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 /// ReplaceCallWith - This function is used when we want to lower an intrinsic
38 /// call to a call of an external function. This handles hard cases such as
39 /// when there was already a prototype for the external function, and if that
40 /// prototype doesn't match the arguments we expect to pass in.
41 template <class ArgIt>
42 static CallInst *ReplaceCallWith(const char *NewFn, CallInst *CI,
43 ArgIt ArgBegin, ArgIt ArgEnd,
44 const Type *RetTy, Constant *&FCache) {
46 // If we haven't already looked up this function, check to see if the
47 // program already contains a function with this name.
48 Module *M = CI->getParent()->getParent()->getParent();
49 // Get or insert the definition now.
50 std::vector<const Type *> ParamTys;
51 for (ArgIt I = ArgBegin; I != ArgEnd; ++I)
52 ParamTys.push_back((*I)->getType());
53 FCache = M->getOrInsertFunction(NewFn,
54 FunctionType::get(RetTy, ParamTys, false));
57 SmallVector<Value *, 8> Args(ArgBegin, ArgEnd);
58 CallInst *NewCI = CallInst::Create(FCache, Args.begin(), Args.end(),
61 CI->replaceAllUsesWith(NewCI);
65 void IntrinsicLowering::AddPrototypes(Module &M) {
66 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
67 if (I->isDeclaration() && !I->use_empty())
68 switch (I->getIntrinsicID()) {
70 case Intrinsic::setjmp:
71 EnsureFunctionExists(M, "setjmp", I->arg_begin(), I->arg_end(),
74 case Intrinsic::longjmp:
75 EnsureFunctionExists(M, "longjmp", I->arg_begin(), I->arg_end(),
78 case Intrinsic::siglongjmp:
79 EnsureFunctionExists(M, "abort", I->arg_end(), I->arg_end(),
82 case Intrinsic::memcpy_i32:
83 case Intrinsic::memcpy_i64:
84 M.getOrInsertFunction("memcpy", PointerType::getUnqual(Type::Int8Ty),
85 PointerType::getUnqual(Type::Int8Ty),
86 PointerType::getUnqual(Type::Int8Ty),
87 TD.getIntPtrType(), (Type *)0);
89 case Intrinsic::memmove_i32:
90 case Intrinsic::memmove_i64:
91 M.getOrInsertFunction("memmove", PointerType::getUnqual(Type::Int8Ty),
92 PointerType::getUnqual(Type::Int8Ty),
93 PointerType::getUnqual(Type::Int8Ty),
94 TD.getIntPtrType(), (Type *)0);
96 case Intrinsic::memset_i32:
97 case Intrinsic::memset_i64:
98 M.getOrInsertFunction("memset", PointerType::getUnqual(Type::Int8Ty),
99 PointerType::getUnqual(Type::Int8Ty),
101 TD.getIntPtrType(), (Type *)0);
103 case Intrinsic::sqrt:
104 switch((int)I->arg_begin()->getType()->getTypeID()) {
105 case Type::FloatTyID:
106 EnsureFunctionExists(M, "sqrtf", I->arg_begin(), I->arg_end(),
108 case Type::DoubleTyID:
109 EnsureFunctionExists(M, "sqrt", I->arg_begin(), I->arg_end(),
111 case Type::X86_FP80TyID:
112 case Type::FP128TyID:
113 case Type::PPC_FP128TyID:
114 EnsureFunctionExists(M, "sqrtl", I->arg_begin(), I->arg_end(),
115 I->arg_begin()->getType());
119 switch((int)I->arg_begin()->getType()->getTypeID()) {
120 case Type::FloatTyID:
121 EnsureFunctionExists(M, "sinf", I->arg_begin(), I->arg_end(),
123 case Type::DoubleTyID:
124 EnsureFunctionExists(M, "sin", I->arg_begin(), I->arg_end(),
126 case Type::X86_FP80TyID:
127 case Type::FP128TyID:
128 case Type::PPC_FP128TyID:
129 EnsureFunctionExists(M, "sinl", I->arg_begin(), I->arg_end(),
130 I->arg_begin()->getType());
134 switch((int)I->arg_begin()->getType()->getTypeID()) {
135 case Type::FloatTyID:
136 EnsureFunctionExists(M, "cosf", I->arg_begin(), I->arg_end(),
138 case Type::DoubleTyID:
139 EnsureFunctionExists(M, "cos", I->arg_begin(), I->arg_end(),
141 case Type::X86_FP80TyID:
142 case Type::FP128TyID:
143 case Type::PPC_FP128TyID:
144 EnsureFunctionExists(M, "cosl", I->arg_begin(), I->arg_end(),
145 I->arg_begin()->getType());
149 switch((int)I->arg_begin()->getType()->getTypeID()) {
150 case Type::FloatTyID:
151 EnsureFunctionExists(M, "powf", I->arg_begin(), I->arg_end(),
153 case Type::DoubleTyID:
154 EnsureFunctionExists(M, "pow", I->arg_begin(), I->arg_end(),
156 case Type::X86_FP80TyID:
157 case Type::FP128TyID:
158 case Type::PPC_FP128TyID:
159 EnsureFunctionExists(M, "powl", I->arg_begin(), I->arg_end(),
160 I->arg_begin()->getType());
164 switch((int)I->arg_begin()->getType()->getTypeID()) {
165 case Type::FloatTyID:
166 EnsureFunctionExists(M, "logf", I->arg_begin(), I->arg_end(),
168 case Type::DoubleTyID:
169 EnsureFunctionExists(M, "log", I->arg_begin(), I->arg_end(),
171 case Type::X86_FP80TyID:
172 case Type::FP128TyID:
173 case Type::PPC_FP128TyID:
174 EnsureFunctionExists(M, "logl", I->arg_begin(), I->arg_end(),
175 I->arg_begin()->getType());
178 case Intrinsic::log2:
179 switch((int)I->arg_begin()->getType()->getTypeID()) {
180 case Type::FloatTyID:
181 EnsureFunctionExists(M, "log2f", I->arg_begin(), I->arg_end(),
183 case Type::DoubleTyID:
184 EnsureFunctionExists(M, "log2", I->arg_begin(), I->arg_end(),
186 case Type::X86_FP80TyID:
187 case Type::FP128TyID:
188 case Type::PPC_FP128TyID:
189 EnsureFunctionExists(M, "log2l", I->arg_begin(), I->arg_end(),
190 I->arg_begin()->getType());
193 case Intrinsic::log10:
194 switch((int)I->arg_begin()->getType()->getTypeID()) {
195 case Type::FloatTyID:
196 EnsureFunctionExists(M, "log10f", I->arg_begin(), I->arg_end(),
198 case Type::DoubleTyID:
199 EnsureFunctionExists(M, "log10", I->arg_begin(), I->arg_end(),
201 case Type::X86_FP80TyID:
202 case Type::FP128TyID:
203 case Type::PPC_FP128TyID:
204 EnsureFunctionExists(M, "log10l", I->arg_begin(), I->arg_end(),
205 I->arg_begin()->getType());
209 switch((int)I->arg_begin()->getType()->getTypeID()) {
210 case Type::FloatTyID:
211 EnsureFunctionExists(M, "expf", I->arg_begin(), I->arg_end(),
213 case Type::DoubleTyID:
214 EnsureFunctionExists(M, "exp", I->arg_begin(), I->arg_end(),
216 case Type::X86_FP80TyID:
217 case Type::FP128TyID:
218 case Type::PPC_FP128TyID:
219 EnsureFunctionExists(M, "expl", I->arg_begin(), I->arg_end(),
220 I->arg_begin()->getType());
223 case Intrinsic::exp2:
224 switch((int)I->arg_begin()->getType()->getTypeID()) {
225 case Type::FloatTyID:
226 EnsureFunctionExists(M, "exp2f", I->arg_begin(), I->arg_end(),
228 case Type::DoubleTyID:
229 EnsureFunctionExists(M, "exp2", I->arg_begin(), I->arg_end(),
231 case Type::X86_FP80TyID:
232 case Type::FP128TyID:
233 case Type::PPC_FP128TyID:
234 EnsureFunctionExists(M, "exp2l", I->arg_begin(), I->arg_end(),
235 I->arg_begin()->getType());
241 /// LowerBSWAP - Emit the code to lower bswap of V before the specified
243 static Value *LowerBSWAP(Value *V, Instruction *IP) {
244 assert(V->getType()->isInteger() && "Can't bswap a non-integer type!");
246 unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
249 default: assert(0 && "Unhandled type size of value to byteswap!");
251 Value *Tmp1 = BinaryOperator::CreateShl(V,
252 ConstantInt::get(V->getType(),8),"bswap.2",IP);
253 Value *Tmp2 = BinaryOperator::CreateLShr(V,
254 ConstantInt::get(V->getType(),8),"bswap.1",IP);
255 V = BinaryOperator::CreateOr(Tmp1, Tmp2, "bswap.i16", IP);
259 Value *Tmp4 = BinaryOperator::CreateShl(V,
260 ConstantInt::get(V->getType(),24),"bswap.4", IP);
261 Value *Tmp3 = BinaryOperator::CreateShl(V,
262 ConstantInt::get(V->getType(),8),"bswap.3",IP);
263 Value *Tmp2 = BinaryOperator::CreateLShr(V,
264 ConstantInt::get(V->getType(),8),"bswap.2",IP);
265 Value *Tmp1 = BinaryOperator::CreateLShr(V,
266 ConstantInt::get(V->getType(),24),"bswap.1", IP);
267 Tmp3 = BinaryOperator::CreateAnd(Tmp3,
268 ConstantInt::get(Type::Int32Ty, 0xFF0000),
270 Tmp2 = BinaryOperator::CreateAnd(Tmp2,
271 ConstantInt::get(Type::Int32Ty, 0xFF00),
273 Tmp4 = BinaryOperator::CreateOr(Tmp4, Tmp3, "bswap.or1", IP);
274 Tmp2 = BinaryOperator::CreateOr(Tmp2, Tmp1, "bswap.or2", IP);
275 V = BinaryOperator::CreateOr(Tmp4, Tmp2, "bswap.i32", IP);
279 Value *Tmp8 = BinaryOperator::CreateShl(V,
280 ConstantInt::get(V->getType(),56),"bswap.8", IP);
281 Value *Tmp7 = BinaryOperator::CreateShl(V,
282 ConstantInt::get(V->getType(),40),"bswap.7", IP);
283 Value *Tmp6 = BinaryOperator::CreateShl(V,
284 ConstantInt::get(V->getType(),24),"bswap.6", IP);
285 Value *Tmp5 = BinaryOperator::CreateShl(V,
286 ConstantInt::get(V->getType(),8),"bswap.5", IP);
287 Value* Tmp4 = BinaryOperator::CreateLShr(V,
288 ConstantInt::get(V->getType(),8),"bswap.4", IP);
289 Value* Tmp3 = BinaryOperator::CreateLShr(V,
290 ConstantInt::get(V->getType(),24),"bswap.3", IP);
291 Value* Tmp2 = BinaryOperator::CreateLShr(V,
292 ConstantInt::get(V->getType(),40),"bswap.2", IP);
293 Value* Tmp1 = BinaryOperator::CreateLShr(V,
294 ConstantInt::get(V->getType(),56),"bswap.1", IP);
295 Tmp7 = BinaryOperator::CreateAnd(Tmp7,
296 ConstantInt::get(Type::Int64Ty,
297 0xFF000000000000ULL),
299 Tmp6 = BinaryOperator::CreateAnd(Tmp6,
300 ConstantInt::get(Type::Int64Ty, 0xFF0000000000ULL),
302 Tmp5 = BinaryOperator::CreateAnd(Tmp5,
303 ConstantInt::get(Type::Int64Ty, 0xFF00000000ULL),
305 Tmp4 = BinaryOperator::CreateAnd(Tmp4,
306 ConstantInt::get(Type::Int64Ty, 0xFF000000ULL),
308 Tmp3 = BinaryOperator::CreateAnd(Tmp3,
309 ConstantInt::get(Type::Int64Ty, 0xFF0000ULL),
311 Tmp2 = BinaryOperator::CreateAnd(Tmp2,
312 ConstantInt::get(Type::Int64Ty, 0xFF00ULL),
314 Tmp8 = BinaryOperator::CreateOr(Tmp8, Tmp7, "bswap.or1", IP);
315 Tmp6 = BinaryOperator::CreateOr(Tmp6, Tmp5, "bswap.or2", IP);
316 Tmp4 = BinaryOperator::CreateOr(Tmp4, Tmp3, "bswap.or3", IP);
317 Tmp2 = BinaryOperator::CreateOr(Tmp2, Tmp1, "bswap.or4", IP);
318 Tmp8 = BinaryOperator::CreateOr(Tmp8, Tmp6, "bswap.or5", IP);
319 Tmp4 = BinaryOperator::CreateOr(Tmp4, Tmp2, "bswap.or6", IP);
320 V = BinaryOperator::CreateOr(Tmp8, Tmp4, "bswap.i64", IP);
327 /// LowerCTPOP - Emit the code to lower ctpop of V before the specified
329 static Value *LowerCTPOP(Value *V, Instruction *IP) {
330 assert(V->getType()->isInteger() && "Can't ctpop a non-integer type!");
332 static const uint64_t MaskValues[6] = {
333 0x5555555555555555ULL, 0x3333333333333333ULL,
334 0x0F0F0F0F0F0F0F0FULL, 0x00FF00FF00FF00FFULL,
335 0x0000FFFF0000FFFFULL, 0x00000000FFFFFFFFULL
338 unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
339 unsigned WordSize = (BitSize + 63) / 64;
340 Value *Count = ConstantInt::get(V->getType(), 0);
342 for (unsigned n = 0; n < WordSize; ++n) {
343 Value *PartValue = V;
344 for (unsigned i = 1, ct = 0; i < (BitSize>64 ? 64 : BitSize);
346 Value *MaskCst = ConstantInt::get(V->getType(), MaskValues[ct]);
347 Value *LHS = BinaryOperator::CreateAnd(
348 PartValue, MaskCst, "cppop.and1", IP);
349 Value *VShift = BinaryOperator::CreateLShr(PartValue,
350 ConstantInt::get(V->getType(), i), "ctpop.sh", IP);
351 Value *RHS = BinaryOperator::CreateAnd(VShift, MaskCst, "cppop.and2", IP);
352 PartValue = BinaryOperator::CreateAdd(LHS, RHS, "ctpop.step", IP);
354 Count = BinaryOperator::CreateAdd(PartValue, Count, "ctpop.part", IP);
356 V = BinaryOperator::CreateLShr(V, ConstantInt::get(V->getType(), 64),
357 "ctpop.part.sh", IP);
365 /// LowerCTLZ - Emit the code to lower ctlz of V before the specified
367 static Value *LowerCTLZ(Value *V, Instruction *IP) {
369 unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
370 for (unsigned i = 1; i < BitSize; i <<= 1) {
371 Value *ShVal = ConstantInt::get(V->getType(), i);
372 ShVal = BinaryOperator::CreateLShr(V, ShVal, "ctlz.sh", IP);
373 V = BinaryOperator::CreateOr(V, ShVal, "ctlz.step", IP);
376 V = BinaryOperator::CreateNot(V, "", IP);
377 return LowerCTPOP(V, IP);
380 /// Convert the llvm.part.select.iX.iY intrinsic. This intrinsic takes
381 /// three integer arguments. The first argument is the Value from which the
382 /// bits will be selected. It may be of any bit width. The second and third
383 /// arguments specify a range of bits to select with the second argument
384 /// specifying the low bit and the third argument specifying the high bit. Both
385 /// must be type i32. The result is the corresponding selected bits from the
386 /// Value in the same width as the Value (first argument). If the low bit index
387 /// is higher than the high bit index then the inverse selection is done and
388 /// the bits are returned in inverse order.
389 /// @brief Lowering of llvm.part.select intrinsic.
390 static Instruction *LowerPartSelect(CallInst *CI) {
391 // Make sure we're dealing with a part select intrinsic here
392 Function *F = CI->getCalledFunction();
393 const FunctionType *FT = F->getFunctionType();
394 if (!F->isDeclaration() || !FT->getReturnType()->isInteger() ||
395 FT->getNumParams() != 3 || !FT->getParamType(0)->isInteger() ||
396 !FT->getParamType(1)->isInteger() || !FT->getParamType(2)->isInteger())
399 // Get the intrinsic implementation function by converting all the . to _
400 // in the intrinsic's function name and then reconstructing the function
402 std::string Name(F->getName());
403 for (unsigned i = 4; i < Name.length(); ++i)
406 Module* M = F->getParent();
407 F = cast<Function>(M->getOrInsertFunction(Name, FT));
408 F->setLinkage(GlobalValue::WeakLinkage);
410 // If we haven't defined the impl function yet, do so now
411 if (F->isDeclaration()) {
413 // Get the arguments to the function
414 Function::arg_iterator args = F->arg_begin();
415 Value* Val = args++; Val->setName("Val");
416 Value* Lo = args++; Lo->setName("Lo");
417 Value* Hi = args++; Hi->setName("High");
419 // We want to select a range of bits here such that [Hi, Lo] is shifted
420 // down to the low bits. However, it is quite possible that Hi is smaller
421 // than Lo in which case the bits have to be reversed.
423 // Create the blocks we will need for the two cases (forward, reverse)
424 BasicBlock* CurBB = BasicBlock::Create("entry", F);
425 BasicBlock *RevSize = BasicBlock::Create("revsize", CurBB->getParent());
426 BasicBlock *FwdSize = BasicBlock::Create("fwdsize", CurBB->getParent());
427 BasicBlock *Compute = BasicBlock::Create("compute", CurBB->getParent());
428 BasicBlock *Reverse = BasicBlock::Create("reverse", CurBB->getParent());
429 BasicBlock *RsltBlk = BasicBlock::Create("result", CurBB->getParent());
431 // Cast Hi and Lo to the size of Val so the widths are all the same
432 if (Hi->getType() != Val->getType())
433 Hi = CastInst::CreateIntegerCast(Hi, Val->getType(), false,
435 if (Lo->getType() != Val->getType())
436 Lo = CastInst::CreateIntegerCast(Lo, Val->getType(), false,
439 // Compute a few things that both cases will need, up front.
440 Constant* Zero = ConstantInt::get(Val->getType(), 0);
441 Constant* One = ConstantInt::get(Val->getType(), 1);
442 Constant* AllOnes = ConstantInt::getAllOnesValue(Val->getType());
444 // Compare the Hi and Lo bit positions. This is used to determine
445 // which case we have (forward or reverse)
446 ICmpInst *Cmp = new ICmpInst(ICmpInst::ICMP_ULT, Hi, Lo, "less",CurBB);
447 BranchInst::Create(RevSize, FwdSize, Cmp, CurBB);
449 // First, copmute the number of bits in the forward case.
450 Instruction* FBitSize =
451 BinaryOperator::CreateSub(Hi, Lo,"fbits", FwdSize);
452 BranchInst::Create(Compute, FwdSize);
454 // Second, compute the number of bits in the reverse case.
455 Instruction* RBitSize =
456 BinaryOperator::CreateSub(Lo, Hi, "rbits", RevSize);
457 BranchInst::Create(Compute, RevSize);
459 // Now, compute the bit range. Start by getting the bitsize and the shift
460 // amount (either Hi or Lo) from PHI nodes. Then we compute a mask for
461 // the number of bits we want in the range. We shift the bits down to the
462 // least significant bits, apply the mask to zero out unwanted high bits,
463 // and we have computed the "forward" result. It may still need to be
466 // Get the BitSize from one of the two subtractions
467 PHINode *BitSize = PHINode::Create(Val->getType(), "bits", Compute);
468 BitSize->reserveOperandSpace(2);
469 BitSize->addIncoming(FBitSize, FwdSize);
470 BitSize->addIncoming(RBitSize, RevSize);
472 // Get the ShiftAmount as the smaller of Hi/Lo
473 PHINode *ShiftAmt = PHINode::Create(Val->getType(), "shiftamt", Compute);
474 ShiftAmt->reserveOperandSpace(2);
475 ShiftAmt->addIncoming(Lo, FwdSize);
476 ShiftAmt->addIncoming(Hi, RevSize);
478 // Increment the bit size
479 Instruction *BitSizePlusOne =
480 BinaryOperator::CreateAdd(BitSize, One, "bits", Compute);
482 // Create a Mask to zero out the high order bits.
484 BinaryOperator::CreateShl(AllOnes, BitSizePlusOne, "mask", Compute);
485 Mask = BinaryOperator::CreateNot(Mask, "mask", Compute);
487 // Shift the bits down and apply the mask
489 BinaryOperator::CreateLShr(Val, ShiftAmt, "fres", Compute);
490 FRes = BinaryOperator::CreateAnd(FRes, Mask, "fres", Compute);
491 BranchInst::Create(Reverse, RsltBlk, Cmp, Compute);
493 // In the Reverse block we have the mask already in FRes but we must reverse
494 // it by shifting FRes bits right and putting them in RRes by shifting them
497 // First set up our loop counters
498 PHINode *Count = PHINode::Create(Val->getType(), "count", Reverse);
499 Count->reserveOperandSpace(2);
500 Count->addIncoming(BitSizePlusOne, Compute);
502 // Next, get the value that we are shifting.
503 PHINode *BitsToShift = PHINode::Create(Val->getType(), "val", Reverse);
504 BitsToShift->reserveOperandSpace(2);
505 BitsToShift->addIncoming(FRes, Compute);
507 // Finally, get the result of the last computation
508 PHINode *RRes = PHINode::Create(Val->getType(), "rres", Reverse);
509 RRes->reserveOperandSpace(2);
510 RRes->addIncoming(Zero, Compute);
512 // Decrement the counter
513 Instruction *Decr = BinaryOperator::CreateSub(Count, One, "decr", Reverse);
514 Count->addIncoming(Decr, Reverse);
516 // Compute the Bit that we want to move
518 BinaryOperator::CreateAnd(BitsToShift, One, "bit", Reverse);
520 // Compute the new value for next iteration.
521 Instruction *NewVal =
522 BinaryOperator::CreateLShr(BitsToShift, One, "rshift", Reverse);
523 BitsToShift->addIncoming(NewVal, Reverse);
525 // Shift the bit into the low bits of the result.
526 Instruction *NewRes =
527 BinaryOperator::CreateShl(RRes, One, "lshift", Reverse);
528 NewRes = BinaryOperator::CreateOr(NewRes, Bit, "addbit", Reverse);
529 RRes->addIncoming(NewRes, Reverse);
531 // Terminate loop if we've moved all the bits.
533 new ICmpInst(ICmpInst::ICMP_EQ, Decr, Zero, "cond", Reverse);
534 BranchInst::Create(RsltBlk, Reverse, Cond, Reverse);
536 // Finally, in the result block, select one of the two results with a PHI
537 // node and return the result;
539 PHINode *BitSelect = PHINode::Create(Val->getType(), "part_select", CurBB);
540 BitSelect->reserveOperandSpace(2);
541 BitSelect->addIncoming(FRes, Compute);
542 BitSelect->addIncoming(NewRes, Reverse);
543 ReturnInst::Create(BitSelect, CurBB);
546 // Return a call to the implementation function
552 return CallInst::Create(F, Args, array_endof(Args), CI->getName(), CI);
555 /// Convert the llvm.part.set.iX.iY.iZ intrinsic. This intrinsic takes
556 /// four integer arguments (iAny %Value, iAny %Replacement, i32 %Low, i32 %High)
557 /// The first two arguments can be any bit width. The result is the same width
558 /// as %Value. The operation replaces bits between %Low and %High with the value
559 /// in %Replacement. If %Replacement is not the same width, it is truncated or
560 /// zero extended as appropriate to fit the bits being replaced. If %Low is
561 /// greater than %High then the inverse set of bits are replaced.
562 /// @brief Lowering of llvm.bit.part.set intrinsic.
563 static Instruction *LowerPartSet(CallInst *CI) {
564 // Make sure we're dealing with a part select intrinsic here
565 Function *F = CI->getCalledFunction();
566 const FunctionType *FT = F->getFunctionType();
567 if (!F->isDeclaration() || !FT->getReturnType()->isInteger() ||
568 FT->getNumParams() != 4 || !FT->getParamType(0)->isInteger() ||
569 !FT->getParamType(1)->isInteger() || !FT->getParamType(2)->isInteger() ||
570 !FT->getParamType(3)->isInteger())
573 // Get the intrinsic implementation function by converting all the . to _
574 // in the intrinsic's function name and then reconstructing the function
576 std::string Name(F->getName());
577 for (unsigned i = 4; i < Name.length(); ++i)
580 Module* M = F->getParent();
581 F = cast<Function>(M->getOrInsertFunction(Name, FT));
582 F->setLinkage(GlobalValue::WeakLinkage);
584 // If we haven't defined the impl function yet, do so now
585 if (F->isDeclaration()) {
586 // Get the arguments for the function.
587 Function::arg_iterator args = F->arg_begin();
588 Value* Val = args++; Val->setName("Val");
589 Value* Rep = args++; Rep->setName("Rep");
590 Value* Lo = args++; Lo->setName("Lo");
591 Value* Hi = args++; Hi->setName("Hi");
593 // Get some types we need
594 const IntegerType* ValTy = cast<IntegerType>(Val->getType());
595 const IntegerType* RepTy = cast<IntegerType>(Rep->getType());
596 uint32_t ValBits = ValTy->getBitWidth();
597 uint32_t RepBits = RepTy->getBitWidth();
599 // Constant Definitions
600 ConstantInt* RepBitWidth = ConstantInt::get(Type::Int32Ty, RepBits);
601 ConstantInt* RepMask = ConstantInt::getAllOnesValue(RepTy);
602 ConstantInt* ValMask = ConstantInt::getAllOnesValue(ValTy);
603 ConstantInt* One = ConstantInt::get(Type::Int32Ty, 1);
604 ConstantInt* ValOne = ConstantInt::get(ValTy, 1);
605 ConstantInt* Zero = ConstantInt::get(Type::Int32Ty, 0);
606 ConstantInt* ValZero = ConstantInt::get(ValTy, 0);
608 // Basic blocks we fill in below.
609 BasicBlock* entry = BasicBlock::Create("entry", F, 0);
610 BasicBlock* large = BasicBlock::Create("large", F, 0);
611 BasicBlock* small = BasicBlock::Create("small", F, 0);
612 BasicBlock* reverse = BasicBlock::Create("reverse", F, 0);
613 BasicBlock* result = BasicBlock::Create("result", F, 0);
615 // BASIC BLOCK: entry
616 // First, get the number of bits that we're placing as an i32
617 ICmpInst* is_forward =
618 new ICmpInst(ICmpInst::ICMP_ULT, Lo, Hi, "", entry);
619 SelectInst* Hi_pn = SelectInst::Create(is_forward, Hi, Lo, "", entry);
620 SelectInst* Lo_pn = SelectInst::Create(is_forward, Lo, Hi, "", entry);
621 BinaryOperator* NumBits = BinaryOperator::CreateSub(Hi_pn, Lo_pn, "",entry);
622 NumBits = BinaryOperator::CreateAdd(NumBits, One, "", entry);
623 // Now, convert Lo and Hi to ValTy bit width
625 Lo = new ZExtInst(Lo_pn, ValTy, "", entry);
626 } else if (ValBits < 32) {
627 Lo = new TruncInst(Lo_pn, ValTy, "", entry);
629 // Determine if the replacement bits are larger than the number of bits we
630 // are replacing and deal with it.
632 new ICmpInst(ICmpInst::ICMP_ULT, NumBits, RepBitWidth, "", entry);
633 BranchInst::Create(large, small, is_large, entry);
635 // BASIC BLOCK: large
636 Instruction* MaskBits =
637 BinaryOperator::CreateSub(RepBitWidth, NumBits, "", large);
638 MaskBits = CastInst::CreateIntegerCast(MaskBits, RepMask->getType(),
640 BinaryOperator* Mask1 =
641 BinaryOperator::CreateLShr(RepMask, MaskBits, "", large);
642 BinaryOperator* Rep2 = BinaryOperator::CreateAnd(Mask1, Rep, "", large);
643 BranchInst::Create(small, large);
645 // BASIC BLOCK: small
646 PHINode* Rep3 = PHINode::Create(RepTy, "", small);
647 Rep3->reserveOperandSpace(2);
648 Rep3->addIncoming(Rep2, large);
649 Rep3->addIncoming(Rep, entry);
651 if (ValBits > RepBits)
652 Rep4 = new ZExtInst(Rep3, ValTy, "", small);
653 else if (ValBits < RepBits)
654 Rep4 = new TruncInst(Rep3, ValTy, "", small);
655 BranchInst::Create(result, reverse, is_forward, small);
657 // BASIC BLOCK: reverse (reverses the bits of the replacement)
658 // Set up our loop counter as a PHI so we can decrement on each iteration.
659 // We will loop for the number of bits in the replacement value.
660 PHINode *Count = PHINode::Create(Type::Int32Ty, "count", reverse);
661 Count->reserveOperandSpace(2);
662 Count->addIncoming(NumBits, small);
664 // Get the value that we are shifting bits out of as a PHI because
665 // we'll change this with each iteration.
666 PHINode *BitsToShift = PHINode::Create(Val->getType(), "val", reverse);
667 BitsToShift->reserveOperandSpace(2);
668 BitsToShift->addIncoming(Rep4, small);
670 // Get the result of the last computation or zero on first iteration
671 PHINode *RRes = PHINode::Create(Val->getType(), "rres", reverse);
672 RRes->reserveOperandSpace(2);
673 RRes->addIncoming(ValZero, small);
675 // Decrement the loop counter by one
676 Instruction *Decr = BinaryOperator::CreateSub(Count, One, "", reverse);
677 Count->addIncoming(Decr, reverse);
679 // Get the bit that we want to move into the result
680 Value *Bit = BinaryOperator::CreateAnd(BitsToShift, ValOne, "", reverse);
682 // Compute the new value of the bits to shift for the next iteration.
683 Value *NewVal = BinaryOperator::CreateLShr(BitsToShift, ValOne,"", reverse);
684 BitsToShift->addIncoming(NewVal, reverse);
686 // Shift the bit we extracted into the low bit of the result.
687 Instruction *NewRes = BinaryOperator::CreateShl(RRes, ValOne, "", reverse);
688 NewRes = BinaryOperator::CreateOr(NewRes, Bit, "", reverse);
689 RRes->addIncoming(NewRes, reverse);
691 // Terminate loop if we've moved all the bits.
692 ICmpInst *Cond = new ICmpInst(ICmpInst::ICMP_EQ, Decr, Zero, "", reverse);
693 BranchInst::Create(result, reverse, Cond, reverse);
695 // BASIC BLOCK: result
696 PHINode *Rplcmnt = PHINode::Create(Val->getType(), "", result);
697 Rplcmnt->reserveOperandSpace(2);
698 Rplcmnt->addIncoming(NewRes, reverse);
699 Rplcmnt->addIncoming(Rep4, small);
700 Value* t0 = CastInst::CreateIntegerCast(NumBits,ValTy,false,"",result);
701 Value* t1 = BinaryOperator::CreateShl(ValMask, Lo, "", result);
702 Value* t2 = BinaryOperator::CreateNot(t1, "", result);
703 Value* t3 = BinaryOperator::CreateShl(t1, t0, "", result);
704 Value* t4 = BinaryOperator::CreateOr(t2, t3, "", result);
705 Value* t5 = BinaryOperator::CreateAnd(t4, Val, "", result);
706 Value* t6 = BinaryOperator::CreateShl(Rplcmnt, Lo, "", result);
707 Value* Rslt = BinaryOperator::CreateOr(t5, t6, "part_set", result);
708 ReturnInst::Create(Rslt, result);
711 // Return a call to the implementation function
718 return CallInst::Create(F, Args, array_endof(Args), CI->getName(), CI);
722 void IntrinsicLowering::LowerIntrinsicCall(CallInst *CI) {
723 Function *Callee = CI->getCalledFunction();
724 assert(Callee && "Cannot lower an indirect call!");
726 switch (Callee->getIntrinsicID()) {
727 case Intrinsic::not_intrinsic:
728 cerr << "Cannot lower a call to a non-intrinsic function '"
729 << Callee->getName() << "'!\n";
732 cerr << "Error: Code generator does not support intrinsic function '"
733 << Callee->getName() << "'!\n";
736 // The setjmp/longjmp intrinsics should only exist in the code if it was
737 // never optimized (ie, right out of the CFE), or if it has been hacked on
738 // by the lowerinvoke pass. In both cases, the right thing to do is to
739 // convert the call to an explicit setjmp or longjmp call.
740 case Intrinsic::setjmp: {
741 static Constant *SetjmpFCache = 0;
742 Value *V = ReplaceCallWith("setjmp", CI, CI->op_begin()+1, CI->op_end(),
743 Type::Int32Ty, SetjmpFCache);
744 if (CI->getType() != Type::VoidTy)
745 CI->replaceAllUsesWith(V);
748 case Intrinsic::sigsetjmp:
749 if (CI->getType() != Type::VoidTy)
750 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
753 case Intrinsic::longjmp: {
754 static Constant *LongjmpFCache = 0;
755 ReplaceCallWith("longjmp", CI, CI->op_begin()+1, CI->op_end(),
756 Type::VoidTy, LongjmpFCache);
760 case Intrinsic::siglongjmp: {
761 // Insert the call to abort
762 static Constant *AbortFCache = 0;
763 ReplaceCallWith("abort", CI, CI->op_end(), CI->op_end(),
764 Type::VoidTy, AbortFCache);
767 case Intrinsic::ctpop:
768 CI->replaceAllUsesWith(LowerCTPOP(CI->getOperand(1), CI));
771 case Intrinsic::bswap:
772 CI->replaceAllUsesWith(LowerBSWAP(CI->getOperand(1), CI));
775 case Intrinsic::ctlz:
776 CI->replaceAllUsesWith(LowerCTLZ(CI->getOperand(1), CI));
779 case Intrinsic::cttz: {
780 // cttz(x) -> ctpop(~X & (X-1))
781 Value *Src = CI->getOperand(1);
782 Value *NotSrc = BinaryOperator::CreateNot(Src, Src->getName()+".not", CI);
783 Value *SrcM1 = ConstantInt::get(Src->getType(), 1);
784 SrcM1 = BinaryOperator::CreateSub(Src, SrcM1, "", CI);
785 Src = LowerCTPOP(BinaryOperator::CreateAnd(NotSrc, SrcM1, "", CI), CI);
786 CI->replaceAllUsesWith(Src);
790 case Intrinsic::part_select:
791 CI->replaceAllUsesWith(LowerPartSelect(CI));
794 case Intrinsic::part_set:
795 CI->replaceAllUsesWith(LowerPartSet(CI));
798 case Intrinsic::stacksave:
799 case Intrinsic::stackrestore: {
800 static bool Warned = false;
802 cerr << "WARNING: this target does not support the llvm.stack"
803 << (Callee->getIntrinsicID() == Intrinsic::stacksave ?
804 "save" : "restore") << " intrinsic.\n";
806 if (Callee->getIntrinsicID() == Intrinsic::stacksave)
807 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
811 case Intrinsic::returnaddress:
812 case Intrinsic::frameaddress:
813 cerr << "WARNING: this target does not support the llvm."
814 << (Callee->getIntrinsicID() == Intrinsic::returnaddress ?
815 "return" : "frame") << "address intrinsic.\n";
816 CI->replaceAllUsesWith(ConstantPointerNull::get(
817 cast<PointerType>(CI->getType())));
820 case Intrinsic::prefetch:
821 break; // Simply strip out prefetches on unsupported architectures
823 case Intrinsic::pcmarker:
824 break; // Simply strip out pcmarker on unsupported architectures
825 case Intrinsic::readcyclecounter: {
826 cerr << "WARNING: this target does not support the llvm.readcyclecoun"
827 << "ter intrinsic. It is being lowered to a constant 0\n";
828 CI->replaceAllUsesWith(ConstantInt::get(Type::Int64Ty, 0));
832 case Intrinsic::dbg_stoppoint:
833 case Intrinsic::dbg_region_start:
834 case Intrinsic::dbg_region_end:
835 case Intrinsic::dbg_func_start:
836 case Intrinsic::dbg_declare:
837 break; // Simply strip out debugging intrinsics
839 case Intrinsic::eh_exception:
840 case Intrinsic::eh_selector_i32:
841 case Intrinsic::eh_selector_i64:
842 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
845 case Intrinsic::eh_typeid_for_i32:
846 case Intrinsic::eh_typeid_for_i64:
847 // Return something different to eh_selector.
848 CI->replaceAllUsesWith(ConstantInt::get(CI->getType(), 1));
851 case Intrinsic::var_annotation:
852 break; // Strip out annotate intrinsic
854 case Intrinsic::memcpy_i32:
855 case Intrinsic::memcpy_i64: {
856 static Constant *MemcpyFCache = 0;
857 Value *Size = CI->getOperand(3);
858 const Type *IntPtr = TD.getIntPtrType();
859 if (Size->getType()->getPrimitiveSizeInBits() <
860 IntPtr->getPrimitiveSizeInBits())
861 Size = new ZExtInst(Size, IntPtr, "", CI);
862 else if (Size->getType()->getPrimitiveSizeInBits() >
863 IntPtr->getPrimitiveSizeInBits())
864 Size = new TruncInst(Size, IntPtr, "", CI);
866 Ops[0] = CI->getOperand(1);
867 Ops[1] = CI->getOperand(2);
869 ReplaceCallWith("memcpy", CI, Ops, Ops+3, CI->getOperand(1)->getType(),
873 case Intrinsic::memmove_i32:
874 case Intrinsic::memmove_i64: {
875 static Constant *MemmoveFCache = 0;
876 Value *Size = CI->getOperand(3);
877 const Type *IntPtr = TD.getIntPtrType();
878 if (Size->getType()->getPrimitiveSizeInBits() <
879 IntPtr->getPrimitiveSizeInBits())
880 Size = new ZExtInst(Size, IntPtr, "", CI);
881 else if (Size->getType()->getPrimitiveSizeInBits() >
882 IntPtr->getPrimitiveSizeInBits())
883 Size = new TruncInst(Size, IntPtr, "", CI);
885 Ops[0] = CI->getOperand(1);
886 Ops[1] = CI->getOperand(2);
888 ReplaceCallWith("memmove", CI, Ops, Ops+3, CI->getOperand(1)->getType(),
892 case Intrinsic::memset_i32:
893 case Intrinsic::memset_i64: {
894 static Constant *MemsetFCache = 0;
895 Value *Size = CI->getOperand(3);
896 const Type *IntPtr = TD.getIntPtrType();
897 if (Size->getType()->getPrimitiveSizeInBits() <
898 IntPtr->getPrimitiveSizeInBits())
899 Size = new ZExtInst(Size, IntPtr, "", CI);
900 else if (Size->getType()->getPrimitiveSizeInBits() >
901 IntPtr->getPrimitiveSizeInBits())
902 Size = new TruncInst(Size, IntPtr, "", CI);
904 Ops[0] = CI->getOperand(1);
905 // Extend the amount to i32.
906 Ops[1] = new ZExtInst(CI->getOperand(2), Type::Int32Ty, "", CI);
908 ReplaceCallWith("memset", CI, Ops, Ops+3, CI->getOperand(1)->getType(),
912 case Intrinsic::sqrt: {
913 static Constant *sqrtfFCache = 0;
914 static Constant *sqrtFCache = 0;
915 static Constant *sqrtLDCache = 0;
916 switch (CI->getOperand(1)->getType()->getTypeID()) {
917 default: assert(0 && "Invalid type in sqrt"); abort();
918 case Type::FloatTyID:
919 ReplaceCallWith("sqrtf", CI, CI->op_begin()+1, CI->op_end(),
920 Type::FloatTy, sqrtfFCache);
922 case Type::DoubleTyID:
923 ReplaceCallWith("sqrt", CI, CI->op_begin()+1, CI->op_end(),
924 Type::DoubleTy, sqrtFCache);
926 case Type::X86_FP80TyID:
927 case Type::FP128TyID:
928 case Type::PPC_FP128TyID:
929 ReplaceCallWith("sqrtl", CI, CI->op_begin()+1, CI->op_end(),
930 CI->getOperand(1)->getType(), sqrtLDCache);
935 case Intrinsic::log: {
936 static Constant *logfFCache = 0;
937 static Constant *logFCache = 0;
938 static Constant *logLDCache = 0;
939 switch (CI->getOperand(1)->getType()->getTypeID()) {
940 default: assert(0 && "Invalid type in log"); abort();
941 case Type::FloatTyID:
942 ReplaceCallWith("logf", CI, CI->op_begin()+1, CI->op_end(),
943 Type::FloatTy, logfFCache);
945 case Type::DoubleTyID:
946 ReplaceCallWith("log", CI, CI->op_begin()+1, CI->op_end(),
947 Type::DoubleTy, logFCache);
949 case Type::X86_FP80TyID:
950 case Type::FP128TyID:
951 case Type::PPC_FP128TyID:
952 ReplaceCallWith("logl", CI, CI->op_begin()+1, CI->op_end(),
953 CI->getOperand(1)->getType(), logLDCache);
958 case Intrinsic::log2: {
959 static Constant *log2fFCache = 0;
960 static Constant *log2FCache = 0;
961 static Constant *log2LDCache = 0;
962 switch (CI->getOperand(1)->getType()->getTypeID()) {
963 default: assert(0 && "Invalid type in log2"); abort();
964 case Type::FloatTyID:
965 ReplaceCallWith("log2f", CI, CI->op_begin()+1, CI->op_end(),
966 Type::FloatTy, log2fFCache);
968 case Type::DoubleTyID:
969 ReplaceCallWith("log2", CI, CI->op_begin()+1, CI->op_end(),
970 Type::DoubleTy, log2FCache);
972 case Type::X86_FP80TyID:
973 case Type::FP128TyID:
974 case Type::PPC_FP128TyID:
975 ReplaceCallWith("log2l", CI, CI->op_begin()+1, CI->op_end(),
976 CI->getOperand(1)->getType(), log2LDCache);
981 case Intrinsic::log10: {
982 static Constant *log10fFCache = 0;
983 static Constant *log10FCache = 0;
984 static Constant *log10LDCache = 0;
985 switch (CI->getOperand(1)->getType()->getTypeID()) {
986 default: assert(0 && "Invalid type in log10"); abort();
987 case Type::FloatTyID:
988 ReplaceCallWith("log10f", CI, CI->op_begin()+1, CI->op_end(),
989 Type::FloatTy, log10fFCache);
991 case Type::DoubleTyID:
992 ReplaceCallWith("log10", CI, CI->op_begin()+1, CI->op_end(),
993 Type::DoubleTy, log10FCache);
995 case Type::X86_FP80TyID:
996 case Type::FP128TyID:
997 case Type::PPC_FP128TyID:
998 ReplaceCallWith("log10l", CI, CI->op_begin()+1, CI->op_end(),
999 CI->getOperand(1)->getType(), log10LDCache);
1004 case Intrinsic::exp: {
1005 static Constant *expfFCache = 0;
1006 static Constant *expFCache = 0;
1007 static Constant *expLDCache = 0;
1008 switch (CI->getOperand(1)->getType()->getTypeID()) {
1009 default: assert(0 && "Invalid type in exp"); abort();
1010 case Type::FloatTyID:
1011 ReplaceCallWith("expf", CI, CI->op_begin()+1, CI->op_end(),
1012 Type::FloatTy, expfFCache);
1014 case Type::DoubleTyID:
1015 ReplaceCallWith("exp", CI, CI->op_begin()+1, CI->op_end(),
1016 Type::DoubleTy, expFCache);
1018 case Type::X86_FP80TyID:
1019 case Type::FP128TyID:
1020 case Type::PPC_FP128TyID:
1021 ReplaceCallWith("expl", CI, CI->op_begin()+1, CI->op_end(),
1022 CI->getOperand(1)->getType(), expLDCache);
1027 case Intrinsic::exp2: {
1028 static Constant *exp2fFCache = 0;
1029 static Constant *exp2FCache = 0;
1030 static Constant *exp2LDCache = 0;
1031 switch (CI->getOperand(1)->getType()->getTypeID()) {
1032 default: assert(0 && "Invalid type in exp2"); abort();
1033 case Type::FloatTyID:
1034 ReplaceCallWith("exp2f", CI, CI->op_begin()+1, CI->op_end(),
1035 Type::FloatTy, exp2fFCache);
1037 case Type::DoubleTyID:
1038 ReplaceCallWith("exp2", CI, CI->op_begin()+1, CI->op_end(),
1039 Type::DoubleTy, exp2FCache);
1041 case Type::X86_FP80TyID:
1042 case Type::FP128TyID:
1043 case Type::PPC_FP128TyID:
1044 ReplaceCallWith("exp2l", CI, CI->op_begin()+1, CI->op_end(),
1045 CI->getOperand(1)->getType(), exp2LDCache);
1050 case Intrinsic::pow: {
1051 static Constant *powfFCache = 0;
1052 static Constant *powFCache = 0;
1053 static Constant *powLDCache = 0;
1054 switch (CI->getOperand(1)->getType()->getTypeID()) {
1055 default: assert(0 && "Invalid type in pow"); abort();
1056 case Type::FloatTyID:
1057 ReplaceCallWith("powf", CI, CI->op_begin()+1, CI->op_end(),
1058 Type::FloatTy, powfFCache);
1060 case Type::DoubleTyID:
1061 ReplaceCallWith("pow", CI, CI->op_begin()+1, CI->op_end(),
1062 Type::DoubleTy, powFCache);
1064 case Type::X86_FP80TyID:
1065 case Type::FP128TyID:
1066 case Type::PPC_FP128TyID:
1067 ReplaceCallWith("powl", CI, CI->op_begin()+1, CI->op_end(),
1068 CI->getOperand(1)->getType(), powLDCache);
1073 case Intrinsic::flt_rounds:
1074 // Lower to "round to the nearest"
1075 if (CI->getType() != Type::VoidTy)
1076 CI->replaceAllUsesWith(ConstantInt::get(CI->getType(), 1));
1080 assert(CI->use_empty() &&
1081 "Lowering should have eliminated any uses of the intrinsic call!");
1082 CI->eraseFromParent();