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::InternalLinkage);
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 Value* Val = F->getOperand(0);
274 Value* Right = F->getOperand(1);
275 Value* Left = F->getOperand(2);
277 // We want to select a range of bits here such that [Left, Right] is shifted
278 // down to the low bits. However, it is quite possible that Left is smaller
279 // than Right in which case the bits have to be reversed.
281 // Create the blocks we will need for the two cases (forward, reverse)
282 BasicBlock* CurBB = new BasicBlock("entry", F);
283 BasicBlock *RevSize = new BasicBlock("revsize", CurBB->getParent());
284 BasicBlock *FwdSize = new BasicBlock("fwdsize", CurBB->getParent());
285 BasicBlock *Compute = new BasicBlock("compute", CurBB->getParent());
286 BasicBlock *Reverse = new BasicBlock("reverse", CurBB->getParent());
287 BasicBlock *RsltBlk = new BasicBlock("result", CurBB->getParent());
289 // Cast Left and Right to the size of Val so the widths are all the same
290 if (Left->getType() != Val->getType())
291 Left = CastInst::createIntegerCast(Left, Val->getType(), false,
293 if (Right->getType() != Val->getType())
294 Right = CastInst::createIntegerCast(Right, Val->getType(), false,
297 // Compute a few things that both cases will need, up front.
298 Constant* Zero = ConstantInt::get(Val->getType(), 0);
299 Constant* One = ConstantInt::get(Val->getType(), 1);
300 Constant* AllOnes = ConstantInt::getAllOnesValue(Val->getType());
302 // Compare the Left and Right bit positions. This is used to determine
303 // which case we have (forward or reverse)
304 ICmpInst *Cmp = new ICmpInst(ICmpInst::ICMP_ULT, Left, Right, "less",CurBB);
305 new BranchInst(RevSize, FwdSize, Cmp, CurBB);
307 // First, copmute the number of bits in the forward case.
308 Instruction* FBitSize =
309 BinaryOperator::createSub(Left, Right,"fbits", FwdSize);
310 new BranchInst(Compute, FwdSize);
312 // Second, compute the number of bits in the reverse case.
313 Instruction* RBitSize =
314 BinaryOperator::createSub(Right, Left, "rbits", RevSize);
315 new BranchInst(Compute, RevSize);
317 // Now, compute the bit range. Start by getting the bitsize and the shift
318 // amount (either Left or Right) from PHI nodes. Then we compute a mask for
319 // the number of bits we want in the range. We shift the bits down to the
320 // least significant bits, apply the mask to zero out unwanted high bits,
321 // and we have computed the "forward" result. It may still need to be
324 // Get the BitSize from one of the two subtractions
325 PHINode *BitSize = new PHINode(Val->getType(), "bits", Compute);
326 BitSize->reserveOperandSpace(2);
327 BitSize->addIncoming(FBitSize, FwdSize);
328 BitSize->addIncoming(RBitSize, RevSize);
330 // Get the ShiftAmount as the smaller of Left/Right
331 PHINode *ShiftAmt = new PHINode(Val->getType(), "shiftamt", Compute);
332 ShiftAmt->reserveOperandSpace(2);
333 ShiftAmt->addIncoming(Right, FwdSize);
334 ShiftAmt->addIncoming(Left, RevSize);
336 // Increment the bit size
337 Instruction *BitSizePlusOne =
338 BinaryOperator::createAdd(BitSize, One, "bits", Compute);
340 // Create a Mask to zero out the high order bits.
342 BinaryOperator::createShl(AllOnes, BitSizePlusOne, "mask", Compute);
343 Mask = BinaryOperator::createNot(Mask, "mask", Compute);
345 // Shift the bits down and apply the mask
347 BinaryOperator::createLShr(Val, ShiftAmt, "fres", Compute);
348 FRes = BinaryOperator::createAnd(FRes, Mask, "fres", Compute);
349 new BranchInst(Reverse, RsltBlk, Cmp, Compute);
351 // In the Reverse block we have the mask already in FRes but we must reverse
352 // it by shifting FRes bits right and putting them in RRes by shifting them
355 // First set up our loop counters
356 PHINode *Count = new PHINode(Val->getType(), "count", Reverse);
357 Count->reserveOperandSpace(2);
358 Count->addIncoming(BitSizePlusOne, Compute);
360 // Next, get the value that we are shifting.
361 PHINode *BitsToShift = new PHINode(Val->getType(), "val", Reverse);
362 BitsToShift->reserveOperandSpace(2);
363 BitsToShift->addIncoming(FRes, Compute);
365 // Finally, get the result of the last computation
366 PHINode *RRes = new PHINode(Val->getType(), "rres", Reverse);
367 RRes->reserveOperandSpace(2);
368 RRes->addIncoming(Zero, Compute);
370 // Decrement the counter
371 Instruction *Decr = BinaryOperator::createSub(Count, One, "decr", Reverse);
372 Count->addIncoming(Decr, Reverse);
374 // Compute the Bit that we want to move
376 BinaryOperator::createAnd(BitsToShift, One, "bit", Reverse);
378 // Compute the new value for next iteration.
379 Instruction *NewVal =
380 BinaryOperator::createLShr(BitsToShift, One, "rshift", Reverse);
381 BitsToShift->addIncoming(NewVal, Reverse);
383 // Shift the bit into the low bits of the result.
384 Instruction *NewRes =
385 BinaryOperator::createShl(RRes, One, "lshift", Reverse);
386 NewRes = BinaryOperator::createOr(NewRes, Bit, "addbit", Reverse);
387 RRes->addIncoming(NewRes, Reverse);
389 // Terminate loop if we've moved all the bits.
391 new ICmpInst(ICmpInst::ICMP_EQ, Decr, Zero, "cond", Reverse);
392 new BranchInst(RsltBlk, Reverse, Cond, Reverse);
394 // Finally, in the result block, select one of the two results with a PHI
395 // node and return the result;
397 PHINode *BitSelect = new PHINode(Val->getType(), "part_select", CurBB);
398 BitSelect->reserveOperandSpace(2);
399 BitSelect->addIncoming(FRes, Compute);
400 BitSelect->addIncoming(NewRes, Reverse);
401 new ReturnInst(BitSelect, CurBB);
404 // Return a call to the implementation function
406 Args[0] = CI->getOperand(0);
407 Args[1] = CI->getOperand(1);
408 Args[2] = CI->getOperand(2);
409 return new CallInst(F, Args, 3, CI->getName(), CI);
412 /// Convert the llvm.part.set.iX.iY.iZ intrinsic. This intrinsic takes
413 /// four integer arguments (iAny %Value, iAny %Replacement, i32 %Low, i32 %High)
414 /// The first two arguments can be any bit width. The result is the same width
415 /// as %Value. The operation replaces bits between %Low and %High with the value
416 /// in %Replacement. If %Replacement is not the same width, it is truncated or
417 /// zero extended as appropriate to fit the bits being replaced. If %Low is
418 /// greater than %High then the inverse set of bits are replaced.
419 /// @brief Lowering of llvm.bit.part.set intrinsic.
420 static Instruction *LowerPartSet(CallInst *CI) {
421 // Make sure we're dealing with a part select intrinsic here
422 Function *F = CI->getCalledFunction();
423 const FunctionType *FT = F->getFunctionType();
424 if (!F->isDeclaration() || !FT->getReturnType()->isInteger() ||
425 FT->getNumParams() != 4 || !FT->getParamType(0)->isInteger() ||
426 !FT->getParamType(1)->isInteger() || !FT->getParamType(2)->isInteger() ||
427 !FT->getParamType(3)->isInteger())
430 // Get the intrinsic implementation function by converting all the . to _
431 // in the intrinsic's function name and then reconstructing the function
433 std::string Name(F->getName());
434 for (unsigned i = 4; i < Name.length(); ++i)
437 Module* M = F->getParent();
438 F = cast<Function>(M->getOrInsertFunction(Name, FT));
439 F->setLinkage(GlobalValue::InternalLinkage);
441 // If we haven't defined the impl function yet, do so now
442 if (F->isDeclaration()) {
443 // Note: the following code is based on code generated by llvm2cpp with
444 // the following input. This is just *one* example of a generated function.
445 // The functions vary by bit width of result and first two arguments.
446 // The generated code has been changed to deal with any bit width not just
447 // the 32/64 bitwidths used in the above sample.
449 // define i64 @part_set(i64 %Val, i32 %Rep, i32 %Lo, i32 %Hi) {
451 // %is_forward = icmp ult i32 %Lo, %Hi
452 // %Lo.pn = select i1 %is_forward, i32 %Hi, i32 %Lo
453 // %Hi.pn = select i1 %is_forward, i32 %Lo, i32 %Hi
454 // %iftmp.16.0 = sub i32 %Lo.pn, %Hi.pn
455 // icmp ult i32 %iftmp.16.0, 32
456 // br i1 %1, label %cond_true11, label %cond_next19
458 // %tmp13 = sub i32 32, %iftmp.16.0
459 // %tmp14 = lshr i32 -1, %tmp13
460 // %tmp16 = and i32 %tmp14, %Rep
461 // br label %cond_next19
463 // %iftmp.17.0 = phi i32 [ %tmp16, %cond_true11 ], [ %Rep, %entry ]
464 // %tmp2021 = zext i32 %iftmp.17.0 to i64
465 // icmp ugt i32 %Lo, %Hi
466 // br i1 %2, label %cond_next60, label %cond_true24
468 // %tmp25.cast = zext i32 %Hi to i64
469 // %tmp26 = lshr i64 -1, %tmp25.cast
470 // %tmp27.cast = zext i32 %Lo to i64
471 // %tmp28 = shl i64 %tmp26, %tmp27.cast
472 // %tmp28not = xor i64 %tmp28, -1
473 // %tmp31 = shl i64 %tmp2021, %tmp27.cast
474 // %tmp34 = and i64 %tmp28not, %Val
475 // %Val_addr.064 = or i64 %tmp31, %tmp34
476 // ret i64 %Val_addr.064
478 // %tmp39.cast = zext i32 %Lo to i64
479 // %tmp40 = shl i64 -1, %tmp39.cast
480 // %tmp41.cast = zext i32 %Hi to i64
481 // %tmp42 = shl i64 -1, %tmp41.cast
482 // %tmp45.demorgan = or i64 %tmp42, %tmp40
483 // %tmp45 = xor i64 %tmp45.demorgan, -1
484 // %tmp47 = and i64 %tmp45, %Val
485 // %tmp50 = shl i64 %tmp2021, %tmp39.cast
486 // %tmp52 = sub i32 32, %Hi
487 // %tmp52.cast = zext i32 %tmp52 to i64
488 // %tmp54 = lshr i64 %tmp2021, %tmp52.cast
489 // %tmp57 = or i64 %tmp50, %tmp47
490 // %Val_addr.0 = or i64 %tmp57, %tmp54
491 // ret i64 %Val_addr.0
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);
512 BasicBlock* entry = new BasicBlock("entry",F,0);
513 BasicBlock* large = new BasicBlock("large",F,0);
514 BasicBlock* small = new BasicBlock("small",F,0);
515 BasicBlock* forward = new BasicBlock("cond_true24",F,0);
516 BasicBlock* reverse = new BasicBlock("cond_next60",F,0);
518 // Block entry (entry)
519 // First, convert Lo and Hi to ValTy bit width
521 Hi = new ZExtInst(Hi, ValTy, "", entry);
522 Lo = new ZExtInst(Lo, ValTy, "", entry);
523 } else if (ValBits < 32) {
524 Hi = new TruncInst(Hi, ValTy, "", entry);
525 Lo = new TruncInst(Lo, ValTy, "", entry);
527 ICmpInst* is_forward =
528 new ICmpInst(ICmpInst::ICMP_ULT, Lo, Hi, "", entry);
529 SelectInst* Lo_pn = new SelectInst(is_forward, Hi, Lo, "", entry);
530 SelectInst* Hi_pn = new SelectInst(is_forward, Lo, Hi, "", entry);
531 BinaryOperator* NumBits = BinaryOperator::createSub(Lo_pn, Hi_pn, "",entry);
533 new ICmpInst(ICmpInst::ICMP_ULT, NumBits, RepBitWidth, "", entry);
534 new BranchInst(large, small, is_large, entry);
537 BinaryOperator* MaskBits =
538 BinaryOperator::createSub(RepBitWidth, NumBits, "", large);
539 BinaryOperator* Mask1 =
540 BinaryOperator::createLShr(RepMask, MaskBits, "", large);
541 BinaryOperator* Rep2 = BinaryOperator::createAnd(Mask1, Rep, "", large);
542 new BranchInst(small, large);
545 PHINode* Rep3 = new PHINode(RepTy, "", small);
546 Rep3->reserveOperandSpace(2);
547 Rep3->addIncoming(Rep2, small);
548 Rep3->addIncoming(Rep, entry);
549 CastInst* Rep4 = new ZExtInst(Rep3, ValTy, "", small);
550 ICmpInst* is_reverse =
551 new ICmpInst(ICmpInst::ICMP_UGT, Lo, Hi, "", small);
552 new BranchInst(reverse, forward, is_reverse, small);
555 Value* t1 = BinaryOperator::createLShr(ValMask, Hi, "", forward);
556 Value* t2 = BinaryOperator::createShl(t1, Lo, "", forward);
557 Value* nott2 = BinaryOperator::createXor(t2, ValMask, "", forward);
558 Value* t3 = BinaryOperator::createShl(Rep4, Lo, "", forward);
559 Value* t4 = BinaryOperator::createAnd(nott2, Val, "", forward);
560 Value* FRslt = BinaryOperator::createOr(t3, t4, "", forward);
561 new ReturnInst(FRslt, forward);
564 Value* t5 = BinaryOperator::createShl(ValMask, Lo, "", reverse);
565 Value* t6 = BinaryOperator::createShl(ValMask, Hi, "", reverse);
566 Value* t7 = BinaryOperator::createOr(t6, t5, "", reverse);
567 Value* t8 = BinaryOperator::createXor(t7, ValMask, "", reverse);
568 Value* t9 = BinaryOperator::createAnd(t8, Val, "", reverse);
569 Value* t10 = BinaryOperator::createShl(Rep4, Lo, "", reverse);
570 Value* t11 = BinaryOperator::createSub(RepBitWidth, Hi, "", reverse);
571 Value* t12 = new ZExtInst(t11, ValTy, "", reverse);
572 Value* t13 = BinaryOperator::createLShr(Rep4, t12, "",reverse);
573 Value* t14 = BinaryOperator::createOr(t10, t9, "", reverse);
574 Value* RRslt = BinaryOperator::createOr(t14, t13, "", reverse);
575 new ReturnInst(RRslt, reverse);
578 // Return a call to the implementation function
580 Args[0] = CI->getOperand(0);
581 Args[1] = CI->getOperand(1);
582 Args[2] = CI->getOperand(2);
583 Args[3] = CI->getOperand(3);
584 return new CallInst(F, Args, 4, CI->getName(), CI);
588 void IntrinsicLowering::LowerIntrinsicCall(CallInst *CI) {
589 Function *Callee = CI->getCalledFunction();
590 assert(Callee && "Cannot lower an indirect call!");
592 switch (Callee->getIntrinsicID()) {
593 case Intrinsic::not_intrinsic:
594 cerr << "Cannot lower a call to a non-intrinsic function '"
595 << Callee->getName() << "'!\n";
598 cerr << "Error: Code generator does not support intrinsic function '"
599 << Callee->getName() << "'!\n";
602 // The setjmp/longjmp intrinsics should only exist in the code if it was
603 // never optimized (ie, right out of the CFE), or if it has been hacked on
604 // by the lowerinvoke pass. In both cases, the right thing to do is to
605 // convert the call to an explicit setjmp or longjmp call.
606 case Intrinsic::setjmp: {
607 static Constant *SetjmpFCache = 0;
608 Value *V = ReplaceCallWith("setjmp", CI, CI->op_begin()+1, CI->op_end(),
609 Type::Int32Ty, SetjmpFCache);
610 if (CI->getType() != Type::VoidTy)
611 CI->replaceAllUsesWith(V);
614 case Intrinsic::sigsetjmp:
615 if (CI->getType() != Type::VoidTy)
616 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
619 case Intrinsic::longjmp: {
620 static Constant *LongjmpFCache = 0;
621 ReplaceCallWith("longjmp", CI, CI->op_begin()+1, CI->op_end(),
622 Type::VoidTy, LongjmpFCache);
626 case Intrinsic::siglongjmp: {
627 // Insert the call to abort
628 static Constant *AbortFCache = 0;
629 ReplaceCallWith("abort", CI, CI->op_end(), CI->op_end(),
630 Type::VoidTy, AbortFCache);
633 case Intrinsic::ctpop:
634 CI->replaceAllUsesWith(LowerCTPOP(CI->getOperand(1), CI));
637 case Intrinsic::bswap:
638 CI->replaceAllUsesWith(LowerBSWAP(CI->getOperand(1), CI));
641 case Intrinsic::ctlz:
642 CI->replaceAllUsesWith(LowerCTLZ(CI->getOperand(1), CI));
645 case Intrinsic::cttz: {
646 // cttz(x) -> ctpop(~X & (X-1))
647 Value *Src = CI->getOperand(1);
648 Value *NotSrc = BinaryOperator::createNot(Src, Src->getName()+".not", CI);
649 Value *SrcM1 = ConstantInt::get(Src->getType(), 1);
650 SrcM1 = BinaryOperator::createSub(Src, SrcM1, "", CI);
651 Src = LowerCTPOP(BinaryOperator::createAnd(NotSrc, SrcM1, "", CI), CI);
652 CI->replaceAllUsesWith(Src);
656 case Intrinsic::part_select:
657 CI->replaceAllUsesWith(LowerPartSelect(CI));
660 case Intrinsic::part_set:
661 CI->replaceAllUsesWith(LowerPartSet(CI));
664 case Intrinsic::stacksave:
665 case Intrinsic::stackrestore: {
666 static bool Warned = false;
668 cerr << "WARNING: this target does not support the llvm.stack"
669 << (Callee->getIntrinsicID() == Intrinsic::stacksave ?
670 "save" : "restore") << " intrinsic.\n";
672 if (Callee->getIntrinsicID() == Intrinsic::stacksave)
673 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
677 case Intrinsic::returnaddress:
678 case Intrinsic::frameaddress:
679 cerr << "WARNING: this target does not support the llvm."
680 << (Callee->getIntrinsicID() == Intrinsic::returnaddress ?
681 "return" : "frame") << "address intrinsic.\n";
682 CI->replaceAllUsesWith(ConstantPointerNull::get(
683 cast<PointerType>(CI->getType())));
686 case Intrinsic::prefetch:
687 break; // Simply strip out prefetches on unsupported architectures
689 case Intrinsic::pcmarker:
690 break; // Simply strip out pcmarker on unsupported architectures
691 case Intrinsic::readcyclecounter: {
692 cerr << "WARNING: this target does not support the llvm.readcyclecoun"
693 << "ter intrinsic. It is being lowered to a constant 0\n";
694 CI->replaceAllUsesWith(ConstantInt::get(Type::Int64Ty, 0));
698 case Intrinsic::dbg_stoppoint:
699 case Intrinsic::dbg_region_start:
700 case Intrinsic::dbg_region_end:
701 case Intrinsic::dbg_func_start:
702 case Intrinsic::dbg_declare:
703 case Intrinsic::eh_exception:
704 case Intrinsic::eh_selector:
705 case Intrinsic::eh_filter:
706 break; // Simply strip out debugging and eh intrinsics
708 case Intrinsic::memcpy_i32:
709 case Intrinsic::memcpy_i64: {
710 static Constant *MemcpyFCache = 0;
711 Value *Size = CI->getOperand(3);
712 const Type *IntPtr = TD.getIntPtrType();
713 if (Size->getType()->getPrimitiveSizeInBits() <
714 IntPtr->getPrimitiveSizeInBits())
715 Size = new ZExtInst(Size, IntPtr, "", CI);
716 else if (Size->getType()->getPrimitiveSizeInBits() >
717 IntPtr->getPrimitiveSizeInBits())
718 Size = new TruncInst(Size, IntPtr, "", CI);
720 Ops[0] = CI->getOperand(1);
721 Ops[1] = CI->getOperand(2);
723 ReplaceCallWith("memcpy", CI, Ops, Ops+3, CI->getOperand(1)->getType(),
727 case Intrinsic::memmove_i32:
728 case Intrinsic::memmove_i64: {
729 static Constant *MemmoveFCache = 0;
730 Value *Size = CI->getOperand(3);
731 const Type *IntPtr = TD.getIntPtrType();
732 if (Size->getType()->getPrimitiveSizeInBits() <
733 IntPtr->getPrimitiveSizeInBits())
734 Size = new ZExtInst(Size, IntPtr, "", CI);
735 else if (Size->getType()->getPrimitiveSizeInBits() >
736 IntPtr->getPrimitiveSizeInBits())
737 Size = new TruncInst(Size, IntPtr, "", CI);
739 Ops[0] = CI->getOperand(1);
740 Ops[1] = CI->getOperand(2);
742 ReplaceCallWith("memmove", CI, Ops, Ops+3, CI->getOperand(1)->getType(),
746 case Intrinsic::memset_i32:
747 case Intrinsic::memset_i64: {
748 static Constant *MemsetFCache = 0;
749 Value *Size = CI->getOperand(3);
750 const Type *IntPtr = TD.getIntPtrType();
751 if (Size->getType()->getPrimitiveSizeInBits() <
752 IntPtr->getPrimitiveSizeInBits())
753 Size = new ZExtInst(Size, IntPtr, "", CI);
754 else if (Size->getType()->getPrimitiveSizeInBits() >
755 IntPtr->getPrimitiveSizeInBits())
756 Size = new TruncInst(Size, IntPtr, "", CI);
758 Ops[0] = CI->getOperand(1);
759 // Extend the amount to i32.
760 Ops[1] = new ZExtInst(CI->getOperand(2), Type::Int32Ty, "", CI);
762 ReplaceCallWith("memset", CI, Ops, Ops+3, CI->getOperand(1)->getType(),
766 case Intrinsic::sqrt_f32: {
767 static Constant *sqrtfFCache = 0;
768 ReplaceCallWith("sqrtf", CI, CI->op_begin()+1, CI->op_end(),
769 Type::FloatTy, sqrtfFCache);
772 case Intrinsic::sqrt_f64: {
773 static Constant *sqrtFCache = 0;
774 ReplaceCallWith("sqrt", CI, CI->op_begin()+1, CI->op_end(),
775 Type::DoubleTy, sqrtFCache);
780 assert(CI->use_empty() &&
781 "Lowering should have eliminated any uses of the intrinsic call!");
782 CI->eraseFromParent();