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"
23 template <class ArgIt>
24 static Function *EnsureFunctionExists(Module &M, const char *Name,
25 ArgIt ArgBegin, ArgIt ArgEnd,
27 if (Function *F = M.getNamedFunction(Name)) return F;
28 // It doesn't already exist in the program, insert a new definition now.
29 std::vector<const Type *> ParamTys;
30 for (ArgIt I = ArgBegin; I != ArgEnd; ++I)
31 ParamTys.push_back(I->getType());
32 return M.getOrInsertFunction(Name, FunctionType::get(RetTy, ParamTys, false));
35 /// ReplaceCallWith - This function is used when we want to lower an intrinsic
36 /// call to a call of an external function. This handles hard cases such as
37 /// when there was already a prototype for the external function, and if that
38 /// prototype doesn't match the arguments we expect to pass in.
39 template <class ArgIt>
40 static CallInst *ReplaceCallWith(const char *NewFn, CallInst *CI,
41 ArgIt ArgBegin, ArgIt ArgEnd,
42 const unsigned *castOpcodes,
43 const Type *RetTy, Function *&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 FCache = M->getNamedFunction(NewFn);
50 // It doesn't already exist in the program, insert a new definition now.
51 std::vector<const Type *> ParamTys;
52 for (ArgIt I = ArgBegin; I != ArgEnd; ++I)
53 ParamTys.push_back((*I)->getType());
54 FCache = M->getOrInsertFunction(NewFn,
55 FunctionType::get(RetTy, ParamTys, false));
59 const FunctionType *FT = FCache->getFunctionType();
60 std::vector<Value*> Operands;
62 for (ArgIt I = ArgBegin; I != ArgEnd && ArgNo != FT->getNumParams();
65 if (Arg->getType() != FT->getParamType(ArgNo))
66 if (castOpcodes[ArgNo])
67 Arg = CastInst::create(Instruction::CastOps(castOpcodes[ArgNo]),
68 Arg, FT->getParamType(ArgNo), Arg->getName(), CI);
70 Instruction::CastOps opcode = CastInst::getCastOpcode(Arg,
71 Arg->getType()->isSigned(), FT->getParamType(ArgNo),
72 FT->getParamType(ArgNo)->isSigned());
73 Arg = CastInst::create(opcode, Arg, FT->getParamType(ArgNo),
76 Operands.push_back(Arg);
78 // Pass nulls into any additional arguments...
79 for (; ArgNo != FT->getNumParams(); ++ArgNo)
80 Operands.push_back(Constant::getNullValue(FT->getParamType(ArgNo)));
82 std::string Name = CI->getName(); CI->setName("");
83 if (FT->getReturnType() == Type::VoidTy) Name.clear();
84 CallInst *NewCI = new CallInst(FCache, Operands, Name, CI);
85 if (!CI->use_empty()) {
87 if (CI->getType() != NewCI->getType()) {
88 Instruction::CastOps opcode = CastInst::getCastOpcode(NewCI,
89 NewCI->getType()->isSigned(), CI->getType(),
90 CI->getType()->isSigned());
91 V = CastInst::create(opcode, NewCI, CI->getType(), Name, CI);
93 CI->replaceAllUsesWith(V);
98 void IntrinsicLowering::AddPrototypes(Module &M) {
99 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
100 if (I->isExternal() && !I->use_empty())
101 switch (I->getIntrinsicID()) {
103 case Intrinsic::setjmp:
104 EnsureFunctionExists(M, "setjmp", I->arg_begin(), I->arg_end(),
107 case Intrinsic::longjmp:
108 EnsureFunctionExists(M, "longjmp", I->arg_begin(), I->arg_end(),
111 case Intrinsic::siglongjmp:
112 EnsureFunctionExists(M, "abort", I->arg_end(), I->arg_end(),
115 case Intrinsic::memcpy_i32:
116 case Intrinsic::memcpy_i64:
117 EnsureFunctionExists(M, "memcpy", I->arg_begin(), --I->arg_end(),
118 I->arg_begin()->getType());
120 case Intrinsic::memmove_i32:
121 case Intrinsic::memmove_i64:
122 EnsureFunctionExists(M, "memmove", I->arg_begin(), --I->arg_end(),
123 I->arg_begin()->getType());
125 case Intrinsic::memset_i32:
126 case Intrinsic::memset_i64:
127 M.getOrInsertFunction("memset", PointerType::get(Type::SByteTy),
128 PointerType::get(Type::SByteTy),
129 Type::IntTy, (--(--I->arg_end()))->getType(),
132 case Intrinsic::isunordered_f32:
133 case Intrinsic::isunordered_f64:
134 EnsureFunctionExists(M, "isunordered", I->arg_begin(), I->arg_end(),
137 case Intrinsic::sqrt_f32:
138 case Intrinsic::sqrt_f64:
139 if(I->arg_begin()->getType() == Type::FloatTy)
140 EnsureFunctionExists(M, "sqrtf", I->arg_begin(), I->arg_end(),
143 EnsureFunctionExists(M, "sqrt", I->arg_begin(), I->arg_end(),
149 /// LowerBSWAP - Emit the code to lower bswap of V before the specified
151 static Value *LowerBSWAP(Value *V, Instruction *IP) {
152 assert(V->getType()->isInteger() && "Can't bswap a non-integer type!");
154 unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
157 default: assert(0 && "Unhandled type size of value to byteswap!");
159 Value *Tmp1 = new ShiftInst(Instruction::Shl, V,
160 ConstantInt::get(Type::UByteTy,8),"bswap.2",IP);
161 Value *Tmp2 = new ShiftInst(Instruction::LShr, V,
162 ConstantInt::get(Type::UByteTy,8),"bswap.1",IP);
163 V = BinaryOperator::createOr(Tmp1, Tmp2, "bswap.i16", IP);
167 Value *Tmp4 = new ShiftInst(Instruction::Shl, V,
168 ConstantInt::get(Type::UByteTy,24),"bswap.4", IP);
169 Value *Tmp3 = new ShiftInst(Instruction::Shl, V,
170 ConstantInt::get(Type::UByteTy,8),"bswap.3",IP);
171 Value *Tmp2 = new ShiftInst(Instruction::LShr, V,
172 ConstantInt::get(Type::UByteTy,8),"bswap.2",IP);
173 Value *Tmp1 = new ShiftInst(Instruction::LShr, V,
174 ConstantInt::get(Type::UByteTy,24),"bswap.1", IP);
175 Tmp3 = BinaryOperator::createAnd(Tmp3,
176 ConstantInt::get(Type::UIntTy, 0xFF0000),
178 Tmp2 = BinaryOperator::createAnd(Tmp2,
179 ConstantInt::get(Type::UIntTy, 0xFF00),
181 Tmp4 = BinaryOperator::createOr(Tmp4, Tmp3, "bswap.or1", IP);
182 Tmp2 = BinaryOperator::createOr(Tmp2, Tmp1, "bswap.or2", IP);
183 V = BinaryOperator::createOr(Tmp4, Tmp3, "bswap.i32", IP);
187 Value *Tmp8 = new ShiftInst(Instruction::Shl, V,
188 ConstantInt::get(Type::UByteTy,56),"bswap.8", IP);
189 Value *Tmp7 = new ShiftInst(Instruction::Shl, V,
190 ConstantInt::get(Type::UByteTy,40),"bswap.7", IP);
191 Value *Tmp6 = new ShiftInst(Instruction::Shl, V,
192 ConstantInt::get(Type::UByteTy,24),"bswap.6", IP);
193 Value *Tmp5 = new ShiftInst(Instruction::Shl, V,
194 ConstantInt::get(Type::UByteTy,8),"bswap.5", IP);
195 Value* Tmp4 = new ShiftInst(Instruction::LShr, V,
196 ConstantInt::get(Type::UByteTy,8),"bswap.4", IP);
197 Value* Tmp3 = new ShiftInst(Instruction::LShr, V,
198 ConstantInt::get(Type::UByteTy,24),"bswap.3", IP);
199 Value* Tmp2 = new ShiftInst(Instruction::LShr, V,
200 ConstantInt::get(Type::UByteTy,40),"bswap.2", IP);
201 Value* Tmp1 = new ShiftInst(Instruction::LShr, V,
202 ConstantInt::get(Type::UByteTy,56),"bswap.1", IP);
203 Tmp7 = BinaryOperator::createAnd(Tmp7,
204 ConstantInt::get(Type::ULongTy,
205 0xFF000000000000ULL),
207 Tmp6 = BinaryOperator::createAnd(Tmp6,
208 ConstantInt::get(Type::ULongTy, 0xFF0000000000ULL),
210 Tmp5 = BinaryOperator::createAnd(Tmp5,
211 ConstantInt::get(Type::ULongTy, 0xFF00000000ULL),
213 Tmp4 = BinaryOperator::createAnd(Tmp4,
214 ConstantInt::get(Type::ULongTy, 0xFF000000ULL),
216 Tmp3 = BinaryOperator::createAnd(Tmp3,
217 ConstantInt::get(Type::ULongTy, 0xFF0000ULL),
219 Tmp2 = BinaryOperator::createAnd(Tmp2,
220 ConstantInt::get(Type::ULongTy, 0xFF00ULL),
222 Tmp8 = BinaryOperator::createOr(Tmp8, Tmp7, "bswap.or1", IP);
223 Tmp6 = BinaryOperator::createOr(Tmp6, Tmp5, "bswap.or2", IP);
224 Tmp4 = BinaryOperator::createOr(Tmp4, Tmp3, "bswap.or3", IP);
225 Tmp2 = BinaryOperator::createOr(Tmp2, Tmp1, "bswap.or4", IP);
226 Tmp8 = BinaryOperator::createOr(Tmp8, Tmp6, "bswap.or5", IP);
227 Tmp4 = BinaryOperator::createOr(Tmp4, Tmp2, "bswap.or6", IP);
228 V = BinaryOperator::createOr(Tmp8, Tmp4, "bswap.i64", IP);
235 /// LowerCTPOP - Emit the code to lower ctpop of V before the specified
237 static Value *LowerCTPOP(Value *V, Instruction *IP) {
238 assert(V->getType()->isInteger() && "Can't ctpop a non-integer type!");
240 static const uint64_t MaskValues[6] = {
241 0x5555555555555555ULL, 0x3333333333333333ULL,
242 0x0F0F0F0F0F0F0F0FULL, 0x00FF00FF00FF00FFULL,
243 0x0000FFFF0000FFFFULL, 0x00000000FFFFFFFFULL
246 unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
248 for (unsigned i = 1, ct = 0; i != BitSize; i <<= 1, ++ct) {
249 Value *MaskCst = ConstantInt::get(V->getType(), MaskValues[ct]);
250 Value *LHS = BinaryOperator::createAnd(V, MaskCst, "cppop.and1", IP);
251 Value *VShift = new ShiftInst(Instruction::LShr, V,
252 ConstantInt::get(Type::UByteTy, i), "ctpop.sh", IP);
253 Value *RHS = BinaryOperator::createAnd(VShift, MaskCst, "cppop.and2", IP);
254 V = BinaryOperator::createAdd(LHS, RHS, "ctpop.step", IP);
260 /// LowerCTLZ - Emit the code to lower ctlz of V before the specified
262 static Value *LowerCTLZ(Value *V, Instruction *IP) {
264 unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
265 for (unsigned i = 1; i != BitSize; i <<= 1) {
266 Value *ShVal = ConstantInt::get(Type::UByteTy, i);
267 ShVal = new ShiftInst(Instruction::LShr, V, ShVal, "ctlz.sh", IP);
268 V = BinaryOperator::createOr(V, ShVal, "ctlz.step", IP);
271 V = BinaryOperator::createNot(V, "", IP);
272 return LowerCTPOP(V, IP);
277 void IntrinsicLowering::LowerIntrinsicCall(CallInst *CI) {
278 Function *Callee = CI->getCalledFunction();
279 assert(Callee && "Cannot lower an indirect call!");
281 switch (Callee->getIntrinsicID()) {
282 case Intrinsic::not_intrinsic:
283 cerr << "Cannot lower a call to a non-intrinsic function '"
284 << Callee->getName() << "'!\n";
287 cerr << "Error: Code generator does not support intrinsic function '"
288 << Callee->getName() << "'!\n";
291 // The setjmp/longjmp intrinsics should only exist in the code if it was
292 // never optimized (ie, right out of the CFE), or if it has been hacked on
293 // by the lowerinvoke pass. In both cases, the right thing to do is to
294 // convert the call to an explicit setjmp or longjmp call.
295 case Intrinsic::setjmp: {
296 static Function *SetjmpFCache = 0;
297 static const unsigned castOpcodes[] = { Instruction::BitCast };
298 Value *V = ReplaceCallWith("setjmp", CI, CI->op_begin()+1, CI->op_end(),
299 castOpcodes, Type::IntTy, SetjmpFCache);
300 if (CI->getType() != Type::VoidTy)
301 CI->replaceAllUsesWith(V);
304 case Intrinsic::sigsetjmp:
305 if (CI->getType() != Type::VoidTy)
306 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
309 case Intrinsic::longjmp: {
310 static Function *LongjmpFCache = 0;
311 static const unsigned castOpcodes[] =
312 { Instruction::BitCast, 0 };
313 ReplaceCallWith("longjmp", CI, CI->op_begin()+1, CI->op_end(),
314 castOpcodes, Type::VoidTy, LongjmpFCache);
318 case Intrinsic::siglongjmp: {
319 // Insert the call to abort
320 static Function *AbortFCache = 0;
321 static const unsigned castOpcodes[] =
322 { Instruction::BitCast, 0 };
323 ReplaceCallWith("abort", CI, CI->op_end(), CI->op_end(),
324 castOpcodes, Type::VoidTy, AbortFCache);
327 case Intrinsic::ctpop_i8:
328 case Intrinsic::ctpop_i16:
329 case Intrinsic::ctpop_i32:
330 case Intrinsic::ctpop_i64:
331 CI->replaceAllUsesWith(LowerCTPOP(CI->getOperand(1), CI));
334 case Intrinsic::bswap_i16:
335 case Intrinsic::bswap_i32:
336 case Intrinsic::bswap_i64:
337 CI->replaceAllUsesWith(LowerBSWAP(CI->getOperand(1), CI));
340 case Intrinsic::ctlz_i8:
341 case Intrinsic::ctlz_i16:
342 case Intrinsic::ctlz_i32:
343 case Intrinsic::ctlz_i64:
344 CI->replaceAllUsesWith(LowerCTLZ(CI->getOperand(1), CI));
347 case Intrinsic::cttz_i8:
348 case Intrinsic::cttz_i16:
349 case Intrinsic::cttz_i32:
350 case Intrinsic::cttz_i64: {
351 // cttz(x) -> ctpop(~X & (X-1))
352 Value *Src = CI->getOperand(1);
353 Value *NotSrc = BinaryOperator::createNot(Src, Src->getName()+".not", CI);
354 Value *SrcM1 = ConstantInt::get(Src->getType(), 1);
355 SrcM1 = BinaryOperator::createSub(Src, SrcM1, "", CI);
356 Src = LowerCTPOP(BinaryOperator::createAnd(NotSrc, SrcM1, "", CI), CI);
357 CI->replaceAllUsesWith(Src);
361 case Intrinsic::stacksave:
362 case Intrinsic::stackrestore: {
363 static bool Warned = false;
365 cerr << "WARNING: this target does not support the llvm.stack"
366 << (Callee->getIntrinsicID() == Intrinsic::stacksave ?
367 "save" : "restore") << " intrinsic.\n";
369 if (Callee->getIntrinsicID() == Intrinsic::stacksave)
370 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
374 case Intrinsic::returnaddress:
375 case Intrinsic::frameaddress:
376 cerr << "WARNING: this target does not support the llvm."
377 << (Callee->getIntrinsicID() == Intrinsic::returnaddress ?
378 "return" : "frame") << "address intrinsic.\n";
379 CI->replaceAllUsesWith(ConstantPointerNull::get(
380 cast<PointerType>(CI->getType())));
383 case Intrinsic::prefetch:
384 break; // Simply strip out prefetches on unsupported architectures
386 case Intrinsic::pcmarker:
387 break; // Simply strip out pcmarker on unsupported architectures
388 case Intrinsic::readcyclecounter: {
389 cerr << "WARNING: this target does not support the llvm.readcyclecoun"
390 << "ter intrinsic. It is being lowered to a constant 0\n";
391 CI->replaceAllUsesWith(ConstantInt::get(Type::ULongTy, 0));
395 case Intrinsic::dbg_stoppoint:
396 case Intrinsic::dbg_region_start:
397 case Intrinsic::dbg_region_end:
398 case Intrinsic::dbg_func_start:
399 case Intrinsic::dbg_declare:
400 break; // Simply strip out debugging intrinsics
402 case Intrinsic::memcpy_i32: {
403 // The memcpy intrinsic take an extra alignment argument that the memcpy
404 // libc function does not.
405 static unsigned opcodes[] =
406 { Instruction::BitCast, Instruction::BitCast, Instruction::BitCast };
408 // if (target_is_64_bit) opcodes[2] = Instruction::ZExt;
409 // else opcodes[2] = Instruction::BitCast;
410 static Function *MemcpyFCache = 0;
411 ReplaceCallWith("memcpy", CI, CI->op_begin()+1, CI->op_end()-1,
412 opcodes, (*(CI->op_begin()+1))->getType(), MemcpyFCache);
415 case Intrinsic::memcpy_i64: {
416 static unsigned opcodes[] =
417 { Instruction::BitCast, Instruction::BitCast, Instruction::Trunc };
419 // if (target_is_64_bit) opcodes[2] = Instruction::BitCast;
420 // else opcodes[2] = Instruction::Trunc;
421 static Function *MemcpyFCache = 0;
422 ReplaceCallWith("memcpy", CI, CI->op_begin()+1, CI->op_end()-1,
423 opcodes, (*(CI->op_begin()+1))->getType(), MemcpyFCache);
426 case Intrinsic::memmove_i32: {
427 // The memmove intrinsic take an extra alignment argument that the memmove
428 // libc function does not.
429 static unsigned opcodes[] =
430 { Instruction::BitCast, Instruction::BitCast, Instruction::BitCast };
432 // if (target_is_64_bit) opcodes[2] = Instruction::ZExt;
433 // else opcodes[2] = Instruction::BitCast;
434 static Function *MemmoveFCache = 0;
435 ReplaceCallWith("memmove", CI, CI->op_begin()+1, CI->op_end()-1,
436 opcodes, (*(CI->op_begin()+1))->getType(), MemmoveFCache);
439 case Intrinsic::memmove_i64: {
440 // The memmove intrinsic take an extra alignment argument that the memmove
441 // libc function does not.
442 static const unsigned opcodes[] =
443 { Instruction::BitCast, Instruction::BitCast, Instruction::Trunc };
444 // if (target_is_64_bit) opcodes[2] = Instruction::BitCast;
445 // else opcodes[2] = Instruction::Trunc;
446 static Function *MemmoveFCache = 0;
447 ReplaceCallWith("memmove", CI, CI->op_begin()+1, CI->op_end()-1,
448 opcodes, (*(CI->op_begin()+1))->getType(), MemmoveFCache);
451 case Intrinsic::memset_i32: {
452 // The memset intrinsic take an extra alignment argument that the memset
453 // libc function does not.
454 static const unsigned opcodes[] =
455 { Instruction::BitCast, Instruction::ZExt, Instruction::ZExt, 0 };
456 // if (target_is_64_bit) opcodes[2] = Instruction::BitCast;
457 // else opcodes[2] = Instruction::ZExt;
458 static Function *MemsetFCache = 0;
459 ReplaceCallWith("memset", CI, CI->op_begin()+1, CI->op_end()-1,
460 opcodes, (*(CI->op_begin()+1))->getType(), MemsetFCache);
462 case Intrinsic::memset_i64: {
463 // The memset intrinsic take an extra alignment argument that the memset
464 // libc function does not.
465 static const unsigned opcodes[] =
466 { Instruction::BitCast, Instruction::ZExt, Instruction::Trunc, 0 };
467 // if (target_is_64_bit) opcodes[2] = Instruction::BitCast;
468 // else opcodes[2] = Instruction::Trunc;
469 static Function *MemsetFCache = 0;
470 ReplaceCallWith("memset", CI, CI->op_begin()+1, CI->op_end()-1,
471 opcodes, (*(CI->op_begin()+1))->getType(), MemsetFCache);
474 case Intrinsic::isunordered_f32:
475 case Intrinsic::isunordered_f64: {
476 Value *L = CI->getOperand(1);
477 Value *R = CI->getOperand(2);
479 Value *LIsNan = new SetCondInst(Instruction::SetNE, L, L, "LIsNan", CI);
480 Value *RIsNan = new SetCondInst(Instruction::SetNE, R, R, "RIsNan", CI);
481 CI->replaceAllUsesWith(
482 BinaryOperator::create(Instruction::Or, LIsNan, RIsNan,
486 case Intrinsic::sqrt_f32: {
487 static const unsigned opcodes[] = { 0 };
488 static Function *sqrtfFCache = 0;
489 ReplaceCallWith("sqrtf", CI, CI->op_begin()+1, CI->op_end(),
490 opcodes, Type::FloatTy, sqrtfFCache);
493 case Intrinsic::sqrt_f64: {
494 static const unsigned opcodes[] = { 0 };
495 static Function *sqrtFCache = 0;
496 ReplaceCallWith("sqrt", CI, CI->op_begin()+1, CI->op_end(),
497 opcodes, Type::DoubleTy, sqrtFCache);
502 assert(CI->use_empty() &&
503 "Lowering should have eliminated any uses of the intrinsic call!");
504 CI->eraseFromParent();