IntrinsicLowering.cpp revision d1fbd142945f5ef0c273c3d756431f8cb9d25ded 
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/IRBuilder.h"
20#include "llvm/Target/TargetData.h"
21#include "llvm/ADT/SmallVector.h"
22using namespace llvm;
23
24template <class ArgIt>
25static void EnsureFunctionExists(Module &M, const char *Name,
26                                 ArgIt ArgBegin, ArgIt ArgEnd,
27                                 const Type *RetTy) {
28  // Insert a correctly-typed definition now.
29  std::vector<const Type *> ParamTys;
30  for (ArgIt I = ArgBegin; I != ArgEnd; ++I)
31    ParamTys.push_back(I->getType());
32  M.getOrInsertFunction(Name, FunctionType::get(RetTy, ParamTys, false));
33}
34
35static void EnsureFPIntrinsicsExist(Module &M, Function *Fn,
36                                    const char *FName,
37                                    const char *DName, const char *LDName) {
38  // Insert definitions for all the floating point types.
39  switch((int)Fn->arg_begin()->getType()->getTypeID()) {
40  case Type::FloatTyID:
41    EnsureFunctionExists(M, FName, Fn->arg_begin(), Fn->arg_end(),
42                         Type::FloatTy);
43    break;
44  case Type::DoubleTyID:
45    EnsureFunctionExists(M, DName, Fn->arg_begin(), Fn->arg_end(),
46                         Type::DoubleTy);
47    break;
48  case Type::X86_FP80TyID:
49  case Type::FP128TyID:
50  case Type::PPC_FP128TyID:
51    EnsureFunctionExists(M, LDName, Fn->arg_begin(), Fn->arg_end(),
52                         Fn->arg_begin()->getType());
53    break;
54  }
55}
56
57/// ReplaceCallWith - This function is used when we want to lower an intrinsic
58/// call to a call of an external function.  This handles hard cases such as
59/// when there was already a prototype for the external function, and if that
60/// prototype doesn't match the arguments we expect to pass in.
61template <class ArgIt>
62static CallInst *ReplaceCallWith(const char *NewFn, CallInst *CI,
63                                 ArgIt ArgBegin, ArgIt ArgEnd,
64                                 const Type *RetTy) {
65  // If we haven't already looked up this function, check to see if the
66  // program already contains a function with this name.
67  Module *M = CI->getParent()->getParent()->getParent();
68  // Get or insert the definition now.
69  std::vector<const Type *> ParamTys;
70  for (ArgIt I = ArgBegin; I != ArgEnd; ++I)
71    ParamTys.push_back((*I)->getType());
72  Constant* FCache = M->getOrInsertFunction(NewFn,
73                                  FunctionType::get(RetTy, ParamTys, false));
74
75  IRBuilder<> Builder(CI->getParent(), CI);
76  SmallVector<Value *, 8> Args(ArgBegin, ArgEnd);
77  CallInst *NewCI = Builder.CreateCall(FCache, Args.begin(), Args.end());
78  NewCI->setName(CI->getName());
79  if (!CI->use_empty())
80    CI->replaceAllUsesWith(NewCI);
81  return NewCI;
82}
83
84void IntrinsicLowering::AddPrototypes(Module &M) {
85  for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
86    if (I->isDeclaration() && !I->use_empty())
87      switch (I->getIntrinsicID()) {
88      default: break;
89      case Intrinsic::setjmp:
90        EnsureFunctionExists(M, "setjmp", I->arg_begin(), I->arg_end(),
91                             Type::Int32Ty);
92        break;
93      case Intrinsic::longjmp:
94        EnsureFunctionExists(M, "longjmp", I->arg_begin(), I->arg_end(),
95                             Type::VoidTy);
96        break;
97      case Intrinsic::siglongjmp:
98        EnsureFunctionExists(M, "abort", I->arg_end(), I->arg_end(),
99                             Type::VoidTy);
100        break;
101      case Intrinsic::memcpy:
102        M.getOrInsertFunction("memcpy", PointerType::getUnqual(Type::Int8Ty),
103                              PointerType::getUnqual(Type::Int8Ty),
104                              PointerType::getUnqual(Type::Int8Ty),
105                              TD.getIntPtrType(), (Type *)0);
106        break;
107      case Intrinsic::memmove:
108        M.getOrInsertFunction("memmove", PointerType::getUnqual(Type::Int8Ty),
109                              PointerType::getUnqual(Type::Int8Ty),
110                              PointerType::getUnqual(Type::Int8Ty),
111                              TD.getIntPtrType(), (Type *)0);
112        break;
113      case Intrinsic::memset:
114        M.getOrInsertFunction("memset", PointerType::getUnqual(Type::Int8Ty),
115                              PointerType::getUnqual(Type::Int8Ty),
116                              Type::Int32Ty,
117                              TD.getIntPtrType(), (Type *)0);
118        break;
119      case Intrinsic::sqrt:
120        EnsureFPIntrinsicsExist(M, I, "sqrtf", "sqrt", "sqrtl");
121        break;
122      case Intrinsic::sin:
123        EnsureFPIntrinsicsExist(M, I, "sinf", "sin", "sinl");
124        break;
125      case Intrinsic::cos:
126        EnsureFPIntrinsicsExist(M, I, "cosf", "cos", "cosl");
127        break;
128      case Intrinsic::pow:
129        EnsureFPIntrinsicsExist(M, I, "powf", "pow", "powl");
130        break;
131      case Intrinsic::log:
132        EnsureFPIntrinsicsExist(M, I, "logf", "log", "logl");
133        break;
134      case Intrinsic::log2:
135        EnsureFPIntrinsicsExist(M, I, "log2f", "log2", "log2l");
136        break;
137      case Intrinsic::log10:
138        EnsureFPIntrinsicsExist(M, I, "log10f", "log10", "log10l");
139        break;
140      case Intrinsic::exp:
141        EnsureFPIntrinsicsExist(M, I, "expf", "exp", "expl");
142        break;
143      case Intrinsic::exp2:
144        EnsureFPIntrinsicsExist(M, I, "exp2f", "exp2", "exp2l");
145        break;
146      }
147}
148
149/// LowerBSWAP - Emit the code to lower bswap of V before the specified
150/// instruction IP.
151static Value *LowerBSWAP(Value *V, Instruction *IP) {
152  assert(V->getType()->isInteger() && "Can't bswap a non-integer type!");
153
154  unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
155
156  IRBuilder<> Builder(IP->getParent(), IP);
157
158  switch(BitSize) {
159  default: assert(0 && "Unhandled type size of value to byteswap!");
160  case 16: {
161    Value *Tmp1 = Builder.CreateShl(V, ConstantInt::get(V->getType(), 8),
162                                    "bswap.2");
163    Value *Tmp2 = Builder.CreateLShr(V, ConstantInt::get(V->getType(), 8),
164                                     "bswap.1");
165    V = Builder.CreateOr(Tmp1, Tmp2, "bswap.i16");
166    break;
167  }
168  case 32: {
169    Value *Tmp4 = Builder.CreateShl(V, ConstantInt::get(V->getType(), 24),
170                                    "bswap.4");
171    Value *Tmp3 = Builder.CreateShl(V, ConstantInt::get(V->getType(), 8),
172                                    "bswap.3");
173    Value *Tmp2 = Builder.CreateLShr(V, ConstantInt::get(V->getType(), 8),
174                                     "bswap.2");
175    Value *Tmp1 = Builder.CreateLShr(V, ConstantInt::get(V->getType(), 24),
176                                     "bswap.1");
177    Tmp3 = Builder.CreateAnd(Tmp3, ConstantInt::get(Type::Int32Ty, 0xFF0000),
178                             "bswap.and3");
179    Tmp2 = Builder.CreateAnd(Tmp2, ConstantInt::get(Type::Int32Ty, 0xFF00),
180                             "bswap.and2");
181    Tmp4 = Builder.CreateOr(Tmp4, Tmp3, "bswap.or1");
182    Tmp2 = Builder.CreateOr(Tmp2, Tmp1, "bswap.or2");
183    V = Builder.CreateOr(Tmp4, Tmp2, "bswap.i32");
184    break;
185  }
186  case 64: {
187    Value *Tmp8 = Builder.CreateShl(V, ConstantInt::get(V->getType(), 56),
188                                    "bswap.8");
189    Value *Tmp7 = Builder.CreateShl(V, ConstantInt::get(V->getType(), 40),
190                                    "bswap.7");
191    Value *Tmp6 = Builder.CreateShl(V, ConstantInt::get(V->getType(), 24),
192                                    "bswap.6");
193    Value *Tmp5 = Builder.CreateShl(V, ConstantInt::get(V->getType(), 8),
194                                    "bswap.5");
195    Value* Tmp4 = Builder.CreateLShr(V, ConstantInt::get(V->getType(), 8),
196                                     "bswap.4");
197    Value* Tmp3 = Builder.CreateLShr(V, ConstantInt::get(V->getType(), 24),
198                                     "bswap.3");
199    Value* Tmp2 = Builder.CreateLShr(V, ConstantInt::get(V->getType(), 40),
200                                     "bswap.2");
201    Value* Tmp1 = Builder.CreateLShr(V, ConstantInt::get(V->getType(), 56),
202                                     "bswap.1");
203    Tmp7 = Builder.CreateAnd(Tmp7,
204                             ConstantInt::get(Type::Int64Ty,
205                                              0xFF000000000000ULL),
206                             "bswap.and7");
207    Tmp6 = Builder.CreateAnd(Tmp6,
208                             ConstantInt::get(Type::Int64Ty,
209                                              0xFF0000000000ULL),
210                             "bswap.and6");
211    Tmp5 = Builder.CreateAnd(Tmp5,
212                             ConstantInt::get(Type::Int64Ty, 0xFF00000000ULL),
213                             "bswap.and5");
214    Tmp4 = Builder.CreateAnd(Tmp4,
215                             ConstantInt::get(Type::Int64Ty, 0xFF000000ULL),
216                             "bswap.and4");
217    Tmp3 = Builder.CreateAnd(Tmp3,
218                             ConstantInt::get(Type::Int64Ty, 0xFF0000ULL),
219                             "bswap.and3");
220    Tmp2 = Builder.CreateAnd(Tmp2,
221                             ConstantInt::get(Type::Int64Ty, 0xFF00ULL),
222                             "bswap.and2");
223    Tmp8 = Builder.CreateOr(Tmp8, Tmp7, "bswap.or1");
224    Tmp6 = Builder.CreateOr(Tmp6, Tmp5, "bswap.or2");
225    Tmp4 = Builder.CreateOr(Tmp4, Tmp3, "bswap.or3");
226    Tmp2 = Builder.CreateOr(Tmp2, Tmp1, "bswap.or4");
227    Tmp8 = Builder.CreateOr(Tmp8, Tmp6, "bswap.or5");
228    Tmp4 = Builder.CreateOr(Tmp4, Tmp2, "bswap.or6");
229    V = Builder.CreateOr(Tmp8, Tmp4, "bswap.i64");
230    break;
231  }
232  }
233  return V;
234}
235
236/// LowerCTPOP - Emit the code to lower ctpop of V before the specified
237/// instruction IP.
238static Value *LowerCTPOP(Value *V, Instruction *IP) {
239  assert(V->getType()->isInteger() && "Can't ctpop a non-integer type!");
240
241  static const uint64_t MaskValues[6] = {
242    0x5555555555555555ULL, 0x3333333333333333ULL,
243    0x0F0F0F0F0F0F0F0FULL, 0x00FF00FF00FF00FFULL,
244    0x0000FFFF0000FFFFULL, 0x00000000FFFFFFFFULL
245  };
246
247  IRBuilder<> Builder(IP->getParent(), IP);
248
249  unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
250  unsigned WordSize = (BitSize + 63) / 64;
251  Value *Count = ConstantInt::get(V->getType(), 0);
252
253  for (unsigned n = 0; n < WordSize; ++n) {
254    Value *PartValue = V;
255    for (unsigned i = 1, ct = 0; i < (BitSize>64 ? 64 : BitSize);
256         i <<= 1, ++ct) {
257      Value *MaskCst = ConstantInt::get(V->getType(), MaskValues[ct]);
258      Value *LHS = Builder.CreateAnd(PartValue, MaskCst, "cppop.and1");
259      Value *VShift = Builder.CreateLShr(PartValue,
260                                         ConstantInt::get(V->getType(), i),
261                                         "ctpop.sh");
262      Value *RHS = Builder.CreateAnd(VShift, MaskCst, "cppop.and2");
263      PartValue = Builder.CreateAdd(LHS, RHS, "ctpop.step");
264    }
265    Count = Builder.CreateAdd(PartValue, Count, "ctpop.part");
266    if (BitSize > 64) {
267      V = Builder.CreateLShr(V, ConstantInt::get(V->getType(), 64),
268                             "ctpop.part.sh");
269      BitSize -= 64;
270    }
271  }
272
273  return Count;
274}
275
276/// LowerCTLZ - Emit the code to lower ctlz of V before the specified
277/// instruction IP.
278static Value *LowerCTLZ(Value *V, Instruction *IP) {
279
280  IRBuilder<> Builder(IP->getParent(), IP);
281
282  unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
283  for (unsigned i = 1; i < BitSize; i <<= 1) {
284    Value *ShVal = ConstantInt::get(V->getType(), i);
285    ShVal = Builder.CreateLShr(V, ShVal, "ctlz.sh");
286    V = Builder.CreateOr(V, ShVal, "ctlz.step");
287  }
288
289  V = Builder.CreateNot(V);
290  return LowerCTPOP(V, IP);
291}
292
293/// Convert the llvm.part.select.iX.iY intrinsic. This intrinsic takes
294/// three integer arguments. The first argument is the Value from which the
295/// bits will be selected. It may be of any bit width. The second and third
296/// arguments specify a range of bits to select with the second argument
297/// specifying the low bit and the third argument specifying the high bit. Both
298/// must be type i32. The result is the corresponding selected bits from the
299/// Value in the same width as the Value (first argument). If the low bit index
300/// is higher than the high bit index then the inverse selection is done and
301/// the bits are returned in inverse order.
302/// @brief Lowering of llvm.part.select intrinsic.
303static Instruction *LowerPartSelect(CallInst *CI) {
304  IRBuilder<> Builder(*CI->getParent()->getContext());
305
306  // Make sure we're dealing with a part select intrinsic here
307  Function *F = CI->getCalledFunction();
308  const FunctionType *FT = F->getFunctionType();
309  if (!F->isDeclaration() || !FT->getReturnType()->isInteger() ||
310      FT->getNumParams() != 3 || !FT->getParamType(0)->isInteger() ||
311      !FT->getParamType(1)->isInteger() || !FT->getParamType(2)->isInteger())
312    return CI;
313
314  // Get the intrinsic implementation function by converting all the . to _
315  // in the intrinsic's function name and then reconstructing the function
316  // declaration.
317  std::string Name(F->getName());
318  for (unsigned i = 4; i < Name.length(); ++i)
319    if (Name[i] == '.')
320      Name[i] = '_';
321  Module* M = F->getParent();
322  F = cast<Function>(M->getOrInsertFunction(Name, FT));
323  F->setLinkage(GlobalValue::WeakAnyLinkage);
324
325  // If we haven't defined the impl function yet, do so now
326  if (F->isDeclaration()) {
327
328    // Get the arguments to the function
329    Function::arg_iterator args = F->arg_begin();
330    Value* Val = args++; Val->setName("Val");
331    Value* Lo = args++; Lo->setName("Lo");
332    Value* Hi = args++; Hi->setName("High");
333
334    // We want to select a range of bits here such that [Hi, Lo] is shifted
335    // down to the low bits. However, it is quite possible that Hi is smaller
336    // than Lo in which case the bits have to be reversed.
337
338    // Create the blocks we will need for the two cases (forward, reverse)
339    BasicBlock* CurBB   = BasicBlock::Create("entry", F);
340    BasicBlock *RevSize = BasicBlock::Create("revsize", CurBB->getParent());
341    BasicBlock *FwdSize = BasicBlock::Create("fwdsize", CurBB->getParent());
342    BasicBlock *Compute = BasicBlock::Create("compute", CurBB->getParent());
343    BasicBlock *Reverse = BasicBlock::Create("reverse", CurBB->getParent());
344    BasicBlock *RsltBlk = BasicBlock::Create("result",  CurBB->getParent());
345
346    Builder.SetInsertPoint(CurBB);
347
348    // Cast Hi and Lo to the size of Val so the widths are all the same
349    if (Hi->getType() != Val->getType())
350      Hi = Builder.CreateIntCast(Hi, Val->getType(), /* isSigned */ false,
351                                 "tmp");
352    if (Lo->getType() != Val->getType())
353      Lo = Builder.CreateIntCast(Lo, Val->getType(), /* isSigned */ false,
354                                 "tmp");
355
356    // Compute a few things that both cases will need, up front.
357    Constant* Zero = ConstantInt::get(Val->getType(), 0);
358    Constant* One = ConstantInt::get(Val->getType(), 1);
359    Constant* AllOnes = ConstantInt::getAllOnesValue(Val->getType());
360
361    // Compare the Hi and Lo bit positions. This is used to determine
362    // which case we have (forward or reverse)
363    Value *Cmp = Builder.CreateICmpULT(Hi, Lo, "less");
364    Builder.CreateCondBr(Cmp, RevSize, FwdSize);
365
366    // First, compute the number of bits in the forward case.
367    Builder.SetInsertPoint(FwdSize);
368    Value* FBitSize = Builder.CreateSub(Hi, Lo, "fbits");
369    Builder.CreateBr(Compute);
370
371    // Second, compute the number of bits in the reverse case.
372    Builder.SetInsertPoint(RevSize);
373    Value* RBitSize = Builder.CreateSub(Lo, Hi, "rbits");
374    Builder.CreateBr(Compute);
375
376    // Now, compute the bit range. Start by getting the bitsize and the shift
377    // amount (either Hi or Lo) from PHI nodes. Then we compute a mask for
378    // the number of bits we want in the range. We shift the bits down to the
379    // least significant bits, apply the mask to zero out unwanted high bits,
380    // and we have computed the "forward" result. It may still need to be
381    // reversed.
382    Builder.SetInsertPoint(Compute);
383
384    // Get the BitSize from one of the two subtractions
385    PHINode *BitSize = Builder.CreatePHI(Val->getType(), "bits");
386    BitSize->reserveOperandSpace(2);
387    BitSize->addIncoming(FBitSize, FwdSize);
388    BitSize->addIncoming(RBitSize, RevSize);
389
390    // Get the ShiftAmount as the smaller of Hi/Lo
391    PHINode *ShiftAmt = Builder.CreatePHI(Val->getType(), "shiftamt");
392    ShiftAmt->reserveOperandSpace(2);
393    ShiftAmt->addIncoming(Lo, FwdSize);
394    ShiftAmt->addIncoming(Hi, RevSize);
395
396    // Increment the bit size
397    Value *BitSizePlusOne = Builder.CreateAdd(BitSize, One, "bits");
398
399    // Create a Mask to zero out the high order bits.
400    Value* Mask = Builder.CreateShl(AllOnes, BitSizePlusOne, "mask");
401    Mask = Builder.CreateNot(Mask, "mask");
402
403    // Shift the bits down and apply the mask
404    Value* FRes = Builder.CreateLShr(Val, ShiftAmt, "fres");
405    FRes = Builder.CreateAnd(FRes, Mask, "fres");
406    Builder.CreateCondBr(Cmp, Reverse, RsltBlk);
407
408    // In the Reverse block we have the mask already in FRes but we must reverse
409    // it by shifting FRes bits right and putting them in RRes by shifting them
410    // in from left.
411    Builder.SetInsertPoint(Reverse);
412
413    // First set up our loop counters
414    PHINode *Count = Builder.CreatePHI(Val->getType(), "count");
415    Count->reserveOperandSpace(2);
416    Count->addIncoming(BitSizePlusOne, Compute);
417
418    // Next, get the value that we are shifting.
419    PHINode *BitsToShift = Builder.CreatePHI(Val->getType(), "val");
420    BitsToShift->reserveOperandSpace(2);
421    BitsToShift->addIncoming(FRes, Compute);
422
423    // Finally, get the result of the last computation
424    PHINode *RRes = Builder.CreatePHI(Val->getType(), "rres");
425    RRes->reserveOperandSpace(2);
426    RRes->addIncoming(Zero, Compute);
427
428    // Decrement the counter
429    Value *Decr = Builder.CreateSub(Count, One, "decr");
430    Count->addIncoming(Decr, Reverse);
431
432    // Compute the Bit that we want to move
433    Value *Bit = Builder.CreateAnd(BitsToShift, One, "bit");
434
435    // Compute the new value for next iteration.
436    Value *NewVal = Builder.CreateLShr(BitsToShift, One, "rshift");
437    BitsToShift->addIncoming(NewVal, Reverse);
438
439    // Shift the bit into the low bits of the result.
440    Value *NewRes = Builder.CreateShl(RRes, One, "lshift");
441    NewRes = Builder.CreateOr(NewRes, Bit, "addbit");
442    RRes->addIncoming(NewRes, Reverse);
443
444    // Terminate loop if we've moved all the bits.
445    Value *Cond = Builder.CreateICmpEQ(Decr, Zero, "cond");
446    Builder.CreateCondBr(Cond, RsltBlk, Reverse);
447
448    // Finally, in the result block, select one of the two results with a PHI
449    // node and return the result;
450    Builder.SetInsertPoint(RsltBlk);
451    PHINode *BitSelect = Builder.CreatePHI(Val->getType(), "part_select");
452    BitSelect->reserveOperandSpace(2);
453    BitSelect->addIncoming(FRes, Compute);
454    BitSelect->addIncoming(NewRes, Reverse);
455    Builder.CreateRet(BitSelect);
456  }
457
458  // Return a call to the implementation function
459  Builder.SetInsertPoint(CI->getParent(), CI);
460  CallInst *NewCI = Builder.CreateCall3(F, CI->getOperand(1),
461                                        CI->getOperand(2), CI->getOperand(3));
462  NewCI->setName(CI->getName());
463  return NewCI;
464}
465
466/// Convert the llvm.part.set.iX.iY.iZ intrinsic. This intrinsic takes
467/// four integer arguments (iAny %Value, iAny %Replacement, i32 %Low, i32 %High)
468/// The first two arguments can be any bit width. The result is the same width
469/// as %Value. The operation replaces bits between %Low and %High with the value
470/// in %Replacement. If %Replacement is not the same width, it is truncated or
471/// zero extended as appropriate to fit the bits being replaced. If %Low is
472/// greater than %High then the inverse set of bits are replaced.
473/// @brief Lowering of llvm.bit.part.set intrinsic.
474static Instruction *LowerPartSet(CallInst *CI) {
475  IRBuilder<> Builder(*CI->getParent()->getContext());
476
477  // Make sure we're dealing with a part select intrinsic here
478  Function *F = CI->getCalledFunction();
479  const FunctionType *FT = F->getFunctionType();
480  if (!F->isDeclaration() || !FT->getReturnType()->isInteger() ||
481      FT->getNumParams() != 4 || !FT->getParamType(0)->isInteger() ||
482      !FT->getParamType(1)->isInteger() || !FT->getParamType(2)->isInteger() ||
483      !FT->getParamType(3)->isInteger())
484    return CI;
485
486  // Get the intrinsic implementation function by converting all the . to _
487  // in the intrinsic's function name and then reconstructing the function
488  // declaration.
489  std::string Name(F->getName());
490  for (unsigned i = 4; i < Name.length(); ++i)
491    if (Name[i] == '.')
492      Name[i] = '_';
493  Module* M = F->getParent();
494  F = cast<Function>(M->getOrInsertFunction(Name, FT));
495  F->setLinkage(GlobalValue::WeakAnyLinkage);
496
497  // If we haven't defined the impl function yet, do so now
498  if (F->isDeclaration()) {
499    // Get the arguments for the function.
500    Function::arg_iterator args = F->arg_begin();
501    Value* Val = args++; Val->setName("Val");
502    Value* Rep = args++; Rep->setName("Rep");
503    Value* Lo  = args++; Lo->setName("Lo");
504    Value* Hi  = args++; Hi->setName("Hi");
505
506    // Get some types we need
507    const IntegerType* ValTy = cast<IntegerType>(Val->getType());
508    const IntegerType* RepTy = cast<IntegerType>(Rep->getType());
509    uint32_t RepBits = RepTy->getBitWidth();
510
511    // Constant Definitions
512    ConstantInt* RepBitWidth = ConstantInt::get(Type::Int32Ty, RepBits);
513    ConstantInt* RepMask = ConstantInt::getAllOnesValue(RepTy);
514    ConstantInt* ValMask = ConstantInt::getAllOnesValue(ValTy);
515    ConstantInt* One = ConstantInt::get(Type::Int32Ty, 1);
516    ConstantInt* ValOne = ConstantInt::get(ValTy, 1);
517    ConstantInt* Zero = ConstantInt::get(Type::Int32Ty, 0);
518    ConstantInt* ValZero = ConstantInt::get(ValTy, 0);
519
520    // Basic blocks we fill in below.
521    BasicBlock* entry = BasicBlock::Create("entry", F, 0);
522    BasicBlock* large = BasicBlock::Create("large", F, 0);
523    BasicBlock* small = BasicBlock::Create("small", F, 0);
524    BasicBlock* reverse = BasicBlock::Create("reverse", F, 0);
525    BasicBlock* result = BasicBlock::Create("result", F, 0);
526
527    // BASIC BLOCK: entry
528    Builder.SetInsertPoint(entry);
529    // First, get the number of bits that we're placing as an i32
530    Value* is_forward = Builder.CreateICmpULT(Lo, Hi);
531    Value* Hi_pn = Builder.CreateSelect(is_forward, Hi, Lo);
532    Value* Lo_pn = Builder.CreateSelect(is_forward, Lo, Hi);
533    Value* NumBits = Builder.CreateSub(Hi_pn, Lo_pn);
534    NumBits = Builder.CreateAdd(NumBits, One);
535    // Now, convert Lo and Hi to ValTy bit width
536    Lo = Builder.CreateIntCast(Lo_pn, ValTy, /* isSigned */ false);
537    // Determine if the replacement bits are larger than the number of bits we
538    // are replacing and deal with it.
539    Value* is_large = Builder.CreateICmpULT(NumBits, RepBitWidth);
540    Builder.CreateCondBr(is_large, large, small);
541
542    // BASIC BLOCK: large
543    Builder.SetInsertPoint(large);
544    Value* MaskBits = Builder.CreateSub(RepBitWidth, NumBits);
545    MaskBits = Builder.CreateIntCast(MaskBits, RepMask->getType(),
546                                     /* isSigned */ false);
547    Value* Mask1 = Builder.CreateLShr(RepMask, MaskBits);
548    Value* Rep2 = Builder.CreateAnd(Mask1, Rep);
549    Builder.CreateBr(small);
550
551    // BASIC BLOCK: small
552    Builder.SetInsertPoint(small);
553    PHINode* Rep3 = Builder.CreatePHI(RepTy);
554    Rep3->reserveOperandSpace(2);
555    Rep3->addIncoming(Rep2, large);
556    Rep3->addIncoming(Rep, entry);
557    Value* Rep4 = Builder.CreateIntCast(Rep3, ValTy, /* isSigned */ false);
558    Builder.CreateCondBr(is_forward, result, reverse);
559
560    // BASIC BLOCK: reverse (reverses the bits of the replacement)
561    Builder.SetInsertPoint(reverse);
562    // Set up our loop counter as a PHI so we can decrement on each iteration.
563    // We will loop for the number of bits in the replacement value.
564    PHINode *Count = Builder.CreatePHI(Type::Int32Ty, "count");
565    Count->reserveOperandSpace(2);
566    Count->addIncoming(NumBits, small);
567
568    // Get the value that we are shifting bits out of as a PHI because
569    // we'll change this with each iteration.
570    PHINode *BitsToShift = Builder.CreatePHI(Val->getType(), "val");
571    BitsToShift->reserveOperandSpace(2);
572    BitsToShift->addIncoming(Rep4, small);
573
574    // Get the result of the last computation or zero on first iteration
575    PHINode *RRes = Builder.CreatePHI(Val->getType(), "rres");
576    RRes->reserveOperandSpace(2);
577    RRes->addIncoming(ValZero, small);
578
579    // Decrement the loop counter by one
580    Value *Decr = Builder.CreateSub(Count, One);
581    Count->addIncoming(Decr, reverse);
582
583    // Get the bit that we want to move into the result
584    Value *Bit = Builder.CreateAnd(BitsToShift, ValOne);
585
586    // Compute the new value of the bits to shift for the next iteration.
587    Value *NewVal = Builder.CreateLShr(BitsToShift, ValOne);
588    BitsToShift->addIncoming(NewVal, reverse);
589
590    // Shift the bit we extracted into the low bit of the result.
591    Value *NewRes = Builder.CreateShl(RRes, ValOne);
592    NewRes = Builder.CreateOr(NewRes, Bit);
593    RRes->addIncoming(NewRes, reverse);
594
595    // Terminate loop if we've moved all the bits.
596    Value *Cond = Builder.CreateICmpEQ(Decr, Zero);
597    Builder.CreateCondBr(Cond, result, reverse);
598
599    // BASIC BLOCK: result
600    Builder.SetInsertPoint(result);
601    PHINode *Rplcmnt = Builder.CreatePHI(Val->getType());
602    Rplcmnt->reserveOperandSpace(2);
603    Rplcmnt->addIncoming(NewRes, reverse);
604    Rplcmnt->addIncoming(Rep4, small);
605    Value* t0   = Builder.CreateIntCast(NumBits, ValTy, /* isSigned */ false);
606    Value* t1   = Builder.CreateShl(ValMask, Lo);
607    Value* t2   = Builder.CreateNot(t1);
608    Value* t3   = Builder.CreateShl(t1, t0);
609    Value* t4   = Builder.CreateOr(t2, t3);
610    Value* t5   = Builder.CreateAnd(t4, Val);
611    Value* t6   = Builder.CreateShl(Rplcmnt, Lo);
612    Value* Rslt = Builder.CreateOr(t5, t6, "part_set");
613    Builder.CreateRet(Rslt);
614  }
615
616  // Return a call to the implementation function
617  Builder.SetInsertPoint(CI->getParent(), CI);
618  CallInst *NewCI = Builder.CreateCall4(F, CI->getOperand(1),
619                                        CI->getOperand(2), CI->getOperand(3),
620                                        CI->getOperand(4));
621  NewCI->setName(CI->getName());
622  return NewCI;
623}
624
625static void ReplaceFPIntrinsicWithCall(CallInst *CI, const char *Fname,
626                                       const char *Dname,
627                                       const char *LDname) {
628  switch (CI->getOperand(1)->getType()->getTypeID()) {
629  default: assert(0 && "Invalid type in intrinsic"); abort();
630  case Type::FloatTyID:
631    ReplaceCallWith(Fname, CI, CI->op_begin() + 1, CI->op_end(),
632                  Type::FloatTy);
633    break;
634  case Type::DoubleTyID:
635    ReplaceCallWith(Dname, CI, CI->op_begin() + 1, CI->op_end(),
636                  Type::DoubleTy);
637    break;
638  case Type::X86_FP80TyID:
639  case Type::FP128TyID:
640  case Type::PPC_FP128TyID:
641    ReplaceCallWith(LDname, CI, CI->op_begin() + 1, CI->op_end(),
642                  CI->getOperand(1)->getType());
643    break;
644  }
645}
646
647void IntrinsicLowering::LowerIntrinsicCall(CallInst *CI) {
648  IRBuilder<> Builder(CI->getParent(), CI);
649
650  Function *Callee = CI->getCalledFunction();
651  assert(Callee && "Cannot lower an indirect call!");
652
653  switch (Callee->getIntrinsicID()) {
654  case Intrinsic::not_intrinsic:
655    cerr << "Cannot lower a call to a non-intrinsic function '"
656         << Callee->getName() << "'!\n";
657    abort();
658  default:
659    cerr << "Error: Code generator does not support intrinsic function '"
660         << Callee->getName() << "'!\n";
661    abort();
662
663    // The setjmp/longjmp intrinsics should only exist in the code if it was
664    // never optimized (ie, right out of the CFE), or if it has been hacked on
665    // by the lowerinvoke pass.  In both cases, the right thing to do is to
666    // convert the call to an explicit setjmp or longjmp call.
667  case Intrinsic::setjmp: {
668    Value *V = ReplaceCallWith("setjmp", CI, CI->op_begin() + 1, CI->op_end(),
669                               Type::Int32Ty);
670    if (CI->getType() != Type::VoidTy)
671      CI->replaceAllUsesWith(V);
672    break;
673  }
674  case Intrinsic::sigsetjmp:
675     if (CI->getType() != Type::VoidTy)
676       CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
677     break;
678
679  case Intrinsic::longjmp: {
680    ReplaceCallWith("longjmp", CI, CI->op_begin() + 1, CI->op_end(),
681                    Type::VoidTy);
682    break;
683  }
684
685  case Intrinsic::siglongjmp: {
686    // Insert the call to abort
687    ReplaceCallWith("abort", CI, CI->op_end(), CI->op_end(),
688                    Type::VoidTy);
689    break;
690  }
691  case Intrinsic::ctpop:
692    CI->replaceAllUsesWith(LowerCTPOP(CI->getOperand(1), CI));
693    break;
694
695  case Intrinsic::bswap:
696    CI->replaceAllUsesWith(LowerBSWAP(CI->getOperand(1), CI));
697    break;
698
699  case Intrinsic::ctlz:
700    CI->replaceAllUsesWith(LowerCTLZ(CI->getOperand(1), CI));
701    break;
702
703  case Intrinsic::cttz: {
704    // cttz(x) -> ctpop(~X & (X-1))
705    Value *Src = CI->getOperand(1);
706    Value *NotSrc = Builder.CreateNot(Src);
707    NotSrc->setName(Src->getName() + ".not");
708    Value *SrcM1 = ConstantInt::get(Src->getType(), 1);
709    SrcM1 = Builder.CreateSub(Src, SrcM1);
710    Src = LowerCTPOP(Builder.CreateAnd(NotSrc, SrcM1), CI);
711    CI->replaceAllUsesWith(Src);
712    break;
713  }
714
715  case Intrinsic::part_select:
716    CI->replaceAllUsesWith(LowerPartSelect(CI));
717    break;
718
719  case Intrinsic::part_set:
720    CI->replaceAllUsesWith(LowerPartSet(CI));
721    break;
722
723  case Intrinsic::stacksave:
724  case Intrinsic::stackrestore: {
725    if (!Warned)
726      cerr << "WARNING: this target does not support the llvm.stack"
727           << (Callee->getIntrinsicID() == Intrinsic::stacksave ?
728               "save" : "restore") << " intrinsic.\n";
729    Warned = true;
730    if (Callee->getIntrinsicID() == Intrinsic::stacksave)
731      CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
732    break;
733  }
734
735  case Intrinsic::returnaddress:
736  case Intrinsic::frameaddress:
737    cerr << "WARNING: this target does not support the llvm."
738         << (Callee->getIntrinsicID() == Intrinsic::returnaddress ?
739             "return" : "frame") << "address intrinsic.\n";
740    CI->replaceAllUsesWith(ConstantPointerNull::get(
741                                            cast<PointerType>(CI->getType())));
742    break;
743
744  case Intrinsic::prefetch:
745    break;    // Simply strip out prefetches on unsupported architectures
746
747  case Intrinsic::pcmarker:
748    break;    // Simply strip out pcmarker on unsupported architectures
749  case Intrinsic::readcyclecounter: {
750    cerr << "WARNING: this target does not support the llvm.readcyclecoun"
751         << "ter intrinsic.  It is being lowered to a constant 0\n";
752    CI->replaceAllUsesWith(ConstantInt::get(Type::Int64Ty, 0));
753    break;
754  }
755
756  case Intrinsic::dbg_stoppoint:
757  case Intrinsic::dbg_region_start:
758  case Intrinsic::dbg_region_end:
759  case Intrinsic::dbg_func_start:
760  case Intrinsic::dbg_declare:
761    break;    // Simply strip out debugging intrinsics
762
763  case Intrinsic::eh_exception:
764  case Intrinsic::eh_selector_i32:
765  case Intrinsic::eh_selector_i64:
766    CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
767    break;
768
769  case Intrinsic::eh_typeid_for_i32:
770  case Intrinsic::eh_typeid_for_i64:
771    // Return something different to eh_selector.
772    CI->replaceAllUsesWith(ConstantInt::get(CI->getType(), 1));
773    break;
774
775  case Intrinsic::var_annotation:
776    break;   // Strip out annotate intrinsic
777
778  case Intrinsic::memcpy: {
779    const IntegerType *IntPtr = TD.getIntPtrType();
780    Value *Size = Builder.CreateIntCast(CI->getOperand(3), IntPtr,
781                                        /* isSigned */ false);
782    Value *Ops[3];
783    Ops[0] = CI->getOperand(1);
784    Ops[1] = CI->getOperand(2);
785    Ops[2] = Size;
786    ReplaceCallWith("memcpy", CI, Ops, Ops+3, CI->getOperand(1)->getType());
787    break;
788  }
789  case Intrinsic::memmove: {
790    const IntegerType *IntPtr = TD.getIntPtrType();
791    Value *Size = Builder.CreateIntCast(CI->getOperand(3), IntPtr,
792                                        /* isSigned */ false);
793    Value *Ops[3];
794    Ops[0] = CI->getOperand(1);
795    Ops[1] = CI->getOperand(2);
796    Ops[2] = Size;
797    ReplaceCallWith("memmove", CI, Ops, Ops+3, CI->getOperand(1)->getType());
798    break;
799  }
800  case Intrinsic::memset: {
801    const IntegerType *IntPtr = TD.getIntPtrType();
802    Value *Size = Builder.CreateIntCast(CI->getOperand(3), IntPtr,
803                                        /* isSigned */ false);
804    Value *Ops[3];
805    Ops[0] = CI->getOperand(1);
806    // Extend the amount to i32.
807    Ops[1] = Builder.CreateIntCast(CI->getOperand(2), Type::Int32Ty,
808                                   /* isSigned */ false);
809    Ops[2] = Size;
810    ReplaceCallWith("memset", CI, Ops, Ops+3, CI->getOperand(1)->getType());
811    break;
812  }
813  case Intrinsic::sqrt: {
814    ReplaceFPIntrinsicWithCall(CI, "sqrtf", "sqrt", "sqrtl");
815    break;
816  }
817  case Intrinsic::log: {
818    ReplaceFPIntrinsicWithCall(CI, "logf", "log", "logl");
819    break;
820  }
821  case Intrinsic::log2: {
822    ReplaceFPIntrinsicWithCall(CI, "log2f", "log2", "log2l");
823    break;
824  }
825  case Intrinsic::log10: {
826    ReplaceFPIntrinsicWithCall(CI, "log10f", "log10", "log10l");
827    break;
828  }
829  case Intrinsic::exp: {
830    ReplaceFPIntrinsicWithCall(CI, "expf", "exp", "expl");
831    break;
832  }
833  case Intrinsic::exp2: {
834    ReplaceFPIntrinsicWithCall(CI, "exp2f", "exp2", "exp2l");
835    break;
836  }
837  case Intrinsic::pow: {
838    ReplaceFPIntrinsicWithCall(CI, "powf", "pow", "powl");
839    break;
840  }
841  case Intrinsic::flt_rounds:
842     // Lower to "round to the nearest"
843     if (CI->getType() != Type::VoidTy)
844       CI->replaceAllUsesWith(ConstantInt::get(CI->getType(), 1));
845     break;
846  }
847
848  assert(CI->use_empty() &&
849         "Lowering should have eliminated any uses of the intrinsic call!");
850  CI->eraseFromParent();
851}
852