clang  9.0.0svn
CGBuiltin.cpp
Go to the documentation of this file.
1 //===---- CGBuiltin.cpp - Emit LLVM Code for builtins ---------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This contains code to emit Builtin calls as LLVM code.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "CGCXXABI.h"
14 #include "CGObjCRuntime.h"
15 #include "CGOpenCLRuntime.h"
16 #include "CGRecordLayout.h"
17 #include "CodeGenFunction.h"
18 #include "CodeGenModule.h"
19 #include "ConstantEmitter.h"
20 #include "PatternInit.h"
21 #include "TargetInfo.h"
22 #include "clang/AST/ASTContext.h"
23 #include "clang/AST/Decl.h"
24 #include "clang/AST/OSLog.h"
26 #include "clang/Basic/TargetInfo.h"
28 #include "llvm/ADT/SmallPtrSet.h"
29 #include "llvm/ADT/StringExtras.h"
30 #include "llvm/IR/DataLayout.h"
31 #include "llvm/IR/InlineAsm.h"
32 #include "llvm/IR/Intrinsics.h"
33 #include "llvm/IR/MDBuilder.h"
34 #include "llvm/Support/ConvertUTF.h"
35 #include "llvm/Support/ScopedPrinter.h"
36 #include "llvm/Support/TargetParser.h"
37 #include <sstream>
38 
39 using namespace clang;
40 using namespace CodeGen;
41 using namespace llvm;
42 
43 static
44 int64_t clamp(int64_t Value, int64_t Low, int64_t High) {
45  return std::min(High, std::max(Low, Value));
46 }
47 
48 static void initializeAlloca(CodeGenFunction &CGF, AllocaInst *AI, Value *Size, unsigned AlignmentInBytes) {
49  ConstantInt *Byte;
50  switch (CGF.getLangOpts().getTrivialAutoVarInit()) {
52  // Nothing to initialize.
53  return;
55  Byte = CGF.Builder.getInt8(0x00);
56  break;
58  llvm::Type *Int8 = llvm::IntegerType::getInt8Ty(CGF.CGM.getLLVMContext());
59  Byte = llvm::dyn_cast<llvm::ConstantInt>(
60  initializationPatternFor(CGF.CGM, Int8));
61  break;
62  }
63  }
64  CGF.Builder.CreateMemSet(AI, Byte, Size, AlignmentInBytes);
65 }
66 
67 /// getBuiltinLibFunction - Given a builtin id for a function like
68 /// "__builtin_fabsf", return a Function* for "fabsf".
70  unsigned BuiltinID) {
71  assert(Context.BuiltinInfo.isLibFunction(BuiltinID));
72 
73  // Get the name, skip over the __builtin_ prefix (if necessary).
74  StringRef Name;
75  GlobalDecl D(FD);
76 
77  // If the builtin has been declared explicitly with an assembler label,
78  // use the mangled name. This differs from the plain label on platforms
79  // that prefix labels.
80  if (FD->hasAttr<AsmLabelAttr>())
81  Name = getMangledName(D);
82  else
83  Name = Context.BuiltinInfo.getName(BuiltinID) + 10;
84 
85  llvm::FunctionType *Ty =
86  cast<llvm::FunctionType>(getTypes().ConvertType(FD->getType()));
87 
88  return GetOrCreateLLVMFunction(Name, Ty, D, /*ForVTable=*/false);
89 }
90 
91 /// Emit the conversions required to turn the given value into an
92 /// integer of the given size.
94  QualType T, llvm::IntegerType *IntType) {
95  V = CGF.EmitToMemory(V, T);
96 
97  if (V->getType()->isPointerTy())
98  return CGF.Builder.CreatePtrToInt(V, IntType);
99 
100  assert(V->getType() == IntType);
101  return V;
102 }
103 
105  QualType T, llvm::Type *ResultType) {
106  V = CGF.EmitFromMemory(V, T);
107 
108  if (ResultType->isPointerTy())
109  return CGF.Builder.CreateIntToPtr(V, ResultType);
110 
111  assert(V->getType() == ResultType);
112  return V;
113 }
114 
115 /// Utility to insert an atomic instruction based on Intrinsic::ID
116 /// and the expression node.
118  CodeGenFunction &CGF, llvm::AtomicRMWInst::BinOp Kind, const CallExpr *E,
119  AtomicOrdering Ordering = AtomicOrdering::SequentiallyConsistent) {
120  QualType T = E->getType();
121  assert(E->getArg(0)->getType()->isPointerType());
122  assert(CGF.getContext().hasSameUnqualifiedType(T,
123  E->getArg(0)->getType()->getPointeeType()));
124  assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType()));
125 
126  llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0));
127  unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace();
128 
129  llvm::IntegerType *IntType =
130  llvm::IntegerType::get(CGF.getLLVMContext(),
131  CGF.getContext().getTypeSize(T));
132  llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace);
133 
134  llvm::Value *Args[2];
135  Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType);
136  Args[1] = CGF.EmitScalarExpr(E->getArg(1));
137  llvm::Type *ValueType = Args[1]->getType();
138  Args[1] = EmitToInt(CGF, Args[1], T, IntType);
139 
140  llvm::Value *Result = CGF.Builder.CreateAtomicRMW(
141  Kind, Args[0], Args[1], Ordering);
142  return EmitFromInt(CGF, Result, T, ValueType);
143 }
144 
146  Value *Val = CGF.EmitScalarExpr(E->getArg(0));
147  Value *Address = CGF.EmitScalarExpr(E->getArg(1));
148 
149  // Convert the type of the pointer to a pointer to the stored type.
150  Val = CGF.EmitToMemory(Val, E->getArg(0)->getType());
151  Value *BC = CGF.Builder.CreateBitCast(
152  Address, llvm::PointerType::getUnqual(Val->getType()), "cast");
153  LValue LV = CGF.MakeNaturalAlignAddrLValue(BC, E->getArg(0)->getType());
154  LV.setNontemporal(true);
155  CGF.EmitStoreOfScalar(Val, LV, false);
156  return nullptr;
157 }
158 
160  Value *Address = CGF.EmitScalarExpr(E->getArg(0));
161 
162  LValue LV = CGF.MakeNaturalAlignAddrLValue(Address, E->getType());
163  LV.setNontemporal(true);
164  return CGF.EmitLoadOfScalar(LV, E->getExprLoc());
165 }
166 
168  llvm::AtomicRMWInst::BinOp Kind,
169  const CallExpr *E) {
170  return RValue::get(MakeBinaryAtomicValue(CGF, Kind, E));
171 }
172 
173 /// Utility to insert an atomic instruction based Intrinsic::ID and
174 /// the expression node, where the return value is the result of the
175 /// operation.
177  llvm::AtomicRMWInst::BinOp Kind,
178  const CallExpr *E,
179  Instruction::BinaryOps Op,
180  bool Invert = false) {
181  QualType T = E->getType();
182  assert(E->getArg(0)->getType()->isPointerType());
183  assert(CGF.getContext().hasSameUnqualifiedType(T,
184  E->getArg(0)->getType()->getPointeeType()));
185  assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType()));
186 
187  llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0));
188  unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace();
189 
190  llvm::IntegerType *IntType =
191  llvm::IntegerType::get(CGF.getLLVMContext(),
192  CGF.getContext().getTypeSize(T));
193  llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace);
194 
195  llvm::Value *Args[2];
196  Args[1] = CGF.EmitScalarExpr(E->getArg(1));
197  llvm::Type *ValueType = Args[1]->getType();
198  Args[1] = EmitToInt(CGF, Args[1], T, IntType);
199  Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType);
200 
201  llvm::Value *Result = CGF.Builder.CreateAtomicRMW(
202  Kind, Args[0], Args[1], llvm::AtomicOrdering::SequentiallyConsistent);
203  Result = CGF.Builder.CreateBinOp(Op, Result, Args[1]);
204  if (Invert)
205  Result = CGF.Builder.CreateBinOp(llvm::Instruction::Xor, Result,
206  llvm::ConstantInt::get(IntType, -1));
207  Result = EmitFromInt(CGF, Result, T, ValueType);
208  return RValue::get(Result);
209 }
210 
211 /// Utility to insert an atomic cmpxchg instruction.
212 ///
213 /// @param CGF The current codegen function.
214 /// @param E Builtin call expression to convert to cmpxchg.
215 /// arg0 - address to operate on
216 /// arg1 - value to compare with
217 /// arg2 - new value
218 /// @param ReturnBool Specifies whether to return success flag of
219 /// cmpxchg result or the old value.
220 ///
221 /// @returns result of cmpxchg, according to ReturnBool
222 ///
223 /// Note: In order to lower Microsoft's _InterlockedCompareExchange* intrinsics
224 /// invoke the function EmitAtomicCmpXchgForMSIntrin.
226  bool ReturnBool) {
227  QualType T = ReturnBool ? E->getArg(1)->getType() : E->getType();
228  llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0));
229  unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace();
230 
231  llvm::IntegerType *IntType = llvm::IntegerType::get(
232  CGF.getLLVMContext(), CGF.getContext().getTypeSize(T));
233  llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace);
234 
235  Value *Args[3];
236  Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType);
237  Args[1] = CGF.EmitScalarExpr(E->getArg(1));
238  llvm::Type *ValueType = Args[1]->getType();
239  Args[1] = EmitToInt(CGF, Args[1], T, IntType);
240  Args[2] = EmitToInt(CGF, CGF.EmitScalarExpr(E->getArg(2)), T, IntType);
241 
242  Value *Pair = CGF.Builder.CreateAtomicCmpXchg(
243  Args[0], Args[1], Args[2], llvm::AtomicOrdering::SequentiallyConsistent,
244  llvm::AtomicOrdering::SequentiallyConsistent);
245  if (ReturnBool)
246  // Extract boolean success flag and zext it to int.
247  return CGF.Builder.CreateZExt(CGF.Builder.CreateExtractValue(Pair, 1),
248  CGF.ConvertType(E->getType()));
249  else
250  // Extract old value and emit it using the same type as compare value.
251  return EmitFromInt(CGF, CGF.Builder.CreateExtractValue(Pair, 0), T,
252  ValueType);
253 }
254 
255 /// This function should be invoked to emit atomic cmpxchg for Microsoft's
256 /// _InterlockedCompareExchange* intrinsics which have the following signature:
257 /// T _InterlockedCompareExchange(T volatile *Destination,
258 /// T Exchange,
259 /// T Comparand);
260 ///
261 /// Whereas the llvm 'cmpxchg' instruction has the following syntax:
262 /// cmpxchg *Destination, Comparand, Exchange.
263 /// So we need to swap Comparand and Exchange when invoking
264 /// CreateAtomicCmpXchg. That is the reason we could not use the above utility
265 /// function MakeAtomicCmpXchgValue since it expects the arguments to be
266 /// already swapped.
267 
268 static
270  AtomicOrdering SuccessOrdering = AtomicOrdering::SequentiallyConsistent) {
271  assert(E->getArg(0)->getType()->isPointerType());
272  assert(CGF.getContext().hasSameUnqualifiedType(
273  E->getType(), E->getArg(0)->getType()->getPointeeType()));
274  assert(CGF.getContext().hasSameUnqualifiedType(E->getType(),
275  E->getArg(1)->getType()));
276  assert(CGF.getContext().hasSameUnqualifiedType(E->getType(),
277  E->getArg(2)->getType()));
278 
279  auto *Destination = CGF.EmitScalarExpr(E->getArg(0));
280  auto *Comparand = CGF.EmitScalarExpr(E->getArg(2));
281  auto *Exchange = CGF.EmitScalarExpr(E->getArg(1));
282 
283  // For Release ordering, the failure ordering should be Monotonic.
284  auto FailureOrdering = SuccessOrdering == AtomicOrdering::Release ?
285  AtomicOrdering::Monotonic :
286  SuccessOrdering;
287 
288  auto *Result = CGF.Builder.CreateAtomicCmpXchg(
289  Destination, Comparand, Exchange,
290  SuccessOrdering, FailureOrdering);
291  Result->setVolatile(true);
292  return CGF.Builder.CreateExtractValue(Result, 0);
293 }
294 
296  AtomicOrdering Ordering = AtomicOrdering::SequentiallyConsistent) {
297  assert(E->getArg(0)->getType()->isPointerType());
298 
299  auto *IntTy = CGF.ConvertType(E->getType());
300  auto *Result = CGF.Builder.CreateAtomicRMW(
301  AtomicRMWInst::Add,
302  CGF.EmitScalarExpr(E->getArg(0)),
303  ConstantInt::get(IntTy, 1),
304  Ordering);
305  return CGF.Builder.CreateAdd(Result, ConstantInt::get(IntTy, 1));
306 }
307 
309  AtomicOrdering Ordering = AtomicOrdering::SequentiallyConsistent) {
310  assert(E->getArg(0)->getType()->isPointerType());
311 
312  auto *IntTy = CGF.ConvertType(E->getType());
313  auto *Result = CGF.Builder.CreateAtomicRMW(
314  AtomicRMWInst::Sub,
315  CGF.EmitScalarExpr(E->getArg(0)),
316  ConstantInt::get(IntTy, 1),
317  Ordering);
318  return CGF.Builder.CreateSub(Result, ConstantInt::get(IntTy, 1));
319 }
320 
321 // Build a plain volatile load.
323  Value *Ptr = CGF.EmitScalarExpr(E->getArg(0));
324  QualType ElTy = E->getArg(0)->getType()->getPointeeType();
325  CharUnits LoadSize = CGF.getContext().getTypeSizeInChars(ElTy);
326  llvm::Type *ITy =
327  llvm::IntegerType::get(CGF.getLLVMContext(), LoadSize.getQuantity() * 8);
328  Ptr = CGF.Builder.CreateBitCast(Ptr, ITy->getPointerTo());
329  llvm::LoadInst *Load = CGF.Builder.CreateAlignedLoad(Ptr, LoadSize);
330  Load->setVolatile(true);
331  return Load;
332 }
333 
334 // Build a plain volatile store.
336  Value *Ptr = CGF.EmitScalarExpr(E->getArg(0));
337  Value *Value = CGF.EmitScalarExpr(E->getArg(1));
338  QualType ElTy = E->getArg(0)->getType()->getPointeeType();
339  CharUnits StoreSize = CGF.getContext().getTypeSizeInChars(ElTy);
340  llvm::Type *ITy =
341  llvm::IntegerType::get(CGF.getLLVMContext(), StoreSize.getQuantity() * 8);
342  Ptr = CGF.Builder.CreateBitCast(Ptr, ITy->getPointerTo());
343  llvm::StoreInst *Store =
344  CGF.Builder.CreateAlignedStore(Value, Ptr, StoreSize);
345  Store->setVolatile(true);
346  return Store;
347 }
348 
349 // Emit a simple mangled intrinsic that has 1 argument and a return type
350 // matching the argument type.
352  const CallExpr *E,
353  unsigned IntrinsicID) {
354  llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0));
355 
356  Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType());
357  return CGF.Builder.CreateCall(F, Src0);
358 }
359 
360 // Emit an intrinsic that has 2 operands of the same type as its result.
362  const CallExpr *E,
363  unsigned IntrinsicID) {
364  llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0));
365  llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1));
366 
367  Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType());
368  return CGF.Builder.CreateCall(F, { Src0, Src1 });
369 }
370 
371 // Emit an intrinsic that has 3 operands of the same type as its result.
373  const CallExpr *E,
374  unsigned IntrinsicID) {
375  llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0));
376  llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1));
377  llvm::Value *Src2 = CGF.EmitScalarExpr(E->getArg(2));
378 
379  Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType());
380  return CGF.Builder.CreateCall(F, { Src0, Src1, Src2 });
381 }
382 
383 // Emit an intrinsic that has 1 float or double operand, and 1 integer.
385  const CallExpr *E,
386  unsigned IntrinsicID) {
387  llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0));
388  llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1));
389 
390  Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType());
391  return CGF.Builder.CreateCall(F, {Src0, Src1});
392 }
393 
394 // Emit an intrinsic that has overloaded integer result and fp operand.
396  const CallExpr *E,
397  unsigned IntrinsicID) {
398  llvm::Type *ResultType = CGF.ConvertType(E->getType());
399  llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0));
400 
401  Function *F = CGF.CGM.getIntrinsic(IntrinsicID,
402  {ResultType, Src0->getType()});
403  return CGF.Builder.CreateCall(F, Src0);
404 }
405 
406 /// EmitFAbs - Emit a call to @llvm.fabs().
408  Function *F = CGF.CGM.getIntrinsic(Intrinsic::fabs, V->getType());
409  llvm::CallInst *Call = CGF.Builder.CreateCall(F, V);
410  Call->setDoesNotAccessMemory();
411  return Call;
412 }
413 
414 /// Emit the computation of the sign bit for a floating point value. Returns
415 /// the i1 sign bit value.
417  LLVMContext &C = CGF.CGM.getLLVMContext();
418 
419  llvm::Type *Ty = V->getType();
420  int Width = Ty->getPrimitiveSizeInBits();
421  llvm::Type *IntTy = llvm::IntegerType::get(C, Width);
422  V = CGF.Builder.CreateBitCast(V, IntTy);
423  if (Ty->isPPC_FP128Ty()) {
424  // We want the sign bit of the higher-order double. The bitcast we just
425  // did works as if the double-double was stored to memory and then
426  // read as an i128. The "store" will put the higher-order double in the
427  // lower address in both little- and big-Endian modes, but the "load"
428  // will treat those bits as a different part of the i128: the low bits in
429  // little-Endian, the high bits in big-Endian. Therefore, on big-Endian
430  // we need to shift the high bits down to the low before truncating.
431  Width >>= 1;
432  if (CGF.getTarget().isBigEndian()) {
433  Value *ShiftCst = llvm::ConstantInt::get(IntTy, Width);
434  V = CGF.Builder.CreateLShr(V, ShiftCst);
435  }
436  // We are truncating value in order to extract the higher-order
437  // double, which we will be using to extract the sign from.
438  IntTy = llvm::IntegerType::get(C, Width);
439  V = CGF.Builder.CreateTrunc(V, IntTy);
440  }
441  Value *Zero = llvm::Constant::getNullValue(IntTy);
442  return CGF.Builder.CreateICmpSLT(V, Zero);
443 }
444 
446  const CallExpr *E, llvm::Constant *calleeValue) {
447  CGCallee callee = CGCallee::forDirect(calleeValue, GlobalDecl(FD));
448  return CGF.EmitCall(E->getCallee()->getType(), callee, E, ReturnValueSlot());
449 }
450 
451 /// Emit a call to llvm.{sadd,uadd,ssub,usub,smul,umul}.with.overflow.*
452 /// depending on IntrinsicID.
453 ///
454 /// \arg CGF The current codegen function.
455 /// \arg IntrinsicID The ID for the Intrinsic we wish to generate.
456 /// \arg X The first argument to the llvm.*.with.overflow.*.
457 /// \arg Y The second argument to the llvm.*.with.overflow.*.
458 /// \arg Carry The carry returned by the llvm.*.with.overflow.*.
459 /// \returns The result (i.e. sum/product) returned by the intrinsic.
461  const llvm::Intrinsic::ID IntrinsicID,
463  llvm::Value *&Carry) {
464  // Make sure we have integers of the same width.
465  assert(X->getType() == Y->getType() &&
466  "Arguments must be the same type. (Did you forget to make sure both "
467  "arguments have the same integer width?)");
468 
469  Function *Callee = CGF.CGM.getIntrinsic(IntrinsicID, X->getType());
470  llvm::Value *Tmp = CGF.Builder.CreateCall(Callee, {X, Y});
471  Carry = CGF.Builder.CreateExtractValue(Tmp, 1);
472  return CGF.Builder.CreateExtractValue(Tmp, 0);
473 }
474 
476  unsigned IntrinsicID,
477  int low, int high) {
478  llvm::MDBuilder MDHelper(CGF.getLLVMContext());
479  llvm::MDNode *RNode = MDHelper.createRange(APInt(32, low), APInt(32, high));
480  Function *F = CGF.CGM.getIntrinsic(IntrinsicID, {});
481  llvm::Instruction *Call = CGF.Builder.CreateCall(F);
482  Call->setMetadata(llvm::LLVMContext::MD_range, RNode);
483  return Call;
484 }
485 
486 namespace {
487  struct WidthAndSignedness {
488  unsigned Width;
489  bool Signed;
490  };
491 }
492 
493 static WidthAndSignedness
495  const clang::QualType Type) {
496  assert(Type->isIntegerType() && "Given type is not an integer.");
497  unsigned Width = Type->isBooleanType() ? 1 : context.getTypeInfo(Type).Width;
498  bool Signed = Type->isSignedIntegerType();
499  return {Width, Signed};
500 }
501 
502 // Given one or more integer types, this function produces an integer type that
503 // encompasses them: any value in one of the given types could be expressed in
504 // the encompassing type.
505 static struct WidthAndSignedness
506 EncompassingIntegerType(ArrayRef<struct WidthAndSignedness> Types) {
507  assert(Types.size() > 0 && "Empty list of types.");
508 
509  // If any of the given types is signed, we must return a signed type.
510  bool Signed = false;
511  for (const auto &Type : Types) {
512  Signed |= Type.Signed;
513  }
514 
515  // The encompassing type must have a width greater than or equal to the width
516  // of the specified types. Additionally, if the encompassing type is signed,
517  // its width must be strictly greater than the width of any unsigned types
518  // given.
519  unsigned Width = 0;
520  for (const auto &Type : Types) {
521  unsigned MinWidth = Type.Width + (Signed && !Type.Signed);
522  if (Width < MinWidth) {
523  Width = MinWidth;
524  }
525  }
526 
527  return {Width, Signed};
528 }
529 
530 Value *CodeGenFunction::EmitVAStartEnd(Value *ArgValue, bool IsStart) {
531  llvm::Type *DestType = Int8PtrTy;
532  if (ArgValue->getType() != DestType)
533  ArgValue =
534  Builder.CreateBitCast(ArgValue, DestType, ArgValue->getName().data());
535 
536  Intrinsic::ID inst = IsStart ? Intrinsic::vastart : Intrinsic::vaend;
537  return Builder.CreateCall(CGM.getIntrinsic(inst), ArgValue);
538 }
539 
540 /// Checks if using the result of __builtin_object_size(p, @p From) in place of
541 /// __builtin_object_size(p, @p To) is correct
542 static bool areBOSTypesCompatible(int From, int To) {
543  // Note: Our __builtin_object_size implementation currently treats Type=0 and
544  // Type=2 identically. Encoding this implementation detail here may make
545  // improving __builtin_object_size difficult in the future, so it's omitted.
546  return From == To || (From == 0 && To == 1) || (From == 3 && To == 2);
547 }
548 
549 static llvm::Value *
550 getDefaultBuiltinObjectSizeResult(unsigned Type, llvm::IntegerType *ResType) {
551  return ConstantInt::get(ResType, (Type & 2) ? 0 : -1, /*isSigned=*/true);
552 }
553 
554 llvm::Value *
555 CodeGenFunction::evaluateOrEmitBuiltinObjectSize(const Expr *E, unsigned Type,
556  llvm::IntegerType *ResType,
557  llvm::Value *EmittedE,
558  bool IsDynamic) {
559  uint64_t ObjectSize;
560  if (!E->tryEvaluateObjectSize(ObjectSize, getContext(), Type))
561  return emitBuiltinObjectSize(E, Type, ResType, EmittedE, IsDynamic);
562  return ConstantInt::get(ResType, ObjectSize, /*isSigned=*/true);
563 }
564 
565 /// Returns a Value corresponding to the size of the given expression.
566 /// This Value may be either of the following:
567 /// - A llvm::Argument (if E is a param with the pass_object_size attribute on
568 /// it)
569 /// - A call to the @llvm.objectsize intrinsic
570 ///
571 /// EmittedE is the result of emitting `E` as a scalar expr. If it's non-null
572 /// and we wouldn't otherwise try to reference a pass_object_size parameter,
573 /// we'll call @llvm.objectsize on EmittedE, rather than emitting E.
574 llvm::Value *
575 CodeGenFunction::emitBuiltinObjectSize(const Expr *E, unsigned Type,
576  llvm::IntegerType *ResType,
577  llvm::Value *EmittedE, bool IsDynamic) {
578  // We need to reference an argument if the pointer is a parameter with the
579  // pass_object_size attribute.
580  if (auto *D = dyn_cast<DeclRefExpr>(E->IgnoreParenImpCasts())) {
581  auto *Param = dyn_cast<ParmVarDecl>(D->getDecl());
582  auto *PS = D->getDecl()->getAttr<PassObjectSizeAttr>();
583  if (Param != nullptr && PS != nullptr &&
584  areBOSTypesCompatible(PS->getType(), Type)) {
585  auto Iter = SizeArguments.find(Param);
586  assert(Iter != SizeArguments.end());
587 
588  const ImplicitParamDecl *D = Iter->second;
589  auto DIter = LocalDeclMap.find(D);
590  assert(DIter != LocalDeclMap.end());
591 
592  return EmitLoadOfScalar(DIter->second, /*volatile=*/false,
593  getContext().getSizeType(), E->getBeginLoc());
594  }
595  }
596 
597  // LLVM can't handle Type=3 appropriately, and __builtin_object_size shouldn't
598  // evaluate E for side-effects. In either case, we shouldn't lower to
599  // @llvm.objectsize.
600  if (Type == 3 || (!EmittedE && E->HasSideEffects(getContext())))
601  return getDefaultBuiltinObjectSizeResult(Type, ResType);
602 
603  Value *Ptr = EmittedE ? EmittedE : EmitScalarExpr(E);
604  assert(Ptr->getType()->isPointerTy() &&
605  "Non-pointer passed to __builtin_object_size?");
606 
607  Function *F =
608  CGM.getIntrinsic(Intrinsic::objectsize, {ResType, Ptr->getType()});
609 
610  // LLVM only supports 0 and 2, make sure that we pass along that as a boolean.
611  Value *Min = Builder.getInt1((Type & 2) != 0);
612  // For GCC compatibility, __builtin_object_size treat NULL as unknown size.
613  Value *NullIsUnknown = Builder.getTrue();
614  Value *Dynamic = Builder.getInt1(IsDynamic);
615  return Builder.CreateCall(F, {Ptr, Min, NullIsUnknown, Dynamic});
616 }
617 
618 namespace {
619 /// A struct to generically describe a bit test intrinsic.
620 struct BitTest {
621  enum ActionKind : uint8_t { TestOnly, Complement, Reset, Set };
622  enum InterlockingKind : uint8_t {
623  Unlocked,
624  Sequential,
625  Acquire,
626  Release,
627  NoFence
628  };
629 
630  ActionKind Action;
631  InterlockingKind Interlocking;
632  bool Is64Bit;
633 
634  static BitTest decodeBitTestBuiltin(unsigned BuiltinID);
635 };
636 } // namespace
637 
638 BitTest BitTest::decodeBitTestBuiltin(unsigned BuiltinID) {
639  switch (BuiltinID) {
640  // Main portable variants.
641  case Builtin::BI_bittest:
642  return {TestOnly, Unlocked, false};
643  case Builtin::BI_bittestandcomplement:
644  return {Complement, Unlocked, false};
645  case Builtin::BI_bittestandreset:
646  return {Reset, Unlocked, false};
647  case Builtin::BI_bittestandset:
648  return {Set, Unlocked, false};
649  case Builtin::BI_interlockedbittestandreset:
650  return {Reset, Sequential, false};
651  case Builtin::BI_interlockedbittestandset:
652  return {Set, Sequential, false};
653 
654  // X86-specific 64-bit variants.
655  case Builtin::BI_bittest64:
656  return {TestOnly, Unlocked, true};
657  case Builtin::BI_bittestandcomplement64:
658  return {Complement, Unlocked, true};
659  case Builtin::BI_bittestandreset64:
660  return {Reset, Unlocked, true};
661  case Builtin::BI_bittestandset64:
662  return {Set, Unlocked, true};
663  case Builtin::BI_interlockedbittestandreset64:
664  return {Reset, Sequential, true};
665  case Builtin::BI_interlockedbittestandset64:
666  return {Set, Sequential, true};
667 
668  // ARM/AArch64-specific ordering variants.
669  case Builtin::BI_interlockedbittestandset_acq:
670  return {Set, Acquire, false};
671  case Builtin::BI_interlockedbittestandset_rel:
672  return {Set, Release, false};
673  case Builtin::BI_interlockedbittestandset_nf:
674  return {Set, NoFence, false};
675  case Builtin::BI_interlockedbittestandreset_acq:
676  return {Reset, Acquire, false};
677  case Builtin::BI_interlockedbittestandreset_rel:
678  return {Reset, Release, false};
679  case Builtin::BI_interlockedbittestandreset_nf:
680  return {Reset, NoFence, false};
681  }
682  llvm_unreachable("expected only bittest intrinsics");
683 }
684 
685 static char bitActionToX86BTCode(BitTest::ActionKind A) {
686  switch (A) {
687  case BitTest::TestOnly: return '\0';
688  case BitTest::Complement: return 'c';
689  case BitTest::Reset: return 'r';
690  case BitTest::Set: return 's';
691  }
692  llvm_unreachable("invalid action");
693 }
694 
696  BitTest BT,
697  const CallExpr *E, Value *BitBase,
698  Value *BitPos) {
699  char Action = bitActionToX86BTCode(BT.Action);
700  char SizeSuffix = BT.Is64Bit ? 'q' : 'l';
701 
702  // Build the assembly.
703  SmallString<64> Asm;
704  raw_svector_ostream AsmOS(Asm);
705  if (BT.Interlocking != BitTest::Unlocked)
706  AsmOS << "lock ";
707  AsmOS << "bt";
708  if (Action)
709  AsmOS << Action;
710  AsmOS << SizeSuffix << " $2, ($1)\n\tsetc ${0:b}";
711 
712  // Build the constraints. FIXME: We should support immediates when possible.
713  std::string Constraints = "=r,r,r,~{cc},~{flags},~{fpsr}";
714  llvm::IntegerType *IntType = llvm::IntegerType::get(
715  CGF.getLLVMContext(),
716  CGF.getContext().getTypeSize(E->getArg(1)->getType()));
717  llvm::Type *IntPtrType = IntType->getPointerTo();
718  llvm::FunctionType *FTy =
719  llvm::FunctionType::get(CGF.Int8Ty, {IntPtrType, IntType}, false);
720 
721  llvm::InlineAsm *IA =
722  llvm::InlineAsm::get(FTy, Asm, Constraints, /*SideEffects=*/true);
723  return CGF.Builder.CreateCall(IA, {BitBase, BitPos});
724 }
725 
726 static llvm::AtomicOrdering
727 getBitTestAtomicOrdering(BitTest::InterlockingKind I) {
728  switch (I) {
729  case BitTest::Unlocked: return llvm::AtomicOrdering::NotAtomic;
730  case BitTest::Sequential: return llvm::AtomicOrdering::SequentiallyConsistent;
731  case BitTest::Acquire: return llvm::AtomicOrdering::Acquire;
732  case BitTest::Release: return llvm::AtomicOrdering::Release;
733  case BitTest::NoFence: return llvm::AtomicOrdering::Monotonic;
734  }
735  llvm_unreachable("invalid interlocking");
736 }
737 
738 /// Emit a _bittest* intrinsic. These intrinsics take a pointer to an array of
739 /// bits and a bit position and read and optionally modify the bit at that
740 /// position. The position index can be arbitrarily large, i.e. it can be larger
741 /// than 31 or 63, so we need an indexed load in the general case.
743  unsigned BuiltinID,
744  const CallExpr *E) {
745  Value *BitBase = CGF.EmitScalarExpr(E->getArg(0));
746  Value *BitPos = CGF.EmitScalarExpr(E->getArg(1));
747 
748  BitTest BT = BitTest::decodeBitTestBuiltin(BuiltinID);
749 
750  // X86 has special BT, BTC, BTR, and BTS instructions that handle the array
751  // indexing operation internally. Use them if possible.
752  llvm::Triple::ArchType Arch = CGF.getTarget().getTriple().getArch();
753  if (Arch == llvm::Triple::x86 || Arch == llvm::Triple::x86_64)
754  return EmitX86BitTestIntrinsic(CGF, BT, E, BitBase, BitPos);
755 
756  // Otherwise, use generic code to load one byte and test the bit. Use all but
757  // the bottom three bits as the array index, and the bottom three bits to form
758  // a mask.
759  // Bit = BitBaseI8[BitPos >> 3] & (1 << (BitPos & 0x7)) != 0;
760  Value *ByteIndex = CGF.Builder.CreateAShr(
761  BitPos, llvm::ConstantInt::get(BitPos->getType(), 3), "bittest.byteidx");
762  Value *BitBaseI8 = CGF.Builder.CreatePointerCast(BitBase, CGF.Int8PtrTy);
763  Address ByteAddr(CGF.Builder.CreateInBoundsGEP(CGF.Int8Ty, BitBaseI8,
764  ByteIndex, "bittest.byteaddr"),
765  CharUnits::One());
766  Value *PosLow =
767  CGF.Builder.CreateAnd(CGF.Builder.CreateTrunc(BitPos, CGF.Int8Ty),
768  llvm::ConstantInt::get(CGF.Int8Ty, 0x7));
769 
770  // The updating instructions will need a mask.
771  Value *Mask = nullptr;
772  if (BT.Action != BitTest::TestOnly) {
773  Mask = CGF.Builder.CreateShl(llvm::ConstantInt::get(CGF.Int8Ty, 1), PosLow,
774  "bittest.mask");
775  }
776 
777  // Check the action and ordering of the interlocked intrinsics.
778  llvm::AtomicOrdering Ordering = getBitTestAtomicOrdering(BT.Interlocking);
779 
780  Value *OldByte = nullptr;
781  if (Ordering != llvm::AtomicOrdering::NotAtomic) {
782  // Emit a combined atomicrmw load/store operation for the interlocked
783  // intrinsics.
784  llvm::AtomicRMWInst::BinOp RMWOp = llvm::AtomicRMWInst::Or;
785  if (BT.Action == BitTest::Reset) {
786  Mask = CGF.Builder.CreateNot(Mask);
787  RMWOp = llvm::AtomicRMWInst::And;
788  }
789  OldByte = CGF.Builder.CreateAtomicRMW(RMWOp, ByteAddr.getPointer(), Mask,
790  Ordering);
791  } else {
792  // Emit a plain load for the non-interlocked intrinsics.
793  OldByte = CGF.Builder.CreateLoad(ByteAddr, "bittest.byte");
794  Value *NewByte = nullptr;
795  switch (BT.Action) {
796  case BitTest::TestOnly:
797  // Don't store anything.
798  break;
799  case BitTest::Complement:
800  NewByte = CGF.Builder.CreateXor(OldByte, Mask);
801  break;
802  case BitTest::Reset:
803  NewByte = CGF.Builder.CreateAnd(OldByte, CGF.Builder.CreateNot(Mask));
804  break;
805  case BitTest::Set:
806  NewByte = CGF.Builder.CreateOr(OldByte, Mask);
807  break;
808  }
809  if (NewByte)
810  CGF.Builder.CreateStore(NewByte, ByteAddr);
811  }
812 
813  // However we loaded the old byte, either by plain load or atomicrmw, shift
814  // the bit into the low position and mask it to 0 or 1.
815  Value *ShiftedByte = CGF.Builder.CreateLShr(OldByte, PosLow, "bittest.shr");
816  return CGF.Builder.CreateAnd(
817  ShiftedByte, llvm::ConstantInt::get(CGF.Int8Ty, 1), "bittest.res");
818 }
819 
820 namespace {
821 enum class MSVCSetJmpKind {
822  _setjmpex,
823  _setjmp3,
824  _setjmp
825 };
826 }
827 
828 /// MSVC handles setjmp a bit differently on different platforms. On every
829 /// architecture except 32-bit x86, the frame address is passed. On x86, extra
830 /// parameters can be passed as variadic arguments, but we always pass none.
832  const CallExpr *E) {
833  llvm::Value *Arg1 = nullptr;
834  llvm::Type *Arg1Ty = nullptr;
835  StringRef Name;
836  bool IsVarArg = false;
837  if (SJKind == MSVCSetJmpKind::_setjmp3) {
838  Name = "_setjmp3";
839  Arg1Ty = CGF.Int32Ty;
840  Arg1 = llvm::ConstantInt::get(CGF.IntTy, 0);
841  IsVarArg = true;
842  } else {
843  Name = SJKind == MSVCSetJmpKind::_setjmp ? "_setjmp" : "_setjmpex";
844  Arg1Ty = CGF.Int8PtrTy;
845  if (CGF.getTarget().getTriple().getArch() == llvm::Triple::aarch64) {
846  Arg1 = CGF.Builder.CreateCall(CGF.CGM.getIntrinsic(Intrinsic::sponentry));
847  } else
848  Arg1 = CGF.Builder.CreateCall(CGF.CGM.getIntrinsic(Intrinsic::frameaddress),
849  llvm::ConstantInt::get(CGF.Int32Ty, 0));
850  }
851 
852  // Mark the call site and declaration with ReturnsTwice.
853  llvm::Type *ArgTypes[2] = {CGF.Int8PtrTy, Arg1Ty};
854  llvm::AttributeList ReturnsTwiceAttr = llvm::AttributeList::get(
855  CGF.getLLVMContext(), llvm::AttributeList::FunctionIndex,
856  llvm::Attribute::ReturnsTwice);
857  llvm::FunctionCallee SetJmpFn = CGF.CGM.CreateRuntimeFunction(
858  llvm::FunctionType::get(CGF.IntTy, ArgTypes, IsVarArg), Name,
859  ReturnsTwiceAttr, /*Local=*/true);
860 
861  llvm::Value *Buf = CGF.Builder.CreateBitOrPointerCast(
862  CGF.EmitScalarExpr(E->getArg(0)), CGF.Int8PtrTy);
863  llvm::Value *Args[] = {Buf, Arg1};
864  llvm::CallBase *CB = CGF.EmitRuntimeCallOrInvoke(SetJmpFn, Args);
865  CB->setAttributes(ReturnsTwiceAttr);
866  return RValue::get(CB);
867 }
868 
869 // Many of MSVC builtins are on x64, ARM and AArch64; to avoid repeating code,
870 // we handle them here.
872  _BitScanForward,
873  _BitScanReverse,
874  _InterlockedAnd,
875  _InterlockedDecrement,
876  _InterlockedExchange,
877  _InterlockedExchangeAdd,
878  _InterlockedExchangeSub,
879  _InterlockedIncrement,
880  _InterlockedOr,
881  _InterlockedXor,
882  _InterlockedExchangeAdd_acq,
883  _InterlockedExchangeAdd_rel,
884  _InterlockedExchangeAdd_nf,
885  _InterlockedExchange_acq,
886  _InterlockedExchange_rel,
887  _InterlockedExchange_nf,
888  _InterlockedCompareExchange_acq,
889  _InterlockedCompareExchange_rel,
890  _InterlockedCompareExchange_nf,
891  _InterlockedOr_acq,
892  _InterlockedOr_rel,
893  _InterlockedOr_nf,
894  _InterlockedXor_acq,
895  _InterlockedXor_rel,
896  _InterlockedXor_nf,
897  _InterlockedAnd_acq,
898  _InterlockedAnd_rel,
899  _InterlockedAnd_nf,
900  _InterlockedIncrement_acq,
901  _InterlockedIncrement_rel,
902  _InterlockedIncrement_nf,
903  _InterlockedDecrement_acq,
904  _InterlockedDecrement_rel,
905  _InterlockedDecrement_nf,
906  __fastfail,
907 };
908 
910  const CallExpr *E) {
911  switch (BuiltinID) {
912  case MSVCIntrin::_BitScanForward:
913  case MSVCIntrin::_BitScanReverse: {
914  Value *ArgValue = EmitScalarExpr(E->getArg(1));
915 
916  llvm::Type *ArgType = ArgValue->getType();
917  llvm::Type *IndexType =
918  EmitScalarExpr(E->getArg(0))->getType()->getPointerElementType();
919  llvm::Type *ResultType = ConvertType(E->getType());
920 
921  Value *ArgZero = llvm::Constant::getNullValue(ArgType);
922  Value *ResZero = llvm::Constant::getNullValue(ResultType);
923  Value *ResOne = llvm::ConstantInt::get(ResultType, 1);
924 
925  BasicBlock *Begin = Builder.GetInsertBlock();
926  BasicBlock *End = createBasicBlock("bitscan_end", this->CurFn);
927  Builder.SetInsertPoint(End);
928  PHINode *Result = Builder.CreatePHI(ResultType, 2, "bitscan_result");
929 
930  Builder.SetInsertPoint(Begin);
931  Value *IsZero = Builder.CreateICmpEQ(ArgValue, ArgZero);
932  BasicBlock *NotZero = createBasicBlock("bitscan_not_zero", this->CurFn);
933  Builder.CreateCondBr(IsZero, End, NotZero);
934  Result->addIncoming(ResZero, Begin);
935 
936  Builder.SetInsertPoint(NotZero);
937  Address IndexAddress = EmitPointerWithAlignment(E->getArg(0));
938 
939  if (BuiltinID == MSVCIntrin::_BitScanForward) {
940  Function *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType);
941  Value *ZeroCount = Builder.CreateCall(F, {ArgValue, Builder.getTrue()});
942  ZeroCount = Builder.CreateIntCast(ZeroCount, IndexType, false);
943  Builder.CreateStore(ZeroCount, IndexAddress, false);
944  } else {
945  unsigned ArgWidth = cast<llvm::IntegerType>(ArgType)->getBitWidth();
946  Value *ArgTypeLastIndex = llvm::ConstantInt::get(IndexType, ArgWidth - 1);
947 
948  Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType);
949  Value *ZeroCount = Builder.CreateCall(F, {ArgValue, Builder.getTrue()});
950  ZeroCount = Builder.CreateIntCast(ZeroCount, IndexType, false);
951  Value *Index = Builder.CreateNSWSub(ArgTypeLastIndex, ZeroCount);
952  Builder.CreateStore(Index, IndexAddress, false);
953  }
954  Builder.CreateBr(End);
955  Result->addIncoming(ResOne, NotZero);
956 
957  Builder.SetInsertPoint(End);
958  return Result;
959  }
960  case MSVCIntrin::_InterlockedAnd:
961  return MakeBinaryAtomicValue(*this, AtomicRMWInst::And, E);
962  case MSVCIntrin::_InterlockedExchange:
963  return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xchg, E);
964  case MSVCIntrin::_InterlockedExchangeAdd:
965  return MakeBinaryAtomicValue(*this, AtomicRMWInst::Add, E);
966  case MSVCIntrin::_InterlockedExchangeSub:
967  return MakeBinaryAtomicValue(*this, AtomicRMWInst::Sub, E);
968  case MSVCIntrin::_InterlockedOr:
969  return MakeBinaryAtomicValue(*this, AtomicRMWInst::Or, E);
970  case MSVCIntrin::_InterlockedXor:
971  return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xor, E);
972  case MSVCIntrin::_InterlockedExchangeAdd_acq:
973  return MakeBinaryAtomicValue(*this, AtomicRMWInst::Add, E,
974  AtomicOrdering::Acquire);
975  case MSVCIntrin::_InterlockedExchangeAdd_rel:
976  return MakeBinaryAtomicValue(*this, AtomicRMWInst::Add, E,
977  AtomicOrdering::Release);
978  case MSVCIntrin::_InterlockedExchangeAdd_nf:
979  return MakeBinaryAtomicValue(*this, AtomicRMWInst::Add, E,
980  AtomicOrdering::Monotonic);
981  case MSVCIntrin::_InterlockedExchange_acq:
982  return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xchg, E,
983  AtomicOrdering::Acquire);
984  case MSVCIntrin::_InterlockedExchange_rel:
985  return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xchg, E,
986  AtomicOrdering::Release);
987  case MSVCIntrin::_InterlockedExchange_nf:
988  return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xchg, E,
989  AtomicOrdering::Monotonic);
990  case MSVCIntrin::_InterlockedCompareExchange_acq:
991  return EmitAtomicCmpXchgForMSIntrin(*this, E, AtomicOrdering::Acquire);
992  case MSVCIntrin::_InterlockedCompareExchange_rel:
993  return EmitAtomicCmpXchgForMSIntrin(*this, E, AtomicOrdering::Release);
994  case MSVCIntrin::_InterlockedCompareExchange_nf:
995  return EmitAtomicCmpXchgForMSIntrin(*this, E, AtomicOrdering::Monotonic);
996  case MSVCIntrin::_InterlockedOr_acq:
997  return MakeBinaryAtomicValue(*this, AtomicRMWInst::Or, E,
998  AtomicOrdering::Acquire);
999  case MSVCIntrin::_InterlockedOr_rel:
1000  return MakeBinaryAtomicValue(*this, AtomicRMWInst::Or, E,
1001  AtomicOrdering::Release);
1002  case MSVCIntrin::_InterlockedOr_nf:
1003  return MakeBinaryAtomicValue(*this, AtomicRMWInst::Or, E,
1004  AtomicOrdering::Monotonic);
1005  case MSVCIntrin::_InterlockedXor_acq:
1006  return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xor, E,
1007  AtomicOrdering::Acquire);
1008  case MSVCIntrin::_InterlockedXor_rel:
1009  return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xor, E,
1010  AtomicOrdering::Release);
1011  case MSVCIntrin::_InterlockedXor_nf:
1012  return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xor, E,
1013  AtomicOrdering::Monotonic);
1014  case MSVCIntrin::_InterlockedAnd_acq:
1015  return MakeBinaryAtomicValue(*this, AtomicRMWInst::And, E,
1016  AtomicOrdering::Acquire);
1017  case MSVCIntrin::_InterlockedAnd_rel:
1018  return MakeBinaryAtomicValue(*this, AtomicRMWInst::And, E,
1019  AtomicOrdering::Release);
1020  case MSVCIntrin::_InterlockedAnd_nf:
1021  return MakeBinaryAtomicValue(*this, AtomicRMWInst::And, E,
1022  AtomicOrdering::Monotonic);
1023  case MSVCIntrin::_InterlockedIncrement_acq:
1024  return EmitAtomicIncrementValue(*this, E, AtomicOrdering::Acquire);
1025  case MSVCIntrin::_InterlockedIncrement_rel:
1026  return EmitAtomicIncrementValue(*this, E, AtomicOrdering::Release);
1027  case MSVCIntrin::_InterlockedIncrement_nf:
1028  return EmitAtomicIncrementValue(*this, E, AtomicOrdering::Monotonic);
1029  case MSVCIntrin::_InterlockedDecrement_acq:
1030  return EmitAtomicDecrementValue(*this, E, AtomicOrdering::Acquire);
1031  case MSVCIntrin::_InterlockedDecrement_rel:
1032  return EmitAtomicDecrementValue(*this, E, AtomicOrdering::Release);
1033  case MSVCIntrin::_InterlockedDecrement_nf:
1034  return EmitAtomicDecrementValue(*this, E, AtomicOrdering::Monotonic);
1035 
1036  case MSVCIntrin::_InterlockedDecrement:
1037  return EmitAtomicDecrementValue(*this, E);
1038  case MSVCIntrin::_InterlockedIncrement:
1039  return EmitAtomicIncrementValue(*this, E);
1040 
1041  case MSVCIntrin::__fastfail: {
1042  // Request immediate process termination from the kernel. The instruction
1043  // sequences to do this are documented on MSDN:
1044  // https://msdn.microsoft.com/en-us/library/dn774154.aspx
1045  llvm::Triple::ArchType ISA = getTarget().getTriple().getArch();
1046  StringRef Asm, Constraints;
1047  switch (ISA) {
1048  default:
1049  ErrorUnsupported(E, "__fastfail call for this architecture");
1050  break;
1051  case llvm::Triple::x86:
1052  case llvm::Triple::x86_64:
1053  Asm = "int $$0x29";
1054  Constraints = "{cx}";
1055  break;
1056  case llvm::Triple::thumb:
1057  Asm = "udf #251";
1058  Constraints = "{r0}";
1059  break;
1060  case llvm::Triple::aarch64:
1061  Asm = "brk #0xF003";
1062  Constraints = "{w0}";
1063  }
1064  llvm::FunctionType *FTy = llvm::FunctionType::get(VoidTy, {Int32Ty}, false);
1065  llvm::InlineAsm *IA =
1066  llvm::InlineAsm::get(FTy, Asm, Constraints, /*SideEffects=*/true);
1067  llvm::AttributeList NoReturnAttr = llvm::AttributeList::get(
1068  getLLVMContext(), llvm::AttributeList::FunctionIndex,
1069  llvm::Attribute::NoReturn);
1070  llvm::CallInst *CI = Builder.CreateCall(IA, EmitScalarExpr(E->getArg(0)));
1071  CI->setAttributes(NoReturnAttr);
1072  return CI;
1073  }
1074  }
1075  llvm_unreachable("Incorrect MSVC intrinsic!");
1076 }
1077 
1078 namespace {
1079 // ARC cleanup for __builtin_os_log_format
1080 struct CallObjCArcUse final : EHScopeStack::Cleanup {
1081  CallObjCArcUse(llvm::Value *object) : object(object) {}
1082  llvm::Value *object;
1083 
1084  void Emit(CodeGenFunction &CGF, Flags flags) override {
1085  CGF.EmitARCIntrinsicUse(object);
1086  }
1087 };
1088 }
1089 
1092  assert((Kind == BCK_CLZPassedZero || Kind == BCK_CTZPassedZero)
1093  && "Unsupported builtin check kind");
1094 
1095  Value *ArgValue = EmitScalarExpr(E);
1096  if (!SanOpts.has(SanitizerKind::Builtin) || !getTarget().isCLZForZeroUndef())
1097  return ArgValue;
1098 
1099  SanitizerScope SanScope(this);
1100  Value *Cond = Builder.CreateICmpNE(
1101  ArgValue, llvm::Constant::getNullValue(ArgValue->getType()));
1102  EmitCheck(std::make_pair(Cond, SanitizerKind::Builtin),
1103  SanitizerHandler::InvalidBuiltin,
1104  {EmitCheckSourceLocation(E->getExprLoc()),
1105  llvm::ConstantInt::get(Builder.getInt8Ty(), Kind)},
1106  None);
1107  return ArgValue;
1108 }
1109 
1110 /// Get the argument type for arguments to os_log_helper.
1112  QualType UnsignedTy = C.getIntTypeForBitwidth(Size * 8, /*Signed=*/false);
1113  return C.getCanonicalType(UnsignedTy);
1114 }
1115 
1117  const analyze_os_log::OSLogBufferLayout &Layout,
1118  CharUnits BufferAlignment) {
1119  ASTContext &Ctx = getContext();
1120 
1121  llvm::SmallString<64> Name;
1122  {
1123  raw_svector_ostream OS(Name);
1124  OS << "__os_log_helper";
1125  OS << "_" << BufferAlignment.getQuantity();
1126  OS << "_" << int(Layout.getSummaryByte());
1127  OS << "_" << int(Layout.getNumArgsByte());
1128  for (const auto &Item : Layout.Items)
1129  OS << "_" << int(Item.getSizeByte()) << "_"
1130  << int(Item.getDescriptorByte());
1131  }
1132 
1133  if (llvm::Function *F = CGM.getModule().getFunction(Name))
1134  return F;
1135 
1137  FunctionArgList Args;
1138  Args.push_back(ImplicitParamDecl::Create(
1139  Ctx, nullptr, SourceLocation(), &Ctx.Idents.get("buffer"), Ctx.VoidPtrTy,
1141  ArgTys.emplace_back(Ctx.VoidPtrTy);
1142 
1143  for (unsigned int I = 0, E = Layout.Items.size(); I < E; ++I) {
1144  char Size = Layout.Items[I].getSizeByte();
1145  if (!Size)
1146  continue;
1147 
1148  QualType ArgTy = getOSLogArgType(Ctx, Size);
1149  Args.push_back(ImplicitParamDecl::Create(
1150  Ctx, nullptr, SourceLocation(),
1151  &Ctx.Idents.get(std::string("arg") + llvm::to_string(I)), ArgTy,
1153  ArgTys.emplace_back(ArgTy);
1154  }
1155 
1156  QualType ReturnTy = Ctx.VoidTy;
1157  QualType FuncionTy = Ctx.getFunctionType(ReturnTy, ArgTys, {});
1158 
1159  // The helper function has linkonce_odr linkage to enable the linker to merge
1160  // identical functions. To ensure the merging always happens, 'noinline' is
1161  // attached to the function when compiling with -Oz.
1162  const CGFunctionInfo &FI =
1163  CGM.getTypes().arrangeBuiltinFunctionDeclaration(ReturnTy, Args);
1164  llvm::FunctionType *FuncTy = CGM.getTypes().GetFunctionType(FI);
1165  llvm::Function *Fn = llvm::Function::Create(
1166  FuncTy, llvm::GlobalValue::LinkOnceODRLinkage, Name, &CGM.getModule());
1167  Fn->setVisibility(llvm::GlobalValue::HiddenVisibility);
1168  CGM.SetLLVMFunctionAttributes(GlobalDecl(), FI, Fn);
1169  CGM.SetLLVMFunctionAttributesForDefinition(nullptr, Fn);
1170  Fn->setDoesNotThrow();
1171 
1172  // Attach 'noinline' at -Oz.
1173  if (CGM.getCodeGenOpts().OptimizeSize == 2)
1174  Fn->addFnAttr(llvm::Attribute::NoInline);
1175 
1176  auto NL = ApplyDebugLocation::CreateEmpty(*this);
1177  IdentifierInfo *II = &Ctx.Idents.get(Name);
1180  FuncionTy, nullptr, SC_PrivateExtern, false, false);
1181 
1182  StartFunction(FD, ReturnTy, Fn, FI, Args);
1183 
1184  // Create a scope with an artificial location for the body of this function.
1185  auto AL = ApplyDebugLocation::CreateArtificial(*this);
1186 
1187  CharUnits Offset;
1188  Address BufAddr(Builder.CreateLoad(GetAddrOfLocalVar(Args[0]), "buf"),
1189  BufferAlignment);
1190  Builder.CreateStore(Builder.getInt8(Layout.getSummaryByte()),
1191  Builder.CreateConstByteGEP(BufAddr, Offset++, "summary"));
1192  Builder.CreateStore(Builder.getInt8(Layout.getNumArgsByte()),
1193  Builder.CreateConstByteGEP(BufAddr, Offset++, "numArgs"));
1194 
1195  unsigned I = 1;
1196  for (const auto &Item : Layout.Items) {
1197  Builder.CreateStore(
1198  Builder.getInt8(Item.getDescriptorByte()),
1199  Builder.CreateConstByteGEP(BufAddr, Offset++, "argDescriptor"));
1200  Builder.CreateStore(
1201  Builder.getInt8(Item.getSizeByte()),
1202  Builder.CreateConstByteGEP(BufAddr, Offset++, "argSize"));
1203 
1204  CharUnits Size = Item.size();
1205  if (!Size.getQuantity())
1206  continue;
1207 
1208  Address Arg = GetAddrOfLocalVar(Args[I]);
1209  Address Addr = Builder.CreateConstByteGEP(BufAddr, Offset, "argData");
1210  Addr = Builder.CreateBitCast(Addr, Arg.getPointer()->getType(),
1211  "argDataCast");
1212  Builder.CreateStore(Builder.CreateLoad(Arg), Addr);
1213  Offset += Size;
1214  ++I;
1215  }
1216 
1217  FinishFunction();
1218 
1219  return Fn;
1220 }
1221 
1223  assert(E.getNumArgs() >= 2 &&
1224  "__builtin_os_log_format takes at least 2 arguments");
1225  ASTContext &Ctx = getContext();
1228  Address BufAddr = EmitPointerWithAlignment(E.getArg(0));
1229  llvm::SmallVector<llvm::Value *, 4> RetainableOperands;
1230 
1231  // Ignore argument 1, the format string. It is not currently used.
1232  CallArgList Args;
1233  Args.add(RValue::get(BufAddr.getPointer()), Ctx.VoidPtrTy);
1234 
1235  for (const auto &Item : Layout.Items) {
1236  int Size = Item.getSizeByte();
1237  if (!Size)
1238  continue;
1239 
1240  llvm::Value *ArgVal;
1241 
1242  if (Item.getKind() == analyze_os_log::OSLogBufferItem::MaskKind) {
1243  uint64_t Val = 0;
1244  for (unsigned I = 0, E = Item.getMaskType().size(); I < E; ++I)
1245  Val |= ((uint64_t)Item.getMaskType()[I]) << I * 8;
1246  ArgVal = llvm::Constant::getIntegerValue(Int64Ty, llvm::APInt(64, Val));
1247  } else if (const Expr *TheExpr = Item.getExpr()) {
1248  ArgVal = EmitScalarExpr(TheExpr, /*Ignore*/ false);
1249 
1250  // Check if this is a retainable type.
1251  if (TheExpr->getType()->isObjCRetainableType()) {
1252  assert(getEvaluationKind(TheExpr->getType()) == TEK_Scalar &&
1253  "Only scalar can be a ObjC retainable type");
1254  // Check if the object is constant, if not, save it in
1255  // RetainableOperands.
1256  if (!isa<Constant>(ArgVal))
1257  RetainableOperands.push_back(ArgVal);
1258  }
1259  } else {
1260  ArgVal = Builder.getInt32(Item.getConstValue().getQuantity());
1261  }
1262 
1263  unsigned ArgValSize =
1264  CGM.getDataLayout().getTypeSizeInBits(ArgVal->getType());
1265  llvm::IntegerType *IntTy = llvm::Type::getIntNTy(getLLVMContext(),
1266  ArgValSize);
1267  ArgVal = Builder.CreateBitOrPointerCast(ArgVal, IntTy);
1268  CanQualType ArgTy = getOSLogArgType(Ctx, Size);
1269  // If ArgVal has type x86_fp80, zero-extend ArgVal.
1270  ArgVal = Builder.CreateZExtOrBitCast(ArgVal, ConvertType(ArgTy));
1271  Args.add(RValue::get(ArgVal), ArgTy);
1272  }
1273 
1274  const CGFunctionInfo &FI =
1275  CGM.getTypes().arrangeBuiltinFunctionCall(Ctx.VoidTy, Args);
1276  llvm::Function *F = CodeGenFunction(CGM).generateBuiltinOSLogHelperFunction(
1277  Layout, BufAddr.getAlignment());
1278  EmitCall(FI, CGCallee::forDirect(F), ReturnValueSlot(), Args);
1279 
1280  // Push a clang.arc.use cleanup for each object in RetainableOperands. The
1281  // cleanup will cause the use to appear after the final log call, keeping
1282  // the object valid while it’s held in the log buffer. Note that if there’s
1283  // a release cleanup on the object, it will already be active; since
1284  // cleanups are emitted in reverse order, the use will occur before the
1285  // object is released.
1286  if (!RetainableOperands.empty() && getLangOpts().ObjCAutoRefCount &&
1287  CGM.getCodeGenOpts().OptimizationLevel != 0)
1288  for (llvm::Value *Object : RetainableOperands)
1289  pushFullExprCleanup<CallObjCArcUse>(getARCCleanupKind(), Object);
1290 
1291  return RValue::get(BufAddr.getPointer());
1292 }
1293 
1294 /// Determine if a binop is a checked mixed-sign multiply we can specialize.
1295 static bool isSpecialMixedSignMultiply(unsigned BuiltinID,
1296  WidthAndSignedness Op1Info,
1297  WidthAndSignedness Op2Info,
1298  WidthAndSignedness ResultInfo) {
1299  return BuiltinID == Builtin::BI__builtin_mul_overflow &&
1300  std::max(Op1Info.Width, Op2Info.Width) >= ResultInfo.Width &&
1301  Op1Info.Signed != Op2Info.Signed;
1302 }
1303 
1304 /// Emit a checked mixed-sign multiply. This is a cheaper specialization of
1305 /// the generic checked-binop irgen.
1306 static RValue
1308  WidthAndSignedness Op1Info, const clang::Expr *Op2,
1309  WidthAndSignedness Op2Info,
1310  const clang::Expr *ResultArg, QualType ResultQTy,
1311  WidthAndSignedness ResultInfo) {
1312  assert(isSpecialMixedSignMultiply(Builtin::BI__builtin_mul_overflow, Op1Info,
1313  Op2Info, ResultInfo) &&
1314  "Not a mixed-sign multipliction we can specialize");
1315 
1316  // Emit the signed and unsigned operands.
1317  const clang::Expr *SignedOp = Op1Info.Signed ? Op1 : Op2;
1318  const clang::Expr *UnsignedOp = Op1Info.Signed ? Op2 : Op1;
1319  llvm::Value *Signed = CGF.EmitScalarExpr(SignedOp);
1320  llvm::Value *Unsigned = CGF.EmitScalarExpr(UnsignedOp);
1321  unsigned SignedOpWidth = Op1Info.Signed ? Op1Info.Width : Op2Info.Width;
1322  unsigned UnsignedOpWidth = Op1Info.Signed ? Op2Info.Width : Op1Info.Width;
1323 
1324  // One of the operands may be smaller than the other. If so, [s|z]ext it.
1325  if (SignedOpWidth < UnsignedOpWidth)
1326  Signed = CGF.Builder.CreateSExt(Signed, Unsigned->getType(), "op.sext");
1327  if (UnsignedOpWidth < SignedOpWidth)
1328  Unsigned = CGF.Builder.CreateZExt(Unsigned, Signed->getType(), "op.zext");
1329 
1330  llvm::Type *OpTy = Signed->getType();
1331  llvm::Value *Zero = llvm::Constant::getNullValue(OpTy);
1332  Address ResultPtr = CGF.EmitPointerWithAlignment(ResultArg);
1333  llvm::Type *ResTy = ResultPtr.getElementType();
1334  unsigned OpWidth = std::max(Op1Info.Width, Op2Info.Width);
1335 
1336  // Take the absolute value of the signed operand.
1337  llvm::Value *IsNegative = CGF.Builder.CreateICmpSLT(Signed, Zero);
1338  llvm::Value *AbsOfNegative = CGF.Builder.CreateSub(Zero, Signed);
1339  llvm::Value *AbsSigned =
1340  CGF.Builder.CreateSelect(IsNegative, AbsOfNegative, Signed);
1341 
1342  // Perform a checked unsigned multiplication.
1343  llvm::Value *UnsignedOverflow;
1344  llvm::Value *UnsignedResult =
1345  EmitOverflowIntrinsic(CGF, llvm::Intrinsic::umul_with_overflow, AbsSigned,
1346  Unsigned, UnsignedOverflow);
1347 
1348  llvm::Value *Overflow, *Result;
1349  if (ResultInfo.Signed) {
1350  // Signed overflow occurs if the result is greater than INT_MAX or lesser
1351  // than INT_MIN, i.e when |Result| > (INT_MAX + IsNegative).
1352  auto IntMax =
1353  llvm::APInt::getSignedMaxValue(ResultInfo.Width).zextOrSelf(OpWidth);
1354  llvm::Value *MaxResult =
1355  CGF.Builder.CreateAdd(llvm::ConstantInt::get(OpTy, IntMax),
1356  CGF.Builder.CreateZExt(IsNegative, OpTy));
1357  llvm::Value *SignedOverflow =
1358  CGF.Builder.CreateICmpUGT(UnsignedResult, MaxResult);
1359  Overflow = CGF.Builder.CreateOr(UnsignedOverflow, SignedOverflow);
1360 
1361  // Prepare the signed result (possibly by negating it).
1362  llvm::Value *NegativeResult = CGF.Builder.CreateNeg(UnsignedResult);
1363  llvm::Value *SignedResult =
1364  CGF.Builder.CreateSelect(IsNegative, NegativeResult, UnsignedResult);
1365  Result = CGF.Builder.CreateTrunc(SignedResult, ResTy);
1366  } else {
1367  // Unsigned overflow occurs if the result is < 0 or greater than UINT_MAX.
1368  llvm::Value *Underflow = CGF.Builder.CreateAnd(
1369  IsNegative, CGF.Builder.CreateIsNotNull(UnsignedResult));
1370  Overflow = CGF.Builder.CreateOr(UnsignedOverflow, Underflow);
1371  if (ResultInfo.Width < OpWidth) {
1372  auto IntMax =
1373  llvm::APInt::getMaxValue(ResultInfo.Width).zext(OpWidth);
1374  llvm::Value *TruncOverflow = CGF.Builder.CreateICmpUGT(
1375  UnsignedResult, llvm::ConstantInt::get(OpTy, IntMax));
1376  Overflow = CGF.Builder.CreateOr(Overflow, TruncOverflow);
1377  }
1378 
1379  // Negate the product if it would be negative in infinite precision.
1380  Result = CGF.Builder.CreateSelect(
1381  IsNegative, CGF.Builder.CreateNeg(UnsignedResult), UnsignedResult);
1382 
1383  Result = CGF.Builder.CreateTrunc(Result, ResTy);
1384  }
1385  assert(Overflow && Result && "Missing overflow or result");
1386 
1387  bool isVolatile =
1388  ResultArg->getType()->getPointeeType().isVolatileQualified();
1389  CGF.Builder.CreateStore(CGF.EmitToMemory(Result, ResultQTy), ResultPtr,
1390  isVolatile);
1391  return RValue::get(Overflow);
1392 }
1393 
1395  Value *&RecordPtr, CharUnits Align,
1396  llvm::FunctionCallee Func, int Lvl) {
1397  const auto *RT = RType->getAs<RecordType>();
1398  ASTContext &Context = CGF.getContext();
1399  RecordDecl *RD = RT->getDecl()->getDefinition();
1400  std::string Pad = std::string(Lvl * 4, ' ');
1401 
1402  Value *GString =
1403  CGF.Builder.CreateGlobalStringPtr(RType.getAsString() + " {\n");
1404  Value *Res = CGF.Builder.CreateCall(Func, {GString});
1405 
1406  static llvm::DenseMap<QualType, const char *> Types;
1407  if (Types.empty()) {
1408  Types[Context.CharTy] = "%c";
1409  Types[Context.BoolTy] = "%d";
1410  Types[Context.SignedCharTy] = "%hhd";
1411  Types[Context.UnsignedCharTy] = "%hhu";
1412  Types[Context.IntTy] = "%d";
1413  Types[Context.UnsignedIntTy] = "%u";
1414  Types[Context.LongTy] = "%ld";
1415  Types[Context.UnsignedLongTy] = "%lu";
1416  Types[Context.LongLongTy] = "%lld";
1417  Types[Context.UnsignedLongLongTy] = "%llu";
1418  Types[Context.ShortTy] = "%hd";
1419  Types[Context.UnsignedShortTy] = "%hu";
1420  Types[Context.VoidPtrTy] = "%p";
1421  Types[Context.FloatTy] = "%f";
1422  Types[Context.DoubleTy] = "%f";
1423  Types[Context.LongDoubleTy] = "%Lf";
1424  Types[Context.getPointerType(Context.CharTy)] = "%s";
1425  Types[Context.getPointerType(Context.getConstType(Context.CharTy))] = "%s";
1426  }
1427 
1428  for (const auto *FD : RD->fields()) {
1429  Value *FieldPtr = RecordPtr;
1430  if (RD->isUnion())
1431  FieldPtr = CGF.Builder.CreatePointerCast(
1432  FieldPtr, CGF.ConvertType(Context.getPointerType(FD->getType())));
1433  else
1434  FieldPtr = CGF.Builder.CreateStructGEP(CGF.ConvertType(RType), FieldPtr,
1435  FD->getFieldIndex());
1436 
1437  GString = CGF.Builder.CreateGlobalStringPtr(
1438  llvm::Twine(Pad)
1439  .concat(FD->getType().getAsString())
1440  .concat(llvm::Twine(' '))
1441  .concat(FD->getNameAsString())
1442  .concat(" : ")
1443  .str());
1444  Value *TmpRes = CGF.Builder.CreateCall(Func, {GString});
1445  Res = CGF.Builder.CreateAdd(Res, TmpRes);
1446 
1447  QualType CanonicalType =
1448  FD->getType().getUnqualifiedType().getCanonicalType();
1449 
1450  // We check whether we are in a recursive type
1451  if (CanonicalType->isRecordType()) {
1452  Value *TmpRes =
1453  dumpRecord(CGF, CanonicalType, FieldPtr, Align, Func, Lvl + 1);
1454  Res = CGF.Builder.CreateAdd(TmpRes, Res);
1455  continue;
1456  }
1457 
1458  // We try to determine the best format to print the current field
1459  llvm::Twine Format = Types.find(CanonicalType) == Types.end()
1460  ? Types[Context.VoidPtrTy]
1461  : Types[CanonicalType];
1462 
1463  Address FieldAddress = Address(FieldPtr, Align);
1464  FieldPtr = CGF.Builder.CreateLoad(FieldAddress);
1465 
1466  // FIXME Need to handle bitfield here
1467  GString = CGF.Builder.CreateGlobalStringPtr(
1468  Format.concat(llvm::Twine('\n')).str());
1469  TmpRes = CGF.Builder.CreateCall(Func, {GString, FieldPtr});
1470  Res = CGF.Builder.CreateAdd(Res, TmpRes);
1471  }
1472 
1473  GString = CGF.Builder.CreateGlobalStringPtr(Pad + "}\n");
1474  Value *TmpRes = CGF.Builder.CreateCall(Func, {GString});
1475  Res = CGF.Builder.CreateAdd(Res, TmpRes);
1476  return Res;
1477 }
1478 
1479 static bool
1481  llvm::SmallPtrSetImpl<const Decl *> &Seen) {
1482  if (const auto *Arr = Ctx.getAsArrayType(Ty))
1483  Ty = Ctx.getBaseElementType(Arr);
1484 
1485  const auto *Record = Ty->getAsCXXRecordDecl();
1486  if (!Record)
1487  return false;
1488 
1489  // We've already checked this type, or are in the process of checking it.
1490  if (!Seen.insert(Record).second)
1491  return false;
1492 
1493  assert(Record->hasDefinition() &&
1494  "Incomplete types should already be diagnosed");
1495 
1496  if (Record->isDynamicClass())
1497  return true;
1498 
1499  for (FieldDecl *F : Record->fields()) {
1500  if (TypeRequiresBuiltinLaunderImp(Ctx, F->getType(), Seen))
1501  return true;
1502  }
1503  return false;
1504 }
1505 
1506 /// Determine if the specified type requires laundering by checking if it is a
1507 /// dynamic class type or contains a subobject which is a dynamic class type.
1509  if (!CGM.getCodeGenOpts().StrictVTablePointers)
1510  return false;
1511  llvm::SmallPtrSet<const Decl *, 16> Seen;
1512  return TypeRequiresBuiltinLaunderImp(CGM.getContext(), Ty, Seen);
1513 }
1514 
1515 RValue CodeGenFunction::emitRotate(const CallExpr *E, bool IsRotateRight) {
1516  llvm::Value *Src = EmitScalarExpr(E->getArg(0));
1517  llvm::Value *ShiftAmt = EmitScalarExpr(E->getArg(1));
1518 
1519  // The builtin's shift arg may have a different type than the source arg and
1520  // result, but the LLVM intrinsic uses the same type for all values.
1521  llvm::Type *Ty = Src->getType();
1522  ShiftAmt = Builder.CreateIntCast(ShiftAmt, Ty, false);
1523 
1524  // Rotate is a special case of LLVM funnel shift - 1st 2 args are the same.
1525  unsigned IID = IsRotateRight ? Intrinsic::fshr : Intrinsic::fshl;
1526  Function *F = CGM.getIntrinsic(IID, Ty);
1527  return RValue::get(Builder.CreateCall(F, { Src, Src, ShiftAmt }));
1528 }
1529 
1531  const CallExpr *E,
1532  ReturnValueSlot ReturnValue) {
1533  const FunctionDecl *FD = GD.getDecl()->getAsFunction();
1534  // See if we can constant fold this builtin. If so, don't emit it at all.
1535  Expr::EvalResult Result;
1536  if (E->EvaluateAsRValue(Result, CGM.getContext()) &&
1537  !Result.hasSideEffects()) {
1538  if (Result.Val.isInt())
1539  return RValue::get(llvm::ConstantInt::get(getLLVMContext(),
1540  Result.Val.getInt()));
1541  if (Result.Val.isFloat())
1542  return RValue::get(llvm::ConstantFP::get(getLLVMContext(),
1543  Result.Val.getFloat()));
1544  }
1545 
1546  // There are LLVM math intrinsics/instructions corresponding to math library
1547  // functions except the LLVM op will never set errno while the math library
1548  // might. Also, math builtins have the same semantics as their math library
1549  // twins. Thus, we can transform math library and builtin calls to their
1550  // LLVM counterparts if the call is marked 'const' (known to never set errno).
1551  if (FD->hasAttr<ConstAttr>()) {
1552  switch (BuiltinID) {
1553  case Builtin::BIceil:
1554  case Builtin::BIceilf:
1555  case Builtin::BIceill:
1556  case Builtin::BI__builtin_ceil:
1557  case Builtin::BI__builtin_ceilf:
1558  case Builtin::BI__builtin_ceill:
1559  return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::ceil));
1560 
1561  case Builtin::BIcopysign:
1562  case Builtin::BIcopysignf:
1563  case Builtin::BIcopysignl:
1564  case Builtin::BI__builtin_copysign:
1565  case Builtin::BI__builtin_copysignf:
1566  case Builtin::BI__builtin_copysignl:
1567  case Builtin::BI__builtin_copysignf128:
1569 
1570  case Builtin::BIcos:
1571  case Builtin::BIcosf:
1572  case Builtin::BIcosl:
1573  case Builtin::BI__builtin_cos:
1574  case Builtin::BI__builtin_cosf:
1575  case Builtin::BI__builtin_cosl:
1576  return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::cos));
1577 
1578  case Builtin::BIexp:
1579  case Builtin::BIexpf:
1580  case Builtin::BIexpl:
1581  case Builtin::BI__builtin_exp:
1582  case Builtin::BI__builtin_expf:
1583  case Builtin::BI__builtin_expl:
1584  return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::exp));
1585 
1586  case Builtin::BIexp2:
1587  case Builtin::BIexp2f:
1588  case Builtin::BIexp2l:
1589  case Builtin::BI__builtin_exp2:
1590  case Builtin::BI__builtin_exp2f:
1591  case Builtin::BI__builtin_exp2l:
1592  return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::exp2));
1593 
1594  case Builtin::BIfabs:
1595  case Builtin::BIfabsf:
1596  case Builtin::BIfabsl:
1597  case Builtin::BI__builtin_fabs:
1598  case Builtin::BI__builtin_fabsf:
1599  case Builtin::BI__builtin_fabsl:
1600  case Builtin::BI__builtin_fabsf128:
1601  return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::fabs));
1602 
1603  case Builtin::BIfloor:
1604  case Builtin::BIfloorf:
1605  case Builtin::BIfloorl:
1606  case Builtin::BI__builtin_floor:
1607  case Builtin::BI__builtin_floorf:
1608  case Builtin::BI__builtin_floorl:
1609  return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::floor));
1610 
1611  case Builtin::BIfma:
1612  case Builtin::BIfmaf:
1613  case Builtin::BIfmal:
1614  case Builtin::BI__builtin_fma:
1615  case Builtin::BI__builtin_fmaf:
1616  case Builtin::BI__builtin_fmal:
1617  return RValue::get(emitTernaryBuiltin(*this, E, Intrinsic::fma));
1618 
1619  case Builtin::BIfmax:
1620  case Builtin::BIfmaxf:
1621  case Builtin::BIfmaxl:
1622  case Builtin::BI__builtin_fmax:
1623  case Builtin::BI__builtin_fmaxf:
1624  case Builtin::BI__builtin_fmaxl:
1625  return RValue::get(emitBinaryBuiltin(*this, E, Intrinsic::maxnum));
1626 
1627  case Builtin::BIfmin:
1628  case Builtin::BIfminf:
1629  case Builtin::BIfminl:
1630  case Builtin::BI__builtin_fmin:
1631  case Builtin::BI__builtin_fminf:
1632  case Builtin::BI__builtin_fminl:
1633  return RValue::get(emitBinaryBuiltin(*this, E, Intrinsic::minnum));
1634 
1635  // fmod() is a special-case. It maps to the frem instruction rather than an
1636  // LLVM intrinsic.
1637  case Builtin::BIfmod:
1638  case Builtin::BIfmodf:
1639  case Builtin::BIfmodl:
1640  case Builtin::BI__builtin_fmod:
1641  case Builtin::BI__builtin_fmodf:
1642  case Builtin::BI__builtin_fmodl: {
1643  Value *Arg1 = EmitScalarExpr(E->getArg(0));
1644  Value *Arg2 = EmitScalarExpr(E->getArg(1));
1645  return RValue::get(Builder.CreateFRem(Arg1, Arg2, "fmod"));
1646  }
1647 
1648  case Builtin::BIlog:
1649  case Builtin::BIlogf:
1650  case Builtin::BIlogl:
1651  case Builtin::BI__builtin_log:
1652  case Builtin::BI__builtin_logf:
1653  case Builtin::BI__builtin_logl:
1654  return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::log));
1655 
1656  case Builtin::BIlog10:
1657  case Builtin::BIlog10f:
1658  case Builtin::BIlog10l:
1659  case Builtin::BI__builtin_log10:
1660  case Builtin::BI__builtin_log10f:
1661  case Builtin::BI__builtin_log10l:
1662  return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::log10));
1663 
1664  case Builtin::BIlog2:
1665  case Builtin::BIlog2f:
1666  case Builtin::BIlog2l:
1667  case Builtin::BI__builtin_log2:
1668  case Builtin::BI__builtin_log2f:
1669  case Builtin::BI__builtin_log2l:
1670  return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::log2));
1671 
1672  case Builtin::BInearbyint:
1673  case Builtin::BInearbyintf:
1674  case Builtin::BInearbyintl:
1675  case Builtin::BI__builtin_nearbyint:
1676  case Builtin::BI__builtin_nearbyintf:
1677  case Builtin::BI__builtin_nearbyintl:
1679 
1680  case Builtin::BIpow:
1681  case Builtin::BIpowf:
1682  case Builtin::BIpowl:
1683  case Builtin::BI__builtin_pow:
1684  case Builtin::BI__builtin_powf:
1685  case Builtin::BI__builtin_powl:
1686  return RValue::get(emitBinaryBuiltin(*this, E, Intrinsic::pow));
1687 
1688  case Builtin::BIrint:
1689  case Builtin::BIrintf:
1690  case Builtin::BIrintl:
1691  case Builtin::BI__builtin_rint:
1692  case Builtin::BI__builtin_rintf:
1693  case Builtin::BI__builtin_rintl:
1694  return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::rint));
1695 
1696  case Builtin::BIround:
1697  case Builtin::BIroundf:
1698  case Builtin::BIroundl:
1699  case Builtin::BI__builtin_round:
1700  case Builtin::BI__builtin_roundf:
1701  case Builtin::BI__builtin_roundl:
1702  return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::round));
1703 
1704  case Builtin::BIsin:
1705  case Builtin::BIsinf:
1706  case Builtin::BIsinl:
1707  case Builtin::BI__builtin_sin:
1708  case Builtin::BI__builtin_sinf:
1709  case Builtin::BI__builtin_sinl:
1710  return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::sin));
1711 
1712  case Builtin::BIsqrt:
1713  case Builtin::BIsqrtf:
1714  case Builtin::BIsqrtl:
1715  case Builtin::BI__builtin_sqrt:
1716  case Builtin::BI__builtin_sqrtf:
1717  case Builtin::BI__builtin_sqrtl:
1718  return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::sqrt));
1719 
1720  case Builtin::BItrunc:
1721  case Builtin::BItruncf:
1722  case Builtin::BItruncl:
1723  case Builtin::BI__builtin_trunc:
1724  case Builtin::BI__builtin_truncf:
1725  case Builtin::BI__builtin_truncl:
1726  return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::trunc));
1727 
1728  case Builtin::BIlround:
1729  case Builtin::BIlroundf:
1730  case Builtin::BIlroundl:
1731  case Builtin::BI__builtin_lround:
1732  case Builtin::BI__builtin_lroundf:
1733  case Builtin::BI__builtin_lroundl:
1735 
1736  case Builtin::BIllround:
1737  case Builtin::BIllroundf:
1738  case Builtin::BIllroundl:
1739  case Builtin::BI__builtin_llround:
1740  case Builtin::BI__builtin_llroundf:
1741  case Builtin::BI__builtin_llroundl:
1743 
1744  case Builtin::BIlrint:
1745  case Builtin::BIlrintf:
1746  case Builtin::BIlrintl:
1747  case Builtin::BI__builtin_lrint:
1748  case Builtin::BI__builtin_lrintf:
1749  case Builtin::BI__builtin_lrintl:
1751 
1752  case Builtin::BIllrint:
1753  case Builtin::BIllrintf:
1754  case Builtin::BIllrintl:
1755  case Builtin::BI__builtin_llrint:
1756  case Builtin::BI__builtin_llrintf:
1757  case Builtin::BI__builtin_llrintl:
1759 
1760  default:
1761  break;
1762  }
1763  }
1764 
1765  switch (BuiltinID) {
1766  default: break;
1767  case Builtin::BI__builtin___CFStringMakeConstantString:
1768  case Builtin::BI__builtin___NSStringMakeConstantString:
1769  return RValue::get(ConstantEmitter(*this).emitAbstract(E, E->getType()));
1770  case Builtin::BI__builtin_stdarg_start:
1771  case Builtin::BI__builtin_va_start:
1772  case Builtin::BI__va_start:
1773  case Builtin::BI__builtin_va_end:
1774  return RValue::get(
1775  EmitVAStartEnd(BuiltinID == Builtin::BI__va_start
1776  ? EmitScalarExpr(E->getArg(0))
1777  : EmitVAListRef(E->getArg(0)).getPointer(),
1778  BuiltinID != Builtin::BI__builtin_va_end));
1779  case Builtin::BI__builtin_va_copy: {
1780  Value *DstPtr = EmitVAListRef(E->getArg(0)).getPointer();
1781  Value *SrcPtr = EmitVAListRef(E->getArg(1)).getPointer();
1782 
1783  llvm::Type *Type = Int8PtrTy;
1784 
1785  DstPtr = Builder.CreateBitCast(DstPtr, Type);
1786  SrcPtr = Builder.CreateBitCast(SrcPtr, Type);
1787  return RValue::get(Builder.CreateCall(CGM.getIntrinsic(Intrinsic::vacopy),
1788  {DstPtr, SrcPtr}));
1789  }
1790  case Builtin::BI__builtin_abs:
1791  case Builtin::BI__builtin_labs:
1792  case Builtin::BI__builtin_llabs: {
1793  // X < 0 ? -X : X
1794  // The negation has 'nsw' because abs of INT_MIN is undefined.
1795  Value *ArgValue = EmitScalarExpr(E->getArg(0));
1796  Value *NegOp = Builder.CreateNSWNeg(ArgValue, "neg");
1797  Constant *Zero = llvm::Constant::getNullValue(ArgValue->getType());
1798  Value *CmpResult = Builder.CreateICmpSLT(ArgValue, Zero, "abscond");
1799  Value *Result = Builder.CreateSelect(CmpResult, NegOp, ArgValue, "abs");
1800  return RValue::get(Result);
1801  }
1802  case Builtin::BI__builtin_conj:
1803  case Builtin::BI__builtin_conjf:
1804  case Builtin::BI__builtin_conjl: {
1805  ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0));
1806  Value *Real = ComplexVal.first;
1807  Value *Imag = ComplexVal.second;
1808  Value *Zero =
1809  Imag->getType()->isFPOrFPVectorTy()
1810  ? llvm::ConstantFP::getZeroValueForNegation(Imag->getType())
1811  : llvm::Constant::getNullValue(Imag->getType());
1812 
1813  Imag = Builder.CreateFSub(Zero, Imag, "sub");
1814  return RValue::getComplex(std::make_pair(Real, Imag));
1815  }
1816  case Builtin::BI__builtin_creal:
1817  case Builtin::BI__builtin_crealf:
1818  case Builtin::BI__builtin_creall:
1819  case Builtin::BIcreal:
1820  case Builtin::BIcrealf:
1821  case Builtin::BIcreall: {
1822  ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0));
1823  return RValue::get(ComplexVal.first);
1824  }
1825 
1826  case Builtin::BI__builtin_dump_struct: {
1827  llvm::Type *LLVMIntTy = getTypes().ConvertType(getContext().IntTy);
1828  llvm::FunctionType *LLVMFuncType = llvm::FunctionType::get(
1829  LLVMIntTy, {llvm::Type::getInt8PtrTy(getLLVMContext())}, true);
1830 
1831  Value *Func = EmitScalarExpr(E->getArg(1)->IgnoreImpCasts());
1832  CharUnits Arg0Align = EmitPointerWithAlignment(E->getArg(0)).getAlignment();
1833 
1834  const Expr *Arg0 = E->getArg(0)->IgnoreImpCasts();
1835  QualType Arg0Type = Arg0->getType()->getPointeeType();
1836 
1837  Value *RecordPtr = EmitScalarExpr(Arg0);
1838  Value *Res = dumpRecord(*this, Arg0Type, RecordPtr, Arg0Align,
1839  {LLVMFuncType, Func}, 0);
1840  return RValue::get(Res);
1841  }
1842 
1843  case Builtin::BI__builtin_cimag:
1844  case Builtin::BI__builtin_cimagf:
1845  case Builtin::BI__builtin_cimagl:
1846  case Builtin::BIcimag:
1847  case Builtin::BIcimagf:
1848  case Builtin::BIcimagl: {
1849  ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0));
1850  return RValue::get(ComplexVal.second);
1851  }
1852 
1853  case Builtin::BI__builtin_clrsb:
1854  case Builtin::BI__builtin_clrsbl:
1855  case Builtin::BI__builtin_clrsbll: {
1856  // clrsb(x) -> clz(x < 0 ? ~x : x) - 1 or
1857  Value *ArgValue = EmitScalarExpr(E->getArg(0));
1858 
1859  llvm::Type *ArgType = ArgValue->getType();
1860  Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType);
1861 
1862  llvm::Type *ResultType = ConvertType(E->getType());
1863  Value *Zero = llvm::Constant::getNullValue(ArgType);
1864  Value *IsNeg = Builder.CreateICmpSLT(ArgValue, Zero, "isneg");
1865  Value *Inverse = Builder.CreateNot(ArgValue, "not");
1866  Value *Tmp = Builder.CreateSelect(IsNeg, Inverse, ArgValue);
1867  Value *Ctlz = Builder.CreateCall(F, {Tmp, Builder.getFalse()});
1868  Value *Result = Builder.CreateSub(Ctlz, llvm::ConstantInt::get(ArgType, 1));
1869  Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
1870  "cast");
1871  return RValue::get(Result);
1872  }
1873  case Builtin::BI__builtin_ctzs:
1874  case Builtin::BI__builtin_ctz:
1875  case Builtin::BI__builtin_ctzl:
1876  case Builtin::BI__builtin_ctzll: {
1877  Value *ArgValue = EmitCheckedArgForBuiltin(E->getArg(0), BCK_CTZPassedZero);
1878 
1879  llvm::Type *ArgType = ArgValue->getType();
1880  Function *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType);
1881 
1882  llvm::Type *ResultType = ConvertType(E->getType());
1883  Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef());
1884  Value *Result = Builder.CreateCall(F, {ArgValue, ZeroUndef});
1885  if (Result->getType() != ResultType)
1886  Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
1887  "cast");
1888  return RValue::get(Result);
1889  }
1890  case Builtin::BI__builtin_clzs:
1891  case Builtin::BI__builtin_clz:
1892  case Builtin::BI__builtin_clzl:
1893  case Builtin::BI__builtin_clzll: {
1894  Value *ArgValue = EmitCheckedArgForBuiltin(E->getArg(0), BCK_CLZPassedZero);
1895 
1896  llvm::Type *ArgType = ArgValue->getType();
1897  Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType);
1898 
1899  llvm::Type *ResultType = ConvertType(E->getType());
1900  Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef());
1901  Value *Result = Builder.CreateCall(F, {ArgValue, ZeroUndef});
1902  if (Result->getType() != ResultType)
1903  Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
1904  "cast");
1905  return RValue::get(Result);
1906  }
1907  case Builtin::BI__builtin_ffs:
1908  case Builtin::BI__builtin_ffsl:
1909  case Builtin::BI__builtin_ffsll: {
1910  // ffs(x) -> x ? cttz(x) + 1 : 0
1911  Value *ArgValue = EmitScalarExpr(E->getArg(0));
1912 
1913  llvm::Type *ArgType = ArgValue->getType();
1914  Function *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType);
1915 
1916  llvm::Type *ResultType = ConvertType(E->getType());
1917  Value *Tmp =
1918  Builder.CreateAdd(Builder.CreateCall(F, {ArgValue, Builder.getTrue()}),
1919  llvm::ConstantInt::get(ArgType, 1));
1920  Value *Zero = llvm::Constant::getNullValue(ArgType);
1921  Value *IsZero = Builder.CreateICmpEQ(ArgValue, Zero, "iszero");
1922  Value *Result = Builder.CreateSelect(IsZero, Zero, Tmp, "ffs");
1923  if (Result->getType() != ResultType)
1924  Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
1925  "cast");
1926  return RValue::get(Result);
1927  }
1928  case Builtin::BI__builtin_parity:
1929  case Builtin::BI__builtin_parityl:
1930  case Builtin::BI__builtin_parityll: {
1931  // parity(x) -> ctpop(x) & 1
1932  Value *ArgValue = EmitScalarExpr(E->getArg(0));
1933 
1934  llvm::Type *ArgType = ArgValue->getType();
1935  Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType);
1936 
1937  llvm::Type *ResultType = ConvertType(E->getType());
1938  Value *Tmp = Builder.CreateCall(F, ArgValue);
1939  Value *Result = Builder.CreateAnd(Tmp, llvm::ConstantInt::get(ArgType, 1));
1940  if (Result->getType() != ResultType)
1941  Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
1942  "cast");
1943  return RValue::get(Result);
1944  }
1945  case Builtin::BI__lzcnt16:
1946  case Builtin::BI__lzcnt:
1947  case Builtin::BI__lzcnt64: {
1948  Value *ArgValue = EmitScalarExpr(E->getArg(0));
1949 
1950  llvm::Type *ArgType = ArgValue->getType();
1951  Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType);
1952 
1953  llvm::Type *ResultType = ConvertType(E->getType());
1954  Value *Result = Builder.CreateCall(F, {ArgValue, Builder.getFalse()});
1955  if (Result->getType() != ResultType)
1956  Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
1957  "cast");
1958  return RValue::get(Result);
1959  }
1960  case Builtin::BI__popcnt16:
1961  case Builtin::BI__popcnt:
1962  case Builtin::BI__popcnt64:
1963  case Builtin::BI__builtin_popcount:
1964  case Builtin::BI__builtin_popcountl:
1965  case Builtin::BI__builtin_popcountll: {
1966  Value *ArgValue = EmitScalarExpr(E->getArg(0));
1967 
1968  llvm::Type *ArgType = ArgValue->getType();
1969  Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType);
1970 
1971  llvm::Type *ResultType = ConvertType(E->getType());
1972  Value *Result = Builder.CreateCall(F, ArgValue);
1973  if (Result->getType() != ResultType)
1974  Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
1975  "cast");
1976  return RValue::get(Result);
1977  }
1978  case Builtin::BI__builtin_unpredictable: {
1979  // Always return the argument of __builtin_unpredictable. LLVM does not
1980  // handle this builtin. Metadata for this builtin should be added directly
1981  // to instructions such as branches or switches that use it.
1982  return RValue::get(EmitScalarExpr(E->getArg(0)));
1983  }
1984  case Builtin::BI__builtin_expect: {
1985  Value *ArgValue = EmitScalarExpr(E->getArg(0));
1986  llvm::Type *ArgType = ArgValue->getType();
1987 
1988  Value *ExpectedValue = EmitScalarExpr(E->getArg(1));
1989  // Don't generate llvm.expect on -O0 as the backend won't use it for
1990  // anything.
1991  // Note, we still IRGen ExpectedValue because it could have side-effects.
1992  if (CGM.getCodeGenOpts().OptimizationLevel == 0)
1993  return RValue::get(ArgValue);
1994 
1995  Function *FnExpect = CGM.getIntrinsic(Intrinsic::expect, ArgType);
1996  Value *Result =
1997  Builder.CreateCall(FnExpect, {ArgValue, ExpectedValue}, "expval");
1998  return RValue::get(Result);
1999  }
2000  case Builtin::BI__builtin_assume_aligned: {
2001  const Expr *Ptr = E->getArg(0);
2002  Value *PtrValue = EmitScalarExpr(Ptr);
2003  Value *OffsetValue =
2004  (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : nullptr;
2005 
2006  Value *AlignmentValue = EmitScalarExpr(E->getArg(1));
2007  ConstantInt *AlignmentCI = cast<ConstantInt>(AlignmentValue);
2008  unsigned Alignment = (unsigned)AlignmentCI->getZExtValue();
2009 
2010  EmitAlignmentAssumption(PtrValue, Ptr,
2011  /*The expr loc is sufficient.*/ SourceLocation(),
2012  Alignment, OffsetValue);
2013  return RValue::get(PtrValue);
2014  }
2015  case Builtin::BI__assume:
2016  case Builtin::BI__builtin_assume: {
2017  if (E->getArg(0)->HasSideEffects(getContext()))
2018  return RValue::get(nullptr);
2019 
2020  Value *ArgValue = EmitScalarExpr(E->getArg(0));
2021  Function *FnAssume = CGM.getIntrinsic(Intrinsic::assume);
2022  return RValue::get(Builder.CreateCall(FnAssume, ArgValue));
2023  }
2024  case Builtin::BI__builtin_bswap16:
2025  case Builtin::BI__builtin_bswap32:
2026  case Builtin::BI__builtin_bswap64: {
2027  return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::bswap));
2028  }
2029  case Builtin::BI__builtin_bitreverse8:
2030  case Builtin::BI__builtin_bitreverse16:
2031  case Builtin::BI__builtin_bitreverse32:
2032  case Builtin::BI__builtin_bitreverse64: {
2033  return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::bitreverse));
2034  }
2035  case Builtin::BI__builtin_rotateleft8:
2036  case Builtin::BI__builtin_rotateleft16:
2037  case Builtin::BI__builtin_rotateleft32:
2038  case Builtin::BI__builtin_rotateleft64:
2039  case Builtin::BI_rotl8: // Microsoft variants of rotate left
2040  case Builtin::BI_rotl16:
2041  case Builtin::BI_rotl:
2042  case Builtin::BI_lrotl:
2043  case Builtin::BI_rotl64:
2044  return emitRotate(E, false);
2045 
2046  case Builtin::BI__builtin_rotateright8:
2047  case Builtin::BI__builtin_rotateright16:
2048  case Builtin::BI__builtin_rotateright32:
2049  case Builtin::BI__builtin_rotateright64:
2050  case Builtin::BI_rotr8: // Microsoft variants of rotate right
2051  case Builtin::BI_rotr16:
2052  case Builtin::BI_rotr:
2053  case Builtin::BI_lrotr:
2054  case Builtin::BI_rotr64:
2055  return emitRotate(E, true);
2056 
2057  case Builtin::BI__builtin_constant_p: {
2058  llvm::Type *ResultType = ConvertType(E->getType());
2059  if (CGM.getCodeGenOpts().OptimizationLevel == 0)
2060  // At -O0, we don't perform inlining, so we don't need to delay the
2061  // processing.
2062  return RValue::get(ConstantInt::get(ResultType, 0));
2063 
2064  const Expr *Arg = E->getArg(0);
2065  QualType ArgType = Arg->getType();
2066  // FIXME: The allowance for Obj-C pointers and block pointers is historical
2067  // and likely a mistake.
2068  if (!ArgType->isIntegralOrEnumerationType() && !ArgType->isFloatingType() &&
2069  !ArgType->isObjCObjectPointerType() && !ArgType->isBlockPointerType())
2070  // Per the GCC documentation, only numeric constants are recognized after
2071  // inlining.
2072  return RValue::get(ConstantInt::get(ResultType, 0));
2073 
2074  if (Arg->HasSideEffects(getContext()))
2075  // The argument is unevaluated, so be conservative if it might have
2076  // side-effects.
2077  return RValue::get(ConstantInt::get(ResultType, 0));
2078 
2079  Value *ArgValue = EmitScalarExpr(Arg);
2080  if (ArgType->isObjCObjectPointerType()) {
2081  // Convert Objective-C objects to id because we cannot distinguish between
2082  // LLVM types for Obj-C classes as they are opaque.
2083  ArgType = CGM.getContext().getObjCIdType();
2084  ArgValue = Builder.CreateBitCast(ArgValue, ConvertType(ArgType));
2085  }
2086  Function *F =
2087  CGM.getIntrinsic(Intrinsic::is_constant, ConvertType(ArgType));
2088  Value *Result = Builder.CreateCall(F, ArgValue);
2089  if (Result->getType() != ResultType)
2090  Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/false);
2091  return RValue::get(Result);
2092  }
2093  case Builtin::BI__builtin_dynamic_object_size:
2094  case Builtin::BI__builtin_object_size: {
2095  unsigned Type =
2096  E->getArg(1)->EvaluateKnownConstInt(getContext()).getZExtValue();
2097  auto *ResType = cast<llvm::IntegerType>(ConvertType(E->getType()));
2098 
2099  // We pass this builtin onto the optimizer so that it can figure out the
2100  // object size in more complex cases.
2101  bool IsDynamic = BuiltinID == Builtin::BI__builtin_dynamic_object_size;
2102  return RValue::get(emitBuiltinObjectSize(E->getArg(0), Type, ResType,
2103  /*EmittedE=*/nullptr, IsDynamic));
2104  }
2105  case Builtin::BI__builtin_prefetch: {
2106  Value *Locality, *RW, *Address = EmitScalarExpr(E->getArg(0));
2107  // FIXME: Technically these constants should of type 'int', yes?
2108  RW = (E->getNumArgs() > 1) ? EmitScalarExpr(E->getArg(1)) :
2109  llvm::ConstantInt::get(Int32Ty, 0);
2110  Locality = (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) :
2111  llvm::ConstantInt::get(Int32Ty, 3);
2112  Value *Data = llvm::ConstantInt::get(Int32Ty, 1);
2113  Function *F = CGM.getIntrinsic(Intrinsic::prefetch);
2114  return RValue::get(Builder.CreateCall(F, {Address, RW, Locality, Data}));
2115  }
2116  case Builtin::BI__builtin_readcyclecounter: {
2117  Function *F = CGM.getIntrinsic(Intrinsic::readcyclecounter);
2118  return RValue::get(Builder.CreateCall(F));
2119  }
2120  case Builtin::BI__builtin___clear_cache: {
2121  Value *Begin = EmitScalarExpr(E->getArg(0));
2122  Value *End = EmitScalarExpr(E->getArg(1));
2123  Function *F = CGM.getIntrinsic(Intrinsic::clear_cache);
2124  return RValue::get(Builder.CreateCall(F, {Begin, End}));
2125  }
2126  case Builtin::BI__builtin_trap:
2127  return RValue::get(EmitTrapCall(Intrinsic::trap));
2128  case Builtin::BI__debugbreak:
2129  return RValue::get(EmitTrapCall(Intrinsic::debugtrap));
2130  case Builtin::BI__builtin_unreachable: {
2131  EmitUnreachable(E->getExprLoc());
2132 
2133  // We do need to preserve an insertion point.
2134  EmitBlock(createBasicBlock("unreachable.cont"));
2135 
2136  return RValue::get(nullptr);
2137  }
2138 
2139  case Builtin::BI__builtin_powi:
2140  case Builtin::BI__builtin_powif:
2141  case Builtin::BI__builtin_powil: {
2142  Value *Base = EmitScalarExpr(E->getArg(0));
2143  Value *Exponent = EmitScalarExpr(E->getArg(1));
2144  llvm::Type *ArgType = Base->getType();
2145  Function *F = CGM.getIntrinsic(Intrinsic::powi, ArgType);
2146  return RValue::get(Builder.CreateCall(F, {Base, Exponent}));
2147  }
2148 
2149  case Builtin::BI__builtin_isgreater:
2150  case Builtin::BI__builtin_isgreaterequal:
2151  case Builtin::BI__builtin_isless:
2152  case Builtin::BI__builtin_islessequal:
2153  case Builtin::BI__builtin_islessgreater:
2154  case Builtin::BI__builtin_isunordered: {
2155  // Ordered comparisons: we know the arguments to these are matching scalar
2156  // floating point values.
2157  Value *LHS = EmitScalarExpr(E->getArg(0));
2158  Value *RHS = EmitScalarExpr(E->getArg(1));
2159 
2160  switch (BuiltinID) {
2161  default: llvm_unreachable("Unknown ordered comparison");
2162  case Builtin::BI__builtin_isgreater:
2163  LHS = Builder.CreateFCmpOGT(LHS, RHS, "cmp");
2164  break;
2165  case Builtin::BI__builtin_isgreaterequal:
2166  LHS = Builder.CreateFCmpOGE(LHS, RHS, "cmp");
2167  break;
2168  case Builtin::BI__builtin_isless:
2169  LHS = Builder.CreateFCmpOLT(LHS, RHS, "cmp");
2170  break;
2171  case Builtin::BI__builtin_islessequal:
2172  LHS = Builder.CreateFCmpOLE(LHS, RHS, "cmp");
2173  break;
2174  case Builtin::BI__builtin_islessgreater:
2175  LHS = Builder.CreateFCmpONE(LHS, RHS, "cmp");
2176  break;
2177  case Builtin::BI__builtin_isunordered:
2178  LHS = Builder.CreateFCmpUNO(LHS, RHS, "cmp");
2179  break;
2180  }
2181  // ZExt bool to int type.
2182  return RValue::get(Builder.CreateZExt(LHS, ConvertType(E->getType())));
2183  }
2184  case Builtin::BI__builtin_isnan: {
2185  Value *V = EmitScalarExpr(E->getArg(0));
2186  V = Builder.CreateFCmpUNO(V, V, "cmp");
2187  return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType())));
2188  }
2189 
2190  case Builtin::BIfinite:
2191  case Builtin::BI__finite:
2192  case Builtin::BIfinitef:
2193  case Builtin::BI__finitef:
2194  case Builtin::BIfinitel:
2195  case Builtin::BI__finitel:
2196  case Builtin::BI__builtin_isinf:
2197  case Builtin::BI__builtin_isfinite: {
2198  // isinf(x) --> fabs(x) == infinity
2199  // isfinite(x) --> fabs(x) != infinity
2200  // x != NaN via the ordered compare in either case.
2201  Value *V = EmitScalarExpr(E->getArg(0));
2202  Value *Fabs = EmitFAbs(*this, V);
2203  Constant *Infinity = ConstantFP::getInfinity(V->getType());
2204  CmpInst::Predicate Pred = (BuiltinID == Builtin::BI__builtin_isinf)
2205  ? CmpInst::FCMP_OEQ
2206  : CmpInst::FCMP_ONE;
2207  Value *FCmp = Builder.CreateFCmp(Pred, Fabs, Infinity, "cmpinf");
2208  return RValue::get(Builder.CreateZExt(FCmp, ConvertType(E->getType())));
2209  }
2210 
2211  case Builtin::BI__builtin_isinf_sign: {
2212  // isinf_sign(x) -> fabs(x) == infinity ? (signbit(x) ? -1 : 1) : 0
2213  Value *Arg = EmitScalarExpr(E->getArg(0));
2214  Value *AbsArg = EmitFAbs(*this, Arg);
2215  Value *IsInf = Builder.CreateFCmpOEQ(
2216  AbsArg, ConstantFP::getInfinity(Arg->getType()), "isinf");
2217  Value *IsNeg = EmitSignBit(*this, Arg);
2218 
2219  llvm::Type *IntTy = ConvertType(E->getType());
2220  Value *Zero = Constant::getNullValue(IntTy);
2221  Value *One = ConstantInt::get(IntTy, 1);
2222  Value *NegativeOne = ConstantInt::get(IntTy, -1);
2223  Value *SignResult = Builder.CreateSelect(IsNeg, NegativeOne, One);
2224  Value *Result = Builder.CreateSelect(IsInf, SignResult, Zero);
2225  return RValue::get(Result);
2226  }
2227 
2228  case Builtin::BI__builtin_isnormal: {
2229  // isnormal(x) --> x == x && fabsf(x) < infinity && fabsf(x) >= float_min
2230  Value *V = EmitScalarExpr(E->getArg(0));
2231  Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq");
2232 
2233  Value *Abs = EmitFAbs(*this, V);
2234  Value *IsLessThanInf =
2235  Builder.CreateFCmpULT(Abs, ConstantFP::getInfinity(V->getType()),"isinf");
2236  APFloat Smallest = APFloat::getSmallestNormalized(
2237  getContext().getFloatTypeSemantics(E->getArg(0)->getType()));
2238  Value *IsNormal =
2239  Builder.CreateFCmpUGE(Abs, ConstantFP::get(V->getContext(), Smallest),
2240  "isnormal");
2241  V = Builder.CreateAnd(Eq, IsLessThanInf, "and");
2242  V = Builder.CreateAnd(V, IsNormal, "and");
2243  return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType())));
2244  }
2245 
2246  case Builtin::BI__builtin_flt_rounds: {
2247  Function *F = CGM.getIntrinsic(Intrinsic::flt_rounds);
2248 
2249  llvm::Type *ResultType = ConvertType(E->getType());
2250  Value *Result = Builder.CreateCall(F);
2251  if (Result->getType() != ResultType)
2252  Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
2253  "cast");
2254  return RValue::get(Result);
2255  }
2256 
2257  case Builtin::BI__builtin_fpclassify: {
2258  Value *V = EmitScalarExpr(E->getArg(5));
2259  llvm::Type *Ty = ConvertType(E->getArg(5)->getType());
2260 
2261  // Create Result
2262  BasicBlock *Begin = Builder.GetInsertBlock();
2263  BasicBlock *End = createBasicBlock("fpclassify_end", this->CurFn);
2264  Builder.SetInsertPoint(End);
2265  PHINode *Result =
2266  Builder.CreatePHI(ConvertType(E->getArg(0)->getType()), 4,
2267  "fpclassify_result");
2268 
2269  // if (V==0) return FP_ZERO
2270  Builder.SetInsertPoint(Begin);
2271  Value *IsZero = Builder.CreateFCmpOEQ(V, Constant::getNullValue(Ty),
2272  "iszero");
2273  Value *ZeroLiteral = EmitScalarExpr(E->getArg(4));
2274  BasicBlock *NotZero = createBasicBlock("fpclassify_not_zero", this->CurFn);
2275  Builder.CreateCondBr(IsZero, End, NotZero);
2276  Result->addIncoming(ZeroLiteral, Begin);
2277 
2278  // if (V != V) return FP_NAN
2279  Builder.SetInsertPoint(NotZero);
2280  Value *IsNan = Builder.CreateFCmpUNO(V, V, "cmp");
2281  Value *NanLiteral = EmitScalarExpr(E->getArg(0));
2282  BasicBlock *NotNan = createBasicBlock("fpclassify_not_nan", this->CurFn);
2283  Builder.CreateCondBr(IsNan, End, NotNan);
2284  Result->addIncoming(NanLiteral, NotZero);
2285 
2286  // if (fabs(V) == infinity) return FP_INFINITY
2287  Builder.SetInsertPoint(NotNan);
2288  Value *VAbs = EmitFAbs(*this, V);
2289  Value *IsInf =
2290  Builder.CreateFCmpOEQ(VAbs, ConstantFP::getInfinity(V->getType()),
2291  "isinf");
2292  Value *InfLiteral = EmitScalarExpr(E->getArg(1));
2293  BasicBlock *NotInf = createBasicBlock("fpclassify_not_inf", this->CurFn);
2294  Builder.CreateCondBr(IsInf, End, NotInf);
2295  Result->addIncoming(InfLiteral, NotNan);
2296 
2297  // if (fabs(V) >= MIN_NORMAL) return FP_NORMAL else FP_SUBNORMAL
2298  Builder.SetInsertPoint(NotInf);
2299  APFloat Smallest = APFloat::getSmallestNormalized(
2300  getContext().getFloatTypeSemantics(E->getArg(5)->getType()));
2301  Value *IsNormal =
2302  Builder.CreateFCmpUGE(VAbs, ConstantFP::get(V->getContext(), Smallest),
2303  "isnormal");
2304  Value *NormalResult =
2305  Builder.CreateSelect(IsNormal, EmitScalarExpr(E->getArg(2)),
2306  EmitScalarExpr(E->getArg(3)));
2307  Builder.CreateBr(End);
2308  Result->addIncoming(NormalResult, NotInf);
2309 
2310  // return Result
2311  Builder.SetInsertPoint(End);
2312  return RValue::get(Result);
2313  }
2314 
2315  case Builtin::BIalloca:
2316  case Builtin::BI_alloca:
2317  case Builtin::BI__builtin_alloca: {
2318  Value *Size = EmitScalarExpr(E->getArg(0));
2319  const TargetInfo &TI = getContext().getTargetInfo();
2320  // The alignment of the alloca should correspond to __BIGGEST_ALIGNMENT__.
2321  unsigned SuitableAlignmentInBytes =
2322  CGM.getContext()
2323  .toCharUnitsFromBits(TI.getSuitableAlign())
2324  .getQuantity();
2325  AllocaInst *AI = Builder.CreateAlloca(Builder.getInt8Ty(), Size);
2326  AI->setAlignment(SuitableAlignmentInBytes);
2327  initializeAlloca(*this, AI, Size, SuitableAlignmentInBytes);
2328  return RValue::get(AI);
2329  }
2330 
2331  case Builtin::BI__builtin_alloca_with_align: {
2332  Value *Size = EmitScalarExpr(E->getArg(0));
2333  Value *AlignmentInBitsValue = EmitScalarExpr(E->getArg(1));
2334  auto *AlignmentInBitsCI = cast<ConstantInt>(AlignmentInBitsValue);
2335  unsigned AlignmentInBits = AlignmentInBitsCI->getZExtValue();
2336  unsigned AlignmentInBytes =
2337  CGM.getContext().toCharUnitsFromBits(AlignmentInBits).getQuantity();
2338  AllocaInst *AI = Builder.CreateAlloca(Builder.getInt8Ty(), Size);
2339  AI->setAlignment(AlignmentInBytes);
2340  initializeAlloca(*this, AI, Size, AlignmentInBytes);
2341  return RValue::get(AI);
2342  }
2343 
2344  case Builtin::BIbzero:
2345  case Builtin::BI__builtin_bzero: {
2346  Address Dest = EmitPointerWithAlignment(E->getArg(0));
2347  Value *SizeVal = EmitScalarExpr(E->getArg(1));
2348  EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(),
2349  E->getArg(0)->getExprLoc(), FD, 0);
2350  Builder.CreateMemSet(Dest, Builder.getInt8(0), SizeVal, false);
2351  return RValue::get(nullptr);
2352  }
2353  case Builtin::BImemcpy:
2354  case Builtin::BI__builtin_memcpy: {
2355  Address Dest = EmitPointerWithAlignment(E->getArg(0));
2356  Address Src = EmitPointerWithAlignment(E->getArg(1));
2357  Value *SizeVal = EmitScalarExpr(E->getArg(2));
2358  EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(),
2359  E->getArg(0)->getExprLoc(), FD, 0);
2360  EmitNonNullArgCheck(RValue::get(Src.getPointer()), E->getArg(1)->getType(),
2361  E->getArg(1)->getExprLoc(), FD, 1);
2362  Builder.CreateMemCpy(Dest, Src, SizeVal, false);
2363  return RValue::get(Dest.getPointer());
2364  }
2365 
2366  case Builtin::BI__builtin_char_memchr:
2367  BuiltinID = Builtin::BI__builtin_memchr;
2368  break;
2369 
2370  case Builtin::BI__builtin___memcpy_chk: {
2371  // fold __builtin_memcpy_chk(x, y, cst1, cst2) to memcpy iff cst1<=cst2.
2372  Expr::EvalResult SizeResult, DstSizeResult;
2373  if (!E->getArg(2)->EvaluateAsInt(SizeResult, CGM.getContext()) ||
2374  !E->getArg(3)->EvaluateAsInt(DstSizeResult, CGM.getContext()))
2375  break;
2376  llvm::APSInt Size = SizeResult.Val.getInt();
2377  llvm::APSInt DstSize = DstSizeResult.Val.getInt();
2378  if (Size.ugt(DstSize))
2379  break;
2380  Address Dest = EmitPointerWithAlignment(E->getArg(0));
2381  Address Src = EmitPointerWithAlignment(E->getArg(1));
2382  Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size);
2383  Builder.CreateMemCpy(Dest, Src, SizeVal, false);
2384  return RValue::get(Dest.getPointer());
2385  }
2386 
2387  case Builtin::BI__builtin_objc_memmove_collectable: {
2388  Address DestAddr = EmitPointerWithAlignment(E->getArg(0));
2389  Address SrcAddr = EmitPointerWithAlignment(E->getArg(1));
2390  Value *SizeVal = EmitScalarExpr(E->getArg(2));
2391  CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this,
2392  DestAddr, SrcAddr, SizeVal);
2393  return RValue::get(DestAddr.getPointer());
2394  }
2395 
2396  case Builtin::BI__builtin___memmove_chk: {
2397  // fold __builtin_memmove_chk(x, y, cst1, cst2) to memmove iff cst1<=cst2.
2398  Expr::EvalResult SizeResult, DstSizeResult;
2399  if (!E->getArg(2)->EvaluateAsInt(SizeResult, CGM.getContext()) ||
2400  !E->getArg(3)->EvaluateAsInt(DstSizeResult, CGM.getContext()))
2401  break;
2402  llvm::APSInt Size = SizeResult.Val.getInt();
2403  llvm::APSInt DstSize = DstSizeResult.Val.getInt();
2404  if (Size.ugt(DstSize))
2405  break;
2406  Address Dest = EmitPointerWithAlignment(E->getArg(0));
2407  Address Src = EmitPointerWithAlignment(E->getArg(1));
2408  Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size);
2409  Builder.CreateMemMove(Dest, Src, SizeVal, false);
2410  return RValue::get(Dest.getPointer());
2411  }
2412 
2413  case Builtin::BImemmove:
2414  case Builtin::BI__builtin_memmove: {
2415  Address Dest = EmitPointerWithAlignment(E->getArg(0));
2416  Address Src = EmitPointerWithAlignment(E->getArg(1));
2417  Value *SizeVal = EmitScalarExpr(E->getArg(2));
2418  EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(),
2419  E->getArg(0)->getExprLoc(), FD, 0);
2420  EmitNonNullArgCheck(RValue::get(Src.getPointer()), E->getArg(1)->getType(),
2421  E->getArg(1)->getExprLoc(), FD, 1);
2422  Builder.CreateMemMove(Dest, Src, SizeVal, false);
2423  return RValue::get(Dest.getPointer());
2424  }
2425  case Builtin::BImemset:
2426  case Builtin::BI__builtin_memset: {
2427  Address Dest = EmitPointerWithAlignment(E->getArg(0));
2428  Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)),
2429  Builder.getInt8Ty());
2430  Value *SizeVal = EmitScalarExpr(E->getArg(2));
2431  EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(),
2432  E->getArg(0)->getExprLoc(), FD, 0);
2433  Builder.CreateMemSet(Dest, ByteVal, SizeVal, false);
2434  return RValue::get(Dest.getPointer());
2435  }
2436  case Builtin::BI__builtin___memset_chk: {
2437  // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2.
2438  Expr::EvalResult SizeResult, DstSizeResult;
2439  if (!E->getArg(2)->EvaluateAsInt(SizeResult, CGM.getContext()) ||
2440  !E->getArg(3)->EvaluateAsInt(DstSizeResult, CGM.getContext()))
2441  break;
2442  llvm::APSInt Size = SizeResult.Val.getInt();
2443  llvm::APSInt DstSize = DstSizeResult.Val.getInt();
2444  if (Size.ugt(DstSize))
2445  break;
2446  Address Dest = EmitPointerWithAlignment(E->getArg(0));
2447  Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)),
2448  Builder.getInt8Ty());
2449  Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size);
2450  Builder.CreateMemSet(Dest, ByteVal, SizeVal, false);
2451  return RValue::get(Dest.getPointer());
2452  }
2453  case Builtin::BI__builtin_wmemcmp: {
2454  // The MSVC runtime library does not provide a definition of wmemcmp, so we
2455  // need an inline implementation.
2456  if (!getTarget().getTriple().isOSMSVCRT())
2457  break;
2458 
2459  llvm::Type *WCharTy = ConvertType(getContext().WCharTy);
2460 
2461  Value *Dst = EmitScalarExpr(E->getArg(0));
2462  Value *Src = EmitScalarExpr(E->getArg(1));
2463  Value *Size = EmitScalarExpr(E->getArg(2));
2464 
2465  BasicBlock *Entry = Builder.GetInsertBlock();
2466  BasicBlock *CmpGT = createBasicBlock("wmemcmp.gt");
2467  BasicBlock *CmpLT = createBasicBlock("wmemcmp.lt");
2468  BasicBlock *Next = createBasicBlock("wmemcmp.next");
2469  BasicBlock *Exit = createBasicBlock("wmemcmp.exit");
2470  Value *SizeEq0 = Builder.CreateICmpEQ(Size, ConstantInt::get(SizeTy, 0));
2471  Builder.CreateCondBr(SizeEq0, Exit, CmpGT);
2472 
2473  EmitBlock(CmpGT);
2474  PHINode *DstPhi = Builder.CreatePHI(Dst->getType(), 2);
2475  DstPhi->addIncoming(Dst, Entry);
2476  PHINode *SrcPhi = Builder.CreatePHI(Src->getType(), 2);
2477  SrcPhi->addIncoming(Src, Entry);
2478  PHINode *SizePhi = Builder.CreatePHI(SizeTy, 2);
2479  SizePhi->addIncoming(Size, Entry);
2480  CharUnits WCharAlign =
2481  getContext().getTypeAlignInChars(getContext().WCharTy);
2482  Value *DstCh = Builder.CreateAlignedLoad(WCharTy, DstPhi, WCharAlign);
2483  Value *SrcCh = Builder.CreateAlignedLoad(WCharTy, SrcPhi, WCharAlign);
2484  Value *DstGtSrc = Builder.CreateICmpUGT(DstCh, SrcCh);
2485  Builder.CreateCondBr(DstGtSrc, Exit, CmpLT);
2486 
2487  EmitBlock(CmpLT);
2488  Value *DstLtSrc = Builder.CreateICmpULT(DstCh, SrcCh);
2489  Builder.CreateCondBr(DstLtSrc, Exit, Next);
2490 
2491  EmitBlock(Next);
2492  Value *NextDst = Builder.CreateConstInBoundsGEP1_32(WCharTy, DstPhi, 1);
2493  Value *NextSrc = Builder.CreateConstInBoundsGEP1_32(WCharTy, SrcPhi, 1);
2494  Value *NextSize = Builder.CreateSub(SizePhi, ConstantInt::get(SizeTy, 1));
2495  Value *NextSizeEq0 =
2496  Builder.CreateICmpEQ(NextSize, ConstantInt::get(SizeTy, 0));
2497  Builder.CreateCondBr(NextSizeEq0, Exit, CmpGT);
2498  DstPhi->addIncoming(NextDst, Next);
2499  SrcPhi->addIncoming(NextSrc, Next);
2500  SizePhi->addIncoming(NextSize, Next);
2501 
2502  EmitBlock(Exit);
2503  PHINode *Ret = Builder.CreatePHI(IntTy, 4);
2504  Ret->addIncoming(ConstantInt::get(IntTy, 0), Entry);
2505  Ret->addIncoming(ConstantInt::get(IntTy, 1), CmpGT);
2506  Ret->addIncoming(ConstantInt::get(IntTy, -1), CmpLT);
2507  Ret->addIncoming(ConstantInt::get(IntTy, 0), Next);
2508  return RValue::get(Ret);
2509  }
2510  case Builtin::BI__builtin_dwarf_cfa: {
2511  // The offset in bytes from the first argument to the CFA.
2512  //
2513  // Why on earth is this in the frontend? Is there any reason at
2514  // all that the backend can't reasonably determine this while
2515  // lowering llvm.eh.dwarf.cfa()?
2516  //
2517  // TODO: If there's a satisfactory reason, add a target hook for
2518  // this instead of hard-coding 0, which is correct for most targets.
2519  int32_t Offset = 0;
2520 
2521  Function *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa);
2522  return RValue::get(Builder.CreateCall(F,
2523  llvm::ConstantInt::get(Int32Ty, Offset)));
2524  }
2525  case Builtin::BI__builtin_return_address: {
2526  Value *Depth = ConstantEmitter(*this).emitAbstract(E->getArg(0),
2527  getContext().UnsignedIntTy);
2528  Function *F = CGM.getIntrinsic(Intrinsic::returnaddress);
2529  return RValue::get(Builder.CreateCall(F, Depth));
2530  }
2531  case Builtin::BI_ReturnAddress: {
2532  Function *F = CGM.getIntrinsic(Intrinsic::returnaddress);
2533  return RValue::get(Builder.CreateCall(F, Builder.getInt32(0)));
2534  }
2535  case Builtin::BI__builtin_frame_address: {
2536  Value *Depth = ConstantEmitter(*this).emitAbstract(E->getArg(0),
2537  getContext().UnsignedIntTy);
2538  Function *F = CGM.getIntrinsic(Intrinsic::frameaddress);
2539  return RValue::get(Builder.CreateCall(F, Depth));
2540  }
2541  case Builtin::BI__builtin_extract_return_addr: {
2542  Value *Address = EmitScalarExpr(E->getArg(0));
2543  Value *Result = getTargetHooks().decodeReturnAddress(*this, Address);
2544  return RValue::get(Result);
2545  }
2546  case Builtin::BI__builtin_frob_return_addr: {
2547  Value *Address = EmitScalarExpr(E->getArg(0));
2548  Value *Result = getTargetHooks().encodeReturnAddress(*this, Address);
2549  return RValue::get(Result);
2550  }
2551  case Builtin::BI__builtin_dwarf_sp_column: {
2552  llvm::IntegerType *Ty
2553  = cast<llvm::IntegerType>(ConvertType(E->getType()));
2554  int Column = getTargetHooks().getDwarfEHStackPointer(CGM);
2555  if (Column == -1) {
2556  CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column");
2557  return RValue::get(llvm::UndefValue::get(Ty));
2558  }
2559  return RValue::get(llvm::ConstantInt::get(Ty, Column, true));
2560  }
2561  case Builtin::BI__builtin_init_dwarf_reg_size_table: {
2562  Value *Address = EmitScalarExpr(E->getArg(0));
2563  if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address))
2564  CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table");
2565  return RValue::get(llvm::UndefValue::get(ConvertType(E->getType())));
2566  }
2567  case Builtin::BI__builtin_eh_return: {
2568  Value *Int = EmitScalarExpr(E->getArg(0));
2569  Value *Ptr = EmitScalarExpr(E->getArg(1));
2570 
2571  llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType());
2572  assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) &&
2573  "LLVM's __builtin_eh_return only supports 32- and 64-bit variants");
2574  Function *F =
2575  CGM.getIntrinsic(IntTy->getBitWidth() == 32 ? Intrinsic::eh_return_i32
2576  : Intrinsic::eh_return_i64);
2577  Builder.CreateCall(F, {Int, Ptr});
2578  Builder.CreateUnreachable();
2579 
2580  // We do need to preserve an insertion point.
2581  EmitBlock(createBasicBlock("builtin_eh_return.cont"));
2582 
2583  return RValue::get(nullptr);
2584  }
2585  case Builtin::BI__builtin_unwind_init: {
2586  Function *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init);
2587  return RValue::get(Builder.CreateCall(F));
2588  }
2589  case Builtin::BI__builtin_extend_pointer: {
2590  // Extends a pointer to the size of an _Unwind_Word, which is
2591  // uint64_t on all platforms. Generally this gets poked into a
2592  // register and eventually used as an address, so if the
2593  // addressing registers are wider than pointers and the platform
2594  // doesn't implicitly ignore high-order bits when doing
2595  // addressing, we need to make sure we zext / sext based on
2596  // the platform's expectations.
2597  //
2598  // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html
2599 
2600  // Cast the pointer to intptr_t.
2601  Value *Ptr = EmitScalarExpr(E->getArg(0));
2602  Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast");
2603 
2604  // If that's 64 bits, we're done.
2605  if (IntPtrTy->getBitWidth() == 64)
2606  return RValue::get(Result);
2607 
2608  // Otherwise, ask the codegen data what to do.
2609  if (getTargetHooks().extendPointerWithSExt())
2610  return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext"));
2611  else
2612  return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext"));
2613  }
2614  case Builtin::BI__builtin_setjmp: {
2615  // Buffer is a void**.
2616  Address Buf = EmitPointerWithAlignment(E->getArg(0));
2617 
2618  // Store the frame pointer to the setjmp buffer.
2619  Value *FrameAddr =
2620  Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress),
2621  ConstantInt::get(Int32Ty, 0));
2622  Builder.CreateStore(FrameAddr, Buf);
2623 
2624  // Store the stack pointer to the setjmp buffer.
2625  Value *StackAddr =
2626  Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave));
2627  Address StackSaveSlot = Builder.CreateConstInBoundsGEP(Buf, 2);
2628  Builder.CreateStore(StackAddr, StackSaveSlot);
2629 
2630  // Call LLVM's EH setjmp, which is lightweight.
2631  Function *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp);
2632  Buf = Builder.CreateBitCast(Buf, Int8PtrTy);
2633  return RValue::get(Builder.CreateCall(F, Buf.getPointer()));
2634  }
2635  case Builtin::BI__builtin_longjmp: {
2636  Value *Buf = EmitScalarExpr(E->getArg(0));
2637  Buf = Builder.CreateBitCast(Buf, Int8PtrTy);
2638 
2639  // Call LLVM's EH longjmp, which is lightweight.
2640  Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf);
2641 
2642  // longjmp doesn't return; mark this as unreachable.
2643  Builder.CreateUnreachable();
2644 
2645  // We do need to preserve an insertion point.
2646  EmitBlock(createBasicBlock("longjmp.cont"));
2647 
2648  return RValue::get(nullptr);
2649  }
2650  case Builtin::BI__builtin_launder: {
2651  const Expr *Arg = E->getArg(0);
2652  QualType ArgTy = Arg->getType()->getPointeeType();
2653  Value *Ptr = EmitScalarExpr(Arg);
2654  if (TypeRequiresBuiltinLaunder(CGM, ArgTy))
2655  Ptr = Builder.CreateLaunderInvariantGroup(Ptr);
2656 
2657  return RValue::get(Ptr);
2658  }
2659  case Builtin::BI__sync_fetch_and_add:
2660  case Builtin::BI__sync_fetch_and_sub:
2661  case Builtin::BI__sync_fetch_and_or:
2662  case Builtin::BI__sync_fetch_and_and:
2663  case Builtin::BI__sync_fetch_and_xor:
2664  case Builtin::BI__sync_fetch_and_nand:
2665  case Builtin::BI__sync_add_and_fetch:
2666  case Builtin::BI__sync_sub_and_fetch:
2667  case Builtin::BI__sync_and_and_fetch:
2668  case Builtin::BI__sync_or_and_fetch:
2669  case Builtin::BI__sync_xor_and_fetch:
2670  case Builtin::BI__sync_nand_and_fetch:
2671  case Builtin::BI__sync_val_compare_and_swap:
2672  case Builtin::BI__sync_bool_compare_and_swap:
2673  case Builtin::BI__sync_lock_test_and_set:
2674  case Builtin::BI__sync_lock_release:
2675  case Builtin::BI__sync_swap:
2676  llvm_unreachable("Shouldn't make it through sema");
2677  case Builtin::BI__sync_fetch_and_add_1:
2678  case Builtin::BI__sync_fetch_and_add_2:
2679  case Builtin::BI__sync_fetch_and_add_4:
2680  case Builtin::BI__sync_fetch_and_add_8:
2681  case Builtin::BI__sync_fetch_and_add_16:
2682  return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E);
2683  case Builtin::BI__sync_fetch_and_sub_1:
2684  case Builtin::BI__sync_fetch_and_sub_2:
2685  case Builtin::BI__sync_fetch_and_sub_4:
2686  case Builtin::BI__sync_fetch_and_sub_8:
2687  case Builtin::BI__sync_fetch_and_sub_16:
2688  return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E);
2689  case Builtin::BI__sync_fetch_and_or_1:
2690  case Builtin::BI__sync_fetch_and_or_2:
2691  case Builtin::BI__sync_fetch_and_or_4:
2692  case Builtin::BI__sync_fetch_and_or_8:
2693  case Builtin::BI__sync_fetch_and_or_16:
2694  return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E);
2695  case Builtin::BI__sync_fetch_and_and_1:
2696  case Builtin::BI__sync_fetch_and_and_2:
2697  case Builtin::BI__sync_fetch_and_and_4:
2698  case Builtin::BI__sync_fetch_and_and_8:
2699  case Builtin::BI__sync_fetch_and_and_16:
2700  return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E);
2701  case Builtin::BI__sync_fetch_and_xor_1:
2702  case Builtin::BI__sync_fetch_and_xor_2:
2703  case Builtin::BI__sync_fetch_and_xor_4:
2704  case Builtin::BI__sync_fetch_and_xor_8:
2705  case Builtin::BI__sync_fetch_and_xor_16:
2706  return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E);
2707  case Builtin::BI__sync_fetch_and_nand_1:
2708  case Builtin::BI__sync_fetch_and_nand_2:
2709  case Builtin::BI__sync_fetch_and_nand_4:
2710  case Builtin::BI__sync_fetch_and_nand_8:
2711  case Builtin::BI__sync_fetch_and_nand_16:
2712  return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Nand, E);
2713 
2714  // Clang extensions: not overloaded yet.
2715  case Builtin::BI__sync_fetch_and_min:
2716  return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E);
2717  case Builtin::BI__sync_fetch_and_max:
2718  return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E);
2719  case Builtin::BI__sync_fetch_and_umin:
2720  return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E);
2721  case Builtin::BI__sync_fetch_and_umax:
2722  return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E);
2723 
2724  case Builtin::BI__sync_add_and_fetch_1:
2725  case Builtin::BI__sync_add_and_fetch_2:
2726  case Builtin::BI__sync_add_and_fetch_4:
2727  case Builtin::BI__sync_add_and_fetch_8:
2728  case Builtin::BI__sync_add_and_fetch_16:
2729  return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E,
2730  llvm::Instruction::Add);
2731  case Builtin::BI__sync_sub_and_fetch_1:
2732  case Builtin::BI__sync_sub_and_fetch_2:
2733  case Builtin::BI__sync_sub_and_fetch_4:
2734  case Builtin::BI__sync_sub_and_fetch_8:
2735  case Builtin::BI__sync_sub_and_fetch_16:
2736  return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E,
2737  llvm::Instruction::Sub);
2738  case Builtin::BI__sync_and_and_fetch_1:
2739  case Builtin::BI__sync_and_and_fetch_2:
2740  case Builtin::BI__sync_and_and_fetch_4:
2741  case Builtin::BI__sync_and_and_fetch_8:
2742  case Builtin::BI__sync_and_and_fetch_16:
2745  case Builtin::BI__sync_or_and_fetch_1:
2746  case Builtin::BI__sync_or_and_fetch_2:
2747  case Builtin::BI__sync_or_and_fetch_4:
2748  case Builtin::BI__sync_or_and_fetch_8:
2749  case Builtin::BI__sync_or_and_fetch_16:
2750  return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E,
2751  llvm::Instruction::Or);
2752  case Builtin::BI__sync_xor_and_fetch_1:
2753  case Builtin::BI__sync_xor_and_fetch_2:
2754  case Builtin::BI__sync_xor_and_fetch_4:
2755  case Builtin::BI__sync_xor_and_fetch_8:
2756  case Builtin::BI__sync_xor_and_fetch_16:
2757  return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E,
2758  llvm::Instruction::Xor);
2759  case Builtin::BI__sync_nand_and_fetch_1:
2760  case Builtin::BI__sync_nand_and_fetch_2:
2761  case Builtin::BI__sync_nand_and_fetch_4:
2762  case Builtin::BI__sync_nand_and_fetch_8:
2763  case Builtin::BI__sync_nand_and_fetch_16:
2764  return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Nand, E,
2765  llvm::Instruction::And, true);
2766 
2767  case Builtin::BI__sync_val_compare_and_swap_1:
2768  case Builtin::BI__sync_val_compare_and_swap_2:
2769  case Builtin::BI__sync_val_compare_and_swap_4:
2770  case Builtin::BI__sync_val_compare_and_swap_8:
2771  case Builtin::BI__sync_val_compare_and_swap_16:
2772  return RValue::get(MakeAtomicCmpXchgValue(*this, E, false));
2773 
2774  case Builtin::BI__sync_bool_compare_and_swap_1:
2775  case Builtin::BI__sync_bool_compare_and_swap_2:
2776  case Builtin::BI__sync_bool_compare_and_swap_4:
2777  case Builtin::BI__sync_bool_compare_and_swap_8:
2778  case Builtin::BI__sync_bool_compare_and_swap_16:
2779  return RValue::get(MakeAtomicCmpXchgValue(*this, E, true));
2780 
2781  case Builtin::BI__sync_swap_1:
2782  case Builtin::BI__sync_swap_2:
2783  case Builtin::BI__sync_swap_4:
2784  case Builtin::BI__sync_swap_8:
2785  case Builtin::BI__sync_swap_16:
2786  return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E);
2787 
2788  case Builtin::BI__sync_lock_test_and_set_1:
2789  case Builtin::BI__sync_lock_test_and_set_2:
2790  case Builtin::BI__sync_lock_test_and_set_4:
2791  case Builtin::BI__sync_lock_test_and_set_8:
2792  case Builtin::BI__sync_lock_test_and_set_16:
2793  return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E);
2794 
2795  case Builtin::BI__sync_lock_release_1:
2796  case Builtin::BI__sync_lock_release_2:
2797  case Builtin::BI__sync_lock_release_4:
2798  case Builtin::BI__sync_lock_release_8:
2799  case Builtin::BI__sync_lock_release_16: {
2800  Value *Ptr = EmitScalarExpr(E->getArg(0));
2801  QualType ElTy = E->getArg(0)->getType()->getPointeeType();
2802  CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy);
2803  llvm::Type *ITy = llvm::IntegerType::get(getLLVMContext(),
2804  StoreSize.getQuantity() * 8);
2805  Ptr = Builder.CreateBitCast(Ptr, ITy->getPointerTo());
2806  llvm::StoreInst *Store =
2807  Builder.CreateAlignedStore(llvm::Constant::getNullValue(ITy), Ptr,
2808  StoreSize);
2809  Store->setAtomic(llvm::AtomicOrdering::Release);
2810  return RValue::get(nullptr);
2811  }
2812 
2813  case Builtin::BI__sync_synchronize: {
2814  // We assume this is supposed to correspond to a C++0x-style
2815  // sequentially-consistent fence (i.e. this is only usable for
2816  // synchronization, not device I/O or anything like that). This intrinsic
2817  // is really badly designed in the sense that in theory, there isn't
2818  // any way to safely use it... but in practice, it mostly works
2819  // to use it with non-atomic loads and stores to get acquire/release
2820  // semantics.
2821  Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent);
2822  return RValue::get(nullptr);
2823  }
2824 
2825  case Builtin::BI__builtin_nontemporal_load:
2826  return RValue::get(EmitNontemporalLoad(*this, E));
2827  case Builtin::BI__builtin_nontemporal_store:
2828  return RValue::get(EmitNontemporalStore(*this, E));
2829  case Builtin::BI__c11_atomic_is_lock_free:
2830  case Builtin::BI__atomic_is_lock_free: {
2831  // Call "bool __atomic_is_lock_free(size_t size, void *ptr)". For the
2832  // __c11 builtin, ptr is 0 (indicating a properly-aligned object), since
2833  // _Atomic(T) is always properly-aligned.
2834  const char *LibCallName = "__atomic_is_lock_free";
2835  CallArgList Args;
2836  Args.add(RValue::get(EmitScalarExpr(E->getArg(0))),
2837  getContext().getSizeType());
2838  if (BuiltinID == Builtin::BI__atomic_is_lock_free)
2839  Args.add(RValue::get(EmitScalarExpr(E->getArg(1))),
2840  getContext().VoidPtrTy);
2841  else
2842  Args.add(RValue::get(llvm::Constant::getNullValue(VoidPtrTy)),
2843  getContext().VoidPtrTy);
2844  const CGFunctionInfo &FuncInfo =
2845  CGM.getTypes().arrangeBuiltinFunctionCall(E->getType(), Args);
2846  llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FuncInfo);
2847  llvm::FunctionCallee Func = CGM.CreateRuntimeFunction(FTy, LibCallName);
2848  return EmitCall(FuncInfo, CGCallee::forDirect(Func),
2849  ReturnValueSlot(), Args);
2850  }
2851 
2852  case Builtin::BI__atomic_test_and_set: {
2853  // Look at the argument type to determine whether this is a volatile
2854  // operation. The parameter type is always volatile.
2855  QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType();
2856  bool Volatile =
2857  PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified();
2858 
2859  Value *Ptr = EmitScalarExpr(E->getArg(0));
2860  unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace();
2861  Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace));
2862  Value *NewVal = Builder.getInt8(1);
2863  Value *Order = EmitScalarExpr(E->getArg(1));
2864  if (isa<llvm::ConstantInt>(Order)) {
2865  int ord = cast<llvm::ConstantInt>(Order)->getZExtValue();
2866  AtomicRMWInst *Result = nullptr;
2867  switch (ord) {
2868  case 0: // memory_order_relaxed
2869  default: // invalid order
2870  Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal,
2871  llvm::AtomicOrdering::Monotonic);
2872  break;
2873  case 1: // memory_order_consume
2874  case 2: // memory_order_acquire
2875  Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal,
2876  llvm::AtomicOrdering::Acquire);
2877  break;
2878  case 3: // memory_order_release
2879  Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal,
2880  llvm::AtomicOrdering::Release);
2881  break;
2882  case 4: // memory_order_acq_rel
2883 
2884  Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal,
2885  llvm::AtomicOrdering::AcquireRelease);
2886  break;
2887  case 5: // memory_order_seq_cst
2888  Result = Builder.CreateAtomicRMW(
2889  llvm::AtomicRMWInst::Xchg, Ptr, NewVal,
2890  llvm::AtomicOrdering::SequentiallyConsistent);
2891  break;
2892  }
2893  Result->setVolatile(Volatile);
2894  return RValue::get(Builder.CreateIsNotNull(Result, "tobool"));
2895  }
2896 
2897  llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn);
2898 
2899  llvm::BasicBlock *BBs[5] = {
2900  createBasicBlock("monotonic", CurFn),
2901  createBasicBlock("acquire", CurFn),
2902  createBasicBlock("release", CurFn),
2903  createBasicBlock("acqrel", CurFn),
2904  createBasicBlock("seqcst", CurFn)
2905  };
2906  llvm::AtomicOrdering Orders[5] = {
2907  llvm::AtomicOrdering::Monotonic, llvm::AtomicOrdering::Acquire,
2908  llvm::AtomicOrdering::Release, llvm::AtomicOrdering::AcquireRelease,
2909  llvm::AtomicOrdering::SequentiallyConsistent};
2910 
2911  Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false);
2912  llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]);
2913 
2914  Builder.SetInsertPoint(ContBB);
2915  PHINode *Result = Builder.CreatePHI(Int8Ty, 5, "was_set");
2916 
2917  for (unsigned i = 0; i < 5; ++i) {
2918  Builder.SetInsertPoint(BBs[i]);
2919  AtomicRMWInst *RMW = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg,
2920  Ptr, NewVal, Orders[i]);
2921  RMW->setVolatile(Volatile);
2922  Result->addIncoming(RMW, BBs[i]);
2923  Builder.CreateBr(ContBB);
2924  }
2925 
2926  SI->addCase(Builder.getInt32(0), BBs[0]);
2927  SI->addCase(Builder.getInt32(1), BBs[1]);
2928  SI->addCase(Builder.getInt32(2), BBs[1]);
2929  SI->addCase(Builder.getInt32(3), BBs[2]);
2930  SI->addCase(Builder.getInt32(4), BBs[3]);
2931  SI->addCase(Builder.getInt32(5), BBs[4]);
2932 
2933  Builder.SetInsertPoint(ContBB);
2934  return RValue::get(Builder.CreateIsNotNull(Result, "tobool"));
2935  }
2936 
2937  case Builtin::BI__atomic_clear: {
2938  QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType();
2939  bool Volatile =
2940  PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified();
2941 
2942  Address Ptr = EmitPointerWithAlignment(E->getArg(0));
2943  unsigned AddrSpace = Ptr.getPointer()->getType()->getPointerAddressSpace();
2944  Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace));
2945  Value *NewVal = Builder.getInt8(0);
2946  Value *Order = EmitScalarExpr(E->getArg(1));
2947  if (isa<llvm::ConstantInt>(Order)) {
2948  int ord = cast<llvm::ConstantInt>(Order)->getZExtValue();
2949  StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile);
2950  switch (ord) {
2951  case 0: // memory_order_relaxed
2952  default: // invalid order
2953  Store->setOrdering(llvm::AtomicOrdering::Monotonic);
2954  break;
2955  case 3: // memory_order_release
2956  Store->setOrdering(llvm::AtomicOrdering::Release);
2957  break;
2958  case 5: // memory_order_seq_cst
2959  Store->setOrdering(llvm::AtomicOrdering::SequentiallyConsistent);
2960  break;
2961  }
2962  return RValue::get(nullptr);
2963  }
2964 
2965  llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn);
2966 
2967  llvm::BasicBlock *BBs[3] = {
2968  createBasicBlock("monotonic", CurFn),
2969  createBasicBlock("release", CurFn),
2970  createBasicBlock("seqcst", CurFn)
2971  };
2972  llvm::AtomicOrdering Orders[3] = {
2973  llvm::AtomicOrdering::Monotonic, llvm::AtomicOrdering::Release,
2974  llvm::AtomicOrdering::SequentiallyConsistent};
2975 
2976  Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false);
2977  llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]);
2978 
2979  for (unsigned i = 0; i < 3; ++i) {
2980  Builder.SetInsertPoint(BBs[i]);
2981  StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile);
2982  Store->setOrdering(Orders[i]);
2983  Builder.CreateBr(ContBB);
2984  }
2985 
2986  SI->addCase(Builder.getInt32(0), BBs[0]);
2987  SI->addCase(Builder.getInt32(3), BBs[1]);
2988  SI->addCase(Builder.getInt32(5), BBs[2]);
2989 
2990  Builder.SetInsertPoint(ContBB);
2991  return RValue::get(nullptr);
2992  }
2993 
2994  case Builtin::BI__atomic_thread_fence:
2995  case Builtin::BI__atomic_signal_fence:
2996  case Builtin::BI__c11_atomic_thread_fence:
2997  case Builtin::BI__c11_atomic_signal_fence: {
2998  llvm::SyncScope::ID SSID;
2999  if (BuiltinID == Builtin::BI__atomic_signal_fence ||
3000  BuiltinID == Builtin::BI__c11_atomic_signal_fence)
3001  SSID = llvm::SyncScope::SingleThread;
3002  else
3003  SSID = llvm::SyncScope::System;
3004  Value *Order = EmitScalarExpr(E->getArg(0));
3005  if (isa<llvm::ConstantInt>(Order)) {
3006  int ord = cast<llvm::ConstantInt>(Order)->getZExtValue();
3007  switch (ord) {
3008  case 0: // memory_order_relaxed
3009  default: // invalid order
3010  break;
3011  case 1: // memory_order_consume
3012  case 2: // memory_order_acquire
3013  Builder.CreateFence(llvm::AtomicOrdering::Acquire, SSID);
3014  break;
3015  case 3: // memory_order_release
3016  Builder.CreateFence(llvm::AtomicOrdering::Release, SSID);
3017  break;
3018  case 4: // memory_order_acq_rel
3019  Builder.CreateFence(llvm::AtomicOrdering::AcquireRelease, SSID);
3020  break;
3021  case 5: // memory_order_seq_cst
3022  Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent, SSID);
3023  break;
3024  }
3025  return RValue::get(nullptr);
3026  }
3027 
3028  llvm::BasicBlock *AcquireBB, *ReleaseBB, *AcqRelBB, *SeqCstBB;
3029  AcquireBB = createBasicBlock("acquire", CurFn);
3030  ReleaseBB = createBasicBlock("release", CurFn);
3031  AcqRelBB = createBasicBlock("acqrel", CurFn);
3032  SeqCstBB = createBasicBlock("seqcst", CurFn);
3033  llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn);
3034 
3035  Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false);
3036  llvm::SwitchInst *SI = Builder.CreateSwitch(Order, ContBB);
3037 
3038  Builder.SetInsertPoint(AcquireBB);
3039  Builder.CreateFence(llvm::AtomicOrdering::Acquire, SSID);
3040  Builder.CreateBr(ContBB);
3041  SI->addCase(Builder.getInt32(1), AcquireBB);
3042  SI->addCase(Builder.getInt32(2), AcquireBB);
3043 
3044  Builder.SetInsertPoint(ReleaseBB);
3045  Builder.CreateFence(llvm::AtomicOrdering::Release, SSID);
3046  Builder.CreateBr(ContBB);
3047  SI->addCase(Builder.getInt32(3), ReleaseBB);
3048 
3049  Builder.SetInsertPoint(AcqRelBB);
3050  Builder.CreateFence(llvm::AtomicOrdering::AcquireRelease, SSID);
3051  Builder.CreateBr(ContBB);
3052  SI->addCase(Builder.getInt32(4), AcqRelBB);
3053 
3054  Builder.SetInsertPoint(SeqCstBB);
3055  Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent, SSID);
3056  Builder.CreateBr(ContBB);
3057  SI->addCase(Builder.getInt32(5), SeqCstBB);
3058 
3059  Builder.SetInsertPoint(ContBB);
3060  return RValue::get(nullptr);
3061  }
3062 
3063  case Builtin::BI__builtin_signbit:
3064  case Builtin::BI__builtin_signbitf:
3065  case Builtin::BI__builtin_signbitl: {
3066  return RValue::get(
3067  Builder.CreateZExt(EmitSignBit(*this, EmitScalarExpr(E->getArg(0))),
3068  ConvertType(E->getType())));
3069  }
3070  case Builtin::BI__annotation: {
3071  // Re-encode each wide string to UTF8 and make an MDString.
3073  for (const Expr *Arg : E->arguments()) {
3074  const auto *Str = cast<StringLiteral>(Arg->IgnoreParenCasts());
3075  assert(Str->getCharByteWidth() == 2);
3076  StringRef WideBytes = Str->getBytes();
3077  std::string StrUtf8;
3078  if (!convertUTF16ToUTF8String(
3079  makeArrayRef(WideBytes.data(), WideBytes.size()), StrUtf8)) {
3080  CGM.ErrorUnsupported(E, "non-UTF16 __annotation argument");
3081  continue;
3082  }
3083  Strings.push_back(llvm::MDString::get(getLLVMContext(), StrUtf8));
3084  }
3085 
3086  // Build and MDTuple of MDStrings and emit the intrinsic call.
3087  llvm::Function *F =
3088  CGM.getIntrinsic(llvm::Intrinsic::codeview_annotation, {});
3089  MDTuple *StrTuple = MDTuple::get(getLLVMContext(), Strings);
3090  Builder.CreateCall(F, MetadataAsValue::get(getLLVMContext(), StrTuple));
3091  return RValue::getIgnored();
3092  }
3093  case Builtin::BI__builtin_annotation: {
3094  llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0));
3095  llvm::Function *F = CGM.getIntrinsic(llvm::Intrinsic::annotation,
3096  AnnVal->getType());
3097 
3098  // Get the annotation string, go through casts. Sema requires this to be a
3099  // non-wide string literal, potentially casted, so the cast<> is safe.
3100  const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts();
3101  StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString();
3102  return RValue::get(EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc()));
3103  }
3104  case Builtin::BI__builtin_addcb:
3105  case Builtin::BI__builtin_addcs:
3106  case Builtin::BI__builtin_addc:
3107  case Builtin::BI__builtin_addcl:
3108  case Builtin::BI__builtin_addcll:
3109  case Builtin::BI__builtin_subcb:
3110  case Builtin::BI__builtin_subcs:
3111  case Builtin::BI__builtin_subc:
3112  case Builtin::BI__builtin_subcl:
3113  case Builtin::BI__builtin_subcll: {
3114 
3115  // We translate all of these builtins from expressions of the form:
3116  // int x = ..., y = ..., carryin = ..., carryout, result;
3117  // result = __builtin_addc(x, y, carryin, &carryout);
3118  //
3119  // to LLVM IR of the form:
3120  //
3121  // %tmp1 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %x, i32 %y)
3122  // %tmpsum1 = extractvalue {i32, i1} %tmp1, 0
3123  // %carry1 = extractvalue {i32, i1} %tmp1, 1
3124  // %tmp2 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %tmpsum1,
3125  // i32 %carryin)
3126  // %result = extractvalue {i32, i1} %tmp2, 0
3127  // %carry2 = extractvalue {i32, i1} %tmp2, 1
3128  // %tmp3 = or i1 %carry1, %carry2
3129  // %tmp4 = zext i1 %tmp3 to i32
3130  // store i32 %tmp4, i32* %carryout
3131 
3132  // Scalarize our inputs.
3133  llvm::Value *X = EmitScalarExpr(E->getArg(0));
3134  llvm::Value *Y = EmitScalarExpr(E->getArg(1));
3135  llvm::Value *Carryin = EmitScalarExpr(E->getArg(2));
3136  Address CarryOutPtr = EmitPointerWithAlignment(E->getArg(3));
3137 
3138  // Decide if we are lowering to a uadd.with.overflow or usub.with.overflow.
3139  llvm::Intrinsic::ID IntrinsicId;
3140  switch (BuiltinID) {
3141  default: llvm_unreachable("Unknown multiprecision builtin id.");
3142  case Builtin::BI__builtin_addcb:
3143  case Builtin::BI__builtin_addcs:
3144  case Builtin::BI__builtin_addc:
3145  case Builtin::BI__builtin_addcl:
3146  case Builtin::BI__builtin_addcll:
3147  IntrinsicId = llvm::Intrinsic::uadd_with_overflow;
3148  break;
3149  case Builtin::BI__builtin_subcb:
3150  case Builtin::BI__builtin_subcs:
3151  case Builtin::BI__builtin_subc:
3152  case Builtin::BI__builtin_subcl:
3153  case Builtin::BI__builtin_subcll:
3154  IntrinsicId = llvm::Intrinsic::usub_with_overflow;
3155  break;
3156  }
3157 
3158  // Construct our resulting LLVM IR expression.
3159  llvm::Value *Carry1;
3160  llvm::Value *Sum1 = EmitOverflowIntrinsic(*this, IntrinsicId,
3161  X, Y, Carry1);
3162  llvm::Value *Carry2;
3163  llvm::Value *Sum2 = EmitOverflowIntrinsic(*this, IntrinsicId,
3164  Sum1, Carryin, Carry2);
3165  llvm::Value *CarryOut = Builder.CreateZExt(Builder.CreateOr(Carry1, Carry2),
3166  X->getType());
3167  Builder.CreateStore(CarryOut, CarryOutPtr);
3168  return RValue::get(Sum2);
3169  }
3170 
3171  case Builtin::BI__builtin_add_overflow:
3172  case Builtin::BI__builtin_sub_overflow:
3173  case Builtin::BI__builtin_mul_overflow: {
3174  const clang::Expr *LeftArg = E->getArg(0);
3175  const clang::Expr *RightArg = E->getArg(1);
3176  const clang::Expr *ResultArg = E->getArg(2);
3177 
3178  clang::QualType ResultQTy =
3179  ResultArg->getType()->castAs<PointerType>()->getPointeeType();
3180 
3181  WidthAndSignedness LeftInfo =
3182  getIntegerWidthAndSignedness(CGM.getContext(), LeftArg->getType());
3183  WidthAndSignedness RightInfo =
3184  getIntegerWidthAndSignedness(CGM.getContext(), RightArg->getType());
3185  WidthAndSignedness ResultInfo =
3186  getIntegerWidthAndSignedness(CGM.getContext(), ResultQTy);
3187 
3188  // Handle mixed-sign multiplication as a special case, because adding
3189  // runtime or backend support for our generic irgen would be too expensive.
3190  if (isSpecialMixedSignMultiply(BuiltinID, LeftInfo, RightInfo, ResultInfo))
3191  return EmitCheckedMixedSignMultiply(*this, LeftArg, LeftInfo, RightArg,
3192  RightInfo, ResultArg, ResultQTy,
3193  ResultInfo);
3194 
3195  WidthAndSignedness EncompassingInfo =
3196  EncompassingIntegerType({LeftInfo, RightInfo, ResultInfo});
3197 
3198  llvm::Type *EncompassingLLVMTy =
3199  llvm::IntegerType::get(CGM.getLLVMContext(), EncompassingInfo.Width);
3200 
3201  llvm::Type *ResultLLVMTy = CGM.getTypes().ConvertType(ResultQTy);
3202 
3203  llvm::Intrinsic::ID IntrinsicId;
3204  switch (BuiltinID) {
3205  default:
3206  llvm_unreachable("Unknown overflow builtin id.");
3207  case Builtin::BI__builtin_add_overflow:
3208  IntrinsicId = EncompassingInfo.Signed
3209  ? llvm::Intrinsic::sadd_with_overflow
3210  : llvm::Intrinsic::uadd_with_overflow;
3211  break;
3212  case Builtin::BI__builtin_sub_overflow:
3213  IntrinsicId = EncompassingInfo.Signed
3214  ? llvm::Intrinsic::ssub_with_overflow
3215  : llvm::Intrinsic::usub_with_overflow;
3216  break;
3217  case Builtin::BI__builtin_mul_overflow:
3218  IntrinsicId = EncompassingInfo.Signed
3219  ? llvm::Intrinsic::smul_with_overflow
3220  : llvm::Intrinsic::umul_with_overflow;
3221  break;
3222  }
3223 
3224  llvm::Value *Left = EmitScalarExpr(LeftArg);
3225  llvm::Value *Right = EmitScalarExpr(RightArg);
3226  Address ResultPtr = EmitPointerWithAlignment(ResultArg);
3227 
3228  // Extend each operand to the encompassing type.
3229  Left = Builder.CreateIntCast(Left, EncompassingLLVMTy, LeftInfo.Signed);
3230  Right = Builder.CreateIntCast(Right, EncompassingLLVMTy, RightInfo.Signed);
3231 
3232  // Perform the operation on the extended values.
3233  llvm::Value *Overflow, *Result;
3234  Result = EmitOverflowIntrinsic(*this, IntrinsicId, Left, Right, Overflow);
3235 
3236  if (EncompassingInfo.Width > ResultInfo.Width) {
3237  // The encompassing type is wider than the result type, so we need to
3238  // truncate it.
3239  llvm::Value *ResultTrunc = Builder.CreateTrunc(Result, ResultLLVMTy);
3240 
3241  // To see if the truncation caused an overflow, we will extend
3242  // the result and then compare it to the original result.
3243  llvm::Value *ResultTruncExt = Builder.CreateIntCast(
3244  ResultTrunc, EncompassingLLVMTy, ResultInfo.Signed);
3245  llvm::Value *TruncationOverflow =
3246  Builder.CreateICmpNE(Result, ResultTruncExt);
3247 
3248  Overflow = Builder.CreateOr(Overflow, TruncationOverflow);
3249  Result = ResultTrunc;
3250  }
3251 
3252  // Finally, store the result using the pointer.
3253  bool isVolatile =
3254  ResultArg->getType()->getPointeeType().isVolatileQualified();
3255  Builder.CreateStore(EmitToMemory(Result, ResultQTy), ResultPtr, isVolatile);
3256 
3257  return RValue::get(Overflow);
3258  }
3259 
3260  case Builtin::BI__builtin_uadd_overflow:
3261  case Builtin::BI__builtin_uaddl_overflow:
3262  case Builtin::BI__builtin_uaddll_overflow:
3263  case Builtin::BI__builtin_usub_overflow:
3264  case Builtin::BI__builtin_usubl_overflow:
3265  case Builtin::BI__builtin_usubll_overflow:
3266  case Builtin::BI__builtin_umul_overflow:
3267  case Builtin::BI__builtin_umull_overflow:
3268  case Builtin::BI__builtin_umulll_overflow:
3269  case Builtin::BI__builtin_sadd_overflow:
3270  case Builtin::BI__builtin_saddl_overflow:
3271  case Builtin::BI__builtin_saddll_overflow:
3272  case Builtin::BI__builtin_ssub_overflow:
3273  case Builtin::BI__builtin_ssubl_overflow:
3274  case Builtin::BI__builtin_ssubll_overflow:
3275  case Builtin::BI__builtin_smul_overflow:
3276  case Builtin::BI__builtin_smull_overflow:
3277  case Builtin::BI__builtin_smulll_overflow: {
3278 
3279  // We translate all of these builtins directly to the relevant llvm IR node.
3280 
3281  // Scalarize our inputs.
3282  llvm::Value *X = EmitScalarExpr(E->getArg(0));
3283  llvm::Value *Y = EmitScalarExpr(E->getArg(1));
3284  Address SumOutPtr = EmitPointerWithAlignment(E->getArg(2));
3285 
3286  // Decide which of the overflow intrinsics we are lowering to:
3287  llvm::Intrinsic::ID IntrinsicId;
3288  switch (BuiltinID) {
3289  default: llvm_unreachable("Unknown overflow builtin id.");
3290  case Builtin::BI__builtin_uadd_overflow:
3291  case Builtin::BI__builtin_uaddl_overflow:
3292  case Builtin::BI__builtin_uaddll_overflow:
3293  IntrinsicId = llvm::Intrinsic::uadd_with_overflow;
3294  break;
3295  case Builtin::BI__builtin_usub_overflow:
3296  case Builtin::BI__builtin_usubl_overflow:
3297  case Builtin::BI__builtin_usubll_overflow:
3298  IntrinsicId = llvm::Intrinsic::usub_with_overflow;
3299  break;
3300  case Builtin::BI__builtin_umul_overflow:
3301  case Builtin::BI__builtin_umull_overflow:
3302  case Builtin::BI__builtin_umulll_overflow:
3303  IntrinsicId = llvm::Intrinsic::umul_with_overflow;
3304  break;
3305  case Builtin::BI__builtin_sadd_overflow:
3306  case Builtin::BI__builtin_saddl_overflow:
3307  case Builtin::BI__builtin_saddll_overflow:
3308  IntrinsicId = llvm::Intrinsic::sadd_with_overflow;
3309  break;
3310  case Builtin::BI__builtin_ssub_overflow:
3311  case Builtin::BI__builtin_ssubl_overflow:
3312  case Builtin::BI__builtin_ssubll_overflow:
3313  IntrinsicId = llvm::Intrinsic::ssub_with_overflow;
3314  break;
3315  case Builtin::BI__builtin_smul_overflow:
3316  case Builtin::BI__builtin_smull_overflow:
3317  case Builtin::BI__builtin_smulll_overflow:
3318  IntrinsicId = llvm::Intrinsic::smul_with_overflow;
3319  break;
3320  }
3321 
3322 
3323  llvm::Value *Carry;
3324  llvm::Value *Sum = EmitOverflowIntrinsic(*this, IntrinsicId, X, Y, Carry);
3325  Builder.CreateStore(Sum, SumOutPtr);
3326 
3327  return RValue::get(Carry);
3328  }
3329  case Builtin::BI__builtin_addressof:
3330  return RValue::get(EmitLValue(E->getArg(0)).getPointer());
3331  case Builtin::BI__builtin_operator_new:
3332  return EmitBuiltinNewDeleteCall(
3333  E->getCallee()->getType()->castAs<FunctionProtoType>(), E, false);
3334  case Builtin::BI__builtin_operator_delete:
3335  return EmitBuiltinNewDeleteCall(
3336  E->getCallee()->getType()->castAs<FunctionProtoType>(), E, true);
3337 
3338  case Builtin::BI__noop:
3339  // __noop always evaluates to an integer literal zero.
3340  return RValue::get(ConstantInt::get(IntTy, 0));
3341  case Builtin::BI__builtin_call_with_static_chain: {
3342  const CallExpr *Call = cast<CallExpr>(E->getArg(0));
3343  const Expr *Chain = E->getArg(1);
3344  return EmitCall(Call->getCallee()->getType(),
3345  EmitCallee(Call->getCallee()), Call, ReturnValue,
3346  EmitScalarExpr(Chain));
3347  }
3348  case Builtin::BI_InterlockedExchange8:
3349  case Builtin::BI_InterlockedExchange16:
3350  case Builtin::BI_InterlockedExchange:
3351  case Builtin::BI_InterlockedExchangePointer:
3352  return RValue::get(
3353  EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchange, E));
3354  case Builtin::BI_InterlockedCompareExchangePointer:
3355  case Builtin::BI_InterlockedCompareExchangePointer_nf: {
3356  llvm::Type *RTy;
3357  llvm::IntegerType *IntType =
3358  IntegerType::get(getLLVMContext(),
3359  getContext().getTypeSize(E->getType()));
3360  llvm::Type *IntPtrType = IntType->getPointerTo();
3361 
3362  llvm::Value *Destination =
3363  Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), IntPtrType);
3364 
3365  llvm::Value *Exchange = EmitScalarExpr(E->getArg(1));
3366  RTy = Exchange->getType();
3367  Exchange = Builder.CreatePtrToInt(Exchange, IntType);
3368 
3369  llvm::Value *Comparand =
3370  Builder.CreatePtrToInt(EmitScalarExpr(E->getArg(2)), IntType);
3371 
3372  auto Ordering =
3373  BuiltinID == Builtin::BI_InterlockedCompareExchangePointer_nf ?
3374  AtomicOrdering::Monotonic : AtomicOrdering::SequentiallyConsistent;
3375 
3376  auto Result = Builder.CreateAtomicCmpXchg(Destination, Comparand, Exchange,
3377  Ordering, Ordering);
3378  Result->setVolatile(true);
3379 
3380  return RValue::get(Builder.CreateIntToPtr(Builder.CreateExtractValue(Result,
3381  0),
3382  RTy));
3383  }
3384  case Builtin::BI_InterlockedCompareExchange8:
3385  case Builtin::BI_InterlockedCompareExchange16:
3386  case Builtin::BI_InterlockedCompareExchange:
3387  case Builtin::BI_InterlockedCompareExchange64:
3388  return RValue::get(EmitAtomicCmpXchgForMSIntrin(*this, E));
3389  case Builtin::BI_InterlockedIncrement16:
3390  case Builtin::BI_InterlockedIncrement:
3391  return RValue::get(
3392  EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedIncrement, E));
3393  case Builtin::BI_InterlockedDecrement16:
3394  case Builtin::BI_InterlockedDecrement:
3395  return RValue::get(
3396  EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedDecrement, E));
3397  case Builtin::BI_InterlockedAnd8:
3398  case Builtin::BI_InterlockedAnd16:
3399  case Builtin::BI_InterlockedAnd:
3400  return RValue::get(EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedAnd, E));
3401  case Builtin::BI_InterlockedExchangeAdd8:
3402  case Builtin::BI_InterlockedExchangeAdd16:
3403  case Builtin::BI_InterlockedExchangeAdd:
3404  return RValue::get(
3405  EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeAdd, E));
3406  case Builtin::BI_InterlockedExchangeSub8:
3407  case Builtin::BI_InterlockedExchangeSub16:
3408  case Builtin::BI_InterlockedExchangeSub:
3409  return RValue::get(
3410  EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeSub, E));
3411  case Builtin::BI_InterlockedOr8:
3412  case Builtin::BI_InterlockedOr16:
3413  case Builtin::BI_InterlockedOr:
3414  return RValue::get(EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedOr, E));
3415  case Builtin::BI_InterlockedXor8:
3416  case Builtin::BI_InterlockedXor16:
3417  case Builtin::BI_InterlockedXor:
3418  return RValue::get(EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedXor, E));
3419 
3420  case Builtin::BI_bittest64:
3421  case Builtin::BI_bittest:
3422  case Builtin::BI_bittestandcomplement64:
3423  case Builtin::BI_bittestandcomplement:
3424  case Builtin::BI_bittestandreset64:
3425  case Builtin::BI_bittestandreset:
3426  case Builtin::BI_bittestandset64:
3427  case Builtin::BI_bittestandset:
3428  case Builtin::BI_interlockedbittestandreset:
3429  case Builtin::BI_interlockedbittestandreset64:
3430  case Builtin::BI_interlockedbittestandset64:
3431  case Builtin::BI_interlockedbittestandset:
3432  case Builtin::BI_interlockedbittestandset_acq:
3433  case Builtin::BI_interlockedbittestandset_rel:
3434  case Builtin::BI_interlockedbittestandset_nf:
3435  case Builtin::BI_interlockedbittestandreset_acq:
3436  case Builtin::BI_interlockedbittestandreset_rel:
3437  case Builtin::BI_interlockedbittestandreset_nf:
3438  return RValue::get(EmitBitTestIntrinsic(*this, BuiltinID, E));
3439 
3440  // These builtins exist to emit regular volatile loads and stores not
3441  // affected by the -fms-volatile setting.
3442  case Builtin::BI__iso_volatile_load8:
3443  case Builtin::BI__iso_volatile_load16:
3444  case Builtin::BI__iso_volatile_load32:
3445  case Builtin::BI__iso_volatile_load64:
3446  return RValue::get(EmitISOVolatileLoad(*this, E));
3447  case Builtin::BI__iso_volatile_store8:
3448  case Builtin::BI__iso_volatile_store16:
3449  case Builtin::BI__iso_volatile_store32:
3450  case Builtin::BI__iso_volatile_store64:
3451  return RValue::get(EmitISOVolatileStore(*this, E));
3452 
3453  case Builtin::BI__exception_code:
3454  case Builtin::BI_exception_code:
3455  return RValue::get(EmitSEHExceptionCode());
3456  case Builtin::BI__exception_info:
3457  case Builtin::BI_exception_info:
3458  return RValue::get(EmitSEHExceptionInfo());
3459  case Builtin::BI__abnormal_termination:
3460  case Builtin::BI_abnormal_termination:
3461  return RValue::get(EmitSEHAbnormalTermination());
3462  case Builtin::BI_setjmpex:
3463  if (getTarget().getTriple().isOSMSVCRT())
3464  return EmitMSVCRTSetJmp(*this, MSVCSetJmpKind::_setjmpex, E);
3465  break;
3466  case Builtin::BI_setjmp:
3467  if (getTarget().getTriple().isOSMSVCRT()) {
3468  if (getTarget().getTriple().getArch() == llvm::Triple::x86)
3469  return EmitMSVCRTSetJmp(*this, MSVCSetJmpKind::_setjmp3, E);
3470  else if (getTarget().getTriple().getArch() == llvm::Triple::aarch64)
3471  return EmitMSVCRTSetJmp(*this, MSVCSetJmpKind::_setjmpex, E);
3472  return EmitMSVCRTSetJmp(*this, MSVCSetJmpKind::_setjmp, E);
3473  }
3474  break;
3475 
3476  case Builtin::BI__GetExceptionInfo: {
3477  if (llvm::GlobalVariable *GV =
3478  CGM.getCXXABI().getThrowInfo(FD->getParamDecl(0)->getType()))
3479  return RValue::get(llvm::ConstantExpr::getBitCast(GV, CGM.Int8PtrTy));
3480  break;
3481  }
3482 
3483  case Builtin::BI__fastfail:
3484  return RValue::get(EmitMSVCBuiltinExpr(MSVCIntrin::__fastfail, E));
3485 
3486  case Builtin::BI__builtin_coro_size: {
3487  auto & Context = getContext();
3488  auto SizeTy = Context.getSizeType();
3489  auto T = Builder.getIntNTy(Context.getTypeSize(SizeTy));
3490  Function *F = CGM.getIntrinsic(Intrinsic::coro_size, T);
3491  return RValue::get(Builder.CreateCall(F));
3492  }
3493 
3494  case Builtin::BI__builtin_coro_id:
3495  return EmitCoroutineIntrinsic(E, Intrinsic::coro_id);
3496  case Builtin::BI__builtin_coro_promise:
3497  return EmitCoroutineIntrinsic(E, Intrinsic::coro_promise);
3498  case Builtin::BI__builtin_coro_resume:
3499  return EmitCoroutineIntrinsic(E, Intrinsic::coro_resume);
3500  case Builtin::BI__builtin_coro_frame:
3501  return EmitCoroutineIntrinsic(E, Intrinsic::coro_frame);
3502  case Builtin::BI__builtin_coro_noop:
3503  return EmitCoroutineIntrinsic(E, Intrinsic::coro_noop);
3504  case Builtin::BI__builtin_coro_free:
3505  return EmitCoroutineIntrinsic(E, Intrinsic::coro_free);
3506  case Builtin::BI__builtin_coro_destroy:
3507  return EmitCoroutineIntrinsic(E, Intrinsic::coro_destroy);
3508  case Builtin::BI__builtin_coro_done:
3509  return EmitCoroutineIntrinsic(E, Intrinsic::coro_done);
3510  case Builtin::BI__builtin_coro_alloc:
3511  return EmitCoroutineIntrinsic(E, Intrinsic::coro_alloc);
3512  case Builtin::BI__builtin_coro_begin:
3513  return EmitCoroutineIntrinsic(E, Intrinsic::coro_begin);
3514  case Builtin::BI__builtin_coro_end:
3515  return EmitCoroutineIntrinsic(E, Intrinsic::coro_end);
3516  case Builtin::BI__builtin_coro_suspend:
3517  return EmitCoroutineIntrinsic(E, Intrinsic::coro_suspend);
3518  case Builtin::BI__builtin_coro_param:
3519  return EmitCoroutineIntrinsic(E, Intrinsic::coro_param);
3520 
3521  // OpenCL v2.0 s6.13.16.2, Built-in pipe read and write functions
3522  case Builtin::BIread_pipe:
3523  case Builtin::BIwrite_pipe: {
3524  Value *Arg0 = EmitScalarExpr(E->getArg(0)),
3525  *Arg1 = EmitScalarExpr(E->getArg(1));
3526  CGOpenCLRuntime OpenCLRT(CGM);
3527  Value *PacketSize = OpenCLRT.getPipeElemSize(E->getArg(0));
3528  Value *PacketAlign = OpenCLRT.getPipeElemAlign(E->getArg(0));
3529 
3530  // Type of the generic packet parameter.
3531  unsigned GenericAS =
3532  getContext().getTargetAddressSpace(LangAS::opencl_generic);
3533  llvm::Type *I8PTy = llvm::PointerType::get(
3534  llvm::Type::getInt8Ty(getLLVMContext()), GenericAS);
3535 
3536  // Testing which overloaded version we should generate the call for.
3537  if (2U == E->getNumArgs()) {
3538  const char *Name = (BuiltinID == Builtin::BIread_pipe) ? "__read_pipe_2"
3539  : "__write_pipe_2";
3540  // Creating a generic function type to be able to call with any builtin or
3541  // user defined type.
3542  llvm::Type *ArgTys[] = {Arg0->getType(), I8PTy, Int32Ty, Int32Ty};
3543  llvm::FunctionType *FTy = llvm::FunctionType::get(
3544  Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false);
3545  Value *BCast = Builder.CreatePointerCast(Arg1, I8PTy);
3546  return RValue::get(
3547  Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name),
3548  {Arg0, BCast, PacketSize, PacketAlign}));
3549  } else {
3550  assert(4 == E->getNumArgs() &&
3551  "Illegal number of parameters to pipe function");
3552  const char *Name = (BuiltinID == Builtin::BIread_pipe) ? "__read_pipe_4"
3553  : "__write_pipe_4";
3554 
3555  llvm::Type *ArgTys[] = {Arg0->getType(), Arg1->getType(), Int32Ty, I8PTy,
3556  Int32Ty, Int32Ty};
3557  Value *Arg2 = EmitScalarExpr(E->getArg(2)),
3558  *Arg3 = EmitScalarExpr(E->getArg(3));
3559  llvm::FunctionType *FTy = llvm::FunctionType::get(
3560  Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false);
3561  Value *BCast = Builder.CreatePointerCast(Arg3, I8PTy);
3562  // We know the third argument is an integer type, but we may need to cast
3563  // it to i32.
3564  if (Arg2->getType() != Int32Ty)
3565  Arg2 = Builder.CreateZExtOrTrunc(Arg2, Int32Ty);
3566  return RValue::get(Builder.CreateCall(
3567  CGM.CreateRuntimeFunction(FTy, Name),
3568  {Arg0, Arg1, Arg2, BCast, PacketSize, PacketAlign}));
3569  }
3570  }
3571  // OpenCL v2.0 s6.13.16 ,s9.17.3.5 - Built-in pipe reserve read and write
3572  // functions
3573  case Builtin::BIreserve_read_pipe:
3574  case Builtin::BIreserve_write_pipe:
3575  case Builtin::BIwork_group_reserve_read_pipe:
3576  case Builtin::BIwork_group_reserve_write_pipe:
3577  case Builtin::BIsub_group_reserve_read_pipe:
3578  case Builtin::BIsub_group_reserve_write_pipe: {
3579  // Composing the mangled name for the function.
3580  const char *Name;
3581  if (BuiltinID == Builtin::BIreserve_read_pipe)
3582  Name = "__reserve_read_pipe";
3583  else if (BuiltinID == Builtin::BIreserve_write_pipe)
3584  Name = "__reserve_write_pipe";
3585  else if (BuiltinID == Builtin::BIwork_group_reserve_read_pipe)
3586  Name = "__work_group_reserve_read_pipe";
3587  else if (BuiltinID == Builtin::BIwork_group_reserve_write_pipe)
3588  Name = "__work_group_reserve_write_pipe";
3589  else if (BuiltinID == Builtin::BIsub_group_reserve_read_pipe)
3590  Name = "__sub_group_reserve_read_pipe";
3591  else
3592  Name = "__sub_group_reserve_write_pipe";
3593 
3594  Value *Arg0 = EmitScalarExpr(E->getArg(0)),
3595  *Arg1 = EmitScalarExpr(E->getArg(1));
3596  llvm::Type *ReservedIDTy = ConvertType(getContext().OCLReserveIDTy);
3597  CGOpenCLRuntime OpenCLRT(CGM);
3598  Value *PacketSize = OpenCLRT.getPipeElemSize(E->getArg(0));
3599  Value *PacketAlign = OpenCLRT.getPipeElemAlign(E->getArg(0));
3600 
3601  // Building the generic function prototype.
3602  llvm::Type *ArgTys[] = {Arg0->getType(), Int32Ty, Int32Ty, Int32Ty};
3603  llvm::FunctionType *FTy = llvm::FunctionType::get(
3604  ReservedIDTy, llvm::ArrayRef<llvm::Type *>(ArgTys), false);
3605  // We know the second argument is an integer type, but we may need to cast
3606  // it to i32.
3607  if (Arg1->getType() != Int32Ty)
3608  Arg1 = Builder.CreateZExtOrTrunc(Arg1, Int32Ty);
3609  return RValue::get(
3610  Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name),
3611  {Arg0, Arg1, PacketSize, PacketAlign}));
3612  }
3613  // OpenCL v2.0 s6.13.16, s9.17.3.5 - Built-in pipe commit read and write
3614  // functions
3615  case Builtin::BIcommit_read_pipe:
3616  case Builtin::BIcommit_write_pipe:
3617  case Builtin::BIwork_group_commit_read_pipe:
3618  case Builtin::BIwork_group_commit_write_pipe:
3619  case Builtin::BIsub_group_commit_read_pipe:
3620  case Builtin::BIsub_group_commit_write_pipe: {
3621  const char *Name;
3622  if (BuiltinID == Builtin::BIcommit_read_pipe)
3623  Name = "__commit_read_pipe";
3624  else if (BuiltinID == Builtin::BIcommit_write_pipe)
3625  Name = "__commit_write_pipe";
3626  else if (BuiltinID == Builtin::BIwork_group_commit_read_pipe)
3627  Name = "__work_group_commit_read_pipe";
3628  else if (BuiltinID == Builtin::BIwork_group_commit_write_pipe)
3629  Name = "__work_group_commit_write_pipe";
3630  else if (BuiltinID == Builtin::BIsub_group_commit_read_pipe)
3631  Name = "__sub_group_commit_read_pipe";
3632  else
3633  Name = "__sub_group_commit_write_pipe";
3634 
3635  Value *Arg0 = EmitScalarExpr(E->getArg(0)),
3636  *Arg1 = EmitScalarExpr(E->getArg(1));
3637  CGOpenCLRuntime OpenCLRT(CGM);
3638  Value *PacketSize = OpenCLRT.getPipeElemSize(E->getArg(0));
3639  Value *PacketAlign = OpenCLRT.getPipeElemAlign(E->getArg(0));
3640 
3641  // Building the generic function prototype.
3642  llvm::Type *ArgTys[] = {Arg0->getType(), Arg1->getType(), Int32Ty, Int32Ty};
3643  llvm::FunctionType *FTy =
3644  llvm::FunctionType::get(llvm::Type::getVoidTy(getLLVMContext()),
3645  llvm::ArrayRef<llvm::Type *>(ArgTys), false);
3646 
3647  return RValue::get(
3648  Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name),
3649  {Arg0, Arg1, PacketSize, PacketAlign}));
3650  }
3651  // OpenCL v2.0 s6.13.16.4 Built-in pipe query functions
3652  case Builtin::BIget_pipe_num_packets:
3653  case Builtin::BIget_pipe_max_packets: {
3654  const char *BaseName;
3655  const PipeType *PipeTy = E->getArg(0)->getType()->getAs<PipeType>();
3656  if (BuiltinID == Builtin::BIget_pipe_num_packets)
3657  BaseName = "__get_pipe_num_packets";
3658  else
3659  BaseName = "__get_pipe_max_packets";
3660  auto Name = std::string(BaseName) +
3661  std::string(PipeTy->isReadOnly() ? "_ro" : "_wo");
3662 
3663  // Building the generic function prototype.
3664  Value *Arg0 = EmitScalarExpr(E->getArg(0));
3665  CGOpenCLRuntime OpenCLRT(CGM);
3666  Value *PacketSize = OpenCLRT.getPipeElemSize(E->getArg(0));
3667  Value *PacketAlign = OpenCLRT.getPipeElemAlign(E->getArg(0));
3668  llvm::Type *ArgTys[] = {Arg0->getType(), Int32Ty, Int32Ty};
3669  llvm::FunctionType *FTy = llvm::FunctionType::get(
3670  Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false);
3671 
3672  return RValue::get(Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name),
3673  {Arg0, PacketSize, PacketAlign}));
3674  }
3675 
3676  // OpenCL v2.0 s6.13.9 - Address space qualifier functions.
3677  case Builtin::BIto_global:
3678  case Builtin::BIto_local:
3679  case Builtin::BIto_private: {
3680  auto Arg0 = EmitScalarExpr(E->getArg(0));
3681  auto NewArgT = llvm::PointerType::get(Int8Ty,
3682  CGM.getContext().getTargetAddressSpace(LangAS::opencl_generic));
3683  auto NewRetT = llvm::PointerType::get(Int8Ty,
3684  CGM.getContext().getTargetAddressSpace(
3686  auto FTy = llvm::FunctionType::get(NewRetT, {NewArgT}, false);
3687  llvm::Value *NewArg;
3688  if (Arg0->getType()->getPointerAddressSpace() !=
3689  NewArgT->getPointerAddressSpace())
3690  NewArg = Builder.CreateAddrSpaceCast(Arg0, NewArgT);
3691  else
3692  NewArg = Builder.CreateBitOrPointerCast(Arg0, NewArgT);
3693  auto NewName = std::string("__") + E->getDirectCallee()->getName().str();
3694  auto NewCall =
3695  Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, NewName), {NewArg});
3696  return RValue::get(Builder.CreateBitOrPointerCast(NewCall,
3697  ConvertType(E->getType())));
3698  }
3699 
3700  // OpenCL v2.0, s6.13.17 - Enqueue kernel function.
3701  // It contains four different overload formats specified in Table 6.13.17.1.
3702  case Builtin::BIenqueue_kernel: {
3703  StringRef Name; // Generated function call name
3704  unsigned NumArgs = E->getNumArgs();
3705 
3706  llvm::Type *QueueTy = ConvertType(getContext().OCLQueueTy);
3707  llvm::Type *GenericVoidPtrTy = Builder.getInt8PtrTy(
3708  getContext().getTargetAddressSpace(LangAS::opencl_generic));
3709 
3710  llvm::Value *Queue = EmitScalarExpr(E->getArg(0));
3711  llvm::Value *Flags = EmitScalarExpr(E->getArg(1));
3712  LValue NDRangeL = EmitAggExprToLValue(E->getArg(2));
3713  llvm::Value *Range = NDRangeL.getAddress().getPointer();
3714  llvm::Type *RangeTy = NDRangeL.getAddress().getType();
3715 
3716  if (NumArgs == 4) {
3717  // The most basic form of the call with parameters:
3718  // queue_t, kernel_enqueue_flags_t, ndrange_t, block(void)
3719  Name = "__enqueue_kernel_basic";
3720  llvm::Type *ArgTys[] = {QueueTy, Int32Ty, RangeTy, GenericVoidPtrTy,
3721  GenericVoidPtrTy};
3722  llvm::FunctionType *FTy = llvm::FunctionType::get(
3723  Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false);
3724 
3725  auto Info =
3726  CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(3));
3727  llvm::Value *Kernel =
3728  Builder.CreatePointerCast(Info.Kernel, GenericVoidPtrTy);
3729  llvm::Value *Block =
3730  Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy);
3731 
3732  AttrBuilder B;
3733  B.addByValAttr(NDRangeL.getAddress().getElementType());
3734  llvm::AttributeList ByValAttrSet =
3735  llvm::AttributeList::get(CGM.getModule().getContext(), 3U, B);
3736 
3737  auto RTCall =
3738  Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name, ByValAttrSet),
3739  {Queue, Flags, Range, Kernel, Block});
3740  RTCall->setAttributes(ByValAttrSet);
3741  return RValue::get(RTCall);
3742  }
3743  assert(NumArgs >= 5 && "Invalid enqueue_kernel signature");
3744 
3745  // Create a temporary array to hold the sizes of local pointer arguments
3746  // for the block. \p First is the position of the first size argument.
3747  auto CreateArrayForSizeVar = [=](unsigned First)
3748  -> std::tuple<llvm::Value *, llvm::Value *, llvm::Value *> {
3749  llvm::APInt ArraySize(32, NumArgs - First);
3750  QualType SizeArrayTy = getContext().getConstantArrayType(
3751  getContext().getSizeType(), ArraySize, ArrayType::Normal,
3752  /*IndexTypeQuals=*/0);
3753  auto Tmp = CreateMemTemp(SizeArrayTy, "block_sizes");
3754  llvm::Value *TmpPtr = Tmp.getPointer();
3755  llvm::Value *TmpSize = EmitLifetimeStart(
3756  CGM.getDataLayout().getTypeAllocSize(Tmp.getElementType()), TmpPtr);
3757  llvm::Value *ElemPtr;
3758  // Each of the following arguments specifies the size of the corresponding
3759  // argument passed to the enqueued block.
3760  auto *Zero = llvm::ConstantInt::get(IntTy, 0);
3761  for (unsigned I = First; I < NumArgs; ++I) {
3762  auto *Index = llvm::ConstantInt::get(IntTy, I - First);
3763  auto *GEP = Builder.CreateGEP(TmpPtr, {Zero, Index});
3764  if (I == First)
3765  ElemPtr = GEP;
3766  auto *V =
3767  Builder.CreateZExtOrTrunc(EmitScalarExpr(E->getArg(I)), SizeTy);
3768  Builder.CreateAlignedStore(
3769  V, GEP, CGM.getDataLayout().getPrefTypeAlignment(SizeTy));
3770  }
3771  return std::tie(ElemPtr, TmpSize, TmpPtr);
3772  };
3773 
3774  // Could have events and/or varargs.
3775  if (E->getArg(3)->getType()->isBlockPointerType()) {
3776  // No events passed, but has variadic arguments.
3777  Name = "__enqueue_kernel_varargs";
3778  auto Info =
3779  CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(3));
3780  llvm::Value *Kernel =
3781  Builder.CreatePointerCast(Info.Kernel, GenericVoidPtrTy);
3782  auto *Block = Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy);
3783  llvm::Value *ElemPtr, *TmpSize, *TmpPtr;
3784  std::tie(ElemPtr, TmpSize, TmpPtr) = CreateArrayForSizeVar(4);
3785 
3786  // Create a vector of the arguments, as well as a constant value to
3787  // express to the runtime the number of variadic arguments.
3788  std::vector<llvm::Value *> Args = {
3789  Queue, Flags, Range,
3790  Kernel, Block, ConstantInt::get(IntTy, NumArgs - 4),
3791  ElemPtr};
3792  std::vector<llvm::Type *> ArgTys = {
3793  QueueTy, IntTy, RangeTy, GenericVoidPtrTy,
3794  GenericVoidPtrTy, IntTy, ElemPtr->getType()};
3795 
3796  llvm::FunctionType *FTy = llvm::FunctionType::get(
3797  Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false);
3798  auto Call =
3799  RValue::get(Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name),
3801  if (TmpSize)
3802  EmitLifetimeEnd(TmpSize, TmpPtr);
3803  return Call;
3804  }
3805  // Any calls now have event arguments passed.
3806  if (NumArgs >= 7) {
3807  llvm::Type *EventTy = ConvertType(getContext().OCLClkEventTy);
3808  llvm::PointerType *EventPtrTy = EventTy->getPointerTo(
3809  CGM.getContext().getTargetAddressSpace(LangAS::opencl_generic));
3810 
3811  llvm::Value *NumEvents =
3812  Builder.CreateZExtOrTrunc(EmitScalarExpr(E->getArg(3)), Int32Ty);
3813 
3814  // Since SemaOpenCLBuiltinEnqueueKernel allows fifth and sixth arguments
3815  // to be a null pointer constant (including `0` literal), we can take it
3816  // into account and emit null pointer directly.
3817  llvm::Value *EventWaitList = nullptr;
3818  if (E->getArg(4)->isNullPointerConstant(
3819  getContext(), Expr::NPC_ValueDependentIsNotNull)) {
3820  EventWaitList = llvm::ConstantPointerNull::get(EventPtrTy);
3821  } else {
3822  EventWaitList = E->getArg(4)->getType()->isArrayType()
3823  ? EmitArrayToPointerDecay(E->getArg(4)).getPointer()
3824  : EmitScalarExpr(E->getArg(4));
3825  // Convert to generic address space.
3826  EventWaitList = Builder.CreatePointerCast(EventWaitList, EventPtrTy);
3827  }
3828  llvm::Value *EventRet = nullptr;
3829  if (E->getArg(5)->isNullPointerConstant(
3830  getContext(), Expr::NPC_ValueDependentIsNotNull)) {
3831  EventRet = llvm::ConstantPointerNull::get(EventPtrTy);
3832  } else {
3833  EventRet =
3834  Builder.CreatePointerCast(EmitScalarExpr(E->getArg(5)), EventPtrTy);
3835  }
3836 
3837  auto Info =
3838  CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(6));
3839  llvm::Value *Kernel =
3840  Builder.CreatePointerCast(Info.Kernel, GenericVoidPtrTy);
3841  llvm::Value *Block =
3842  Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy);
3843 
3844  std::vector<llvm::Type *> ArgTys = {
3845  QueueTy, Int32Ty, RangeTy, Int32Ty,
3846  EventPtrTy, EventPtrTy, GenericVoidPtrTy, GenericVoidPtrTy};
3847 
3848  std::vector<llvm::Value *> Args = {Queue, Flags, Range,
3849  NumEvents, EventWaitList, EventRet,
3850  Kernel, Block};
3851 
3852  if (NumArgs == 7) {
3853  // Has events but no variadics.
3854  Name = "__enqueue_kernel_basic_events";
3855  llvm::FunctionType *FTy = llvm::FunctionType::get(
3856  Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false);
3857  return RValue::get(
3858  Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name),
3860  }
3861  // Has event info and variadics
3862  // Pass the number of variadics to the runtime function too.
3863  Args.push_back(ConstantInt::get(Int32Ty, NumArgs - 7));
3864  ArgTys.push_back(Int32Ty);
3865  Name = "__enqueue_kernel_events_varargs";
3866 
3867  llvm::Value *ElemPtr, *TmpSize, *TmpPtr;
3868  std::tie(ElemPtr, TmpSize, TmpPtr) = CreateArrayForSizeVar(7);
3869  Args.push_back(ElemPtr);
3870  ArgTys.push_back(ElemPtr->getType());
3871 
3872  llvm::FunctionType *FTy = llvm::FunctionType::get(
3873  Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false);
3874  auto Call =
3875  RValue::get(Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name),
3877  if (TmpSize)
3878  EmitLifetimeEnd(TmpSize, TmpPtr);
3879  return Call;
3880  }
3881  LLVM_FALLTHROUGH;
3882  }
3883  // OpenCL v2.0 s6.13.17.6 - Kernel query functions need bitcast of block
3884  // parameter.
3885  case Builtin::BIget_kernel_work_group_size: {
3886  llvm::Type *GenericVoidPtrTy = Builder.getInt8PtrTy(
3887  getContext().getTargetAddressSpace(LangAS::opencl_generic));
3888  auto Info =
3889  CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(0));
3890  Value *Kernel = Builder.CreatePointerCast(Info.Kernel, GenericVoidPtrTy);
3891  Value *Arg = Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy);
3892  return RValue::get(Builder.CreateCall(
3893  CGM.CreateRuntimeFunction(
3894  llvm::FunctionType::get(IntTy, {GenericVoidPtrTy, GenericVoidPtrTy},
3895  false),
3896  "__get_kernel_work_group_size_impl"),
3897  {Kernel, Arg}));
3898  }
3899  case Builtin::BIget_kernel_preferred_work_group_size_multiple: {
3900  llvm::Type *GenericVoidPtrTy = Builder.getInt8PtrTy(
3901  getContext().getTargetAddressSpace(LangAS::opencl_generic));
3902  auto Info =
3903  CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(0));
3904  Value *Kernel = Builder.CreatePointerCast(Info.Kernel, GenericVoidPtrTy);
3905  Value *Arg = Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy);
3906  return RValue::get(Builder.CreateCall(
3907  CGM.CreateRuntimeFunction(
3908  llvm::FunctionType::get(IntTy, {GenericVoidPtrTy, GenericVoidPtrTy},
3909  false),
3910  "__get_kernel_preferred_work_group_size_multiple_impl"),
3911  {Kernel, Arg}));
3912  }
3913  case Builtin::BIget_kernel_max_sub_group_size_for_ndrange:
3914  case Builtin::BIget_kernel_sub_group_count_for_ndrange: {
3915  llvm::Type *GenericVoidPtrTy = Builder.getInt8PtrTy(
3916  getContext().getTargetAddressSpace(LangAS::opencl_generic));
3917  LValue NDRangeL = EmitAggExprToLValue(E->getArg(0));
3918  llvm::Value *NDRange = NDRangeL.getAddress().getPointer();
3919  auto Info =
3920  CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(1));
3921  Value *Kernel = Builder.CreatePointerCast(Info.Kernel, GenericVoidPtrTy);
3922  Value *Block = Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy);
3923  const char *Name =
3924  BuiltinID == Builtin::BIget_kernel_max_sub_group_size_for_ndrange
3925  ? "__get_kernel_max_sub_group_size_for_ndrange_impl"
3926  : "__get_kernel_sub_group_count_for_ndrange_impl";
3927  return RValue::get(Builder.CreateCall(
3928  CGM.CreateRuntimeFunction(
3929  llvm::FunctionType::get(
3930  IntTy, {NDRange->getType(), GenericVoidPtrTy, GenericVoidPtrTy},
3931  false),
3932  Name),
3933  {NDRange, Kernel, Block}));
3934  }
3935 
3936  case Builtin::BI__builtin_store_half:
3937  case Builtin::BI__builtin_store_halff: {
3938  Value *Val = EmitScalarExpr(E->getArg(0));
3939  Address Address = EmitPointerWithAlignment(E->getArg(1));
3940  Value *HalfVal = Builder.CreateFPTrunc(Val, Builder.getHalfTy());
3941  return RValue::get(Builder.CreateStore(HalfVal, Address));
3942  }
3943  case Builtin::BI__builtin_load_half: {
3944  Address Address = EmitPointerWithAlignment(E->getArg(0));
3945  Value *HalfVal = Builder.CreateLoad(Address);
3946  return RValue::get(Builder.CreateFPExt(HalfVal, Builder.getDoubleTy()));
3947  }
3948  case Builtin::BI__builtin_load_halff: {
3949  Address Address = EmitPointerWithAlignment(E->getArg(0));
3950  Value *HalfVal = Builder.CreateLoad(Address);
3951  return RValue::get(Builder.CreateFPExt(HalfVal, Builder.getFloatTy()));
3952  }
3953  case Builtin::BIprintf:
3954  if (getTarget().getTriple().isNVPTX())
3955  return EmitNVPTXDevicePrintfCallExpr(E, ReturnValue);
3956  break;
3957  case Builtin::BI__builtin_canonicalize:
3958  case Builtin::BI__builtin_canonicalizef:
3959  case Builtin::BI__builtin_canonicalizel:
3960  return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::canonicalize));
3961 
3962  case Builtin::BI__builtin_thread_pointer: {
3963  if (!getContext().getTargetInfo().isTLSSupported())
3964  CGM.ErrorUnsupported(E, "__builtin_thread_pointer");
3965  // Fall through - it's already mapped to the intrinsic by GCCBuiltin.
3966  break;
3967  }
3968  case Builtin::BI__builtin_os_log_format:
3969  return emitBuiltinOSLogFormat(*E);
3970 
3971  case Builtin::BI__xray_customevent: {
3972  if (!ShouldXRayInstrumentFunction())
3973  return RValue::getIgnored();
3974 
3975  if (!CGM.getCodeGenOpts().XRayInstrumentationBundle.has(
3977  return RValue::getIgnored();
3978 
3979  if (const auto *XRayAttr = CurFuncDecl->getAttr<XRayInstrumentAttr>())
3980  if (XRayAttr->neverXRayInstrument() && !AlwaysEmitXRayCustomEvents())
3981  return RValue::getIgnored();
3982 
3983  Function *F = CGM.getIntrinsic(Intrinsic::xray_customevent);
3984  auto FTy = F->getFunctionType();
3985  auto Arg0 = E->getArg(0);
3986  auto Arg0Val = EmitScalarExpr(Arg0);
3987  auto Arg0Ty = Arg0->getType();
3988  auto PTy0 = FTy->getParamType(0);
3989  if (PTy0 != Arg0Val->getType()) {
3990  if (Arg0Ty->isArrayType())
3991  Arg0Val = EmitArrayToPointerDecay(Arg0).getPointer();
3992  else
3993  Arg0Val = Builder.CreatePointerCast(Arg0Val, PTy0);
3994  }
3995  auto Arg1 = EmitScalarExpr(E->getArg(1));
3996  auto PTy1 = FTy->getParamType(1);
3997  if (PTy1 != Arg1->getType())
3998  Arg1 = Builder.CreateTruncOrBitCast(Arg1, PTy1);
3999  return RValue::get(Builder.CreateCall(F, {Arg0Val, Arg1}));
4000  }
4001 
4002  case Builtin::BI__xray_typedevent: {
4003  // TODO: There should be a way to always emit events even if the current
4004  // function is not instrumented. Losing events in a stream can cripple
4005  // a trace.
4006  if (!ShouldXRayInstrumentFunction())
4007  return RValue::getIgnored();
4008 
4009  if (!CGM.getCodeGenOpts().XRayInstrumentationBundle.has(
4011  return RValue::getIgnored();
4012 
4013  if (const auto *XRayAttr = CurFuncDecl->getAttr<XRayInstrumentAttr>())
4014  if (XRayAttr->neverXRayInstrument() && !AlwaysEmitXRayTypedEvents())
4015  return RValue::getIgnored();
4016 
4017  Function *F = CGM.getIntrinsic(Intrinsic::xray_typedevent);
4018  auto FTy = F->getFunctionType();
4019  auto Arg0 = EmitScalarExpr(E->getArg(0));
4020  auto PTy0 = FTy->getParamType(0);
4021  if (PTy0 != Arg0->getType())
4022  Arg0 = Builder.CreateTruncOrBitCast(Arg0, PTy0);
4023  auto Arg1 = E->getArg(1);
4024  auto Arg1Val = EmitScalarExpr(Arg1);
4025  auto Arg1Ty = Arg1->getType();
4026  auto PTy1 = FTy->getParamType(1);
4027  if (PTy1 != Arg1Val->getType()) {
4028  if (Arg1Ty->isArrayType())
4029  Arg1Val = EmitArrayToPointerDecay(Arg1).getPointer();
4030  else
4031  Arg1Val = Builder.CreatePointerCast(Arg1Val, PTy1);
4032  }
4033  auto Arg2 = EmitScalarExpr(E->getArg(2));
4034  auto PTy2 = FTy->getParamType(2);
4035  if (PTy2 != Arg2->getType())
4036  Arg2 = Builder.CreateTruncOrBitCast(Arg2, PTy2);
4037  return RValue::get(Builder.CreateCall(F, {Arg0, Arg1Val, Arg2}));
4038  }
4039 
4040  case Builtin::BI__builtin_ms_va_start:
4041  case Builtin::BI__builtin_ms_va_end:
4042  return RValue::get(
4043  EmitVAStartEnd(EmitMSVAListRef(E->getArg(0)).getPointer(),
4044  BuiltinID == Builtin::BI__builtin_ms_va_start));
4045 
4046  case Builtin::BI__builtin_ms_va_copy: {
4047  // Lower this manually. We can't reliably determine whether or not any
4048  // given va_copy() is for a Win64 va_list from the calling convention
4049  // alone, because it's legal to do this from a System V ABI function.
4050  // With opaque pointer types, we won't have enough information in LLVM
4051  // IR to determine this from the argument types, either. Best to do it
4052  // now, while we have enough information.
4053  Address DestAddr = EmitMSVAListRef(E->getArg(0));
4054  Address SrcAddr = EmitMSVAListRef(E->getArg(1));
4055 
4056  llvm::Type *BPP = Int8PtrPtrTy;
4057 
4058  DestAddr = Address(Builder.CreateBitCast(DestAddr.getPointer(), BPP, "cp"),
4059  DestAddr.getAlignment());
4060  SrcAddr = Address(Builder.CreateBitCast(SrcAddr.getPointer(), BPP, "ap"),
4061  SrcAddr.getAlignment());
4062 
4063  Value *ArgPtr = Builder.CreateLoad(SrcAddr, "ap.val");
4064  return RValue::get(Builder.CreateStore(ArgPtr, DestAddr));
4065  }
4066  }
4067 
4068  // If this is an alias for a lib function (e.g. __builtin_sin), emit
4069  // the call using the normal call path, but using the unmangled
4070  // version of the function name.
4071  if (getContext().BuiltinInfo.isLibFunction(BuiltinID))
4072  return emitLibraryCall(*this, FD, E,
4073  CGM.getBuiltinLibFunction(FD, BuiltinID));
4074 
4075  // If this is a predefined lib function (e.g. malloc), emit the call
4076  // using exactly the normal call path.
4077  if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID))
4078  return emitLibraryCall(*this, FD, E,
4079  cast<llvm::Constant>(EmitScalarExpr(E->getCallee())));
4080 
4081  // Check that a call to a target specific builtin has the correct target
4082  // features.
4083  // This is down here to avoid non-target specific builtins, however, if
4084  // generic builtins start to require generic target features then we
4085  // can move this up to the beginning of the function.
4086  checkTargetFeatures(E, FD);
4087 
4088  if (unsigned VectorWidth = getContext().BuiltinInfo.getRequiredVectorWidth(BuiltinID))
4089  LargestVectorWidth = std::max(LargestVectorWidth, VectorWidth);
4090 
4091  // See if we have a target specific intrinsic.
4092  const char *Name = getContext().BuiltinInfo.getName(BuiltinID);
4093  Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic;
4094  StringRef Prefix =
4095  llvm::Triple::getArchTypePrefix(getTarget().getTriple().getArch());
4096  if (!Prefix.empty()) {
4097  IntrinsicID = Intrinsic::getIntrinsicForGCCBuiltin(Prefix.data(), Name);
4098  // NOTE we don't need to perform a compatibility flag check here since the
4099  // intrinsics are declared in Builtins*.def via LANGBUILTIN which filter the
4100  // MS builtins via ALL_MS_LANGUAGES and are filtered earlier.
4101  if (IntrinsicID == Intrinsic::not_intrinsic)
4102  IntrinsicID = Intrinsic::getIntrinsicForMSBuiltin(Prefix.data(), Name);
4103  }
4104 
4105  if (IntrinsicID != Intrinsic::not_intrinsic) {
4107 
4108  // Find out if any arguments are required to be integer constant
4109  // expressions.
4110  unsigned ICEArguments = 0;
4112  getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments);
4113  assert(Error == ASTContext::GE_None && "Should not codegen an error");
4114 
4115  Function *F = CGM.getIntrinsic(IntrinsicID);
4116  llvm::FunctionType *FTy = F->getFunctionType();
4117 
4118  for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) {
4119  Value *ArgValue;
4120  // If this is a normal argument, just emit it as a scalar.
4121  if ((ICEArguments & (1 << i)) == 0) {
4122  ArgValue = EmitScalarExpr(E->getArg(i));
4123  } else {
4124  // If this is required to be a constant, constant fold it so that we
4125  // know that the generated intrinsic gets a ConstantInt.
4126  llvm::APSInt Result;
4127  bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result,getContext());
4128  assert(IsConst && "Constant arg isn't actually constant?");
4129  (void)IsConst;
4130  ArgValue = llvm::ConstantInt::get(getLLVMContext(), Result);
4131  }
4132 
4133  // If the intrinsic arg type is different from the builtin arg type
4134  // we need to do a bit cast.
4135  llvm::Type *PTy = FTy->getParamType(i);
4136  if (PTy != ArgValue->getType()) {
4137  // XXX - vector of pointers?
4138  if (auto *PtrTy = dyn_cast<llvm::PointerType>(PTy)) {
4139  if (PtrTy->getAddressSpace() !=
4140  ArgValue->getType()->getPointerAddressSpace()) {
4141  ArgValue = Builder.CreateAddrSpaceCast(
4142  ArgValue,
4143  ArgValue->getType()->getPointerTo(PtrTy->getAddressSpace()));
4144  }
4145  }
4146 
4147  assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) &&
4148  "Must be able to losslessly bit cast to param");
4149  ArgValue = Builder.CreateBitCast(ArgValue, PTy);
4150  }
4151 
4152  Args.push_back(ArgValue);
4153  }
4154 
4155  Value *V = Builder.CreateCall(F, Args);
4156  QualType BuiltinRetType = E->getType();
4157 
4158  llvm::Type *RetTy = VoidTy;
4159  if (!BuiltinRetType->isVoidType())
4160  RetTy = ConvertType(BuiltinRetType);
4161 
4162  if (RetTy != V->getType()) {
4163  // XXX - vector of pointers?
4164  if (auto *PtrTy = dyn_cast<llvm::PointerType>(RetTy)) {
4165  if (PtrTy->getAddressSpace() != V->getType()->getPointerAddressSpace()) {
4166  V = Builder.CreateAddrSpaceCast(
4167  V, V->getType()->getPointerTo(PtrTy->getAddressSpace()));
4168  }
4169  }
4170 
4171  assert(V->getType()->canLosslesslyBitCastTo(RetTy) &&
4172  "Must be able to losslessly bit cast result type");
4173  V = Builder.CreateBitCast(V, RetTy);
4174  }
4175 
4176  return RValue::get(V);
4177  }
4178 
4179  // See if we have a target specific builtin that needs to be lowered.
4180  if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E))
4181  return RValue::get(V);
4182 
4183  ErrorUnsupported(E, "builtin function");
4184 
4185  // Unknown builtin, for now just dump it out and return undef.
4186  return GetUndefRValue(E->getType());
4187 }
4188 
4190  unsigned BuiltinID, const CallExpr *E,
4191  llvm::Triple::ArchType Arch) {
4192  switch (Arch) {
4193  case llvm::Triple::arm:
4194  case llvm::Triple::armeb:
4195  case llvm::Triple::thumb:
4196  case llvm::Triple::thumbeb:
4197  return CGF->EmitARMBuiltinExpr(BuiltinID, E, Arch);
4198  case llvm::Triple::aarch64:
4199  case llvm::Triple::aarch64_be:
4200  return CGF->EmitAArch64BuiltinExpr(BuiltinID, E, Arch);
4201  case llvm::Triple::x86:
4202  case llvm::Triple::x86_64:
4203  return CGF->EmitX86BuiltinExpr(BuiltinID, E);
4204  case llvm::Triple::ppc:
4205  case llvm::Triple::ppc64:
4206  case llvm::Triple::ppc64le:
4207  return CGF->EmitPPCBuiltinExpr(BuiltinID, E);
4208  case llvm::Triple::r600:
4209  case llvm::Triple::amdgcn:
4210  return CGF->EmitAMDGPUBuiltinExpr(BuiltinID, E);
4211  case llvm::Triple::systemz:
4212  return CGF->EmitSystemZBuiltinExpr(BuiltinID, E);
4213  case llvm::Triple::nvptx:
4214  case llvm::Triple::nvptx64:
4215  return CGF->EmitNVPTXBuiltinExpr(BuiltinID, E);
4216  case llvm::Triple::wasm32:
4217  case llvm::Triple::wasm64:
4218  return CGF->EmitWebAssemblyBuiltinExpr(BuiltinID, E);
4219  case llvm::Triple::hexagon:
4220  return CGF->EmitHexagonBuiltinExpr(BuiltinID, E);
4221  default:
4222  return nullptr;
4223  }
4224 }
4225 
4227  const CallExpr *E) {
4228  if (getContext().BuiltinInfo.isAuxBuiltinID(BuiltinID)) {
4229  assert(getContext().getAuxTargetInfo() && "Missing aux target info");
4231  this, getContext().BuiltinInfo.getAuxBuiltinID(BuiltinID), E,
4232  getContext().getAuxTargetInfo()->getTriple().getArch());
4233  }
4234 
4235  return EmitTargetArchBuiltinExpr(this, BuiltinID, E,
4236  getTarget().getTriple().getArch());
4237 }
4238 
4239 static llvm::VectorType *GetNeonType(CodeGenFunction *CGF,
4240  NeonTypeFlags TypeFlags,
4241  bool HasLegalHalfType=true,
4242  bool V1Ty=false) {
4243  int IsQuad = TypeFlags.isQuad();
4244  switch (TypeFlags.getEltType()) {
4245  case NeonTypeFlags::Int8:
4246  case NeonTypeFlags::Poly8:
4247  return llvm::VectorType::get(CGF->Int8Ty, V1Ty ? 1 : (8 << IsQuad));
4248  case NeonTypeFlags::Int16:
4249  case NeonTypeFlags::Poly16:
4250  return llvm::VectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad));
4252  if (HasLegalHalfType)
4253  return llvm::VectorType::get(CGF->HalfTy, V1Ty ? 1 : (4 << IsQuad));
4254  else
4255  return llvm::VectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad));
4256  case NeonTypeFlags::Int32:
4257  return llvm::VectorType::get(CGF->Int32Ty, V1Ty ? 1 : (2 << IsQuad));
4258  case NeonTypeFlags::Int64:
4259  case NeonTypeFlags::Poly64:
4260  return llvm::VectorType::get(CGF->Int64Ty, V1Ty ? 1 : (1 << IsQuad));
4262  // FIXME: i128 and f128 doesn't get fully support in Clang and llvm.
4263  // There is a lot of i128 and f128 API missing.
4264  // so we use v16i8 to represent poly128 and get pattern matched.
4265  return llvm::VectorType::get(CGF->Int8Ty, 16);
4267  return llvm::VectorType::get(CGF->FloatTy, V1Ty ? 1 : (2 << IsQuad));
4269  return llvm::VectorType::get(CGF->DoubleTy, V1Ty ? 1 : (1 << IsQuad));
4270  }
4271  llvm_unreachable("Unknown vector element type!");
4272 }
4273 
4274 static llvm::VectorType *GetFloatNeonType(CodeGenFunction *CGF,
4275  NeonTypeFlags IntTypeFlags) {
4276  int IsQuad = IntTypeFlags.isQuad();
4277  switch (IntTypeFlags.getEltType()) {
4278  case NeonTypeFlags::Int16:
4279  return llvm::VectorType::get(CGF->HalfTy, (4 << IsQuad));
4280  case NeonTypeFlags::Int32:
4281  return llvm::VectorType::get(CGF->FloatTy, (2 << IsQuad));
4282  case NeonTypeFlags::Int64:
4283  return llvm::VectorType::get(CGF->DoubleTy, (1 << IsQuad));
4284  default:
4285  llvm_unreachable("Type can't be converted to floating-point!");
4286  }
4287 }
4288 
4290  unsigned nElts = V->getType()->getVectorNumElements();
4291  Value* SV = llvm::ConstantVector::getSplat(nElts, C);
4292  return Builder.CreateShuffleVector(V, V, SV, "lane");
4293 }
4294 
4296  const char *name,
4297  unsigned shift, bool rightshift) {
4298  unsigned j = 0;
4299  for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end();
4300  ai != ae; ++ai, ++j)
4301  if (shift > 0 && shift == j)
4302  Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift);
4303  else
4304  Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name);
4305 
4306  return Builder.CreateCall(F, Ops, name);
4307 }
4308 
4310  bool neg) {
4311  int SV = cast<ConstantInt>(V)->getSExtValue();
4312  return ConstantInt::get(Ty, neg ? -SV : SV);
4313 }
4314 
4315 // Right-shift a vector by a constant.
4317  llvm::Type *Ty, bool usgn,
4318  const char *name) {
4319  llvm::VectorType *VTy = cast<llvm::VectorType>(Ty);
4320 
4321  int ShiftAmt = cast<ConstantInt>(Shift)->getSExtValue();
4322  int EltSize = VTy->getScalarSizeInBits();
4323 
4324  Vec = Builder.CreateBitCast(Vec, Ty);
4325 
4326  // lshr/ashr are undefined when the shift amount is equal to the vector
4327  // element size.
4328  if (ShiftAmt == EltSize) {
4329  if (usgn) {
4330  // Right-shifting an unsigned value by its size yields 0.
4331  return llvm::ConstantAggregateZero::get(VTy);
4332  } else {
4333  // Right-shifting a signed value by its size is equivalent
4334  // to a shift of size-1.
4335  --ShiftAmt;
4336  Shift = ConstantInt::get(VTy->getElementType(), ShiftAmt);
4337  }
4338  }
4339 
4340  Shift = EmitNeonShiftVector(Shift, Ty, false);
4341  if (usgn)
4342  return Builder.CreateLShr(Vec, Shift, name);
4343  else
4344  return Builder.CreateAShr(Vec, Shift, name);
4345 }
4346 
4347 enum {
4348  AddRetType = (1 << 0),
4349  Add1ArgType = (1 << 1),
4350  Add2ArgTypes = (1 << 2),
4351 
4352  VectorizeRetType = (1 << 3),
4353  VectorizeArgTypes = (1 << 4),
4354 
4355  InventFloatType = (1 << 5),
4356  UnsignedAlts = (1 << 6),
4357 
4358  Use64BitVectors = (1 << 7),
4359  Use128BitVectors = (1 << 8),
4360 
4367 };
4368 
4369 namespace {
4370 struct NeonIntrinsicInfo {
4371  const char *NameHint;
4372  unsigned BuiltinID;
4373  unsigned LLVMIntrinsic;
4374  unsigned AltLLVMIntrinsic;
4375  unsigned TypeModifier;
4376 
4377  bool operator<(unsigned RHSBuiltinID) const {
4378  return BuiltinID < RHSBuiltinID;
4379  }
4380  bool operator<(const NeonIntrinsicInfo &TE) const {
4381  return BuiltinID < TE.BuiltinID;
4382  }
4383 };
4384 } // end anonymous namespace
4385 
4386 #define NEONMAP0(NameBase) \
4387  { #NameBase, NEON::BI__builtin_neon_ ## NameBase, 0, 0, 0 }
4388 
4389 #define NEONMAP1(NameBase, LLVMIntrinsic, TypeModifier) \
4390  { #NameBase, NEON:: BI__builtin_neon_ ## NameBase, \
4391  Intrinsic::LLVMIntrinsic, 0, TypeModifier }
4392 
4393 #define NEONMAP2(NameBase, LLVMIntrinsic, AltLLVMIntrinsic, TypeModifier) \
4394  { #NameBase, NEON:: BI__builtin_neon_ ## NameBase, \
4395  Intrinsic::LLVMIntrinsic, Intrinsic::AltLLVMIntrinsic, \
4396  TypeModifier }
4397 
4398 static const NeonIntrinsicInfo ARMSIMDIntrinsicMap [] = {
4399  NEONMAP2(vabd_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts),
4400  NEONMAP2(vabdq_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts),
4401  NEONMAP1(vabs_v, arm_neon_vabs, 0),
4402  NEONMAP1(vabsq_v, arm_neon_vabs, 0),
4403  NEONMAP0(vaddhn_v),
4404  NEONMAP1(vaesdq_v, arm_neon_aesd, 0),
4405  NEONMAP1(vaeseq_v, arm_neon_aese, 0),
4406  NEONMAP1(vaesimcq_v, arm_neon_aesimc, 0),
4407  NEONMAP1(vaesmcq_v, arm_neon_aesmc, 0),
4408  NEONMAP1(vbsl_v, arm_neon_vbsl, AddRetType),
4409  NEONMAP1(vbslq_v, arm_neon_vbsl, AddRetType),
4410  NEONMAP1(vcage_v, arm_neon_vacge, 0),
4411  NEONMAP1(vcageq_v, arm_neon_vacge, 0),
4412  NEONMAP1(vcagt_v, arm_neon_vacgt, 0),
4413  NEONMAP1(vcagtq_v, arm_neon_vacgt, 0),
4414  NEONMAP1(vcale_v, arm_neon_vacge, 0),
4415  NEONMAP1(vcaleq_v, arm_neon_vacge, 0),
4416  NEONMAP1(vcalt_v, arm_neon_vacgt, 0),
4417  NEONMAP1(vcaltq_v, arm_neon_vacgt, 0),
4418  NEONMAP0(vceqz_v),
4419  NEONMAP0(vceqzq_v),
4420  NEONMAP0(vcgez_v),
4421  NEONMAP0(vcgezq_v),
4422  NEONMAP0(vcgtz_v),
4423  NEONMAP0(vcgtzq_v),
4424  NEONMAP0(vclez_v),
4425  NEONMAP0(vclezq_v),
4426  NEONMAP1(vcls_v, arm_neon_vcls, Add1ArgType),
4427  NEONMAP1(vclsq_v, arm_neon_vcls, Add1ArgType),
4428  NEONMAP0(vcltz_v),
4429  NEONMAP0(vcltzq_v),
4430  NEONMAP1(vclz_v, ctlz, Add1ArgType),
4431  NEONMAP1(vclzq_v, ctlz, Add1ArgType),
4432  NEONMAP1(vcnt_v, ctpop, Add1ArgType),
4433  NEONMAP1(vcntq_v, ctpop, Add1ArgType),
4434  NEONMAP1(vcvt_f16_f32, arm_neon_vcvtfp2hf, 0),
4435  NEONMAP0(vcvt_f16_v),
4436  NEONMAP1(vcvt_f32_f16, arm_neon_vcvthf2fp, 0),
4437  NEONMAP0(vcvt_f32_v),
4438  NEONMAP2(vcvt_n_f16_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0),
4439  NEONMAP2(vcvt_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0),
4440  NEONMAP1(vcvt_n_s16_v, arm_neon_vcvtfp2fxs, 0),
4441  NEONMAP1(vcvt_n_s32_v, arm_neon_vcvtfp2fxs, 0),
4442  NEONMAP1(vcvt_n_s64_v, arm_neon_vcvtfp2fxs, 0),
4443  NEONMAP1(vcvt_n_u16_v, arm_neon_vcvtfp2fxu, 0),
4444  NEONMAP1(vcvt_n_u32_v, arm_neon_vcvtfp2fxu, 0),
4445  NEONMAP1(vcvt_n_u64_v, arm_neon_vcvtfp2fxu, 0),
4446  NEONMAP0(vcvt_s16_v),
4447  NEONMAP0(vcvt_s32_v),
4448  NEONMAP0(vcvt_s64_v),
4449  NEONMAP0(vcvt_u16_v),
4450  NEONMAP0(vcvt_u32_v),
4451  NEONMAP0(vcvt_u64_v),
4452  NEONMAP1(vcvta_s16_v, arm_neon_vcvtas, 0),
4453  NEONMAP1(vcvta_s32_v, arm_neon_vcvtas, 0),
4454  NEONMAP1(vcvta_s64_v, arm_neon_vcvtas, 0),
4455  NEONMAP1(vcvta_u16_v, arm_neon_vcvtau, 0),
4456  NEONMAP1(vcvta_u32_v, arm_neon_vcvtau, 0),
4457  NEONMAP1(vcvta_u64_v, arm_neon_vcvtau, 0),
4458  NEONMAP1(vcvtaq_s16_v, arm_neon_vcvtas, 0),
4459  NEONMAP1(vcvtaq_s32_v, arm_neon_vcvtas, 0),
4460  NEONMAP1(vcvtaq_s64_v, arm_neon_vcvtas, 0),
4461  NEONMAP1(vcvtaq_u16_v, arm_neon_vcvtau, 0),
4462  NEONMAP1(vcvtaq_u32_v, arm_neon_vcvtau, 0),
4463  NEONMAP1(vcvtaq_u64_v, arm_neon_vcvtau, 0),
4464  NEONMAP1(vcvtm_s16_v, arm_neon_vcvtms, 0),
4465  NEONMAP1(vcvtm_s32_v, arm_neon_vcvtms, 0),
4466  NEONMAP1(vcvtm_s64_v, arm_neon_vcvtms, 0),
4467  NEONMAP1(vcvtm_u16_v, arm_neon_vcvtmu, 0),
4468  NEONMAP1(vcvtm_u32_v, arm_neon_vcvtmu, 0),
4469  NEONMAP1(vcvtm_u64_v, arm_neon_vcvtmu, 0),
4470  NEONMAP1(vcvtmq_s16_v, arm_neon_vcvtms, 0),
4471  NEONMAP1(vcvtmq_s32_v, arm_neon_vcvtms, 0),
4472  NEONMAP1(vcvtmq_s64_v, arm_neon_vcvtms, 0),
4473  NEONMAP1(vcvtmq_u16_v, arm_neon_vcvtmu, 0),
4474  NEONMAP1(vcvtmq_u32_v, arm_neon_vcvtmu, 0),
4475  NEONMAP1(vcvtmq_u64_v, arm_neon_vcvtmu, 0),
4476  NEONMAP1(vcvtn_s16_v, arm_neon_vcvtns, 0),
4477  NEONMAP1(vcvtn_s32_v, arm_neon_vcvtns, 0),
4478  NEONMAP1(vcvtn_s64_v, arm_neon_vcvtns, 0),
4479  NEONMAP1(vcvtn_u16_v, arm_neon_vcvtnu, 0),
4480  NEONMAP1(vcvtn_u32_v, arm_neon_vcvtnu, 0),
4481  NEONMAP1(vcvtn_u64_v, arm_neon_vcvtnu, 0),
4482  NEONMAP1(vcvtnq_s16_v, arm_neon_vcvtns, 0),
4483  NEONMAP1(vcvtnq_s32_v, arm_neon_vcvtns, 0),
4484  NEONMAP1(vcvtnq_s64_v, arm_neon_vcvtns, 0),
4485  NEONMAP1(vcvtnq_u16_v, arm_neon_vcvtnu, 0),
4486  NEONMAP1(vcvtnq_u32_v, arm_neon_vcvtnu, 0),
4487  NEONMAP1(vcvtnq_u64_v, arm_neon_vcvtnu, 0),
4488  NEONMAP1(vcvtp_s16_v, arm_neon_vcvtps, 0),
4489  NEONMAP1(vcvtp_s32_v, arm_neon_vcvtps, 0),
4490  NEONMAP1(vcvtp_s64_v, arm_neon_vcvtps, 0),
4491  NEONMAP1(vcvtp_u16_v, arm_neon_vcvtpu, 0),
4492  NEONMAP1(vcvtp_u32_v, arm_neon_vcvtpu, 0),
4493  NEONMAP1(vcvtp_u64_v, arm_neon_vcvtpu, 0),
4494  NEONMAP1(vcvtpq_s16_v, arm_neon_vcvtps, 0),
4495  NEONMAP1(vcvtpq_s32_v, arm_neon_vcvtps, 0),
4496  NEONMAP1(vcvtpq_s64_v, arm_neon_vcvtps, 0),
4497  NEONMAP1(vcvtpq_u16_v, arm_neon_vcvtpu, 0),
4498  NEONMAP1(vcvtpq_u32_v, arm_neon_vcvtpu, 0),
4499  NEONMAP1(vcvtpq_u64_v, arm_neon_vcvtpu, 0),
4500  NEONMAP0(vcvtq_f16_v),
4501  NEONMAP0(vcvtq_f32_v),
4502  NEONMAP2(vcvtq_n_f16_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0),
4503  NEONMAP2(vcvtq_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0),
4504  NEONMAP1(vcvtq_n_s16_v, arm_neon_vcvtfp2fxs, 0),
4505  NEONMAP1(vcvtq_n_s32_v, arm_neon_vcvtfp2fxs, 0),
4506  NEONMAP1(vcvtq_n_s64_v, arm_neon_vcvtfp2fxs, 0),
4507  NEONMAP1(vcvtq_n_u16_v, arm_neon_vcvtfp2fxu, 0),
4508  NEONMAP1(vcvtq_n_u32_v, arm_neon_vcvtfp2fxu, 0),
4509  NEONMAP1(vcvtq_n_u64_v, arm_neon_vcvtfp2fxu, 0),
4510  NEONMAP0(vcvtq_s16_v),
4511  NEONMAP0(vcvtq_s32_v),
4512  NEONMAP0(vcvtq_s64_v),
4513  NEONMAP0(vcvtq_u16_v),
4514  NEONMAP0(vcvtq_u32_v),
4515  NEONMAP0(vcvtq_u64_v),
4516  NEONMAP2(vdot_v, arm_neon_udot, arm_neon_sdot, 0),
4517  NEONMAP2(vdotq_v, arm_neon_udot, arm_neon_sdot, 0),
4518  NEONMAP0(vext_v),
4519  NEONMAP0(vextq_v),
4520  NEONMAP0(vfma_v),
4521  NEONMAP0(vfmaq_v),
4522  NEONMAP2(vhadd_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts),
4523  NEONMAP2(vhaddq_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts),
4524  NEONMAP2(vhsub_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts),
4525  NEONMAP2(vhsubq_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts),
4526  NEONMAP0(vld1_dup_v),
4527  NEONMAP1(vld1_v, arm_neon_vld1, 0),
4528  NEONMAP1(vld1_x2_v, arm_neon_vld1x2, 0),
4529  NEONMAP1(vld1_x3_v, arm_neon_vld1x3, 0),
4530  NEONMAP1(vld1_x4_v, arm_neon_vld1x4, 0),
4531  NEONMAP0(vld1q_dup_v),
4532  NEONMAP1(vld1q_v, arm_neon_vld1, 0),
4533  NEONMAP1(vld1q_x2_v, arm_neon_vld1x2, 0),
4534  NEONMAP1(vld1q_x3_v, arm_neon_vld1x3, 0),
4535  NEONMAP1(vld1q_x4_v, arm_neon_vld1x4, 0),
4536  NEONMAP1(vld2_dup_v, arm_neon_vld2dup, 0),
4537  NEONMAP1(vld2_lane_v, arm_neon_vld2lane, 0),
4538  NEONMAP1(vld2_v, arm_neon_vld2, 0),
4539  NEONMAP1(vld2q_dup_v, arm_neon_vld2dup, 0),
4540  NEONMAP1(vld2q_lane_v, arm_neon_vld2lane, 0),
4541  NEONMAP1(vld2q_v, arm_neon_vld2, 0),
4542  NEONMAP1(vld3_dup_v, arm_neon_vld3dup, 0),
4543  NEONMAP1(vld3_lane_v, arm_neon_vld3lane, 0),
4544  NEONMAP1(vld3_v, arm_neon_vld3, 0),
4545  NEONMAP1(vld3q_dup_v, arm_neon_vld3dup, 0),
4546  NEONMAP1(vld3q_lane_v, arm_neon_vld3lane, 0),
4547  NEONMAP1(vld3q_v, arm_neon_vld3, 0),
4548  NEONMAP1(vld4_dup_v, arm_neon_vld4dup, 0),
4549  NEONMAP1(vld4_lane_v, arm_neon_vld4lane, 0),
4550  NEONMAP1(vld4_v, arm_neon_vld4, 0),
4551  NEONMAP1(vld4q_dup_v, arm_neon_vld4dup, 0),
4552  NEONMAP1(vld4q_lane_v, arm_neon_vld4lane, 0),
4553  NEONMAP1(vld4q_v, arm_neon_vld4, 0),
4554  NEONMAP2(vmax_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts),
4555  NEONMAP1(vmaxnm_v, arm_neon_vmaxnm, Add1ArgType),
4556  NEONMAP1(vmaxnmq_v, arm_neon_vmaxnm, Add1ArgType),
4557  NEONMAP2(vmaxq_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts),
4558  NEONMAP2(vmin_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts),
4559  NEONMAP1(vminnm_v, arm_neon_vminnm, Add1ArgType),
4560  NEONMAP1(vminnmq_v, arm_neon_vminnm, Add1ArgType),
4561  NEONMAP2(vminq_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts),
4562  NEONMAP0(vmovl_v),
4563  NEONMAP0(vmovn_v),
4564  NEONMAP1(vmul_v, arm_neon_vmulp, Add1ArgType),
4565  NEONMAP0(vmull_v),
4566  NEONMAP1(vmulq_v, arm_neon_vmulp, Add1ArgType),
4567  NEONMAP2(vpadal_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts),
4568  NEONMAP2(vpadalq_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts),
4569  NEONMAP1(vpadd_v, arm_neon_vpadd, Add1ArgType),
4570  NEONMAP2(vpaddl_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts),
4571  NEONMAP2(vpaddlq_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts),
4572  NEONMAP1(vpaddq_v, arm_neon_vpadd, Add1ArgType),
4573  NEONMAP2(vpmax_v, arm_neon_vpmaxu, arm_neon_vpmaxs, Add1ArgType | UnsignedAlts),
4574  NEONMAP2(vpmin_v, arm_neon_vpminu, arm_neon_vpmins, Add1ArgType | UnsignedAlts),
4575  NEONMAP1(vqabs_v, arm_neon_vqabs, Add1ArgType),
4576  NEONMAP1(vqabsq_v, arm_neon_vqabs, Add1ArgType),
4577  NEONMAP2(vqadd_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts),
4578  NEONMAP2(vqaddq_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts),
4579  NEONMAP2(vqdmlal_v, arm_neon_vqdmull, arm_neon_vqadds, 0),
4580  NEONMAP2(vqdmlsl_v, arm_neon_vqdmull, arm_neon_vqsubs, 0),
4581  NEONMAP1(vqdmulh_v, arm_neon_vqdmulh, Add1ArgType),
4582  NEONMAP1(vqdmulhq_v, arm_neon_vqdmulh, Add1ArgType),
4583  NEONMAP1(vqdmull_v, arm_neon_vqdmull, Add1ArgType),
4584  NEONMAP2(vqmovn_v, arm_neon_vqmovnu, arm_neon_vqmovns, Add1ArgType | UnsignedAlts),
4585  NEONMAP1(vqmovun_v, arm_neon_vqmovnsu, Add1ArgType),
4586  NEONMAP1(vqneg_v, arm_neon_vqneg, Add1ArgType),
4587  NEONMAP1(vqnegq_v, arm_neon_vqneg, Add1ArgType),
4588  NEONMAP1(vqrdmulh_v, arm_neon_vqrdmulh, Add1ArgType),
4589  NEONMAP1(vqrdmulhq_v, arm_neon_vqrdmulh, Add1ArgType),
4590  NEONMAP2(vqrshl_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts),
4591  NEONMAP2(vqrshlq_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts),
4592  NEONMAP2(vqshl_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts),
4593  NEONMAP2(vqshl_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts),
4594  NEONMAP2(vqshlq_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts),
4595  NEONMAP2(vqshlq_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts),
4596  NEONMAP1(vqshlu_n_v, arm_neon_vqshiftsu, 0),
4597  NEONMAP1(vqshluq_n_v, arm_neon_vqshiftsu, 0),
4598  NEONMAP2(vqsub_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts),
4599  NEONMAP2(vqsubq_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts),
4600  NEONMAP1(vraddhn_v, arm_neon_vraddhn, Add1ArgType),
4601  NEONMAP2(vrecpe_v, arm_neon_vrecpe, arm_neon_vrecpe, 0),
4602  NEONMAP2(vrecpeq_v, arm_neon_vrecpe, arm_neon_vrecpe, 0),
4603  NEONMAP1(vrecps_v, arm_neon_vrecps, Add1ArgType),
4604  NEONMAP1(vrecpsq_v, arm_neon_vrecps, Add1ArgType),
4605  NEONMAP2(vrhadd_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts),
4606  NEONMAP2(vrhaddq_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts),
4607  NEONMAP1(vrnd_v, arm_neon_vrintz, Add1ArgType),
4608  NEONMAP1(vrnda_v, arm_neon_vrinta, Add1ArgType),
4609  NEONMAP1(vrndaq_v, arm_neon_vrinta, Add1ArgType),
4610  NEONMAP0(vrndi_v),
4611  NEONMAP0(vrndiq_v),
4612  NEONMAP1(vrndm_v, arm_neon_vrintm, Add1ArgType),
4613  NEONMAP1(vrndmq_v, arm_neon_vrintm, Add1ArgType),
4614  NEONMAP1(vrndn_v, arm_neon_vrintn, Add1ArgType),
4615  NEONMAP1(vrndnq_v, arm_neon_vrintn, Add1ArgType),
4616  NEONMAP1(vrndp_v, arm_neon_vrintp, Add1ArgType),
4617  NEONMAP1(vrndpq_v, arm_neon_vrintp, Add1ArgType),
4618  NEONMAP1(vrndq_v, arm_neon_vrintz, Add1ArgType),
4619  NEONMAP1(vrndx_v, arm_neon_vrintx, Add1ArgType),
4620  NEONMAP1(vrndxq_v, arm_neon_vrintx, Add1ArgType),
4621  NEONMAP2(vrshl_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts),
4622  NEONMAP2(vrshlq_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts),
4623  NEONMAP2(vrshr_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts),
4624  NEONMAP2(vrshrq_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts),
4625  NEONMAP2(vrsqrte_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0),
4626  NEONMAP2(vrsqrteq_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0),
4627  NEONMAP1(vrsqrts_v, arm_neon_vrsqrts, Add1ArgType),
4628  NEONMAP1(vrsqrtsq_v, arm_neon_vrsqrts, Add1ArgType),
4629  NEONMAP1(vrsubhn_v, arm_neon_vrsubhn, Add1ArgType),
4630  NEONMAP1(vsha1su0q_v, arm_neon_sha1su0, 0),
4631  NEONMAP1(vsha1su1q_v, arm_neon_sha1su1, 0),
4632  NEONMAP1(vsha256h2q_v, arm_neon_sha256h2, 0),
4633  NEONMAP1(vsha256hq_v, arm_neon_sha256h, 0),
4634  NEONMAP1(vsha256su0q_v, arm_neon_sha256su0, 0),
4635  NEONMAP1(vsha256su1q_v, arm_neon_sha256su1, 0),
4636  NEONMAP0(vshl_n_v),
4637  NEONMAP2(vshl_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts),
4638  NEONMAP0(vshll_n_v),
4639  NEONMAP0(vshlq_n_v),
4640  NEONMAP2(vshlq_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts),
4641  NEONMAP0(vshr_n_v),
4642  NEONMAP0(vshrn_n_v),
4643  NEONMAP0(vshrq_n_v),
4644  NEONMAP1(vst1_v, arm_neon_vst1, 0),
4645  NEONMAP1(vst1_x2_v, arm_neon_vst1x2, 0),
4646  NEONMAP1(vst1_x3_v, arm_neon_vst1x3, 0),
4647  NEONMAP1(vst1_x4_v, arm_neon_vst1x4, 0),
4648  NEONMAP1(vst1q_v, arm_neon_vst1, 0),
4649  NEONMAP1(vst1q_x2_v, arm_neon_vst1x2, 0),
4650  NEONMAP1(vst1q_x3_v, arm_neon_vst1x3, 0),
4651  NEONMAP1(vst1q_x4_v, arm_neon_vst1x4, 0),
4652  NEONMAP1(vst2_lane_v, arm_neon_vst2lane, 0),
4653  NEONMAP1(vst2_v, arm_neon_vst2, 0),
4654  NEONMAP1(vst2q_lane_v, arm_neon_vst2lane, 0),
4655  NEONMAP1(vst2q_v, arm_neon_vst2, 0),
4656  NEONMAP1(vst3_lane_v, arm_neon_vst3lane, 0),
4657  NEONMAP1(vst3_v, arm_neon_vst3, 0),
4658  NEONMAP1(vst3q_lane_v, arm_neon_vst3lane, 0),
4659  NEONMAP1(vst3q_v, arm_neon_vst3, 0),
4660  NEONMAP1(vst4_lane_v, arm_neon_vst4lane, 0),
4661  NEONMAP1(vst4_v, arm_neon_vst4, 0),
4662  NEONMAP1(vst4q_lane_v, arm_neon_vst4lane, 0),
4663  NEONMAP1(vst4q_v, arm_neon_vst4, 0),
4664  NEONMAP0(vsubhn_v),
4665  NEONMAP0(vtrn_v),
4666  NEONMAP0(vtrnq_v),
4667  NEONMAP0(vtst_v),
4668  NEONMAP0(vtstq_v),
4669  NEONMAP0(vuzp_v),
4670  NEONMAP0(vuzpq_v),
4671  NEONMAP0(vzip_v),
4672  NEONMAP0(vzipq_v)
4673 };
4674 
4675 static const NeonIntrinsicInfo AArch64SIMDIntrinsicMap[] = {
4676  NEONMAP1(vabs_v, aarch64_neon_abs, 0),
4677  NEONMAP1(vabsq_v, aarch64_neon_abs, 0),
4678  NEONMAP0(vaddhn_v),
4679  NEONMAP1(vaesdq_v, aarch64_crypto_aesd, 0),
4680  NEONMAP1(vaeseq_v, aarch64_crypto_aese, 0),
4681  NEONMAP1(vaesimcq_v, aarch64_crypto_aesimc, 0),
4682  NEONMAP1(vaesmcq_v, aarch64_crypto_aesmc, 0),
4683  NEONMAP1(vcage_v, aarch64_neon_facge, 0),
4684  NEONMAP1(vcageq_v, aarch64_neon_facge, 0),
4685  NEONMAP1(vcagt_v, aarch64_neon_facgt, 0),
4686  NEONMAP1(vcagtq_v, aarch64_neon_facgt, 0),
4687  NEONMAP1(vcale_v, aarch64_neon_facge, 0),
4688  NEONMAP1(vcaleq_v, aarch64_neon_facge, 0),
4689  NEONMAP1(vcalt_v, aarch64_neon_facgt, 0),
4690  NEONMAP1(vcaltq_v, aarch64_neon_facgt, 0),
4691  NEONMAP0(vceqz_v),
4692  NEONMAP0(vceqzq_v),
4693  NEONMAP0(vcgez_v),
4694  NEONMAP0(vcgezq_v),
4695  NEONMAP0(vcgtz_v),
4696  NEONMAP0(vcgtzq_v),
4697  NEONMAP0(vclez_v),
4698  NEONMAP0(vclezq_v),
4699  NEONMAP1(vcls_v, aarch64_neon_cls, Add1ArgType),
4700  NEONMAP1(vclsq_v, aarch64_neon_cls, Add1ArgType),
4701  NEONMAP0(vcltz_v),
4702  NEONMAP0(vcltzq_v),
4703  NEONMAP1(vclz_v, ctlz, Add1ArgType),
4704  NEONMAP1(vclzq_v, ctlz, Add1ArgType),
4705  NEONMAP1(vcnt_v, ctpop, Add1ArgType),
4706  NEONMAP1(vcntq_v, ctpop, Add1ArgType),
4707  NEONMAP1(vcvt_f16_f32, aarch64_neon_vcvtfp2hf, 0),
4708  NEONMAP0(vcvt_f16_v),
4709  NEONMAP1(vcvt_f32_f16, aarch64_neon_vcvthf2fp, 0),
4710  NEONMAP0(vcvt_f32_v),
4711  NEONMAP2(vcvt_n_f16_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
4712  NEONMAP2(vcvt_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
4713  NEONMAP2(vcvt_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
4714  NEONMAP1(vcvt_n_s16_v, aarch64_neon_vcvtfp2fxs, 0),
4715  NEONMAP1(vcvt_n_s32_v, aarch64_neon_vcvtfp2fxs, 0),
4716  NEONMAP1(vcvt_n_s64_v, aarch64_neon_vcvtfp2fxs, 0),
4717  NEONMAP1(vcvt_n_u16_v, aarch64_neon_vcvtfp2fxu, 0),
4718  NEONMAP1(vcvt_n_u32_v, aarch64_neon_vcvtfp2fxu, 0),
4719  NEONMAP1(vcvt_n_u64_v, aarch64_neon_vcvtfp2fxu, 0),
4720  NEONMAP0(vcvtq_f16_v),
4721  NEONMAP0(vcvtq_f32_v),
4722  NEONMAP2(vcvtq_n_f16_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
4723  NEONMAP2(vcvtq_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
4724  NEONMAP2(vcvtq_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
4725  NEONMAP1(vcvtq_n_s16_v, aarch64_neon_vcvtfp2fxs, 0),
4726  NEONMAP1(vcvtq_n_s32_v, aarch64_neon_vcvtfp2fxs, 0),
4727  NEONMAP1(vcvtq_n_s64_v, aarch64_neon_vcvtfp2fxs, 0),
4728  NEONMAP1(vcvtq_n_u16_v, aarch64_neon_vcvtfp2fxu, 0),
4729  NEONMAP1(vcvtq_n_u32_v, aarch64_neon_vcvtfp2fxu, 0),
4730  NEONMAP1(vcvtq_n_u64_v, aarch64_neon_vcvtfp2fxu, 0),
4731  NEONMAP1(vcvtx_f32_v, aarch64_neon_fcvtxn, AddRetType | Add1ArgType),
4732  NEONMAP2(vdot_v, aarch64_neon_udot, aarch64_neon_sdot, 0),
4733  NEONMAP2(vdotq_v, aarch64_neon_udot, aarch64_neon_sdot, 0),
4734  NEONMAP0(vext_v),
4735  NEONMAP0(vextq_v),
4736  NEONMAP0(vfma_v),
4737  NEONMAP0(vfmaq_v),
4738  NEONMAP1(vfmlal_high_v, aarch64_neon_fmlal2, 0),
4739  NEONMAP1(vfmlal_low_v, aarch64_neon_fmlal, 0),
4740  NEONMAP1(vfmlalq_high_v, aarch64_neon_fmlal2, 0),
4741  NEONMAP1(vfmlalq_low_v, aarch64_neon_fmlal, 0),
4742  NEONMAP1(vfmlsl_high_v, aarch64_neon_fmlsl2, 0),
4743  NEONMAP1(vfmlsl_low_v, aarch64_neon_fmlsl, 0),
4744  NEONMAP1(vfmlslq_high_v, aarch64_neon_fmlsl2, 0),
4745  NEONMAP1(vfmlslq_low_v, aarch64_neon_fmlsl, 0),
4746  NEONMAP2(vhadd_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts),
4747  NEONMAP2(vhaddq_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts),
4748  NEONMAP2(vhsub_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts),
4749  NEONMAP2(vhsubq_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts),
4750  NEONMAP1(vld1_x2_v, aarch64_neon_ld1x2, 0),
4751  NEONMAP1(vld1_x3_v, aarch64_neon_ld1x3, 0),
4752  NEONMAP1(vld1_x4_v, aarch64_neon_ld1x4, 0),
4753  NEONMAP1(vld1q_x2_v, aarch64_neon_ld1x2, 0),
4754  NEONMAP1(vld1q_x3_v, aarch64_neon_ld1x3, 0),
4755  NEONMAP1(vld1q_x4_v, aarch64_neon_ld1x4, 0),
4756  NEONMAP0(vmovl_v),
4757  NEONMAP0(vmovn_v),
4758  NEONMAP1(vmul_v, aarch64_neon_pmul, Add1ArgType),
4759  NEONMAP1(vmulq_v, aarch64_neon_pmul, Add1ArgType),
4760  NEONMAP1(vpadd_v, aarch64_neon_addp, Add1ArgType),
4761  NEONMAP2(vpaddl_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts),
4762  NEONMAP2(vpaddlq_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts),
4763  NEONMAP1(vpaddq_v, aarch64_neon_addp, Add1ArgType),
4764  NEONMAP1(vqabs_v, aarch64_neon_sqabs, Add1ArgType),
4765  NEONMAP1(vqabsq_v, aarch64_neon_sqabs, Add1ArgType),
4766  NEONMAP2(vqadd_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts),
4767  NEONMAP2(vqaddq_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts),
4768  NEONMAP2(vqdmlal_v, aarch64_neon_sqdmull, aarch64_neon_sqadd, 0),
4769  NEONMAP2(vqdmlsl_v, aarch64_neon_sqdmull, aarch64_neon_sqsub, 0),
4770  NEONMAP1(vqdmulh_v, aarch64_neon_sqdmulh, Add1ArgType),
4771  NEONMAP1(vqdmulhq_v, aarch64_neon_sqdmulh, Add1ArgType),
4772  NEONMAP1(vqdmull_v, aarch64_neon_sqdmull, Add1ArgType),
4773  NEONMAP2(vqmovn_v, aarch64_neon_uqxtn, aarch64_neon_sqxtn, Add1ArgType | UnsignedAlts),
4774  NEONMAP1(vqmovun_v, aarch64_neon_sqxtun, Add1ArgType),
4775  NEONMAP1(vqneg_v, aarch64_neon_sqneg, Add1ArgType),
4776  NEONMAP1(vqnegq_v, aarch64_neon_sqneg, Add1ArgType),
4777  NEONMAP1(vqrdmulh_v, aarch64_neon_sqrdmulh, Add1ArgType),
4778  NEONMAP1(vqrdmulhq_v, aarch64_neon_sqrdmulh, Add1ArgType),
4779  NEONMAP2(vqrshl_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts),
4780  NEONMAP2(vqrshlq_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts),
4781  NEONMAP2(vqshl_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl, UnsignedAlts),
4782  NEONMAP2(vqshl_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts),
4783  NEONMAP2(vqshlq_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl,UnsignedAlts),
4784  NEONMAP2(vqshlq_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts),
4785  NEONMAP1(vqshlu_n_v, aarch64_neon_sqshlu, 0),
4786  NEONMAP1(vqshluq_n_v, aarch64_neon_sqshlu, 0),
4787  NEONMAP2(vqsub_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts),
4788  NEONMAP2(vqsubq_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts),
4789  NEONMAP1(vraddhn_v, aarch64_neon_raddhn, Add1ArgType),
4790  NEONMAP2(vrecpe_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0),
4791  NEONMAP2(vrecpeq_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0),
4792  NEONMAP1(vrecps_v, aarch64_neon_frecps, Add1ArgType),
4793  NEONMAP1(vrecpsq_v, aarch64_neon_frecps, Add1ArgType),
4794  NEONMAP2(vrhadd_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts),
4795  NEONMAP2(vrhaddq_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts),
4796  NEONMAP0(vrndi_v),
4797  NEONMAP0(vrndiq_v),
4798  NEONMAP2(vrshl_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts),
4799  NEONMAP2(vrshlq_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts),
4800  NEONMAP2(vrshr_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts),
4801  NEONMAP2(vrshrq_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts),
4802  NEONMAP2(vrsqrte_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0),
4803  NEONMAP2(vrsqrteq_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0),
4804  NEONMAP1(vrsqrts_v, aarch64_neon_frsqrts, Add1ArgType),
4805  NEONMAP1(vrsqrtsq_v, aarch64_neon_frsqrts, Add1ArgType),
4806  NEONMAP1(vrsubhn_v, aarch64_neon_rsubhn, Add1ArgType),
4807  NEONMAP1(vsha1su0q_v, aarch64_crypto_sha1su0, 0),
4808  NEONMAP1(vsha1su1q_v, aarch64_crypto_sha1su1, 0),
4809  NEONMAP1(vsha256h2q_v, aarch64_crypto_sha256h2, 0),
4810  NEONMAP1(vsha256hq_v, aarch64_crypto_sha256h, 0),
4811  NEONMAP1(vsha256su0q_v, aarch64_crypto_sha256su0, 0),
4812  NEONMAP1(vsha256su1q_v, aarch64_crypto_sha256su1, 0),
4813  NEONMAP0(vshl_n_v),
4814  NEONMAP2(vshl_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts),
4815  NEONMAP0(vshll_n_v),
4816  NEONMAP0(vshlq_n_v),
4817  NEONMAP2(vshlq_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts),
4818  NEONMAP0(vshr_n_v),
4819  NEONMAP0(vshrn_n_v),
4820  NEONMAP0(vshrq_n_v),
4821  NEONMAP1(vst1_x2_v, aarch64_neon_st1x2, 0),
4822  NEONMAP1(vst1_x3_v, aarch64_neon_st1x3, 0),
4823  NEONMAP1(vst1_x4_v, aarch64_neon_st1x4, 0),
4824  NEONMAP1(vst1q_x2_v, aarch64_neon_st1x2, 0),
4825  NEONMAP1(vst1q_x3_v, aarch64_neon_st1x3, 0),
4826  NEONMAP1(vst1q_x4_v, aarch64_neon_st1x4, 0),
4827  NEONMAP0(vsubhn_v),
4828  NEONMAP0(vtst_v),
4829  NEONMAP0(vtstq_v),
4830 };
4831 
4832 static const NeonIntrinsicInfo AArch64SISDIntrinsicMap[] = {
4833  NEONMAP1(vabdd_f64, aarch64_sisd_fabd, Add1ArgType),
4834  NEONMAP1(vabds_f32, aarch64_sisd_fabd, Add1ArgType),
4835  NEONMAP1(vabsd_s64, aarch64_neon_abs, Add1ArgType),
4836  NEONMAP1(vaddlv_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType),
4837  NEONMAP1(vaddlv_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType),
4838  NEONMAP1(vaddlvq_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType),
4839  NEONMAP1(vaddlvq_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType),
4840  NEONMAP1(vaddv_f32, aarch64_neon_faddv, AddRetType | Add1ArgType),
4841  NEONMAP1(vaddv_s32, aarch64_neon_saddv, AddRetType | Add1ArgType),
4842  NEONMAP1(vaddv_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType),
4843  NEONMAP1(vaddvq_f32, aarch64_neon_faddv, AddRetType | Add1ArgType),
4844  NEONMAP1(vaddvq_f64, aarch64_neon_faddv, AddRetType | Add1ArgType),
4845  NEONMAP1(vaddvq_s32, aarch64_neon_saddv, AddRetType | Add1ArgType),
4846  NEONMAP1(vaddvq_s64, aarch64_neon_saddv, AddRetType | Add1ArgType),
4847  NEONMAP1(vaddvq_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType),
4848  NEONMAP1(vaddvq_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType),
4849  NEONMAP1(vcaged_f64, aarch64_neon_facge, AddRetType | Add1ArgType),
4850  NEONMAP1(vcages_f32, aarch64_neon_facge, AddRetType | Add1ArgType),
4851  NEONMAP1(vcagtd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType),
4852  NEONMAP1(vcagts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType),
4853  NEONMAP1(vcaled_f64, aarch64_neon_facge, AddRetType | Add1ArgType),
4854  NEONMAP1(vcales_f32, aarch64_neon_facge, AddRetType | Add1ArgType),
4855  NEONMAP1(vcaltd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType),
4856  NEONMAP1(vcalts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType),
4857  NEONMAP1(vcvtad_s64_f64, aarch64_neon_fcvtas, AddRetType | Add1ArgType),
4858  NEONMAP1(vcvtad_u64_f64, aarch64_neon_fcvtau, AddRetType | Add1ArgType),
4859  NEONMAP1(vcvtas_s32_f32, aarch64_neon_fcvtas, AddRetType | Add1ArgType),
4860  NEONMAP1(vcvtas_u32_f32, aarch64_neon_fcvtau, AddRetType | Add1ArgType),
4861  NEONMAP1(vcvtd_n_f64_s64, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType),
4862  NEONMAP1(vcvtd_n_f64_u64, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType),
4863  NEONMAP1(vcvtd_n_s64_f64, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType),
4864  NEONMAP1(vcvtd_n_u64_f64, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType),
4865  NEONMAP1(vcvtmd_s64_f64, aarch64_neon_fcvtms, AddRetType | Add1ArgType),
4866  NEONMAP1(vcvtmd_u64_f64, aarch64_neon_fcvtmu, AddRetType | Add1ArgType),
4867  NEONMAP1(vcvtms_s32_f32, aarch64_neon_fcvtms, AddRetType | Add1ArgType),
4868  NEONMAP1(vcvtms_u32_f32, aarch64_neon_fcvtmu, AddRetType | Add1ArgType),
4869  NEONMAP1(vcvtnd_s64_f64, aarch64_neon_fcvtns, AddRetType | Add1ArgType),
4870  NEONMAP1(vcvtnd_u64_f64, aarch64_neon_fcvtnu, AddRetType | Add1ArgType),
4871  NEONMAP1(vcvtns_s32_f32, aarch64_neon_fcvtns, AddRetType | Add1ArgType),
4872  NEONMAP1(vcvtns_u32_f32, aarch64_neon_fcvtnu, AddRetType | Add1ArgType),
4873  NEONMAP1(vcvtpd_s64_f64, aarch64_neon_fcvtps, AddRetType | Add1ArgType),
4874  NEONMAP1(vcvtpd_u64_f64, aarch64_neon_fcvtpu, AddRetType | Add1ArgType),
4875  NEONMAP1(vcvtps_s32_f32, aarch64_neon_fcvtps, AddRetType | Add1ArgType),
4876  NEONMAP1(vcvtps_u32_f32, aarch64_neon_fcvtpu, AddRetType | Add1ArgType),
4877  NEONMAP1(vcvts_n_f32_s32, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType),
4878  NEONMAP1(vcvts_n_f32_u32, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType),
4879  NEONMAP1(vcvts_n_s32_f32, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType),
4880  NEONMAP1(vcvts_n_u32_f32, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType),
4881  NEONMAP1(vcvtxd_f32_f64, aarch64_sisd_fcvtxn, 0),
4882  NEONMAP1(vmaxnmv_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType),
4883  NEONMAP1(vmaxnmvq_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType),
4884  NEONMAP1(vmaxnmvq_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType),
4885  NEONMAP1(vmaxv_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
4886  NEONMAP1(vmaxv_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType),
4887  NEONMAP1(vmaxv_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType),
4888  NEONMAP1(vmaxvq_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
4889  NEONMAP1(vmaxvq_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
4890  NEONMAP1(vmaxvq_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType),
4891  NEONMAP1(vmaxvq_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType),
4892  NEONMAP1(vminnmv_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
4893  NEONMAP1(vminnmvq_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
4894  NEONMAP1(vminnmvq_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
4895  NEONMAP1(vminv_f32, aarch64_neon_fminv, AddRetType | Add1ArgType),
4896  NEONMAP1(vminv_s32, aarch64_neon_sminv, AddRetType | Add1ArgType),
4897  NEONMAP1(vminv_u32, aarch64_neon_uminv, AddRetType | Add1ArgType),
4898  NEONMAP1(vminvq_f32, aarch64_neon_fminv, AddRetType | Add1ArgType),
4899  NEONMAP1(vminvq_f64, aarch64_neon_fminv, AddRetType | Add1ArgType),
4900  NEONMAP1(vminvq_s32, aarch64_neon_sminv, AddRetType | Add1ArgType),
4901  NEONMAP1(vminvq_u32, aarch64_neon_uminv, AddRetType | Add1ArgType),
4902  NEONMAP1(vmull_p64, aarch64_neon_pmull64, 0),
4903  NEONMAP1(vmulxd_f64, aarch64_neon_fmulx, Add1ArgType),
4904  NEONMAP1(vmulxs_f32, aarch64_neon_fmulx, Add1ArgType),
4905  NEONMAP1(vpaddd_s64, aarch64_neon_uaddv, AddRetType | Add1ArgType),
4906  NEONMAP1(vpaddd_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType),
4907  NEONMAP1(vpmaxnmqd_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType),
4908  NEONMAP1(vpmaxnms_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType),
4909  NEONMAP1(vpmaxqd_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
4910  NEONMAP1(vpmaxs_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
4911  NEONMAP1(vpminnmqd_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
4912  NEONMAP1(vpminnms_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
4913  NEONMAP1(vpminqd_f64, aarch64_neon_fminv, AddRetType | Add1ArgType),
4914  NEONMAP1(vpmins_f32, aarch64_neon_fminv, AddRetType | Add1ArgType),
4915  NEONMAP1(vqabsb_s8, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors),
4916  NEONMAP1(vqabsd_s64, aarch64_neon_sqabs, Add1ArgType),
4917  NEONMAP1(vqabsh_s16, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors),
4918  NEONMAP1(vqabss_s32, aarch64_neon_sqabs, Add1ArgType),
4919  NEONMAP1(vqaddb_s8, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors),
4920  NEONMAP1(vqaddb_u8, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors),
4921  NEONMAP1(vqaddd_s64, aarch64_neon_sqadd, Add1ArgType),
4922  NEONMAP1(vqaddd_u64, aarch64_neon_uqadd, Add1ArgType),
4923  NEONMAP1(vqaddh_s16, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors),
4924  NEONMAP1(vqaddh_u16, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors),
4925  NEONMAP1(vqadds_s32, aarch64_neon_sqadd, Add1ArgType),
4926  NEONMAP1(vqadds_u32, aarch64_neon_uqadd, Add1ArgType),
4927  NEONMAP1(vqdmulhh_s16, aarch64_neon_sqdmulh, Vectorize1ArgType | Use64BitVectors),
4928  NEONMAP1(vqdmulhs_s32, aarch64_neon_sqdmulh, Add1ArgType),
4929  NEONMAP1(vqdmullh_s16, aarch64_neon_sqdmull, VectorRet | Use128BitVectors),
4930  NEONMAP1(vqdmulls_s32, aarch64_neon_sqdmulls_scalar, 0),
4931  NEONMAP1(vqmovnd_s64, aarch64_neon_scalar_sqxtn, AddRetType | Add1ArgType),
4932  NEONMAP1(vqmovnd_u64, aarch64_neon_scalar_uqxtn, AddRetType | Add1ArgType),
4933  NEONMAP1(vqmovnh_s16, aarch64_neon_sqxtn, VectorRet | Use64BitVectors),
4934  NEONMAP1(vqmovnh_u16, aarch64_neon_uqxtn, VectorRet | Use64BitVectors),
4935  NEONMAP1(vqmovns_s32, aarch64_neon_sqxtn, VectorRet | Use64BitVectors),
4936  NEONMAP1(vqmovns_u32, aarch64_neon_uqxtn, VectorRet | Use64BitVectors),
4937  NEONMAP1(vqmovund_s64, aarch64_neon_scalar_sqxtun, AddRetType | Add1ArgType),
4938  NEONMAP1(vqmovunh_s16, aarch64_neon_sqxtun, VectorRet | Use64BitVectors),
4939  NEONMAP1(vqmovuns_s32, aarch64_neon_sqxtun, VectorRet | Use64BitVectors),
4940  NEONMAP1(vqnegb_s8, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors),
4941  NEONMAP1(vqnegd_s64, aarch64_neon_sqneg, Add1ArgType),
4942  NEONMAP1(vqnegh_s16, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors),
4943  NEONMAP1(vqnegs_s32, aarch64_neon_sqneg, Add1ArgType),
4944  NEONMAP1(vqrdmulhh_s16, aarch64_neon_sqrdmulh, Vectorize1ArgType | Use64BitVectors),
4945  NEONMAP1(vqrdmulhs_s32, aarch64_neon_sqrdmulh, Add1ArgType),
4946  NEONMAP1(vqrshlb_s8, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors),
4947  NEONMAP1(vqrshlb_u8, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors),
4948  NEONMAP1(vqrshld_s64, aarch64_neon_sqrshl, Add1ArgType),
4949  NEONMAP1(vqrshld_u64, aarch64_neon_uqrshl, Add1ArgType),
4950  NEONMAP1(vqrshlh_s16, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors),
4951  NEONMAP1(vqrshlh_u16, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors),
4952  NEONMAP1(vqrshls_s32, aarch64_neon_sqrshl, Add1ArgType),
4953  NEONMAP1(vqrshls_u32, aarch64_neon_uqrshl, Add1ArgType),
4954  NEONMAP1(vqrshrnd_n_s64, aarch64_neon_sqrshrn, AddRetType),
4955  NEONMAP1(vqrshrnd_n_u64, aarch64_neon_uqrshrn, AddRetType),
4956  NEONMAP1(vqrshrnh_n_s16, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors),
4957  NEONMAP1(vqrshrnh_n_u16, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors),
4958  NEONMAP1(vqrshrns_n_s32, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors),
4959  NEONMAP1(vqrshrns_n_u32, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors),
4960  NEONMAP1(vqrshrund_n_s64, aarch64_neon_sqrshrun, AddRetType),
4961  NEONMAP1(vqrshrunh_n_s16, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors),
4962  NEONMAP1(vqrshruns_n_s32, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors),
4963  NEONMAP1(vqshlb_n_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors),
4964  NEONMAP1(vqshlb_n_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors),
4965  NEONMAP1(vqshlb_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors),
4966  NEONMAP1(vqshlb_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors),
4967  NEONMAP1(vqshld_s64, aarch64_neon_sqshl, Add1ArgType),
4968  NEONMAP1(vqshld_u64, aarch64_neon_uqshl, Add1ArgType),
4969  NEONMAP1(vqshlh_n_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors),
4970  NEONMAP1(vqshlh_n_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors),
4971  NEONMAP1(vqshlh_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors),
4972  NEONMAP1(vqshlh_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors),
4973  NEONMAP1(vqshls_n_s32, aarch64_neon_sqshl, Add1ArgType),
4974  NEONMAP1(vqshls_n_u32, aarch64_neon_uqshl, Add1ArgType),
4975  NEONMAP1(vqshls_s32, aarch64_neon_sqshl, Add1ArgType),
4976  NEONMAP1(vqshls_u32, aarch64_neon_uqshl, Add1ArgType),
4977  NEONMAP1(vqshlub_n_s8, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors),
4978  NEONMAP1(vqshluh_n_s16, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors),
4979  NEONMAP1(vqshlus_n_s32, aarch64_neon_sqshlu, Add1ArgType),
4980  NEONMAP1(vqshrnd_n_s64, aarch64_neon_sqshrn, AddRetType),
4981  NEONMAP1(vqshrnd_n_u64, aarch64_neon_uqshrn, AddRetType),
4982  NEONMAP1(vqshrnh_n_s16, aarch64_neon_sqshrn, VectorRet | Use64BitVectors),
4983  NEONMAP1(vqshrnh_n_u16, aarch64_neon_uqshrn, VectorRet | Use64BitVectors),
4984  NEONMAP1(vqshrns_n_s32, aarch64_neon_sqshrn, VectorRet | Use64BitVectors),
4985  NEONMAP1(vqshrns_n_u32, aarch64_neon_uqshrn, VectorRet | Use64BitVectors),
4986  NEONMAP1(vqshrund_n_s64, aarch64_neon_sqshrun, AddRetType),
4987  NEONMAP1(vqshrunh_n_s16, aarch64_neon_sqshrun, VectorRet | Use64BitVectors),
4988  NEONMAP1(vqshruns_n_s32, aarch64_neon_sqshrun, VectorRet | Use64BitVectors),
4989  NEONMAP1(vqsubb_s8, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors),
4990  NEONMAP1(vqsubb_u8, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors),
4991  NEONMAP1(vqsubd_s64, aarch64_neon_sqsub, Add1ArgType),
4992  NEONMAP1(vqsubd_u64, aarch64_neon_uqsub, Add1ArgType),
4993  NEONMAP1(vqsubh_s16, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors),
4994  NEONMAP1(vqsubh_u16, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors),
4995  NEONMAP1(vqsubs_s32, aarch64_neon_sqsub, Add1ArgType),
4996  NEONMAP1(vqsubs_u32, aarch64_neon_uqsub, Add1ArgType),
4997  NEONMAP1(vrecped_f64, aarch64_neon_frecpe, Add1ArgType),
4998  NEONMAP1(vrecpes_f32, aarch64_neon_frecpe, Add1ArgType),
4999  NEONMAP1(vrecpxd_f64, aarch64_neon_frecpx, Add1ArgType),
5000  NEONMAP1(vrecpxs_f32, aarch64_neon_frecpx, Add1ArgType),
5001  NEONMAP1(vrshld_s64, aarch64_neon_srshl, Add1ArgType),
5002  NEONMAP1(vrshld_u64, aarch64_neon_urshl, Add1ArgType),
5003  NEONMAP1(vrsqrted_f64, aarch64_neon_frsqrte, Add1ArgType),
5004  NEONMAP1(vrsqrtes_f32, aarch64_neon_frsqrte, Add1ArgType),
5005  NEONMAP1(vrsqrtsd_f64, aarch64_neon_frsqrts, Add1ArgType),
5006  NEONMAP1(vrsqrtss_f32, aarch64_neon_frsqrts, Add1ArgType),
5007  NEONMAP1(vsha1cq_u32, aarch64_crypto_sha1c, 0),
5008  NEONMAP1(vsha1h_u32, aarch64_crypto_sha1h, 0),
5009  NEONMAP1(vsha1mq_u32, aarch64_crypto_sha1m, 0),
5010  NEONMAP1(vsha1pq_u32, aarch64_crypto_sha1p, 0),
5011  NEONMAP1(vshld_s64, aarch64_neon_sshl, Add1ArgType),
5012  NEONMAP1(vshld_u64, aarch64_neon_ushl, Add1ArgType),
5013  NEONMAP1(vslid_n_s64, aarch64_neon_vsli, Vectorize1ArgType),
5014  NEONMAP1(vslid_n_u64, aarch64_neon_vsli, Vectorize1ArgType),
5015  NEONMAP1(vsqaddb_u8, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors),
5016  NEONMAP1(vsqaddd_u64, aarch64_neon_usqadd, Add1ArgType),
5017  NEONMAP1(vsqaddh_u16, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors),
5018  NEONMAP1(vsqadds_u32, aarch64_neon_usqadd, Add1ArgType),
5019  NEONMAP1(vsrid_n_s64, aarch64_neon_vsri, Vectorize1ArgType),
5020  NEONMAP1(vsrid_n_u64, aarch64_neon_vsri, Vectorize1ArgType),
5021  NEONMAP1(vuqaddb_s8, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors),
5022  NEONMAP1(vuqaddd_s64, aarch64_neon_suqadd, Add1ArgType),
5023  NEONMAP1(vuqaddh_s16, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors),
5024  NEONMAP1(vuqadds_s32, aarch64_neon_suqadd, Add1ArgType),
5025  // FP16 scalar intrinisics go here.
5026  NEONMAP1(vabdh_f16, aarch64_sisd_fabd, Add1ArgType),
5027  NEONMAP1(vcvtah_s32_f16, aarch64_neon_fcvtas, AddRetType | Add1ArgType),
5028  NEONMAP1(vcvtah_s64_f16, aarch64_neon_fcvtas, AddRetType | Add1ArgType),
5029  NEONMAP1(vcvtah_u32_f16, aarch64_neon_fcvtau, AddRetType | Add1ArgType),
5030  NEONMAP1(vcvtah_u64_f16, aarch64_neon_fcvtau, AddRetType | Add1ArgType),
5031  NEONMAP1(vcvth_n_f16_s32, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType),
5032  NEONMAP1(vcvth_n_f16_s64, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType),
5033  NEONMAP1(vcvth_n_f16_u32, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType),
5034  NEONMAP1(vcvth_n_f16_u64, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType),
5035  NEONMAP1(vcvth_n_s32_f16, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType),
5036  NEONMAP1(vcvth_n_s64_f16, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType),
5037  NEONMAP1(vcvth_n_u32_f16, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType),
5038  NEONMAP1(vcvth_n_u64_f16, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType),
5039  NEONMAP1(vcvtmh_s32_f16, aarch64_neon_fcvtms, AddRetType | Add1ArgType),
5040  NEONMAP1(vcvtmh_s64_f16, aarch64_neon_fcvtms, AddRetType | Add1ArgType),
5041  NEONMAP1(vcvtmh_u32_f16, aarch64_neon_fcvtmu, AddRetType | Add1ArgType),
5042  NEONMAP1(vcvtmh_u64_f16, aarch64_neon_fcvtmu, AddRetType | Add1ArgType),
5043  NEONMAP1(vcvtnh_s32_f16, aarch64_neon_fcvtns, AddRetType | Add1ArgType),
5044  NEONMAP1(vcvtnh_s64_f16, aarch64_neon_fcvtns, AddRetType | Add1ArgType),
5045  NEONMAP1(vcvtnh_u32_f16, aarch64_neon_fcvtnu, AddRetType | Add1ArgType),
5046  NEONMAP1(vcvtnh_u64_f16, aarch64_neon_fcvtnu, AddRetType | Add1ArgType),
5047  NEONMAP1(vcvtph_s32_f16, aarch64_neon_fcvtps, AddRetType | Add1ArgType),
5048  NEONMAP1(vcvtph_s64_f16, aarch64_neon_fcvtps, AddRetType | Add1ArgType),
5049  NEONMAP1(vcvtph_u32_f16, aarch64_neon_fcvtpu, AddRetType | Add1ArgType),
5050  NEONMAP1(vcvtph_u64_f16, aarch64_neon_fcvtpu, AddRetType | Add1ArgType),
5051  NEONMAP1(vmulxh_f16, aarch64_neon_fmulx, Add1ArgType),
5052  NEONMAP1(vrecpeh_f16, aarch64_neon_frecpe, Add1ArgType),
5053  NEONMAP1(vrecpxh_f16, aarch64_neon_frecpx, Add1ArgType),
5054  NEONMAP1(vrsqrteh_f16, aarch64_neon_frsqrte, Add1ArgType),
5055  NEONMAP1(vrsqrtsh_f16, aarch64_neon_frsqrts, Add1ArgType),
5056 };
5057 
5058 #undef NEONMAP0
5059 #undef NEONMAP1
5060 #undef NEONMAP2
5061 
5063 
5066 
5067 
5068 static const NeonIntrinsicInfo *
5070  unsigned BuiltinID, bool &MapProvenSorted) {
5071 
5072 #ifndef NDEBUG
5073  if (!MapProvenSorted) {
5074  assert(std::is_sorted(std::begin(IntrinsicMap), std::end(IntrinsicMap)));
5075  MapProvenSorted = true;
5076  }
5077 #endif
5078 
5079  const NeonIntrinsicInfo *Builtin =
5080  std::lower_bound(IntrinsicMap.begin(), IntrinsicMap.end(), BuiltinID);
5081 
5082  if (Builtin != IntrinsicMap.end() && Builtin->BuiltinID == BuiltinID)
5083  return Builtin;
5084 
5085  return nullptr;
5086 }
5087 
5088 Function *CodeGenFunction::LookupNeonLLVMIntrinsic(unsigned IntrinsicID,
5089  unsigned Modifier,
5090  llvm::Type *ArgType,
5091  const CallExpr *E) {
5092  int VectorSize = 0;
5093  if (Modifier & Use64BitVectors)
5094  VectorSize = 64;
5095  else if (Modifier & Use128BitVectors)
5096  VectorSize = 128;