Spillers may alter MachineLoopInfo when breaking critical edges, so make it
[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/CallSite.h"
20 #include "llvm/Support/ErrorHandling.h"
21 #include "llvm/Support/IRBuilder.h"
22 #include "llvm/Support/raw_ostream.h"
23 #include "llvm/Target/TargetData.h"
24 #include "llvm/ADT/SmallVector.h"
25 using namespace llvm;
26
27 template <class ArgIt>
28 static void EnsureFunctionExists(Module &M, const char *Name,
29                                  ArgIt ArgBegin, ArgIt ArgEnd,
30                                  const Type *RetTy) {
31   // Insert a correctly-typed definition now.
32   std::vector<const Type *> ParamTys;
33   for (ArgIt I = ArgBegin; I != ArgEnd; ++I)
34     ParamTys.push_back(I->getType());
35   M.getOrInsertFunction(Name, FunctionType::get(RetTy, ParamTys, false));
36 }
37
38 static void EnsureFPIntrinsicsExist(Module &M, Function *Fn,
39                                     const char *FName,
40                                     const char *DName, const char *LDName) {
41   // Insert definitions for all the floating point types.
42   switch((int)Fn->arg_begin()->getType()->getTypeID()) {
43   case Type::FloatTyID:
44     EnsureFunctionExists(M, FName, Fn->arg_begin(), Fn->arg_end(),
45                          Type::getFloatTy(M.getContext()));
46     break;
47   case Type::DoubleTyID:
48     EnsureFunctionExists(M, DName, Fn->arg_begin(), Fn->arg_end(),
49                          Type::getDoubleTy(M.getContext()));
50     break;
51   case Type::X86_FP80TyID:
52   case Type::FP128TyID:
53   case Type::PPC_FP128TyID:
54     EnsureFunctionExists(M, LDName, Fn->arg_begin(), Fn->arg_end(),
55                          Fn->arg_begin()->getType());
56     break;
57   }
58 }
59
60 /// ReplaceCallWith - This function is used when we want to lower an intrinsic
61 /// call to a call of an external function.  This handles hard cases such as
62 /// when there was already a prototype for the external function, and if that
63 /// prototype doesn't match the arguments we expect to pass in.
64 template <class ArgIt>
65 static CallInst *ReplaceCallWith(const char *NewFn, CallInst *CI,
66                                  ArgIt ArgBegin, ArgIt ArgEnd,
67                                  const Type *RetTy) {
68   // If we haven't already looked up this function, check to see if the
69   // program already contains a function with this name.
70   Module *M = CI->getParent()->getParent()->getParent();
71   // Get or insert the definition now.
72   std::vector<const Type *> ParamTys;
73   for (ArgIt I = ArgBegin; I != ArgEnd; ++I)
74     ParamTys.push_back((*I)->getType());
75   Constant* FCache = M->getOrInsertFunction(NewFn,
76                                   FunctionType::get(RetTy, ParamTys, false));
77
78   IRBuilder<> Builder(CI->getParent(), CI);
79   SmallVector<Value *, 8> Args(ArgBegin, ArgEnd);
80   CallInst *NewCI = Builder.CreateCall(FCache, Args.begin(), Args.end());
81   NewCI->setName(CI->getName());
82   if (!CI->use_empty())
83     CI->replaceAllUsesWith(NewCI);
84   return NewCI;
85 }
86
87 // VisualStudio defines setjmp as _setjmp
88 #if defined(_MSC_VER) && defined(setjmp)
89 #define setjmp_undefined_for_visual_studio
90 #undef setjmp
91 #endif
92
93 void IntrinsicLowering::AddPrototypes(Module &M) {
94   LLVMContext &Context = M.getContext();
95   for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
96     if (I->isDeclaration() && !I->use_empty())
97       switch (I->getIntrinsicID()) {
98       default: break;
99       case Intrinsic::setjmp:
100         EnsureFunctionExists(M, "setjmp", I->arg_begin(), I->arg_end(),
101                              Type::getInt32Ty(M.getContext()));
102         break;
103       case Intrinsic::longjmp:
104         EnsureFunctionExists(M, "longjmp", I->arg_begin(), I->arg_end(),
105                              Type::getVoidTy(M.getContext()));
106         break;
107       case Intrinsic::siglongjmp:
108         EnsureFunctionExists(M, "abort", I->arg_end(), I->arg_end(),
109                              Type::getVoidTy(M.getContext()));
110         break;
111       case Intrinsic::memcpy:
112         M.getOrInsertFunction("memcpy",
113           Type::getInt8PtrTy(Context),
114                               Type::getInt8PtrTy(Context), 
115                               Type::getInt8PtrTy(Context), 
116                               TD.getIntPtrType(Context), (Type *)0);
117         break;
118       case Intrinsic::memmove:
119         M.getOrInsertFunction("memmove",
120           Type::getInt8PtrTy(Context),
121                               Type::getInt8PtrTy(Context), 
122                               Type::getInt8PtrTy(Context), 
123                               TD.getIntPtrType(Context), (Type *)0);
124         break;
125       case Intrinsic::memset:
126         M.getOrInsertFunction("memset",
127           Type::getInt8PtrTy(Context),
128                               Type::getInt8PtrTy(Context), 
129                               Type::getInt32Ty(M.getContext()), 
130                               TD.getIntPtrType(Context), (Type *)0);
131         break;
132       case Intrinsic::sqrt:
133         EnsureFPIntrinsicsExist(M, I, "sqrtf", "sqrt", "sqrtl");
134         break;
135       case Intrinsic::sin:
136         EnsureFPIntrinsicsExist(M, I, "sinf", "sin", "sinl");
137         break;
138       case Intrinsic::cos:
139         EnsureFPIntrinsicsExist(M, I, "cosf", "cos", "cosl");
140         break;
141       case Intrinsic::pow:
142         EnsureFPIntrinsicsExist(M, I, "powf", "pow", "powl");
143         break;
144       case Intrinsic::log:
145         EnsureFPIntrinsicsExist(M, I, "logf", "log", "logl");
146         break;
147       case Intrinsic::log2:
148         EnsureFPIntrinsicsExist(M, I, "log2f", "log2", "log2l");
149         break;
150       case Intrinsic::log10:
151         EnsureFPIntrinsicsExist(M, I, "log10f", "log10", "log10l");
152         break;
153       case Intrinsic::exp:
154         EnsureFPIntrinsicsExist(M, I, "expf", "exp", "expl");
155         break;
156       case Intrinsic::exp2:
157         EnsureFPIntrinsicsExist(M, I, "exp2f", "exp2", "exp2l");
158         break;
159       }
160 }
161
162 /// LowerBSWAP - Emit the code to lower bswap of V before the specified
163 /// instruction IP.
164 static Value *LowerBSWAP(LLVMContext &Context, Value *V, Instruction *IP) {
165   assert(V->getType()->isIntegerTy() && "Can't bswap a non-integer type!");
166
167   unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
168   
169   IRBuilder<> Builder(IP->getParent(), IP);
170
171   switch(BitSize) {
172   default: llvm_unreachable("Unhandled type size of value to byteswap!");
173   case 16: {
174     Value *Tmp1 = Builder.CreateShl(V, ConstantInt::get(V->getType(), 8),
175                                     "bswap.2");
176     Value *Tmp2 = Builder.CreateLShr(V, ConstantInt::get(V->getType(), 8),
177                                      "bswap.1");
178     V = Builder.CreateOr(Tmp1, Tmp2, "bswap.i16");
179     break;
180   }
181   case 32: {
182     Value *Tmp4 = Builder.CreateShl(V, ConstantInt::get(V->getType(), 24),
183                                     "bswap.4");
184     Value *Tmp3 = Builder.CreateShl(V, ConstantInt::get(V->getType(), 8),
185                                     "bswap.3");
186     Value *Tmp2 = Builder.CreateLShr(V, ConstantInt::get(V->getType(), 8),
187                                      "bswap.2");
188     Value *Tmp1 = Builder.CreateLShr(V,ConstantInt::get(V->getType(), 24),
189                                      "bswap.1");
190     Tmp3 = Builder.CreateAnd(Tmp3,
191                          ConstantInt::get(Type::getInt32Ty(Context), 0xFF0000),
192                              "bswap.and3");
193     Tmp2 = Builder.CreateAnd(Tmp2,
194                            ConstantInt::get(Type::getInt32Ty(Context), 0xFF00),
195                              "bswap.and2");
196     Tmp4 = Builder.CreateOr(Tmp4, Tmp3, "bswap.or1");
197     Tmp2 = Builder.CreateOr(Tmp2, Tmp1, "bswap.or2");
198     V = Builder.CreateOr(Tmp4, Tmp2, "bswap.i32");
199     break;
200   }
201   case 64: {
202     Value *Tmp8 = Builder.CreateShl(V, ConstantInt::get(V->getType(), 56),
203                                     "bswap.8");
204     Value *Tmp7 = Builder.CreateShl(V, ConstantInt::get(V->getType(), 40),
205                                     "bswap.7");
206     Value *Tmp6 = Builder.CreateShl(V, ConstantInt::get(V->getType(), 24),
207                                     "bswap.6");
208     Value *Tmp5 = Builder.CreateShl(V, ConstantInt::get(V->getType(), 8),
209                                     "bswap.5");
210     Value* Tmp4 = Builder.CreateLShr(V, ConstantInt::get(V->getType(), 8),
211                                      "bswap.4");
212     Value* Tmp3 = Builder.CreateLShr(V, 
213                                      ConstantInt::get(V->getType(), 24),
214                                      "bswap.3");
215     Value* Tmp2 = Builder.CreateLShr(V, 
216                                      ConstantInt::get(V->getType(), 40),
217                                      "bswap.2");
218     Value* Tmp1 = Builder.CreateLShr(V, 
219                                      ConstantInt::get(V->getType(), 56),
220                                      "bswap.1");
221     Tmp7 = Builder.CreateAnd(Tmp7,
222                              ConstantInt::get(Type::getInt64Ty(Context),
223                                               0xFF000000000000ULL),
224                              "bswap.and7");
225     Tmp6 = Builder.CreateAnd(Tmp6,
226                              ConstantInt::get(Type::getInt64Ty(Context),
227                                               0xFF0000000000ULL),
228                              "bswap.and6");
229     Tmp5 = Builder.CreateAnd(Tmp5,
230                         ConstantInt::get(Type::getInt64Ty(Context),
231                              0xFF00000000ULL),
232                              "bswap.and5");
233     Tmp4 = Builder.CreateAnd(Tmp4,
234                         ConstantInt::get(Type::getInt64Ty(Context),
235                              0xFF000000ULL),
236                              "bswap.and4");
237     Tmp3 = Builder.CreateAnd(Tmp3,
238                              ConstantInt::get(Type::getInt64Ty(Context),
239                              0xFF0000ULL),
240                              "bswap.and3");
241     Tmp2 = Builder.CreateAnd(Tmp2,
242                              ConstantInt::get(Type::getInt64Ty(Context),
243                              0xFF00ULL),
244                              "bswap.and2");
245     Tmp8 = Builder.CreateOr(Tmp8, Tmp7, "bswap.or1");
246     Tmp6 = Builder.CreateOr(Tmp6, Tmp5, "bswap.or2");
247     Tmp4 = Builder.CreateOr(Tmp4, Tmp3, "bswap.or3");
248     Tmp2 = Builder.CreateOr(Tmp2, Tmp1, "bswap.or4");
249     Tmp8 = Builder.CreateOr(Tmp8, Tmp6, "bswap.or5");
250     Tmp4 = Builder.CreateOr(Tmp4, Tmp2, "bswap.or6");
251     V = Builder.CreateOr(Tmp8, Tmp4, "bswap.i64");
252     break;
253   }
254   }
255   return V;
256 }
257
258 /// LowerCTPOP - Emit the code to lower ctpop of V before the specified
259 /// instruction IP.
260 static Value *LowerCTPOP(LLVMContext &Context, Value *V, Instruction *IP) {
261   assert(V->getType()->isIntegerTy() && "Can't ctpop a non-integer type!");
262
263   static const uint64_t MaskValues[6] = {
264     0x5555555555555555ULL, 0x3333333333333333ULL,
265     0x0F0F0F0F0F0F0F0FULL, 0x00FF00FF00FF00FFULL,
266     0x0000FFFF0000FFFFULL, 0x00000000FFFFFFFFULL
267   };
268
269   IRBuilder<> Builder(IP->getParent(), IP);
270
271   unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
272   unsigned WordSize = (BitSize + 63) / 64;
273   Value *Count = ConstantInt::get(V->getType(), 0);
274
275   for (unsigned n = 0; n < WordSize; ++n) {
276     Value *PartValue = V;
277     for (unsigned i = 1, ct = 0; i < (BitSize>64 ? 64 : BitSize); 
278          i <<= 1, ++ct) {
279       Value *MaskCst = ConstantInt::get(V->getType(), MaskValues[ct]);
280       Value *LHS = Builder.CreateAnd(PartValue, MaskCst, "cppop.and1");
281       Value *VShift = Builder.CreateLShr(PartValue,
282                                         ConstantInt::get(V->getType(), i),
283                                          "ctpop.sh");
284       Value *RHS = Builder.CreateAnd(VShift, MaskCst, "cppop.and2");
285       PartValue = Builder.CreateAdd(LHS, RHS, "ctpop.step");
286     }
287     Count = Builder.CreateAdd(PartValue, Count, "ctpop.part");
288     if (BitSize > 64) {
289       V = Builder.CreateLShr(V, ConstantInt::get(V->getType(), 64),
290                              "ctpop.part.sh");
291       BitSize -= 64;
292     }
293   }
294
295   return Count;
296 }
297
298 /// LowerCTLZ - Emit the code to lower ctlz of V before the specified
299 /// instruction IP.
300 static Value *LowerCTLZ(LLVMContext &Context, Value *V, Instruction *IP) {
301
302   IRBuilder<> Builder(IP->getParent(), IP);
303
304   unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
305   for (unsigned i = 1; i < BitSize; i <<= 1) {
306     Value *ShVal = ConstantInt::get(V->getType(), i);
307     ShVal = Builder.CreateLShr(V, ShVal, "ctlz.sh");
308     V = Builder.CreateOr(V, ShVal, "ctlz.step");
309   }
310
311   V = Builder.CreateNot(V);
312   return LowerCTPOP(Context, V, IP);
313 }
314
315 static void ReplaceFPIntrinsicWithCall(CallInst *CI, const char *Fname,
316                                        const char *Dname,
317                                        const char *LDname) {
318   CallSite CS(CI);
319   switch (CI->getArgOperand(0)->getType()->getTypeID()) {
320   default: llvm_unreachable("Invalid type in intrinsic");
321   case Type::FloatTyID:
322     ReplaceCallWith(Fname, CI, CS.arg_begin(), CS.arg_end(),
323                   Type::getFloatTy(CI->getContext()));
324     break;
325   case Type::DoubleTyID:
326     ReplaceCallWith(Dname, CI, CS.arg_begin(), CS.arg_end(),
327                   Type::getDoubleTy(CI->getContext()));
328     break;
329   case Type::X86_FP80TyID:
330   case Type::FP128TyID:
331   case Type::PPC_FP128TyID:
332     ReplaceCallWith(LDname, CI, CS.arg_begin(), CS.arg_end(),
333                   CI->getArgOperand(0)->getType());
334     break;
335   }
336 }
337
338 void IntrinsicLowering::LowerIntrinsicCall(CallInst *CI) {
339   IRBuilder<> Builder(CI->getParent(), CI);
340   LLVMContext &Context = CI->getContext();
341
342   const Function *Callee = CI->getCalledFunction();
343   assert(Callee && "Cannot lower an indirect call!");
344
345   CallSite CS(CI);
346   switch (Callee->getIntrinsicID()) {
347   case Intrinsic::not_intrinsic:
348     report_fatal_error("Cannot lower a call to a non-intrinsic function '"+
349                       Callee->getName() + "'!");
350   default:
351     report_fatal_error("Code generator does not support intrinsic function '"+
352                       Callee->getName()+"'!");
353
354     // The setjmp/longjmp intrinsics should only exist in the code if it was
355     // never optimized (ie, right out of the CFE), or if it has been hacked on
356     // by the lowerinvoke pass.  In both cases, the right thing to do is to
357     // convert the call to an explicit setjmp or longjmp call.
358   case Intrinsic::setjmp: {
359     Value *V = ReplaceCallWith("setjmp", CI, CS.arg_begin(), CS.arg_end(),
360                                Type::getInt32Ty(Context));
361     if (!CI->getType()->isVoidTy())
362       CI->replaceAllUsesWith(V);
363     break;
364   }
365   case Intrinsic::sigsetjmp:
366      if (!CI->getType()->isVoidTy())
367        CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
368      break;
369
370   case Intrinsic::longjmp: {
371     ReplaceCallWith("longjmp", CI, CS.arg_begin(), CS.arg_end(),
372                     Type::getVoidTy(Context));
373     break;
374   }
375
376   case Intrinsic::siglongjmp: {
377     // Insert the call to abort
378     ReplaceCallWith("abort", CI, CS.arg_end(), CS.arg_end(), 
379                     Type::getVoidTy(Context));
380     break;
381   }
382   case Intrinsic::ctpop:
383     CI->replaceAllUsesWith(LowerCTPOP(Context, CI->getArgOperand(0), CI));
384     break;
385
386   case Intrinsic::bswap:
387     CI->replaceAllUsesWith(LowerBSWAP(Context, CI->getArgOperand(0), CI));
388     break;
389     
390   case Intrinsic::ctlz:
391     CI->replaceAllUsesWith(LowerCTLZ(Context, CI->getArgOperand(0), CI));
392     break;
393
394   case Intrinsic::cttz: {
395     // cttz(x) -> ctpop(~X & (X-1))
396     Value *Src = CI->getArgOperand(0);
397     Value *NotSrc = Builder.CreateNot(Src);
398     NotSrc->setName(Src->getName() + ".not");
399     Value *SrcM1 = ConstantInt::get(Src->getType(), 1);
400     SrcM1 = Builder.CreateSub(Src, SrcM1);
401     Src = LowerCTPOP(Context, Builder.CreateAnd(NotSrc, SrcM1), CI);
402     CI->replaceAllUsesWith(Src);
403     break;
404   }
405
406   case Intrinsic::stacksave:
407   case Intrinsic::stackrestore: {
408     if (!Warned)
409       errs() << "WARNING: this target does not support the llvm.stack"
410              << (Callee->getIntrinsicID() == Intrinsic::stacksave ?
411                "save" : "restore") << " intrinsic.\n";
412     Warned = true;
413     if (Callee->getIntrinsicID() == Intrinsic::stacksave)
414       CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
415     break;
416   }
417     
418   case Intrinsic::returnaddress:
419   case Intrinsic::frameaddress:
420     errs() << "WARNING: this target does not support the llvm."
421            << (Callee->getIntrinsicID() == Intrinsic::returnaddress ?
422              "return" : "frame") << "address intrinsic.\n";
423     CI->replaceAllUsesWith(ConstantPointerNull::get(
424                                             cast<PointerType>(CI->getType())));
425     break;
426
427   case Intrinsic::prefetch:
428     break;    // Simply strip out prefetches on unsupported architectures
429
430   case Intrinsic::pcmarker:
431     break;    // Simply strip out pcmarker on unsupported architectures
432   case Intrinsic::readcyclecounter: {
433     errs() << "WARNING: this target does not support the llvm.readcyclecoun"
434            << "ter intrinsic.  It is being lowered to a constant 0\n";
435     CI->replaceAllUsesWith(ConstantInt::get(Type::getInt64Ty(Context), 0));
436     break;
437   }
438
439   case Intrinsic::dbg_declare:
440     break;    // Simply strip out debugging intrinsics
441
442   case Intrinsic::eh_exception:
443   case Intrinsic::eh_selector:
444     CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
445     break;
446
447   case Intrinsic::eh_typeid_for:
448     // Return something different to eh_selector.
449     CI->replaceAllUsesWith(ConstantInt::get(CI->getType(), 1));
450     break;
451
452   case Intrinsic::var_annotation:
453     break;   // Strip out annotate intrinsic
454     
455   case Intrinsic::memcpy: {
456     const IntegerType *IntPtr = TD.getIntPtrType(Context);
457     Value *Size = Builder.CreateIntCast(CI->getArgOperand(2), IntPtr,
458                                         /* isSigned */ false);
459     Value *Ops[3];
460     Ops[0] = CI->getArgOperand(0);
461     Ops[1] = CI->getArgOperand(1);
462     Ops[2] = Size;
463     ReplaceCallWith("memcpy", CI, Ops, Ops+3, CI->getArgOperand(0)->getType());
464     break;
465   }
466   case Intrinsic::memmove: {
467     const IntegerType *IntPtr = TD.getIntPtrType(Context);
468     Value *Size = Builder.CreateIntCast(CI->getArgOperand(2), IntPtr,
469                                         /* isSigned */ false);
470     Value *Ops[3];
471     Ops[0] = CI->getArgOperand(0);
472     Ops[1] = CI->getArgOperand(1);
473     Ops[2] = Size;
474     ReplaceCallWith("memmove", CI, Ops, Ops+3, CI->getArgOperand(0)->getType());
475     break;
476   }
477   case Intrinsic::memset: {
478     const IntegerType *IntPtr = TD.getIntPtrType(Context);
479     Value *Size = Builder.CreateIntCast(CI->getArgOperand(2), IntPtr,
480                                         /* isSigned */ false);
481     Value *Ops[3];
482     Ops[0] = CI->getArgOperand(0);
483     // Extend the amount to i32.
484     Ops[1] = Builder.CreateIntCast(CI->getArgOperand(1), Type::getInt32Ty(Context),
485                                    /* isSigned */ false);
486     Ops[2] = Size;
487     ReplaceCallWith("memset", CI, Ops, Ops+3, CI->getArgOperand(0)->getType());
488     break;
489   }
490   case Intrinsic::sqrt: {
491     ReplaceFPIntrinsicWithCall(CI, "sqrtf", "sqrt", "sqrtl");
492     break;
493   }
494   case Intrinsic::log: {
495     ReplaceFPIntrinsicWithCall(CI, "logf", "log", "logl");
496     break;
497   }
498   case Intrinsic::log2: {
499     ReplaceFPIntrinsicWithCall(CI, "log2f", "log2", "log2l");
500     break;
501   }
502   case Intrinsic::log10: {
503     ReplaceFPIntrinsicWithCall(CI, "log10f", "log10", "log10l");
504     break;
505   }
506   case Intrinsic::exp: {
507     ReplaceFPIntrinsicWithCall(CI, "expf", "exp", "expl");
508     break;
509   }
510   case Intrinsic::exp2: {
511     ReplaceFPIntrinsicWithCall(CI, "exp2f", "exp2", "exp2l");
512     break;
513   }
514   case Intrinsic::pow: {
515     ReplaceFPIntrinsicWithCall(CI, "powf", "pow", "powl");
516     break;
517   }
518   case Intrinsic::flt_rounds:
519      // Lower to "round to the nearest"
520      if (!CI->getType()->isVoidTy())
521        CI->replaceAllUsesWith(ConstantInt::get(CI->getType(), 1));
522      break;
523   case Intrinsic::invariant_start:
524   case Intrinsic::lifetime_start:
525     // Discard region information.
526     CI->replaceAllUsesWith(UndefValue::get(CI->getType()));
527     break;
528   case Intrinsic::invariant_end:
529   case Intrinsic::lifetime_end:
530     // Discard region information.
531     break;
532   }
533
534   assert(CI->use_empty() &&
535          "Lowering should have eliminated any uses of the intrinsic call!");
536   CI->eraseFromParent();
537 }