinsicLowering.cpp
1 //===-- IntrinsicLowering.cpp - Intrinsic Lowering default implementation -===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements the IntrinsicLowering class.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "llvm/Constants.h"
15 #include "llvm/DerivedTypes.h"
16 #include "llvm/Module.h"
17 #include "llvm/Type.h"
18 #include "llvm/CodeGen/IntrinsicLowering.h"
19 #include "llvm/Support/ErrorHandling.h"
20 #include "llvm/Support/IRBuilder.h"
21 #include "llvm/Support/raw_ostream.h"
22 #include "llvm/Target/TargetData.h"
23 #include "llvm/ADT/SmallVector.h"
24 using namespace llvm;
25
26 template <class ArgIt>
27 static void EnsureFunctionExists(Module &M, const char *Name,
28                                  ArgIt ArgBegin, ArgIt ArgEnd,
29                                  const Type *RetTy) {
30   // Insert a correctly-typed definition now.
31   std::vector<const Type *> ParamTys;
32   for (ArgIt I = ArgBegin; I != ArgEnd; ++I)
33     ParamTys.push_back(I->getType());
34   M.getOrInsertFunction(Name, FunctionType::get(RetTy, ParamTys, false));
35 }
36
37 static void EnsureFPIntrinsicsExist(Module &M, Function *Fn,
38                                     const char *FName,
39                                     const char *DName, const char *LDName) {
40   // Insert definitions for all the floating point types.
41   switch((int)Fn->arg_begin()->getType()->getTypeID()) {
42   case Type::FloatTyID:
43     EnsureFunctionExists(M, FName, Fn->arg_begin(), Fn->arg_end(),
44                          Type::getFloatTy(M.getContext()));
45     break;
46   case Type::DoubleTyID:
47     EnsureFunctionExists(M, DName, Fn->arg_begin(), Fn->arg_end(),
48                          Type::getDoubleTy(M.getContext()));
49     break;
50   case Type::X86_FP80TyID:
51   case Type::FP128TyID:
52   case Type::PPC_FP128TyID:
53     EnsureFunctionExists(M, LDName, Fn->arg_begin(), Fn->arg_end(),
54                          Fn->arg_begin()->getType());
55     break;
56   }
57 }
58
59 /// ReplaceCallWith - This function is used when we want to lower an intrinsic
60 /// call to a call of an external function.  This handles hard cases such as
61 /// when there was already a prototype for the external function, and if that
62 /// prototype doesn't match the arguments we expect to pass in.
63 template <class ArgIt>
64 static CallInst *ReplaceCallWith(const char *NewFn, CallInst *CI,
65                                  ArgIt ArgBegin, ArgIt ArgEnd,
66                                  const Type *RetTy) {
67   // If we haven't already looked up this function, check to see if the
68   // program already contains a function with this name.
69   Module *M = CI->getParent()->getParent()->getParent();
70   // Get or insert the definition now.
71   std::vector<const Type *> ParamTys;
72   for (ArgIt I = ArgBegin; I != ArgEnd; ++I)
73     ParamTys.push_back((*I)->getType());
74   Constant* FCache = M->getOrInsertFunction(NewFn,
75                                   FunctionType::get(RetTy, ParamTys, false));
76
77   IRBuilder<> Builder(CI->getParent(), CI);
78   SmallVector<Value *, 8> Args(ArgBegin, ArgEnd);
79   CallInst *NewCI = Builder.CreateCall(FCache, Args.begin(), Args.end());
80   NewCI->setName(CI->getName());
81   if (!CI->use_empty())
82     CI->replaceAllUsesWith(NewCI);
83   return NewCI;
84 }
85
86 // VisualStudio defines setjmp as _setjmp
87 #if defined(_MSC_VER) && defined(setjmp)
88 #define setjmp_undefined_for_visual_studio
89 #undef setjmp
90 #endif
91
92 void IntrinsicLowering::AddPrototypes(Module &M) {
93   LLVMContext &Context = M.getContext();
94   for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
95     if (I->isDeclaration() && !I->use_empty())
96       switch (I->getIntrinsicID()) {
97       default: break;
98       case Intrinsic::setjmp:
99         EnsureFunctionExists(M, "setjmp", I->arg_begin(), I->arg_end(),
100                              Type::getInt32Ty(M.getContext()));
101         break;
102       case Intrinsic::longjmp:
103         EnsureFunctionExists(M, "longjmp", I->arg_begin(), I->arg_end(),
104                              Type::getVoidTy(M.getContext()));
105         break;
106       case Intrinsic::siglongjmp:
107         EnsureFunctionExists(M, "abort", I->arg_end(), I->arg_end(),
108                              Type::getVoidTy(M.getContext()));
109         break;
110       case Intrinsic::memcpy:
111         M.getOrInsertFunction("memcpy",
112           Type::getInt8PtrTy(Context),
113                               Type::getInt8PtrTy(Context), 
114                               Type::getInt8PtrTy(Context), 
115                               TD.getIntPtrType(Context), (Type *)0);
116         break;
117       case Intrinsic::memmove:
118         M.getOrInsertFunction("memmove",
119           Type::getInt8PtrTy(Context),
120                               Type::getInt8PtrTy(Context), 
121                               Type::getInt8PtrTy(Context), 
122                               TD.getIntPtrType(Context), (Type *)0);
123         break;
124       case Intrinsic::memset:
125         M.getOrInsertFunction("memset",
126           Type::getInt8PtrTy(Context),
127                               Type::getInt8PtrTy(Context), 
128                               Type::getInt32Ty(M.getContext()), 
129                               TD.getIntPtrType(Context), (Type *)0);
130         break;
131       case Intrinsic::sqrt:
132         EnsureFPIntrinsicsExist(M, I, "sqrtf", "sqrt", "sqrtl");
133         break;
134       case Intrinsic::sin:
135         EnsureFPIntrinsicsExist(M, I, "sinf", "sin", "sinl");
136         break;
137       case Intrinsic::cos:
138         EnsureFPIntrinsicsExist(M, I, "cosf", "cos", "cosl");
139         break;
140       case Intrinsic::pow:
141         EnsureFPIntrinsicsExist(M, I, "powf", "pow", "powl");
142         break;
143       case Intrinsic::log:
144         EnsureFPIntrinsicsExist(M, I, "logf", "log", "logl");
145         break;
146       case Intrinsic::log2:
147         EnsureFPIntrinsicsExist(M, I, "log2f", "log2", "log2l");
148         break;
149       case Intrinsic::log10:
150         EnsureFPIntrinsicsExist(M, I, "log10f", "log10", "log10l");
151         break;
152       case Intrinsic::exp:
153         EnsureFPIntrinsicsExist(M, I, "expf", "exp", "expl");
154         break;
155       case Intrinsic::exp2:
156         EnsureFPIntrinsicsExist(M, I, "exp2f", "exp2", "exp2l");
157         break;
158       }
159 }
160
161 /// LowerBSWAP - Emit the code to lower bswap of V before the specified
162 /// instruction IP.
163 static Value *LowerBSWAP(LLVMContext &Context, Value *V, Instruction *IP) {
164   assert(V->getType()->isIntegerTy() && "Can't bswap a non-integer type!");
165
166   unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
167   
168   IRBuilder<> Builder(IP->getParent(), IP);
169
170   switch(BitSize) {
171   default: llvm_unreachable("Unhandled type size of value to byteswap!");
172   case 16: {
173     Value *Tmp1 = Builder.CreateShl(V, ConstantInt::get(V->getType(), 8),
174                                     "bswap.2");
175     Value *Tmp2 = Builder.CreateLShr(V, ConstantInt::get(V->getType(), 8),
176                                      "bswap.1");
177     V = Builder.CreateOr(Tmp1, Tmp2, "bswap.i16");
178     break;
179   }
180   case 32: {
181     Value *Tmp4 = Builder.CreateShl(V, ConstantInt::get(V->getType(), 24),
182                                     "bswap.4");
183     Value *Tmp3 = Builder.CreateShl(V, ConstantInt::get(V->getType(), 8),
184                                     "bswap.3");
185     Value *Tmp2 = Builder.CreateLShr(V, ConstantInt::get(V->getType(), 8),
186                                      "bswap.2");
187     Value *Tmp1 = Builder.CreateLShr(V,ConstantInt::get(V->getType(), 24),
188                                      "bswap.1");
189     Tmp3 = Builder.CreateAnd(Tmp3,
190                          ConstantInt::get(Type::getInt32Ty(Context), 0xFF0000),
191                              "bswap.and3");
192     Tmp2 = Builder.CreateAnd(Tmp2,
193                            ConstantInt::get(Type::getInt32Ty(Context), 0xFF00),
194                              "bswap.and2");
195     Tmp4 = Builder.CreateOr(Tmp4, Tmp3, "bswap.or1");
196     Tmp2 = Builder.CreateOr(Tmp2, Tmp1, "bswap.or2");
197     V = Builder.CreateOr(Tmp4, Tmp2, "bswap.i32");
198     break;
199   }
200   case 64: {
201     Value *Tmp8 = Builder.CreateShl(V, ConstantInt::get(V->getType(), 56),
202                                     "bswap.8");
203     Value *Tmp7 = Builder.CreateShl(V, ConstantInt::get(V->getType(), 40),
204                                     "bswap.7");
205     Value *Tmp6 = Builder.CreateShl(V, ConstantInt::get(V->getType(), 24),
206                                     "bswap.6");
207     Value *Tmp5 = Builder.CreateShl(V, ConstantInt::get(V->getType(), 8),
208                                     "bswap.5");
209     Value* Tmp4 = Builder.CreateLShr(V, ConstantInt::get(V->getType(), 8),
210                                      "bswap.4");
211     Value* Tmp3 = Builder.CreateLShr(V, 
212                                      ConstantInt::get(V->getType(), 24),
213                                      "bswap.3");
214     Value* Tmp2 = Builder.CreateLShr(V, 
215                                      ConstantInt::get(V->getType(), 40),
216                                      "bswap.2");
217     Value* Tmp1 = Builder.CreateLShr(V, 
218                                      ConstantInt::get(V->getType(), 56),
219                                      "bswap.1");
220     Tmp7 = Builder.CreateAnd(Tmp7,
221                              ConstantInt::get(Type::getInt64Ty(Context),
222                                               0xFF000000000000ULL),
223                              "bswap.and7");
224     Tmp6 = Builder.CreateAnd(Tmp6,
225                              ConstantInt::get(Type::getInt64Ty(Context),
226                                               0xFF0000000000ULL),
227                              "bswap.and6");
228     Tmp5 = Builder.CreateAnd(Tmp5,
229                         ConstantInt::get(Type::getInt64Ty(Context),
230                              0xFF00000000ULL),
231                              "bswap.and5");
232     Tmp4 = Builder.CreateAnd(Tmp4,
233                         ConstantInt::get(Type::getInt64Ty(Context),
234                              0xFF000000ULL),
235                              "bswap.and4");
236     Tmp3 = Builder.CreateAnd(Tmp3,
237                              ConstantInt::get(Type::getInt64Ty(Context),
238                              0xFF0000ULL),
239                              "bswap.and3");
240     Tmp2 = Builder.CreateAnd(Tmp2,
241                              ConstantInt::get(Type::getInt64Ty(Context),
242                              0xFF00ULL),
243                              "bswap.and2");
244     Tmp8 = Builder.CreateOr(Tmp8, Tmp7, "bswap.or1");
245     Tmp6 = Builder.CreateOr(Tmp6, Tmp5, "bswap.or2");
246     Tmp4 = Builder.CreateOr(Tmp4, Tmp3, "bswap.or3");
247     Tmp2 = Builder.CreateOr(Tmp2, Tmp1, "bswap.or4");
248     Tmp8 = Builder.CreateOr(Tmp8, Tmp6, "bswap.or5");
249     Tmp4 = Builder.CreateOr(Tmp4, Tmp2, "bswap.or6");
250     V = Builder.CreateOr(Tmp8, Tmp4, "bswap.i64");
251     break;
252   }
253   }
254   return V;
255 }
256
257 /// LowerCTPOP - Emit the code to lower ctpop of V before the specified
258 /// instruction IP.
259 static Value *LowerCTPOP(LLVMContext &Context, Value *V, Instruction *IP) {
260   assert(V->getType()->isIntegerTy() && "Can't ctpop a non-integer type!");
261
262   static const uint64_t MaskValues[6] = {
263     0x5555555555555555ULL, 0x3333333333333333ULL,
264     0x0F0F0F0F0F0F0F0FULL, 0x00FF00FF00FF00FFULL,
265     0x0000FFFF0000FFFFULL, 0x00000000FFFFFFFFULL
266   };
267
268   IRBuilder<> Builder(IP->getParent(), IP);
269
270   unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
271   unsigned WordSize = (BitSize + 63) / 64;
272   Value *Count = ConstantInt::get(V->getType(), 0);
273
274   for (unsigned n = 0; n < WordSize; ++n) {
275     Value *PartValue = V;
276     for (unsigned i = 1, ct = 0; i < (BitSize>64 ? 64 : BitSize); 
277          i <<= 1, ++ct) {
278       Value *MaskCst = ConstantInt::get(V->getType(), MaskValues[ct]);
279       Value *LHS = Builder.CreateAnd(PartValue, MaskCst, "cppop.and1");
280       Value *VShift = Builder.CreateLShr(PartValue,
281                                         ConstantInt::get(V->getType(), i),
282                                          "ctpop.sh");
283       Value *RHS = Builder.CreateAnd(VShift, MaskCst, "cppop.and2");
284       PartValue = Builder.CreateAdd(LHS, RHS, "ctpop.step");
285     }
286     Count = Builder.CreateAdd(PartValue, Count, "ctpop.part");
287     if (BitSize > 64) {
288       V = Builder.CreateLShr(V, ConstantInt::get(V->getType(), 64),
289                              "ctpop.part.sh");
290       BitSize -= 64;
291     }
292   }
293
294   return Count;
295 }
296
297 /// LowerCTLZ - Emit the code to lower ctlz of V before the specified
298 /// instruction IP.
299 static Value *LowerCTLZ(LLVMContext &Context, Value *V, Instruction *IP) {
300
301   IRBuilder<> Builder(IP->getParent(), IP);
302
303   unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
304   for (unsigned i = 1; i < BitSize; i <<= 1) {
305     Value *ShVal = ConstantInt::get(V->getType(), i);
306     ShVal = Builder.CreateLShr(V, ShVal, "ctlz.sh");
307     V = Builder.CreateOr(V, ShVal, "ctlz.step");
308   }
309
310   V = Builder.CreateNot(V);
311   return LowerCTPOP(Context, V, IP);
312 }
313
314 static void ReplaceFPIntrinsicWithCall(CallInst *CI, const char *Fname,
315                                        const char *Dname,
316                                        const char *LDname) {
317   switch (CI->getOperand(1)->getType()->getTypeID()) {
318   default: llvm_unreachable("Invalid type in intrinsic");
319   case Type::FloatTyID:
320     ReplaceCallWith(Fname, CI, CI->op_begin() + 1, CI->op_end(),
321                   Type::getFloatTy(CI->getContext()));
322     break;
323   case Type::DoubleTyID:
324     ReplaceCallWith(Dname, CI, CI->op_begin() + 1, CI->op_end(),
325                   Type::getDoubleTy(CI->getContext()));
326     break;
327   case Type::X86_FP80TyID:
328   case Type::FP128TyID:
329   case Type::PPC_FP128TyID:
330     ReplaceCallWith(LDname, CI, CI->op_begin() + 1, CI->op_end(),
331                   CI->getOperand(1)->getType());
332     break;
333   }
334 }
335
336 void IntrinsicLowering::LowerIntrinsicCall(CallInst *CI) {
337   IRBuilder<> Builder(CI->getParent(), CI);
338   LLVMContext &Context = CI->getContext();
339
340   const Function *Callee = CI->getCalledFunction();
341   assert(Callee && "Cannot lower an indirect call!");
342
343   switch (Callee->getIntrinsicID()) {
344   case Intrinsic::not_intrinsic:
345     report_fatal_error("Cannot lower a call to a non-intrinsic function '"+
346                       Callee->getName() + "'!");
347   default:
348     report_fatal_error("Code generator does not support intrinsic function '"+
349                       Callee->getName()+"'!");
350
351     // The setjmp/longjmp intrinsics should only exist in the code if it was
352     // never optimized (ie, right out of the CFE), or if it has been hacked on
353     // by the lowerinvoke pass.  In both cases, the right thing to do is to
354     // convert the call to an explicit setjmp or longjmp call.
355   case Intrinsic::setjmp: {
356     Value *V = ReplaceCallWith("setjmp", CI, CI->op_begin() + 1, CI->op_end(),
357                                Type::getInt32Ty(Context));
358     if (!CI->getType()->isVoidTy())
359       CI->replaceAllUsesWith(V);
360     break;
361   }
362   case Intrinsic::sigsetjmp:
363      if (!CI->getType()->isVoidTy())
364        CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
365      break;
366
367   case Intrinsic::longjmp: {
368     ReplaceCallWith("longjmp", CI, CI->op_begin() + 1, CI->op_end(),
369                     Type::getVoidTy(Context));
370     break;
371   }
372
373   case Intrinsic::siglongjmp: {
374     // Insert the call to abort
375     ReplaceCallWith("abort", CI, CI->op_end(), CI->op_end(), 
376                     Type::getVoidTy(Context));
377     break;
378   }
379   case Intrinsic::ctpop:
380     CI->replaceAllUsesWith(LowerCTPOP(Context, CI->getOperand(1), CI));
381     break;
382
383   case Intrinsic::bswap:
384     CI->replaceAllUsesWith(LowerBSWAP(Context, CI->getOperand(1), CI));
385     break;
386     
387   case Intrinsic::ctlz:
388     CI->replaceAllUsesWith(LowerCTLZ(Context, CI->getOperand(1), CI));
389     break;
390
391   case Intrinsic::cttz: {
392     // cttz(x) -> ctpop(~X & (X-1))
393     Value *Src = CI->getOperand(1);
394     Value *NotSrc = Builder.CreateNot(Src);
395     NotSrc->setName(Src->getName() + ".not");
396     Value *SrcM1 = ConstantInt::get(Src->getType(), 1);
397     SrcM1 = Builder.CreateSub(Src, SrcM1);
398     Src = LowerCTPOP(Context, Builder.CreateAnd(NotSrc, SrcM1), CI);
399     CI->replaceAllUsesWith(Src);
400     break;
401   }
402
403   case Intrinsic::stacksave:
404   case Intrinsic::stackrestore: {
405     if (!Warned)
406       errs() << "WARNING: this target does not support the llvm.stack"
407              << (Callee->getIntrinsicID() == Intrinsic::stacksave ?
408                "save" : "restore") << " intrinsic.\n";
409     Warned = true;
410     if (Callee->getIntrinsicID() == Intrinsic::stacksave)
411       CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
412     break;
413   }
414     
415   case Intrinsic::returnaddress:
416   case Intrinsic::frameaddress:
417     errs() << "WARNING: this target does not support the llvm."
418            << (Callee->getIntrinsicID() == Intrinsic::returnaddress ?
419              "return" : "frame") << "address intrinsic.\n";
420     CI->replaceAllUsesWith(ConstantPointerNull::get(
421                                             cast<PointerType>(CI->getType())));
422     break;
423
424   case Intrinsic::prefetch:
425     break;    // Simply strip out prefetches on unsupported architectures
426
427   case Intrinsic::pcmarker:
428     break;    // Simply strip out pcmarker on unsupported architectures
429   case Intrinsic::readcyclecounter: {
430     errs() << "WARNING: this target does not support the llvm.readcyclecoun"
431            << "ter intrinsic.  It is being lowered to a constant 0\n";
432     CI->replaceAllUsesWith(ConstantInt::get(Type::getInt64Ty(Context), 0));
433     break;
434   }
435
436   case Intrinsic::dbg_declare:
437     break;    // Simply strip out debugging intrinsics
438
439   case Intrinsic::eh_exception:
440   case Intrinsic::eh_selector:
441     CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
442     break;
443
444   case Intrinsic::eh_typeid_for:
445     // Return something different to eh_selector.
446     CI->replaceAllUsesWith(ConstantInt::get(CI->getType(), 1));
447     break;
448
449   case Intrinsic::var_annotation:
450     break;   // Strip out annotate intrinsic
451     
452   case Intrinsic::memcpy: {
453     const IntegerType *IntPtr = TD.getIntPtrType(Context);
454     Value *Size = Builder.CreateIntCast(CI->getOperand(3), IntPtr,
455                                         /* isSigned */ false);
456     Value *Ops[3];
457     Ops[0] = CI->getOperand(1);
458     Ops[1] = CI->getOperand(2);
459     Ops[2] = Size;
460     ReplaceCallWith("memcpy", CI, Ops, Ops+3, CI->getOperand(1)->getType());
461     break;
462   }
463   case Intrinsic::memmove: {
464     const IntegerType *IntPtr = TD.getIntPtrType(Context);
465     Value *Size = Builder.CreateIntCast(CI->getOperand(3), IntPtr,
466                                         /* isSigned */ false);
467     Value *Ops[3];
468     Ops[0] = CI->getOperand(1);
469     Ops[1] = CI->getOperand(2);
470     Ops[2] = Size;
471     ReplaceCallWith("memmove", CI, Ops, Ops+3, CI->getOperand(1)->getType());
472     break;
473   }
474   case Intrinsic::memset: {
475     const IntegerType *IntPtr = TD.getIntPtrType(Context);
476     Value *Size = Builder.CreateIntCast(CI->getOperand(3), IntPtr,
477                                         /* isSigned */ false);
478     Value *Ops[3];
479     Ops[0] = CI->getOperand(1);
480     // Extend the amount to i32.
481     Ops[1] = Builder.CreateIntCast(CI->getOperand(2), Type::getInt32Ty(Context),
482                                    /* isSigned */ false);
483     Ops[2] = Size;
484     ReplaceCallWith("memset", CI, Ops, Ops+3, CI->getOperand(1)->getType());
485     break;
486   }
487   case Intrinsic::sqrt: {
488     ReplaceFPIntrinsicWithCall(CI, "sqrtf", "sqrt", "sqrtl");
489     break;
490   }
491   case Intrinsic::log: {
492     ReplaceFPIntrinsicWithCall(CI, "logf", "log", "logl");
493     break;
494   }
495   case Intrinsic::log2: {
496     ReplaceFPIntrinsicWithCall(CI, "log2f", "log2", "log2l");
497     break;
498   }
499   case Intrinsic::log10: {
500     ReplaceFPIntrinsicWithCall(CI, "log10f", "log10", "log10l");
501     break;
502   }
503   case Intrinsic::exp: {
504     ReplaceFPIntrinsicWithCall(CI, "expf", "exp", "expl");
505     break;
506   }
507   case Intrinsic::exp2: {
508     ReplaceFPIntrinsicWithCall(CI, "exp2f", "exp2", "exp2l");
509     break;
510   }
511   case Intrinsic::pow: {
512     ReplaceFPIntrinsicWithCall(CI, "powf", "pow", "powl");
513     break;
514   }
515   case Intrinsic::flt_rounds:
516      // Lower to "round to the nearest"
517      if (!CI->getType()->isVoidTy())
518        CI->replaceAllUsesWith(ConstantInt::get(CI->getType(), 1));
519      break;
520   case Intrinsic::invariant_start:
521   case Intrinsic::lifetime_start:
522     // Discard region information.
523     CI->replaceAllUsesWith(UndefValue::get(CI->getType()));
524     break;
525   case Intrinsic::invariant_end:
526   case Intrinsic::lifetime_end:
527     // Discard region information.
528     break;
529   }
530
531   assert(CI->use_empty() &&
532          "Lowering should have eliminated any uses of the intrinsic call!");
533   CI->eraseFromParent();
534 }