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/CodeGen/IntrinsicLowering.h"
15 #include "llvm/Constants.h"
16 #include "llvm/DerivedTypes.h"
17 #include "llvm/Module.h"
18 #include "llvm/Instructions.h"
19 #include "llvm/Type.h"
24 template <class ArgIt>
25 static Function *EnsureFunctionExists(Module &M, const char *Name,
26 ArgIt ArgBegin, ArgIt ArgEnd,
28 if (Function *F = M.getNamedFunction(Name)) return F;
29 // It doesn't already exist in the program, insert a new definition now.
30 std::vector<const Type *> ParamTys;
31 for (ArgIt I = ArgBegin; I != ArgEnd; ++I)
32 ParamTys.push_back(I->getType());
33 return 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 unsigned *castOpcodes,
44 const Type *RetTy, Function *&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 FCache = M->getNamedFunction(NewFn);
51 // It doesn't already exist in the program, insert a new definition now.
52 std::vector<const Type *> ParamTys;
53 for (ArgIt I = ArgBegin; I != ArgEnd; ++I)
54 ParamTys.push_back((*I)->getType());
55 FCache = M->getOrInsertFunction(NewFn,
56 FunctionType::get(RetTy, ParamTys, false));
60 const FunctionType *FT = FCache->getFunctionType();
61 std::vector<Value*> Operands;
63 for (ArgIt I = ArgBegin; I != ArgEnd && ArgNo != FT->getNumParams();
66 if (Arg->getType() != FT->getParamType(ArgNo))
67 if (castOpcodes[ArgNo])
68 Arg = CastInst::create(Instruction::CastOps(castOpcodes[ArgNo]),
69 Arg, FT->getParamType(ArgNo), Arg->getName(), CI);
71 Arg = CastInst::createInferredCast(Arg, FT->getParamType(ArgNo),
73 Operands.push_back(Arg);
75 // Pass nulls into any additional arguments...
76 for (; ArgNo != FT->getNumParams(); ++ArgNo)
77 Operands.push_back(Constant::getNullValue(FT->getParamType(ArgNo)));
79 std::string Name = CI->getName(); CI->setName("");
80 if (FT->getReturnType() == Type::VoidTy) Name.clear();
81 CallInst *NewCI = new CallInst(FCache, Operands, Name, CI);
82 if (!CI->use_empty()) {
84 if (CI->getType() != NewCI->getType())
85 V = CastInst::createInferredCast(NewCI, CI->getType(), Name, CI);
86 CI->replaceAllUsesWith(V);
91 void IntrinsicLowering::AddPrototypes(Module &M) {
92 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
93 if (I->isExternal() && !I->use_empty())
94 switch (I->getIntrinsicID()) {
96 case Intrinsic::setjmp:
97 EnsureFunctionExists(M, "setjmp", I->arg_begin(), I->arg_end(),
100 case Intrinsic::longjmp:
101 EnsureFunctionExists(M, "longjmp", I->arg_begin(), I->arg_end(),
104 case Intrinsic::siglongjmp:
105 EnsureFunctionExists(M, "abort", I->arg_end(), I->arg_end(),
108 case Intrinsic::memcpy_i32:
109 case Intrinsic::memcpy_i64:
110 EnsureFunctionExists(M, "memcpy", I->arg_begin(), --I->arg_end(),
111 I->arg_begin()->getType());
113 case Intrinsic::memmove_i32:
114 case Intrinsic::memmove_i64:
115 EnsureFunctionExists(M, "memmove", I->arg_begin(), --I->arg_end(),
116 I->arg_begin()->getType());
118 case Intrinsic::memset_i32:
119 case Intrinsic::memset_i64:
120 M.getOrInsertFunction("memset", PointerType::get(Type::SByteTy),
121 PointerType::get(Type::SByteTy),
122 Type::IntTy, (--(--I->arg_end()))->getType(),
125 case Intrinsic::isunordered_f32:
126 case Intrinsic::isunordered_f64:
127 EnsureFunctionExists(M, "isunordered", I->arg_begin(), I->arg_end(),
130 case Intrinsic::sqrt_f32:
131 case Intrinsic::sqrt_f64:
132 if(I->arg_begin()->getType() == Type::FloatTy)
133 EnsureFunctionExists(M, "sqrtf", I->arg_begin(), I->arg_end(),
136 EnsureFunctionExists(M, "sqrt", I->arg_begin(), I->arg_end(),
142 /// LowerBSWAP - Emit the code to lower bswap of V before the specified
144 static Value *LowerBSWAP(Value *V, Instruction *IP) {
145 assert(V->getType()->isInteger() && "Can't bswap a non-integer type!");
147 unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
150 default: assert(0 && "Unhandled type size of value to byteswap!");
152 Value *Tmp1 = new ShiftInst(Instruction::Shl, V,
153 ConstantInt::get(Type::UByteTy,8),"bswap.2",IP);
154 Value *Tmp2 = new ShiftInst(Instruction::LShr, V,
155 ConstantInt::get(Type::UByteTy,8),"bswap.1",IP);
156 V = BinaryOperator::createOr(Tmp1, Tmp2, "bswap.i16", IP);
160 Value *Tmp4 = new ShiftInst(Instruction::Shl, V,
161 ConstantInt::get(Type::UByteTy,24),"bswap.4", IP);
162 Value *Tmp3 = new ShiftInst(Instruction::Shl, V,
163 ConstantInt::get(Type::UByteTy,8),"bswap.3",IP);
164 Value *Tmp2 = new ShiftInst(Instruction::LShr, V,
165 ConstantInt::get(Type::UByteTy,8),"bswap.2",IP);
166 Value *Tmp1 = new ShiftInst(Instruction::LShr, V,
167 ConstantInt::get(Type::UByteTy,24),"bswap.1", IP);
168 Tmp3 = BinaryOperator::createAnd(Tmp3,
169 ConstantInt::get(Type::UIntTy, 0xFF0000),
171 Tmp2 = BinaryOperator::createAnd(Tmp2,
172 ConstantInt::get(Type::UIntTy, 0xFF00),
174 Tmp4 = BinaryOperator::createOr(Tmp4, Tmp3, "bswap.or1", IP);
175 Tmp2 = BinaryOperator::createOr(Tmp2, Tmp1, "bswap.or2", IP);
176 V = BinaryOperator::createOr(Tmp4, Tmp3, "bswap.i32", IP);
180 Value *Tmp8 = new ShiftInst(Instruction::Shl, V,
181 ConstantInt::get(Type::UByteTy,56),"bswap.8", IP);
182 Value *Tmp7 = new ShiftInst(Instruction::Shl, V,
183 ConstantInt::get(Type::UByteTy,40),"bswap.7", IP);
184 Value *Tmp6 = new ShiftInst(Instruction::Shl, V,
185 ConstantInt::get(Type::UByteTy,24),"bswap.6", IP);
186 Value *Tmp5 = new ShiftInst(Instruction::Shl, V,
187 ConstantInt::get(Type::UByteTy,8),"bswap.5", IP);
188 Value* Tmp4 = new ShiftInst(Instruction::LShr, V,
189 ConstantInt::get(Type::UByteTy,8),"bswap.4", IP);
190 Value* Tmp3 = new ShiftInst(Instruction::LShr, V,
191 ConstantInt::get(Type::UByteTy,24),"bswap.3", IP);
192 Value* Tmp2 = new ShiftInst(Instruction::LShr, V,
193 ConstantInt::get(Type::UByteTy,40),"bswap.2", IP);
194 Value* Tmp1 = new ShiftInst(Instruction::LShr, V,
195 ConstantInt::get(Type::UByteTy,56),"bswap.1", IP);
196 Tmp7 = BinaryOperator::createAnd(Tmp7,
197 ConstantInt::get(Type::ULongTy,
198 0xFF000000000000ULL),
200 Tmp6 = BinaryOperator::createAnd(Tmp6,
201 ConstantInt::get(Type::ULongTy, 0xFF0000000000ULL),
203 Tmp5 = BinaryOperator::createAnd(Tmp5,
204 ConstantInt::get(Type::ULongTy, 0xFF00000000ULL),
206 Tmp4 = BinaryOperator::createAnd(Tmp4,
207 ConstantInt::get(Type::ULongTy, 0xFF000000ULL),
209 Tmp3 = BinaryOperator::createAnd(Tmp3,
210 ConstantInt::get(Type::ULongTy, 0xFF0000ULL),
212 Tmp2 = BinaryOperator::createAnd(Tmp2,
213 ConstantInt::get(Type::ULongTy, 0xFF00ULL),
215 Tmp8 = BinaryOperator::createOr(Tmp8, Tmp7, "bswap.or1", IP);
216 Tmp6 = BinaryOperator::createOr(Tmp6, Tmp5, "bswap.or2", IP);
217 Tmp4 = BinaryOperator::createOr(Tmp4, Tmp3, "bswap.or3", IP);
218 Tmp2 = BinaryOperator::createOr(Tmp2, Tmp1, "bswap.or4", IP);
219 Tmp8 = BinaryOperator::createOr(Tmp8, Tmp6, "bswap.or5", IP);
220 Tmp4 = BinaryOperator::createOr(Tmp4, Tmp2, "bswap.or6", IP);
221 V = BinaryOperator::createOr(Tmp8, Tmp4, "bswap.i64", IP);
228 /// LowerCTPOP - Emit the code to lower ctpop of V before the specified
230 static Value *LowerCTPOP(Value *V, Instruction *IP) {
231 assert(V->getType()->isInteger() && "Can't ctpop a non-integer type!");
233 static const uint64_t MaskValues[6] = {
234 0x5555555555555555ULL, 0x3333333333333333ULL,
235 0x0F0F0F0F0F0F0F0FULL, 0x00FF00FF00FF00FFULL,
236 0x0000FFFF0000FFFFULL, 0x00000000FFFFFFFFULL
239 unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
241 for (unsigned i = 1, ct = 0; i != BitSize; i <<= 1, ++ct) {
243 ConstantExpr::getCast(ConstantInt::get(Type::ULongTy, MaskValues[ct]),
245 Value *LHS = BinaryOperator::createAnd(V, MaskCst, "cppop.and1", IP);
246 Value *VShift = new ShiftInst(Instruction::LShr, V,
247 ConstantInt::get(Type::UByteTy, i), "ctpop.sh", IP);
248 Value *RHS = BinaryOperator::createAnd(VShift, MaskCst, "cppop.and2", IP);
249 V = BinaryOperator::createAdd(LHS, RHS, "ctpop.step", IP);
255 /// LowerCTLZ - Emit the code to lower ctlz of V before the specified
257 static Value *LowerCTLZ(Value *V, Instruction *IP) {
259 unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
260 for (unsigned i = 1; i != BitSize; i <<= 1) {
261 Value *ShVal = ConstantInt::get(Type::UByteTy, i);
262 ShVal = new ShiftInst(Instruction::LShr, V, ShVal, "ctlz.sh", IP);
263 V = BinaryOperator::createOr(V, ShVal, "ctlz.step", IP);
266 V = BinaryOperator::createNot(V, "", IP);
267 return LowerCTPOP(V, IP);
272 void IntrinsicLowering::LowerIntrinsicCall(CallInst *CI) {
273 Function *Callee = CI->getCalledFunction();
274 assert(Callee && "Cannot lower an indirect call!");
276 switch (Callee->getIntrinsicID()) {
277 case Intrinsic::not_intrinsic:
278 std::cerr << "Cannot lower a call to a non-intrinsic function '"
279 << Callee->getName() << "'!\n";
282 std::cerr << "Error: Code generator does not support intrinsic function '"
283 << Callee->getName() << "'!\n";
286 // The setjmp/longjmp intrinsics should only exist in the code if it was
287 // never optimized (ie, right out of the CFE), or if it has been hacked on
288 // by the lowerinvoke pass. In both cases, the right thing to do is to
289 // convert the call to an explicit setjmp or longjmp call.
290 case Intrinsic::setjmp: {
291 static Function *SetjmpFCache = 0;
292 static const unsigned castOpcodes[] = { Instruction::BitCast };
293 Value *V = ReplaceCallWith("setjmp", CI, CI->op_begin()+1, CI->op_end(),
294 castOpcodes, Type::IntTy, SetjmpFCache);
295 if (CI->getType() != Type::VoidTy)
296 CI->replaceAllUsesWith(V);
299 case Intrinsic::sigsetjmp:
300 if (CI->getType() != Type::VoidTy)
301 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
304 case Intrinsic::longjmp: {
305 static Function *LongjmpFCache = 0;
306 static const unsigned castOpcodes[] =
307 { Instruction::BitCast, 0 };
308 ReplaceCallWith("longjmp", CI, CI->op_begin()+1, CI->op_end(),
309 castOpcodes, Type::VoidTy, LongjmpFCache);
313 case Intrinsic::siglongjmp: {
314 // Insert the call to abort
315 static Function *AbortFCache = 0;
316 static const unsigned castOpcodes[] =
317 { Instruction::BitCast, 0 };
318 ReplaceCallWith("abort", CI, CI->op_end(), CI->op_end(),
319 castOpcodes, Type::VoidTy, AbortFCache);
322 case Intrinsic::ctpop_i8:
323 case Intrinsic::ctpop_i16:
324 case Intrinsic::ctpop_i32:
325 case Intrinsic::ctpop_i64:
326 CI->replaceAllUsesWith(LowerCTPOP(CI->getOperand(1), CI));
329 case Intrinsic::bswap_i16:
330 case Intrinsic::bswap_i32:
331 case Intrinsic::bswap_i64:
332 CI->replaceAllUsesWith(LowerBSWAP(CI->getOperand(1), CI));
335 case Intrinsic::ctlz_i8:
336 case Intrinsic::ctlz_i16:
337 case Intrinsic::ctlz_i32:
338 case Intrinsic::ctlz_i64:
339 CI->replaceAllUsesWith(LowerCTLZ(CI->getOperand(1), CI));
342 case Intrinsic::cttz_i8:
343 case Intrinsic::cttz_i16:
344 case Intrinsic::cttz_i32:
345 case Intrinsic::cttz_i64: {
346 // cttz(x) -> ctpop(~X & (X-1))
347 Value *Src = CI->getOperand(1);
348 Value *NotSrc = BinaryOperator::createNot(Src, Src->getName()+".not", CI);
349 Value *SrcM1 = ConstantInt::get(Src->getType(), 1);
350 SrcM1 = BinaryOperator::createSub(Src, SrcM1, "", CI);
351 Src = LowerCTPOP(BinaryOperator::createAnd(NotSrc, SrcM1, "", CI), CI);
352 CI->replaceAllUsesWith(Src);
356 case Intrinsic::stacksave:
357 case Intrinsic::stackrestore: {
358 static bool Warned = false;
360 std::cerr << "WARNING: this target does not support the llvm.stack"
361 << (Callee->getIntrinsicID() == Intrinsic::stacksave ?
362 "save" : "restore") << " intrinsic.\n";
364 if (Callee->getIntrinsicID() == Intrinsic::stacksave)
365 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
369 case Intrinsic::returnaddress:
370 case Intrinsic::frameaddress:
371 std::cerr << "WARNING: this target does not support the llvm."
372 << (Callee->getIntrinsicID() == Intrinsic::returnaddress ?
373 "return" : "frame") << "address intrinsic.\n";
374 CI->replaceAllUsesWith(ConstantPointerNull::get(
375 cast<PointerType>(CI->getType())));
378 case Intrinsic::prefetch:
379 break; // Simply strip out prefetches on unsupported architectures
381 case Intrinsic::pcmarker:
382 break; // Simply strip out pcmarker on unsupported architectures
383 case Intrinsic::readcyclecounter: {
384 std::cerr << "WARNING: this target does not support the llvm.readcyclecoun"
385 << "ter intrinsic. It is being lowered to a constant 0\n";
386 CI->replaceAllUsesWith(ConstantInt::get(Type::ULongTy, 0));
390 case Intrinsic::dbg_stoppoint:
391 case Intrinsic::dbg_region_start:
392 case Intrinsic::dbg_region_end:
393 case Intrinsic::dbg_func_start:
394 case Intrinsic::dbg_declare:
395 break; // Simply strip out debugging intrinsics
397 case Intrinsic::memcpy_i32: {
398 // The memcpy intrinsic take an extra alignment argument that the memcpy
399 // libc function does not.
400 static unsigned opcodes[] =
401 { Instruction::BitCast, Instruction::BitCast, Instruction::BitCast };
403 // if (target_is_64_bit) opcodes[2] = Instruction::ZExt;
404 // else opcodes[2] = Instruction::BitCast;
405 static Function *MemcpyFCache = 0;
406 ReplaceCallWith("memcpy", CI, CI->op_begin()+1, CI->op_end()-1,
407 opcodes, (*(CI->op_begin()+1))->getType(), MemcpyFCache);
410 case Intrinsic::memcpy_i64: {
411 static unsigned opcodes[] =
412 { Instruction::BitCast, Instruction::BitCast, Instruction::Trunc };
414 // if (target_is_64_bit) opcodes[2] = Instruction::BitCast;
415 // else opcodes[2] = Instruction::Trunc;
416 static Function *MemcpyFCache = 0;
417 ReplaceCallWith("memcpy", CI, CI->op_begin()+1, CI->op_end()-1,
418 opcodes, (*(CI->op_begin()+1))->getType(), MemcpyFCache);
421 case Intrinsic::memmove_i32: {
422 // The memmove intrinsic take an extra alignment argument that the memmove
423 // libc function does not.
424 static unsigned opcodes[] =
425 { Instruction::BitCast, Instruction::BitCast, Instruction::BitCast };
427 // if (target_is_64_bit) opcodes[2] = Instruction::ZExt;
428 // else opcodes[2] = Instruction::BitCast;
429 static Function *MemmoveFCache = 0;
430 ReplaceCallWith("memmove", CI, CI->op_begin()+1, CI->op_end()-1,
431 opcodes, (*(CI->op_begin()+1))->getType(), MemmoveFCache);
434 case Intrinsic::memmove_i64: {
435 // The memmove intrinsic take an extra alignment argument that the memmove
436 // libc function does not.
437 static const unsigned opcodes[] =
438 { Instruction::BitCast, Instruction::BitCast, Instruction::Trunc };
439 // if (target_is_64_bit) opcodes[2] = Instruction::BitCast;
440 // else opcodes[2] = Instruction::Trunc;
441 static Function *MemmoveFCache = 0;
442 ReplaceCallWith("memmove", CI, CI->op_begin()+1, CI->op_end()-1,
443 opcodes, (*(CI->op_begin()+1))->getType(), MemmoveFCache);
446 case Intrinsic::memset_i32: {
447 // The memset intrinsic take an extra alignment argument that the memset
448 // libc function does not.
449 static const unsigned opcodes[] =
450 { Instruction::BitCast, Instruction::ZExt, Instruction::ZExt, 0 };
451 // if (target_is_64_bit) opcodes[2] = Instruction::BitCast;
452 // else opcodes[2] = Instruction::ZExt;
453 static Function *MemsetFCache = 0;
454 ReplaceCallWith("memset", CI, CI->op_begin()+1, CI->op_end()-1,
455 opcodes, (*(CI->op_begin()+1))->getType(), MemsetFCache);
457 case Intrinsic::memset_i64: {
458 // The memset intrinsic take an extra alignment argument that the memset
459 // libc function does not.
460 static const unsigned opcodes[] =
461 { Instruction::BitCast, Instruction::ZExt, Instruction::Trunc, 0 };
462 // if (target_is_64_bit) opcodes[2] = Instruction::BitCast;
463 // else opcodes[2] = Instruction::Trunc;
464 static Function *MemsetFCache = 0;
465 ReplaceCallWith("memset", CI, CI->op_begin()+1, CI->op_end()-1,
466 opcodes, (*(CI->op_begin()+1))->getType(), MemsetFCache);
469 case Intrinsic::isunordered_f32:
470 case Intrinsic::isunordered_f64: {
471 Value *L = CI->getOperand(1);
472 Value *R = CI->getOperand(2);
474 Value *LIsNan = new SetCondInst(Instruction::SetNE, L, L, "LIsNan", CI);
475 Value *RIsNan = new SetCondInst(Instruction::SetNE, R, R, "RIsNan", CI);
476 CI->replaceAllUsesWith(
477 BinaryOperator::create(Instruction::Or, LIsNan, RIsNan,
481 case Intrinsic::sqrt_f32: {
482 static const unsigned opcodes[] = { 0 };
483 static Function *sqrtfFCache = 0;
484 ReplaceCallWith("sqrtf", CI, CI->op_begin()+1, CI->op_end(),
485 opcodes, Type::FloatTy, sqrtfFCache);
488 case Intrinsic::sqrt_f64: {
489 static const unsigned opcodes[] = { 0 };
490 static Function *sqrtFCache = 0;
491 ReplaceCallWith("sqrt", CI, CI->op_begin()+1, CI->op_end(),
492 opcodes, Type::DoubleTy, sqrtFCache);
497 assert(CI->use_empty() &&
498 "Lowering should have eliminated any uses of the intrinsic call!");
499 CI->eraseFromParent();