1 //===-- IntrinsicLowering.cpp - Intrinsic Lowering default implementation -===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the IntrinsicLowering class.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Constants.h"
15 #include "llvm/DerivedTypes.h"
16 #include "llvm/Module.h"
17 #include "llvm/Type.h"
18 #include "llvm/CodeGen/IntrinsicLowering.h"
19 #include "llvm/Support/IRBuilder.h"
20 #include "llvm/Support/ErrorHandling.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 static void EnsureFPIntrinsicsExist(Module &M, Function *Fn,
38 const char *DName, const char *LDName) {
39 // Insert definitions for all the floating point types.
40 switch((int)Fn->arg_begin()->getType()->getTypeID()) {
42 EnsureFunctionExists(M, FName, Fn->arg_begin(), Fn->arg_end(),
45 case Type::DoubleTyID:
46 EnsureFunctionExists(M, DName, Fn->arg_begin(), Fn->arg_end(),
49 case Type::X86_FP80TyID:
51 case Type::PPC_FP128TyID:
52 EnsureFunctionExists(M, LDName, Fn->arg_begin(), Fn->arg_end(),
53 Fn->arg_begin()->getType());
58 /// ReplaceCallWith - This function is used when we want to lower an intrinsic
59 /// call to a call of an external function. This handles hard cases such as
60 /// when there was already a prototype for the external function, and if that
61 /// prototype doesn't match the arguments we expect to pass in.
62 template <class ArgIt>
63 static CallInst *ReplaceCallWith(const char *NewFn, CallInst *CI,
64 ArgIt ArgBegin, ArgIt ArgEnd,
66 // If we haven't already looked up this function, check to see if the
67 // program already contains a function with this name.
68 Module *M = CI->getParent()->getParent()->getParent();
69 // Get or insert the definition now.
70 std::vector<const Type *> ParamTys;
71 for (ArgIt I = ArgBegin; I != ArgEnd; ++I)
72 ParamTys.push_back((*I)->getType());
73 Constant* FCache = M->getOrInsertFunction(NewFn,
74 FunctionType::get(RetTy, ParamTys, false));
76 IRBuilder<> Builder(CI->getParent(), CI);
77 SmallVector<Value *, 8> Args(ArgBegin, ArgEnd);
78 CallInst *NewCI = Builder.CreateCall(FCache, Args.begin(), Args.end());
79 NewCI->setName(CI->getName());
81 CI->replaceAllUsesWith(NewCI);
85 void IntrinsicLowering::AddPrototypes(Module &M) {
86 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
87 if (I->isDeclaration() && !I->use_empty())
88 switch (I->getIntrinsicID()) {
90 case Intrinsic::setjmp:
91 EnsureFunctionExists(M, "setjmp", I->arg_begin(), I->arg_end(),
94 case Intrinsic::longjmp:
95 EnsureFunctionExists(M, "longjmp", I->arg_begin(), I->arg_end(),
98 case Intrinsic::siglongjmp:
99 EnsureFunctionExists(M, "abort", I->arg_end(), I->arg_end(),
102 case Intrinsic::memcpy:
103 M.getOrInsertFunction("memcpy", PointerType::getUnqual(Type::Int8Ty),
104 PointerType::getUnqual(Type::Int8Ty),
105 PointerType::getUnqual(Type::Int8Ty),
106 TD.getIntPtrType(), (Type *)0);
108 case Intrinsic::memmove:
109 M.getOrInsertFunction("memmove", PointerType::getUnqual(Type::Int8Ty),
110 PointerType::getUnqual(Type::Int8Ty),
111 PointerType::getUnqual(Type::Int8Ty),
112 TD.getIntPtrType(), (Type *)0);
114 case Intrinsic::memset:
115 M.getOrInsertFunction("memset", PointerType::getUnqual(Type::Int8Ty),
116 PointerType::getUnqual(Type::Int8Ty),
118 TD.getIntPtrType(), (Type *)0);
120 case Intrinsic::sqrt:
121 EnsureFPIntrinsicsExist(M, I, "sqrtf", "sqrt", "sqrtl");
124 EnsureFPIntrinsicsExist(M, I, "sinf", "sin", "sinl");
127 EnsureFPIntrinsicsExist(M, I, "cosf", "cos", "cosl");
130 EnsureFPIntrinsicsExist(M, I, "powf", "pow", "powl");
133 EnsureFPIntrinsicsExist(M, I, "logf", "log", "logl");
135 case Intrinsic::log2:
136 EnsureFPIntrinsicsExist(M, I, "log2f", "log2", "log2l");
138 case Intrinsic::log10:
139 EnsureFPIntrinsicsExist(M, I, "log10f", "log10", "log10l");
142 EnsureFPIntrinsicsExist(M, I, "expf", "exp", "expl");
144 case Intrinsic::exp2:
145 EnsureFPIntrinsicsExist(M, I, "exp2f", "exp2", "exp2l");
150 /// LowerBSWAP - Emit the code to lower bswap of V before the specified
152 static Value *LowerBSWAP(Value *V, Instruction *IP) {
153 assert(V->getType()->isInteger() && "Can't bswap a non-integer type!");
155 unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
157 IRBuilder<> Builder(IP->getParent(), IP);
160 default: llvm_unreachable("Unhandled type size of value to byteswap!");
162 Value *Tmp1 = Builder.CreateShl(V, ConstantInt::get(V->getType(), 8),
164 Value *Tmp2 = Builder.CreateLShr(V, ConstantInt::get(V->getType(), 8),
166 V = Builder.CreateOr(Tmp1, Tmp2, "bswap.i16");
170 Value *Tmp4 = Builder.CreateShl(V, ConstantInt::get(V->getType(), 24),
172 Value *Tmp3 = Builder.CreateShl(V, ConstantInt::get(V->getType(), 8),
174 Value *Tmp2 = Builder.CreateLShr(V, ConstantInt::get(V->getType(), 8),
176 Value *Tmp1 = Builder.CreateLShr(V, ConstantInt::get(V->getType(), 24),
178 Tmp3 = Builder.CreateAnd(Tmp3, ConstantInt::get(Type::Int32Ty, 0xFF0000),
180 Tmp2 = Builder.CreateAnd(Tmp2, ConstantInt::get(Type::Int32Ty, 0xFF00),
182 Tmp4 = Builder.CreateOr(Tmp4, Tmp3, "bswap.or1");
183 Tmp2 = Builder.CreateOr(Tmp2, Tmp1, "bswap.or2");
184 V = Builder.CreateOr(Tmp4, Tmp2, "bswap.i32");
188 Value *Tmp8 = Builder.CreateShl(V, ConstantInt::get(V->getType(), 56),
190 Value *Tmp7 = Builder.CreateShl(V, ConstantInt::get(V->getType(), 40),
192 Value *Tmp6 = Builder.CreateShl(V, ConstantInt::get(V->getType(), 24),
194 Value *Tmp5 = Builder.CreateShl(V, ConstantInt::get(V->getType(), 8),
196 Value* Tmp4 = Builder.CreateLShr(V, ConstantInt::get(V->getType(), 8),
198 Value* Tmp3 = Builder.CreateLShr(V, ConstantInt::get(V->getType(), 24),
200 Value* Tmp2 = Builder.CreateLShr(V, ConstantInt::get(V->getType(), 40),
202 Value* Tmp1 = Builder.CreateLShr(V, ConstantInt::get(V->getType(), 56),
204 Tmp7 = Builder.CreateAnd(Tmp7,
205 ConstantInt::get(Type::Int64Ty,
206 0xFF000000000000ULL),
208 Tmp6 = Builder.CreateAnd(Tmp6,
209 ConstantInt::get(Type::Int64Ty,
212 Tmp5 = Builder.CreateAnd(Tmp5,
213 ConstantInt::get(Type::Int64Ty, 0xFF00000000ULL),
215 Tmp4 = Builder.CreateAnd(Tmp4,
216 ConstantInt::get(Type::Int64Ty, 0xFF000000ULL),
218 Tmp3 = Builder.CreateAnd(Tmp3,
219 ConstantInt::get(Type::Int64Ty, 0xFF0000ULL),
221 Tmp2 = Builder.CreateAnd(Tmp2,
222 ConstantInt::get(Type::Int64Ty, 0xFF00ULL),
224 Tmp8 = Builder.CreateOr(Tmp8, Tmp7, "bswap.or1");
225 Tmp6 = Builder.CreateOr(Tmp6, Tmp5, "bswap.or2");
226 Tmp4 = Builder.CreateOr(Tmp4, Tmp3, "bswap.or3");
227 Tmp2 = Builder.CreateOr(Tmp2, Tmp1, "bswap.or4");
228 Tmp8 = Builder.CreateOr(Tmp8, Tmp6, "bswap.or5");
229 Tmp4 = Builder.CreateOr(Tmp4, Tmp2, "bswap.or6");
230 V = Builder.CreateOr(Tmp8, Tmp4, "bswap.i64");
237 /// LowerCTPOP - Emit the code to lower ctpop of V before the specified
239 static Value *LowerCTPOP(Value *V, Instruction *IP) {
240 assert(V->getType()->isInteger() && "Can't ctpop a non-integer type!");
242 static const uint64_t MaskValues[6] = {
243 0x5555555555555555ULL, 0x3333333333333333ULL,
244 0x0F0F0F0F0F0F0F0FULL, 0x00FF00FF00FF00FFULL,
245 0x0000FFFF0000FFFFULL, 0x00000000FFFFFFFFULL
248 IRBuilder<> Builder(IP->getParent(), IP);
250 unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
251 unsigned WordSize = (BitSize + 63) / 64;
252 Value *Count = ConstantInt::get(V->getType(), 0);
254 for (unsigned n = 0; n < WordSize; ++n) {
255 Value *PartValue = V;
256 for (unsigned i = 1, ct = 0; i < (BitSize>64 ? 64 : BitSize);
258 Value *MaskCst = ConstantInt::get(V->getType(), MaskValues[ct]);
259 Value *LHS = Builder.CreateAnd(PartValue, MaskCst, "cppop.and1");
260 Value *VShift = Builder.CreateLShr(PartValue,
261 ConstantInt::get(V->getType(), i),
263 Value *RHS = Builder.CreateAnd(VShift, MaskCst, "cppop.and2");
264 PartValue = Builder.CreateAdd(LHS, RHS, "ctpop.step");
266 Count = Builder.CreateAdd(PartValue, Count, "ctpop.part");
268 V = Builder.CreateLShr(V, ConstantInt::get(V->getType(), 64),
277 /// LowerCTLZ - Emit the code to lower ctlz of V before the specified
279 static Value *LowerCTLZ(Value *V, Instruction *IP) {
281 IRBuilder<> Builder(IP->getParent(), IP);
283 unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
284 for (unsigned i = 1; i < BitSize; i <<= 1) {
285 Value *ShVal = ConstantInt::get(V->getType(), i);
286 ShVal = Builder.CreateLShr(V, ShVal, "ctlz.sh");
287 V = Builder.CreateOr(V, ShVal, "ctlz.step");
290 V = Builder.CreateNot(V);
291 return LowerCTPOP(V, IP);
294 static void ReplaceFPIntrinsicWithCall(CallInst *CI, const char *Fname,
296 const char *LDname) {
297 switch (CI->getOperand(1)->getType()->getTypeID()) {
298 default: llvm_unreachable("Invalid type in intrinsic");
299 case Type::FloatTyID:
300 ReplaceCallWith(Fname, CI, CI->op_begin() + 1, CI->op_end(),
303 case Type::DoubleTyID:
304 ReplaceCallWith(Dname, CI, CI->op_begin() + 1, CI->op_end(),
307 case Type::X86_FP80TyID:
308 case Type::FP128TyID:
309 case Type::PPC_FP128TyID:
310 ReplaceCallWith(LDname, CI, CI->op_begin() + 1, CI->op_end(),
311 CI->getOperand(1)->getType());
316 void IntrinsicLowering::LowerIntrinsicCall(CallInst *CI) {
317 IRBuilder<> Builder(CI->getParent(), CI);
318 LLVMContext *Context = CI->getParent()->getContext();
320 Function *Callee = CI->getCalledFunction();
321 assert(Callee && "Cannot lower an indirect call!");
323 switch (Callee->getIntrinsicID()) {
324 case Intrinsic::not_intrinsic:
325 llvm_report_error("Cannot lower a call to a non-intrinsic function '"+
326 Callee->getName() + "'!");
328 llvm_report_error("Code generator does not support intrinsic function '"+
329 Callee->getName()+"'!");
331 // The setjmp/longjmp intrinsics should only exist in the code if it was
332 // never optimized (ie, right out of the CFE), or if it has been hacked on
333 // by the lowerinvoke pass. In both cases, the right thing to do is to
334 // convert the call to an explicit setjmp or longjmp call.
335 case Intrinsic::setjmp: {
336 Value *V = ReplaceCallWith("setjmp", CI, CI->op_begin() + 1, CI->op_end(),
338 if (CI->getType() != Type::VoidTy)
339 CI->replaceAllUsesWith(V);
342 case Intrinsic::sigsetjmp:
343 if (CI->getType() != Type::VoidTy)
344 CI->replaceAllUsesWith(Context->getNullValue(CI->getType()));
347 case Intrinsic::longjmp: {
348 ReplaceCallWith("longjmp", CI, CI->op_begin() + 1, CI->op_end(),
353 case Intrinsic::siglongjmp: {
354 // Insert the call to abort
355 ReplaceCallWith("abort", CI, CI->op_end(), CI->op_end(),
359 case Intrinsic::ctpop:
360 CI->replaceAllUsesWith(LowerCTPOP(CI->getOperand(1), CI));
363 case Intrinsic::bswap:
364 CI->replaceAllUsesWith(LowerBSWAP(CI->getOperand(1), CI));
367 case Intrinsic::ctlz:
368 CI->replaceAllUsesWith(LowerCTLZ(CI->getOperand(1), CI));
371 case Intrinsic::cttz: {
372 // cttz(x) -> ctpop(~X & (X-1))
373 Value *Src = CI->getOperand(1);
374 Value *NotSrc = Builder.CreateNot(Src);
375 NotSrc->setName(Src->getName() + ".not");
376 Value *SrcM1 = ConstantInt::get(Src->getType(), 1);
377 SrcM1 = Builder.CreateSub(Src, SrcM1);
378 Src = LowerCTPOP(Builder.CreateAnd(NotSrc, SrcM1), CI);
379 CI->replaceAllUsesWith(Src);
383 case Intrinsic::stacksave:
384 case Intrinsic::stackrestore: {
386 cerr << "WARNING: this target does not support the llvm.stack"
387 << (Callee->getIntrinsicID() == Intrinsic::stacksave ?
388 "save" : "restore") << " intrinsic.\n";
390 if (Callee->getIntrinsicID() == Intrinsic::stacksave)
391 CI->replaceAllUsesWith(Context->getNullValue(CI->getType()));
395 case Intrinsic::returnaddress:
396 case Intrinsic::frameaddress:
397 cerr << "WARNING: this target does not support the llvm."
398 << (Callee->getIntrinsicID() == Intrinsic::returnaddress ?
399 "return" : "frame") << "address intrinsic.\n";
400 CI->replaceAllUsesWith(ConstantPointerNull::get(
401 cast<PointerType>(CI->getType())));
404 case Intrinsic::prefetch:
405 break; // Simply strip out prefetches on unsupported architectures
407 case Intrinsic::pcmarker:
408 break; // Simply strip out pcmarker on unsupported architectures
409 case Intrinsic::readcyclecounter: {
410 cerr << "WARNING: this target does not support the llvm.readcyclecoun"
411 << "ter intrinsic. It is being lowered to a constant 0\n";
412 CI->replaceAllUsesWith(ConstantInt::get(Type::Int64Ty, 0));
416 case Intrinsic::dbg_stoppoint:
417 case Intrinsic::dbg_region_start:
418 case Intrinsic::dbg_region_end:
419 case Intrinsic::dbg_func_start:
420 case Intrinsic::dbg_declare:
421 break; // Simply strip out debugging intrinsics
423 case Intrinsic::eh_exception:
424 case Intrinsic::eh_selector_i32:
425 case Intrinsic::eh_selector_i64:
426 CI->replaceAllUsesWith(Context->getNullValue(CI->getType()));
429 case Intrinsic::eh_typeid_for_i32:
430 case Intrinsic::eh_typeid_for_i64:
431 // Return something different to eh_selector.
432 CI->replaceAllUsesWith(ConstantInt::get(CI->getType(), 1));
435 case Intrinsic::var_annotation:
436 break; // Strip out annotate intrinsic
438 case Intrinsic::memcpy: {
439 const IntegerType *IntPtr = TD.getIntPtrType();
440 Value *Size = Builder.CreateIntCast(CI->getOperand(3), IntPtr,
441 /* isSigned */ false);
443 Ops[0] = CI->getOperand(1);
444 Ops[1] = CI->getOperand(2);
446 ReplaceCallWith("memcpy", CI, Ops, Ops+3, CI->getOperand(1)->getType());
449 case Intrinsic::memmove: {
450 const IntegerType *IntPtr = TD.getIntPtrType();
451 Value *Size = Builder.CreateIntCast(CI->getOperand(3), IntPtr,
452 /* isSigned */ false);
454 Ops[0] = CI->getOperand(1);
455 Ops[1] = CI->getOperand(2);
457 ReplaceCallWith("memmove", CI, Ops, Ops+3, CI->getOperand(1)->getType());
460 case Intrinsic::memset: {
461 const IntegerType *IntPtr = TD.getIntPtrType();
462 Value *Size = Builder.CreateIntCast(CI->getOperand(3), IntPtr,
463 /* isSigned */ false);
465 Ops[0] = CI->getOperand(1);
466 // Extend the amount to i32.
467 Ops[1] = Builder.CreateIntCast(CI->getOperand(2), Type::Int32Ty,
468 /* isSigned */ false);
470 ReplaceCallWith("memset", CI, Ops, Ops+3, CI->getOperand(1)->getType());
473 case Intrinsic::sqrt: {
474 ReplaceFPIntrinsicWithCall(CI, "sqrtf", "sqrt", "sqrtl");
477 case Intrinsic::log: {
478 ReplaceFPIntrinsicWithCall(CI, "logf", "log", "logl");
481 case Intrinsic::log2: {
482 ReplaceFPIntrinsicWithCall(CI, "log2f", "log2", "log2l");
485 case Intrinsic::log10: {
486 ReplaceFPIntrinsicWithCall(CI, "log10f", "log10", "log10l");
489 case Intrinsic::exp: {
490 ReplaceFPIntrinsicWithCall(CI, "expf", "exp", "expl");
493 case Intrinsic::exp2: {
494 ReplaceFPIntrinsicWithCall(CI, "exp2f", "exp2", "exp2l");
497 case Intrinsic::pow: {
498 ReplaceFPIntrinsicWithCall(CI, "powf", "pow", "powl");
501 case Intrinsic::flt_rounds:
502 // Lower to "round to the nearest"
503 if (CI->getType() != Type::VoidTy)
504 CI->replaceAllUsesWith(ConstantInt::get(CI->getType(), 1));
508 assert(CI->use_empty() &&
509 "Lowering should have eliminated any uses of the intrinsic call!");
510 CI->eraseFromParent();