1 //===-- IntrinsicLowering.cpp - Intrinsic Lowering default implementation -===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source 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"
25 template <class ArgIt>
26 static void EnsureFunctionExists(Module &M, const char *Name,
27 ArgIt ArgBegin, ArgIt ArgEnd,
29 // Insert a correctly-typed definition now.
30 std::vector<const Type *> ParamTys;
31 for (ArgIt I = ArgBegin; I != ArgEnd; ++I)
32 ParamTys.push_back(I->getType());
33 M.getOrInsertFunction(Name, FunctionType::get(RetTy, ParamTys, false));
36 /// ReplaceCallWith - This function is used when we want to lower an intrinsic
37 /// call to a call of an external function. This handles hard cases such as
38 /// when there was already a prototype for the external function, and if that
39 /// prototype doesn't match the arguments we expect to pass in.
40 template <class ArgIt>
41 static CallInst *ReplaceCallWith(const char *NewFn, CallInst *CI,
42 ArgIt ArgBegin, ArgIt ArgEnd,
43 const Type *RetTy, Constant *&FCache) {
45 // If we haven't already looked up this function, check to see if the
46 // program already contains a function with this name.
47 Module *M = CI->getParent()->getParent()->getParent();
48 // Get or insert the definition now.
49 std::vector<const Type *> ParamTys;
50 for (ArgIt I = ArgBegin; I != ArgEnd; ++I)
51 ParamTys.push_back((*I)->getType());
52 FCache = M->getOrInsertFunction(NewFn,
53 FunctionType::get(RetTy, ParamTys, false));
56 SmallVector<Value*, 8> Operands(ArgBegin, ArgEnd);
57 CallInst *NewCI = new CallInst(FCache, &Operands[0], Operands.size(),
60 CI->replaceAllUsesWith(NewCI);
64 void IntrinsicLowering::AddPrototypes(Module &M) {
65 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
66 if (I->isDeclaration() && !I->use_empty())
67 switch (I->getIntrinsicID()) {
69 case Intrinsic::setjmp:
70 EnsureFunctionExists(M, "setjmp", I->arg_begin(), I->arg_end(),
73 case Intrinsic::longjmp:
74 EnsureFunctionExists(M, "longjmp", I->arg_begin(), I->arg_end(),
77 case Intrinsic::siglongjmp:
78 EnsureFunctionExists(M, "abort", I->arg_end(), I->arg_end(),
81 case Intrinsic::memcpy_i32:
82 case Intrinsic::memcpy_i64:
83 M.getOrInsertFunction("memcpy", PointerType::get(Type::Int8Ty),
84 PointerType::get(Type::Int8Ty),
85 PointerType::get(Type::Int8Ty),
86 TD.getIntPtrType(), (Type *)0);
88 case Intrinsic::memmove_i32:
89 case Intrinsic::memmove_i64:
90 M.getOrInsertFunction("memmove", PointerType::get(Type::Int8Ty),
91 PointerType::get(Type::Int8Ty),
92 PointerType::get(Type::Int8Ty),
93 TD.getIntPtrType(), (Type *)0);
95 case Intrinsic::memset_i32:
96 case Intrinsic::memset_i64:
97 M.getOrInsertFunction("memset", PointerType::get(Type::Int8Ty),
98 PointerType::get(Type::Int8Ty), Type::Int32Ty,
99 TD.getIntPtrType(), (Type *)0);
101 case Intrinsic::sqrt_f32:
102 case Intrinsic::sqrt_f64:
103 if(I->arg_begin()->getType() == Type::FloatTy)
104 EnsureFunctionExists(M, "sqrtf", I->arg_begin(), I->arg_end(),
107 EnsureFunctionExists(M, "sqrt", I->arg_begin(), I->arg_end(),
113 /// LowerBSWAP - Emit the code to lower bswap of V before the specified
115 static Value *LowerBSWAP(Value *V, Instruction *IP) {
116 assert(V->getType()->isInteger() && "Can't bswap a non-integer type!");
118 unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
121 default: assert(0 && "Unhandled type size of value to byteswap!");
123 Value *Tmp1 = BinaryOperator::createShl(V,
124 ConstantInt::get(V->getType(),8),"bswap.2",IP);
125 Value *Tmp2 = BinaryOperator::createLShr(V,
126 ConstantInt::get(V->getType(),8),"bswap.1",IP);
127 V = BinaryOperator::createOr(Tmp1, Tmp2, "bswap.i16", IP);
131 Value *Tmp4 = BinaryOperator::createShl(V,
132 ConstantInt::get(V->getType(),24),"bswap.4", IP);
133 Value *Tmp3 = BinaryOperator::createShl(V,
134 ConstantInt::get(V->getType(),8),"bswap.3",IP);
135 Value *Tmp2 = BinaryOperator::createLShr(V,
136 ConstantInt::get(V->getType(),8),"bswap.2",IP);
137 Value *Tmp1 = BinaryOperator::createLShr(V,
138 ConstantInt::get(V->getType(),24),"bswap.1", IP);
139 Tmp3 = BinaryOperator::createAnd(Tmp3,
140 ConstantInt::get(Type::Int32Ty, 0xFF0000),
142 Tmp2 = BinaryOperator::createAnd(Tmp2,
143 ConstantInt::get(Type::Int32Ty, 0xFF00),
145 Tmp4 = BinaryOperator::createOr(Tmp4, Tmp3, "bswap.or1", IP);
146 Tmp2 = BinaryOperator::createOr(Tmp2, Tmp1, "bswap.or2", IP);
147 V = BinaryOperator::createOr(Tmp4, Tmp3, "bswap.i32", IP);
151 Value *Tmp8 = BinaryOperator::createShl(V,
152 ConstantInt::get(V->getType(),56),"bswap.8", IP);
153 Value *Tmp7 = BinaryOperator::createShl(V,
154 ConstantInt::get(V->getType(),40),"bswap.7", IP);
155 Value *Tmp6 = BinaryOperator::createShl(V,
156 ConstantInt::get(V->getType(),24),"bswap.6", IP);
157 Value *Tmp5 = BinaryOperator::createShl(V,
158 ConstantInt::get(V->getType(),8),"bswap.5", IP);
159 Value* Tmp4 = BinaryOperator::createLShr(V,
160 ConstantInt::get(V->getType(),8),"bswap.4", IP);
161 Value* Tmp3 = BinaryOperator::createLShr(V,
162 ConstantInt::get(V->getType(),24),"bswap.3", IP);
163 Value* Tmp2 = BinaryOperator::createLShr(V,
164 ConstantInt::get(V->getType(),40),"bswap.2", IP);
165 Value* Tmp1 = BinaryOperator::createLShr(V,
166 ConstantInt::get(V->getType(),56),"bswap.1", IP);
167 Tmp7 = BinaryOperator::createAnd(Tmp7,
168 ConstantInt::get(Type::Int64Ty,
169 0xFF000000000000ULL),
171 Tmp6 = BinaryOperator::createAnd(Tmp6,
172 ConstantInt::get(Type::Int64Ty, 0xFF0000000000ULL),
174 Tmp5 = BinaryOperator::createAnd(Tmp5,
175 ConstantInt::get(Type::Int64Ty, 0xFF00000000ULL),
177 Tmp4 = BinaryOperator::createAnd(Tmp4,
178 ConstantInt::get(Type::Int64Ty, 0xFF000000ULL),
180 Tmp3 = BinaryOperator::createAnd(Tmp3,
181 ConstantInt::get(Type::Int64Ty, 0xFF0000ULL),
183 Tmp2 = BinaryOperator::createAnd(Tmp2,
184 ConstantInt::get(Type::Int64Ty, 0xFF00ULL),
186 Tmp8 = BinaryOperator::createOr(Tmp8, Tmp7, "bswap.or1", IP);
187 Tmp6 = BinaryOperator::createOr(Tmp6, Tmp5, "bswap.or2", IP);
188 Tmp4 = BinaryOperator::createOr(Tmp4, Tmp3, "bswap.or3", IP);
189 Tmp2 = BinaryOperator::createOr(Tmp2, Tmp1, "bswap.or4", IP);
190 Tmp8 = BinaryOperator::createOr(Tmp8, Tmp6, "bswap.or5", IP);
191 Tmp4 = BinaryOperator::createOr(Tmp4, Tmp2, "bswap.or6", IP);
192 V = BinaryOperator::createOr(Tmp8, Tmp4, "bswap.i64", IP);
199 /// LowerCTPOP - Emit the code to lower ctpop of V before the specified
201 static Value *LowerCTPOP(Value *V, Instruction *IP) {
202 assert(V->getType()->isInteger() && "Can't ctpop a non-integer type!");
204 static const uint64_t MaskValues[6] = {
205 0x5555555555555555ULL, 0x3333333333333333ULL,
206 0x0F0F0F0F0F0F0F0FULL, 0x00FF00FF00FF00FFULL,
207 0x0000FFFF0000FFFFULL, 0x00000000FFFFFFFFULL
210 unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
212 for (unsigned i = 1, ct = 0; i != BitSize; i <<= 1, ++ct) {
213 Value *MaskCst = ConstantInt::get(V->getType(), MaskValues[ct]);
214 Value *LHS = BinaryOperator::createAnd(V, MaskCst, "cppop.and1", IP);
215 Value *VShift = BinaryOperator::createLShr(V,
216 ConstantInt::get(V->getType(), i), "ctpop.sh", IP);
217 Value *RHS = BinaryOperator::createAnd(VShift, MaskCst, "cppop.and2", IP);
218 V = BinaryOperator::createAdd(LHS, RHS, "ctpop.step", IP);
221 return CastInst::createIntegerCast(V, Type::Int32Ty, false, "ctpop", IP);
224 /// LowerCTLZ - Emit the code to lower ctlz of V before the specified
226 static Value *LowerCTLZ(Value *V, Instruction *IP) {
228 unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
229 for (unsigned i = 1; i != BitSize; i <<= 1) {
230 Value *ShVal = ConstantInt::get(V->getType(), i);
231 ShVal = BinaryOperator::createLShr(V, ShVal, "ctlz.sh", IP);
232 V = BinaryOperator::createOr(V, ShVal, "ctlz.step", IP);
235 V = BinaryOperator::createNot(V, "", IP);
236 return LowerCTPOP(V, IP);
239 /// Convert the llvm.part.select.iX.iY intrinsic. This intrinsic takes
240 /// three integer arguments. The first argument is the Value from which the
241 /// bits will be selected. It may be of any bit width. The second and third
242 /// arguments specify a range of bits to select with the second argument
243 /// specifying the low bit and the third argument specifying the high bit. Both
244 /// must be type i32. The result is the corresponding selected bits from the
245 /// Value in the same width as the Value (first argument). If the low bit index
246 /// is higher than the high bit index then the inverse selection is done and
247 /// the bits are returned in inverse order.
248 /// @brief Lowering of llvm.part.select intrinsic.
249 static Instruction *LowerPartSelect(CallInst *CI) {
250 // Make sure we're dealing with a part select intrinsic here
251 Function *F = CI->getCalledFunction();
252 const FunctionType *FT = F->getFunctionType();
253 if (!F->isDeclaration() || !FT->getReturnType()->isInteger() ||
254 FT->getNumParams() != 3 || !FT->getParamType(0)->isInteger() ||
255 !FT->getParamType(1)->isInteger() || !FT->getParamType(2)->isInteger())
258 // Get the intrinsic implementation function by converting all the . to _
259 // in the intrinsic's function name and then reconstructing the function
261 std::string Name(F->getName());
262 for (unsigned i = 4; i < Name.length(); ++i)
265 Module* M = F->getParent();
266 F = cast<Function>(M->getOrInsertFunction(Name, FT));
267 F->setLinkage(GlobalValue::WeakLinkage);
269 // If we haven't defined the impl function yet, do so now
270 if (F->isDeclaration()) {
272 // Get the arguments to the function
273 Function::arg_iterator args = F->arg_begin();
274 Value* Val = args++; Val->setName("Val");
275 Value* Lo = args++; Lo->setName("Lo");
276 Value* Hi = args++; Hi->setName("High");
278 // We want to select a range of bits here such that [Hi, Lo] is shifted
279 // down to the low bits. However, it is quite possible that Hi is smaller
280 // than Lo in which case the bits have to be reversed.
282 // Create the blocks we will need for the two cases (forward, reverse)
283 BasicBlock* CurBB = new BasicBlock("entry", F);
284 BasicBlock *RevSize = new BasicBlock("revsize", CurBB->getParent());
285 BasicBlock *FwdSize = new BasicBlock("fwdsize", CurBB->getParent());
286 BasicBlock *Compute = new BasicBlock("compute", CurBB->getParent());
287 BasicBlock *Reverse = new BasicBlock("reverse", CurBB->getParent());
288 BasicBlock *RsltBlk = new BasicBlock("result", CurBB->getParent());
290 // Cast Hi and Lo to the size of Val so the widths are all the same
291 if (Hi->getType() != Val->getType())
292 Hi = CastInst::createIntegerCast(Hi, Val->getType(), false,
294 if (Lo->getType() != Val->getType())
295 Lo = CastInst::createIntegerCast(Lo, Val->getType(), false,
298 // Compute a few things that both cases will need, up front.
299 Constant* Zero = ConstantInt::get(Val->getType(), 0);
300 Constant* One = ConstantInt::get(Val->getType(), 1);
301 Constant* AllOnes = ConstantInt::getAllOnesValue(Val->getType());
303 // Compare the Hi and Lo bit positions. This is used to determine
304 // which case we have (forward or reverse)
305 ICmpInst *Cmp = new ICmpInst(ICmpInst::ICMP_ULT, Hi, Lo, "less",CurBB);
306 new BranchInst(RevSize, FwdSize, Cmp, CurBB);
308 // First, copmute the number of bits in the forward case.
309 Instruction* FBitSize =
310 BinaryOperator::createSub(Hi, Lo,"fbits", FwdSize);
311 new BranchInst(Compute, FwdSize);
313 // Second, compute the number of bits in the reverse case.
314 Instruction* RBitSize =
315 BinaryOperator::createSub(Lo, Hi, "rbits", RevSize);
316 new BranchInst(Compute, RevSize);
318 // Now, compute the bit range. Start by getting the bitsize and the shift
319 // amount (either Hi or Lo) from PHI nodes. Then we compute a mask for
320 // the number of bits we want in the range. We shift the bits down to the
321 // least significant bits, apply the mask to zero out unwanted high bits,
322 // and we have computed the "forward" result. It may still need to be
325 // Get the BitSize from one of the two subtractions
326 PHINode *BitSize = new PHINode(Val->getType(), "bits", Compute);
327 BitSize->reserveOperandSpace(2);
328 BitSize->addIncoming(FBitSize, FwdSize);
329 BitSize->addIncoming(RBitSize, RevSize);
331 // Get the ShiftAmount as the smaller of Hi/Lo
332 PHINode *ShiftAmt = new PHINode(Val->getType(), "shiftamt", Compute);
333 ShiftAmt->reserveOperandSpace(2);
334 ShiftAmt->addIncoming(Lo, FwdSize);
335 ShiftAmt->addIncoming(Hi, RevSize);
337 // Increment the bit size
338 Instruction *BitSizePlusOne =
339 BinaryOperator::createAdd(BitSize, One, "bits", Compute);
341 // Create a Mask to zero out the high order bits.
343 BinaryOperator::createShl(AllOnes, BitSizePlusOne, "mask", Compute);
344 Mask = BinaryOperator::createNot(Mask, "mask", Compute);
346 // Shift the bits down and apply the mask
348 BinaryOperator::createLShr(Val, ShiftAmt, "fres", Compute);
349 FRes = BinaryOperator::createAnd(FRes, Mask, "fres", Compute);
350 new BranchInst(Reverse, RsltBlk, Cmp, Compute);
352 // In the Reverse block we have the mask already in FRes but we must reverse
353 // it by shifting FRes bits right and putting them in RRes by shifting them
356 // First set up our loop counters
357 PHINode *Count = new PHINode(Val->getType(), "count", Reverse);
358 Count->reserveOperandSpace(2);
359 Count->addIncoming(BitSizePlusOne, Compute);
361 // Next, get the value that we are shifting.
362 PHINode *BitsToShift = new PHINode(Val->getType(), "val", Reverse);
363 BitsToShift->reserveOperandSpace(2);
364 BitsToShift->addIncoming(FRes, Compute);
366 // Finally, get the result of the last computation
367 PHINode *RRes = new PHINode(Val->getType(), "rres", Reverse);
368 RRes->reserveOperandSpace(2);
369 RRes->addIncoming(Zero, Compute);
371 // Decrement the counter
372 Instruction *Decr = BinaryOperator::createSub(Count, One, "decr", Reverse);
373 Count->addIncoming(Decr, Reverse);
375 // Compute the Bit that we want to move
377 BinaryOperator::createAnd(BitsToShift, One, "bit", Reverse);
379 // Compute the new value for next iteration.
380 Instruction *NewVal =
381 BinaryOperator::createLShr(BitsToShift, One, "rshift", Reverse);
382 BitsToShift->addIncoming(NewVal, Reverse);
384 // Shift the bit into the low bits of the result.
385 Instruction *NewRes =
386 BinaryOperator::createShl(RRes, One, "lshift", Reverse);
387 NewRes = BinaryOperator::createOr(NewRes, Bit, "addbit", Reverse);
388 RRes->addIncoming(NewRes, Reverse);
390 // Terminate loop if we've moved all the bits.
392 new ICmpInst(ICmpInst::ICMP_EQ, Decr, Zero, "cond", Reverse);
393 new BranchInst(RsltBlk, Reverse, Cond, Reverse);
395 // Finally, in the result block, select one of the two results with a PHI
396 // node and return the result;
398 PHINode *BitSelect = new PHINode(Val->getType(), "part_select", CurBB);
399 BitSelect->reserveOperandSpace(2);
400 BitSelect->addIncoming(FRes, Compute);
401 BitSelect->addIncoming(NewRes, Reverse);
402 new ReturnInst(BitSelect, CurBB);
405 // Return a call to the implementation function
411 return new CallInst(F, Args, sizeof(Args)/sizeof(Args[0]), CI->getName(), CI);
414 /// Convert the llvm.part.set.iX.iY.iZ intrinsic. This intrinsic takes
415 /// four integer arguments (iAny %Value, iAny %Replacement, i32 %Low, i32 %High)
416 /// The first two arguments can be any bit width. The result is the same width
417 /// as %Value. The operation replaces bits between %Low and %High with the value
418 /// in %Replacement. If %Replacement is not the same width, it is truncated or
419 /// zero extended as appropriate to fit the bits being replaced. If %Low is
420 /// greater than %High then the inverse set of bits are replaced.
421 /// @brief Lowering of llvm.bit.part.set intrinsic.
422 static Instruction *LowerPartSet(CallInst *CI) {
423 // Make sure we're dealing with a part select intrinsic here
424 Function *F = CI->getCalledFunction();
425 const FunctionType *FT = F->getFunctionType();
426 if (!F->isDeclaration() || !FT->getReturnType()->isInteger() ||
427 FT->getNumParams() != 4 || !FT->getParamType(0)->isInteger() ||
428 !FT->getParamType(1)->isInteger() || !FT->getParamType(2)->isInteger() ||
429 !FT->getParamType(3)->isInteger())
432 // Get the intrinsic implementation function by converting all the . to _
433 // in the intrinsic's function name and then reconstructing the function
435 std::string Name(F->getName());
436 for (unsigned i = 4; i < Name.length(); ++i)
439 Module* M = F->getParent();
440 F = cast<Function>(M->getOrInsertFunction(Name, FT));
441 F->setLinkage(GlobalValue::WeakLinkage);
443 // If we haven't defined the impl function yet, do so now
444 if (F->isDeclaration()) {
445 // Get the arguments for the function.
446 Function::arg_iterator args = F->arg_begin();
447 Value* Val = args++; Val->setName("Val");
448 Value* Rep = args++; Rep->setName("Rep");
449 Value* Lo = args++; Lo->setName("Lo");
450 Value* Hi = args++; Hi->setName("Hi");
452 // Get some types we need
453 const IntegerType* ValTy = cast<IntegerType>(Val->getType());
454 const IntegerType* RepTy = cast<IntegerType>(Rep->getType());
455 uint32_t ValBits = ValTy->getBitWidth();
456 uint32_t RepBits = RepTy->getBitWidth();
458 // Constant Definitions
459 ConstantInt* RepBitWidth = ConstantInt::get(Type::Int32Ty, RepBits);
460 ConstantInt* RepMask = ConstantInt::getAllOnesValue(RepTy);
461 ConstantInt* ValMask = ConstantInt::getAllOnesValue(ValTy);
462 ConstantInt* One = ConstantInt::get(Type::Int32Ty, 1);
463 ConstantInt* ValOne = ConstantInt::get(ValTy, 1);
464 ConstantInt* Zero = ConstantInt::get(Type::Int32Ty, 0);
465 ConstantInt* ValZero = ConstantInt::get(ValTy, 0);
467 // Basic blocks we fill in below.
468 BasicBlock* entry = new BasicBlock("entry", F, 0);
469 BasicBlock* large = new BasicBlock("large", F, 0);
470 BasicBlock* small = new BasicBlock("small", F, 0);
471 BasicBlock* reverse = new BasicBlock("reverse", F, 0);
472 BasicBlock* result = new BasicBlock("result", F, 0);
474 // BASIC BLOCK: entry
475 // First, get the number of bits that we're placing as an i32
476 ICmpInst* is_forward =
477 new ICmpInst(ICmpInst::ICMP_ULT, Lo, Hi, "", entry);
478 SelectInst* Hi_pn = new SelectInst(is_forward, Hi, Lo, "", entry);
479 SelectInst* Lo_pn = new SelectInst(is_forward, Lo, Hi, "", entry);
480 BinaryOperator* NumBits = BinaryOperator::createSub(Hi_pn, Lo_pn, "",entry);
481 NumBits = BinaryOperator::createAdd(NumBits, One, "", entry);
482 // Now, convert Lo and Hi to ValTy bit width
484 Lo = new ZExtInst(Lo_pn, ValTy, "", entry);
485 } else if (ValBits < 32) {
486 Lo = new TruncInst(Lo_pn, ValTy, "", entry);
488 // Determine if the replacement bits are larger than the number of bits we
489 // are replacing and deal with it.
491 new ICmpInst(ICmpInst::ICMP_ULT, NumBits, RepBitWidth, "", entry);
492 new BranchInst(large, small, is_large, entry);
494 // BASIC BLOCK: large
495 Instruction* MaskBits =
496 BinaryOperator::createSub(RepBitWidth, NumBits, "", large);
497 MaskBits = CastInst::createIntegerCast(MaskBits, RepMask->getType(),
499 BinaryOperator* Mask1 =
500 BinaryOperator::createLShr(RepMask, MaskBits, "", large);
501 BinaryOperator* Rep2 = BinaryOperator::createAnd(Mask1, Rep, "", large);
502 new BranchInst(small, large);
504 // BASIC BLOCK: small
505 PHINode* Rep3 = new PHINode(RepTy, "", small);
506 Rep3->reserveOperandSpace(2);
507 Rep3->addIncoming(Rep2, large);
508 Rep3->addIncoming(Rep, entry);
510 if (ValBits > RepBits)
511 Rep4 = new ZExtInst(Rep3, ValTy, "", small);
512 else if (ValBits < RepBits)
513 Rep4 = new TruncInst(Rep3, ValTy, "", small);
514 new BranchInst(result, reverse, is_forward, small);
516 // BASIC BLOCK: reverse (reverses the bits of the replacement)
517 // Set up our loop counter as a PHI so we can decrement on each iteration.
518 // We will loop for the number of bits in the replacement value.
519 PHINode *Count = new PHINode(Type::Int32Ty, "count", reverse);
520 Count->reserveOperandSpace(2);
521 Count->addIncoming(NumBits, small);
523 // Get the value that we are shifting bits out of as a PHI because
524 // we'll change this with each iteration.
525 PHINode *BitsToShift = new PHINode(Val->getType(), "val", reverse);
526 BitsToShift->reserveOperandSpace(2);
527 BitsToShift->addIncoming(Rep4, small);
529 // Get the result of the last computation or zero on first iteration
530 PHINode *RRes = new PHINode(Val->getType(), "rres", reverse);
531 RRes->reserveOperandSpace(2);
532 RRes->addIncoming(ValZero, small);
534 // Decrement the loop counter by one
535 Instruction *Decr = BinaryOperator::createSub(Count, One, "", reverse);
536 Count->addIncoming(Decr, reverse);
538 // Get the bit that we want to move into the result
539 Value *Bit = BinaryOperator::createAnd(BitsToShift, ValOne, "", reverse);
541 // Compute the new value of the bits to shift for the next iteration.
542 Value *NewVal = BinaryOperator::createLShr(BitsToShift, ValOne,"", reverse);
543 BitsToShift->addIncoming(NewVal, reverse);
545 // Shift the bit we extracted into the low bit of the result.
546 Instruction *NewRes = BinaryOperator::createShl(RRes, ValOne, "", reverse);
547 NewRes = BinaryOperator::createOr(NewRes, Bit, "", reverse);
548 RRes->addIncoming(NewRes, reverse);
550 // Terminate loop if we've moved all the bits.
551 ICmpInst *Cond = new ICmpInst(ICmpInst::ICMP_EQ, Decr, Zero, "", reverse);
552 new BranchInst(result, reverse, Cond, reverse);
554 // BASIC BLOCK: result
555 PHINode *Rplcmnt = new PHINode(Val->getType(), "", result);
556 Rplcmnt->reserveOperandSpace(2);
557 Rplcmnt->addIncoming(NewRes, reverse);
558 Rplcmnt->addIncoming(Rep4, small);
559 Value* t0 = CastInst::createIntegerCast(NumBits,ValTy,false,"",result);
560 Value* t1 = BinaryOperator::createShl(ValMask, Lo, "", result);
561 Value* t2 = BinaryOperator::createNot(t1, "", result);
562 Value* t3 = BinaryOperator::createShl(t1, t0, "", result);
563 Value* t4 = BinaryOperator::createOr(t2, t3, "", result);
564 Value* t5 = BinaryOperator::createAnd(t4, Val, "", result);
565 Value* t6 = BinaryOperator::createShl(Rplcmnt, Lo, "", result);
566 Value* Rslt = BinaryOperator::createOr(t5, t6, "part_set", result);
567 new ReturnInst(Rslt, result);
570 // Return a call to the implementation function
577 return new CallInst(F, Args, sizeof(Args)/sizeof(Args[0]), CI->getName(), CI);
581 void IntrinsicLowering::LowerIntrinsicCall(CallInst *CI) {
582 Function *Callee = CI->getCalledFunction();
583 assert(Callee && "Cannot lower an indirect call!");
585 switch (Callee->getIntrinsicID()) {
586 case Intrinsic::not_intrinsic:
587 cerr << "Cannot lower a call to a non-intrinsic function '"
588 << Callee->getName() << "'!\n";
591 cerr << "Error: Code generator does not support intrinsic function '"
592 << Callee->getName() << "'!\n";
595 // The setjmp/longjmp intrinsics should only exist in the code if it was
596 // never optimized (ie, right out of the CFE), or if it has been hacked on
597 // by the lowerinvoke pass. In both cases, the right thing to do is to
598 // convert the call to an explicit setjmp or longjmp call.
599 case Intrinsic::setjmp: {
600 static Constant *SetjmpFCache = 0;
601 Value *V = ReplaceCallWith("setjmp", CI, CI->op_begin()+1, CI->op_end(),
602 Type::Int32Ty, SetjmpFCache);
603 if (CI->getType() != Type::VoidTy)
604 CI->replaceAllUsesWith(V);
607 case Intrinsic::sigsetjmp:
608 if (CI->getType() != Type::VoidTy)
609 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
612 case Intrinsic::longjmp: {
613 static Constant *LongjmpFCache = 0;
614 ReplaceCallWith("longjmp", CI, CI->op_begin()+1, CI->op_end(),
615 Type::VoidTy, LongjmpFCache);
619 case Intrinsic::siglongjmp: {
620 // Insert the call to abort
621 static Constant *AbortFCache = 0;
622 ReplaceCallWith("abort", CI, CI->op_end(), CI->op_end(),
623 Type::VoidTy, AbortFCache);
626 case Intrinsic::ctpop:
627 CI->replaceAllUsesWith(LowerCTPOP(CI->getOperand(1), CI));
630 case Intrinsic::bswap:
631 CI->replaceAllUsesWith(LowerBSWAP(CI->getOperand(1), CI));
634 case Intrinsic::ctlz:
635 CI->replaceAllUsesWith(LowerCTLZ(CI->getOperand(1), CI));
638 case Intrinsic::cttz: {
639 // cttz(x) -> ctpop(~X & (X-1))
640 Value *Src = CI->getOperand(1);
641 Value *NotSrc = BinaryOperator::createNot(Src, Src->getName()+".not", CI);
642 Value *SrcM1 = ConstantInt::get(Src->getType(), 1);
643 SrcM1 = BinaryOperator::createSub(Src, SrcM1, "", CI);
644 Src = LowerCTPOP(BinaryOperator::createAnd(NotSrc, SrcM1, "", CI), CI);
645 CI->replaceAllUsesWith(Src);
649 case Intrinsic::part_select:
650 CI->replaceAllUsesWith(LowerPartSelect(CI));
653 case Intrinsic::part_set:
654 CI->replaceAllUsesWith(LowerPartSet(CI));
657 case Intrinsic::stacksave:
658 case Intrinsic::stackrestore: {
659 static bool Warned = false;
661 cerr << "WARNING: this target does not support the llvm.stack"
662 << (Callee->getIntrinsicID() == Intrinsic::stacksave ?
663 "save" : "restore") << " intrinsic.\n";
665 if (Callee->getIntrinsicID() == Intrinsic::stacksave)
666 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
670 case Intrinsic::returnaddress:
671 case Intrinsic::frameaddress:
672 cerr << "WARNING: this target does not support the llvm."
673 << (Callee->getIntrinsicID() == Intrinsic::returnaddress ?
674 "return" : "frame") << "address intrinsic.\n";
675 CI->replaceAllUsesWith(ConstantPointerNull::get(
676 cast<PointerType>(CI->getType())));
679 case Intrinsic::prefetch:
680 break; // Simply strip out prefetches on unsupported architectures
682 case Intrinsic::pcmarker:
683 break; // Simply strip out pcmarker on unsupported architectures
684 case Intrinsic::readcyclecounter: {
685 cerr << "WARNING: this target does not support the llvm.readcyclecoun"
686 << "ter intrinsic. It is being lowered to a constant 0\n";
687 CI->replaceAllUsesWith(ConstantInt::get(Type::Int64Ty, 0));
691 case Intrinsic::dbg_stoppoint:
692 case Intrinsic::dbg_region_start:
693 case Intrinsic::dbg_region_end:
694 case Intrinsic::dbg_func_start:
695 case Intrinsic::dbg_declare:
696 case Intrinsic::eh_exception:
697 case Intrinsic::eh_selector:
698 case Intrinsic::eh_filter:
699 break; // Simply strip out debugging and eh intrinsics
701 case Intrinsic::memcpy_i32:
702 case Intrinsic::memcpy_i64: {
703 static Constant *MemcpyFCache = 0;
704 Value *Size = CI->getOperand(3);
705 const Type *IntPtr = TD.getIntPtrType();
706 if (Size->getType()->getPrimitiveSizeInBits() <
707 IntPtr->getPrimitiveSizeInBits())
708 Size = new ZExtInst(Size, IntPtr, "", CI);
709 else if (Size->getType()->getPrimitiveSizeInBits() >
710 IntPtr->getPrimitiveSizeInBits())
711 Size = new TruncInst(Size, IntPtr, "", CI);
713 Ops[0] = CI->getOperand(1);
714 Ops[1] = CI->getOperand(2);
716 ReplaceCallWith("memcpy", CI, Ops, Ops+3, CI->getOperand(1)->getType(),
720 case Intrinsic::memmove_i32:
721 case Intrinsic::memmove_i64: {
722 static Constant *MemmoveFCache = 0;
723 Value *Size = CI->getOperand(3);
724 const Type *IntPtr = TD.getIntPtrType();
725 if (Size->getType()->getPrimitiveSizeInBits() <
726 IntPtr->getPrimitiveSizeInBits())
727 Size = new ZExtInst(Size, IntPtr, "", CI);
728 else if (Size->getType()->getPrimitiveSizeInBits() >
729 IntPtr->getPrimitiveSizeInBits())
730 Size = new TruncInst(Size, IntPtr, "", CI);
732 Ops[0] = CI->getOperand(1);
733 Ops[1] = CI->getOperand(2);
735 ReplaceCallWith("memmove", CI, Ops, Ops+3, CI->getOperand(1)->getType(),
739 case Intrinsic::memset_i32:
740 case Intrinsic::memset_i64: {
741 static Constant *MemsetFCache = 0;
742 Value *Size = CI->getOperand(3);
743 const Type *IntPtr = TD.getIntPtrType();
744 if (Size->getType()->getPrimitiveSizeInBits() <
745 IntPtr->getPrimitiveSizeInBits())
746 Size = new ZExtInst(Size, IntPtr, "", CI);
747 else if (Size->getType()->getPrimitiveSizeInBits() >
748 IntPtr->getPrimitiveSizeInBits())
749 Size = new TruncInst(Size, IntPtr, "", CI);
751 Ops[0] = CI->getOperand(1);
752 // Extend the amount to i32.
753 Ops[1] = new ZExtInst(CI->getOperand(2), Type::Int32Ty, "", CI);
755 ReplaceCallWith("memset", CI, Ops, Ops+3, CI->getOperand(1)->getType(),
759 case Intrinsic::sqrt_f32: {
760 static Constant *sqrtfFCache = 0;
761 ReplaceCallWith("sqrtf", CI, CI->op_begin()+1, CI->op_end(),
762 Type::FloatTy, sqrtfFCache);
765 case Intrinsic::sqrt_f64: {
766 static Constant *sqrtFCache = 0;
767 ReplaceCallWith("sqrt", CI, CI->op_begin()+1, CI->op_end(),
768 Type::DoubleTy, sqrtFCache);
773 assert(CI->use_empty() &&
774 "Lowering should have eliminated any uses of the intrinsic call!");
775 CI->eraseFromParent();