1 //===-- IntrinsicLowering.cpp - Intrinsic Lowering default implementation -===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the IntrinsicLowering class.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Constants.h"
15 #include "llvm/DerivedTypes.h"
16 #include "llvm/Module.h"
17 #include "llvm/Instructions.h"
18 #include "llvm/Type.h"
19 #include "llvm/CodeGen/IntrinsicLowering.h"
20 #include "llvm/Support/Streams.h"
21 #include "llvm/Target/TargetData.h"
22 #include "llvm/ADT/SmallVector.h"
23 #include "llvm/ADT/STLExtras.h"
26 template <class ArgIt>
27 static void EnsureFunctionExists(Module &M, const char *Name,
28 ArgIt ArgBegin, ArgIt ArgEnd,
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));
37 static void EnsureFPIntrinsicsExist(Module &M, Module::iterator I,
39 const char *DName, const char *LDName) {
40 // Insert definitions for all the floating point types.
41 switch((int)I->arg_begin()->getType()->getTypeID()) {
43 EnsureFunctionExists(M, FName, I->arg_begin(), I->arg_end(),
45 case Type::DoubleTyID:
46 EnsureFunctionExists(M, DName, I->arg_begin(), I->arg_end(),
48 case Type::X86_FP80TyID:
50 case Type::PPC_FP128TyID:
51 EnsureFunctionExists(M, LDName, I->arg_begin(), I->arg_end(),
52 I->arg_begin()->getType());
56 /// ReplaceCallWith - This function is used when we want to lower an intrinsic
57 /// call to a call of an external function. This handles hard cases such as
58 /// when there was already a prototype for the external function, and if that
59 /// prototype doesn't match the arguments we expect to pass in.
60 template <class ArgIt>
61 static CallInst *ReplaceCallWith(const char *NewFn, CallInst *CI,
62 ArgIt ArgBegin, ArgIt ArgEnd,
63 const Type *RetTy, Constant *&FCache) {
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 FCache = M->getOrInsertFunction(NewFn,
73 FunctionType::get(RetTy, ParamTys, false));
76 SmallVector<Value *, 8> Args(ArgBegin, ArgEnd);
77 CallInst *NewCI = CallInst::Create(FCache, Args.begin(), Args.end(),
80 CI->replaceAllUsesWith(NewCI);
84 void 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()) {
89 case Intrinsic::setjmp:
90 EnsureFunctionExists(M, "setjmp", I->arg_begin(), I->arg_end(),
93 case Intrinsic::longjmp:
94 EnsureFunctionExists(M, "longjmp", I->arg_begin(), I->arg_end(),
97 case Intrinsic::siglongjmp:
98 EnsureFunctionExists(M, "abort", I->arg_end(), I->arg_end(),
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);
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);
113 case Intrinsic::memset:
114 M.getOrInsertFunction("memset", PointerType::getUnqual(Type::Int8Ty),
115 PointerType::getUnqual(Type::Int8Ty),
117 TD.getIntPtrType(), (Type *)0);
119 case Intrinsic::sqrt:
120 EnsureFPIntrinsicsExist(M, I, "sqrtf", "sqrt", "sqrtl");
123 EnsureFPIntrinsicsExist(M, I, "sinf", "sin", "sinl");
126 EnsureFPIntrinsicsExist(M, I, "cosf", "cos", "cosl");
129 EnsureFPIntrinsicsExist(M, I, "powf", "pow", "powl");
132 EnsureFPIntrinsicsExist(M, I, "logf", "log", "logl");
134 case Intrinsic::log2:
135 EnsureFPIntrinsicsExist(M, I, "log2f", "log2", "log2l");
137 case Intrinsic::log10:
138 EnsureFPIntrinsicsExist(M, I, "log10f", "log10", "log10l");
141 EnsureFPIntrinsicsExist(M, I, "expf", "exp", "expl");
143 case Intrinsic::exp2:
144 EnsureFPIntrinsicsExist(M, I, "exp2f", "exp2", "exp2l");
149 /// LowerBSWAP - Emit the code to lower bswap of V before the specified
151 static Value *LowerBSWAP(Value *V, Instruction *IP) {
152 assert(V->getType()->isInteger() && "Can't bswap a non-integer type!");
154 unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
157 default: assert(0 && "Unhandled type size of value to byteswap!");
159 Value *Tmp1 = BinaryOperator::CreateShl(V,
160 ConstantInt::get(V->getType(),8),"bswap.2",IP);
161 Value *Tmp2 = BinaryOperator::CreateLShr(V,
162 ConstantInt::get(V->getType(),8),"bswap.1",IP);
163 V = BinaryOperator::CreateOr(Tmp1, Tmp2, "bswap.i16", IP);
167 Value *Tmp4 = BinaryOperator::CreateShl(V,
168 ConstantInt::get(V->getType(),24),"bswap.4", IP);
169 Value *Tmp3 = BinaryOperator::CreateShl(V,
170 ConstantInt::get(V->getType(),8),"bswap.3",IP);
171 Value *Tmp2 = BinaryOperator::CreateLShr(V,
172 ConstantInt::get(V->getType(),8),"bswap.2",IP);
173 Value *Tmp1 = BinaryOperator::CreateLShr(V,
174 ConstantInt::get(V->getType(),24),"bswap.1", IP);
175 Tmp3 = BinaryOperator::CreateAnd(Tmp3,
176 ConstantInt::get(Type::Int32Ty, 0xFF0000),
178 Tmp2 = BinaryOperator::CreateAnd(Tmp2,
179 ConstantInt::get(Type::Int32Ty, 0xFF00),
181 Tmp4 = BinaryOperator::CreateOr(Tmp4, Tmp3, "bswap.or1", IP);
182 Tmp2 = BinaryOperator::CreateOr(Tmp2, Tmp1, "bswap.or2", IP);
183 V = BinaryOperator::CreateOr(Tmp4, Tmp2, "bswap.i32", IP);
187 Value *Tmp8 = BinaryOperator::CreateShl(V,
188 ConstantInt::get(V->getType(),56),"bswap.8", IP);
189 Value *Tmp7 = BinaryOperator::CreateShl(V,
190 ConstantInt::get(V->getType(),40),"bswap.7", IP);
191 Value *Tmp6 = BinaryOperator::CreateShl(V,
192 ConstantInt::get(V->getType(),24),"bswap.6", IP);
193 Value *Tmp5 = BinaryOperator::CreateShl(V,
194 ConstantInt::get(V->getType(),8),"bswap.5", IP);
195 Value* Tmp4 = BinaryOperator::CreateLShr(V,
196 ConstantInt::get(V->getType(),8),"bswap.4", IP);
197 Value* Tmp3 = BinaryOperator::CreateLShr(V,
198 ConstantInt::get(V->getType(),24),"bswap.3", IP);
199 Value* Tmp2 = BinaryOperator::CreateLShr(V,
200 ConstantInt::get(V->getType(),40),"bswap.2", IP);
201 Value* Tmp1 = BinaryOperator::CreateLShr(V,
202 ConstantInt::get(V->getType(),56),"bswap.1", IP);
203 Tmp7 = BinaryOperator::CreateAnd(Tmp7,
204 ConstantInt::get(Type::Int64Ty,
205 0xFF000000000000ULL),
207 Tmp6 = BinaryOperator::CreateAnd(Tmp6,
208 ConstantInt::get(Type::Int64Ty, 0xFF0000000000ULL),
210 Tmp5 = BinaryOperator::CreateAnd(Tmp5,
211 ConstantInt::get(Type::Int64Ty, 0xFF00000000ULL),
213 Tmp4 = BinaryOperator::CreateAnd(Tmp4,
214 ConstantInt::get(Type::Int64Ty, 0xFF000000ULL),
216 Tmp3 = BinaryOperator::CreateAnd(Tmp3,
217 ConstantInt::get(Type::Int64Ty, 0xFF0000ULL),
219 Tmp2 = BinaryOperator::CreateAnd(Tmp2,
220 ConstantInt::get(Type::Int64Ty, 0xFF00ULL),
222 Tmp8 = BinaryOperator::CreateOr(Tmp8, Tmp7, "bswap.or1", IP);
223 Tmp6 = BinaryOperator::CreateOr(Tmp6, Tmp5, "bswap.or2", IP);
224 Tmp4 = BinaryOperator::CreateOr(Tmp4, Tmp3, "bswap.or3", IP);
225 Tmp2 = BinaryOperator::CreateOr(Tmp2, Tmp1, "bswap.or4", IP);
226 Tmp8 = BinaryOperator::CreateOr(Tmp8, Tmp6, "bswap.or5", IP);
227 Tmp4 = BinaryOperator::CreateOr(Tmp4, Tmp2, "bswap.or6", IP);
228 V = BinaryOperator::CreateOr(Tmp8, Tmp4, "bswap.i64", IP);
235 /// LowerCTPOP - Emit the code to lower ctpop of V before the specified
237 static Value *LowerCTPOP(Value *V, Instruction *IP) {
238 assert(V->getType()->isInteger() && "Can't ctpop a non-integer type!");
240 static const uint64_t MaskValues[6] = {
241 0x5555555555555555ULL, 0x3333333333333333ULL,
242 0x0F0F0F0F0F0F0F0FULL, 0x00FF00FF00FF00FFULL,
243 0x0000FFFF0000FFFFULL, 0x00000000FFFFFFFFULL
246 unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
247 unsigned WordSize = (BitSize + 63) / 64;
248 Value *Count = ConstantInt::get(V->getType(), 0);
250 for (unsigned n = 0; n < WordSize; ++n) {
251 Value *PartValue = V;
252 for (unsigned i = 1, ct = 0; i < (BitSize>64 ? 64 : BitSize);
254 Value *MaskCst = ConstantInt::get(V->getType(), MaskValues[ct]);
255 Value *LHS = BinaryOperator::CreateAnd(
256 PartValue, MaskCst, "cppop.and1", IP);
257 Value *VShift = BinaryOperator::CreateLShr(PartValue,
258 ConstantInt::get(V->getType(), i), "ctpop.sh", IP);
259 Value *RHS = BinaryOperator::CreateAnd(VShift, MaskCst, "cppop.and2", IP);
260 PartValue = BinaryOperator::CreateAdd(LHS, RHS, "ctpop.step", IP);
262 Count = BinaryOperator::CreateAdd(PartValue, Count, "ctpop.part", IP);
264 V = BinaryOperator::CreateLShr(V, ConstantInt::get(V->getType(), 64),
265 "ctpop.part.sh", IP);
273 /// LowerCTLZ - Emit the code to lower ctlz of V before the specified
275 static Value *LowerCTLZ(Value *V, Instruction *IP) {
277 unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
278 for (unsigned i = 1; i < BitSize; i <<= 1) {
279 Value *ShVal = ConstantInt::get(V->getType(), i);
280 ShVal = BinaryOperator::CreateLShr(V, ShVal, "ctlz.sh", IP);
281 V = BinaryOperator::CreateOr(V, ShVal, "ctlz.step", IP);
284 V = BinaryOperator::CreateNot(V, "", IP);
285 return LowerCTPOP(V, IP);
288 /// Convert the llvm.part.select.iX.iY intrinsic. This intrinsic takes
289 /// three integer arguments. The first argument is the Value from which the
290 /// bits will be selected. It may be of any bit width. The second and third
291 /// arguments specify a range of bits to select with the second argument
292 /// specifying the low bit and the third argument specifying the high bit. Both
293 /// must be type i32. The result is the corresponding selected bits from the
294 /// Value in the same width as the Value (first argument). If the low bit index
295 /// is higher than the high bit index then the inverse selection is done and
296 /// the bits are returned in inverse order.
297 /// @brief Lowering of llvm.part.select intrinsic.
298 static Instruction *LowerPartSelect(CallInst *CI) {
299 // Make sure we're dealing with a part select intrinsic here
300 Function *F = CI->getCalledFunction();
301 const FunctionType *FT = F->getFunctionType();
302 if (!F->isDeclaration() || !FT->getReturnType()->isInteger() ||
303 FT->getNumParams() != 3 || !FT->getParamType(0)->isInteger() ||
304 !FT->getParamType(1)->isInteger() || !FT->getParamType(2)->isInteger())
307 // Get the intrinsic implementation function by converting all the . to _
308 // in the intrinsic's function name and then reconstructing the function
310 std::string Name(F->getName());
311 for (unsigned i = 4; i < Name.length(); ++i)
314 Module* M = F->getParent();
315 F = cast<Function>(M->getOrInsertFunction(Name, FT));
316 F->setLinkage(GlobalValue::WeakLinkage);
318 // If we haven't defined the impl function yet, do so now
319 if (F->isDeclaration()) {
321 // Get the arguments to the function
322 Function::arg_iterator args = F->arg_begin();
323 Value* Val = args++; Val->setName("Val");
324 Value* Lo = args++; Lo->setName("Lo");
325 Value* Hi = args++; Hi->setName("High");
327 // We want to select a range of bits here such that [Hi, Lo] is shifted
328 // down to the low bits. However, it is quite possible that Hi is smaller
329 // than Lo in which case the bits have to be reversed.
331 // Create the blocks we will need for the two cases (forward, reverse)
332 BasicBlock* CurBB = BasicBlock::Create("entry", F);
333 BasicBlock *RevSize = BasicBlock::Create("revsize", CurBB->getParent());
334 BasicBlock *FwdSize = BasicBlock::Create("fwdsize", CurBB->getParent());
335 BasicBlock *Compute = BasicBlock::Create("compute", CurBB->getParent());
336 BasicBlock *Reverse = BasicBlock::Create("reverse", CurBB->getParent());
337 BasicBlock *RsltBlk = BasicBlock::Create("result", CurBB->getParent());
339 // Cast Hi and Lo to the size of Val so the widths are all the same
340 if (Hi->getType() != Val->getType())
341 Hi = CastInst::CreateIntegerCast(Hi, Val->getType(), false,
343 if (Lo->getType() != Val->getType())
344 Lo = CastInst::CreateIntegerCast(Lo, Val->getType(), false,
347 // Compute a few things that both cases will need, up front.
348 Constant* Zero = ConstantInt::get(Val->getType(), 0);
349 Constant* One = ConstantInt::get(Val->getType(), 1);
350 Constant* AllOnes = ConstantInt::getAllOnesValue(Val->getType());
352 // Compare the Hi and Lo bit positions. This is used to determine
353 // which case we have (forward or reverse)
354 ICmpInst *Cmp = new ICmpInst(ICmpInst::ICMP_ULT, Hi, Lo, "less",CurBB);
355 BranchInst::Create(RevSize, FwdSize, Cmp, CurBB);
357 // First, copmute the number of bits in the forward case.
358 Instruction* FBitSize =
359 BinaryOperator::CreateSub(Hi, Lo,"fbits", FwdSize);
360 BranchInst::Create(Compute, FwdSize);
362 // Second, compute the number of bits in the reverse case.
363 Instruction* RBitSize =
364 BinaryOperator::CreateSub(Lo, Hi, "rbits", RevSize);
365 BranchInst::Create(Compute, RevSize);
367 // Now, compute the bit range. Start by getting the bitsize and the shift
368 // amount (either Hi or Lo) from PHI nodes. Then we compute a mask for
369 // the number of bits we want in the range. We shift the bits down to the
370 // least significant bits, apply the mask to zero out unwanted high bits,
371 // and we have computed the "forward" result. It may still need to be
374 // Get the BitSize from one of the two subtractions
375 PHINode *BitSize = PHINode::Create(Val->getType(), "bits", Compute);
376 BitSize->reserveOperandSpace(2);
377 BitSize->addIncoming(FBitSize, FwdSize);
378 BitSize->addIncoming(RBitSize, RevSize);
380 // Get the ShiftAmount as the smaller of Hi/Lo
381 PHINode *ShiftAmt = PHINode::Create(Val->getType(), "shiftamt", Compute);
382 ShiftAmt->reserveOperandSpace(2);
383 ShiftAmt->addIncoming(Lo, FwdSize);
384 ShiftAmt->addIncoming(Hi, RevSize);
386 // Increment the bit size
387 Instruction *BitSizePlusOne =
388 BinaryOperator::CreateAdd(BitSize, One, "bits", Compute);
390 // Create a Mask to zero out the high order bits.
392 BinaryOperator::CreateShl(AllOnes, BitSizePlusOne, "mask", Compute);
393 Mask = BinaryOperator::CreateNot(Mask, "mask", Compute);
395 // Shift the bits down and apply the mask
397 BinaryOperator::CreateLShr(Val, ShiftAmt, "fres", Compute);
398 FRes = BinaryOperator::CreateAnd(FRes, Mask, "fres", Compute);
399 BranchInst::Create(Reverse, RsltBlk, Cmp, Compute);
401 // In the Reverse block we have the mask already in FRes but we must reverse
402 // it by shifting FRes bits right and putting them in RRes by shifting them
405 // First set up our loop counters
406 PHINode *Count = PHINode::Create(Val->getType(), "count", Reverse);
407 Count->reserveOperandSpace(2);
408 Count->addIncoming(BitSizePlusOne, Compute);
410 // Next, get the value that we are shifting.
411 PHINode *BitsToShift = PHINode::Create(Val->getType(), "val", Reverse);
412 BitsToShift->reserveOperandSpace(2);
413 BitsToShift->addIncoming(FRes, Compute);
415 // Finally, get the result of the last computation
416 PHINode *RRes = PHINode::Create(Val->getType(), "rres", Reverse);
417 RRes->reserveOperandSpace(2);
418 RRes->addIncoming(Zero, Compute);
420 // Decrement the counter
421 Instruction *Decr = BinaryOperator::CreateSub(Count, One, "decr", Reverse);
422 Count->addIncoming(Decr, Reverse);
424 // Compute the Bit that we want to move
426 BinaryOperator::CreateAnd(BitsToShift, One, "bit", Reverse);
428 // Compute the new value for next iteration.
429 Instruction *NewVal =
430 BinaryOperator::CreateLShr(BitsToShift, One, "rshift", Reverse);
431 BitsToShift->addIncoming(NewVal, Reverse);
433 // Shift the bit into the low bits of the result.
434 Instruction *NewRes =
435 BinaryOperator::CreateShl(RRes, One, "lshift", Reverse);
436 NewRes = BinaryOperator::CreateOr(NewRes, Bit, "addbit", Reverse);
437 RRes->addIncoming(NewRes, Reverse);
439 // Terminate loop if we've moved all the bits.
441 new ICmpInst(ICmpInst::ICMP_EQ, Decr, Zero, "cond", Reverse);
442 BranchInst::Create(RsltBlk, Reverse, Cond, Reverse);
444 // Finally, in the result block, select one of the two results with a PHI
445 // node and return the result;
447 PHINode *BitSelect = PHINode::Create(Val->getType(), "part_select", CurBB);
448 BitSelect->reserveOperandSpace(2);
449 BitSelect->addIncoming(FRes, Compute);
450 BitSelect->addIncoming(NewRes, Reverse);
451 ReturnInst::Create(BitSelect, CurBB);
454 // Return a call to the implementation function
460 return CallInst::Create(F, Args, array_endof(Args), CI->getName(), CI);
463 /// Convert the llvm.part.set.iX.iY.iZ intrinsic. This intrinsic takes
464 /// four integer arguments (iAny %Value, iAny %Replacement, i32 %Low, i32 %High)
465 /// The first two arguments can be any bit width. The result is the same width
466 /// as %Value. The operation replaces bits between %Low and %High with the value
467 /// in %Replacement. If %Replacement is not the same width, it is truncated or
468 /// zero extended as appropriate to fit the bits being replaced. If %Low is
469 /// greater than %High then the inverse set of bits are replaced.
470 /// @brief Lowering of llvm.bit.part.set intrinsic.
471 static Instruction *LowerPartSet(CallInst *CI) {
472 // Make sure we're dealing with a part select intrinsic here
473 Function *F = CI->getCalledFunction();
474 const FunctionType *FT = F->getFunctionType();
475 if (!F->isDeclaration() || !FT->getReturnType()->isInteger() ||
476 FT->getNumParams() != 4 || !FT->getParamType(0)->isInteger() ||
477 !FT->getParamType(1)->isInteger() || !FT->getParamType(2)->isInteger() ||
478 !FT->getParamType(3)->isInteger())
481 // Get the intrinsic implementation function by converting all the . to _
482 // in the intrinsic's function name and then reconstructing the function
484 std::string Name(F->getName());
485 for (unsigned i = 4; i < Name.length(); ++i)
488 Module* M = F->getParent();
489 F = cast<Function>(M->getOrInsertFunction(Name, FT));
490 F->setLinkage(GlobalValue::WeakLinkage);
492 // If we haven't defined the impl function yet, do so now
493 if (F->isDeclaration()) {
494 // Get the arguments for the function.
495 Function::arg_iterator args = F->arg_begin();
496 Value* Val = args++; Val->setName("Val");
497 Value* Rep = args++; Rep->setName("Rep");
498 Value* Lo = args++; Lo->setName("Lo");
499 Value* Hi = args++; Hi->setName("Hi");
501 // Get some types we need
502 const IntegerType* ValTy = cast<IntegerType>(Val->getType());
503 const IntegerType* RepTy = cast<IntegerType>(Rep->getType());
504 uint32_t ValBits = ValTy->getBitWidth();
505 uint32_t RepBits = RepTy->getBitWidth();
507 // Constant Definitions
508 ConstantInt* RepBitWidth = ConstantInt::get(Type::Int32Ty, RepBits);
509 ConstantInt* RepMask = ConstantInt::getAllOnesValue(RepTy);
510 ConstantInt* ValMask = ConstantInt::getAllOnesValue(ValTy);
511 ConstantInt* One = ConstantInt::get(Type::Int32Ty, 1);
512 ConstantInt* ValOne = ConstantInt::get(ValTy, 1);
513 ConstantInt* Zero = ConstantInt::get(Type::Int32Ty, 0);
514 ConstantInt* ValZero = ConstantInt::get(ValTy, 0);
516 // Basic blocks we fill in below.
517 BasicBlock* entry = BasicBlock::Create("entry", F, 0);
518 BasicBlock* large = BasicBlock::Create("large", F, 0);
519 BasicBlock* small = BasicBlock::Create("small", F, 0);
520 BasicBlock* reverse = BasicBlock::Create("reverse", F, 0);
521 BasicBlock* result = BasicBlock::Create("result", F, 0);
523 // BASIC BLOCK: entry
524 // First, get the number of bits that we're placing as an i32
525 ICmpInst* is_forward =
526 new ICmpInst(ICmpInst::ICMP_ULT, Lo, Hi, "", entry);
527 SelectInst* Hi_pn = SelectInst::Create(is_forward, Hi, Lo, "", entry);
528 SelectInst* Lo_pn = SelectInst::Create(is_forward, Lo, Hi, "", entry);
529 BinaryOperator* NumBits = BinaryOperator::CreateSub(Hi_pn, Lo_pn, "",entry);
530 NumBits = BinaryOperator::CreateAdd(NumBits, One, "", entry);
531 // Now, convert Lo and Hi to ValTy bit width
533 Lo = new ZExtInst(Lo_pn, ValTy, "", entry);
534 } else if (ValBits < 32) {
535 Lo = new TruncInst(Lo_pn, ValTy, "", entry);
537 // Determine if the replacement bits are larger than the number of bits we
538 // are replacing and deal with it.
540 new ICmpInst(ICmpInst::ICMP_ULT, NumBits, RepBitWidth, "", entry);
541 BranchInst::Create(large, small, is_large, entry);
543 // BASIC BLOCK: large
544 Instruction* MaskBits =
545 BinaryOperator::CreateSub(RepBitWidth, NumBits, "", large);
546 MaskBits = CastInst::CreateIntegerCast(MaskBits, RepMask->getType(),
548 BinaryOperator* Mask1 =
549 BinaryOperator::CreateLShr(RepMask, MaskBits, "", large);
550 BinaryOperator* Rep2 = BinaryOperator::CreateAnd(Mask1, Rep, "", large);
551 BranchInst::Create(small, large);
553 // BASIC BLOCK: small
554 PHINode* Rep3 = PHINode::Create(RepTy, "", small);
555 Rep3->reserveOperandSpace(2);
556 Rep3->addIncoming(Rep2, large);
557 Rep3->addIncoming(Rep, entry);
559 if (ValBits > RepBits)
560 Rep4 = new ZExtInst(Rep3, ValTy, "", small);
561 else if (ValBits < RepBits)
562 Rep4 = new TruncInst(Rep3, ValTy, "", small);
563 BranchInst::Create(result, reverse, is_forward, small);
565 // BASIC BLOCK: reverse (reverses the bits of the replacement)
566 // Set up our loop counter as a PHI so we can decrement on each iteration.
567 // We will loop for the number of bits in the replacement value.
568 PHINode *Count = PHINode::Create(Type::Int32Ty, "count", reverse);
569 Count->reserveOperandSpace(2);
570 Count->addIncoming(NumBits, small);
572 // Get the value that we are shifting bits out of as a PHI because
573 // we'll change this with each iteration.
574 PHINode *BitsToShift = PHINode::Create(Val->getType(), "val", reverse);
575 BitsToShift->reserveOperandSpace(2);
576 BitsToShift->addIncoming(Rep4, small);
578 // Get the result of the last computation or zero on first iteration
579 PHINode *RRes = PHINode::Create(Val->getType(), "rres", reverse);
580 RRes->reserveOperandSpace(2);
581 RRes->addIncoming(ValZero, small);
583 // Decrement the loop counter by one
584 Instruction *Decr = BinaryOperator::CreateSub(Count, One, "", reverse);
585 Count->addIncoming(Decr, reverse);
587 // Get the bit that we want to move into the result
588 Value *Bit = BinaryOperator::CreateAnd(BitsToShift, ValOne, "", reverse);
590 // Compute the new value of the bits to shift for the next iteration.
591 Value *NewVal = BinaryOperator::CreateLShr(BitsToShift, ValOne,"", reverse);
592 BitsToShift->addIncoming(NewVal, reverse);
594 // Shift the bit we extracted into the low bit of the result.
595 Instruction *NewRes = BinaryOperator::CreateShl(RRes, ValOne, "", reverse);
596 NewRes = BinaryOperator::CreateOr(NewRes, Bit, "", reverse);
597 RRes->addIncoming(NewRes, reverse);
599 // Terminate loop if we've moved all the bits.
600 ICmpInst *Cond = new ICmpInst(ICmpInst::ICMP_EQ, Decr, Zero, "", reverse);
601 BranchInst::Create(result, reverse, Cond, reverse);
603 // BASIC BLOCK: result
604 PHINode *Rplcmnt = PHINode::Create(Val->getType(), "", result);
605 Rplcmnt->reserveOperandSpace(2);
606 Rplcmnt->addIncoming(NewRes, reverse);
607 Rplcmnt->addIncoming(Rep4, small);
608 Value* t0 = CastInst::CreateIntegerCast(NumBits,ValTy,false,"",result);
609 Value* t1 = BinaryOperator::CreateShl(ValMask, Lo, "", result);
610 Value* t2 = BinaryOperator::CreateNot(t1, "", result);
611 Value* t3 = BinaryOperator::CreateShl(t1, t0, "", result);
612 Value* t4 = BinaryOperator::CreateOr(t2, t3, "", result);
613 Value* t5 = BinaryOperator::CreateAnd(t4, Val, "", result);
614 Value* t6 = BinaryOperator::CreateShl(Rplcmnt, Lo, "", result);
615 Value* Rslt = BinaryOperator::CreateOr(t5, t6, "part_set", result);
616 ReturnInst::Create(Rslt, result);
619 // Return a call to the implementation function
626 return CallInst::Create(F, Args, array_endof(Args), CI->getName(), CI);
629 static void ReplaceFPIntrinsicWithCall(CallInst *CI, Constant *FCache,
630 Constant *DCache, Constant *LDCache,
631 const char *Fname, const char *Dname,
632 const char *LDname) {
633 switch (CI->getOperand(1)->getType()->getTypeID()) {
634 default: assert(0 && "Invalid type in intrinsic"); abort();
635 case Type::FloatTyID:
636 ReplaceCallWith(Fname, CI, CI->op_begin()+1, CI->op_end(),
637 Type::FloatTy, FCache);
639 case Type::DoubleTyID:
640 ReplaceCallWith(Dname, CI, CI->op_begin()+1, CI->op_end(),
641 Type::DoubleTy, DCache);
643 case Type::X86_FP80TyID:
644 case Type::FP128TyID:
645 case Type::PPC_FP128TyID:
646 ReplaceCallWith(LDname, CI, CI->op_begin()+1, CI->op_end(),
647 CI->getOperand(1)->getType(), LDCache);
652 void IntrinsicLowering::LowerIntrinsicCall(CallInst *CI) {
653 Function *Callee = CI->getCalledFunction();
654 assert(Callee && "Cannot lower an indirect call!");
656 switch (Callee->getIntrinsicID()) {
657 case Intrinsic::not_intrinsic:
658 cerr << "Cannot lower a call to a non-intrinsic function '"
659 << Callee->getName() << "'!\n";
662 cerr << "Error: Code generator does not support intrinsic function '"
663 << Callee->getName() << "'!\n";
666 // The setjmp/longjmp intrinsics should only exist in the code if it was
667 // never optimized (ie, right out of the CFE), or if it has been hacked on
668 // by the lowerinvoke pass. In both cases, the right thing to do is to
669 // convert the call to an explicit setjmp or longjmp call.
670 case Intrinsic::setjmp: {
671 static Constant *SetjmpFCache = 0;
672 Value *V = ReplaceCallWith("setjmp", CI, CI->op_begin()+1, CI->op_end(),
673 Type::Int32Ty, SetjmpFCache);
674 if (CI->getType() != Type::VoidTy)
675 CI->replaceAllUsesWith(V);
678 case Intrinsic::sigsetjmp:
679 if (CI->getType() != Type::VoidTy)
680 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
683 case Intrinsic::longjmp: {
684 static Constant *LongjmpFCache = 0;
685 ReplaceCallWith("longjmp", CI, CI->op_begin()+1, CI->op_end(),
686 Type::VoidTy, LongjmpFCache);
690 case Intrinsic::siglongjmp: {
691 // Insert the call to abort
692 static Constant *AbortFCache = 0;
693 ReplaceCallWith("abort", CI, CI->op_end(), CI->op_end(),
694 Type::VoidTy, AbortFCache);
697 case Intrinsic::ctpop:
698 CI->replaceAllUsesWith(LowerCTPOP(CI->getOperand(1), CI));
701 case Intrinsic::bswap:
702 CI->replaceAllUsesWith(LowerBSWAP(CI->getOperand(1), CI));
705 case Intrinsic::ctlz:
706 CI->replaceAllUsesWith(LowerCTLZ(CI->getOperand(1), CI));
709 case Intrinsic::cttz: {
710 // cttz(x) -> ctpop(~X & (X-1))
711 Value *Src = CI->getOperand(1);
712 Value *NotSrc = BinaryOperator::CreateNot(Src, Src->getName()+".not", CI);
713 Value *SrcM1 = ConstantInt::get(Src->getType(), 1);
714 SrcM1 = BinaryOperator::CreateSub(Src, SrcM1, "", CI);
715 Src = LowerCTPOP(BinaryOperator::CreateAnd(NotSrc, SrcM1, "", CI), CI);
716 CI->replaceAllUsesWith(Src);
720 case Intrinsic::part_select:
721 CI->replaceAllUsesWith(LowerPartSelect(CI));
724 case Intrinsic::part_set:
725 CI->replaceAllUsesWith(LowerPartSet(CI));
728 case Intrinsic::stacksave:
729 case Intrinsic::stackrestore: {
730 static bool Warned = false;
732 cerr << "WARNING: this target does not support the llvm.stack"
733 << (Callee->getIntrinsicID() == Intrinsic::stacksave ?
734 "save" : "restore") << " intrinsic.\n";
736 if (Callee->getIntrinsicID() == Intrinsic::stacksave)
737 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
741 case Intrinsic::returnaddress:
742 case Intrinsic::frameaddress:
743 cerr << "WARNING: this target does not support the llvm."
744 << (Callee->getIntrinsicID() == Intrinsic::returnaddress ?
745 "return" : "frame") << "address intrinsic.\n";
746 CI->replaceAllUsesWith(ConstantPointerNull::get(
747 cast<PointerType>(CI->getType())));
750 case Intrinsic::prefetch:
751 break; // Simply strip out prefetches on unsupported architectures
753 case Intrinsic::pcmarker:
754 break; // Simply strip out pcmarker on unsupported architectures
755 case Intrinsic::readcyclecounter: {
756 cerr << "WARNING: this target does not support the llvm.readcyclecoun"
757 << "ter intrinsic. It is being lowered to a constant 0\n";
758 CI->replaceAllUsesWith(ConstantInt::get(Type::Int64Ty, 0));
762 case Intrinsic::dbg_stoppoint:
763 case Intrinsic::dbg_region_start:
764 case Intrinsic::dbg_region_end:
765 case Intrinsic::dbg_func_start:
766 case Intrinsic::dbg_declare:
767 break; // Simply strip out debugging intrinsics
769 case Intrinsic::eh_exception:
770 case Intrinsic::eh_selector_i32:
771 case Intrinsic::eh_selector_i64:
772 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
775 case Intrinsic::eh_typeid_for_i32:
776 case Intrinsic::eh_typeid_for_i64:
777 // Return something different to eh_selector.
778 CI->replaceAllUsesWith(ConstantInt::get(CI->getType(), 1));
781 case Intrinsic::var_annotation:
782 break; // Strip out annotate intrinsic
784 case Intrinsic::memcpy: {
785 static Constant *MemcpyFCache = 0;
786 Value *Size = CI->getOperand(3);
787 const Type *IntPtr = TD.getIntPtrType();
788 if (Size->getType()->getPrimitiveSizeInBits() <
789 IntPtr->getPrimitiveSizeInBits())
790 Size = new ZExtInst(Size, IntPtr, "", CI);
791 else if (Size->getType()->getPrimitiveSizeInBits() >
792 IntPtr->getPrimitiveSizeInBits())
793 Size = new TruncInst(Size, IntPtr, "", CI);
795 Ops[0] = CI->getOperand(1);
796 Ops[1] = CI->getOperand(2);
798 ReplaceCallWith("memcpy", CI, Ops, Ops+3, CI->getOperand(1)->getType(),
802 case Intrinsic::memmove: {
803 static Constant *MemmoveFCache = 0;
804 Value *Size = CI->getOperand(3);
805 const Type *IntPtr = TD.getIntPtrType();
806 if (Size->getType()->getPrimitiveSizeInBits() <
807 IntPtr->getPrimitiveSizeInBits())
808 Size = new ZExtInst(Size, IntPtr, "", CI);
809 else if (Size->getType()->getPrimitiveSizeInBits() >
810 IntPtr->getPrimitiveSizeInBits())
811 Size = new TruncInst(Size, IntPtr, "", CI);
813 Ops[0] = CI->getOperand(1);
814 Ops[1] = CI->getOperand(2);
816 ReplaceCallWith("memmove", CI, Ops, Ops+3, CI->getOperand(1)->getType(),
820 case Intrinsic::memset: {
821 static Constant *MemsetFCache = 0;
822 Value *Size = CI->getOperand(3);
823 const Type *IntPtr = TD.getIntPtrType();
824 if (Size->getType()->getPrimitiveSizeInBits() <
825 IntPtr->getPrimitiveSizeInBits())
826 Size = new ZExtInst(Size, IntPtr, "", CI);
827 else if (Size->getType()->getPrimitiveSizeInBits() >
828 IntPtr->getPrimitiveSizeInBits())
829 Size = new TruncInst(Size, IntPtr, "", CI);
831 Ops[0] = CI->getOperand(1);
832 // Extend the amount to i32.
833 Ops[1] = new ZExtInst(CI->getOperand(2), Type::Int32Ty, "", CI);
835 ReplaceCallWith("memset", CI, Ops, Ops+3, CI->getOperand(1)->getType(),
839 case Intrinsic::sqrt: {
840 static Constant *sqrtFCache = 0;
841 static Constant *sqrtDCache = 0;
842 static Constant *sqrtLDCache = 0;
843 ReplaceFPIntrinsicWithCall(CI, sqrtFCache, sqrtDCache, sqrtLDCache,
844 "sqrtf", "sqrt", "sqrtl");
847 case Intrinsic::log: {
848 static Constant *logFCache = 0;
849 static Constant *logDCache = 0;
850 static Constant *logLDCache = 0;
851 ReplaceFPIntrinsicWithCall(CI, logFCache, logDCache, logLDCache,
852 "logf", "log", "logl");
855 case Intrinsic::log2: {
856 static Constant *log2FCache = 0;
857 static Constant *log2DCache = 0;
858 static Constant *log2LDCache = 0;
859 ReplaceFPIntrinsicWithCall(CI, log2FCache, log2DCache, log2LDCache,
860 "log2f", "log2", "log2l");
863 case Intrinsic::log10: {
864 static Constant *log10FCache = 0;
865 static Constant *log10DCache = 0;
866 static Constant *log10LDCache = 0;
867 ReplaceFPIntrinsicWithCall(CI, log10FCache, log10DCache, log10LDCache,
868 "log10f", "log10", "log10l");
871 case Intrinsic::exp: {
872 static Constant *expFCache = 0;
873 static Constant *expDCache = 0;
874 static Constant *expLDCache = 0;
875 ReplaceFPIntrinsicWithCall(CI, expFCache, expDCache, expLDCache,
876 "expf", "exp", "expl");
879 case Intrinsic::exp2: {
880 static Constant *exp2FCache = 0;
881 static Constant *exp2DCache = 0;
882 static Constant *exp2LDCache = 0;
883 ReplaceFPIntrinsicWithCall(CI, exp2FCache, exp2DCache, exp2LDCache,
884 "exp2f", "exp2", "exp2l");
887 case Intrinsic::pow: {
888 static Constant *powFCache = 0;
889 static Constant *powDCache = 0;
890 static Constant *powLDCache = 0;
891 ReplaceFPIntrinsicWithCall(CI, powFCache, powDCache, powLDCache,
892 "powf", "pow", "powl");
895 case Intrinsic::flt_rounds:
896 // Lower to "round to the nearest"
897 if (CI->getType() != Type::VoidTy)
898 CI->replaceAllUsesWith(ConstantInt::get(CI->getType(), 1));
902 assert(CI->use_empty() &&
903 "Lowering should have eliminated any uses of the intrinsic call!");
904 CI->eraseFromParent();