1 //===---- CGBuiltin.cpp - Emit LLVM Code for builtins ---------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This contains code to emit Builtin calls as LLVM code.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "CodeGenFunction.h"
15 #include "CGCXXABI.h"
16 #include "CGObjCRuntime.h"
17 #include "CodeGenModule.h"
18 #include "TargetInfo.h"
19 #include "clang/AST/ASTContext.h"
20 #include "clang/AST/Decl.h"
21 #include "clang/Basic/TargetBuiltins.h"
22 #include "clang/Basic/TargetInfo.h"
23 #include "clang/CodeGen/CGFunctionInfo.h"
24 #include "llvm/ADT/StringExtras.h"
25 #include "llvm/IR/CallSite.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/InlineAsm.h"
28 #include "llvm/IR/Intrinsics.h"
29 #include <sstream>
30
31 using namespace clang;
32 using namespace CodeGen;
33 using namespace llvm;
34
35 /// getBuiltinLibFunction - Given a builtin id for a function like
36 /// "__builtin_fabsf", return a Function* for "fabsf".
getBuiltinLibFunction(const FunctionDecl * FD,unsigned BuiltinID)37 llvm::Value *CodeGenModule::getBuiltinLibFunction(const FunctionDecl *FD,
38 unsigned BuiltinID) {
39 assert(Context.BuiltinInfo.isLibFunction(BuiltinID));
40
41 // Get the name, skip over the __builtin_ prefix (if necessary).
42 StringRef Name;
43 GlobalDecl D(FD);
44
45 // If the builtin has been declared explicitly with an assembler label,
46 // use the mangled name. This differs from the plain label on platforms
47 // that prefix labels.
48 if (FD->hasAttr<AsmLabelAttr>())
49 Name = getMangledName(D);
50 else
51 Name = Context.BuiltinInfo.GetName(BuiltinID) + 10;
52
53 llvm::FunctionType *Ty =
54 cast<llvm::FunctionType>(getTypes().ConvertType(FD->getType()));
55
56 return GetOrCreateLLVMFunction(Name, Ty, D, /*ForVTable=*/false);
57 }
58
59 /// Emit the conversions required to turn the given value into an
60 /// integer of the given size.
EmitToInt(CodeGenFunction & CGF,llvm::Value * V,QualType T,llvm::IntegerType * IntType)61 static Value *EmitToInt(CodeGenFunction &CGF, llvm::Value *V,
62 QualType T, llvm::IntegerType *IntType) {
63 V = CGF.EmitToMemory(V, T);
64
65 if (V->getType()->isPointerTy())
66 return CGF.Builder.CreatePtrToInt(V, IntType);
67
68 assert(V->getType() == IntType);
69 return V;
70 }
71
EmitFromInt(CodeGenFunction & CGF,llvm::Value * V,QualType T,llvm::Type * ResultType)72 static Value *EmitFromInt(CodeGenFunction &CGF, llvm::Value *V,
73 QualType T, llvm::Type *ResultType) {
74 V = CGF.EmitFromMemory(V, T);
75
76 if (ResultType->isPointerTy())
77 return CGF.Builder.CreateIntToPtr(V, ResultType);
78
79 assert(V->getType() == ResultType);
80 return V;
81 }
82
83 /// Utility to insert an atomic instruction based on Instrinsic::ID
84 /// and the expression node.
EmitBinaryAtomic(CodeGenFunction & CGF,llvm::AtomicRMWInst::BinOp Kind,const CallExpr * E)85 static RValue EmitBinaryAtomic(CodeGenFunction &CGF,
86 llvm::AtomicRMWInst::BinOp Kind,
87 const CallExpr *E) {
88 QualType T = E->getType();
89 assert(E->getArg(0)->getType()->isPointerType());
90 assert(CGF.getContext().hasSameUnqualifiedType(T,
91 E->getArg(0)->getType()->getPointeeType()));
92 assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType()));
93
94 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0));
95 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace();
96
97 llvm::IntegerType *IntType =
98 llvm::IntegerType::get(CGF.getLLVMContext(),
99 CGF.getContext().getTypeSize(T));
100 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace);
101
102 llvm::Value *Args[2];
103 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType);
104 Args[1] = CGF.EmitScalarExpr(E->getArg(1));
105 llvm::Type *ValueType = Args[1]->getType();
106 Args[1] = EmitToInt(CGF, Args[1], T, IntType);
107
108 llvm::Value *Result =
109 CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1],
110 llvm::SequentiallyConsistent);
111 Result = EmitFromInt(CGF, Result, T, ValueType);
112 return RValue::get(Result);
113 }
114
115 /// Utility to insert an atomic instruction based Instrinsic::ID and
116 /// the expression node, where the return value is the result of the
117 /// operation.
EmitBinaryAtomicPost(CodeGenFunction & CGF,llvm::AtomicRMWInst::BinOp Kind,const CallExpr * E,Instruction::BinaryOps Op,bool Invert=false)118 static RValue EmitBinaryAtomicPost(CodeGenFunction &CGF,
119 llvm::AtomicRMWInst::BinOp Kind,
120 const CallExpr *E,
121 Instruction::BinaryOps Op,
122 bool Invert = false) {
123 QualType T = E->getType();
124 assert(E->getArg(0)->getType()->isPointerType());
125 assert(CGF.getContext().hasSameUnqualifiedType(T,
126 E->getArg(0)->getType()->getPointeeType()));
127 assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType()));
128
129 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0));
130 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace();
131
132 llvm::IntegerType *IntType =
133 llvm::IntegerType::get(CGF.getLLVMContext(),
134 CGF.getContext().getTypeSize(T));
135 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace);
136
137 llvm::Value *Args[2];
138 Args[1] = CGF.EmitScalarExpr(E->getArg(1));
139 llvm::Type *ValueType = Args[1]->getType();
140 Args[1] = EmitToInt(CGF, Args[1], T, IntType);
141 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType);
142
143 llvm::Value *Result =
144 CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1],
145 llvm::SequentiallyConsistent);
146 Result = CGF.Builder.CreateBinOp(Op, Result, Args[1]);
147 if (Invert)
148 Result = CGF.Builder.CreateBinOp(llvm::Instruction::Xor, Result,
149 llvm::ConstantInt::get(IntType, -1));
150 Result = EmitFromInt(CGF, Result, T, ValueType);
151 return RValue::get(Result);
152 }
153
154 /// EmitFAbs - Emit a call to @llvm.fabs().
EmitFAbs(CodeGenFunction & CGF,Value * V)155 static Value *EmitFAbs(CodeGenFunction &CGF, Value *V) {
156 Value *F = CGF.CGM.getIntrinsic(Intrinsic::fabs, V->getType());
157 llvm::CallInst *Call = CGF.Builder.CreateCall(F, V);
158 Call->setDoesNotAccessMemory();
159 return Call;
160 }
161
162 /// Emit the computation of the sign bit for a floating point value. Returns
163 /// the i1 sign bit value.
EmitSignBit(CodeGenFunction & CGF,Value * V)164 static Value *EmitSignBit(CodeGenFunction &CGF, Value *V) {
165 LLVMContext &C = CGF.CGM.getLLVMContext();
166
167 llvm::Type *Ty = V->getType();
168 int Width = Ty->getPrimitiveSizeInBits();
169 llvm::Type *IntTy = llvm::IntegerType::get(C, Width);
170 V = CGF.Builder.CreateBitCast(V, IntTy);
171 if (Ty->isPPC_FP128Ty()) {
172 // The higher-order double comes first, and so we need to truncate the
173 // pair to extract the overall sign. The order of the pair is the same
174 // in both little- and big-Endian modes.
175 Width >>= 1;
176 IntTy = llvm::IntegerType::get(C, Width);
177 V = CGF.Builder.CreateTrunc(V, IntTy);
178 }
179 Value *Zero = llvm::Constant::getNullValue(IntTy);
180 return CGF.Builder.CreateICmpSLT(V, Zero);
181 }
182
emitLibraryCall(CodeGenFunction & CGF,const FunctionDecl * Fn,const CallExpr * E,llvm::Value * calleeValue)183 static RValue emitLibraryCall(CodeGenFunction &CGF, const FunctionDecl *Fn,
184 const CallExpr *E, llvm::Value *calleeValue) {
185 return CGF.EmitCall(E->getCallee()->getType(), calleeValue, E,
186 ReturnValueSlot(), Fn);
187 }
188
189 /// \brief Emit a call to llvm.{sadd,uadd,ssub,usub,smul,umul}.with.overflow.*
190 /// depending on IntrinsicID.
191 ///
192 /// \arg CGF The current codegen function.
193 /// \arg IntrinsicID The ID for the Intrinsic we wish to generate.
194 /// \arg X The first argument to the llvm.*.with.overflow.*.
195 /// \arg Y The second argument to the llvm.*.with.overflow.*.
196 /// \arg Carry The carry returned by the llvm.*.with.overflow.*.
197 /// \returns The result (i.e. sum/product) returned by the intrinsic.
EmitOverflowIntrinsic(CodeGenFunction & CGF,const llvm::Intrinsic::ID IntrinsicID,llvm::Value * X,llvm::Value * Y,llvm::Value * & Carry)198 static llvm::Value *EmitOverflowIntrinsic(CodeGenFunction &CGF,
199 const llvm::Intrinsic::ID IntrinsicID,
200 llvm::Value *X, llvm::Value *Y,
201 llvm::Value *&Carry) {
202 // Make sure we have integers of the same width.
203 assert(X->getType() == Y->getType() &&
204 "Arguments must be the same type. (Did you forget to make sure both "
205 "arguments have the same integer width?)");
206
207 llvm::Value *Callee = CGF.CGM.getIntrinsic(IntrinsicID, X->getType());
208 llvm::Value *Tmp = CGF.Builder.CreateCall2(Callee, X, Y);
209 Carry = CGF.Builder.CreateExtractValue(Tmp, 1);
210 return CGF.Builder.CreateExtractValue(Tmp, 0);
211 }
212
EmitBuiltinExpr(const FunctionDecl * FD,unsigned BuiltinID,const CallExpr * E,ReturnValueSlot ReturnValue)213 RValue CodeGenFunction::EmitBuiltinExpr(const FunctionDecl *FD,
214 unsigned BuiltinID, const CallExpr *E,
215 ReturnValueSlot ReturnValue) {
216 // See if we can constant fold this builtin. If so, don't emit it at all.
217 Expr::EvalResult Result;
218 if (E->EvaluateAsRValue(Result, CGM.getContext()) &&
219 !Result.hasSideEffects()) {
220 if (Result.Val.isInt())
221 return RValue::get(llvm::ConstantInt::get(getLLVMContext(),
222 Result.Val.getInt()));
223 if (Result.Val.isFloat())
224 return RValue::get(llvm::ConstantFP::get(getLLVMContext(),
225 Result.Val.getFloat()));
226 }
227
228 switch (BuiltinID) {
229 default: break; // Handle intrinsics and libm functions below.
230 case Builtin::BI__builtin___CFStringMakeConstantString:
231 case Builtin::BI__builtin___NSStringMakeConstantString:
232 return RValue::get(CGM.EmitConstantExpr(E, E->getType(), nullptr));
233 case Builtin::BI__builtin_stdarg_start:
234 case Builtin::BI__builtin_va_start:
235 case Builtin::BI__va_start:
236 case Builtin::BI__builtin_va_end: {
237 Value *ArgValue = (BuiltinID == Builtin::BI__va_start)
238 ? EmitScalarExpr(E->getArg(0))
239 : EmitVAListRef(E->getArg(0));
240 llvm::Type *DestType = Int8PtrTy;
241 if (ArgValue->getType() != DestType)
242 ArgValue = Builder.CreateBitCast(ArgValue, DestType,
243 ArgValue->getName().data());
244
245 Intrinsic::ID inst = (BuiltinID == Builtin::BI__builtin_va_end) ?
246 Intrinsic::vaend : Intrinsic::vastart;
247 return RValue::get(Builder.CreateCall(CGM.getIntrinsic(inst), ArgValue));
248 }
249 case Builtin::BI__builtin_va_copy: {
250 Value *DstPtr = EmitVAListRef(E->getArg(0));
251 Value *SrcPtr = EmitVAListRef(E->getArg(1));
252
253 llvm::Type *Type = Int8PtrTy;
254
255 DstPtr = Builder.CreateBitCast(DstPtr, Type);
256 SrcPtr = Builder.CreateBitCast(SrcPtr, Type);
257 return RValue::get(Builder.CreateCall2(CGM.getIntrinsic(Intrinsic::vacopy),
258 DstPtr, SrcPtr));
259 }
260 case Builtin::BI__builtin_abs:
261 case Builtin::BI__builtin_labs:
262 case Builtin::BI__builtin_llabs: {
263 Value *ArgValue = EmitScalarExpr(E->getArg(0));
264
265 Value *NegOp = Builder.CreateNeg(ArgValue, "neg");
266 Value *CmpResult =
267 Builder.CreateICmpSGE(ArgValue,
268 llvm::Constant::getNullValue(ArgValue->getType()),
269 "abscond");
270 Value *Result =
271 Builder.CreateSelect(CmpResult, ArgValue, NegOp, "abs");
272
273 return RValue::get(Result);
274 }
275 case Builtin::BI__builtin_fabs:
276 case Builtin::BI__builtin_fabsf:
277 case Builtin::BI__builtin_fabsl: {
278 Value *Arg1 = EmitScalarExpr(E->getArg(0));
279 Value *Result = EmitFAbs(*this, Arg1);
280 return RValue::get(Result);
281 }
282 case Builtin::BI__builtin_fmod:
283 case Builtin::BI__builtin_fmodf:
284 case Builtin::BI__builtin_fmodl: {
285 Value *Arg1 = EmitScalarExpr(E->getArg(0));
286 Value *Arg2 = EmitScalarExpr(E->getArg(1));
287 Value *Result = Builder.CreateFRem(Arg1, Arg2, "fmod");
288 return RValue::get(Result);
289 }
290
291 case Builtin::BI__builtin_conj:
292 case Builtin::BI__builtin_conjf:
293 case Builtin::BI__builtin_conjl: {
294 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0));
295 Value *Real = ComplexVal.first;
296 Value *Imag = ComplexVal.second;
297 Value *Zero =
298 Imag->getType()->isFPOrFPVectorTy()
299 ? llvm::ConstantFP::getZeroValueForNegation(Imag->getType())
300 : llvm::Constant::getNullValue(Imag->getType());
301
302 Imag = Builder.CreateFSub(Zero, Imag, "sub");
303 return RValue::getComplex(std::make_pair(Real, Imag));
304 }
305 case Builtin::BI__builtin_creal:
306 case Builtin::BI__builtin_crealf:
307 case Builtin::BI__builtin_creall:
308 case Builtin::BIcreal:
309 case Builtin::BIcrealf:
310 case Builtin::BIcreall: {
311 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0));
312 return RValue::get(ComplexVal.first);
313 }
314
315 case Builtin::BI__builtin_cimag:
316 case Builtin::BI__builtin_cimagf:
317 case Builtin::BI__builtin_cimagl:
318 case Builtin::BIcimag:
319 case Builtin::BIcimagf:
320 case Builtin::BIcimagl: {
321 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0));
322 return RValue::get(ComplexVal.second);
323 }
324
325 case Builtin::BI__builtin_ctzs:
326 case Builtin::BI__builtin_ctz:
327 case Builtin::BI__builtin_ctzl:
328 case Builtin::BI__builtin_ctzll: {
329 Value *ArgValue = EmitScalarExpr(E->getArg(0));
330
331 llvm::Type *ArgType = ArgValue->getType();
332 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType);
333
334 llvm::Type *ResultType = ConvertType(E->getType());
335 Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef());
336 Value *Result = Builder.CreateCall2(F, ArgValue, ZeroUndef);
337 if (Result->getType() != ResultType)
338 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
339 "cast");
340 return RValue::get(Result);
341 }
342 case Builtin::BI__builtin_clzs:
343 case Builtin::BI__builtin_clz:
344 case Builtin::BI__builtin_clzl:
345 case Builtin::BI__builtin_clzll: {
346 Value *ArgValue = EmitScalarExpr(E->getArg(0));
347
348 llvm::Type *ArgType = ArgValue->getType();
349 Value *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType);
350
351 llvm::Type *ResultType = ConvertType(E->getType());
352 Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef());
353 Value *Result = Builder.CreateCall2(F, ArgValue, ZeroUndef);
354 if (Result->getType() != ResultType)
355 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
356 "cast");
357 return RValue::get(Result);
358 }
359 case Builtin::BI__builtin_ffs:
360 case Builtin::BI__builtin_ffsl:
361 case Builtin::BI__builtin_ffsll: {
362 // ffs(x) -> x ? cttz(x) + 1 : 0
363 Value *ArgValue = EmitScalarExpr(E->getArg(0));
364
365 llvm::Type *ArgType = ArgValue->getType();
366 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType);
367
368 llvm::Type *ResultType = ConvertType(E->getType());
369 Value *Tmp = Builder.CreateAdd(Builder.CreateCall2(F, ArgValue,
370 Builder.getTrue()),
371 llvm::ConstantInt::get(ArgType, 1));
372 Value *Zero = llvm::Constant::getNullValue(ArgType);
373 Value *IsZero = Builder.CreateICmpEQ(ArgValue, Zero, "iszero");
374 Value *Result = Builder.CreateSelect(IsZero, Zero, Tmp, "ffs");
375 if (Result->getType() != ResultType)
376 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
377 "cast");
378 return RValue::get(Result);
379 }
380 case Builtin::BI__builtin_parity:
381 case Builtin::BI__builtin_parityl:
382 case Builtin::BI__builtin_parityll: {
383 // parity(x) -> ctpop(x) & 1
384 Value *ArgValue = EmitScalarExpr(E->getArg(0));
385
386 llvm::Type *ArgType = ArgValue->getType();
387 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType);
388
389 llvm::Type *ResultType = ConvertType(E->getType());
390 Value *Tmp = Builder.CreateCall(F, ArgValue);
391 Value *Result = Builder.CreateAnd(Tmp, llvm::ConstantInt::get(ArgType, 1));
392 if (Result->getType() != ResultType)
393 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
394 "cast");
395 return RValue::get(Result);
396 }
397 case Builtin::BI__builtin_popcount:
398 case Builtin::BI__builtin_popcountl:
399 case Builtin::BI__builtin_popcountll: {
400 Value *ArgValue = EmitScalarExpr(E->getArg(0));
401
402 llvm::Type *ArgType = ArgValue->getType();
403 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType);
404
405 llvm::Type *ResultType = ConvertType(E->getType());
406 Value *Result = Builder.CreateCall(F, ArgValue);
407 if (Result->getType() != ResultType)
408 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
409 "cast");
410 return RValue::get(Result);
411 }
412 case Builtin::BI__builtin_expect: {
413 Value *ArgValue = EmitScalarExpr(E->getArg(0));
414 llvm::Type *ArgType = ArgValue->getType();
415
416 Value *ExpectedValue = EmitScalarExpr(E->getArg(1));
417 // Don't generate llvm.expect on -O0 as the backend won't use it for
418 // anything.
419 // Note, we still IRGen ExpectedValue because it could have side-effects.
420 if (CGM.getCodeGenOpts().OptimizationLevel == 0)
421 return RValue::get(ArgValue);
422
423 Value *FnExpect = CGM.getIntrinsic(Intrinsic::expect, ArgType);
424 Value *Result = Builder.CreateCall2(FnExpect, ArgValue, ExpectedValue,
425 "expval");
426 return RValue::get(Result);
427 }
428 case Builtin::BI__builtin_assume_aligned: {
429 Value *PtrValue = EmitScalarExpr(E->getArg(0));
430 Value *OffsetValue =
431 (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : nullptr;
432
433 Value *AlignmentValue = EmitScalarExpr(E->getArg(1));
434 ConstantInt *AlignmentCI = cast<ConstantInt>(AlignmentValue);
435 unsigned Alignment = (unsigned) AlignmentCI->getZExtValue();
436
437 EmitAlignmentAssumption(PtrValue, Alignment, OffsetValue);
438 return RValue::get(PtrValue);
439 }
440 case Builtin::BI__assume:
441 case Builtin::BI__builtin_assume: {
442 if (E->getArg(0)->HasSideEffects(getContext()))
443 return RValue::get(nullptr);
444
445 Value *ArgValue = EmitScalarExpr(E->getArg(0));
446 Value *FnAssume = CGM.getIntrinsic(Intrinsic::assume);
447 return RValue::get(Builder.CreateCall(FnAssume, ArgValue));
448 }
449 case Builtin::BI__builtin_bswap16:
450 case Builtin::BI__builtin_bswap32:
451 case Builtin::BI__builtin_bswap64: {
452 Value *ArgValue = EmitScalarExpr(E->getArg(0));
453 llvm::Type *ArgType = ArgValue->getType();
454 Value *F = CGM.getIntrinsic(Intrinsic::bswap, ArgType);
455 return RValue::get(Builder.CreateCall(F, ArgValue));
456 }
457 case Builtin::BI__builtin_object_size: {
458 // We rely on constant folding to deal with expressions with side effects.
459 assert(!E->getArg(0)->HasSideEffects(getContext()) &&
460 "should have been constant folded");
461
462 // We pass this builtin onto the optimizer so that it can
463 // figure out the object size in more complex cases.
464 llvm::Type *ResType = ConvertType(E->getType());
465
466 // LLVM only supports 0 and 2, make sure that we pass along that
467 // as a boolean.
468 Value *Ty = EmitScalarExpr(E->getArg(1));
469 ConstantInt *CI = dyn_cast<ConstantInt>(Ty);
470 assert(CI);
471 uint64_t val = CI->getZExtValue();
472 CI = ConstantInt::get(Builder.getInt1Ty(), (val & 0x2) >> 1);
473 // FIXME: Get right address space.
474 llvm::Type *Tys[] = { ResType, Builder.getInt8PtrTy(0) };
475 Value *F = CGM.getIntrinsic(Intrinsic::objectsize, Tys);
476 return RValue::get(Builder.CreateCall2(F, EmitScalarExpr(E->getArg(0)),CI));
477 }
478 case Builtin::BI__builtin_prefetch: {
479 Value *Locality, *RW, *Address = EmitScalarExpr(E->getArg(0));
480 // FIXME: Technically these constants should of type 'int', yes?
481 RW = (E->getNumArgs() > 1) ? EmitScalarExpr(E->getArg(1)) :
482 llvm::ConstantInt::get(Int32Ty, 0);
483 Locality = (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) :
484 llvm::ConstantInt::get(Int32Ty, 3);
485 Value *Data = llvm::ConstantInt::get(Int32Ty, 1);
486 Value *F = CGM.getIntrinsic(Intrinsic::prefetch);
487 return RValue::get(Builder.CreateCall4(F, Address, RW, Locality, Data));
488 }
489 case Builtin::BI__builtin_readcyclecounter: {
490 Value *F = CGM.getIntrinsic(Intrinsic::readcyclecounter);
491 return RValue::get(Builder.CreateCall(F));
492 }
493 case Builtin::BI__builtin___clear_cache: {
494 Value *Begin = EmitScalarExpr(E->getArg(0));
495 Value *End = EmitScalarExpr(E->getArg(1));
496 Value *F = CGM.getIntrinsic(Intrinsic::clear_cache);
497 return RValue::get(Builder.CreateCall2(F, Begin, End));
498 }
499 case Builtin::BI__builtin_trap: {
500 Value *F = CGM.getIntrinsic(Intrinsic::trap);
501 return RValue::get(Builder.CreateCall(F));
502 }
503 case Builtin::BI__debugbreak: {
504 Value *F = CGM.getIntrinsic(Intrinsic::debugtrap);
505 return RValue::get(Builder.CreateCall(F));
506 }
507 case Builtin::BI__builtin_unreachable: {
508 if (SanOpts.has(SanitizerKind::Unreachable)) {
509 SanitizerScope SanScope(this);
510 EmitCheck(std::make_pair(static_cast<llvm::Value *>(Builder.getFalse()),
511 SanitizerKind::Unreachable),
512 "builtin_unreachable", EmitCheckSourceLocation(E->getExprLoc()),
513 None);
514 } else
515 Builder.CreateUnreachable();
516
517 // We do need to preserve an insertion point.
518 EmitBlock(createBasicBlock("unreachable.cont"));
519
520 return RValue::get(nullptr);
521 }
522
523 case Builtin::BI__builtin_powi:
524 case Builtin::BI__builtin_powif:
525 case Builtin::BI__builtin_powil: {
526 Value *Base = EmitScalarExpr(E->getArg(0));
527 Value *Exponent = EmitScalarExpr(E->getArg(1));
528 llvm::Type *ArgType = Base->getType();
529 Value *F = CGM.getIntrinsic(Intrinsic::powi, ArgType);
530 return RValue::get(Builder.CreateCall2(F, Base, Exponent));
531 }
532
533 case Builtin::BI__builtin_isgreater:
534 case Builtin::BI__builtin_isgreaterequal:
535 case Builtin::BI__builtin_isless:
536 case Builtin::BI__builtin_islessequal:
537 case Builtin::BI__builtin_islessgreater:
538 case Builtin::BI__builtin_isunordered: {
539 // Ordered comparisons: we know the arguments to these are matching scalar
540 // floating point values.
541 Value *LHS = EmitScalarExpr(E->getArg(0));
542 Value *RHS = EmitScalarExpr(E->getArg(1));
543
544 switch (BuiltinID) {
545 default: llvm_unreachable("Unknown ordered comparison");
546 case Builtin::BI__builtin_isgreater:
547 LHS = Builder.CreateFCmpOGT(LHS, RHS, "cmp");
548 break;
549 case Builtin::BI__builtin_isgreaterequal:
550 LHS = Builder.CreateFCmpOGE(LHS, RHS, "cmp");
551 break;
552 case Builtin::BI__builtin_isless:
553 LHS = Builder.CreateFCmpOLT(LHS, RHS, "cmp");
554 break;
555 case Builtin::BI__builtin_islessequal:
556 LHS = Builder.CreateFCmpOLE(LHS, RHS, "cmp");
557 break;
558 case Builtin::BI__builtin_islessgreater:
559 LHS = Builder.CreateFCmpONE(LHS, RHS, "cmp");
560 break;
561 case Builtin::BI__builtin_isunordered:
562 LHS = Builder.CreateFCmpUNO(LHS, RHS, "cmp");
563 break;
564 }
565 // ZExt bool to int type.
566 return RValue::get(Builder.CreateZExt(LHS, ConvertType(E->getType())));
567 }
568 case Builtin::BI__builtin_isnan: {
569 Value *V = EmitScalarExpr(E->getArg(0));
570 V = Builder.CreateFCmpUNO(V, V, "cmp");
571 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType())));
572 }
573
574 case Builtin::BI__builtin_isinf: {
575 // isinf(x) --> fabs(x) == infinity
576 Value *V = EmitScalarExpr(E->getArg(0));
577 V = EmitFAbs(*this, V);
578
579 V = Builder.CreateFCmpOEQ(V, ConstantFP::getInfinity(V->getType()),"isinf");
580 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType())));
581 }
582
583 case Builtin::BI__builtin_isinf_sign: {
584 // isinf_sign(x) -> fabs(x) == infinity ? (signbit(x) ? -1 : 1) : 0
585 Value *Arg = EmitScalarExpr(E->getArg(0));
586 Value *AbsArg = EmitFAbs(*this, Arg);
587 Value *IsInf = Builder.CreateFCmpOEQ(
588 AbsArg, ConstantFP::getInfinity(Arg->getType()), "isinf");
589 Value *IsNeg = EmitSignBit(*this, Arg);
590
591 llvm::Type *IntTy = ConvertType(E->getType());
592 Value *Zero = Constant::getNullValue(IntTy);
593 Value *One = ConstantInt::get(IntTy, 1);
594 Value *NegativeOne = ConstantInt::get(IntTy, -1);
595 Value *SignResult = Builder.CreateSelect(IsNeg, NegativeOne, One);
596 Value *Result = Builder.CreateSelect(IsInf, SignResult, Zero);
597 return RValue::get(Result);
598 }
599
600 case Builtin::BI__builtin_isnormal: {
601 // isnormal(x) --> x == x && fabsf(x) < infinity && fabsf(x) >= float_min
602 Value *V = EmitScalarExpr(E->getArg(0));
603 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq");
604
605 Value *Abs = EmitFAbs(*this, V);
606 Value *IsLessThanInf =
607 Builder.CreateFCmpULT(Abs, ConstantFP::getInfinity(V->getType()),"isinf");
608 APFloat Smallest = APFloat::getSmallestNormalized(
609 getContext().getFloatTypeSemantics(E->getArg(0)->getType()));
610 Value *IsNormal =
611 Builder.CreateFCmpUGE(Abs, ConstantFP::get(V->getContext(), Smallest),
612 "isnormal");
613 V = Builder.CreateAnd(Eq, IsLessThanInf, "and");
614 V = Builder.CreateAnd(V, IsNormal, "and");
615 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType())));
616 }
617
618 case Builtin::BI__builtin_isfinite: {
619 // isfinite(x) --> x == x && fabs(x) != infinity;
620 Value *V = EmitScalarExpr(E->getArg(0));
621 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq");
622
623 Value *Abs = EmitFAbs(*this, V);
624 Value *IsNotInf =
625 Builder.CreateFCmpUNE(Abs, ConstantFP::getInfinity(V->getType()),"isinf");
626
627 V = Builder.CreateAnd(Eq, IsNotInf, "and");
628 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType())));
629 }
630
631 case Builtin::BI__builtin_fpclassify: {
632 Value *V = EmitScalarExpr(E->getArg(5));
633 llvm::Type *Ty = ConvertType(E->getArg(5)->getType());
634
635 // Create Result
636 BasicBlock *Begin = Builder.GetInsertBlock();
637 BasicBlock *End = createBasicBlock("fpclassify_end", this->CurFn);
638 Builder.SetInsertPoint(End);
639 PHINode *Result =
640 Builder.CreatePHI(ConvertType(E->getArg(0)->getType()), 4,
641 "fpclassify_result");
642
643 // if (V==0) return FP_ZERO
644 Builder.SetInsertPoint(Begin);
645 Value *IsZero = Builder.CreateFCmpOEQ(V, Constant::getNullValue(Ty),
646 "iszero");
647 Value *ZeroLiteral = EmitScalarExpr(E->getArg(4));
648 BasicBlock *NotZero = createBasicBlock("fpclassify_not_zero", this->CurFn);
649 Builder.CreateCondBr(IsZero, End, NotZero);
650 Result->addIncoming(ZeroLiteral, Begin);
651
652 // if (V != V) return FP_NAN
653 Builder.SetInsertPoint(NotZero);
654 Value *IsNan = Builder.CreateFCmpUNO(V, V, "cmp");
655 Value *NanLiteral = EmitScalarExpr(E->getArg(0));
656 BasicBlock *NotNan = createBasicBlock("fpclassify_not_nan", this->CurFn);
657 Builder.CreateCondBr(IsNan, End, NotNan);
658 Result->addIncoming(NanLiteral, NotZero);
659
660 // if (fabs(V) == infinity) return FP_INFINITY
661 Builder.SetInsertPoint(NotNan);
662 Value *VAbs = EmitFAbs(*this, V);
663 Value *IsInf =
664 Builder.CreateFCmpOEQ(VAbs, ConstantFP::getInfinity(V->getType()),
665 "isinf");
666 Value *InfLiteral = EmitScalarExpr(E->getArg(1));
667 BasicBlock *NotInf = createBasicBlock("fpclassify_not_inf", this->CurFn);
668 Builder.CreateCondBr(IsInf, End, NotInf);
669 Result->addIncoming(InfLiteral, NotNan);
670
671 // if (fabs(V) >= MIN_NORMAL) return FP_NORMAL else FP_SUBNORMAL
672 Builder.SetInsertPoint(NotInf);
673 APFloat Smallest = APFloat::getSmallestNormalized(
674 getContext().getFloatTypeSemantics(E->getArg(5)->getType()));
675 Value *IsNormal =
676 Builder.CreateFCmpUGE(VAbs, ConstantFP::get(V->getContext(), Smallest),
677 "isnormal");
678 Value *NormalResult =
679 Builder.CreateSelect(IsNormal, EmitScalarExpr(E->getArg(2)),
680 EmitScalarExpr(E->getArg(3)));
681 Builder.CreateBr(End);
682 Result->addIncoming(NormalResult, NotInf);
683
684 // return Result
685 Builder.SetInsertPoint(End);
686 return RValue::get(Result);
687 }
688
689 case Builtin::BIalloca:
690 case Builtin::BI_alloca:
691 case Builtin::BI__builtin_alloca: {
692 Value *Size = EmitScalarExpr(E->getArg(0));
693 return RValue::get(Builder.CreateAlloca(Builder.getInt8Ty(), Size));
694 }
695 case Builtin::BIbzero:
696 case Builtin::BI__builtin_bzero: {
697 std::pair<llvm::Value*, unsigned> Dest =
698 EmitPointerWithAlignment(E->getArg(0));
699 Value *SizeVal = EmitScalarExpr(E->getArg(1));
700 Builder.CreateMemSet(Dest.first, Builder.getInt8(0), SizeVal,
701 Dest.second, false);
702 return RValue::get(Dest.first);
703 }
704 case Builtin::BImemcpy:
705 case Builtin::BI__builtin_memcpy: {
706 std::pair<llvm::Value*, unsigned> Dest =
707 EmitPointerWithAlignment(E->getArg(0));
708 std::pair<llvm::Value*, unsigned> Src =
709 EmitPointerWithAlignment(E->getArg(1));
710 Value *SizeVal = EmitScalarExpr(E->getArg(2));
711 unsigned Align = std::min(Dest.second, Src.second);
712 Builder.CreateMemCpy(Dest.first, Src.first, SizeVal, Align, false);
713 return RValue::get(Dest.first);
714 }
715
716 case Builtin::BI__builtin___memcpy_chk: {
717 // fold __builtin_memcpy_chk(x, y, cst1, cst2) to memcpy iff cst1<=cst2.
718 llvm::APSInt Size, DstSize;
719 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) ||
720 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext()))
721 break;
722 if (Size.ugt(DstSize))
723 break;
724 std::pair<llvm::Value*, unsigned> Dest =
725 EmitPointerWithAlignment(E->getArg(0));
726 std::pair<llvm::Value*, unsigned> Src =
727 EmitPointerWithAlignment(E->getArg(1));
728 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size);
729 unsigned Align = std::min(Dest.second, Src.second);
730 Builder.CreateMemCpy(Dest.first, Src.first, SizeVal, Align, false);
731 return RValue::get(Dest.first);
732 }
733
734 case Builtin::BI__builtin_objc_memmove_collectable: {
735 Value *Address = EmitScalarExpr(E->getArg(0));
736 Value *SrcAddr = EmitScalarExpr(E->getArg(1));
737 Value *SizeVal = EmitScalarExpr(E->getArg(2));
738 CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this,
739 Address, SrcAddr, SizeVal);
740 return RValue::get(Address);
741 }
742
743 case Builtin::BI__builtin___memmove_chk: {
744 // fold __builtin_memmove_chk(x, y, cst1, cst2) to memmove iff cst1<=cst2.
745 llvm::APSInt Size, DstSize;
746 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) ||
747 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext()))
748 break;
749 if (Size.ugt(DstSize))
750 break;
751 std::pair<llvm::Value*, unsigned> Dest =
752 EmitPointerWithAlignment(E->getArg(0));
753 std::pair<llvm::Value*, unsigned> Src =
754 EmitPointerWithAlignment(E->getArg(1));
755 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size);
756 unsigned Align = std::min(Dest.second, Src.second);
757 Builder.CreateMemMove(Dest.first, Src.first, SizeVal, Align, false);
758 return RValue::get(Dest.first);
759 }
760
761 case Builtin::BImemmove:
762 case Builtin::BI__builtin_memmove: {
763 std::pair<llvm::Value*, unsigned> Dest =
764 EmitPointerWithAlignment(E->getArg(0));
765 std::pair<llvm::Value*, unsigned> Src =
766 EmitPointerWithAlignment(E->getArg(1));
767 Value *SizeVal = EmitScalarExpr(E->getArg(2));
768 unsigned Align = std::min(Dest.second, Src.second);
769 Builder.CreateMemMove(Dest.first, Src.first, SizeVal, Align, false);
770 return RValue::get(Dest.first);
771 }
772 case Builtin::BImemset:
773 case Builtin::BI__builtin_memset: {
774 std::pair<llvm::Value*, unsigned> Dest =
775 EmitPointerWithAlignment(E->getArg(0));
776 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)),
777 Builder.getInt8Ty());
778 Value *SizeVal = EmitScalarExpr(E->getArg(2));
779 Builder.CreateMemSet(Dest.first, ByteVal, SizeVal, Dest.second, false);
780 return RValue::get(Dest.first);
781 }
782 case Builtin::BI__builtin___memset_chk: {
783 // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2.
784 llvm::APSInt Size, DstSize;
785 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) ||
786 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext()))
787 break;
788 if (Size.ugt(DstSize))
789 break;
790 std::pair<llvm::Value*, unsigned> Dest =
791 EmitPointerWithAlignment(E->getArg(0));
792 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)),
793 Builder.getInt8Ty());
794 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size);
795 Builder.CreateMemSet(Dest.first, ByteVal, SizeVal, Dest.second, false);
796 return RValue::get(Dest.first);
797 }
798 case Builtin::BI__builtin_dwarf_cfa: {
799 // The offset in bytes from the first argument to the CFA.
800 //
801 // Why on earth is this in the frontend? Is there any reason at
802 // all that the backend can't reasonably determine this while
803 // lowering llvm.eh.dwarf.cfa()?
804 //
805 // TODO: If there's a satisfactory reason, add a target hook for
806 // this instead of hard-coding 0, which is correct for most targets.
807 int32_t Offset = 0;
808
809 Value *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa);
810 return RValue::get(Builder.CreateCall(F,
811 llvm::ConstantInt::get(Int32Ty, Offset)));
812 }
813 case Builtin::BI__builtin_return_address: {
814 Value *Depth = EmitScalarExpr(E->getArg(0));
815 Depth = Builder.CreateIntCast(Depth, Int32Ty, false);
816 Value *F = CGM.getIntrinsic(Intrinsic::returnaddress);
817 return RValue::get(Builder.CreateCall(F, Depth));
818 }
819 case Builtin::BI__builtin_frame_address: {
820 Value *Depth = EmitScalarExpr(E->getArg(0));
821 Depth = Builder.CreateIntCast(Depth, Int32Ty, false);
822 Value *F = CGM.getIntrinsic(Intrinsic::frameaddress);
823 return RValue::get(Builder.CreateCall(F, Depth));
824 }
825 case Builtin::BI__builtin_extract_return_addr: {
826 Value *Address = EmitScalarExpr(E->getArg(0));
827 Value *Result = getTargetHooks().decodeReturnAddress(*this, Address);
828 return RValue::get(Result);
829 }
830 case Builtin::BI__builtin_frob_return_addr: {
831 Value *Address = EmitScalarExpr(E->getArg(0));
832 Value *Result = getTargetHooks().encodeReturnAddress(*this, Address);
833 return RValue::get(Result);
834 }
835 case Builtin::BI__builtin_dwarf_sp_column: {
836 llvm::IntegerType *Ty
837 = cast<llvm::IntegerType>(ConvertType(E->getType()));
838 int Column = getTargetHooks().getDwarfEHStackPointer(CGM);
839 if (Column == -1) {
840 CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column");
841 return RValue::get(llvm::UndefValue::get(Ty));
842 }
843 return RValue::get(llvm::ConstantInt::get(Ty, Column, true));
844 }
845 case Builtin::BI__builtin_init_dwarf_reg_size_table: {
846 Value *Address = EmitScalarExpr(E->getArg(0));
847 if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address))
848 CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table");
849 return RValue::get(llvm::UndefValue::get(ConvertType(E->getType())));
850 }
851 case Builtin::BI__builtin_eh_return: {
852 Value *Int = EmitScalarExpr(E->getArg(0));
853 Value *Ptr = EmitScalarExpr(E->getArg(1));
854
855 llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType());
856 assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) &&
857 "LLVM's __builtin_eh_return only supports 32- and 64-bit variants");
858 Value *F = CGM.getIntrinsic(IntTy->getBitWidth() == 32
859 ? Intrinsic::eh_return_i32
860 : Intrinsic::eh_return_i64);
861 Builder.CreateCall2(F, Int, Ptr);
862 Builder.CreateUnreachable();
863
864 // We do need to preserve an insertion point.
865 EmitBlock(createBasicBlock("builtin_eh_return.cont"));
866
867 return RValue::get(nullptr);
868 }
869 case Builtin::BI__builtin_unwind_init: {
870 Value *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init);
871 return RValue::get(Builder.CreateCall(F));
872 }
873 case Builtin::BI__builtin_extend_pointer: {
874 // Extends a pointer to the size of an _Unwind_Word, which is
875 // uint64_t on all platforms. Generally this gets poked into a
876 // register and eventually used as an address, so if the
877 // addressing registers are wider than pointers and the platform
878 // doesn't implicitly ignore high-order bits when doing
879 // addressing, we need to make sure we zext / sext based on
880 // the platform's expectations.
881 //
882 // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html
883
884 // Cast the pointer to intptr_t.
885 Value *Ptr = EmitScalarExpr(E->getArg(0));
886 Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast");
887
888 // If that's 64 bits, we're done.
889 if (IntPtrTy->getBitWidth() == 64)
890 return RValue::get(Result);
891
892 // Otherwise, ask the codegen data what to do.
893 if (getTargetHooks().extendPointerWithSExt())
894 return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext"));
895 else
896 return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext"));
897 }
898 case Builtin::BI__builtin_setjmp: {
899 // Buffer is a void**.
900 Value *Buf = EmitScalarExpr(E->getArg(0));
901
902 // Store the frame pointer to the setjmp buffer.
903 Value *FrameAddr =
904 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress),
905 ConstantInt::get(Int32Ty, 0));
906 Builder.CreateStore(FrameAddr, Buf);
907
908 // Store the stack pointer to the setjmp buffer.
909 Value *StackAddr =
910 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave));
911 Value *StackSaveSlot =
912 Builder.CreateGEP(Buf, ConstantInt::get(Int32Ty, 2));
913 Builder.CreateStore(StackAddr, StackSaveSlot);
914
915 // Call LLVM's EH setjmp, which is lightweight.
916 Value *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp);
917 Buf = Builder.CreateBitCast(Buf, Int8PtrTy);
918 return RValue::get(Builder.CreateCall(F, Buf));
919 }
920 case Builtin::BI__builtin_longjmp: {
921 Value *Buf = EmitScalarExpr(E->getArg(0));
922 Buf = Builder.CreateBitCast(Buf, Int8PtrTy);
923
924 // Call LLVM's EH longjmp, which is lightweight.
925 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf);
926
927 // longjmp doesn't return; mark this as unreachable.
928 Builder.CreateUnreachable();
929
930 // We do need to preserve an insertion point.
931 EmitBlock(createBasicBlock("longjmp.cont"));
932
933 return RValue::get(nullptr);
934 }
935 case Builtin::BI__sync_fetch_and_add:
936 case Builtin::BI__sync_fetch_and_sub:
937 case Builtin::BI__sync_fetch_and_or:
938 case Builtin::BI__sync_fetch_and_and:
939 case Builtin::BI__sync_fetch_and_xor:
940 case Builtin::BI__sync_fetch_and_nand:
941 case Builtin::BI__sync_add_and_fetch:
942 case Builtin::BI__sync_sub_and_fetch:
943 case Builtin::BI__sync_and_and_fetch:
944 case Builtin::BI__sync_or_and_fetch:
945 case Builtin::BI__sync_xor_and_fetch:
946 case Builtin::BI__sync_nand_and_fetch:
947 case Builtin::BI__sync_val_compare_and_swap:
948 case Builtin::BI__sync_bool_compare_and_swap:
949 case Builtin::BI__sync_lock_test_and_set:
950 case Builtin::BI__sync_lock_release:
951 case Builtin::BI__sync_swap:
952 llvm_unreachable("Shouldn't make it through sema");
953 case Builtin::BI__sync_fetch_and_add_1:
954 case Builtin::BI__sync_fetch_and_add_2:
955 case Builtin::BI__sync_fetch_and_add_4:
956 case Builtin::BI__sync_fetch_and_add_8:
957 case Builtin::BI__sync_fetch_and_add_16:
958 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E);
959 case Builtin::BI__sync_fetch_and_sub_1:
960 case Builtin::BI__sync_fetch_and_sub_2:
961 case Builtin::BI__sync_fetch_and_sub_4:
962 case Builtin::BI__sync_fetch_and_sub_8:
963 case Builtin::BI__sync_fetch_and_sub_16:
964 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E);
965 case Builtin::BI__sync_fetch_and_or_1:
966 case Builtin::BI__sync_fetch_and_or_2:
967 case Builtin::BI__sync_fetch_and_or_4:
968 case Builtin::BI__sync_fetch_and_or_8:
969 case Builtin::BI__sync_fetch_and_or_16:
970 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E);
971 case Builtin::BI__sync_fetch_and_and_1:
972 case Builtin::BI__sync_fetch_and_and_2:
973 case Builtin::BI__sync_fetch_and_and_4:
974 case Builtin::BI__sync_fetch_and_and_8:
975 case Builtin::BI__sync_fetch_and_and_16:
976 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E);
977 case Builtin::BI__sync_fetch_and_xor_1:
978 case Builtin::BI__sync_fetch_and_xor_2:
979 case Builtin::BI__sync_fetch_and_xor_4:
980 case Builtin::BI__sync_fetch_and_xor_8:
981 case Builtin::BI__sync_fetch_and_xor_16:
982 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E);
983 case Builtin::BI__sync_fetch_and_nand_1:
984 case Builtin::BI__sync_fetch_and_nand_2:
985 case Builtin::BI__sync_fetch_and_nand_4:
986 case Builtin::BI__sync_fetch_and_nand_8:
987 case Builtin::BI__sync_fetch_and_nand_16:
988 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Nand, E);
989
990 // Clang extensions: not overloaded yet.
991 case Builtin::BI__sync_fetch_and_min:
992 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E);
993 case Builtin::BI__sync_fetch_and_max:
994 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E);
995 case Builtin::BI__sync_fetch_and_umin:
996 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E);
997 case Builtin::BI__sync_fetch_and_umax:
998 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E);
999
1000 case Builtin::BI__sync_add_and_fetch_1:
1001 case Builtin::BI__sync_add_and_fetch_2:
1002 case Builtin::BI__sync_add_and_fetch_4:
1003 case Builtin::BI__sync_add_and_fetch_8:
1004 case Builtin::BI__sync_add_and_fetch_16:
1005 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E,
1006 llvm::Instruction::Add);
1007 case Builtin::BI__sync_sub_and_fetch_1:
1008 case Builtin::BI__sync_sub_and_fetch_2:
1009 case Builtin::BI__sync_sub_and_fetch_4:
1010 case Builtin::BI__sync_sub_and_fetch_8:
1011 case Builtin::BI__sync_sub_and_fetch_16:
1012 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E,
1013 llvm::Instruction::Sub);
1014 case Builtin::BI__sync_and_and_fetch_1:
1015 case Builtin::BI__sync_and_and_fetch_2:
1016 case Builtin::BI__sync_and_and_fetch_4:
1017 case Builtin::BI__sync_and_and_fetch_8:
1018 case Builtin::BI__sync_and_and_fetch_16:
1019 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::And, E,
1020 llvm::Instruction::And);
1021 case Builtin::BI__sync_or_and_fetch_1:
1022 case Builtin::BI__sync_or_and_fetch_2:
1023 case Builtin::BI__sync_or_and_fetch_4:
1024 case Builtin::BI__sync_or_and_fetch_8:
1025 case Builtin::BI__sync_or_and_fetch_16:
1026 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E,
1027 llvm::Instruction::Or);
1028 case Builtin::BI__sync_xor_and_fetch_1:
1029 case Builtin::BI__sync_xor_and_fetch_2:
1030 case Builtin::BI__sync_xor_and_fetch_4:
1031 case Builtin::BI__sync_xor_and_fetch_8:
1032 case Builtin::BI__sync_xor_and_fetch_16:
1033 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E,
1034 llvm::Instruction::Xor);
1035 case Builtin::BI__sync_nand_and_fetch_1:
1036 case Builtin::BI__sync_nand_and_fetch_2:
1037 case Builtin::BI__sync_nand_and_fetch_4:
1038 case Builtin::BI__sync_nand_and_fetch_8:
1039 case Builtin::BI__sync_nand_and_fetch_16:
1040 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Nand, E,
1041 llvm::Instruction::And, true);
1042
1043 case Builtin::BI__sync_val_compare_and_swap_1:
1044 case Builtin::BI__sync_val_compare_and_swap_2:
1045 case Builtin::BI__sync_val_compare_and_swap_4:
1046 case Builtin::BI__sync_val_compare_and_swap_8:
1047 case Builtin::BI__sync_val_compare_and_swap_16: {
1048 QualType T = E->getType();
1049 llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0));
1050 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace();
1051
1052 llvm::IntegerType *IntType =
1053 llvm::IntegerType::get(getLLVMContext(),
1054 getContext().getTypeSize(T));
1055 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace);
1056
1057 Value *Args[3];
1058 Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType);
1059 Args[1] = EmitScalarExpr(E->getArg(1));
1060 llvm::Type *ValueType = Args[1]->getType();
1061 Args[1] = EmitToInt(*this, Args[1], T, IntType);
1062 Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType);
1063
1064 Value *Result = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2],
1065 llvm::SequentiallyConsistent,
1066 llvm::SequentiallyConsistent);
1067 Result = Builder.CreateExtractValue(Result, 0);
1068 Result = EmitFromInt(*this, Result, T, ValueType);
1069 return RValue::get(Result);
1070 }
1071
1072 case Builtin::BI__sync_bool_compare_and_swap_1:
1073 case Builtin::BI__sync_bool_compare_and_swap_2:
1074 case Builtin::BI__sync_bool_compare_and_swap_4:
1075 case Builtin::BI__sync_bool_compare_and_swap_8:
1076 case Builtin::BI__sync_bool_compare_and_swap_16: {
1077 QualType T = E->getArg(1)->getType();
1078 llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0));
1079 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace();
1080
1081 llvm::IntegerType *IntType =
1082 llvm::IntegerType::get(getLLVMContext(),
1083 getContext().getTypeSize(T));
1084 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace);
1085
1086 Value *Args[3];
1087 Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType);
1088 Args[1] = EmitToInt(*this, EmitScalarExpr(E->getArg(1)), T, IntType);
1089 Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType);
1090
1091 Value *Pair = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2],
1092 llvm::SequentiallyConsistent,
1093 llvm::SequentiallyConsistent);
1094 Value *Result = Builder.CreateExtractValue(Pair, 1);
1095 // zext bool to int.
1096 Result = Builder.CreateZExt(Result, ConvertType(E->getType()));
1097 return RValue::get(Result);
1098 }
1099
1100 case Builtin::BI__sync_swap_1:
1101 case Builtin::BI__sync_swap_2:
1102 case Builtin::BI__sync_swap_4:
1103 case Builtin::BI__sync_swap_8:
1104 case Builtin::BI__sync_swap_16:
1105 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E);
1106
1107 case Builtin::BI__sync_lock_test_and_set_1:
1108 case Builtin::BI__sync_lock_test_and_set_2:
1109 case Builtin::BI__sync_lock_test_and_set_4:
1110 case Builtin::BI__sync_lock_test_and_set_8:
1111 case Builtin::BI__sync_lock_test_and_set_16:
1112 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E);
1113
1114 case Builtin::BI__sync_lock_release_1:
1115 case Builtin::BI__sync_lock_release_2:
1116 case Builtin::BI__sync_lock_release_4:
1117 case Builtin::BI__sync_lock_release_8:
1118 case Builtin::BI__sync_lock_release_16: {
1119 Value *Ptr = EmitScalarExpr(E->getArg(0));
1120 QualType ElTy = E->getArg(0)->getType()->getPointeeType();
1121 CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy);
1122 llvm::Type *ITy = llvm::IntegerType::get(getLLVMContext(),
1123 StoreSize.getQuantity() * 8);
1124 Ptr = Builder.CreateBitCast(Ptr, ITy->getPointerTo());
1125 llvm::StoreInst *Store =
1126 Builder.CreateStore(llvm::Constant::getNullValue(ITy), Ptr);
1127 Store->setAlignment(StoreSize.getQuantity());
1128 Store->setAtomic(llvm::Release);
1129 return RValue::get(nullptr);
1130 }
1131
1132 case Builtin::BI__sync_synchronize: {
1133 // We assume this is supposed to correspond to a C++0x-style
1134 // sequentially-consistent fence (i.e. this is only usable for
1135 // synchonization, not device I/O or anything like that). This intrinsic
1136 // is really badly designed in the sense that in theory, there isn't
1137 // any way to safely use it... but in practice, it mostly works
1138 // to use it with non-atomic loads and stores to get acquire/release
1139 // semantics.
1140 Builder.CreateFence(llvm::SequentiallyConsistent);
1141 return RValue::get(nullptr);
1142 }
1143
1144 case Builtin::BI__c11_atomic_is_lock_free:
1145 case Builtin::BI__atomic_is_lock_free: {
1146 // Call "bool __atomic_is_lock_free(size_t size, void *ptr)". For the
1147 // __c11 builtin, ptr is 0 (indicating a properly-aligned object), since
1148 // _Atomic(T) is always properly-aligned.
1149 const char *LibCallName = "__atomic_is_lock_free";
1150 CallArgList Args;
1151 Args.add(RValue::get(EmitScalarExpr(E->getArg(0))),
1152 getContext().getSizeType());
1153 if (BuiltinID == Builtin::BI__atomic_is_lock_free)
1154 Args.add(RValue::get(EmitScalarExpr(E->getArg(1))),
1155 getContext().VoidPtrTy);
1156 else
1157 Args.add(RValue::get(llvm::Constant::getNullValue(VoidPtrTy)),
1158 getContext().VoidPtrTy);
1159 const CGFunctionInfo &FuncInfo =
1160 CGM.getTypes().arrangeFreeFunctionCall(E->getType(), Args,
1161 FunctionType::ExtInfo(),
1162 RequiredArgs::All);
1163 llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FuncInfo);
1164 llvm::Constant *Func = CGM.CreateRuntimeFunction(FTy, LibCallName);
1165 return EmitCall(FuncInfo, Func, ReturnValueSlot(), Args);
1166 }
1167
1168 case Builtin::BI__atomic_test_and_set: {
1169 // Look at the argument type to determine whether this is a volatile
1170 // operation. The parameter type is always volatile.
1171 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType();
1172 bool Volatile =
1173 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified();
1174
1175 Value *Ptr = EmitScalarExpr(E->getArg(0));
1176 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace();
1177 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace));
1178 Value *NewVal = Builder.getInt8(1);
1179 Value *Order = EmitScalarExpr(E->getArg(1));
1180 if (isa<llvm::ConstantInt>(Order)) {
1181 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue();
1182 AtomicRMWInst *Result = nullptr;
1183 switch (ord) {
1184 case 0: // memory_order_relaxed
1185 default: // invalid order
1186 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg,
1187 Ptr, NewVal,
1188 llvm::Monotonic);
1189 break;
1190 case 1: // memory_order_consume
1191 case 2: // memory_order_acquire
1192 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg,
1193 Ptr, NewVal,
1194 llvm::Acquire);
1195 break;
1196 case 3: // memory_order_release
1197 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg,
1198 Ptr, NewVal,
1199 llvm::Release);
1200 break;
1201 case 4: // memory_order_acq_rel
1202 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg,
1203 Ptr, NewVal,
1204 llvm::AcquireRelease);
1205 break;
1206 case 5: // memory_order_seq_cst
1207 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg,
1208 Ptr, NewVal,
1209 llvm::SequentiallyConsistent);
1210 break;
1211 }
1212 Result->setVolatile(Volatile);
1213 return RValue::get(Builder.CreateIsNotNull(Result, "tobool"));
1214 }
1215
1216 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn);
1217
1218 llvm::BasicBlock *BBs[5] = {
1219 createBasicBlock("monotonic", CurFn),
1220 createBasicBlock("acquire", CurFn),
1221 createBasicBlock("release", CurFn),
1222 createBasicBlock("acqrel", CurFn),
1223 createBasicBlock("seqcst", CurFn)
1224 };
1225 llvm::AtomicOrdering Orders[5] = {
1226 llvm::Monotonic, llvm::Acquire, llvm::Release,
1227 llvm::AcquireRelease, llvm::SequentiallyConsistent
1228 };
1229
1230 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false);
1231 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]);
1232
1233 Builder.SetInsertPoint(ContBB);
1234 PHINode *Result = Builder.CreatePHI(Int8Ty, 5, "was_set");
1235
1236 for (unsigned i = 0; i < 5; ++i) {
1237 Builder.SetInsertPoint(BBs[i]);
1238 AtomicRMWInst *RMW = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg,
1239 Ptr, NewVal, Orders[i]);
1240 RMW->setVolatile(Volatile);
1241 Result->addIncoming(RMW, BBs[i]);
1242 Builder.CreateBr(ContBB);
1243 }
1244
1245 SI->addCase(Builder.getInt32(0), BBs[0]);
1246 SI->addCase(Builder.getInt32(1), BBs[1]);
1247 SI->addCase(Builder.getInt32(2), BBs[1]);
1248 SI->addCase(Builder.getInt32(3), BBs[2]);
1249 SI->addCase(Builder.getInt32(4), BBs[3]);
1250 SI->addCase(Builder.getInt32(5), BBs[4]);
1251
1252 Builder.SetInsertPoint(ContBB);
1253 return RValue::get(Builder.CreateIsNotNull(Result, "tobool"));
1254 }
1255
1256 case Builtin::BI__atomic_clear: {
1257 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType();
1258 bool Volatile =
1259 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified();
1260
1261 Value *Ptr = EmitScalarExpr(E->getArg(0));
1262 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace();
1263 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace));
1264 Value *NewVal = Builder.getInt8(0);
1265 Value *Order = EmitScalarExpr(E->getArg(1));
1266 if (isa<llvm::ConstantInt>(Order)) {
1267 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue();
1268 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile);
1269 Store->setAlignment(1);
1270 switch (ord) {
1271 case 0: // memory_order_relaxed
1272 default: // invalid order
1273 Store->setOrdering(llvm::Monotonic);
1274 break;
1275 case 3: // memory_order_release
1276 Store->setOrdering(llvm::Release);
1277 break;
1278 case 5: // memory_order_seq_cst
1279 Store->setOrdering(llvm::SequentiallyConsistent);
1280 break;
1281 }
1282 return RValue::get(nullptr);
1283 }
1284
1285 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn);
1286
1287 llvm::BasicBlock *BBs[3] = {
1288 createBasicBlock("monotonic", CurFn),
1289 createBasicBlock("release", CurFn),
1290 createBasicBlock("seqcst", CurFn)
1291 };
1292 llvm::AtomicOrdering Orders[3] = {
1293 llvm::Monotonic, llvm::Release, llvm::SequentiallyConsistent
1294 };
1295
1296 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false);
1297 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]);
1298
1299 for (unsigned i = 0; i < 3; ++i) {
1300 Builder.SetInsertPoint(BBs[i]);
1301 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile);
1302 Store->setAlignment(1);
1303 Store->setOrdering(Orders[i]);
1304 Builder.CreateBr(ContBB);
1305 }
1306
1307 SI->addCase(Builder.getInt32(0), BBs[0]);
1308 SI->addCase(Builder.getInt32(3), BBs[1]);
1309 SI->addCase(Builder.getInt32(5), BBs[2]);
1310
1311 Builder.SetInsertPoint(ContBB);
1312 return RValue::get(nullptr);
1313 }
1314
1315 case Builtin::BI__atomic_thread_fence:
1316 case Builtin::BI__atomic_signal_fence:
1317 case Builtin::BI__c11_atomic_thread_fence:
1318 case Builtin::BI__c11_atomic_signal_fence: {
1319 llvm::SynchronizationScope Scope;
1320 if (BuiltinID == Builtin::BI__atomic_signal_fence ||
1321 BuiltinID == Builtin::BI__c11_atomic_signal_fence)
1322 Scope = llvm::SingleThread;
1323 else
1324 Scope = llvm::CrossThread;
1325 Value *Order = EmitScalarExpr(E->getArg(0));
1326 if (isa<llvm::ConstantInt>(Order)) {
1327 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue();
1328 switch (ord) {
1329 case 0: // memory_order_relaxed
1330 default: // invalid order
1331 break;
1332 case 1: // memory_order_consume
1333 case 2: // memory_order_acquire
1334 Builder.CreateFence(llvm::Acquire, Scope);
1335 break;
1336 case 3: // memory_order_release
1337 Builder.CreateFence(llvm::Release, Scope);
1338 break;
1339 case 4: // memory_order_acq_rel
1340 Builder.CreateFence(llvm::AcquireRelease, Scope);
1341 break;
1342 case 5: // memory_order_seq_cst
1343 Builder.CreateFence(llvm::SequentiallyConsistent, Scope);
1344 break;
1345 }
1346 return RValue::get(nullptr);
1347 }
1348
1349 llvm::BasicBlock *AcquireBB, *ReleaseBB, *AcqRelBB, *SeqCstBB;
1350 AcquireBB = createBasicBlock("acquire", CurFn);
1351 ReleaseBB = createBasicBlock("release", CurFn);
1352 AcqRelBB = createBasicBlock("acqrel", CurFn);
1353 SeqCstBB = createBasicBlock("seqcst", CurFn);
1354 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn);
1355
1356 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false);
1357 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, ContBB);
1358
1359 Builder.SetInsertPoint(AcquireBB);
1360 Builder.CreateFence(llvm::Acquire, Scope);
1361 Builder.CreateBr(ContBB);
1362 SI->addCase(Builder.getInt32(1), AcquireBB);
1363 SI->addCase(Builder.getInt32(2), AcquireBB);
1364
1365 Builder.SetInsertPoint(ReleaseBB);
1366 Builder.CreateFence(llvm::Release, Scope);
1367 Builder.CreateBr(ContBB);
1368 SI->addCase(Builder.getInt32(3), ReleaseBB);
1369
1370 Builder.SetInsertPoint(AcqRelBB);
1371 Builder.CreateFence(llvm::AcquireRelease, Scope);
1372 Builder.CreateBr(ContBB);
1373 SI->addCase(Builder.getInt32(4), AcqRelBB);
1374
1375 Builder.SetInsertPoint(SeqCstBB);
1376 Builder.CreateFence(llvm::SequentiallyConsistent, Scope);
1377 Builder.CreateBr(ContBB);
1378 SI->addCase(Builder.getInt32(5), SeqCstBB);
1379
1380 Builder.SetInsertPoint(ContBB);
1381 return RValue::get(nullptr);
1382 }
1383
1384 // Library functions with special handling.
1385 case Builtin::BIsqrt:
1386 case Builtin::BIsqrtf:
1387 case Builtin::BIsqrtl: {
1388 // Transform a call to sqrt* into a @llvm.sqrt.* intrinsic call, but only
1389 // in finite- or unsafe-math mode (the intrinsic has different semantics
1390 // for handling negative numbers compared to the library function, so
1391 // -fmath-errno=0 is not enough).
1392 if (!FD->hasAttr<ConstAttr>())
1393 break;
1394 if (!(CGM.getCodeGenOpts().UnsafeFPMath ||
1395 CGM.getCodeGenOpts().NoNaNsFPMath))
1396 break;
1397 Value *Arg0 = EmitScalarExpr(E->getArg(0));
1398 llvm::Type *ArgType = Arg0->getType();
1399 Value *F = CGM.getIntrinsic(Intrinsic::sqrt, ArgType);
1400 return RValue::get(Builder.CreateCall(F, Arg0));
1401 }
1402
1403 case Builtin::BI__builtin_pow:
1404 case Builtin::BI__builtin_powf:
1405 case Builtin::BI__builtin_powl:
1406 case Builtin::BIpow:
1407 case Builtin::BIpowf:
1408 case Builtin::BIpowl: {
1409 // Transform a call to pow* into a @llvm.pow.* intrinsic call.
1410 if (!FD->hasAttr<ConstAttr>())
1411 break;
1412 Value *Base = EmitScalarExpr(E->getArg(0));
1413 Value *Exponent = EmitScalarExpr(E->getArg(1));
1414 llvm::Type *ArgType = Base->getType();
1415 Value *F = CGM.getIntrinsic(Intrinsic::pow, ArgType);
1416 return RValue::get(Builder.CreateCall2(F, Base, Exponent));
1417 }
1418
1419 case Builtin::BIfma:
1420 case Builtin::BIfmaf:
1421 case Builtin::BIfmal:
1422 case Builtin::BI__builtin_fma:
1423 case Builtin::BI__builtin_fmaf:
1424 case Builtin::BI__builtin_fmal: {
1425 // Rewrite fma to intrinsic.
1426 Value *FirstArg = EmitScalarExpr(E->getArg(0));
1427 llvm::Type *ArgType = FirstArg->getType();
1428 Value *F = CGM.getIntrinsic(Intrinsic::fma, ArgType);
1429 return RValue::get(Builder.CreateCall3(F, FirstArg,
1430 EmitScalarExpr(E->getArg(1)),
1431 EmitScalarExpr(E->getArg(2))));
1432 }
1433
1434 case Builtin::BI__builtin_signbit:
1435 case Builtin::BI__builtin_signbitf:
1436 case Builtin::BI__builtin_signbitl: {
1437 return RValue::get(
1438 Builder.CreateZExt(EmitSignBit(*this, EmitScalarExpr(E->getArg(0))),
1439 ConvertType(E->getType())));
1440 }
1441 case Builtin::BI__builtin_annotation: {
1442 llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0));
1443 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::annotation,
1444 AnnVal->getType());
1445
1446 // Get the annotation string, go through casts. Sema requires this to be a
1447 // non-wide string literal, potentially casted, so the cast<> is safe.
1448 const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts();
1449 StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString();
1450 return RValue::get(EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc()));
1451 }
1452 case Builtin::BI__builtin_addcb:
1453 case Builtin::BI__builtin_addcs:
1454 case Builtin::BI__builtin_addc:
1455 case Builtin::BI__builtin_addcl:
1456 case Builtin::BI__builtin_addcll:
1457 case Builtin::BI__builtin_subcb:
1458 case Builtin::BI__builtin_subcs:
1459 case Builtin::BI__builtin_subc:
1460 case Builtin::BI__builtin_subcl:
1461 case Builtin::BI__builtin_subcll: {
1462
1463 // We translate all of these builtins from expressions of the form:
1464 // int x = ..., y = ..., carryin = ..., carryout, result;
1465 // result = __builtin_addc(x, y, carryin, &carryout);
1466 //
1467 // to LLVM IR of the form:
1468 //
1469 // %tmp1 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %x, i32 %y)
1470 // %tmpsum1 = extractvalue {i32, i1} %tmp1, 0
1471 // %carry1 = extractvalue {i32, i1} %tmp1, 1
1472 // %tmp2 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %tmpsum1,
1473 // i32 %carryin)
1474 // %result = extractvalue {i32, i1} %tmp2, 0
1475 // %carry2 = extractvalue {i32, i1} %tmp2, 1
1476 // %tmp3 = or i1 %carry1, %carry2
1477 // %tmp4 = zext i1 %tmp3 to i32
1478 // store i32 %tmp4, i32* %carryout
1479
1480 // Scalarize our inputs.
1481 llvm::Value *X = EmitScalarExpr(E->getArg(0));
1482 llvm::Value *Y = EmitScalarExpr(E->getArg(1));
1483 llvm::Value *Carryin = EmitScalarExpr(E->getArg(2));
1484 std::pair<llvm::Value*, unsigned> CarryOutPtr =
1485 EmitPointerWithAlignment(E->getArg(3));
1486
1487 // Decide if we are lowering to a uadd.with.overflow or usub.with.overflow.
1488 llvm::Intrinsic::ID IntrinsicId;
1489 switch (BuiltinID) {
1490 default: llvm_unreachable("Unknown multiprecision builtin id.");
1491 case Builtin::BI__builtin_addcb:
1492 case Builtin::BI__builtin_addcs:
1493 case Builtin::BI__builtin_addc:
1494 case Builtin::BI__builtin_addcl:
1495 case Builtin::BI__builtin_addcll:
1496 IntrinsicId = llvm::Intrinsic::uadd_with_overflow;
1497 break;
1498 case Builtin::BI__builtin_subcb:
1499 case Builtin::BI__builtin_subcs:
1500 case Builtin::BI__builtin_subc:
1501 case Builtin::BI__builtin_subcl:
1502 case Builtin::BI__builtin_subcll:
1503 IntrinsicId = llvm::Intrinsic::usub_with_overflow;
1504 break;
1505 }
1506
1507 // Construct our resulting LLVM IR expression.
1508 llvm::Value *Carry1;
1509 llvm::Value *Sum1 = EmitOverflowIntrinsic(*this, IntrinsicId,
1510 X, Y, Carry1);
1511 llvm::Value *Carry2;
1512 llvm::Value *Sum2 = EmitOverflowIntrinsic(*this, IntrinsicId,
1513 Sum1, Carryin, Carry2);
1514 llvm::Value *CarryOut = Builder.CreateZExt(Builder.CreateOr(Carry1, Carry2),
1515 X->getType());
1516 llvm::StoreInst *CarryOutStore = Builder.CreateStore(CarryOut,
1517 CarryOutPtr.first);
1518 CarryOutStore->setAlignment(CarryOutPtr.second);
1519 return RValue::get(Sum2);
1520 }
1521 case Builtin::BI__builtin_uadd_overflow:
1522 case Builtin::BI__builtin_uaddl_overflow:
1523 case Builtin::BI__builtin_uaddll_overflow:
1524 case Builtin::BI__builtin_usub_overflow:
1525 case Builtin::BI__builtin_usubl_overflow:
1526 case Builtin::BI__builtin_usubll_overflow:
1527 case Builtin::BI__builtin_umul_overflow:
1528 case Builtin::BI__builtin_umull_overflow:
1529 case Builtin::BI__builtin_umulll_overflow:
1530 case Builtin::BI__builtin_sadd_overflow:
1531 case Builtin::BI__builtin_saddl_overflow:
1532 case Builtin::BI__builtin_saddll_overflow:
1533 case Builtin::BI__builtin_ssub_overflow:
1534 case Builtin::BI__builtin_ssubl_overflow:
1535 case Builtin::BI__builtin_ssubll_overflow:
1536 case Builtin::BI__builtin_smul_overflow:
1537 case Builtin::BI__builtin_smull_overflow:
1538 case Builtin::BI__builtin_smulll_overflow: {
1539
1540 // We translate all of these builtins directly to the relevant llvm IR node.
1541
1542 // Scalarize our inputs.
1543 llvm::Value *X = EmitScalarExpr(E->getArg(0));
1544 llvm::Value *Y = EmitScalarExpr(E->getArg(1));
1545 std::pair<llvm::Value *, unsigned> SumOutPtr =
1546 EmitPointerWithAlignment(E->getArg(2));
1547
1548 // Decide which of the overflow intrinsics we are lowering to:
1549 llvm::Intrinsic::ID IntrinsicId;
1550 switch (BuiltinID) {
1551 default: llvm_unreachable("Unknown security overflow builtin id.");
1552 case Builtin::BI__builtin_uadd_overflow:
1553 case Builtin::BI__builtin_uaddl_overflow:
1554 case Builtin::BI__builtin_uaddll_overflow:
1555 IntrinsicId = llvm::Intrinsic::uadd_with_overflow;
1556 break;
1557 case Builtin::BI__builtin_usub_overflow:
1558 case Builtin::BI__builtin_usubl_overflow:
1559 case Builtin::BI__builtin_usubll_overflow:
1560 IntrinsicId = llvm::Intrinsic::usub_with_overflow;
1561 break;
1562 case Builtin::BI__builtin_umul_overflow:
1563 case Builtin::BI__builtin_umull_overflow:
1564 case Builtin::BI__builtin_umulll_overflow:
1565 IntrinsicId = llvm::Intrinsic::umul_with_overflow;
1566 break;
1567 case Builtin::BI__builtin_sadd_overflow:
1568 case Builtin::BI__builtin_saddl_overflow:
1569 case Builtin::BI__builtin_saddll_overflow:
1570 IntrinsicId = llvm::Intrinsic::sadd_with_overflow;
1571 break;
1572 case Builtin::BI__builtin_ssub_overflow:
1573 case Builtin::BI__builtin_ssubl_overflow:
1574 case Builtin::BI__builtin_ssubll_overflow:
1575 IntrinsicId = llvm::Intrinsic::ssub_with_overflow;
1576 break;
1577 case Builtin::BI__builtin_smul_overflow:
1578 case Builtin::BI__builtin_smull_overflow:
1579 case Builtin::BI__builtin_smulll_overflow:
1580 IntrinsicId = llvm::Intrinsic::smul_with_overflow;
1581 break;
1582 }
1583
1584
1585 llvm::Value *Carry;
1586 llvm::Value *Sum = EmitOverflowIntrinsic(*this, IntrinsicId, X, Y, Carry);
1587 llvm::StoreInst *SumOutStore = Builder.CreateStore(Sum, SumOutPtr.first);
1588 SumOutStore->setAlignment(SumOutPtr.second);
1589
1590 return RValue::get(Carry);
1591 }
1592 case Builtin::BI__builtin_addressof:
1593 return RValue::get(EmitLValue(E->getArg(0)).getAddress());
1594 case Builtin::BI__builtin_operator_new:
1595 return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(),
1596 E->getArg(0), false);
1597 case Builtin::BI__builtin_operator_delete:
1598 return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(),
1599 E->getArg(0), true);
1600 case Builtin::BI__noop:
1601 // __noop always evaluates to an integer literal zero.
1602 return RValue::get(ConstantInt::get(IntTy, 0));
1603 case Builtin::BI__builtin_call_with_static_chain: {
1604 const CallExpr *Call = cast<CallExpr>(E->getArg(0));
1605 const Expr *Chain = E->getArg(1);
1606 return EmitCall(Call->getCallee()->getType(),
1607 EmitScalarExpr(Call->getCallee()), Call, ReturnValue,
1608 Call->getCalleeDecl(), EmitScalarExpr(Chain));
1609 }
1610 case Builtin::BI_InterlockedExchange:
1611 case Builtin::BI_InterlockedExchangePointer:
1612 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E);
1613 case Builtin::BI_InterlockedCompareExchangePointer: {
1614 llvm::Type *RTy;
1615 llvm::IntegerType *IntType =
1616 IntegerType::get(getLLVMContext(),
1617 getContext().getTypeSize(E->getType()));
1618 llvm::Type *IntPtrType = IntType->getPointerTo();
1619
1620 llvm::Value *Destination =
1621 Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), IntPtrType);
1622
1623 llvm::Value *Exchange = EmitScalarExpr(E->getArg(1));
1624 RTy = Exchange->getType();
1625 Exchange = Builder.CreatePtrToInt(Exchange, IntType);
1626
1627 llvm::Value *Comparand =
1628 Builder.CreatePtrToInt(EmitScalarExpr(E->getArg(2)), IntType);
1629
1630 auto Result = Builder.CreateAtomicCmpXchg(Destination, Comparand, Exchange,
1631 SequentiallyConsistent,
1632 SequentiallyConsistent);
1633 Result->setVolatile(true);
1634
1635 return RValue::get(Builder.CreateIntToPtr(Builder.CreateExtractValue(Result,
1636 0),
1637 RTy));
1638 }
1639 case Builtin::BI_InterlockedCompareExchange: {
1640 AtomicCmpXchgInst *CXI = Builder.CreateAtomicCmpXchg(
1641 EmitScalarExpr(E->getArg(0)),
1642 EmitScalarExpr(E->getArg(2)),
1643 EmitScalarExpr(E->getArg(1)),
1644 SequentiallyConsistent,
1645 SequentiallyConsistent);
1646 CXI->setVolatile(true);
1647 return RValue::get(Builder.CreateExtractValue(CXI, 0));
1648 }
1649 case Builtin::BI_InterlockedIncrement: {
1650 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW(
1651 AtomicRMWInst::Add,
1652 EmitScalarExpr(E->getArg(0)),
1653 ConstantInt::get(Int32Ty, 1),
1654 llvm::SequentiallyConsistent);
1655 RMWI->setVolatile(true);
1656 return RValue::get(Builder.CreateAdd(RMWI, ConstantInt::get(Int32Ty, 1)));
1657 }
1658 case Builtin::BI_InterlockedDecrement: {
1659 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW(
1660 AtomicRMWInst::Sub,
1661 EmitScalarExpr(E->getArg(0)),
1662 ConstantInt::get(Int32Ty, 1),
1663 llvm::SequentiallyConsistent);
1664 RMWI->setVolatile(true);
1665 return RValue::get(Builder.CreateSub(RMWI, ConstantInt::get(Int32Ty, 1)));
1666 }
1667 case Builtin::BI_InterlockedExchangeAdd: {
1668 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW(
1669 AtomicRMWInst::Add,
1670 EmitScalarExpr(E->getArg(0)),
1671 EmitScalarExpr(E->getArg(1)),
1672 llvm::SequentiallyConsistent);
1673 RMWI->setVolatile(true);
1674 return RValue::get(RMWI);
1675 }
1676 case Builtin::BI__readfsdword: {
1677 Value *IntToPtr =
1678 Builder.CreateIntToPtr(EmitScalarExpr(E->getArg(0)),
1679 llvm::PointerType::get(CGM.Int32Ty, 257));
1680 LoadInst *Load =
1681 Builder.CreateAlignedLoad(IntToPtr, /*Align=*/4, /*isVolatile=*/true);
1682 return RValue::get(Load);
1683 }
1684
1685 case Builtin::BI__exception_code:
1686 case Builtin::BI_exception_code:
1687 return RValue::get(EmitSEHExceptionCode());
1688 case Builtin::BI__exception_info:
1689 case Builtin::BI_exception_info:
1690 return RValue::get(EmitSEHExceptionInfo());
1691 case Builtin::BI__abnormal_termination:
1692 case Builtin::BI_abnormal_termination:
1693 return RValue::get(EmitSEHAbnormalTermination());
1694 case Builtin::BI_setjmpex: {
1695 if (getTarget().getTriple().isOSMSVCRT()) {
1696 llvm::Type *ArgTypes[] = {Int8PtrTy, Int8PtrTy};
1697 llvm::AttributeSet ReturnsTwiceAttr =
1698 AttributeSet::get(getLLVMContext(), llvm::AttributeSet::FunctionIndex,
1699 llvm::Attribute::ReturnsTwice);
1700 llvm::Constant *SetJmpEx = CGM.CreateRuntimeFunction(
1701 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/false),
1702 "_setjmpex", ReturnsTwiceAttr);
1703 llvm::Value *Buf = Builder.CreateBitOrPointerCast(
1704 EmitScalarExpr(E->getArg(0)), Int8PtrTy);
1705 llvm::Value *FrameAddr =
1706 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress),
1707 ConstantInt::get(Int32Ty, 0));
1708 llvm::Value *Args[] = {Buf, FrameAddr};
1709 llvm::CallSite CS = EmitRuntimeCallOrInvoke(SetJmpEx, Args);
1710 CS.setAttributes(ReturnsTwiceAttr);
1711 return RValue::get(CS.getInstruction());
1712 }
1713 break;
1714 }
1715 case Builtin::BI_setjmp: {
1716 if (getTarget().getTriple().isOSMSVCRT()) {
1717 llvm::AttributeSet ReturnsTwiceAttr =
1718 AttributeSet::get(getLLVMContext(), llvm::AttributeSet::FunctionIndex,
1719 llvm::Attribute::ReturnsTwice);
1720 llvm::Value *Buf = Builder.CreateBitOrPointerCast(
1721 EmitScalarExpr(E->getArg(0)), Int8PtrTy);
1722 llvm::CallSite CS;
1723 if (getTarget().getTriple().getArch() == llvm::Triple::x86) {
1724 llvm::Type *ArgTypes[] = {Int8PtrTy, IntTy};
1725 llvm::Constant *SetJmp3 = CGM.CreateRuntimeFunction(
1726 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/true),
1727 "_setjmp3", ReturnsTwiceAttr);
1728 llvm::Value *Count = ConstantInt::get(IntTy, 0);
1729 llvm::Value *Args[] = {Buf, Count};
1730 CS = EmitRuntimeCallOrInvoke(SetJmp3, Args);
1731 } else {
1732 llvm::Type *ArgTypes[] = {Int8PtrTy, Int8PtrTy};
1733 llvm::Constant *SetJmp = CGM.CreateRuntimeFunction(
1734 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/false),
1735 "_setjmp", ReturnsTwiceAttr);
1736 llvm::Value *FrameAddr =
1737 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress),
1738 ConstantInt::get(Int32Ty, 0));
1739 llvm::Value *Args[] = {Buf, FrameAddr};
1740 CS = EmitRuntimeCallOrInvoke(SetJmp, Args);
1741 }
1742 CS.setAttributes(ReturnsTwiceAttr);
1743 return RValue::get(CS.getInstruction());
1744 }
1745 break;
1746 }
1747
1748 case Builtin::BI__GetExceptionInfo: {
1749 if (llvm::GlobalVariable *GV =
1750 CGM.getCXXABI().getThrowInfo(FD->getParamDecl(0)->getType()))
1751 return RValue::get(llvm::ConstantExpr::getBitCast(GV, CGM.Int8PtrTy));
1752 break;
1753 }
1754 }
1755
1756 // If this is an alias for a lib function (e.g. __builtin_sin), emit
1757 // the call using the normal call path, but using the unmangled
1758 // version of the function name.
1759 if (getContext().BuiltinInfo.isLibFunction(BuiltinID))
1760 return emitLibraryCall(*this, FD, E,
1761 CGM.getBuiltinLibFunction(FD, BuiltinID));
1762
1763 // If this is a predefined lib function (e.g. malloc), emit the call
1764 // using exactly the normal call path.
1765 if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID))
1766 return emitLibraryCall(*this, FD, E, EmitScalarExpr(E->getCallee()));
1767
1768 // See if we have a target specific intrinsic.
1769 const char *Name = getContext().BuiltinInfo.GetName(BuiltinID);
1770 Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic;
1771 if (const char *Prefix =
1772 llvm::Triple::getArchTypePrefix(getTarget().getTriple().getArch())) {
1773 IntrinsicID = Intrinsic::getIntrinsicForGCCBuiltin(Prefix, Name);
1774 // NOTE we dont need to perform a compatibility flag check here since the
1775 // intrinsics are declared in Builtins*.def via LANGBUILTIN which filter the
1776 // MS builtins via ALL_MS_LANGUAGES and are filtered earlier.
1777 if (IntrinsicID == Intrinsic::not_intrinsic)
1778 IntrinsicID = Intrinsic::getIntrinsicForMSBuiltin(Prefix, Name);
1779 }
1780
1781 if (IntrinsicID != Intrinsic::not_intrinsic) {
1782 SmallVector<Value*, 16> Args;
1783
1784 // Find out if any arguments are required to be integer constant
1785 // expressions.
1786 unsigned ICEArguments = 0;
1787 ASTContext::GetBuiltinTypeError Error;
1788 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments);
1789 assert(Error == ASTContext::GE_None && "Should not codegen an error");
1790
1791 Function *F = CGM.getIntrinsic(IntrinsicID);
1792 llvm::FunctionType *FTy = F->getFunctionType();
1793
1794 for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) {
1795 Value *ArgValue;
1796 // If this is a normal argument, just emit it as a scalar.
1797 if ((ICEArguments & (1 << i)) == 0) {
1798 ArgValue = EmitScalarExpr(E->getArg(i));
1799 } else {
1800 // If this is required to be a constant, constant fold it so that we
1801 // know that the generated intrinsic gets a ConstantInt.
1802 llvm::APSInt Result;
1803 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result,getContext());
1804 assert(IsConst && "Constant arg isn't actually constant?");
1805 (void)IsConst;
1806 ArgValue = llvm::ConstantInt::get(getLLVMContext(), Result);
1807 }
1808
1809 // If the intrinsic arg type is different from the builtin arg type
1810 // we need to do a bit cast.
1811 llvm::Type *PTy = FTy->getParamType(i);
1812 if (PTy != ArgValue->getType()) {
1813 assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) &&
1814 "Must be able to losslessly bit cast to param");
1815 ArgValue = Builder.CreateBitCast(ArgValue, PTy);
1816 }
1817
1818 Args.push_back(ArgValue);
1819 }
1820
1821 Value *V = Builder.CreateCall(F, Args);
1822 QualType BuiltinRetType = E->getType();
1823
1824 llvm::Type *RetTy = VoidTy;
1825 if (!BuiltinRetType->isVoidType())
1826 RetTy = ConvertType(BuiltinRetType);
1827
1828 if (RetTy != V->getType()) {
1829 assert(V->getType()->canLosslesslyBitCastTo(RetTy) &&
1830 "Must be able to losslessly bit cast result type");
1831 V = Builder.CreateBitCast(V, RetTy);
1832 }
1833
1834 return RValue::get(V);
1835 }
1836
1837 // See if we have a target specific builtin that needs to be lowered.
1838 if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E))
1839 return RValue::get(V);
1840
1841 ErrorUnsupported(E, "builtin function");
1842
1843 // Unknown builtin, for now just dump it out and return undef.
1844 return GetUndefRValue(E->getType());
1845 }
1846
EmitTargetBuiltinExpr(unsigned BuiltinID,const CallExpr * E)1847 Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID,
1848 const CallExpr *E) {
1849 switch (getTarget().getTriple().getArch()) {
1850 case llvm::Triple::arm:
1851 case llvm::Triple::armeb:
1852 case llvm::Triple::thumb:
1853 case llvm::Triple::thumbeb:
1854 return EmitARMBuiltinExpr(BuiltinID, E);
1855 case llvm::Triple::aarch64:
1856 case llvm::Triple::aarch64_be:
1857 return EmitAArch64BuiltinExpr(BuiltinID, E);
1858 case llvm::Triple::x86:
1859 case llvm::Triple::x86_64:
1860 return EmitX86BuiltinExpr(BuiltinID, E);
1861 case llvm::Triple::ppc:
1862 case llvm::Triple::ppc64:
1863 case llvm::Triple::ppc64le:
1864 return EmitPPCBuiltinExpr(BuiltinID, E);
1865 case llvm::Triple::r600:
1866 case llvm::Triple::amdgcn:
1867 return EmitR600BuiltinExpr(BuiltinID, E);
1868 case llvm::Triple::systemz:
1869 return EmitSystemZBuiltinExpr(BuiltinID, E);
1870 default:
1871 return nullptr;
1872 }
1873 }
1874
GetNeonType(CodeGenFunction * CGF,NeonTypeFlags TypeFlags,bool V1Ty=false)1875 static llvm::VectorType *GetNeonType(CodeGenFunction *CGF,
1876 NeonTypeFlags TypeFlags,
1877 bool V1Ty=false) {
1878 int IsQuad = TypeFlags.isQuad();
1879 switch (TypeFlags.getEltType()) {
1880 case NeonTypeFlags::Int8:
1881 case NeonTypeFlags::Poly8:
1882 return llvm::VectorType::get(CGF->Int8Ty, V1Ty ? 1 : (8 << IsQuad));
1883 case NeonTypeFlags::Int16:
1884 case NeonTypeFlags::Poly16:
1885 case NeonTypeFlags::Float16:
1886 return llvm::VectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad));
1887 case NeonTypeFlags::Int32:
1888 return llvm::VectorType::get(CGF->Int32Ty, V1Ty ? 1 : (2 << IsQuad));
1889 case NeonTypeFlags::Int64:
1890 case NeonTypeFlags::Poly64:
1891 return llvm::VectorType::get(CGF->Int64Ty, V1Ty ? 1 : (1 << IsQuad));
1892 case NeonTypeFlags::Poly128:
1893 // FIXME: i128 and f128 doesn't get fully support in Clang and llvm.
1894 // There is a lot of i128 and f128 API missing.
1895 // so we use v16i8 to represent poly128 and get pattern matched.
1896 return llvm::VectorType::get(CGF->Int8Ty, 16);
1897 case NeonTypeFlags::Float32:
1898 return llvm::VectorType::get(CGF->FloatTy, V1Ty ? 1 : (2 << IsQuad));
1899 case NeonTypeFlags::Float64:
1900 return llvm::VectorType::get(CGF->DoubleTy, V1Ty ? 1 : (1 << IsQuad));
1901 }
1902 llvm_unreachable("Unknown vector element type!");
1903 }
1904
EmitNeonSplat(Value * V,Constant * C)1905 Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) {
1906 unsigned nElts = cast<llvm::VectorType>(V->getType())->getNumElements();
1907 Value* SV = llvm::ConstantVector::getSplat(nElts, C);
1908 return Builder.CreateShuffleVector(V, V, SV, "lane");
1909 }
1910
EmitNeonCall(Function * F,SmallVectorImpl<Value * > & Ops,const char * name,unsigned shift,bool rightshift)1911 Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops,
1912 const char *name,
1913 unsigned shift, bool rightshift) {
1914 unsigned j = 0;
1915 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end();
1916 ai != ae; ++ai, ++j)
1917 if (shift > 0 && shift == j)
1918 Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift);
1919 else
1920 Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name);
1921
1922 return Builder.CreateCall(F, Ops, name);
1923 }
1924
EmitNeonShiftVector(Value * V,llvm::Type * Ty,bool neg)1925 Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty,
1926 bool neg) {
1927 int SV = cast<ConstantInt>(V)->getSExtValue();
1928
1929 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty);
1930 llvm::Constant *C = ConstantInt::get(VTy->getElementType(), neg ? -SV : SV);
1931 return llvm::ConstantVector::getSplat(VTy->getNumElements(), C);
1932 }
1933
1934 // \brief Right-shift a vector by a constant.
EmitNeonRShiftImm(Value * Vec,Value * Shift,llvm::Type * Ty,bool usgn,const char * name)1935 Value *CodeGenFunction::EmitNeonRShiftImm(Value *Vec, Value *Shift,
1936 llvm::Type *Ty, bool usgn,
1937 const char *name) {
1938 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty);
1939
1940 int ShiftAmt = cast<ConstantInt>(Shift)->getSExtValue();
1941 int EltSize = VTy->getScalarSizeInBits();
1942
1943 Vec = Builder.CreateBitCast(Vec, Ty);
1944
1945 // lshr/ashr are undefined when the shift amount is equal to the vector
1946 // element size.
1947 if (ShiftAmt == EltSize) {
1948 if (usgn) {
1949 // Right-shifting an unsigned value by its size yields 0.
1950 llvm::Constant *Zero = ConstantInt::get(VTy->getElementType(), 0);
1951 return llvm::ConstantVector::getSplat(VTy->getNumElements(), Zero);
1952 } else {
1953 // Right-shifting a signed value by its size is equivalent
1954 // to a shift of size-1.
1955 --ShiftAmt;
1956 Shift = ConstantInt::get(VTy->getElementType(), ShiftAmt);
1957 }
1958 }
1959
1960 Shift = EmitNeonShiftVector(Shift, Ty, false);
1961 if (usgn)
1962 return Builder.CreateLShr(Vec, Shift, name);
1963 else
1964 return Builder.CreateAShr(Vec, Shift, name);
1965 }
1966
1967 /// GetPointeeAlignment - Given an expression with a pointer type, find the
1968 /// alignment of the type referenced by the pointer. Skip over implicit
1969 /// casts.
1970 std::pair<llvm::Value*, unsigned>
EmitPointerWithAlignment(const Expr * Addr)1971 CodeGenFunction::EmitPointerWithAlignment(const Expr *Addr) {
1972 assert(Addr->getType()->isPointerType());
1973 Addr = Addr->IgnoreParens();
1974 if (const ImplicitCastExpr *ICE = dyn_cast<ImplicitCastExpr>(Addr)) {
1975 if ((ICE->getCastKind() == CK_BitCast || ICE->getCastKind() == CK_NoOp) &&
1976 ICE->getSubExpr()->getType()->isPointerType()) {
1977 std::pair<llvm::Value*, unsigned> Ptr =
1978 EmitPointerWithAlignment(ICE->getSubExpr());
1979 Ptr.first = Builder.CreateBitCast(Ptr.first,
1980 ConvertType(Addr->getType()));
1981 return Ptr;
1982 } else if (ICE->getCastKind() == CK_ArrayToPointerDecay) {
1983 LValue LV = EmitLValue(ICE->getSubExpr());
1984 unsigned Align = LV.getAlignment().getQuantity();
1985 if (!Align) {
1986 // FIXME: Once LValues are fixed to always set alignment,
1987 // zap this code.
1988 QualType PtTy = ICE->getSubExpr()->getType();
1989 if (!PtTy->isIncompleteType())
1990 Align = getContext().getTypeAlignInChars(PtTy).getQuantity();
1991 else
1992 Align = 1;
1993 }
1994 return std::make_pair(LV.getAddress(), Align);
1995 }
1996 }
1997 if (const UnaryOperator *UO = dyn_cast<UnaryOperator>(Addr)) {
1998 if (UO->getOpcode() == UO_AddrOf) {
1999 LValue LV = EmitLValue(UO->getSubExpr());
2000 unsigned Align = LV.getAlignment().getQuantity();
2001 if (!Align) {
2002 // FIXME: Once LValues are fixed to always set alignment,
2003 // zap this code.
2004 QualType PtTy = UO->getSubExpr()->getType();
2005 if (!PtTy->isIncompleteType())
2006 Align = getContext().getTypeAlignInChars(PtTy).getQuantity();
2007 else
2008 Align = 1;
2009 }
2010 return std::make_pair(LV.getAddress(), Align);
2011 }
2012 }
2013
2014 unsigned Align = 1;
2015 QualType PtTy = Addr->getType()->getPointeeType();
2016 if (!PtTy->isIncompleteType())
2017 Align = getContext().getTypeAlignInChars(PtTy).getQuantity();
2018
2019 return std::make_pair(EmitScalarExpr(Addr), Align);
2020 }
2021
2022 enum {
2023 AddRetType = (1 << 0),
2024 Add1ArgType = (1 << 1),
2025 Add2ArgTypes = (1 << 2),
2026
2027 VectorizeRetType = (1 << 3),
2028 VectorizeArgTypes = (1 << 4),
2029
2030 InventFloatType = (1 << 5),
2031 UnsignedAlts = (1 << 6),
2032
2033 Use64BitVectors = (1 << 7),
2034 Use128BitVectors = (1 << 8),
2035
2036 Vectorize1ArgType = Add1ArgType | VectorizeArgTypes,
2037 VectorRet = AddRetType | VectorizeRetType,
2038 VectorRetGetArgs01 =
2039 AddRetType | Add2ArgTypes | VectorizeRetType | VectorizeArgTypes,
2040 FpCmpzModifiers =
2041 AddRetType | VectorizeRetType | Add1ArgType | InventFloatType
2042 };
2043
2044 struct NeonIntrinsicInfo {
2045 unsigned BuiltinID;
2046 unsigned LLVMIntrinsic;
2047 unsigned AltLLVMIntrinsic;
2048 const char *NameHint;
2049 unsigned TypeModifier;
2050
operator <NeonIntrinsicInfo2051 bool operator<(unsigned RHSBuiltinID) const {
2052 return BuiltinID < RHSBuiltinID;
2053 }
2054 };
2055
2056 #define NEONMAP0(NameBase) \
2057 { NEON::BI__builtin_neon_ ## NameBase, 0, 0, #NameBase, 0 }
2058
2059 #define NEONMAP1(NameBase, LLVMIntrinsic, TypeModifier) \
2060 { NEON:: BI__builtin_neon_ ## NameBase, \
2061 Intrinsic::LLVMIntrinsic, 0, #NameBase, TypeModifier }
2062
2063 #define NEONMAP2(NameBase, LLVMIntrinsic, AltLLVMIntrinsic, TypeModifier) \
2064 { NEON:: BI__builtin_neon_ ## NameBase, \
2065 Intrinsic::LLVMIntrinsic, Intrinsic::AltLLVMIntrinsic, \
2066 #NameBase, TypeModifier }
2067
2068 static NeonIntrinsicInfo ARMSIMDIntrinsicMap [] = {
2069 NEONMAP2(vabd_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts),
2070 NEONMAP2(vabdq_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts),
2071 NEONMAP1(vabs_v, arm_neon_vabs, 0),
2072 NEONMAP1(vabsq_v, arm_neon_vabs, 0),
2073 NEONMAP0(vaddhn_v),
2074 NEONMAP1(vaesdq_v, arm_neon_aesd, 0),
2075 NEONMAP1(vaeseq_v, arm_neon_aese, 0),
2076 NEONMAP1(vaesimcq_v, arm_neon_aesimc, 0),
2077 NEONMAP1(vaesmcq_v, arm_neon_aesmc, 0),
2078 NEONMAP1(vbsl_v, arm_neon_vbsl, AddRetType),
2079 NEONMAP1(vbslq_v, arm_neon_vbsl, AddRetType),
2080 NEONMAP1(vcage_v, arm_neon_vacge, 0),
2081 NEONMAP1(vcageq_v, arm_neon_vacge, 0),
2082 NEONMAP1(vcagt_v, arm_neon_vacgt, 0),
2083 NEONMAP1(vcagtq_v, arm_neon_vacgt, 0),
2084 NEONMAP1(vcale_v, arm_neon_vacge, 0),
2085 NEONMAP1(vcaleq_v, arm_neon_vacge, 0),
2086 NEONMAP1(vcalt_v, arm_neon_vacgt, 0),
2087 NEONMAP1(vcaltq_v, arm_neon_vacgt, 0),
2088 NEONMAP1(vcls_v, arm_neon_vcls, Add1ArgType),
2089 NEONMAP1(vclsq_v, arm_neon_vcls, Add1ArgType),
2090 NEONMAP1(vclz_v, ctlz, Add1ArgType),
2091 NEONMAP1(vclzq_v, ctlz, Add1ArgType),
2092 NEONMAP1(vcnt_v, ctpop, Add1ArgType),
2093 NEONMAP1(vcntq_v, ctpop, Add1ArgType),
2094 NEONMAP1(vcvt_f16_v, arm_neon_vcvtfp2hf, 0),
2095 NEONMAP1(vcvt_f32_f16, arm_neon_vcvthf2fp, 0),
2096 NEONMAP0(vcvt_f32_v),
2097 NEONMAP2(vcvt_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0),
2098 NEONMAP1(vcvt_n_s32_v, arm_neon_vcvtfp2fxs, 0),
2099 NEONMAP1(vcvt_n_s64_v, arm_neon_vcvtfp2fxs, 0),
2100 NEONMAP1(vcvt_n_u32_v, arm_neon_vcvtfp2fxu, 0),
2101 NEONMAP1(vcvt_n_u64_v, arm_neon_vcvtfp2fxu, 0),
2102 NEONMAP0(vcvt_s32_v),
2103 NEONMAP0(vcvt_s64_v),
2104 NEONMAP0(vcvt_u32_v),
2105 NEONMAP0(vcvt_u64_v),
2106 NEONMAP1(vcvta_s32_v, arm_neon_vcvtas, 0),
2107 NEONMAP1(vcvta_s64_v, arm_neon_vcvtas, 0),
2108 NEONMAP1(vcvta_u32_v, arm_neon_vcvtau, 0),
2109 NEONMAP1(vcvta_u64_v, arm_neon_vcvtau, 0),
2110 NEONMAP1(vcvtaq_s32_v, arm_neon_vcvtas, 0),
2111 NEONMAP1(vcvtaq_s64_v, arm_neon_vcvtas, 0),
2112 NEONMAP1(vcvtaq_u32_v, arm_neon_vcvtau, 0),
2113 NEONMAP1(vcvtaq_u64_v, arm_neon_vcvtau, 0),
2114 NEONMAP1(vcvtm_s32_v, arm_neon_vcvtms, 0),
2115 NEONMAP1(vcvtm_s64_v, arm_neon_vcvtms, 0),
2116 NEONMAP1(vcvtm_u32_v, arm_neon_vcvtmu, 0),
2117 NEONMAP1(vcvtm_u64_v, arm_neon_vcvtmu, 0),
2118 NEONMAP1(vcvtmq_s32_v, arm_neon_vcvtms, 0),
2119 NEONMAP1(vcvtmq_s64_v, arm_neon_vcvtms, 0),
2120 NEONMAP1(vcvtmq_u32_v, arm_neon_vcvtmu, 0),
2121 NEONMAP1(vcvtmq_u64_v, arm_neon_vcvtmu, 0),
2122 NEONMAP1(vcvtn_s32_v, arm_neon_vcvtns, 0),
2123 NEONMAP1(vcvtn_s64_v, arm_neon_vcvtns, 0),
2124 NEONMAP1(vcvtn_u32_v, arm_neon_vcvtnu, 0),
2125 NEONMAP1(vcvtn_u64_v, arm_neon_vcvtnu, 0),
2126 NEONMAP1(vcvtnq_s32_v, arm_neon_vcvtns, 0),
2127 NEONMAP1(vcvtnq_s64_v, arm_neon_vcvtns, 0),
2128 NEONMAP1(vcvtnq_u32_v, arm_neon_vcvtnu, 0),
2129 NEONMAP1(vcvtnq_u64_v, arm_neon_vcvtnu, 0),
2130 NEONMAP1(vcvtp_s32_v, arm_neon_vcvtps, 0),
2131 NEONMAP1(vcvtp_s64_v, arm_neon_vcvtps, 0),
2132 NEONMAP1(vcvtp_u32_v, arm_neon_vcvtpu, 0),
2133 NEONMAP1(vcvtp_u64_v, arm_neon_vcvtpu, 0),
2134 NEONMAP1(vcvtpq_s32_v, arm_neon_vcvtps, 0),
2135 NEONMAP1(vcvtpq_s64_v, arm_neon_vcvtps, 0),
2136 NEONMAP1(vcvtpq_u32_v, arm_neon_vcvtpu, 0),
2137 NEONMAP1(vcvtpq_u64_v, arm_neon_vcvtpu, 0),
2138 NEONMAP0(vcvtq_f32_v),
2139 NEONMAP2(vcvtq_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0),
2140 NEONMAP1(vcvtq_n_s32_v, arm_neon_vcvtfp2fxs, 0),
2141 NEONMAP1(vcvtq_n_s64_v, arm_neon_vcvtfp2fxs, 0),
2142 NEONMAP1(vcvtq_n_u32_v, arm_neon_vcvtfp2fxu, 0),
2143 NEONMAP1(vcvtq_n_u64_v, arm_neon_vcvtfp2fxu, 0),
2144 NEONMAP0(vcvtq_s32_v),
2145 NEONMAP0(vcvtq_s64_v),
2146 NEONMAP0(vcvtq_u32_v),
2147 NEONMAP0(vcvtq_u64_v),
2148 NEONMAP0(vext_v),
2149 NEONMAP0(vextq_v),
2150 NEONMAP0(vfma_v),
2151 NEONMAP0(vfmaq_v),
2152 NEONMAP2(vhadd_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts),
2153 NEONMAP2(vhaddq_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts),
2154 NEONMAP2(vhsub_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts),
2155 NEONMAP2(vhsubq_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts),
2156 NEONMAP0(vld1_dup_v),
2157 NEONMAP1(vld1_v, arm_neon_vld1, 0),
2158 NEONMAP0(vld1q_dup_v),
2159 NEONMAP1(vld1q_v, arm_neon_vld1, 0),
2160 NEONMAP1(vld2_lane_v, arm_neon_vld2lane, 0),
2161 NEONMAP1(vld2_v, arm_neon_vld2, 0),
2162 NEONMAP1(vld2q_lane_v, arm_neon_vld2lane, 0),
2163 NEONMAP1(vld2q_v, arm_neon_vld2, 0),
2164 NEONMAP1(vld3_lane_v, arm_neon_vld3lane, 0),
2165 NEONMAP1(vld3_v, arm_neon_vld3, 0),
2166 NEONMAP1(vld3q_lane_v, arm_neon_vld3lane, 0),
2167 NEONMAP1(vld3q_v, arm_neon_vld3, 0),
2168 NEONMAP1(vld4_lane_v, arm_neon_vld4lane, 0),
2169 NEONMAP1(vld4_v, arm_neon_vld4, 0),
2170 NEONMAP1(vld4q_lane_v, arm_neon_vld4lane, 0),
2171 NEONMAP1(vld4q_v, arm_neon_vld4, 0),
2172 NEONMAP2(vmax_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts),
2173 NEONMAP1(vmaxnm_v, arm_neon_vmaxnm, Add1ArgType),
2174 NEONMAP1(vmaxnmq_v, arm_neon_vmaxnm, Add1ArgType),
2175 NEONMAP2(vmaxq_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts),
2176 NEONMAP2(vmin_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts),
2177 NEONMAP1(vminnm_v, arm_neon_vminnm, Add1ArgType),
2178 NEONMAP1(vminnmq_v, arm_neon_vminnm, Add1ArgType),
2179 NEONMAP2(vminq_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts),
2180 NEONMAP0(vmovl_v),
2181 NEONMAP0(vmovn_v),
2182 NEONMAP1(vmul_v, arm_neon_vmulp, Add1ArgType),
2183 NEONMAP0(vmull_v),
2184 NEONMAP1(vmulq_v, arm_neon_vmulp, Add1ArgType),
2185 NEONMAP2(vpadal_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts),
2186 NEONMAP2(vpadalq_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts),
2187 NEONMAP1(vpadd_v, arm_neon_vpadd, Add1ArgType),
2188 NEONMAP2(vpaddl_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts),
2189 NEONMAP2(vpaddlq_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts),
2190 NEONMAP1(vpaddq_v, arm_neon_vpadd, Add1ArgType),
2191 NEONMAP2(vpmax_v, arm_neon_vpmaxu, arm_neon_vpmaxs, Add1ArgType | UnsignedAlts),
2192 NEONMAP2(vpmin_v, arm_neon_vpminu, arm_neon_vpmins, Add1ArgType | UnsignedAlts),
2193 NEONMAP1(vqabs_v, arm_neon_vqabs, Add1ArgType),
2194 NEONMAP1(vqabsq_v, arm_neon_vqabs, Add1ArgType),
2195 NEONMAP2(vqadd_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts),
2196 NEONMAP2(vqaddq_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts),
2197 NEONMAP2(vqdmlal_v, arm_neon_vqdmull, arm_neon_vqadds, 0),
2198 NEONMAP2(vqdmlsl_v, arm_neon_vqdmull, arm_neon_vqsubs, 0),
2199 NEONMAP1(vqdmulh_v, arm_neon_vqdmulh, Add1ArgType),
2200 NEONMAP1(vqdmulhq_v, arm_neon_vqdmulh, Add1ArgType),
2201 NEONMAP1(vqdmull_v, arm_neon_vqdmull, Add1ArgType),
2202 NEONMAP2(vqmovn_v, arm_neon_vqmovnu, arm_neon_vqmovns, Add1ArgType | UnsignedAlts),
2203 NEONMAP1(vqmovun_v, arm_neon_vqmovnsu, Add1ArgType),
2204 NEONMAP1(vqneg_v, arm_neon_vqneg, Add1ArgType),
2205 NEONMAP1(vqnegq_v, arm_neon_vqneg, Add1ArgType),
2206 NEONMAP1(vqrdmulh_v, arm_neon_vqrdmulh, Add1ArgType),
2207 NEONMAP1(vqrdmulhq_v, arm_neon_vqrdmulh, Add1ArgType),
2208 NEONMAP2(vqrshl_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts),
2209 NEONMAP2(vqrshlq_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts),
2210 NEONMAP2(vqshl_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts),
2211 NEONMAP2(vqshl_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts),
2212 NEONMAP2(vqshlq_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts),
2213 NEONMAP2(vqshlq_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts),
2214 NEONMAP1(vqshlu_n_v, arm_neon_vqshiftsu, 0),
2215 NEONMAP1(vqshluq_n_v, arm_neon_vqshiftsu, 0),
2216 NEONMAP2(vqsub_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts),
2217 NEONMAP2(vqsubq_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts),
2218 NEONMAP1(vraddhn_v, arm_neon_vraddhn, Add1ArgType),
2219 NEONMAP2(vrecpe_v, arm_neon_vrecpe, arm_neon_vrecpe, 0),
2220 NEONMAP2(vrecpeq_v, arm_neon_vrecpe, arm_neon_vrecpe, 0),
2221 NEONMAP1(vrecps_v, arm_neon_vrecps, Add1ArgType),
2222 NEONMAP1(vrecpsq_v, arm_neon_vrecps, Add1ArgType),
2223 NEONMAP2(vrhadd_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts),
2224 NEONMAP2(vrhaddq_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts),
2225 NEONMAP1(vrnd_v, arm_neon_vrintz, Add1ArgType),
2226 NEONMAP1(vrnda_v, arm_neon_vrinta, Add1ArgType),
2227 NEONMAP1(vrndaq_v, arm_neon_vrinta, Add1ArgType),
2228 NEONMAP1(vrndm_v, arm_neon_vrintm, Add1ArgType),
2229 NEONMAP1(vrndmq_v, arm_neon_vrintm, Add1ArgType),
2230 NEONMAP1(vrndn_v, arm_neon_vrintn, Add1ArgType),
2231 NEONMAP1(vrndnq_v, arm_neon_vrintn, Add1ArgType),
2232 NEONMAP1(vrndp_v, arm_neon_vrintp, Add1ArgType),
2233 NEONMAP1(vrndpq_v, arm_neon_vrintp, Add1ArgType),
2234 NEONMAP1(vrndq_v, arm_neon_vrintz, Add1ArgType),
2235 NEONMAP1(vrndx_v, arm_neon_vrintx, Add1ArgType),
2236 NEONMAP1(vrndxq_v, arm_neon_vrintx, Add1ArgType),
2237 NEONMAP2(vrshl_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts),
2238 NEONMAP2(vrshlq_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts),
2239 NEONMAP2(vrshr_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts),
2240 NEONMAP2(vrshrq_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts),
2241 NEONMAP2(vrsqrte_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0),
2242 NEONMAP2(vrsqrteq_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0),
2243 NEONMAP1(vrsqrts_v, arm_neon_vrsqrts, Add1ArgType),
2244 NEONMAP1(vrsqrtsq_v, arm_neon_vrsqrts, Add1ArgType),
2245 NEONMAP1(vrsubhn_v, arm_neon_vrsubhn, Add1ArgType),
2246 NEONMAP1(vsha1su0q_v, arm_neon_sha1su0, 0),
2247 NEONMAP1(vsha1su1q_v, arm_neon_sha1su1, 0),
2248 NEONMAP1(vsha256h2q_v, arm_neon_sha256h2, 0),
2249 NEONMAP1(vsha256hq_v, arm_neon_sha256h, 0),
2250 NEONMAP1(vsha256su0q_v, arm_neon_sha256su0, 0),
2251 NEONMAP1(vsha256su1q_v, arm_neon_sha256su1, 0),
2252 NEONMAP0(vshl_n_v),
2253 NEONMAP2(vshl_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts),
2254 NEONMAP0(vshll_n_v),
2255 NEONMAP0(vshlq_n_v),
2256 NEONMAP2(vshlq_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts),
2257 NEONMAP0(vshr_n_v),
2258 NEONMAP0(vshrn_n_v),
2259 NEONMAP0(vshrq_n_v),
2260 NEONMAP1(vst1_v, arm_neon_vst1, 0),
2261 NEONMAP1(vst1q_v, arm_neon_vst1, 0),
2262 NEONMAP1(vst2_lane_v, arm_neon_vst2lane, 0),
2263 NEONMAP1(vst2_v, arm_neon_vst2, 0),
2264 NEONMAP1(vst2q_lane_v, arm_neon_vst2lane, 0),
2265 NEONMAP1(vst2q_v, arm_neon_vst2, 0),
2266 NEONMAP1(vst3_lane_v, arm_neon_vst3lane, 0),
2267 NEONMAP1(vst3_v, arm_neon_vst3, 0),
2268 NEONMAP1(vst3q_lane_v, arm_neon_vst3lane, 0),
2269 NEONMAP1(vst3q_v, arm_neon_vst3, 0),
2270 NEONMAP1(vst4_lane_v, arm_neon_vst4lane, 0),
2271 NEONMAP1(vst4_v, arm_neon_vst4, 0),
2272 NEONMAP1(vst4q_lane_v, arm_neon_vst4lane, 0),
2273 NEONMAP1(vst4q_v, arm_neon_vst4, 0),
2274 NEONMAP0(vsubhn_v),
2275 NEONMAP0(vtrn_v),
2276 NEONMAP0(vtrnq_v),
2277 NEONMAP0(vtst_v),
2278 NEONMAP0(vtstq_v),
2279 NEONMAP0(vuzp_v),
2280 NEONMAP0(vuzpq_v),
2281 NEONMAP0(vzip_v),
2282 NEONMAP0(vzipq_v)
2283 };
2284
2285 static NeonIntrinsicInfo AArch64SIMDIntrinsicMap[] = {
2286 NEONMAP1(vabs_v, aarch64_neon_abs, 0),
2287 NEONMAP1(vabsq_v, aarch64_neon_abs, 0),
2288 NEONMAP0(vaddhn_v),
2289 NEONMAP1(vaesdq_v, aarch64_crypto_aesd, 0),
2290 NEONMAP1(vaeseq_v, aarch64_crypto_aese, 0),
2291 NEONMAP1(vaesimcq_v, aarch64_crypto_aesimc, 0),
2292 NEONMAP1(vaesmcq_v, aarch64_crypto_aesmc, 0),
2293 NEONMAP1(vcage_v, aarch64_neon_facge, 0),
2294 NEONMAP1(vcageq_v, aarch64_neon_facge, 0),
2295 NEONMAP1(vcagt_v, aarch64_neon_facgt, 0),
2296 NEONMAP1(vcagtq_v, aarch64_neon_facgt, 0),
2297 NEONMAP1(vcale_v, aarch64_neon_facge, 0),
2298 NEONMAP1(vcaleq_v, aarch64_neon_facge, 0),
2299 NEONMAP1(vcalt_v, aarch64_neon_facgt, 0),
2300 NEONMAP1(vcaltq_v, aarch64_neon_facgt, 0),
2301 NEONMAP1(vcls_v, aarch64_neon_cls, Add1ArgType),
2302 NEONMAP1(vclsq_v, aarch64_neon_cls, Add1ArgType),
2303 NEONMAP1(vclz_v, ctlz, Add1ArgType),
2304 NEONMAP1(vclzq_v, ctlz, Add1ArgType),
2305 NEONMAP1(vcnt_v, ctpop, Add1ArgType),
2306 NEONMAP1(vcntq_v, ctpop, Add1ArgType),
2307 NEONMAP1(vcvt_f16_v, aarch64_neon_vcvtfp2hf, 0),
2308 NEONMAP1(vcvt_f32_f16, aarch64_neon_vcvthf2fp, 0),
2309 NEONMAP0(vcvt_f32_v),
2310 NEONMAP2(vcvt_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
2311 NEONMAP2(vcvt_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
2312 NEONMAP1(vcvt_n_s32_v, aarch64_neon_vcvtfp2fxs, 0),
2313 NEONMAP1(vcvt_n_s64_v, aarch64_neon_vcvtfp2fxs, 0),
2314 NEONMAP1(vcvt_n_u32_v, aarch64_neon_vcvtfp2fxu, 0),
2315 NEONMAP1(vcvt_n_u64_v, aarch64_neon_vcvtfp2fxu, 0),
2316 NEONMAP0(vcvtq_f32_v),
2317 NEONMAP2(vcvtq_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
2318 NEONMAP2(vcvtq_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
2319 NEONMAP1(vcvtq_n_s32_v, aarch64_neon_vcvtfp2fxs, 0),
2320 NEONMAP1(vcvtq_n_s64_v, aarch64_neon_vcvtfp2fxs, 0),
2321 NEONMAP1(vcvtq_n_u32_v, aarch64_neon_vcvtfp2fxu, 0),
2322 NEONMAP1(vcvtq_n_u64_v, aarch64_neon_vcvtfp2fxu, 0),
2323 NEONMAP1(vcvtx_f32_v, aarch64_neon_fcvtxn, AddRetType | Add1ArgType),
2324 NEONMAP0(vext_v),
2325 NEONMAP0(vextq_v),
2326 NEONMAP0(vfma_v),
2327 NEONMAP0(vfmaq_v),
2328 NEONMAP2(vhadd_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts),
2329 NEONMAP2(vhaddq_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts),
2330 NEONMAP2(vhsub_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts),
2331 NEONMAP2(vhsubq_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts),
2332 NEONMAP0(vmovl_v),
2333 NEONMAP0(vmovn_v),
2334 NEONMAP1(vmul_v, aarch64_neon_pmul, Add1ArgType),
2335 NEONMAP1(vmulq_v, aarch64_neon_pmul, Add1ArgType),
2336 NEONMAP1(vpadd_v, aarch64_neon_addp, Add1ArgType),
2337 NEONMAP2(vpaddl_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts),
2338 NEONMAP2(vpaddlq_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts),
2339 NEONMAP1(vpaddq_v, aarch64_neon_addp, Add1ArgType),
2340 NEONMAP1(vqabs_v, aarch64_neon_sqabs, Add1ArgType),
2341 NEONMAP1(vqabsq_v, aarch64_neon_sqabs, Add1ArgType),
2342 NEONMAP2(vqadd_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts),
2343 NEONMAP2(vqaddq_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts),
2344 NEONMAP2(vqdmlal_v, aarch64_neon_sqdmull, aarch64_neon_sqadd, 0),
2345 NEONMAP2(vqdmlsl_v, aarch64_neon_sqdmull, aarch64_neon_sqsub, 0),
2346 NEONMAP1(vqdmulh_v, aarch64_neon_sqdmulh, Add1ArgType),
2347 NEONMAP1(vqdmulhq_v, aarch64_neon_sqdmulh, Add1ArgType),
2348 NEONMAP1(vqdmull_v, aarch64_neon_sqdmull, Add1ArgType),
2349 NEONMAP2(vqmovn_v, aarch64_neon_uqxtn, aarch64_neon_sqxtn, Add1ArgType | UnsignedAlts),
2350 NEONMAP1(vqmovun_v, aarch64_neon_sqxtun, Add1ArgType),
2351 NEONMAP1(vqneg_v, aarch64_neon_sqneg, Add1ArgType),
2352 NEONMAP1(vqnegq_v, aarch64_neon_sqneg, Add1ArgType),
2353 NEONMAP1(vqrdmulh_v, aarch64_neon_sqrdmulh, Add1ArgType),
2354 NEONMAP1(vqrdmulhq_v, aarch64_neon_sqrdmulh, Add1ArgType),
2355 NEONMAP2(vqrshl_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts),
2356 NEONMAP2(vqrshlq_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts),
2357 NEONMAP2(vqshl_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl, UnsignedAlts),
2358 NEONMAP2(vqshl_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts),
2359 NEONMAP2(vqshlq_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl,UnsignedAlts),
2360 NEONMAP2(vqshlq_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts),
2361 NEONMAP1(vqshlu_n_v, aarch64_neon_sqshlu, 0),
2362 NEONMAP1(vqshluq_n_v, aarch64_neon_sqshlu, 0),
2363 NEONMAP2(vqsub_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts),
2364 NEONMAP2(vqsubq_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts),
2365 NEONMAP1(vraddhn_v, aarch64_neon_raddhn, Add1ArgType),
2366 NEONMAP2(vrecpe_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0),
2367 NEONMAP2(vrecpeq_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0),
2368 NEONMAP1(vrecps_v, aarch64_neon_frecps, Add1ArgType),
2369 NEONMAP1(vrecpsq_v, aarch64_neon_frecps, Add1ArgType),
2370 NEONMAP2(vrhadd_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts),
2371 NEONMAP2(vrhaddq_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts),
2372 NEONMAP2(vrshl_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts),
2373 NEONMAP2(vrshlq_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts),
2374 NEONMAP2(vrshr_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts),
2375 NEONMAP2(vrshrq_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts),
2376 NEONMAP2(vrsqrte_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0),
2377 NEONMAP2(vrsqrteq_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0),
2378 NEONMAP1(vrsqrts_v, aarch64_neon_frsqrts, Add1ArgType),
2379 NEONMAP1(vrsqrtsq_v, aarch64_neon_frsqrts, Add1ArgType),
2380 NEONMAP1(vrsubhn_v, aarch64_neon_rsubhn, Add1ArgType),
2381 NEONMAP1(vsha1su0q_v, aarch64_crypto_sha1su0, 0),
2382 NEONMAP1(vsha1su1q_v, aarch64_crypto_sha1su1, 0),
2383 NEONMAP1(vsha256h2q_v, aarch64_crypto_sha256h2, 0),
2384 NEONMAP1(vsha256hq_v, aarch64_crypto_sha256h, 0),
2385 NEONMAP1(vsha256su0q_v, aarch64_crypto_sha256su0, 0),
2386 NEONMAP1(vsha256su1q_v, aarch64_crypto_sha256su1, 0),
2387 NEONMAP0(vshl_n_v),
2388 NEONMAP2(vshl_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts),
2389 NEONMAP0(vshll_n_v),
2390 NEONMAP0(vshlq_n_v),
2391 NEONMAP2(vshlq_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts),
2392 NEONMAP0(vshr_n_v),
2393 NEONMAP0(vshrn_n_v),
2394 NEONMAP0(vshrq_n_v),
2395 NEONMAP0(vsubhn_v),
2396 NEONMAP0(vtst_v),
2397 NEONMAP0(vtstq_v),
2398 };
2399
2400 static NeonIntrinsicInfo AArch64SISDIntrinsicMap[] = {
2401 NEONMAP1(vabdd_f64, aarch64_sisd_fabd, Add1ArgType),
2402 NEONMAP1(vabds_f32, aarch64_sisd_fabd, Add1ArgType),
2403 NEONMAP1(vabsd_s64, aarch64_neon_abs, Add1ArgType),
2404 NEONMAP1(vaddlv_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType),
2405 NEONMAP1(vaddlv_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType),
2406 NEONMAP1(vaddlvq_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType),
2407 NEONMAP1(vaddlvq_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType),
2408 NEONMAP1(vaddv_f32, aarch64_neon_faddv, AddRetType | Add1ArgType),
2409 NEONMAP1(vaddv_s32, aarch64_neon_saddv, AddRetType | Add1ArgType),
2410 NEONMAP1(vaddv_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType),
2411 NEONMAP1(vaddvq_f32, aarch64_neon_faddv, AddRetType | Add1ArgType),
2412 NEONMAP1(vaddvq_f64, aarch64_neon_faddv, AddRetType | Add1ArgType),
2413 NEONMAP1(vaddvq_s32, aarch64_neon_saddv, AddRetType | Add1ArgType),
2414 NEONMAP1(vaddvq_s64, aarch64_neon_saddv, AddRetType | Add1ArgType),
2415 NEONMAP1(vaddvq_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType),
2416 NEONMAP1(vaddvq_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType),
2417 NEONMAP1(vcaged_f64, aarch64_neon_facge, AddRetType | Add1ArgType),
2418 NEONMAP1(vcages_f32, aarch64_neon_facge, AddRetType | Add1ArgType),
2419 NEONMAP1(vcagtd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType),
2420 NEONMAP1(vcagts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType),
2421 NEONMAP1(vcaled_f64, aarch64_neon_facge, AddRetType | Add1ArgType),
2422 NEONMAP1(vcales_f32, aarch64_neon_facge, AddRetType | Add1ArgType),
2423 NEONMAP1(vcaltd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType),
2424 NEONMAP1(vcalts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType),
2425 NEONMAP1(vcvtad_s64_f64, aarch64_neon_fcvtas, AddRetType | Add1ArgType),
2426 NEONMAP1(vcvtad_u64_f64, aarch64_neon_fcvtau, AddRetType | Add1ArgType),
2427 NEONMAP1(vcvtas_s32_f32, aarch64_neon_fcvtas, AddRetType | Add1ArgType),
2428 NEONMAP1(vcvtas_u32_f32, aarch64_neon_fcvtau, AddRetType | Add1ArgType),
2429 NEONMAP1(vcvtd_n_f64_s64, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType),
2430 NEONMAP1(vcvtd_n_f64_u64, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType),
2431 NEONMAP1(vcvtd_n_s64_f64, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType),
2432 NEONMAP1(vcvtd_n_u64_f64, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType),
2433 NEONMAP1(vcvtmd_s64_f64, aarch64_neon_fcvtms, AddRetType | Add1ArgType),
2434 NEONMAP1(vcvtmd_u64_f64, aarch64_neon_fcvtmu, AddRetType | Add1ArgType),
2435 NEONMAP1(vcvtms_s32_f32, aarch64_neon_fcvtms, AddRetType | Add1ArgType),
2436 NEONMAP1(vcvtms_u32_f32, aarch64_neon_fcvtmu, AddRetType | Add1ArgType),
2437 NEONMAP1(vcvtnd_s64_f64, aarch64_neon_fcvtns, AddRetType | Add1ArgType),
2438 NEONMAP1(vcvtnd_u64_f64, aarch64_neon_fcvtnu, AddRetType | Add1ArgType),
2439 NEONMAP1(vcvtns_s32_f32, aarch64_neon_fcvtns, AddRetType | Add1ArgType),
2440 NEONMAP1(vcvtns_u32_f32, aarch64_neon_fcvtnu, AddRetType | Add1ArgType),
2441 NEONMAP1(vcvtpd_s64_f64, aarch64_neon_fcvtps, AddRetType | Add1ArgType),
2442 NEONMAP1(vcvtpd_u64_f64, aarch64_neon_fcvtpu, AddRetType | Add1ArgType),
2443 NEONMAP1(vcvtps_s32_f32, aarch64_neon_fcvtps, AddRetType | Add1ArgType),
2444 NEONMAP1(vcvtps_u32_f32, aarch64_neon_fcvtpu, AddRetType | Add1ArgType),
2445 NEONMAP1(vcvts_n_f32_s32, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType),
2446 NEONMAP1(vcvts_n_f32_u32, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType),
2447 NEONMAP1(vcvts_n_s32_f32, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType),
2448 NEONMAP1(vcvts_n_u32_f32, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType),
2449 NEONMAP1(vcvtxd_f32_f64, aarch64_sisd_fcvtxn, 0),
2450 NEONMAP1(vmaxnmv_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType),
2451 NEONMAP1(vmaxnmvq_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType),
2452 NEONMAP1(vmaxnmvq_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType),
2453 NEONMAP1(vmaxv_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
2454 NEONMAP1(vmaxv_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType),
2455 NEONMAP1(vmaxv_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType),
2456 NEONMAP1(vmaxvq_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
2457 NEONMAP1(vmaxvq_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
2458 NEONMAP1(vmaxvq_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType),
2459 NEONMAP1(vmaxvq_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType),
2460 NEONMAP1(vminnmv_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
2461 NEONMAP1(vminnmvq_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
2462 NEONMAP1(vminnmvq_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
2463 NEONMAP1(vminv_f32, aarch64_neon_fminv, AddRetType | Add1ArgType),
2464 NEONMAP1(vminv_s32, aarch64_neon_sminv, AddRetType | Add1ArgType),
2465 NEONMAP1(vminv_u32, aarch64_neon_uminv, AddRetType | Add1ArgType),
2466 NEONMAP1(vminvq_f32, aarch64_neon_fminv, AddRetType | Add1ArgType),
2467 NEONMAP1(vminvq_f64, aarch64_neon_fminv, AddRetType | Add1ArgType),
2468 NEONMAP1(vminvq_s32, aarch64_neon_sminv, AddRetType | Add1ArgType),
2469 NEONMAP1(vminvq_u32, aarch64_neon_uminv, AddRetType | Add1ArgType),
2470 NEONMAP1(vmull_p64, aarch64_neon_pmull64, 0),
2471 NEONMAP1(vmulxd_f64, aarch64_neon_fmulx, Add1ArgType),
2472 NEONMAP1(vmulxs_f32, aarch64_neon_fmulx, Add1ArgType),
2473 NEONMAP1(vpaddd_s64, aarch64_neon_uaddv, AddRetType | Add1ArgType),
2474 NEONMAP1(vpaddd_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType),
2475 NEONMAP1(vpmaxnmqd_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType),
2476 NEONMAP1(vpmaxnms_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType),
2477 NEONMAP1(vpmaxqd_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
2478 NEONMAP1(vpmaxs_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
2479 NEONMAP1(vpminnmqd_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
2480 NEONMAP1(vpminnms_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
2481 NEONMAP1(vpminqd_f64, aarch64_neon_fminv, AddRetType | Add1ArgType),
2482 NEONMAP1(vpmins_f32, aarch64_neon_fminv, AddRetType | Add1ArgType),
2483 NEONMAP1(vqabsb_s8, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors),
2484 NEONMAP1(vqabsd_s64, aarch64_neon_sqabs, Add1ArgType),
2485 NEONMAP1(vqabsh_s16, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors),
2486 NEONMAP1(vqabss_s32, aarch64_neon_sqabs, Add1ArgType),
2487 NEONMAP1(vqaddb_s8, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors),
2488 NEONMAP1(vqaddb_u8, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors),
2489 NEONMAP1(vqaddd_s64, aarch64_neon_sqadd, Add1ArgType),
2490 NEONMAP1(vqaddd_u64, aarch64_neon_uqadd, Add1ArgType),
2491 NEONMAP1(vqaddh_s16, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors),
2492 NEONMAP1(vqaddh_u16, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors),
2493 NEONMAP1(vqadds_s32, aarch64_neon_sqadd, Add1ArgType),
2494 NEONMAP1(vqadds_u32, aarch64_neon_uqadd, Add1ArgType),
2495 NEONMAP1(vqdmulhh_s16, aarch64_neon_sqdmulh, Vectorize1ArgType | Use64BitVectors),
2496 NEONMAP1(vqdmulhs_s32, aarch64_neon_sqdmulh, Add1ArgType),
2497 NEONMAP1(vqdmullh_s16, aarch64_neon_sqdmull, VectorRet | Use128BitVectors),
2498 NEONMAP1(vqdmulls_s32, aarch64_neon_sqdmulls_scalar, 0),
2499 NEONMAP1(vqmovnd_s64, aarch64_neon_scalar_sqxtn, AddRetType | Add1ArgType),
2500 NEONMAP1(vqmovnd_u64, aarch64_neon_scalar_uqxtn, AddRetType | Add1ArgType),
2501 NEONMAP1(vqmovnh_s16, aarch64_neon_sqxtn, VectorRet | Use64BitVectors),
2502 NEONMAP1(vqmovnh_u16, aarch64_neon_uqxtn, VectorRet | Use64BitVectors),
2503 NEONMAP1(vqmovns_s32, aarch64_neon_sqxtn, VectorRet | Use64BitVectors),
2504 NEONMAP1(vqmovns_u32, aarch64_neon_uqxtn, VectorRet | Use64BitVectors),
2505 NEONMAP1(vqmovund_s64, aarch64_neon_scalar_sqxtun, AddRetType | Add1ArgType),
2506 NEONMAP1(vqmovunh_s16, aarch64_neon_sqxtun, VectorRet | Use64BitVectors),
2507 NEONMAP1(vqmovuns_s32, aarch64_neon_sqxtun, VectorRet | Use64BitVectors),
2508 NEONMAP1(vqnegb_s8, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors),
2509 NEONMAP1(vqnegd_s64, aarch64_neon_sqneg, Add1ArgType),
2510 NEONMAP1(vqnegh_s16, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors),
2511 NEONMAP1(vqnegs_s32, aarch64_neon_sqneg, Add1ArgType),
2512 NEONMAP1(vqrdmulhh_s16, aarch64_neon_sqrdmulh, Vectorize1ArgType | Use64BitVectors),
2513 NEONMAP1(vqrdmulhs_s32, aarch64_neon_sqrdmulh, Add1ArgType),
2514 NEONMAP1(vqrshlb_s8, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors),
2515 NEONMAP1(vqrshlb_u8, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors),
2516 NEONMAP1(vqrshld_s64, aarch64_neon_sqrshl, Add1ArgType),
2517 NEONMAP1(vqrshld_u64, aarch64_neon_uqrshl, Add1ArgType),
2518 NEONMAP1(vqrshlh_s16, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors),
2519 NEONMAP1(vqrshlh_u16, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors),
2520 NEONMAP1(vqrshls_s32, aarch64_neon_sqrshl, Add1ArgType),
2521 NEONMAP1(vqrshls_u32, aarch64_neon_uqrshl, Add1ArgType),
2522 NEONMAP1(vqrshrnd_n_s64, aarch64_neon_sqrshrn, AddRetType),
2523 NEONMAP1(vqrshrnd_n_u64, aarch64_neon_uqrshrn, AddRetType),
2524 NEONMAP1(vqrshrnh_n_s16, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors),
2525 NEONMAP1(vqrshrnh_n_u16, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors),
2526 NEONMAP1(vqrshrns_n_s32, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors),
2527 NEONMAP1(vqrshrns_n_u32, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors),
2528 NEONMAP1(vqrshrund_n_s64, aarch64_neon_sqrshrun, AddRetType),
2529 NEONMAP1(vqrshrunh_n_s16, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors),
2530 NEONMAP1(vqrshruns_n_s32, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors),
2531 NEONMAP1(vqshlb_n_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors),
2532 NEONMAP1(vqshlb_n_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors),
2533 NEONMAP1(vqshlb_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors),
2534 NEONMAP1(vqshlb_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors),
2535 NEONMAP1(vqshld_s64, aarch64_neon_sqshl, Add1ArgType),
2536 NEONMAP1(vqshld_u64, aarch64_neon_uqshl, Add1ArgType),
2537 NEONMAP1(vqshlh_n_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors),
2538 NEONMAP1(vqshlh_n_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors),
2539 NEONMAP1(vqshlh_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors),
2540 NEONMAP1(vqshlh_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors),
2541 NEONMAP1(vqshls_n_s32, aarch64_neon_sqshl, Add1ArgType),
2542 NEONMAP1(vqshls_n_u32, aarch64_neon_uqshl, Add1ArgType),
2543 NEONMAP1(vqshls_s32, aarch64_neon_sqshl, Add1ArgType),
2544 NEONMAP1(vqshls_u32, aarch64_neon_uqshl, Add1ArgType),
2545 NEONMAP1(vqshlub_n_s8, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors),
2546 NEONMAP1(vqshluh_n_s16, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors),
2547 NEONMAP1(vqshlus_n_s32, aarch64_neon_sqshlu, Add1ArgType),
2548 NEONMAP1(vqshrnd_n_s64, aarch64_neon_sqshrn, AddRetType),
2549 NEONMAP1(vqshrnd_n_u64, aarch64_neon_uqshrn, AddRetType),
2550 NEONMAP1(vqshrnh_n_s16, aarch64_neon_sqshrn, VectorRet | Use64BitVectors),
2551 NEONMAP1(vqshrnh_n_u16, aarch64_neon_uqshrn, VectorRet | Use64BitVectors),
2552 NEONMAP1(vqshrns_n_s32, aarch64_neon_sqshrn, VectorRet | Use64BitVectors),
2553 NEONMAP1(vqshrns_n_u32, aarch64_neon_uqshrn, VectorRet | Use64BitVectors),
2554 NEONMAP1(vqshrund_n_s64, aarch64_neon_sqshrun, AddRetType),
2555 NEONMAP1(vqshrunh_n_s16, aarch64_neon_sqshrun, VectorRet | Use64BitVectors),
2556 NEONMAP1(vqshruns_n_s32, aarch64_neon_sqshrun, VectorRet | Use64BitVectors),
2557 NEONMAP1(vqsubb_s8, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors),
2558 NEONMAP1(vqsubb_u8, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors),
2559 NEONMAP1(vqsubd_s64, aarch64_neon_sqsub, Add1ArgType),
2560 NEONMAP1(vqsubd_u64, aarch64_neon_uqsub, Add1ArgType),
2561 NEONMAP1(vqsubh_s16, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors),
2562 NEONMAP1(vqsubh_u16, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors),
2563 NEONMAP1(vqsubs_s32, aarch64_neon_sqsub, Add1ArgType),
2564 NEONMAP1(vqsubs_u32, aarch64_neon_uqsub, Add1ArgType),
2565 NEONMAP1(vrecped_f64, aarch64_neon_frecpe, Add1ArgType),
2566 NEONMAP1(vrecpes_f32, aarch64_neon_frecpe, Add1ArgType),
2567 NEONMAP1(vrecpxd_f64, aarch64_neon_frecpx, Add1ArgType),
2568 NEONMAP1(vrecpxs_f32, aarch64_neon_frecpx, Add1ArgType),
2569 NEONMAP1(vrshld_s64, aarch64_neon_srshl, Add1ArgType),
2570 NEONMAP1(vrshld_u64, aarch64_neon_urshl, Add1ArgType),
2571 NEONMAP1(vrsqrted_f64, aarch64_neon_frsqrte, Add1ArgType),
2572 NEONMAP1(vrsqrtes_f32, aarch64_neon_frsqrte, Add1ArgType),
2573 NEONMAP1(vrsqrtsd_f64, aarch64_neon_frsqrts, Add1ArgType),
2574 NEONMAP1(vrsqrtss_f32, aarch64_neon_frsqrts, Add1ArgType),
2575 NEONMAP1(vsha1cq_u32, aarch64_crypto_sha1c, 0),
2576 NEONMAP1(vsha1h_u32, aarch64_crypto_sha1h, 0),
2577 NEONMAP1(vsha1mq_u32, aarch64_crypto_sha1m, 0),
2578 NEONMAP1(vsha1pq_u32, aarch64_crypto_sha1p, 0),
2579 NEONMAP1(vshld_s64, aarch64_neon_sshl, Add1ArgType),
2580 NEONMAP1(vshld_u64, aarch64_neon_ushl, Add1ArgType),
2581 NEONMAP1(vslid_n_s64, aarch64_neon_vsli, Vectorize1ArgType),
2582 NEONMAP1(vslid_n_u64, aarch64_neon_vsli, Vectorize1ArgType),
2583 NEONMAP1(vsqaddb_u8, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors),
2584 NEONMAP1(vsqaddd_u64, aarch64_neon_usqadd, Add1ArgType),
2585 NEONMAP1(vsqaddh_u16, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors),
2586 NEONMAP1(vsqadds_u32, aarch64_neon_usqadd, Add1ArgType),
2587 NEONMAP1(vsrid_n_s64, aarch64_neon_vsri, Vectorize1ArgType),
2588 NEONMAP1(vsrid_n_u64, aarch64_neon_vsri, Vectorize1ArgType),
2589 NEONMAP1(vuqaddb_s8, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors),
2590 NEONMAP1(vuqaddd_s64, aarch64_neon_suqadd, Add1ArgType),
2591 NEONMAP1(vuqaddh_s16, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors),
2592 NEONMAP1(vuqadds_s32, aarch64_neon_suqadd, Add1ArgType),
2593 };
2594
2595 #undef NEONMAP0
2596 #undef NEONMAP1
2597 #undef NEONMAP2
2598
2599 static bool NEONSIMDIntrinsicsProvenSorted = false;
2600
2601 static bool AArch64SIMDIntrinsicsProvenSorted = false;
2602 static bool AArch64SISDIntrinsicsProvenSorted = false;
2603
2604
2605 static const NeonIntrinsicInfo *
findNeonIntrinsicInMap(ArrayRef<NeonIntrinsicInfo> IntrinsicMap,unsigned BuiltinID,bool & MapProvenSorted)2606 findNeonIntrinsicInMap(ArrayRef<NeonIntrinsicInfo> IntrinsicMap,
2607 unsigned BuiltinID, bool &MapProvenSorted) {
2608
2609 #ifndef NDEBUG
2610 if (!MapProvenSorted) {
2611 // FIXME: use std::is_sorted once C++11 is allowed
2612 for (unsigned i = 0; i < IntrinsicMap.size() - 1; ++i)
2613 assert(IntrinsicMap[i].BuiltinID <= IntrinsicMap[i + 1].BuiltinID);
2614 MapProvenSorted = true;
2615 }
2616 #endif
2617
2618 const NeonIntrinsicInfo *Builtin =
2619 std::lower_bound(IntrinsicMap.begin(), IntrinsicMap.end(), BuiltinID);
2620
2621 if (Builtin != IntrinsicMap.end() && Builtin->BuiltinID == BuiltinID)
2622 return Builtin;
2623
2624 return nullptr;
2625 }
2626
LookupNeonLLVMIntrinsic(unsigned IntrinsicID,unsigned Modifier,llvm::Type * ArgType,const CallExpr * E)2627 Function *CodeGenFunction::LookupNeonLLVMIntrinsic(unsigned IntrinsicID,
2628 unsigned Modifier,
2629 llvm::Type *ArgType,
2630 const CallExpr *E) {
2631 int VectorSize = 0;
2632 if (Modifier & Use64BitVectors)
2633 VectorSize = 64;
2634 else if (Modifier & Use128BitVectors)
2635 VectorSize = 128;
2636
2637 // Return type.
2638 SmallVector<llvm::Type *, 3> Tys;
2639 if (Modifier & AddRetType) {
2640 llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext()));
2641 if (Modifier & VectorizeRetType)
2642 Ty = llvm::VectorType::get(
2643 Ty, VectorSize ? VectorSize / Ty->getPrimitiveSizeInBits() : 1);
2644
2645 Tys.push_back(Ty);
2646 }
2647
2648 // Arguments.
2649 if (Modifier & VectorizeArgTypes) {
2650 int Elts = VectorSize ? VectorSize / ArgType->getPrimitiveSizeInBits() : 1;
2651 ArgType = llvm::VectorType::get(ArgType, Elts);
2652 }
2653
2654 if (Modifier & (Add1ArgType | Add2ArgTypes))
2655 Tys.push_back(ArgType);
2656
2657 if (Modifier & Add2ArgTypes)
2658 Tys.push_back(ArgType);
2659
2660 if (Modifier & InventFloatType)
2661 Tys.push_back(FloatTy);
2662
2663 return CGM.getIntrinsic(IntrinsicID, Tys);
2664 }
2665
EmitCommonNeonSISDBuiltinExpr(CodeGenFunction & CGF,const NeonIntrinsicInfo & SISDInfo,SmallVectorImpl<Value * > & Ops,const CallExpr * E)2666 static Value *EmitCommonNeonSISDBuiltinExpr(CodeGenFunction &CGF,
2667 const NeonIntrinsicInfo &SISDInfo,
2668 SmallVectorImpl<Value *> &Ops,
2669 const CallExpr *E) {
2670 unsigned BuiltinID = SISDInfo.BuiltinID;
2671 unsigned int Int = SISDInfo.LLVMIntrinsic;
2672 unsigned Modifier = SISDInfo.TypeModifier;
2673 const char *s = SISDInfo.NameHint;
2674
2675 switch (BuiltinID) {
2676 case NEON::BI__builtin_neon_vcled_s64:
2677 case NEON::BI__builtin_neon_vcled_u64:
2678 case NEON::BI__builtin_neon_vcles_f32:
2679 case NEON::BI__builtin_neon_vcled_f64:
2680 case NEON::BI__builtin_neon_vcltd_s64:
2681 case NEON::BI__builtin_neon_vcltd_u64:
2682 case NEON::BI__builtin_neon_vclts_f32:
2683 case NEON::BI__builtin_neon_vcltd_f64:
2684 case NEON::BI__builtin_neon_vcales_f32:
2685 case NEON::BI__builtin_neon_vcaled_f64:
2686 case NEON::BI__builtin_neon_vcalts_f32:
2687 case NEON::BI__builtin_neon_vcaltd_f64:
2688 // Only one direction of comparisons actually exist, cmle is actually a cmge
2689 // with swapped operands. The table gives us the right intrinsic but we
2690 // still need to do the swap.
2691 std::swap(Ops[0], Ops[1]);
2692 break;
2693 }
2694
2695 assert(Int && "Generic code assumes a valid intrinsic");
2696
2697 // Determine the type(s) of this overloaded AArch64 intrinsic.
2698 const Expr *Arg = E->getArg(0);
2699 llvm::Type *ArgTy = CGF.ConvertType(Arg->getType());
2700 Function *F = CGF.LookupNeonLLVMIntrinsic(Int, Modifier, ArgTy, E);
2701
2702 int j = 0;
2703 ConstantInt *C0 = ConstantInt::get(CGF.SizeTy, 0);
2704 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end();
2705 ai != ae; ++ai, ++j) {
2706 llvm::Type *ArgTy = ai->getType();
2707 if (Ops[j]->getType()->getPrimitiveSizeInBits() ==
2708 ArgTy->getPrimitiveSizeInBits())
2709 continue;
2710
2711 assert(ArgTy->isVectorTy() && !Ops[j]->getType()->isVectorTy());
2712 // The constant argument to an _n_ intrinsic always has Int32Ty, so truncate
2713 // it before inserting.
2714 Ops[j] =
2715 CGF.Builder.CreateTruncOrBitCast(Ops[j], ArgTy->getVectorElementType());
2716 Ops[j] =
2717 CGF.Builder.CreateInsertElement(UndefValue::get(ArgTy), Ops[j], C0);
2718 }
2719
2720 Value *Result = CGF.EmitNeonCall(F, Ops, s);
2721 llvm::Type *ResultType = CGF.ConvertType(E->getType());
2722 if (ResultType->getPrimitiveSizeInBits() <
2723 Result->getType()->getPrimitiveSizeInBits())
2724 return CGF.Builder.CreateExtractElement(Result, C0);
2725
2726 return CGF.Builder.CreateBitCast(Result, ResultType, s);
2727 }
2728
EmitCommonNeonBuiltinExpr(unsigned BuiltinID,unsigned LLVMIntrinsic,unsigned AltLLVMIntrinsic,const char * NameHint,unsigned Modifier,const CallExpr * E,SmallVectorImpl<llvm::Value * > & Ops,llvm::Value * Align)2729 Value *CodeGenFunction::EmitCommonNeonBuiltinExpr(
2730 unsigned BuiltinID, unsigned LLVMIntrinsic, unsigned AltLLVMIntrinsic,
2731 const char *NameHint, unsigned Modifier, const CallExpr *E,
2732 SmallVectorImpl<llvm::Value *> &Ops, llvm::Value *Align) {
2733 // Get the last argument, which specifies the vector type.
2734 llvm::APSInt NeonTypeConst;
2735 const Expr *Arg = E->getArg(E->getNumArgs() - 1);
2736 if (!Arg->isIntegerConstantExpr(NeonTypeConst, getContext()))
2737 return nullptr;
2738
2739 // Determine the type of this overloaded NEON intrinsic.
2740 NeonTypeFlags Type(NeonTypeConst.getZExtValue());
2741 bool Usgn = Type.isUnsigned();
2742 bool Quad = Type.isQuad();
2743
2744 llvm::VectorType *VTy = GetNeonType(this, Type);
2745 llvm::Type *Ty = VTy;
2746 if (!Ty)
2747 return nullptr;
2748
2749 unsigned Int = LLVMIntrinsic;
2750 if ((Modifier & UnsignedAlts) && !Usgn)
2751 Int = AltLLVMIntrinsic;
2752
2753 switch (BuiltinID) {
2754 default: break;
2755 case NEON::BI__builtin_neon_vabs_v:
2756 case NEON::BI__builtin_neon_vabsq_v:
2757 if (VTy->getElementType()->isFloatingPointTy())
2758 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::fabs, Ty), Ops, "vabs");
2759 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vabs");
2760 case NEON::BI__builtin_neon_vaddhn_v: {
2761 llvm::VectorType *SrcTy =
2762 llvm::VectorType::getExtendedElementVectorType(VTy);
2763
2764 // %sum = add <4 x i32> %lhs, %rhs
2765 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy);
2766 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy);
2767 Ops[0] = Builder.CreateAdd(Ops[0], Ops[1], "vaddhn");
2768
2769 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16>
2770 Constant *ShiftAmt = ConstantInt::get(SrcTy->getElementType(),
2771 SrcTy->getScalarSizeInBits() / 2);
2772 ShiftAmt = ConstantVector::getSplat(VTy->getNumElements(), ShiftAmt);
2773 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vaddhn");
2774
2775 // %res = trunc <4 x i32> %high to <4 x i16>
2776 return Builder.CreateTrunc(Ops[0], VTy, "vaddhn");
2777 }
2778 case NEON::BI__builtin_neon_vcale_v:
2779 case NEON::BI__builtin_neon_vcaleq_v:
2780 case NEON::BI__builtin_neon_vcalt_v:
2781 case NEON::BI__builtin_neon_vcaltq_v:
2782 std::swap(Ops[0], Ops[1]);
2783 case NEON::BI__builtin_neon_vcage_v:
2784 case NEON::BI__builtin_neon_vcageq_v:
2785 case NEON::BI__builtin_neon_vcagt_v:
2786 case NEON::BI__builtin_neon_vcagtq_v: {
2787 llvm::Type *VecFlt = llvm::VectorType::get(
2788 VTy->getScalarSizeInBits() == 32 ? FloatTy : DoubleTy,
2789 VTy->getNumElements());
2790 llvm::Type *Tys[] = { VTy, VecFlt };
2791 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys);
2792 return EmitNeonCall(F, Ops, NameHint);
2793 }
2794 case NEON::BI__builtin_neon_vclz_v:
2795 case NEON::BI__builtin_neon_vclzq_v:
2796 // We generate target-independent intrinsic, which needs a second argument
2797 // for whether or not clz of zero is undefined; on ARM it isn't.
2798 Ops.push_back(Builder.getInt1(getTarget().isCLZForZeroUndef()));
2799 break;
2800 case NEON::BI__builtin_neon_vcvt_f32_v:
2801 case NEON::BI__builtin_neon_vcvtq_f32_v:
2802 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2803 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, Quad));
2804 return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt")
2805 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt");
2806 case NEON::BI__builtin_neon_vcvt_n_f32_v:
2807 case NEON::BI__builtin_neon_vcvt_n_f64_v:
2808 case NEON::BI__builtin_neon_vcvtq_n_f32_v:
2809 case NEON::BI__builtin_neon_vcvtq_n_f64_v: {
2810 bool Double =
2811 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64);
2812 llvm::Type *FloatTy =
2813 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64
2814 : NeonTypeFlags::Float32,
2815 false, Quad));
2816 llvm::Type *Tys[2] = { FloatTy, Ty };
2817 Int = Usgn ? LLVMIntrinsic : AltLLVMIntrinsic;
2818 Function *F = CGM.getIntrinsic(Int, Tys);
2819 return EmitNeonCall(F, Ops, "vcvt_n");
2820 }
2821 case NEON::BI__builtin_neon_vcvt_n_s32_v:
2822 case NEON::BI__builtin_neon_vcvt_n_u32_v:
2823 case NEON::BI__builtin_neon_vcvt_n_s64_v:
2824 case NEON::BI__builtin_neon_vcvt_n_u64_v:
2825 case NEON::BI__builtin_neon_vcvtq_n_s32_v:
2826 case NEON::BI__builtin_neon_vcvtq_n_u32_v:
2827 case NEON::BI__builtin_neon_vcvtq_n_s64_v:
2828 case NEON::BI__builtin_neon_vcvtq_n_u64_v: {
2829 bool Double =
2830 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64);
2831 llvm::Type *FloatTy =
2832 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64
2833 : NeonTypeFlags::Float32,
2834 false, Quad));
2835 llvm::Type *Tys[2] = { Ty, FloatTy };
2836 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys);
2837 return EmitNeonCall(F, Ops, "vcvt_n");
2838 }
2839 case NEON::BI__builtin_neon_vcvt_s32_v:
2840 case NEON::BI__builtin_neon_vcvt_u32_v:
2841 case NEON::BI__builtin_neon_vcvt_s64_v:
2842 case NEON::BI__builtin_neon_vcvt_u64_v:
2843 case NEON::BI__builtin_neon_vcvtq_s32_v:
2844 case NEON::BI__builtin_neon_vcvtq_u32_v:
2845 case NEON::BI__builtin_neon_vcvtq_s64_v:
2846 case NEON::BI__builtin_neon_vcvtq_u64_v: {
2847 bool Double =
2848 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64);
2849 llvm::Type *FloatTy =
2850 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64
2851 : NeonTypeFlags::Float32,
2852 false, Quad));
2853 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy);
2854 return Usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt")
2855 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt");
2856 }
2857 case NEON::BI__builtin_neon_vcvta_s32_v:
2858 case NEON::BI__builtin_neon_vcvta_s64_v:
2859 case NEON::BI__builtin_neon_vcvta_u32_v:
2860 case NEON::BI__builtin_neon_vcvta_u64_v:
2861 case NEON::BI__builtin_neon_vcvtaq_s32_v:
2862 case NEON::BI__builtin_neon_vcvtaq_s64_v:
2863 case NEON::BI__builtin_neon_vcvtaq_u32_v:
2864 case NEON::BI__builtin_neon_vcvtaq_u64_v:
2865 case NEON::BI__builtin_neon_vcvtn_s32_v:
2866 case NEON::BI__builtin_neon_vcvtn_s64_v:
2867 case NEON::BI__builtin_neon_vcvtn_u32_v:
2868 case NEON::BI__builtin_neon_vcvtn_u64_v:
2869 case NEON::BI__builtin_neon_vcvtnq_s32_v:
2870 case NEON::BI__builtin_neon_vcvtnq_s64_v:
2871 case NEON::BI__builtin_neon_vcvtnq_u32_v:
2872 case NEON::BI__builtin_neon_vcvtnq_u64_v:
2873 case NEON::BI__builtin_neon_vcvtp_s32_v:
2874 case NEON::BI__builtin_neon_vcvtp_s64_v:
2875 case NEON::BI__builtin_neon_vcvtp_u32_v:
2876 case NEON::BI__builtin_neon_vcvtp_u64_v:
2877 case NEON::BI__builtin_neon_vcvtpq_s32_v:
2878 case NEON::BI__builtin_neon_vcvtpq_s64_v:
2879 case NEON::BI__builtin_neon_vcvtpq_u32_v:
2880 case NEON::BI__builtin_neon_vcvtpq_u64_v:
2881 case NEON::BI__builtin_neon_vcvtm_s32_v:
2882 case NEON::BI__builtin_neon_vcvtm_s64_v:
2883 case NEON::BI__builtin_neon_vcvtm_u32_v:
2884 case NEON::BI__builtin_neon_vcvtm_u64_v:
2885 case NEON::BI__builtin_neon_vcvtmq_s32_v:
2886 case NEON::BI__builtin_neon_vcvtmq_s64_v:
2887 case NEON::BI__builtin_neon_vcvtmq_u32_v:
2888 case NEON::BI__builtin_neon_vcvtmq_u64_v: {
2889 bool Double =
2890 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64);
2891 llvm::Type *InTy =
2892 GetNeonType(this,
2893 NeonTypeFlags(Double ? NeonTypeFlags::Float64
2894 : NeonTypeFlags::Float32, false, Quad));
2895 llvm::Type *Tys[2] = { Ty, InTy };
2896 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, NameHint);
2897 }
2898 case NEON::BI__builtin_neon_vext_v:
2899 case NEON::BI__builtin_neon_vextq_v: {
2900 int CV = cast<ConstantInt>(Ops[2])->getSExtValue();
2901 SmallVector<Constant*, 16> Indices;
2902 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i)
2903 Indices.push_back(ConstantInt::get(Int32Ty, i+CV));
2904
2905 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2906 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
2907 Value *SV = llvm::ConstantVector::get(Indices);
2908 return Builder.CreateShuffleVector(Ops[0], Ops[1], SV, "vext");
2909 }
2910 case NEON::BI__builtin_neon_vfma_v:
2911 case NEON::BI__builtin_neon_vfmaq_v: {
2912 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty);
2913 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2914 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
2915 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
2916
2917 // NEON intrinsic puts accumulator first, unlike the LLVM fma.
2918 return Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]);
2919 }
2920 case NEON::BI__builtin_neon_vld1_v:
2921 case NEON::BI__builtin_neon_vld1q_v:
2922 Ops.push_back(Align);
2923 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vld1");
2924 case NEON::BI__builtin_neon_vld2_v:
2925 case NEON::BI__builtin_neon_vld2q_v:
2926 case NEON::BI__builtin_neon_vld3_v:
2927 case NEON::BI__builtin_neon_vld3q_v:
2928 case NEON::BI__builtin_neon_vld4_v:
2929 case NEON::BI__builtin_neon_vld4q_v: {
2930 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Ty);
2931 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, NameHint);
2932 Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
2933 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2934 return Builder.CreateStore(Ops[1], Ops[0]);
2935 }
2936 case NEON::BI__builtin_neon_vld1_dup_v:
2937 case NEON::BI__builtin_neon_vld1q_dup_v: {
2938 Value *V = UndefValue::get(Ty);
2939 Ty = llvm::PointerType::getUnqual(VTy->getElementType());
2940 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2941 LoadInst *Ld = Builder.CreateLoad(Ops[0]);
2942 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue());
2943 llvm::Constant *CI = ConstantInt::get(SizeTy, 0);
2944 Ops[0] = Builder.CreateInsertElement(V, Ld, CI);
2945 return EmitNeonSplat(Ops[0], CI);
2946 }
2947 case NEON::BI__builtin_neon_vld2_lane_v:
2948 case NEON::BI__builtin_neon_vld2q_lane_v:
2949 case NEON::BI__builtin_neon_vld3_lane_v:
2950 case NEON::BI__builtin_neon_vld3q_lane_v:
2951 case NEON::BI__builtin_neon_vld4_lane_v:
2952 case NEON::BI__builtin_neon_vld4q_lane_v: {
2953 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Ty);
2954 for (unsigned I = 2; I < Ops.size() - 1; ++I)
2955 Ops[I] = Builder.CreateBitCast(Ops[I], Ty);
2956 Ops.push_back(Align);
2957 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), NameHint);
2958 Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
2959 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2960 return Builder.CreateStore(Ops[1], Ops[0]);
2961 }
2962 case NEON::BI__builtin_neon_vmovl_v: {
2963 llvm::Type *DTy =llvm::VectorType::getTruncatedElementVectorType(VTy);
2964 Ops[0] = Builder.CreateBitCast(Ops[0], DTy);
2965 if (Usgn)
2966 return Builder.CreateZExt(Ops[0], Ty, "vmovl");
2967 return Builder.CreateSExt(Ops[0], Ty, "vmovl");
2968 }
2969 case NEON::BI__builtin_neon_vmovn_v: {
2970 llvm::Type *QTy = llvm::VectorType::getExtendedElementVectorType(VTy);
2971 Ops[0] = Builder.CreateBitCast(Ops[0], QTy);
2972 return Builder.CreateTrunc(Ops[0], Ty, "vmovn");
2973 }
2974 case NEON::BI__builtin_neon_vmull_v:
2975 // FIXME: the integer vmull operations could be emitted in terms of pure
2976 // LLVM IR (2 exts followed by a mul). Unfortunately LLVM has a habit of
2977 // hoisting the exts outside loops. Until global ISel comes along that can
2978 // see through such movement this leads to bad CodeGen. So we need an
2979 // intrinsic for now.
2980 Int = Usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls;
2981 Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp : Int;
2982 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull");
2983 case NEON::BI__builtin_neon_vpadal_v:
2984 case NEON::BI__builtin_neon_vpadalq_v: {
2985 // The source operand type has twice as many elements of half the size.
2986 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits();
2987 llvm::Type *EltTy =
2988 llvm::IntegerType::get(getLLVMContext(), EltBits / 2);
2989 llvm::Type *NarrowTy =
2990 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2);
2991 llvm::Type *Tys[2] = { Ty, NarrowTy };
2992 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, NameHint);
2993 }
2994 case NEON::BI__builtin_neon_vpaddl_v:
2995 case NEON::BI__builtin_neon_vpaddlq_v: {
2996 // The source operand type has twice as many elements of half the size.
2997 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits();
2998 llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2);
2999 llvm::Type *NarrowTy =
3000 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2);
3001 llvm::Type *Tys[2] = { Ty, NarrowTy };
3002 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl");
3003 }
3004 case NEON::BI__builtin_neon_vqdmlal_v:
3005 case NEON::BI__builtin_neon_vqdmlsl_v: {
3006 SmallVector<Value *, 2> MulOps(Ops.begin() + 1, Ops.end());
3007 Value *Mul = EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty),
3008 MulOps, "vqdmlal");
3009
3010 SmallVector<Value *, 2> AccumOps;
3011 AccumOps.push_back(Ops[0]);
3012 AccumOps.push_back(Mul);
3013 return EmitNeonCall(CGM.getIntrinsic(AltLLVMIntrinsic, Ty),
3014 AccumOps, NameHint);
3015 }
3016 case NEON::BI__builtin_neon_vqshl_n_v:
3017 case NEON::BI__builtin_neon_vqshlq_n_v:
3018 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n",
3019 1, false);
3020 case NEON::BI__builtin_neon_vqshlu_n_v:
3021 case NEON::BI__builtin_neon_vqshluq_n_v:
3022 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshlu_n",
3023 1, false);
3024 case NEON::BI__builtin_neon_vrecpe_v:
3025 case NEON::BI__builtin_neon_vrecpeq_v:
3026 case NEON::BI__builtin_neon_vrsqrte_v:
3027 case NEON::BI__builtin_neon_vrsqrteq_v:
3028 Int = Ty->isFPOrFPVectorTy() ? LLVMIntrinsic : AltLLVMIntrinsic;
3029 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, NameHint);
3030
3031 case NEON::BI__builtin_neon_vrshr_n_v:
3032 case NEON::BI__builtin_neon_vrshrq_n_v:
3033 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n",
3034 1, true);
3035 case NEON::BI__builtin_neon_vshl_n_v:
3036 case NEON::BI__builtin_neon_vshlq_n_v:
3037 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false);
3038 return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1],
3039 "vshl_n");
3040 case NEON::BI__builtin_neon_vshll_n_v: {
3041 llvm::Type *SrcTy = llvm::VectorType::getTruncatedElementVectorType(VTy);
3042 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy);
3043 if (Usgn)
3044 Ops[0] = Builder.CreateZExt(Ops[0], VTy);
3045 else
3046 Ops[0] = Builder.CreateSExt(Ops[0], VTy);
3047 Ops[1] = EmitNeonShiftVector(Ops[1], VTy, false);
3048 return Builder.CreateShl(Ops[0], Ops[1], "vshll_n");
3049 }
3050 case NEON::BI__builtin_neon_vshrn_n_v: {
3051 llvm::Type *SrcTy = llvm::VectorType::getExtendedElementVectorType(VTy);
3052 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy);
3053 Ops[1] = EmitNeonShiftVector(Ops[1], SrcTy, false);
3054 if (Usgn)
3055 Ops[0] = Builder.CreateLShr(Ops[0], Ops[1]);
3056 else
3057 Ops[0] = Builder.CreateAShr(Ops[0], Ops[1]);
3058 return Builder.CreateTrunc(Ops[0], Ty, "vshrn_n");
3059 }
3060 case NEON::BI__builtin_neon_vshr_n_v:
3061 case NEON::BI__builtin_neon_vshrq_n_v:
3062 return EmitNeonRShiftImm(Ops[0], Ops[1], Ty, Usgn, "vshr_n");
3063 case NEON::BI__builtin_neon_vst1_v:
3064 case NEON::BI__builtin_neon_vst1q_v:
3065 case NEON::BI__builtin_neon_vst2_v:
3066 case NEON::BI__builtin_neon_vst2q_v:
3067 case NEON::BI__builtin_neon_vst3_v:
3068 case NEON::BI__builtin_neon_vst3q_v:
3069 case NEON::BI__builtin_neon_vst4_v:
3070 case NEON::BI__builtin_neon_vst4q_v:
3071 case NEON::BI__builtin_neon_vst2_lane_v:
3072 case NEON::BI__builtin_neon_vst2q_lane_v:
3073 case NEON::BI__builtin_neon_vst3_lane_v:
3074 case NEON::BI__builtin_neon_vst3q_lane_v:
3075 case NEON::BI__builtin_neon_vst4_lane_v:
3076 case NEON::BI__builtin_neon_vst4q_lane_v:
3077 Ops.push_back(Align);
3078 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "");
3079 case NEON::BI__builtin_neon_vsubhn_v: {
3080 llvm::VectorType *SrcTy =
3081 llvm::VectorType::getExtendedElementVectorType(VTy);
3082
3083 // %sum = add <4 x i32> %lhs, %rhs
3084 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy);
3085 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy);
3086 Ops[0] = Builder.CreateSub(Ops[0], Ops[1], "vsubhn");
3087
3088 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16>
3089 Constant *ShiftAmt = ConstantInt::get(SrcTy->getElementType(),
3090 SrcTy->getScalarSizeInBits() / 2);
3091 ShiftAmt = ConstantVector::getSplat(VTy->getNumElements(), ShiftAmt);
3092 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vsubhn");
3093
3094 // %res = trunc <4 x i32> %high to <4 x i16>
3095 return Builder.CreateTrunc(Ops[0], VTy, "vsubhn");
3096 }
3097 case NEON::BI__builtin_neon_vtrn_v:
3098 case NEON::BI__builtin_neon_vtrnq_v: {
3099 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
3100 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
3101 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
3102 Value *SV = nullptr;
3103
3104 for (unsigned vi = 0; vi != 2; ++vi) {
3105 SmallVector<Constant*, 16> Indices;
3106 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) {
3107 Indices.push_back(Builder.getInt32(i+vi));
3108 Indices.push_back(Builder.getInt32(i+e+vi));
3109 }
3110 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi);
3111 SV = llvm::ConstantVector::get(Indices);
3112 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn");
3113 SV = Builder.CreateStore(SV, Addr);
3114 }
3115 return SV;
3116 }
3117 case NEON::BI__builtin_neon_vtst_v:
3118 case NEON::BI__builtin_neon_vtstq_v: {
3119 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
3120 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
3121 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]);
3122 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0],
3123 ConstantAggregateZero::get(Ty));
3124 return Builder.CreateSExt(Ops[0], Ty, "vtst");
3125 }
3126 case NEON::BI__builtin_neon_vuzp_v:
3127 case NEON::BI__builtin_neon_vuzpq_v: {
3128 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
3129 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
3130 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
3131 Value *SV = nullptr;
3132
3133 for (unsigned vi = 0; vi != 2; ++vi) {
3134 SmallVector<Constant*, 16> Indices;
3135 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i)
3136 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi));
3137
3138 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi);
3139 SV = llvm::ConstantVector::get(Indices);
3140 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp");
3141 SV = Builder.CreateStore(SV, Addr);
3142 }
3143 return SV;
3144 }
3145 case NEON::BI__builtin_neon_vzip_v:
3146 case NEON::BI__builtin_neon_vzipq_v: {
3147 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
3148 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
3149 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
3150 Value *SV = nullptr;
3151
3152 for (unsigned vi = 0; vi != 2; ++vi) {
3153 SmallVector<Constant*, 16> Indices;
3154 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) {
3155 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1));
3156 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e));
3157 }
3158 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi);
3159 SV = llvm::ConstantVector::get(Indices);
3160 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip");
3161 SV = Builder.CreateStore(SV, Addr);
3162 }
3163 return SV;
3164 }
3165 }
3166
3167 assert(Int && "Expected valid intrinsic number");
3168
3169 // Determine the type(s) of this overloaded AArch64 intrinsic.
3170 Function *F = LookupNeonLLVMIntrinsic(Int, Modifier, Ty, E);
3171
3172 Value *Result = EmitNeonCall(F, Ops, NameHint);
3173 llvm::Type *ResultType = ConvertType(E->getType());
3174 // AArch64 intrinsic one-element vector type cast to
3175 // scalar type expected by the builtin
3176 return Builder.CreateBitCast(Result, ResultType, NameHint);
3177 }
3178
EmitAArch64CompareBuiltinExpr(Value * Op,llvm::Type * Ty,const CmpInst::Predicate Fp,const CmpInst::Predicate Ip,const Twine & Name)3179 Value *CodeGenFunction::EmitAArch64CompareBuiltinExpr(
3180 Value *Op, llvm::Type *Ty, const CmpInst::Predicate Fp,
3181 const CmpInst::Predicate Ip, const Twine &Name) {
3182 llvm::Type *OTy = Op->getType();
3183
3184 // FIXME: this is utterly horrific. We should not be looking at previous
3185 // codegen context to find out what needs doing. Unfortunately TableGen
3186 // currently gives us exactly the same calls for vceqz_f32 and vceqz_s32
3187 // (etc).
3188 if (BitCastInst *BI = dyn_cast<BitCastInst>(Op))
3189 OTy = BI->getOperand(0)->getType();
3190
3191 Op = Builder.CreateBitCast(Op, OTy);
3192 if (OTy->getScalarType()->isFloatingPointTy()) {
3193 Op = Builder.CreateFCmp(Fp, Op, Constant::getNullValue(OTy));
3194 } else {
3195 Op = Builder.CreateICmp(Ip, Op, Constant::getNullValue(OTy));
3196 }
3197 return Builder.CreateSExt(Op, Ty, Name);
3198 }
3199
packTBLDVectorList(CodeGenFunction & CGF,ArrayRef<Value * > Ops,Value * ExtOp,Value * IndexOp,llvm::Type * ResTy,unsigned IntID,const char * Name)3200 static Value *packTBLDVectorList(CodeGenFunction &CGF, ArrayRef<Value *> Ops,
3201 Value *ExtOp, Value *IndexOp,
3202 llvm::Type *ResTy, unsigned IntID,
3203 const char *Name) {
3204 SmallVector<Value *, 2> TblOps;
3205 if (ExtOp)
3206 TblOps.push_back(ExtOp);
3207
3208 // Build a vector containing sequential number like (0, 1, 2, ..., 15)
3209 SmallVector<Constant*, 16> Indices;
3210 llvm::VectorType *TblTy = cast<llvm::VectorType>(Ops[0]->getType());
3211 for (unsigned i = 0, e = TblTy->getNumElements(); i != e; ++i) {
3212 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i));
3213 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i+1));
3214 }
3215 Value *SV = llvm::ConstantVector::get(Indices);
3216
3217 int PairPos = 0, End = Ops.size() - 1;
3218 while (PairPos < End) {
3219 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos],
3220 Ops[PairPos+1], SV, Name));
3221 PairPos += 2;
3222 }
3223
3224 // If there's an odd number of 64-bit lookup table, fill the high 64-bit
3225 // of the 128-bit lookup table with zero.
3226 if (PairPos == End) {
3227 Value *ZeroTbl = ConstantAggregateZero::get(TblTy);
3228 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos],
3229 ZeroTbl, SV, Name));
3230 }
3231
3232 Function *TblF;
3233 TblOps.push_back(IndexOp);
3234 TblF = CGF.CGM.getIntrinsic(IntID, ResTy);
3235
3236 return CGF.EmitNeonCall(TblF, TblOps, Name);
3237 }
3238
GetValueForARMHint(unsigned BuiltinID)3239 Value *CodeGenFunction::GetValueForARMHint(unsigned BuiltinID) {
3240 switch (BuiltinID) {
3241 default:
3242 return nullptr;
3243 case ARM::BI__builtin_arm_nop:
3244 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint),
3245 llvm::ConstantInt::get(Int32Ty, 0));
3246 case ARM::BI__builtin_arm_yield:
3247 case ARM::BI__yield:
3248 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint),
3249 llvm::ConstantInt::get(Int32Ty, 1));
3250 case ARM::BI__builtin_arm_wfe:
3251 case ARM::BI__wfe:
3252 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint),
3253 llvm::ConstantInt::get(Int32Ty, 2));
3254 case ARM::BI__builtin_arm_wfi:
3255 case ARM::BI__wfi:
3256 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint),
3257 llvm::ConstantInt::get(Int32Ty, 3));
3258 case ARM::BI__builtin_arm_sev:
3259 case ARM::BI__sev:
3260 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint),
3261 llvm::ConstantInt::get(Int32Ty, 4));
3262 case ARM::BI__builtin_arm_sevl:
3263 case ARM::BI__sevl:
3264 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint),
3265 llvm::ConstantInt::get(Int32Ty, 5));
3266 }
3267 }
3268
EmitARMBuiltinExpr(unsigned BuiltinID,const CallExpr * E)3269 Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID,
3270 const CallExpr *E) {
3271 if (auto Hint = GetValueForARMHint(BuiltinID))
3272 return Hint;
3273
3274 if (BuiltinID == ARM::BI__emit) {
3275 bool IsThumb = getTarget().getTriple().getArch() == llvm::Triple::thumb;
3276 llvm::FunctionType *FTy =
3277 llvm::FunctionType::get(VoidTy, /*Variadic=*/false);
3278
3279 APSInt Value;
3280 if (!E->getArg(0)->EvaluateAsInt(Value, CGM.getContext()))
3281 llvm_unreachable("Sema will ensure that the parameter is constant");
3282
3283 uint64_t ZExtValue = Value.zextOrTrunc(IsThumb ? 16 : 32).getZExtValue();
3284
3285 llvm::InlineAsm *Emit =
3286 IsThumb ? InlineAsm::get(FTy, ".inst.n 0x" + utohexstr(ZExtValue), "",
3287 /*SideEffects=*/true)
3288 : InlineAsm::get(FTy, ".inst 0x" + utohexstr(ZExtValue), "",
3289 /*SideEffects=*/true);
3290
3291 return Builder.CreateCall(Emit);
3292 }
3293
3294 if (BuiltinID == ARM::BI__builtin_arm_dbg) {
3295 Value *Option = EmitScalarExpr(E->getArg(0));
3296 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_dbg), Option);
3297 }
3298
3299 if (BuiltinID == ARM::BI__builtin_arm_prefetch) {
3300 Value *Address = EmitScalarExpr(E->getArg(0));
3301 Value *RW = EmitScalarExpr(E->getArg(1));
3302 Value *IsData = EmitScalarExpr(E->getArg(2));
3303
3304 // Locality is not supported on ARM target
3305 Value *Locality = llvm::ConstantInt::get(Int32Ty, 3);
3306
3307 Value *F = CGM.getIntrinsic(Intrinsic::prefetch);
3308 return Builder.CreateCall4(F, Address, RW, Locality, IsData);
3309 }
3310
3311 if (BuiltinID == ARM::BI__builtin_arm_rbit) {
3312 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_rbit),
3313 EmitScalarExpr(E->getArg(0)),
3314 "rbit");
3315 }
3316
3317 if (BuiltinID == ARM::BI__clear_cache) {
3318 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments");
3319 const FunctionDecl *FD = E->getDirectCallee();
3320 SmallVector<Value*, 2> Ops;
3321 for (unsigned i = 0; i < 2; i++)
3322 Ops.push_back(EmitScalarExpr(E->getArg(i)));
3323 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType());
3324 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty);
3325 StringRef Name = FD->getName();
3326 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops);
3327 }
3328
3329 if (BuiltinID == ARM::BI__builtin_arm_ldrexd ||
3330 ((BuiltinID == ARM::BI__builtin_arm_ldrex ||
3331 BuiltinID == ARM::BI__builtin_arm_ldaex) &&
3332 getContext().getTypeSize(E->getType()) == 64) ||
3333 BuiltinID == ARM::BI__ldrexd) {
3334 Function *F;
3335
3336 switch (BuiltinID) {
3337 default: llvm_unreachable("unexpected builtin");
3338 case ARM::BI__builtin_arm_ldaex:
3339 F = CGM.getIntrinsic(Intrinsic::arm_ldaexd);
3340 break;
3341 case ARM::BI__builtin_arm_ldrexd:
3342 case ARM::BI__builtin_arm_ldrex:
3343 case ARM::BI__ldrexd:
3344 F = CGM.getIntrinsic(Intrinsic::arm_ldrexd);
3345 break;
3346 }
3347
3348 Value *LdPtr = EmitScalarExpr(E->getArg(0));
3349 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy),
3350 "ldrexd");
3351
3352 Value *Val0 = Builder.CreateExtractValue(Val, 1);
3353 Value *Val1 = Builder.CreateExtractValue(Val, 0);
3354 Val0 = Builder.CreateZExt(Val0, Int64Ty);
3355 Val1 = Builder.CreateZExt(Val1, Int64Ty);
3356
3357 Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32);
3358 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */);
3359 Val = Builder.CreateOr(Val, Val1);
3360 return Builder.CreateBitCast(Val, ConvertType(E->getType()));
3361 }
3362
3363 if (BuiltinID == ARM::BI__builtin_arm_ldrex ||
3364 BuiltinID == ARM::BI__builtin_arm_ldaex) {
3365 Value *LoadAddr = EmitScalarExpr(E->getArg(0));
3366
3367 QualType Ty = E->getType();
3368 llvm::Type *RealResTy = ConvertType(Ty);
3369 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(),
3370 getContext().getTypeSize(Ty));
3371 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo());
3372
3373 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_ldaex
3374 ? Intrinsic::arm_ldaex
3375 : Intrinsic::arm_ldrex,
3376 LoadAddr->getType());
3377 Value *Val = Builder.CreateCall(F, LoadAddr, "ldrex");
3378
3379 if (RealResTy->isPointerTy())
3380 return Builder.CreateIntToPtr(Val, RealResTy);
3381 else {
3382 Val = Builder.CreateTruncOrBitCast(Val, IntResTy);
3383 return Builder.CreateBitCast(Val, RealResTy);
3384 }
3385 }
3386
3387 if (BuiltinID == ARM::BI__builtin_arm_strexd ||
3388 ((BuiltinID == ARM::BI__builtin_arm_stlex ||
3389 BuiltinID == ARM::BI__builtin_arm_strex) &&
3390 getContext().getTypeSize(E->getArg(0)->getType()) == 64)) {
3391 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex
3392 ? Intrinsic::arm_stlexd
3393 : Intrinsic::arm_strexd);
3394 llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, nullptr);
3395
3396 Value *Tmp = CreateMemTemp(E->getArg(0)->getType());
3397 Value *Val = EmitScalarExpr(E->getArg(0));
3398 Builder.CreateStore(Val, Tmp);
3399
3400 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy));
3401 Val = Builder.CreateLoad(LdPtr);
3402
3403 Value *Arg0 = Builder.CreateExtractValue(Val, 0);
3404 Value *Arg1 = Builder.CreateExtractValue(Val, 1);
3405 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), Int8PtrTy);
3406 return Builder.CreateCall3(F, Arg0, Arg1, StPtr, "strexd");
3407 }
3408
3409 if (BuiltinID == ARM::BI__builtin_arm_strex ||
3410 BuiltinID == ARM::BI__builtin_arm_stlex) {
3411 Value *StoreVal = EmitScalarExpr(E->getArg(0));
3412 Value *StoreAddr = EmitScalarExpr(E->getArg(1));
3413
3414 QualType Ty = E->getArg(0)->getType();
3415 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(),
3416 getContext().getTypeSize(Ty));
3417 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo());
3418
3419 if (StoreVal->getType()->isPointerTy())
3420 StoreVal = Builder.CreatePtrToInt(StoreVal, Int32Ty);
3421 else {
3422 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy);
3423 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int32Ty);
3424 }
3425
3426 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex
3427 ? Intrinsic::arm_stlex
3428 : Intrinsic::arm_strex,
3429 StoreAddr->getType());
3430 return Builder.CreateCall2(F, StoreVal, StoreAddr, "strex");
3431 }
3432
3433 if (BuiltinID == ARM::BI__builtin_arm_clrex) {
3434 Function *F = CGM.getIntrinsic(Intrinsic::arm_clrex);
3435 return Builder.CreateCall(F);
3436 }
3437
3438 // CRC32
3439 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic;
3440 switch (BuiltinID) {
3441 case ARM::BI__builtin_arm_crc32b:
3442 CRCIntrinsicID = Intrinsic::arm_crc32b; break;
3443 case ARM::BI__builtin_arm_crc32cb:
3444 CRCIntrinsicID = Intrinsic::arm_crc32cb; break;
3445 case ARM::BI__builtin_arm_crc32h:
3446 CRCIntrinsicID = Intrinsic::arm_crc32h; break;
3447 case ARM::BI__builtin_arm_crc32ch:
3448 CRCIntrinsicID = Intrinsic::arm_crc32ch; break;
3449 case ARM::BI__builtin_arm_crc32w:
3450 case ARM::BI__builtin_arm_crc32d:
3451 CRCIntrinsicID = Intrinsic::arm_crc32w; break;
3452 case ARM::BI__builtin_arm_crc32cw:
3453 case ARM::BI__builtin_arm_crc32cd:
3454 CRCIntrinsicID = Intrinsic::arm_crc32cw; break;
3455 }
3456
3457 if (CRCIntrinsicID != Intrinsic::not_intrinsic) {
3458 Value *Arg0 = EmitScalarExpr(E->getArg(0));
3459 Value *Arg1 = EmitScalarExpr(E->getArg(1));
3460
3461 // crc32{c,}d intrinsics are implemnted as two calls to crc32{c,}w
3462 // intrinsics, hence we need different codegen for these cases.
3463 if (BuiltinID == ARM::BI__builtin_arm_crc32d ||
3464 BuiltinID == ARM::BI__builtin_arm_crc32cd) {
3465 Value *C1 = llvm::ConstantInt::get(Int64Ty, 32);
3466 Value *Arg1a = Builder.CreateTruncOrBitCast(Arg1, Int32Ty);
3467 Value *Arg1b = Builder.CreateLShr(Arg1, C1);
3468 Arg1b = Builder.CreateTruncOrBitCast(Arg1b, Int32Ty);
3469
3470 Function *F = CGM.getIntrinsic(CRCIntrinsicID);
3471 Value *Res = Builder.CreateCall2(F, Arg0, Arg1a);
3472 return Builder.CreateCall2(F, Res, Arg1b);
3473 } else {
3474 Arg1 = Builder.CreateZExtOrBitCast(Arg1, Int32Ty);
3475
3476 Function *F = CGM.getIntrinsic(CRCIntrinsicID);
3477 return Builder.CreateCall2(F, Arg0, Arg1);
3478 }
3479 }
3480
3481 SmallVector<Value*, 4> Ops;
3482 llvm::Value *Align = nullptr;
3483 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) {
3484 if (i == 0) {
3485 switch (BuiltinID) {
3486 case NEON::BI__builtin_neon_vld1_v:
3487 case NEON::BI__builtin_neon_vld1q_v:
3488 case NEON::BI__builtin_neon_vld1q_lane_v:
3489 case NEON::BI__builtin_neon_vld1_lane_v:
3490 case NEON::BI__builtin_neon_vld1_dup_v:
3491 case NEON::BI__builtin_neon_vld1q_dup_v:
3492 case NEON::BI__builtin_neon_vst1_v:
3493 case NEON::BI__builtin_neon_vst1q_v:
3494 case NEON::BI__builtin_neon_vst1q_lane_v:
3495 case NEON::BI__builtin_neon_vst1_lane_v:
3496 case NEON::BI__builtin_neon_vst2_v:
3497 case NEON::BI__builtin_neon_vst2q_v:
3498 case NEON::BI__builtin_neon_vst2_lane_v:
3499 case NEON::BI__builtin_neon_vst2q_lane_v:
3500 case NEON::BI__builtin_neon_vst3_v:
3501 case NEON::BI__builtin_neon_vst3q_v:
3502 case NEON::BI__builtin_neon_vst3_lane_v:
3503 case NEON::BI__builtin_neon_vst3q_lane_v:
3504 case NEON::BI__builtin_neon_vst4_v:
3505 case NEON::BI__builtin_neon_vst4q_v:
3506 case NEON::BI__builtin_neon_vst4_lane_v:
3507 case NEON::BI__builtin_neon_vst4q_lane_v:
3508 // Get the alignment for the argument in addition to the value;
3509 // we'll use it later.
3510 std::pair<llvm::Value*, unsigned> Src =
3511 EmitPointerWithAlignment(E->getArg(0));
3512 Ops.push_back(Src.first);
3513 Align = Builder.getInt32(Src.second);
3514 continue;
3515 }
3516 }
3517 if (i == 1) {
3518 switch (BuiltinID) {
3519 case NEON::BI__builtin_neon_vld2_v:
3520 case NEON::BI__builtin_neon_vld2q_v:
3521 case NEON::BI__builtin_neon_vld3_v:
3522 case NEON::BI__builtin_neon_vld3q_v:
3523 case NEON::BI__builtin_neon_vld4_v:
3524 case NEON::BI__builtin_neon_vld4q_v:
3525 case NEON::BI__builtin_neon_vld2_lane_v:
3526 case NEON::BI__builtin_neon_vld2q_lane_v:
3527 case NEON::BI__builtin_neon_vld3_lane_v:
3528 case NEON::BI__builtin_neon_vld3q_lane_v:
3529 case NEON::BI__builtin_neon_vld4_lane_v:
3530 case NEON::BI__builtin_neon_vld4q_lane_v:
3531 case NEON::BI__builtin_neon_vld2_dup_v:
3532 case NEON::BI__builtin_neon_vld3_dup_v:
3533 case NEON::BI__builtin_neon_vld4_dup_v:
3534 // Get the alignment for the argument in addition to the value;
3535 // we'll use it later.
3536 std::pair<llvm::Value*, unsigned> Src =
3537 EmitPointerWithAlignment(E->getArg(1));
3538 Ops.push_back(Src.first);
3539 Align = Builder.getInt32(Src.second);
3540 continue;
3541 }
3542 }
3543 Ops.push_back(EmitScalarExpr(E->getArg(i)));
3544 }
3545
3546 switch (BuiltinID) {
3547 default: break;
3548 // vget_lane and vset_lane are not overloaded and do not have an extra
3549 // argument that specifies the vector type.
3550 case NEON::BI__builtin_neon_vget_lane_i8:
3551 case NEON::BI__builtin_neon_vget_lane_i16:
3552 case NEON::BI__builtin_neon_vget_lane_i32:
3553 case NEON::BI__builtin_neon_vget_lane_i64:
3554 case NEON::BI__builtin_neon_vget_lane_f32:
3555 case NEON::BI__builtin_neon_vgetq_lane_i8:
3556 case NEON::BI__builtin_neon_vgetq_lane_i16:
3557 case NEON::BI__builtin_neon_vgetq_lane_i32:
3558 case NEON::BI__builtin_neon_vgetq_lane_i64:
3559 case NEON::BI__builtin_neon_vgetq_lane_f32:
3560 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
3561 "vget_lane");
3562 case NEON::BI__builtin_neon_vset_lane_i8:
3563 case NEON::BI__builtin_neon_vset_lane_i16:
3564 case NEON::BI__builtin_neon_vset_lane_i32:
3565 case NEON::BI__builtin_neon_vset_lane_i64:
3566 case NEON::BI__builtin_neon_vset_lane_f32:
3567 case NEON::BI__builtin_neon_vsetq_lane_i8:
3568 case NEON::BI__builtin_neon_vsetq_lane_i16:
3569 case NEON::BI__builtin_neon_vsetq_lane_i32:
3570 case NEON::BI__builtin_neon_vsetq_lane_i64:
3571 case NEON::BI__builtin_neon_vsetq_lane_f32:
3572 Ops.push_back(EmitScalarExpr(E->getArg(2)));
3573 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane");
3574
3575 // Non-polymorphic crypto instructions also not overloaded
3576 case NEON::BI__builtin_neon_vsha1h_u32:
3577 Ops.push_back(EmitScalarExpr(E->getArg(0)));
3578 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1h), Ops,
3579 "vsha1h");
3580 case NEON::BI__builtin_neon_vsha1cq_u32:
3581 Ops.push_back(EmitScalarExpr(E->getArg(2)));
3582 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1c), Ops,
3583 "vsha1h");
3584 case NEON::BI__builtin_neon_vsha1pq_u32:
3585 Ops.push_back(EmitScalarExpr(E->getArg(2)));
3586 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1p), Ops,
3587 "vsha1h");
3588 case NEON::BI__builtin_neon_vsha1mq_u32:
3589 Ops.push_back(EmitScalarExpr(E->getArg(2)));
3590 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1m), Ops,
3591 "vsha1h");
3592 }
3593
3594 // Get the last argument, which specifies the vector type.
3595 llvm::APSInt Result;
3596 const Expr *Arg = E->getArg(E->getNumArgs()-1);
3597 if (!Arg->isIntegerConstantExpr(Result, getContext()))
3598 return nullptr;
3599
3600 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f ||
3601 BuiltinID == ARM::BI__builtin_arm_vcvtr_d) {
3602 // Determine the overloaded type of this builtin.
3603 llvm::Type *Ty;
3604 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f)
3605 Ty = FloatTy;
3606 else
3607 Ty = DoubleTy;
3608
3609 // Determine whether this is an unsigned conversion or not.
3610 bool usgn = Result.getZExtValue() == 1;
3611 unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr;
3612
3613 // Call the appropriate intrinsic.
3614 Function *F = CGM.getIntrinsic(Int, Ty);
3615 return Builder.CreateCall(F, Ops, "vcvtr");
3616 }
3617
3618 // Determine the type of this overloaded NEON intrinsic.
3619 NeonTypeFlags Type(Result.getZExtValue());
3620 bool usgn = Type.isUnsigned();
3621 bool rightShift = false;
3622
3623 llvm::VectorType *VTy = GetNeonType(this, Type);
3624 llvm::Type *Ty = VTy;
3625 if (!Ty)
3626 return nullptr;
3627
3628 // Many NEON builtins have identical semantics and uses in ARM and
3629 // AArch64. Emit these in a single function.
3630 auto IntrinsicMap = makeArrayRef(ARMSIMDIntrinsicMap);
3631 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap(
3632 IntrinsicMap, BuiltinID, NEONSIMDIntrinsicsProvenSorted);
3633 if (Builtin)
3634 return EmitCommonNeonBuiltinExpr(
3635 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic,
3636 Builtin->NameHint, Builtin->TypeModifier, E, Ops, Align);
3637
3638 unsigned Int;
3639 switch (BuiltinID) {
3640 default: return nullptr;
3641 case NEON::BI__builtin_neon_vld1q_lane_v:
3642 // Handle 64-bit integer elements as a special case. Use shuffles of
3643 // one-element vectors to avoid poor code for i64 in the backend.
3644 if (VTy->getElementType()->isIntegerTy(64)) {
3645 // Extract the other lane.
3646 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
3647 int Lane = cast<ConstantInt>(Ops[2])->getZExtValue();
3648 Value *SV = llvm::ConstantVector::get(ConstantInt::get(Int32Ty, 1-Lane));
3649 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV);
3650 // Load the value as a one-element vector.
3651 Ty = llvm::VectorType::get(VTy->getElementType(), 1);
3652 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Ty);
3653 Value *Ld = Builder.CreateCall2(F, Ops[0], Align);
3654 // Combine them.
3655 SmallVector<Constant*, 2> Indices;
3656 Indices.push_back(ConstantInt::get(Int32Ty, 1-Lane));
3657 Indices.push_back(ConstantInt::get(Int32Ty, Lane));
3658 SV = llvm::ConstantVector::get(Indices);
3659 return Builder.CreateShuffleVector(Ops[1], Ld, SV, "vld1q_lane");
3660 }
3661 // fall through
3662 case NEON::BI__builtin_neon_vld1_lane_v: {
3663 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
3664 Ty = llvm::PointerType::getUnqual(VTy->getElementType());
3665 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
3666 LoadInst *Ld = Builder.CreateLoad(Ops[0]);
3667 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue());
3668 return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane");
3669 }
3670 case NEON::BI__builtin_neon_vld2_dup_v:
3671 case NEON::BI__builtin_neon_vld3_dup_v:
3672 case NEON::BI__builtin_neon_vld4_dup_v: {
3673 // Handle 64-bit elements as a special-case. There is no "dup" needed.
3674 if (VTy->getElementType()->getPrimitiveSizeInBits() == 64) {
3675 switch (BuiltinID) {
3676 case NEON::BI__builtin_neon_vld2_dup_v:
3677 Int = Intrinsic::arm_neon_vld2;
3678 break;
3679 case NEON::BI__builtin_neon_vld3_dup_v:
3680 Int = Intrinsic::arm_neon_vld3;
3681 break;
3682 case NEON::BI__builtin_neon_vld4_dup_v:
3683 Int = Intrinsic::arm_neon_vld4;
3684 break;
3685 default: llvm_unreachable("unknown vld_dup intrinsic?");
3686 }
3687 Function *F = CGM.getIntrinsic(Int, Ty);
3688 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld_dup");
3689 Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
3690 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
3691 return Builder.CreateStore(Ops[1], Ops[0]);
3692 }
3693 switch (BuiltinID) {
3694 case NEON::BI__builtin_neon_vld2_dup_v:
3695 Int = Intrinsic::arm_neon_vld2lane;
3696 break;
3697 case NEON::BI__builtin_neon_vld3_dup_v:
3698 Int = Intrinsic::arm_neon_vld3lane;
3699 break;
3700 case NEON::BI__builtin_neon_vld4_dup_v:
3701 Int = Intrinsic::arm_neon_vld4lane;
3702 break;
3703 default: llvm_unreachable("unknown vld_dup intrinsic?");
3704 }
3705 Function *F = CGM.getIntrinsic(Int, Ty);
3706 llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType());
3707
3708 SmallVector<Value*, 6> Args;
3709 Args.push_back(Ops[1]);
3710 Args.append(STy->getNumElements(), UndefValue::get(Ty));
3711
3712 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0);
3713 Args.push_back(CI);
3714 Args.push_back(Align);
3715
3716 Ops[1] = Builder.CreateCall(F, Args, "vld_dup");
3717 // splat lane 0 to all elts in each vector of the result.
3718 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
3719 Value *Val = Builder.CreateExtractValue(Ops[1], i);
3720 Value *Elt = Builder.CreateBitCast(Val, Ty);
3721 Elt = EmitNeonSplat(Elt, CI);
3722 Elt = Builder.CreateBitCast(Elt, Val->getType());
3723 Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i);
3724 }
3725 Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
3726 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
3727 return Builder.CreateStore(Ops[1], Ops[0]);
3728 }
3729 case NEON::BI__builtin_neon_vqrshrn_n_v:
3730 Int =
3731 usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns;
3732 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n",
3733 1, true);
3734 case NEON::BI__builtin_neon_vqrshrun_n_v:
3735 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty),
3736 Ops, "vqrshrun_n", 1, true);
3737 case NEON::BI__builtin_neon_vqshrn_n_v:
3738 Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns;
3739 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n",
3740 1, true);
3741 case NEON::BI__builtin_neon_vqshrun_n_v:
3742 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty),
3743 Ops, "vqshrun_n", 1, true);
3744 case NEON::BI__builtin_neon_vrecpe_v:
3745 case NEON::BI__builtin_neon_vrecpeq_v:
3746 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty),
3747 Ops, "vrecpe");
3748 case NEON::BI__builtin_neon_vrshrn_n_v:
3749 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty),
3750 Ops, "vrshrn_n", 1, true);
3751 case NEON::BI__builtin_neon_vrsra_n_v:
3752 case NEON::BI__builtin_neon_vrsraq_n_v:
3753 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
3754 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
3755 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true);
3756 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts;
3757 Ops[1] = Builder.CreateCall2(CGM.getIntrinsic(Int, Ty), Ops[1], Ops[2]);
3758 return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n");
3759 case NEON::BI__builtin_neon_vsri_n_v:
3760 case NEON::BI__builtin_neon_vsriq_n_v:
3761 rightShift = true;
3762 case NEON::BI__builtin_neon_vsli_n_v:
3763 case NEON::BI__builtin_neon_vsliq_n_v:
3764 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift);
3765 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty),
3766 Ops, "vsli_n");
3767 case NEON::BI__builtin_neon_vsra_n_v:
3768 case NEON::BI__builtin_neon_vsraq_n_v:
3769 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
3770 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n");
3771 return Builder.CreateAdd(Ops[0], Ops[1]);
3772 case NEON::BI__builtin_neon_vst1q_lane_v:
3773 // Handle 64-bit integer elements as a special case. Use a shuffle to get
3774 // a one-element vector and avoid poor code for i64 in the backend.
3775 if (VTy->getElementType()->isIntegerTy(64)) {
3776 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
3777 Value *SV = llvm::ConstantVector::get(cast<llvm::Constant>(Ops[2]));
3778 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV);
3779 Ops[2] = Align;
3780 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1,
3781 Ops[1]->getType()), Ops);
3782 }
3783 // fall through
3784 case NEON::BI__builtin_neon_vst1_lane_v: {
3785 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
3786 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]);
3787 Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
3788 StoreInst *St = Builder.CreateStore(Ops[1],
3789 Builder.CreateBitCast(Ops[0], Ty));
3790 St->setAlignment(cast<ConstantInt>(Align)->getZExtValue());
3791 return St;
3792 }
3793 case NEON::BI__builtin_neon_vtbl1_v:
3794 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1),
3795 Ops, "vtbl1");
3796 case NEON::BI__builtin_neon_vtbl2_v:
3797 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2),
3798 Ops, "vtbl2");
3799 case NEON::BI__builtin_neon_vtbl3_v:
3800 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3),
3801 Ops, "vtbl3");
3802 case NEON::BI__builtin_neon_vtbl4_v:
3803 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4),
3804 Ops, "vtbl4");
3805 case NEON::BI__builtin_neon_vtbx1_v:
3806 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1),
3807 Ops, "vtbx1");
3808 case NEON::BI__builtin_neon_vtbx2_v:
3809 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2),
3810 Ops, "vtbx2");
3811 case NEON::BI__builtin_neon_vtbx3_v:
3812 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3),
3813 Ops, "vtbx3");
3814 case NEON::BI__builtin_neon_vtbx4_v:
3815 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4),
3816 Ops, "vtbx4");
3817 }
3818 }
3819
EmitAArch64TblBuiltinExpr(CodeGenFunction & CGF,unsigned BuiltinID,const CallExpr * E,SmallVectorImpl<Value * > & Ops)3820 static Value *EmitAArch64TblBuiltinExpr(CodeGenFunction &CGF, unsigned BuiltinID,
3821 const CallExpr *E,
3822 SmallVectorImpl<Value *> &Ops) {
3823 unsigned int Int = 0;
3824 const char *s = nullptr;
3825
3826 switch (BuiltinID) {
3827 default:
3828 return nullptr;
3829 case NEON::BI__builtin_neon_vtbl1_v:
3830 case NEON::BI__builtin_neon_vqtbl1_v:
3831 case NEON::BI__builtin_neon_vqtbl1q_v:
3832 case NEON::BI__builtin_neon_vtbl2_v:
3833 case NEON::BI__builtin_neon_vqtbl2_v:
3834 case NEON::BI__builtin_neon_vqtbl2q_v:
3835 case NEON::BI__builtin_neon_vtbl3_v:
3836 case NEON::BI__builtin_neon_vqtbl3_v:
3837 case NEON::BI__builtin_neon_vqtbl3q_v:
3838 case NEON::BI__builtin_neon_vtbl4_v:
3839 case NEON::BI__builtin_neon_vqtbl4_v:
3840 case NEON::BI__builtin_neon_vqtbl4q_v:
3841 break;
3842 case NEON::BI__builtin_neon_vtbx1_v:
3843 case NEON::BI__builtin_neon_vqtbx1_v:
3844 case NEON::BI__builtin_neon_vqtbx1q_v:
3845 case NEON::BI__builtin_neon_vtbx2_v:
3846 case NEON::BI__builtin_neon_vqtbx2_v:
3847 case NEON::BI__builtin_neon_vqtbx2q_v:
3848 case NEON::BI__builtin_neon_vtbx3_v:
3849 case NEON::BI__builtin_neon_vqtbx3_v:
3850 case NEON::BI__builtin_neon_vqtbx3q_v:
3851 case NEON::BI__builtin_neon_vtbx4_v:
3852 case NEON::BI__builtin_neon_vqtbx4_v:
3853 case NEON::BI__builtin_neon_vqtbx4q_v:
3854 break;
3855 }
3856
3857 assert(E->getNumArgs() >= 3);
3858
3859 // Get the last argument, which specifies the vector type.
3860 llvm::APSInt Result;
3861 const Expr *Arg = E->getArg(E->getNumArgs() - 1);
3862 if (!Arg->isIntegerConstantExpr(Result, CGF.getContext()))
3863 return nullptr;
3864
3865 // Determine the type of this overloaded NEON intrinsic.
3866 NeonTypeFlags Type(Result.getZExtValue());
3867 llvm::VectorType *VTy = GetNeonType(&CGF, Type);
3868 llvm::Type *Ty = VTy;
3869 if (!Ty)
3870 return nullptr;
3871
3872 unsigned nElts = VTy->getNumElements();
3873
3874 CodeGen::CGBuilderTy &Builder = CGF.Builder;
3875
3876 // AArch64 scalar builtins are not overloaded, they do not have an extra
3877 // argument that specifies the vector type, need to handle each case.
3878 SmallVector<Value *, 2> TblOps;
3879 switch (BuiltinID) {
3880 case NEON::BI__builtin_neon_vtbl1_v: {
3881 TblOps.push_back(Ops[0]);
3882 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[1], Ty,
3883 Intrinsic::aarch64_neon_tbl1, "vtbl1");
3884 }
3885 case NEON::BI__builtin_neon_vtbl2_v: {
3886 TblOps.push_back(Ops[0]);
3887 TblOps.push_back(Ops[1]);
3888 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[2], Ty,
3889 Intrinsic::aarch64_neon_tbl1, "vtbl1");
3890 }
3891 case NEON::BI__builtin_neon_vtbl3_v: {
3892 TblOps.push_back(Ops[0]);
3893 TblOps.push_back(Ops[1]);
3894 TblOps.push_back(Ops[2]);
3895 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[3], Ty,
3896 Intrinsic::aarch64_neon_tbl2, "vtbl2");
3897 }
3898 case NEON::BI__builtin_neon_vtbl4_v: {
3899 TblOps.push_back(Ops[0]);
3900 TblOps.push_back(Ops[1]);
3901 TblOps.push_back(Ops[2]);
3902 TblOps.push_back(Ops[3]);
3903 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[4], Ty,
3904 Intrinsic::aarch64_neon_tbl2, "vtbl2");
3905 }
3906 case NEON::BI__builtin_neon_vtbx1_v: {
3907 TblOps.push_back(Ops[1]);
3908 Value *TblRes = packTBLDVectorList(CGF, TblOps, nullptr, Ops[2], Ty,
3909 Intrinsic::aarch64_neon_tbl1, "vtbl1");
3910
3911 llvm::Constant *Eight = ConstantInt::get(VTy->getElementType(), 8);
3912 Value* EightV = llvm::ConstantVector::getSplat(nElts, Eight);
3913 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[2], EightV);
3914 CmpRes = Builder.CreateSExt(CmpRes, Ty);
3915
3916 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]);
3917 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes);
3918 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx");
3919 }
3920 case NEON::BI__builtin_neon_vtbx2_v: {
3921 TblOps.push_back(Ops[1]);
3922 TblOps.push_back(Ops[2]);
3923 return packTBLDVectorList(CGF, TblOps, Ops[0], Ops[3], Ty,
3924 Intrinsic::aarch64_neon_tbx1, "vtbx1");
3925 }
3926 case NEON::BI__builtin_neon_vtbx3_v: {
3927 TblOps.push_back(Ops[1]);
3928 TblOps.push_back(Ops[2]);
3929 TblOps.push_back(Ops[3]);
3930 Value *TblRes = packTBLDVectorList(CGF, TblOps, nullptr, Ops[4], Ty,
3931 Intrinsic::aarch64_neon_tbl2, "vtbl2");
3932
3933 llvm::Constant *TwentyFour = ConstantInt::get(VTy->getElementType(), 24);
3934 Value* TwentyFourV = llvm::ConstantVector::getSplat(nElts, TwentyFour);
3935 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[4],
3936 TwentyFourV);
3937 CmpRes = Builder.CreateSExt(CmpRes, Ty);
3938
3939 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]);
3940 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes);
3941 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx");
3942 }
3943 case NEON::BI__builtin_neon_vtbx4_v: {
3944 TblOps.push_back(Ops[1]);
3945 TblOps.push_back(Ops[2]);
3946 TblOps.push_back(Ops[3]);
3947 TblOps.push_back(Ops[4]);
3948 return packTBLDVectorList(CGF, TblOps, Ops[0], Ops[5], Ty,
3949 Intrinsic::aarch64_neon_tbx2, "vtbx2");
3950 }
3951 case NEON::BI__builtin_neon_vqtbl1_v:
3952 case NEON::BI__builtin_neon_vqtbl1q_v:
3953 Int = Intrinsic::aarch64_neon_tbl1; s = "vtbl1"; break;
3954 case NEON::BI__builtin_neon_vqtbl2_v:
3955 case NEON::BI__builtin_neon_vqtbl2q_v: {
3956 Int = Intrinsic::aarch64_neon_tbl2; s = "vtbl2"; break;
3957 case NEON::BI__builtin_neon_vqtbl3_v:
3958 case NEON::BI__builtin_neon_vqtbl3q_v:
3959 Int = Intrinsic::aarch64_neon_tbl3; s = "vtbl3"; break;
3960 case NEON::BI__builtin_neon_vqtbl4_v:
3961 case NEON::BI__builtin_neon_vqtbl4q_v:
3962 Int = Intrinsic::aarch64_neon_tbl4; s = "vtbl4"; break;
3963 case NEON::BI__builtin_neon_vqtbx1_v:
3964 case NEON::BI__builtin_neon_vqtbx1q_v:
3965 Int = Intrinsic::aarch64_neon_tbx1; s = "vtbx1"; break;
3966 case NEON::BI__builtin_neon_vqtbx2_v:
3967 case NEON::BI__builtin_neon_vqtbx2q_v:
3968 Int = Intrinsic::aarch64_neon_tbx2; s = "vtbx2"; break;
3969 case NEON::BI__builtin_neon_vqtbx3_v:
3970 case NEON::BI__builtin_neon_vqtbx3q_v:
3971 Int = Intrinsic::aarch64_neon_tbx3; s = "vtbx3"; break;
3972 case NEON::BI__builtin_neon_vqtbx4_v:
3973 case NEON::BI__builtin_neon_vqtbx4q_v:
3974 Int = Intrinsic::aarch64_neon_tbx4; s = "vtbx4"; break;
3975 }
3976 }
3977
3978 if (!Int)
3979 return nullptr;
3980
3981 Function *F = CGF.CGM.getIntrinsic(Int, Ty);
3982 return CGF.EmitNeonCall(F, Ops, s);
3983 }
3984
vectorWrapScalar16(Value * Op)3985 Value *CodeGenFunction::vectorWrapScalar16(Value *Op) {
3986 llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4);
3987 Op = Builder.CreateBitCast(Op, Int16Ty);
3988 Value *V = UndefValue::get(VTy);
3989 llvm::Constant *CI = ConstantInt::get(SizeTy, 0);
3990 Op = Builder.CreateInsertElement(V, Op, CI);
3991 return Op;
3992 }
3993
vectorWrapScalar8(Value * Op)3994 Value *CodeGenFunction::vectorWrapScalar8(Value *Op) {
3995 llvm::Type *VTy = llvm::VectorType::get(Int8Ty, 8);
3996 Op = Builder.CreateBitCast(Op, Int8Ty);
3997 Value *V = UndefValue::get(VTy);
3998 llvm::Constant *CI = ConstantInt::get(SizeTy, 0);
3999 Op = Builder.CreateInsertElement(V, Op, CI);
4000 return Op;
4001 }
4002
4003 Value *CodeGenFunction::
emitVectorWrappedScalar8Intrinsic(unsigned Int,SmallVectorImpl<Value * > & Ops,const char * Name)4004 emitVectorWrappedScalar8Intrinsic(unsigned Int, SmallVectorImpl<Value*> &Ops,
4005 const char *Name) {
4006 // i8 is not a legal types for AArch64, so we can't just use
4007 // a normal overloaded intrinsic call for these scalar types. Instead
4008 // we'll build 64-bit vectors w/ lane zero being our input values and
4009 // perform the operation on that. The back end can pattern match directly
4010 // to the scalar instruction.
4011 Ops[0] = vectorWrapScalar8(Ops[0]);
4012 Ops[1] = vectorWrapScalar8(Ops[1]);
4013 llvm::Type *VTy = llvm::VectorType::get(Int8Ty, 8);
4014 Value *V = EmitNeonCall(CGM.getIntrinsic(Int, VTy), Ops, Name);
4015 Constant *CI = ConstantInt::get(SizeTy, 0);
4016 return Builder.CreateExtractElement(V, CI, "lane0");
4017 }
4018
4019 Value *CodeGenFunction::
emitVectorWrappedScalar16Intrinsic(unsigned Int,SmallVectorImpl<Value * > & Ops,const char * Name)4020 emitVectorWrappedScalar16Intrinsic(unsigned Int, SmallVectorImpl<Value*> &Ops,
4021 const char *Name) {
4022 // i16 is not a legal types for AArch64, so we can't just use
4023 // a normal overloaded intrinsic call for these scalar types. Instead
4024 // we'll build 64-bit vectors w/ lane zero being our input values and
4025 // perform the operation on that. The back end can pattern match directly
4026 // to the scalar instruction.
4027 Ops[0] = vectorWrapScalar16(Ops[0]);
4028 Ops[1] = vectorWrapScalar16(Ops[1]);
4029 llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4);
4030 Value *V = EmitNeonCall(CGM.getIntrinsic(Int, VTy), Ops, Name);
4031 Constant *CI = ConstantInt::get(SizeTy, 0);
4032 return Builder.CreateExtractElement(V, CI, "lane0");
4033 }
4034
EmitAArch64BuiltinExpr(unsigned BuiltinID,const CallExpr * E)4035 Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID,
4036 const CallExpr *E) {
4037 unsigned HintID = static_cast<unsigned>(-1);
4038 switch (BuiltinID) {
4039 default: break;
4040 case AArch64::BI__builtin_arm_nop:
4041 HintID = 0;
4042 break;
4043 case AArch64::BI__builtin_arm_yield:
4044 HintID = 1;
4045 break;
4046 case AArch64::BI__builtin_arm_wfe:
4047 HintID = 2;
4048 break;
4049 case AArch64::BI__builtin_arm_wfi:
4050 HintID = 3;
4051 break;
4052 case AArch64::BI__builtin_arm_sev:
4053 HintID = 4;
4054 break;
4055 case AArch64::BI__builtin_arm_sevl:
4056 HintID = 5;
4057 break;
4058 }
4059
4060 if (HintID != static_cast<unsigned>(-1)) {
4061 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_hint);
4062 return Builder.CreateCall(F, llvm::ConstantInt::get(Int32Ty, HintID));
4063 }
4064
4065 if (BuiltinID == AArch64::BI__builtin_arm_prefetch) {
4066 Value *Address = EmitScalarExpr(E->getArg(0));
4067 Value *RW = EmitScalarExpr(E->getArg(1));
4068 Value *CacheLevel = EmitScalarExpr(E->getArg(2));
4069 Value *RetentionPolicy = EmitScalarExpr(E->getArg(3));
4070 Value *IsData = EmitScalarExpr(E->getArg(4));
4071
4072 Value *Locality = nullptr;
4073 if (cast<llvm::ConstantInt>(RetentionPolicy)->isZero()) {
4074 // Temporal fetch, needs to convert cache level to locality.
4075 Locality = llvm::ConstantInt::get(Int32Ty,
4076 -cast<llvm::ConstantInt>(CacheLevel)->getValue() + 3);
4077 } else {
4078 // Streaming fetch.
4079 Locality = llvm::ConstantInt::get(Int32Ty, 0);
4080 }
4081
4082 // FIXME: We need AArch64 specific LLVM intrinsic if we want to specify
4083 // PLDL3STRM or PLDL2STRM.
4084 Value *F = CGM.getIntrinsic(Intrinsic::prefetch);
4085 return Builder.CreateCall4(F, Address, RW, Locality, IsData);
4086 }
4087
4088 if (BuiltinID == AArch64::BI__builtin_arm_rbit) {
4089 assert((getContext().getTypeSize(E->getType()) == 32) &&
4090 "rbit of unusual size!");
4091 llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
4092 return Builder.CreateCall(
4093 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit");
4094 }
4095 if (BuiltinID == AArch64::BI__builtin_arm_rbit64) {
4096 assert((getContext().getTypeSize(E->getType()) == 64) &&
4097 "rbit of unusual size!");
4098 llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
4099 return Builder.CreateCall(
4100 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit");
4101 }
4102
4103 if (BuiltinID == AArch64::BI__clear_cache) {
4104 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments");
4105 const FunctionDecl *FD = E->getDirectCallee();
4106 SmallVector<Value*, 2> Ops;
4107 for (unsigned i = 0; i < 2; i++)
4108 Ops.push_back(EmitScalarExpr(E->getArg(i)));
4109 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType());
4110 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty);
4111 StringRef Name = FD->getName();
4112 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops);
4113 }
4114
4115 if ((BuiltinID == AArch64::BI__builtin_arm_ldrex ||
4116 BuiltinID == AArch64::BI__builtin_arm_ldaex) &&
4117 getContext().getTypeSize(E->getType()) == 128) {
4118 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex
4119 ? Intrinsic::aarch64_ldaxp
4120 : Intrinsic::aarch64_ldxp);
4121
4122 Value *LdPtr = EmitScalarExpr(E->getArg(0));
4123 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy),
4124 "ldxp");
4125
4126 Value *Val0 = Builder.CreateExtractValue(Val, 1);
4127 Value *Val1 = Builder.CreateExtractValue(Val, 0);
4128 llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128);
4129 Val0 = Builder.CreateZExt(Val0, Int128Ty);
4130 Val1 = Builder.CreateZExt(Val1, Int128Ty);
4131
4132 Value *ShiftCst = llvm::ConstantInt::get(Int128Ty, 64);
4133 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */);
4134 Val = Builder.CreateOr(Val, Val1);
4135 return Builder.CreateBitCast(Val, ConvertType(E->getType()));
4136 } else if (BuiltinID == AArch64::BI__builtin_arm_ldrex ||
4137 BuiltinID == AArch64::BI__builtin_arm_ldaex) {
4138 Value *LoadAddr = EmitScalarExpr(E->getArg(0));
4139
4140 QualType Ty = E->getType();
4141 llvm::Type *RealResTy = ConvertType(Ty);
4142 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(),
4143 getContext().getTypeSize(Ty));
4144 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo());
4145
4146 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex
4147 ? Intrinsic::aarch64_ldaxr
4148 : Intrinsic::aarch64_ldxr,
4149 LoadAddr->getType());
4150 Value *Val = Builder.CreateCall(F, LoadAddr, "ldxr");
4151
4152 if (RealResTy->isPointerTy())
4153 return Builder.CreateIntToPtr(Val, RealResTy);
4154
4155 Val = Builder.CreateTruncOrBitCast(Val, IntResTy);
4156 return Builder.CreateBitCast(Val, RealResTy);
4157 }
4158
4159 if ((BuiltinID == AArch64::BI__builtin_arm_strex ||
4160 BuiltinID == AArch64::BI__builtin_arm_stlex) &&
4161 getContext().getTypeSize(E->getArg(0)->getType()) == 128) {
4162 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex
4163 ? Intrinsic::aarch64_stlxp
4164 : Intrinsic::aarch64_stxp);
4165 llvm::Type *STy = llvm::StructType::get(Int64Ty, Int64Ty, nullptr);
4166
4167 Value *One = llvm::ConstantInt::get(Int32Ty, 1);
4168 Value *Tmp = Builder.CreateAlloca(ConvertType(E->getArg(0)->getType()),
4169 One);
4170 Value *Val = EmitScalarExpr(E->getArg(0));
4171 Builder.CreateStore(Val, Tmp);
4172
4173 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy));
4174 Val = Builder.CreateLoad(LdPtr);
4175
4176 Value *Arg0 = Builder.CreateExtractValue(Val, 0);
4177 Value *Arg1 = Builder.CreateExtractValue(Val, 1);
4178 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)),
4179 Int8PtrTy);
4180 return Builder.CreateCall3(F, Arg0, Arg1, StPtr, "stxp");
4181 } else if (BuiltinID == AArch64::BI__builtin_arm_strex ||
4182 BuiltinID == AArch64::BI__builtin_arm_stlex) {
4183 Value *StoreVal = EmitScalarExpr(E->getArg(0));
4184 Value *StoreAddr = EmitScalarExpr(E->getArg(1));
4185
4186 QualType Ty = E->getArg(0)->getType();
4187 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(),
4188 getContext().getTypeSize(Ty));
4189 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo());
4190
4191 if (StoreVal->getType()->isPointerTy())
4192 StoreVal = Builder.CreatePtrToInt(StoreVal, Int64Ty);
4193 else {
4194 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy);
4195 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int64Ty);
4196 }
4197
4198 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex
4199 ? Intrinsic::aarch64_stlxr
4200 : Intrinsic::aarch64_stxr,
4201 StoreAddr->getType());
4202 return Builder.CreateCall2(F, StoreVal, StoreAddr, "stxr");
4203 }
4204
4205 if (BuiltinID == AArch64::BI__builtin_arm_clrex) {
4206 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_clrex);
4207 return Builder.CreateCall(F);
4208 }
4209
4210 // CRC32
4211 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic;
4212 switch (BuiltinID) {
4213 case AArch64::BI__builtin_arm_crc32b:
4214 CRCIntrinsicID = Intrinsic::aarch64_crc32b; break;
4215 case AArch64::BI__builtin_arm_crc32cb:
4216 CRCIntrinsicID = Intrinsic::aarch64_crc32cb; break;
4217 case AArch64::BI__builtin_arm_crc32h:
4218 CRCIntrinsicID = Intrinsic::aarch64_crc32h; break;
4219 case AArch64::BI__builtin_arm_crc32ch:
4220 CRCIntrinsicID = Intrinsic::aarch64_crc32ch; break;
4221 case AArch64::BI__builtin_arm_crc32w:
4222 CRCIntrinsicID = Intrinsic::aarch64_crc32w; break;
4223 case AArch64::BI__builtin_arm_crc32cw:
4224 CRCIntrinsicID = Intrinsic::aarch64_crc32cw; break;
4225 case AArch64::BI__builtin_arm_crc32d:
4226 CRCIntrinsicID = Intrinsic::aarch64_crc32x; break;
4227 case AArch64::BI__builtin_arm_crc32cd:
4228 CRCIntrinsicID = Intrinsic::aarch64_crc32cx; break;
4229 }
4230
4231 if (CRCIntrinsicID != Intrinsic::not_intrinsic) {
4232 Value *Arg0 = EmitScalarExpr(E->getArg(0));
4233 Value *Arg1 = EmitScalarExpr(E->getArg(1));
4234 Function *F = CGM.getIntrinsic(CRCIntrinsicID);
4235
4236 llvm::Type *DataTy = F->getFunctionType()->getParamType(1);
4237 Arg1 = Builder.CreateZExtOrBitCast(Arg1, DataTy);
4238
4239 return Builder.CreateCall2(F, Arg0, Arg1);
4240 }
4241
4242 llvm::SmallVector<Value*, 4> Ops;
4243 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++)
4244 Ops.push_back(EmitScalarExpr(E->getArg(i)));
4245
4246 auto SISDMap = makeArrayRef(AArch64SISDIntrinsicMap);
4247 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap(
4248 SISDMap, BuiltinID, AArch64SISDIntrinsicsProvenSorted);
4249
4250 if (Builtin) {
4251 Ops.push_back(EmitScalarExpr(E->getArg(E->getNumArgs() - 1)));
4252 Value *Result = EmitCommonNeonSISDBuiltinExpr(*this, *Builtin, Ops, E);
4253 assert(Result && "SISD intrinsic should have been handled");
4254 return Result;
4255 }
4256
4257 llvm::APSInt Result;
4258 const Expr *Arg = E->getArg(E->getNumArgs()-1);
4259 NeonTypeFlags Type(0);
4260 if (Arg->isIntegerConstantExpr(Result, getContext()))
4261 // Determine the type of this overloaded NEON intrinsic.
4262 Type = NeonTypeFlags(Result.getZExtValue());
4263
4264 bool usgn = Type.isUnsigned();
4265 bool quad = Type.isQuad();
4266
4267 // Handle non-overloaded intrinsics first.
4268 switch (BuiltinID) {
4269 default: break;
4270 case NEON::BI__builtin_neon_vldrq_p128: {
4271 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128);
4272 Value *Ptr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int128PTy);
4273 return Builder.CreateLoad(Ptr);
4274 }
4275 case NEON::BI__builtin_neon_vstrq_p128: {
4276 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128);
4277 Value *Ptr = Builder.CreateBitCast(Ops[0], Int128PTy);
4278 return Builder.CreateStore(EmitScalarExpr(E->getArg(1)), Ptr);
4279 }
4280 case NEON::BI__builtin_neon_vcvts_u32_f32:
4281 case NEON::BI__builtin_neon_vcvtd_u64_f64:
4282 usgn = true;
4283 // FALL THROUGH
4284 case NEON::BI__builtin_neon_vcvts_s32_f32:
4285 case NEON::BI__builtin_neon_vcvtd_s64_f64: {
4286 Ops.push_back(EmitScalarExpr(E->getArg(0)));
4287 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64;
4288 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty;
4289 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy;
4290 Ops[0] = Builder.CreateBitCast(Ops[0], FTy);
4291 if (usgn)
4292 return Builder.CreateFPToUI(Ops[0], InTy);
4293 return Builder.CreateFPToSI(Ops[0], InTy);
4294 }
4295 case NEON::BI__builtin_neon_vcvts_f32_u32:
4296 case NEON::BI__builtin_neon_vcvtd_f64_u64:
4297 usgn = true;
4298 // FALL THROUGH
4299 case NEON::BI__builtin_neon_vcvts_f32_s32:
4300 case NEON::BI__builtin_neon_vcvtd_f64_s64: {
4301 Ops.push_back(EmitScalarExpr(E->getArg(0)));
4302 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64;
4303 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty;
4304 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy;
4305 Ops[0] = Builder.CreateBitCast(Ops[0], InTy);
4306 if (usgn)
4307 return Builder.CreateUIToFP(Ops[0], FTy);
4308 return Builder.CreateSIToFP(Ops[0], FTy);
4309 }
4310 case NEON::BI__builtin_neon_vpaddd_s64: {
4311 llvm::Type *Ty =
4312 llvm::VectorType::get(llvm::Type::getInt64Ty(getLLVMContext()), 2);
4313 Value *Vec = EmitScalarExpr(E->getArg(0));
4314 // The vector is v2f64, so make sure it's bitcast to that.
4315 Vec = Builder.CreateBitCast(Vec, Ty, "v2i64");
4316 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0);
4317 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1);
4318 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0");
4319 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1");
4320 // Pairwise addition of a v2f64 into a scalar f64.
4321 return Builder.CreateAdd(Op0, Op1, "vpaddd");
4322 }
4323 case NEON::BI__builtin_neon_vpaddd_f64: {
4324 llvm::Type *Ty =
4325 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2);
4326 Value *Vec = EmitScalarExpr(E->getArg(0));
4327 // The vector is v2f64, so make sure it's bitcast to that.
4328 Vec = Builder.CreateBitCast(Vec, Ty, "v2f64");
4329 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0);
4330 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1);
4331 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0");
4332 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1");
4333 // Pairwise addition of a v2f64 into a scalar f64.
4334 return Builder.CreateFAdd(Op0, Op1, "vpaddd");
4335 }
4336 case NEON::BI__builtin_neon_vpadds_f32: {
4337 llvm::Type *Ty =
4338 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2);
4339 Value *Vec = EmitScalarExpr(E->getArg(0));
4340 // The vector is v2f32, so make sure it's bitcast to that.
4341 Vec = Builder.CreateBitCast(Vec, Ty, "v2f32");
4342 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0);
4343 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1);
4344 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0");
4345 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1");
4346 // Pairwise addition of a v2f32 into a scalar f32.
4347 return Builder.CreateFAdd(Op0, Op1, "vpaddd");
4348 }
4349 case NEON::BI__builtin_neon_vceqzd_s64:
4350 case NEON::BI__builtin_neon_vceqzd_f64:
4351 case NEON::BI__builtin_neon_vceqzs_f32:
4352 Ops.push_back(EmitScalarExpr(E->getArg(0)));
4353 return EmitAArch64CompareBuiltinExpr(
4354 Ops[0], ConvertType(E->getCallReturnType(getContext())),
4355 ICmpInst::FCMP_OEQ, ICmpInst::ICMP_EQ, "vceqz");
4356 case NEON::BI__builtin_neon_vcgezd_s64:
4357 case NEON::BI__builtin_neon_vcgezd_f64:
4358 case NEON::BI__builtin_neon_vcgezs_f32:
4359 Ops.push_back(EmitScalarExpr(E->getArg(0)));
4360 return EmitAArch64CompareBuiltinExpr(
4361 Ops[0], ConvertType(E->getCallReturnType(getContext())),
4362 ICmpInst::FCMP_OGE, ICmpInst::ICMP_SGE, "vcgez");
4363 case NEON::BI__builtin_neon_vclezd_s64:
4364 case NEON::BI__builtin_neon_vclezd_f64:
4365 case NEON::BI__builtin_neon_vclezs_f32:
4366 Ops.push_back(EmitScalarExpr(E->getArg(0)));
4367 return EmitAArch64CompareBuiltinExpr(
4368 Ops[0], ConvertType(E->getCallReturnType(getContext())),
4369 ICmpInst::FCMP_OLE, ICmpInst::ICMP_SLE, "vclez");
4370 case NEON::BI__builtin_neon_vcgtzd_s64:
4371 case NEON::BI__builtin_neon_vcgtzd_f64:
4372 case NEON::BI__builtin_neon_vcgtzs_f32:
4373 Ops.push_back(EmitScalarExpr(E->getArg(0)));
4374 return EmitAArch64CompareBuiltinExpr(
4375 Ops[0], ConvertType(E->getCallReturnType(getContext())),
4376 ICmpInst::FCMP_OGT, ICmpInst::ICMP_SGT, "vcgtz");
4377 case NEON::BI__builtin_neon_vcltzd_s64:
4378 case NEON::BI__builtin_neon_vcltzd_f64:
4379 case NEON::BI__builtin_neon_vcltzs_f32:
4380 Ops.push_back(EmitScalarExpr(E->getArg(0)));
4381 return EmitAArch64CompareBuiltinExpr(
4382 Ops[0], ConvertType(E->getCallReturnType(getContext())),
4383 ICmpInst::FCMP_OLT, ICmpInst::ICMP_SLT, "vcltz");
4384
4385 case NEON::BI__builtin_neon_vceqzd_u64: {
4386 llvm::Type *Ty = llvm::Type::getInt64Ty(getLLVMContext());
4387 Ops.push_back(EmitScalarExpr(E->getArg(0)));
4388 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
4389 Ops[0] = Builder.CreateICmp(llvm::ICmpInst::ICMP_EQ, Ops[0],
4390 llvm::Constant::getNullValue(Ty));
4391 return Builder.CreateSExt(Ops[0], Ty, "vceqzd");
4392 }
4393 case NEON::BI__builtin_neon_vceqd_f64:
4394 case NEON::BI__builtin_neon_vcled_f64:
4395 case NEON::BI__builtin_neon_vcltd_f64:
4396 case NEON::BI__builtin_neon_vcged_f64:
4397 case NEON::BI__builtin_neon_vcgtd_f64: {
4398 llvm::CmpInst::Predicate P;
4399 switch (BuiltinID) {
4400 default: llvm_unreachable("missing builtin ID in switch!");
4401 case NEON::BI__builtin_neon_vceqd_f64: P = llvm::FCmpInst::FCMP_OEQ; break;
4402 case NEON::BI__builtin_neon_vcled_f64: P = llvm::FCmpInst::FCMP_OLE; break;
4403 case NEON::BI__builtin_neon_vcltd_f64: P = llvm::FCmpInst::FCMP_OLT; break;
4404 case NEON::BI__builtin_neon_vcged_f64: P = llvm::FCmpInst::FCMP_OGE; break;
4405 case NEON::BI__builtin_neon_vcgtd_f64: P = llvm::FCmpInst::FCMP_OGT; break;
4406 }
4407 Ops.push_back(EmitScalarExpr(E->getArg(1)));
4408 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy);
4409 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy);
4410 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]);
4411 return Builder.CreateSExt(Ops[0], Int64Ty, "vcmpd");
4412 }
4413 case NEON::BI__builtin_neon_vceqs_f32:
4414 case NEON::BI__builtin_neon_vcles_f32:
4415 case NEON::BI__builtin_neon_vclts_f32:
4416 case NEON::BI__builtin_neon_vcges_f32:
4417 case NEON::BI__builtin_neon_vcgts_f32: {
4418 llvm::CmpInst::Predicate P;
4419 switch (BuiltinID) {
4420 default: llvm_unreachable("missing builtin ID in switch!");
4421 case NEON::BI__builtin_neon_vceqs_f32: P = llvm::FCmpInst::FCMP_OEQ; break;
4422 case NEON::BI__builtin_neon_vcles_f32: P = llvm::FCmpInst::FCMP_OLE; break;
4423 case NEON::BI__builtin_neon_vclts_f32: P = llvm::FCmpInst::FCMP_OLT; break;
4424 case NEON::BI__builtin_neon_vcges_f32: P = llvm::FCmpInst::FCMP_OGE; break;
4425 case NEON::BI__builtin_neon_vcgts_f32: P = llvm::FCmpInst::FCMP_OGT; break;
4426 }
4427 Ops.push_back(EmitScalarExpr(E->getArg(1)));
4428 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy);
4429 Ops[1] = Builder.CreateBitCast(Ops[1], FloatTy);
4430 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]);
4431 return Builder.CreateSExt(Ops[0], Int32Ty, "vcmpd");
4432 }
4433 case NEON::BI__builtin_neon_vceqd_s64:
4434 case NEON::BI__builtin_neon_vceqd_u64:
4435 case NEON::BI__builtin_neon_vcgtd_s64:
4436 case NEON::BI__builtin_neon_vcgtd_u64:
4437 case NEON::BI__builtin_neon_vcltd_s64:
4438 case NEON::BI__builtin_neon_vcltd_u64:
4439 case NEON::BI__builtin_neon_vcged_u64:
4440 case NEON::BI__builtin_neon_vcged_s64:
4441 case NEON::BI__builtin_neon_vcled_u64:
4442 case NEON::BI__builtin_neon_vcled_s64: {
4443 llvm::CmpInst::Predicate P;
4444 switch (BuiltinID) {
4445 default: llvm_unreachable("missing builtin ID in switch!");
4446 case NEON::BI__builtin_neon_vceqd_s64:
4447 case NEON::BI__builtin_neon_vceqd_u64:P = llvm::ICmpInst::ICMP_EQ;break;
4448 case NEON::BI__builtin_neon_vcgtd_s64:P = llvm::ICmpInst::ICMP_SGT;break;
4449 case NEON::BI__builtin_neon_vcgtd_u64:P = llvm::ICmpInst::ICMP_UGT;break;
4450 case NEON::BI__builtin_neon_vcltd_s64:P = llvm::ICmpInst::ICMP_SLT;break;
4451 case NEON::BI__builtin_neon_vcltd_u64:P = llvm::ICmpInst::ICMP_ULT;break;
4452 case NEON::BI__builtin_neon_vcged_u64:P = llvm::ICmpInst::ICMP_UGE;break;
4453 case NEON::BI__builtin_neon_vcged_s64:P = llvm::ICmpInst::ICMP_SGE;break;
4454 case NEON::BI__builtin_neon_vcled_u64:P = llvm::ICmpInst::ICMP_ULE;break;
4455 case NEON::BI__builtin_neon_vcled_s64:P = llvm::ICmpInst::ICMP_SLE;break;
4456 }
4457 Ops.push_back(EmitScalarExpr(E->getArg(1)));
4458 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty);
4459 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty);
4460 Ops[0] = Builder.CreateICmp(P, Ops[0], Ops[1]);
4461 return Builder.CreateSExt(Ops[0], Int64Ty, "vceqd");
4462 }
4463 case NEON::BI__builtin_neon_vtstd_s64:
4464 case NEON::BI__builtin_neon_vtstd_u64: {
4465 llvm::Type *Ty = llvm::Type::getInt64Ty(getLLVMContext());
4466 Ops.push_back(EmitScalarExpr(E->getArg(1)));
4467 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
4468 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
4469 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]);
4470 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0],
4471 llvm::Constant::getNullValue(Ty));
4472 return Builder.CreateSExt(Ops[0], Ty, "vtstd");
4473 }
4474 case NEON::BI__builtin_neon_vset_lane_i8:
4475 case NEON::BI__builtin_neon_vset_lane_i16:
4476 case NEON::BI__builtin_neon_vset_lane_i32:
4477 case NEON::BI__builtin_neon_vset_lane_i64:
4478 case NEON::BI__builtin_neon_vset_lane_f32:
4479 case NEON::BI__builtin_neon_vsetq_lane_i8:
4480 case NEON::BI__builtin_neon_vsetq_lane_i16:
4481 case NEON::BI__builtin_neon_vsetq_lane_i32:
4482 case NEON::BI__builtin_neon_vsetq_lane_i64:
4483 case NEON::BI__builtin_neon_vsetq_lane_f32:
4484 Ops.push_back(EmitScalarExpr(E->getArg(2)));
4485 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane");
4486 case NEON::BI__builtin_neon_vset_lane_f64:
4487 // The vector type needs a cast for the v1f64 variant.
4488 Ops[1] = Builder.CreateBitCast(Ops[1],
4489 llvm::VectorType::get(DoubleTy, 1));
4490 Ops.push_back(EmitScalarExpr(E->getArg(2)));
4491 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane");
4492 case NEON::BI__builtin_neon_vsetq_lane_f64:
4493 // The vector type needs a cast for the v2f64 variant.
4494 Ops[1] = Builder.CreateBitCast(Ops[1],
4495 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2));
4496 Ops.push_back(EmitScalarExpr(E->getArg(2)));
4497 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane");
4498
4499 case NEON::BI__builtin_neon_vget_lane_i8:
4500 case NEON::BI__builtin_neon_vdupb_lane_i8:
4501 Ops[0] = Builder.CreateBitCast(Ops[0],
4502 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8));
4503 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
4504 "vget_lane");
4505 case NEON::BI__builtin_neon_vgetq_lane_i8:
4506 case NEON::BI__builtin_neon_vdupb_laneq_i8:
4507 Ops[0] = Builder.CreateBitCast(Ops[0],
4508 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16));
4509 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
4510 "vgetq_lane");
4511 case NEON::BI__builtin_neon_vget_lane_i16:
4512 case NEON::BI__builtin_neon_vduph_lane_i16:
4513 Ops[0] = Builder.CreateBitCast(Ops[0],
4514 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4));
4515 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
4516 "vget_lane");
4517 case NEON::BI__builtin_neon_vgetq_lane_i16:
4518 case NEON::BI__builtin_neon_vduph_laneq_i16:
4519 Ops[0] = Builder.CreateBitCast(Ops[0],
4520 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8));
4521 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
4522 "vgetq_lane");
4523 case NEON::BI__builtin_neon_vget_lane_i32:
4524 case NEON::BI__builtin_neon_vdups_lane_i32:
4525 Ops[0] = Builder.CreateBitCast(
4526 Ops[0],
4527 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 32), 2));
4528 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
4529 "vget_lane");
4530 case NEON::BI__builtin_neon_vdups_lane_f32:
4531 Ops[0] = Builder.CreateBitCast(Ops[0],
4532 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2));
4533 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
4534 "vdups_lane");
4535 case NEON::BI__builtin_neon_vgetq_lane_i32:
4536 case NEON::BI__builtin_neon_vdups_laneq_i32:
4537 Ops[0] = Builder.CreateBitCast(Ops[0],
4538 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 32), 4));
4539 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
4540 "vgetq_lane");
4541 case NEON::BI__builtin_neon_vget_lane_i64:
4542 case NEON::BI__builtin_neon_vdupd_lane_i64:
4543 Ops[0] = Builder.CreateBitCast(Ops[0],
4544 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 64), 1));
4545 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
4546 "vget_lane");
4547 case NEON::BI__builtin_neon_vdupd_lane_f64:
4548 Ops[0] = Builder.CreateBitCast(Ops[0],
4549 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 1));
4550 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
4551 "vdupd_lane");
4552 case NEON::BI__builtin_neon_vgetq_lane_i64:
4553 case NEON::BI__builtin_neon_vdupd_laneq_i64:
4554 Ops[0] = Builder.CreateBitCast(Ops[0],
4555 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 64), 2));
4556 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
4557 "vgetq_lane");
4558 case NEON::BI__builtin_neon_vget_lane_f32:
4559 Ops[0] = Builder.CreateBitCast(Ops[0],
4560 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2));
4561 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
4562 "vget_lane");
4563 case NEON::BI__builtin_neon_vget_lane_f64:
4564 Ops[0] = Builder.CreateBitCast(Ops[0],
4565 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 1));
4566 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
4567 "vget_lane");
4568 case NEON::BI__builtin_neon_vgetq_lane_f32:
4569 case NEON::BI__builtin_neon_vdups_laneq_f32:
4570 Ops[0] = Builder.CreateBitCast(Ops[0],
4571 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 4));
4572 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
4573 "vgetq_lane");
4574 case NEON::BI__builtin_neon_vgetq_lane_f64:
4575 case NEON::BI__builtin_neon_vdupd_laneq_f64:
4576 Ops[0] = Builder.CreateBitCast(Ops[0],
4577 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2));
4578 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
4579 "vgetq_lane");
4580 case NEON::BI__builtin_neon_vaddd_s64:
4581 case NEON::BI__builtin_neon_vaddd_u64:
4582 return Builder.CreateAdd(Ops[0], EmitScalarExpr(E->getArg(1)), "vaddd");
4583 case NEON::BI__builtin_neon_vsubd_s64:
4584 case NEON::BI__builtin_neon_vsubd_u64:
4585 return Builder.CreateSub(Ops[0], EmitScalarExpr(E->getArg(1)), "vsubd");
4586 case NEON::BI__builtin_neon_vqdmlalh_s16:
4587 case NEON::BI__builtin_neon_vqdmlslh_s16: {
4588 SmallVector<Value *, 2> ProductOps;
4589 ProductOps.push_back(vectorWrapScalar16(Ops[1]));
4590 ProductOps.push_back(vectorWrapScalar16(EmitScalarExpr(E->getArg(2))));
4591 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4);
4592 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy),
4593 ProductOps, "vqdmlXl");
4594 Constant *CI = ConstantInt::get(SizeTy, 0);
4595 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0");
4596
4597 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlalh_s16
4598 ? Intrinsic::aarch64_neon_sqadd
4599 : Intrinsic::aarch64_neon_sqsub;
4600 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int32Ty), Ops, "vqdmlXl");
4601 }
4602 case NEON::BI__builtin_neon_vqshlud_n_s64: {
4603 Ops.push_back(EmitScalarExpr(E->getArg(1)));
4604 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty);
4605 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqshlu, Int64Ty),
4606 Ops, "vqshlu_n");
4607 }
4608 case NEON::BI__builtin_neon_vqshld_n_u64:
4609 case NEON::BI__builtin_neon_vqshld_n_s64: {
4610 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vqshld_n_u64
4611 ? Intrinsic::aarch64_neon_uqshl
4612 : Intrinsic::aarch64_neon_sqshl;
4613 Ops.push_back(EmitScalarExpr(E->getArg(1)));
4614 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty);
4615 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vqshl_n");
4616 }
4617 case NEON::BI__builtin_neon_vrshrd_n_u64:
4618 case NEON::BI__builtin_neon_vrshrd_n_s64: {
4619 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrshrd_n_u64
4620 ? Intrinsic::aarch64_neon_urshl
4621 : Intrinsic::aarch64_neon_srshl;
4622 Ops.push_back(EmitScalarExpr(E->getArg(1)));
4623 int SV = cast<ConstantInt>(Ops[1])->getSExtValue();
4624 Ops[1] = ConstantInt::get(Int64Ty, -SV);
4625 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vrshr_n");
4626 }
4627 case NEON::BI__builtin_neon_vrsrad_n_u64:
4628 case NEON::BI__builtin_neon_vrsrad_n_s64: {
4629 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrsrad_n_u64
4630 ? Intrinsic::aarch64_neon_urshl
4631 : Intrinsic::aarch64_neon_srshl;
4632 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty);
4633 Ops.push_back(Builder.CreateNeg(EmitScalarExpr(E->getArg(2))));
4634 Ops[1] = Builder.CreateCall2(CGM.getIntrinsic(Int, Int64Ty), Ops[1],
4635 Builder.CreateSExt(Ops[2], Int64Ty));
4636 return Builder.CreateAdd(Ops[0], Builder.CreateBitCast(Ops[1], Int64Ty));
4637 }
4638 case NEON::BI__builtin_neon_vshld_n_s64:
4639 case NEON::BI__builtin_neon_vshld_n_u64: {
4640 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1)));
4641 return Builder.CreateShl(
4642 Ops[0], ConstantInt::get(Int64Ty, Amt->getZExtValue()), "shld_n");
4643 }
4644 case NEON::BI__builtin_neon_vshrd_n_s64: {
4645 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1)));
4646 return Builder.CreateAShr(
4647 Ops[0], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63),
4648 Amt->getZExtValue())),
4649 "shrd_n");
4650 }
4651 case NEON::BI__builtin_neon_vshrd_n_u64: {
4652 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1)));
4653 uint64_t ShiftAmt = Amt->getZExtValue();
4654 // Right-shifting an unsigned value by its size yields 0.
4655 if (ShiftAmt == 64)
4656 return ConstantInt::get(Int64Ty, 0);
4657 return Builder.CreateLShr(Ops[0], ConstantInt::get(Int64Ty, ShiftAmt),
4658 "shrd_n");
4659 }
4660 case NEON::BI__builtin_neon_vsrad_n_s64: {
4661 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2)));
4662 Ops[1] = Builder.CreateAShr(
4663 Ops[1], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63),
4664 Amt->getZExtValue())),
4665 "shrd_n");
4666 return Builder.CreateAdd(Ops[0], Ops[1]);
4667 }
4668 case NEON::BI__builtin_neon_vsrad_n_u64: {
4669 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2)));
4670 uint64_t ShiftAmt = Amt->getZExtValue();
4671 // Right-shifting an unsigned value by its size yields 0.
4672 // As Op + 0 = Op, return Ops[0] directly.
4673 if (ShiftAmt == 64)
4674 return Ops[0];
4675 Ops[1] = Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, ShiftAmt),
4676 "shrd_n");
4677 return Builder.CreateAdd(Ops[0], Ops[1]);
4678 }
4679 case NEON::BI__builtin_neon_vqdmlalh_lane_s16:
4680 case NEON::BI__builtin_neon_vqdmlalh_laneq_s16:
4681 case NEON::BI__builtin_neon_vqdmlslh_lane_s16:
4682 case NEON::BI__builtin_neon_vqdmlslh_laneq_s16: {
4683 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)),
4684 "lane");
4685 SmallVector<Value *, 2> ProductOps;
4686 ProductOps.push_back(vectorWrapScalar16(Ops[1]));
4687 ProductOps.push_back(vectorWrapScalar16(Ops[2]));
4688 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4);
4689 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy),
4690 ProductOps, "vqdmlXl");
4691 Constant *CI = ConstantInt::get(SizeTy, 0);
4692 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0");
4693 Ops.pop_back();
4694
4695 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlalh_lane_s16 ||
4696 BuiltinID == NEON::BI__builtin_neon_vqdmlalh_laneq_s16)
4697 ? Intrinsic::aarch64_neon_sqadd
4698 : Intrinsic::aarch64_neon_sqsub;
4699 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int32Ty), Ops, "vqdmlXl");
4700 }
4701 case NEON::BI__builtin_neon_vqdmlals_s32:
4702 case NEON::BI__builtin_neon_vqdmlsls_s32: {
4703 SmallVector<Value *, 2> ProductOps;
4704 ProductOps.push_back(Ops[1]);
4705 ProductOps.push_back(EmitScalarExpr(E->getArg(2)));
4706 Ops[1] =
4707 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar),
4708 ProductOps, "vqdmlXl");
4709
4710 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlals_s32
4711 ? Intrinsic::aarch64_neon_sqadd
4712 : Intrinsic::aarch64_neon_sqsub;
4713 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int64Ty), Ops, "vqdmlXl");
4714 }
4715 case NEON::BI__builtin_neon_vqdmlals_lane_s32:
4716 case NEON::BI__builtin_neon_vqdmlals_laneq_s32:
4717 case NEON::BI__builtin_neon_vqdmlsls_lane_s32:
4718 case NEON::BI__builtin_neon_vqdmlsls_laneq_s32: {
4719 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)),
4720 "lane");
4721 SmallVector<Value *, 2> ProductOps;
4722 ProductOps.push_back(Ops[1]);
4723 ProductOps.push_back(Ops[2]);
4724 Ops[1] =
4725 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar),
4726 ProductOps, "vqdmlXl");
4727 Ops.pop_back();
4728
4729 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlals_lane_s32 ||
4730 BuiltinID == NEON::BI__builtin_neon_vqdmlals_laneq_s32)
4731 ? Intrinsic::aarch64_neon_sqadd
4732 : Intrinsic::aarch64_neon_sqsub;
4733 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int64Ty), Ops, "vqdmlXl");
4734 }
4735 }
4736
4737 llvm::VectorType *VTy = GetNeonType(this, Type);
4738 llvm::Type *Ty = VTy;
4739 if (!Ty)
4740 return nullptr;
4741
4742 // Not all intrinsics handled by the common case work for AArch64 yet, so only
4743 // defer to common code if it's been added to our special map.
4744 Builtin = findNeonIntrinsicInMap(AArch64SIMDIntrinsicMap, BuiltinID,
4745 AArch64SIMDIntrinsicsProvenSorted);
4746
4747 if (Builtin)
4748 return EmitCommonNeonBuiltinExpr(
4749 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic,
4750 Builtin->NameHint, Builtin->TypeModifier, E, Ops, nullptr);
4751
4752 if (Value *V = EmitAArch64TblBuiltinExpr(*this, BuiltinID, E, Ops))
4753 return V;
4754
4755 unsigned Int;
4756 switch (BuiltinID) {
4757 default: return nullptr;
4758 case NEON::BI__builtin_neon_vbsl_v:
4759 case NEON::BI__builtin_neon_vbslq_v: {
4760 llvm::Type *BitTy = llvm::VectorType::getInteger(VTy);
4761 Ops[0] = Builder.CreateBitCast(Ops[0], BitTy, "vbsl");
4762 Ops[1] = Builder.CreateBitCast(Ops[1], BitTy, "vbsl");
4763 Ops[2] = Builder.CreateBitCast(Ops[2], BitTy, "vbsl");
4764
4765 Ops[1] = Builder.CreateAnd(Ops[0], Ops[1], "vbsl");
4766 Ops[2] = Builder.CreateAnd(Builder.CreateNot(Ops[0]), Ops[2], "vbsl");
4767 Ops[0] = Builder.CreateOr(Ops[1], Ops[2], "vbsl");
4768 return Builder.CreateBitCast(Ops[0], Ty);
4769 }
4770 case NEON::BI__builtin_neon_vfma_lane_v:
4771 case NEON::BI__builtin_neon_vfmaq_lane_v: { // Only used for FP types
4772 // The ARM builtins (and instructions) have the addend as the first
4773 // operand, but the 'fma' intrinsics have it last. Swap it around here.
4774 Value *Addend = Ops[0];
4775 Value *Multiplicand = Ops[1];
4776 Value *LaneSource = Ops[2];
4777 Ops[0] = Multiplicand;
4778 Ops[1] = LaneSource;
4779 Ops[2] = Addend;
4780
4781 // Now adjust things to handle the lane access.
4782 llvm::Type *SourceTy = BuiltinID == NEON::BI__builtin_neon_vfmaq_lane_v ?
4783 llvm::VectorType::get(VTy->getElementType(), VTy->getNumElements() / 2) :
4784 VTy;
4785 llvm::Constant *cst = cast<Constant>(Ops[3]);
4786 Value *SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), cst);
4787 Ops[1] = Builder.CreateBitCast(Ops[1], SourceTy);
4788 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV, "lane");
4789
4790 Ops.pop_back();
4791 Int = Intrinsic::fma;
4792 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmla");
4793 }
4794 case NEON::BI__builtin_neon_vfma_laneq_v: {
4795 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty);
4796 // v1f64 fma should be mapped to Neon scalar f64 fma
4797 if (VTy && VTy->getElementType() == DoubleTy) {
4798 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy);
4799 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy);
4800 llvm::Type *VTy = GetNeonType(this,
4801 NeonTypeFlags(NeonTypeFlags::Float64, false, true));
4802 Ops[2] = Builder.CreateBitCast(Ops[2], VTy);
4803 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract");
4804 Value *F = CGM.getIntrinsic(Intrinsic::fma, DoubleTy);
4805 Value *Result = Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]);
4806 return Builder.CreateBitCast(Result, Ty);
4807 }
4808 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty);
4809 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
4810 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
4811
4812 llvm::Type *STy = llvm::VectorType::get(VTy->getElementType(),
4813 VTy->getNumElements() * 2);
4814 Ops[2] = Builder.CreateBitCast(Ops[2], STy);
4815 Value* SV = llvm::ConstantVector::getSplat(VTy->getNumElements(),
4816 cast<ConstantInt>(Ops[3]));
4817 Ops[2] = Builder.CreateShuffleVector(Ops[2], Ops[2], SV, "lane");
4818
4819 return Builder.CreateCall3(F, Ops[2], Ops[1], Ops[0]);
4820 }
4821 case NEON::BI__builtin_neon_vfmaq_laneq_v: {
4822 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty);
4823 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
4824 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
4825
4826 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
4827 Ops[2] = EmitNeonSplat(Ops[2], cast<ConstantInt>(Ops[3]));
4828 return Builder.CreateCall3(F, Ops[2], Ops[1], Ops[0]);
4829 }
4830 case NEON::BI__builtin_neon_vfmas_lane_f32:
4831 case NEON::BI__builtin_neon_vfmas_laneq_f32:
4832 case NEON::BI__builtin_neon_vfmad_lane_f64:
4833 case NEON::BI__builtin_neon_vfmad_laneq_f64: {
4834 Ops.push_back(EmitScalarExpr(E->getArg(3)));
4835 llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext()));
4836 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty);
4837 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract");
4838 return Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]);
4839 }
4840 case NEON::BI__builtin_neon_vfms_v:
4841 case NEON::BI__builtin_neon_vfmsq_v: { // Only used for FP types
4842 // FIXME: probably remove when we no longer support aarch64_simd.h
4843 // (arm_neon.h delegates to vfma).
4844
4845 // The ARM builtins (and instructions) have the addend as the first
4846 // operand, but the 'fma' intrinsics have it last. Swap it around here.
4847 Value *Subtrahend = Ops[0];
4848 Value *Multiplicand = Ops[2];
4849 Ops[0] = Multiplicand;
4850 Ops[2] = Subtrahend;
4851 Ops[1] = Builder.CreateBitCast(Ops[1], VTy);
4852 Ops[1] = Builder.CreateFNeg(Ops[1]);
4853 Int = Intrinsic::fma;
4854 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmls");
4855 }
4856 case NEON::BI__builtin_neon_vmull_v:
4857 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics.
4858 Int = usgn ? Intrinsic::aarch64_neon_umull : Intrinsic::aarch64_neon_smull;
4859 if (Type.isPoly()) Int = Intrinsic::aarch64_neon_pmull;
4860 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull");
4861 case NEON::BI__builtin_neon_vmax_v:
4862 case NEON::BI__builtin_neon_vmaxq_v:
4863 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics.
4864 Int = usgn ? Intrinsic::aarch64_neon_umax : Intrinsic::aarch64_neon_smax;
4865 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmax;
4866 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax");
4867 case NEON::BI__builtin_neon_vmin_v:
4868 case NEON::BI__builtin_neon_vminq_v:
4869 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics.
4870 Int = usgn ? Intrinsic::aarch64_neon_umin : Intrinsic::aarch64_neon_smin;
4871 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmin;
4872 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin");
4873 case NEON::BI__builtin_neon_vabd_v:
4874 case NEON::BI__builtin_neon_vabdq_v:
4875 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics.
4876 Int = usgn ? Intrinsic::aarch64_neon_uabd : Intrinsic::aarch64_neon_sabd;
4877 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fabd;
4878 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd");
4879 case NEON::BI__builtin_neon_vpadal_v:
4880 case NEON::BI__builtin_neon_vpadalq_v: {
4881 unsigned ArgElts = VTy->getNumElements();
4882 llvm::IntegerType *EltTy = cast<IntegerType>(VTy->getElementType());
4883 unsigned BitWidth = EltTy->getBitWidth();
4884 llvm::Type *ArgTy = llvm::VectorType::get(
4885 llvm::IntegerType::get(getLLVMContext(), BitWidth/2), 2*ArgElts);
4886 llvm::Type* Tys[2] = { VTy, ArgTy };
4887 Int = usgn ? Intrinsic::aarch64_neon_uaddlp : Intrinsic::aarch64_neon_saddlp;
4888 SmallVector<llvm::Value*, 1> TmpOps;
4889 TmpOps.push_back(Ops[1]);
4890 Function *F = CGM.getIntrinsic(Int, Tys);
4891 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vpadal");
4892 llvm::Value *addend = Builder.CreateBitCast(Ops[0], tmp->getType());
4893 return Builder.CreateAdd(tmp, addend);
4894 }
4895 case NEON::BI__builtin_neon_vpmin_v:
4896 case NEON::BI__builtin_neon_vpminq_v:
4897 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics.
4898 Int = usgn ? Intrinsic::aarch64_neon_uminp : Intrinsic::aarch64_neon_sminp;
4899 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fminp;
4900 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin");
4901 case NEON::BI__builtin_neon_vpmax_v:
4902 case NEON::BI__builtin_neon_vpmaxq_v:
4903 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics.
4904 Int = usgn ? Intrinsic::aarch64_neon_umaxp : Intrinsic::aarch64_neon_smaxp;
4905 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmaxp;
4906 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax");
4907 case NEON::BI__builtin_neon_vminnm_v:
4908 case NEON::BI__builtin_neon_vminnmq_v:
4909 Int = Intrinsic::aarch64_neon_fminnm;
4910 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vminnm");
4911 case NEON::BI__builtin_neon_vmaxnm_v:
4912 case NEON::BI__builtin_neon_vmaxnmq_v:
4913 Int = Intrinsic::aarch64_neon_fmaxnm;
4914 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmaxnm");
4915 case NEON::BI__builtin_neon_vrecpss_f32: {
4916 llvm::Type *f32Type = llvm::Type::getFloatTy(getLLVMContext());
4917 Ops.push_back(EmitScalarExpr(E->getArg(1)));
4918 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, f32Type),
4919 Ops, "vrecps");
4920 }
4921 case NEON::BI__builtin_neon_vrecpsd_f64: {
4922 llvm::Type *f64Type = llvm::Type::getDoubleTy(getLLVMContext());
4923 Ops.push_back(EmitScalarExpr(E->getArg(1)));
4924 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, f64Type),
4925 Ops, "vrecps");
4926 }
4927 case NEON::BI__builtin_neon_vqshrun_n_v:
4928 Int = Intrinsic::aarch64_neon_sqshrun;
4929 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrun_n");
4930 case NEON::BI__builtin_neon_vqrshrun_n_v:
4931 Int = Intrinsic::aarch64_neon_sqrshrun;
4932 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrun_n");
4933 case NEON::BI__builtin_neon_vqshrn_n_v:
4934 Int = usgn ? Intrinsic::aarch64_neon_uqshrn : Intrinsic::aarch64_neon_sqshrn;
4935 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n");
4936 case NEON::BI__builtin_neon_vrshrn_n_v:
4937 Int = Intrinsic::aarch64_neon_rshrn;
4938 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshrn_n");
4939 case NEON::BI__builtin_neon_vqrshrn_n_v:
4940 Int = usgn ? Intrinsic::aarch64_neon_uqrshrn : Intrinsic::aarch64_neon_sqrshrn;
4941 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n");
4942 case NEON::BI__builtin_neon_vrnda_v:
4943 case NEON::BI__builtin_neon_vrndaq_v: {
4944 Int = Intrinsic::round;
4945 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnda");
4946 }
4947 case NEON::BI__builtin_neon_vrndi_v:
4948 case NEON::BI__builtin_neon_vrndiq_v: {
4949 Int = Intrinsic::nearbyint;
4950 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndi");
4951 }
4952 case NEON::BI__builtin_neon_vrndm_v:
4953 case NEON::BI__builtin_neon_vrndmq_v: {
4954 Int = Intrinsic::floor;
4955 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndm");
4956 }
4957 case NEON::BI__builtin_neon_vrndn_v:
4958 case NEON::BI__builtin_neon_vrndnq_v: {
4959 Int = Intrinsic::aarch64_neon_frintn;
4960 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndn");
4961 }
4962 case NEON::BI__builtin_neon_vrndp_v:
4963 case NEON::BI__builtin_neon_vrndpq_v: {
4964 Int = Intrinsic::ceil;
4965 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndp");
4966 }
4967 case NEON::BI__builtin_neon_vrndx_v:
4968 case NEON::BI__builtin_neon_vrndxq_v: {
4969 Int = Intrinsic::rint;
4970 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndx");
4971 }
4972 case NEON::BI__builtin_neon_vrnd_v:
4973 case NEON::BI__builtin_neon_vrndq_v: {
4974 Int = Intrinsic::trunc;
4975 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndz");
4976 }
4977 case NEON::BI__builtin_neon_vceqz_v:
4978 case NEON::BI__builtin_neon_vceqzq_v:
4979 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OEQ,
4980 ICmpInst::ICMP_EQ, "vceqz");
4981 case NEON::BI__builtin_neon_vcgez_v:
4982 case NEON::BI__builtin_neon_vcgezq_v:
4983 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGE,
4984 ICmpInst::ICMP_SGE, "vcgez");
4985 case NEON::BI__builtin_neon_vclez_v:
4986 case NEON::BI__builtin_neon_vclezq_v:
4987 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLE,
4988 ICmpInst::ICMP_SLE, "vclez");
4989 case NEON::BI__builtin_neon_vcgtz_v:
4990 case NEON::BI__builtin_neon_vcgtzq_v:
4991 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGT,
4992 ICmpInst::ICMP_SGT, "vcgtz");
4993 case NEON::BI__builtin_neon_vcltz_v:
4994 case NEON::BI__builtin_neon_vcltzq_v:
4995 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLT,
4996 ICmpInst::ICMP_SLT, "vcltz");
4997 case NEON::BI__builtin_neon_vcvt_f64_v:
4998 case NEON::BI__builtin_neon_vcvtq_f64_v:
4999 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
5000 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, quad));
5001 return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt")
5002 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt");
5003 case NEON::BI__builtin_neon_vcvt_f64_f32: {
5004 assert(Type.getEltType() == NeonTypeFlags::Float64 && quad &&
5005 "unexpected vcvt_f64_f32 builtin");
5006 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float32, false, false);
5007 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag));
5008
5009 return Builder.CreateFPExt(Ops[0], Ty, "vcvt");
5010 }
5011 case NEON::BI__builtin_neon_vcvt_f32_f64: {
5012 assert(Type.getEltType() == NeonTypeFlags::Float32 &&
5013 "unexpected vcvt_f32_f64 builtin");
5014 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float64, false, true);
5015 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag));
5016
5017 return Builder.CreateFPTrunc(Ops[0], Ty, "vcvt");
5018 }
5019 case NEON::BI__builtin_neon_vcvt_s32_v:
5020 case NEON::BI__builtin_neon_vcvt_u32_v:
5021 case NEON::BI__builtin_neon_vcvt_s64_v:
5022 case NEON::BI__builtin_neon_vcvt_u64_v:
5023 case NEON::BI__builtin_neon_vcvtq_s32_v:
5024 case NEON::BI__builtin_neon_vcvtq_u32_v:
5025 case NEON::BI__builtin_neon_vcvtq_s64_v:
5026 case NEON::BI__builtin_neon_vcvtq_u64_v: {
5027 bool Double =
5028 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64);
5029 llvm::Type *InTy =
5030 GetNeonType(this,
5031 NeonTypeFlags(Double ? NeonTypeFlags::Float64
5032 : NeonTypeFlags::Float32, false, quad));
5033 Ops[0] = Builder.CreateBitCast(Ops[0], InTy);
5034 if (usgn)
5035 return Builder.CreateFPToUI(Ops[0], Ty);
5036 return Builder.CreateFPToSI(Ops[0], Ty);
5037 }
5038 case NEON::BI__builtin_neon_vcvta_s32_v:
5039 case NEON::BI__builtin_neon_vcvtaq_s32_v:
5040 case NEON::BI__builtin_neon_vcvta_u32_v:
5041 case NEON::BI__builtin_neon_vcvtaq_u32_v:
5042 case NEON::BI__builtin_neon_vcvta_s64_v:
5043 case NEON::BI__builtin_neon_vcvtaq_s64_v:
5044 case NEON::BI__builtin_neon_vcvta_u64_v:
5045 case NEON::BI__builtin_neon_vcvtaq_u64_v: {
5046 Int = usgn ? Intrinsic::aarch64_neon_fcvtau : Intrinsic::aarch64_neon_fcvtas;
5047 bool Double =
5048 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64);
5049 llvm::Type *InTy =
5050 GetNeonType(this,
5051 NeonTypeFlags(Double ? NeonTypeFlags::Float64
5052 : NeonTypeFlags::Float32, false, quad));
5053 llvm::Type *Tys[2] = { Ty, InTy };
5054 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvta");
5055 }
5056 case NEON::BI__builtin_neon_vcvtm_s32_v:
5057 case NEON::BI__builtin_neon_vcvtmq_s32_v:
5058 case NEON::BI__builtin_neon_vcvtm_u32_v:
5059 case NEON::BI__builtin_neon_vcvtmq_u32_v:
5060 case NEON::BI__builtin_neon_vcvtm_s64_v:
5061 case NEON::BI__builtin_neon_vcvtmq_s64_v:
5062 case NEON::BI__builtin_neon_vcvtm_u64_v:
5063 case NEON::BI__builtin_neon_vcvtmq_u64_v: {
5064 Int = usgn ? Intrinsic::aarch64_neon_fcvtmu : Intrinsic::aarch64_neon_fcvtms;
5065 bool Double =
5066 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64);
5067 llvm::Type *InTy =
5068 GetNeonType(this,
5069 NeonTypeFlags(Double ? NeonTypeFlags::Float64
5070 : NeonTypeFlags::Float32, false, quad));
5071 llvm::Type *Tys[2] = { Ty, InTy };
5072 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtm");
5073 }
5074 case NEON::BI__builtin_neon_vcvtn_s32_v:
5075 case NEON::BI__builtin_neon_vcvtnq_s32_v:
5076 case NEON::BI__builtin_neon_vcvtn_u32_v:
5077 case NEON::BI__builtin_neon_vcvtnq_u32_v:
5078 case NEON::BI__builtin_neon_vcvtn_s64_v:
5079 case NEON::BI__builtin_neon_vcvtnq_s64_v:
5080 case NEON::BI__builtin_neon_vcvtn_u64_v:
5081 case NEON::BI__builtin_neon_vcvtnq_u64_v: {
5082 Int = usgn ? Intrinsic::aarch64_neon_fcvtnu : Intrinsic::aarch64_neon_fcvtns;
5083 bool Double =
5084 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64);
5085 llvm::Type *InTy =
5086 GetNeonType(this,
5087 NeonTypeFlags(Double ? NeonTypeFlags::Float64
5088 : NeonTypeFlags::Float32, false, quad));
5089 llvm::Type *Tys[2] = { Ty, InTy };
5090 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtn");
5091 }
5092 case NEON::BI__builtin_neon_vcvtp_s32_v:
5093 case NEON::BI__builtin_neon_vcvtpq_s32_v:
5094 case NEON::BI__builtin_neon_vcvtp_u32_v:
5095 case NEON::BI__builtin_neon_vcvtpq_u32_v:
5096 case NEON::BI__builtin_neon_vcvtp_s64_v:
5097 case NEON::BI__builtin_neon_vcvtpq_s64_v:
5098 case NEON::BI__builtin_neon_vcvtp_u64_v:
5099 case NEON::BI__builtin_neon_vcvtpq_u64_v: {
5100 Int = usgn ? Intrinsic::aarch64_neon_fcvtpu : Intrinsic::aarch64_neon_fcvtps;
5101 bool Double =
5102 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64);
5103 llvm::Type *InTy =
5104 GetNeonType(this,
5105 NeonTypeFlags(Double ? NeonTypeFlags::Float64
5106 : NeonTypeFlags::Float32, false, quad));
5107 llvm::Type *Tys[2] = { Ty, InTy };
5108 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtp");
5109 }
5110 case NEON::BI__builtin_neon_vmulx_v:
5111 case NEON::BI__builtin_neon_vmulxq_v: {
5112 Int = Intrinsic::aarch64_neon_fmulx;
5113 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmulx");
5114 }
5115 case NEON::BI__builtin_neon_vmul_lane_v:
5116 case NEON::BI__builtin_neon_vmul_laneq_v: {
5117 // v1f64 vmul_lane should be mapped to Neon scalar mul lane
5118 bool Quad = false;
5119 if (BuiltinID == NEON::BI__builtin_neon_vmul_laneq_v)
5120 Quad = true;
5121 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy);
5122 llvm::Type *VTy = GetNeonType(this,
5123 NeonTypeFlags(NeonTypeFlags::Float64, false, Quad));
5124 Ops[1] = Builder.CreateBitCast(Ops[1], VTy);
5125 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract");
5126 Value *Result = Builder.CreateFMul(Ops[0], Ops[1]);
5127 return Builder.CreateBitCast(Result, Ty);
5128 }
5129 case NEON::BI__builtin_neon_vnegd_s64:
5130 return Builder.CreateNeg(EmitScalarExpr(E->getArg(0)), "vnegd");
5131 case NEON::BI__builtin_neon_vpmaxnm_v:
5132 case NEON::BI__builtin_neon_vpmaxnmq_v: {
5133 Int = Intrinsic::aarch64_neon_fmaxnmp;
5134 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmaxnm");
5135 }
5136 case NEON::BI__builtin_neon_vpminnm_v:
5137 case NEON::BI__builtin_neon_vpminnmq_v: {
5138 Int = Intrinsic::aarch64_neon_fminnmp;
5139 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpminnm");
5140 }
5141 case NEON::BI__builtin_neon_vsqrt_v:
5142 case NEON::BI__builtin_neon_vsqrtq_v: {
5143 Int = Intrinsic::sqrt;
5144 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
5145 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqrt");
5146 }
5147 case NEON::BI__builtin_neon_vrbit_v:
5148 case NEON::BI__builtin_neon_vrbitq_v: {
5149 Int = Intrinsic::aarch64_neon_rbit;
5150 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrbit");
5151 }
5152 case NEON::BI__builtin_neon_vaddv_u8:
5153 // FIXME: These are handled by the AArch64 scalar code.
5154 usgn = true;
5155 // FALLTHROUGH
5156 case NEON::BI__builtin_neon_vaddv_s8: {
5157 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv;
5158 Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5159 VTy =
5160 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8);
5161 llvm::Type *Tys[2] = { Ty, VTy };
5162 Ops.push_back(EmitScalarExpr(E->getArg(0)));
5163 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv");
5164 return Builder.CreateTrunc(Ops[0],
5165 llvm::IntegerType::get(getLLVMContext(), 8));
5166 }
5167 case NEON::BI__builtin_neon_vaddv_u16:
5168 usgn = true;
5169 // FALLTHROUGH
5170 case NEON::BI__builtin_neon_vaddv_s16: {
5171 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv;
5172 Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5173 VTy =
5174 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4);
5175 llvm::Type *Tys[2] = { Ty, VTy };
5176 Ops.push_back(EmitScalarExpr(E->getArg(0)));
5177 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv");
5178 return Builder.CreateTrunc(Ops[0],
5179 llvm::IntegerType::get(getLLVMContext(), 16));
5180 }
5181 case NEON::BI__builtin_neon_vaddvq_u8:
5182 usgn = true;
5183 // FALLTHROUGH
5184 case NEON::BI__builtin_neon_vaddvq_s8: {
5185 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv;
5186 Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5187 VTy =
5188 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16);
5189 llvm::Type *Tys[2] = { Ty, VTy };
5190 Ops.push_back(EmitScalarExpr(E->getArg(0)));
5191 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv");
5192 return Builder.CreateTrunc(Ops[0],
5193 llvm::IntegerType::get(getLLVMContext(), 8));
5194 }
5195 case NEON::BI__builtin_neon_vaddvq_u16:
5196 usgn = true;
5197 // FALLTHROUGH
5198 case NEON::BI__builtin_neon_vaddvq_s16: {
5199 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv;
5200 Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5201 VTy =
5202 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8);
5203 llvm::Type *Tys[2] = { Ty, VTy };
5204 Ops.push_back(EmitScalarExpr(E->getArg(0)));
5205 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv");
5206 return Builder.CreateTrunc(Ops[0],
5207 llvm::IntegerType::get(getLLVMContext(), 16));
5208 }
5209 case NEON::BI__builtin_neon_vmaxv_u8: {
5210 Int = Intrinsic::aarch64_neon_umaxv;
5211 Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5212 VTy =
5213 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8);
5214 llvm::Type *Tys[2] = { Ty, VTy };
5215 Ops.push_back(EmitScalarExpr(E->getArg(0)));
5216 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
5217 return Builder.CreateTrunc(Ops[0],
5218 llvm::IntegerType::get(getLLVMContext(), 8));
5219 }
5220 case NEON::BI__builtin_neon_vmaxv_u16: {
5221 Int = Intrinsic::aarch64_neon_umaxv;
5222 Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5223 VTy =
5224 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4);
5225 llvm::Type *Tys[2] = { Ty, VTy };
5226 Ops.push_back(EmitScalarExpr(E->getArg(0)));
5227 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
5228 return Builder.CreateTrunc(Ops[0],
5229 llvm::IntegerType::get(getLLVMContext(), 16));
5230 }
5231 case NEON::BI__builtin_neon_vmaxvq_u8: {
5232 Int = Intrinsic::aarch64_neon_umaxv;
5233 Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5234 VTy =
5235 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16);
5236 llvm::Type *Tys[2] = { Ty, VTy };
5237 Ops.push_back(EmitScalarExpr(E->getArg(0)));
5238 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
5239 return Builder.CreateTrunc(Ops[0],
5240 llvm::IntegerType::get(getLLVMContext(), 8));
5241 }
5242 case NEON::BI__builtin_neon_vmaxvq_u16: {
5243 Int = Intrinsic::aarch64_neon_umaxv;
5244 Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5245 VTy =
5246 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8);
5247 llvm::Type *Tys[2] = { Ty, VTy };
5248 Ops.push_back(EmitScalarExpr(E->getArg(0)));
5249 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
5250 return Builder.CreateTrunc(Ops[0],
5251 llvm::IntegerType::get(getLLVMContext(), 16));
5252 }
5253 case NEON::BI__builtin_neon_vmaxv_s8: {
5254 Int = Intrinsic::aarch64_neon_smaxv;
5255 Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5256 VTy =
5257 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8);
5258 llvm::Type *Tys[2] = { Ty, VTy };
5259 Ops.push_back(EmitScalarExpr(E->getArg(0)));
5260 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
5261 return Builder.CreateTrunc(Ops[0],
5262 llvm::IntegerType::get(getLLVMContext(), 8));
5263 }
5264 case NEON::BI__builtin_neon_vmaxv_s16: {
5265 Int = Intrinsic::aarch64_neon_smaxv;
5266 Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5267 VTy =
5268 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4);
5269 llvm::Type *Tys[2] = { Ty, VTy };
5270 Ops.push_back(EmitScalarExpr(E->getArg(0)));
5271 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
5272 return Builder.CreateTrunc(Ops[0],
5273 llvm::IntegerType::get(getLLVMContext(), 16));
5274 }
5275 case NEON::BI__builtin_neon_vmaxvq_s8: {
5276 Int = Intrinsic::aarch64_neon_smaxv;
5277 Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5278 VTy =
5279 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16);
5280 llvm::Type *Tys[2] = { Ty, VTy };
5281 Ops.push_back(EmitScalarExpr(E->getArg(0)));
5282 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
5283 return Builder.CreateTrunc(Ops[0],
5284 llvm::IntegerType::get(getLLVMContext(), 8));
5285 }
5286 case NEON::BI__builtin_neon_vmaxvq_s16: {
5287 Int = Intrinsic::aarch64_neon_smaxv;
5288 Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5289 VTy =
5290 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8);
5291 llvm::Type *Tys[2] = { Ty, VTy };
5292 Ops.push_back(EmitScalarExpr(E->getArg(0)));
5293 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
5294 return Builder.CreateTrunc(Ops[0],
5295 llvm::IntegerType::get(getLLVMContext(), 16));
5296 }
5297 case NEON::BI__builtin_neon_vminv_u8: {
5298 Int = Intrinsic::aarch64_neon_uminv;
5299 Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5300 VTy =
5301 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8);
5302 llvm::Type *Tys[2] = { Ty, VTy };
5303 Ops.push_back(EmitScalarExpr(E->getArg(0)));
5304 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
5305 return Builder.CreateTrunc(Ops[0],
5306 llvm::IntegerType::get(getLLVMContext(), 8));
5307 }
5308 case NEON::BI__builtin_neon_vminv_u16: {
5309 Int = Intrinsic::aarch64_neon_uminv;
5310 Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5311 VTy =
5312 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4);
5313 llvm::Type *Tys[2] = { Ty, VTy };
5314 Ops.push_back(EmitScalarExpr(E->getArg(0)));
5315 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
5316 return Builder.CreateTrunc(Ops[0],
5317 llvm::IntegerType::get(getLLVMContext(), 16));
5318 }
5319 case NEON::BI__builtin_neon_vminvq_u8: {
5320 Int = Intrinsic::aarch64_neon_uminv;
5321 Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5322 VTy =
5323 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16);
5324 llvm::Type *Tys[2] = { Ty, VTy };
5325 Ops.push_back(EmitScalarExpr(E->getArg(0)));
5326 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
5327 return Builder.CreateTrunc(Ops[0],
5328 llvm::IntegerType::get(getLLVMContext(), 8));
5329 }
5330 case NEON::BI__builtin_neon_vminvq_u16: {
5331 Int = Intrinsic::aarch64_neon_uminv;
5332 Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5333 VTy =
5334 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8);
5335 llvm::Type *Tys[2] = { Ty, VTy };
5336 Ops.push_back(EmitScalarExpr(E->getArg(0)));
5337 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
5338 return Builder.CreateTrunc(Ops[0],
5339 llvm::IntegerType::get(getLLVMContext(), 16));
5340 }
5341 case NEON::BI__builtin_neon_vminv_s8: {
5342 Int = Intrinsic::aarch64_neon_sminv;
5343 Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5344 VTy =
5345 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8);
5346 llvm::Type *Tys[2] = { Ty, VTy };
5347 Ops.push_back(EmitScalarExpr(E->getArg(0)));
5348 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
5349 return Builder.CreateTrunc(Ops[0],
5350 llvm::IntegerType::get(getLLVMContext(), 8));
5351 }
5352 case NEON::BI__builtin_neon_vminv_s16: {
5353 Int = Intrinsic::aarch64_neon_sminv;
5354 Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5355 VTy =
5356 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4);
5357 llvm::Type *Tys[2] = { Ty, VTy };
5358 Ops.push_back(EmitScalarExpr(E->getArg(0)));
5359 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
5360 return Builder.CreateTrunc(Ops[0],
5361 llvm::IntegerType::get(getLLVMContext(), 16));
5362 }
5363 case NEON::BI__builtin_neon_vminvq_s8: {
5364 Int = Intrinsic::aarch64_neon_sminv;
5365 Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5366 VTy =
5367 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16);
5368 llvm::Type *Tys[2] = { Ty, VTy };
5369 Ops.push_back(EmitScalarExpr(E->getArg(0)));
5370 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
5371 return Builder.CreateTrunc(Ops[0],
5372 llvm::IntegerType::get(getLLVMContext(), 8));
5373 }
5374 case NEON::BI__builtin_neon_vminvq_s16: {
5375 Int = Intrinsic::aarch64_neon_sminv;
5376 Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5377 VTy =
5378 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8);
5379 llvm::Type *Tys[2] = { Ty, VTy };
5380 Ops.push_back(EmitScalarExpr(E->getArg(0)));
5381 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
5382 return Builder.CreateTrunc(Ops[0],
5383 llvm::IntegerType::get(getLLVMContext(), 16));
5384 }
5385 case NEON::BI__builtin_neon_vmul_n_f64: {
5386 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy);
5387 Value *RHS = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), DoubleTy);
5388 return Builder.CreateFMul(Ops[0], RHS);
5389 }
5390 case NEON::BI__builtin_neon_vaddlv_u8: {
5391 Int = Intrinsic::aarch64_neon_uaddlv;
5392 Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5393 VTy =
5394 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8);
5395 llvm::Type *Tys[2] = { Ty, VTy };
5396 Ops.push_back(EmitScalarExpr(E->getArg(0)));
5397 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
5398 return Builder.CreateTrunc(Ops[0],
5399 llvm::IntegerType::get(getLLVMContext(), 16));
5400 }
5401 case NEON::BI__builtin_neon_vaddlv_u16: {
5402 Int = Intrinsic::aarch64_neon_uaddlv;
5403 Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5404 VTy =
5405 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4);
5406 llvm::Type *Tys[2] = { Ty, VTy };
5407 Ops.push_back(EmitScalarExpr(E->getArg(0)));
5408 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
5409 }
5410 case NEON::BI__builtin_neon_vaddlvq_u8: {
5411 Int = Intrinsic::aarch64_neon_uaddlv;
5412 Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5413 VTy =
5414 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16);
5415 llvm::Type *Tys[2] = { Ty, VTy };
5416 Ops.push_back(EmitScalarExpr(E->getArg(0)));
5417 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
5418 return Builder.CreateTrunc(Ops[0],
5419 llvm::IntegerType::get(getLLVMContext(), 16));
5420 }
5421 case NEON::BI__builtin_neon_vaddlvq_u16: {
5422 Int = Intrinsic::aarch64_neon_uaddlv;
5423 Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5424 VTy =
5425 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8);
5426 llvm::Type *Tys[2] = { Ty, VTy };
5427 Ops.push_back(EmitScalarExpr(E->getArg(0)));
5428 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
5429 }
5430 case NEON::BI__builtin_neon_vaddlv_s8: {
5431 Int = Intrinsic::aarch64_neon_saddlv;
5432 Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5433 VTy =
5434 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8);
5435 llvm::Type *Tys[2] = { Ty, VTy };
5436 Ops.push_back(EmitScalarExpr(E->getArg(0)));
5437 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
5438 return Builder.CreateTrunc(Ops[0],
5439 llvm::IntegerType::get(getLLVMContext(), 16));
5440 }
5441 case NEON::BI__builtin_neon_vaddlv_s16: {
5442 Int = Intrinsic::aarch64_neon_saddlv;
5443 Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5444 VTy =
5445 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4);
5446 llvm::Type *Tys[2] = { Ty, VTy };
5447 Ops.push_back(EmitScalarExpr(E->getArg(0)));
5448 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
5449 }
5450 case NEON::BI__builtin_neon_vaddlvq_s8: {
5451 Int = Intrinsic::aarch64_neon_saddlv;
5452 Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5453 VTy =
5454 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16);
5455 llvm::Type *Tys[2] = { Ty, VTy };
5456 Ops.push_back(EmitScalarExpr(E->getArg(0)));
5457 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
5458 return Builder.CreateTrunc(Ops[0],
5459 llvm::IntegerType::get(getLLVMContext(), 16));
5460 }
5461 case NEON::BI__builtin_neon_vaddlvq_s16: {
5462 Int = Intrinsic::aarch64_neon_saddlv;
5463 Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5464 VTy =
5465 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8);
5466 llvm::Type *Tys[2] = { Ty, VTy };
5467 Ops.push_back(EmitScalarExpr(E->getArg(0)));
5468 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
5469 }
5470 case NEON::BI__builtin_neon_vsri_n_v:
5471 case NEON::BI__builtin_neon_vsriq_n_v: {
5472 Int = Intrinsic::aarch64_neon_vsri;
5473 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty);
5474 return EmitNeonCall(Intrin, Ops, "vsri_n");
5475 }
5476 case NEON::BI__builtin_neon_vsli_n_v:
5477 case NEON::BI__builtin_neon_vsliq_n_v: {
5478 Int = Intrinsic::aarch64_neon_vsli;
5479 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty);
5480 return EmitNeonCall(Intrin, Ops, "vsli_n");
5481 }
5482 case NEON::BI__builtin_neon_vsra_n_v:
5483 case NEON::BI__builtin_neon_vsraq_n_v:
5484 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
5485 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n");
5486 return Builder.CreateAdd(Ops[0], Ops[1]);
5487 case NEON::BI__builtin_neon_vrsra_n_v:
5488 case NEON::BI__builtin_neon_vrsraq_n_v: {
5489 Int = usgn ? Intrinsic::aarch64_neon_urshl : Intrinsic::aarch64_neon_srshl;
5490 SmallVector<llvm::Value*,2> TmpOps;
5491 TmpOps.push_back(Ops[1]);
5492 TmpOps.push_back(Ops[2]);
5493 Function* F = CGM.getIntrinsic(Int, Ty);
5494 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vrshr_n", 1, true);
5495 Ops[0] = Builder.CreateBitCast(Ops[0], VTy);
5496 return Builder.CreateAdd(Ops[0], tmp);
5497 }
5498 // FIXME: Sharing loads & stores with 32-bit is complicated by the absence
5499 // of an Align parameter here.
5500 case NEON::BI__builtin_neon_vld1_x2_v:
5501 case NEON::BI__builtin_neon_vld1q_x2_v:
5502 case NEON::BI__builtin_neon_vld1_x3_v:
5503 case NEON::BI__builtin_neon_vld1q_x3_v:
5504 case NEON::BI__builtin_neon_vld1_x4_v:
5505 case NEON::BI__builtin_neon_vld1q_x4_v: {
5506 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType());
5507 Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
5508 llvm::Type *Tys[2] = { VTy, PTy };
5509 unsigned Int;
5510 switch (BuiltinID) {
5511 case NEON::BI__builtin_neon_vld1_x2_v:
5512 case NEON::BI__builtin_neon_vld1q_x2_v:
5513 Int = Intrinsic::aarch64_neon_ld1x2;
5514 break;
5515 case NEON::BI__builtin_neon_vld1_x3_v:
5516 case NEON::BI__builtin_neon_vld1q_x3_v:
5517 Int = Intrinsic::aarch64_neon_ld1x3;
5518 break;
5519 case NEON::BI__builtin_neon_vld1_x4_v:
5520 case NEON::BI__builtin_neon_vld1q_x4_v:
5521 Int = Intrinsic::aarch64_neon_ld1x4;
5522 break;
5523 }
5524 Function *F = CGM.getIntrinsic(Int, Tys);
5525 Ops[1] = Builder.CreateCall(F, Ops[1], "vld1xN");
5526 Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
5527 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
5528 return Builder.CreateStore(Ops[1], Ops[0]);
5529 }
5530 case NEON::BI__builtin_neon_vst1_x2_v:
5531 case NEON::BI__builtin_neon_vst1q_x2_v:
5532 case NEON::BI__builtin_neon_vst1_x3_v:
5533 case NEON::BI__builtin_neon_vst1q_x3_v:
5534 case NEON::BI__builtin_neon_vst1_x4_v:
5535 case NEON::BI__builtin_neon_vst1q_x4_v: {
5536 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType());
5537 llvm::Type *Tys[2] = { VTy, PTy };
5538 unsigned Int;
5539 switch (BuiltinID) {
5540 case NEON::BI__builtin_neon_vst1_x2_v:
5541 case NEON::BI__builtin_neon_vst1q_x2_v:
5542 Int = Intrinsic::aarch64_neon_st1x2;
5543 break;
5544 case NEON::BI__builtin_neon_vst1_x3_v:
5545 case NEON::BI__builtin_neon_vst1q_x3_v:
5546 Int = Intrinsic::aarch64_neon_st1x3;
5547 break;
5548 case NEON::BI__builtin_neon_vst1_x4_v:
5549 case NEON::BI__builtin_neon_vst1q_x4_v:
5550 Int = Intrinsic::aarch64_neon_st1x4;
5551 break;
5552 }
5553 SmallVector<Value *, 4> IntOps(Ops.begin()+1, Ops.end());
5554 IntOps.push_back(Ops[0]);
5555 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), IntOps, "");
5556 }
5557 case NEON::BI__builtin_neon_vld1_v:
5558 case NEON::BI__builtin_neon_vld1q_v:
5559 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy));
5560 return Builder.CreateLoad(Ops[0]);
5561 case NEON::BI__builtin_neon_vst1_v:
5562 case NEON::BI__builtin_neon_vst1q_v:
5563 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy));
5564 Ops[1] = Builder.CreateBitCast(Ops[1], VTy);
5565 return Builder.CreateStore(Ops[1], Ops[0]);
5566 case NEON::BI__builtin_neon_vld1_lane_v:
5567 case NEON::BI__builtin_neon_vld1q_lane_v:
5568 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
5569 Ty = llvm::PointerType::getUnqual(VTy->getElementType());
5570 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
5571 Ops[0] = Builder.CreateLoad(Ops[0]);
5572 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vld1_lane");
5573 case NEON::BI__builtin_neon_vld1_dup_v:
5574 case NEON::BI__builtin_neon_vld1q_dup_v: {
5575 Value *V = UndefValue::get(Ty);
5576 Ty = llvm::PointerType::getUnqual(VTy->getElementType());
5577 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
5578 Ops[0] = Builder.CreateLoad(Ops[0]);
5579 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0);
5580 Ops[0] = Builder.CreateInsertElement(V, Ops[0], CI);
5581 return EmitNeonSplat(Ops[0], CI);
5582 }
5583 case NEON::BI__builtin_neon_vst1_lane_v:
5584 case NEON::BI__builtin_neon_vst1q_lane_v:
5585 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
5586 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]);
5587 Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
5588 return Builder.CreateStore(Ops[1], Builder.CreateBitCast(Ops[0], Ty));
5589 case NEON::BI__builtin_neon_vld2_v:
5590 case NEON::BI__builtin_neon_vld2q_v: {
5591 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy);
5592 Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
5593 llvm::Type *Tys[2] = { VTy, PTy };
5594 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2, Tys);
5595 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2");
5596 Ops[0] = Builder.CreateBitCast(Ops[0],
5597 llvm::PointerType::getUnqual(Ops[1]->getType()));
5598 return Builder.CreateStore(Ops[1], Ops[0]);
5599 }
5600 case NEON::BI__builtin_neon_vld3_v:
5601 case NEON::BI__builtin_neon_vld3q_v: {
5602 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy);
5603 Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
5604 llvm::Type *Tys[2] = { VTy, PTy };
5605 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3, Tys);
5606 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3");
5607 Ops[0] = Builder.CreateBitCast(Ops[0],
5608 llvm::PointerType::getUnqual(Ops[1]->getType()));
5609 return Builder.CreateStore(Ops[1], Ops[0]);
5610 }
5611 case NEON::BI__builtin_neon_vld4_v:
5612 case NEON::BI__builtin_neon_vld4q_v: {
5613 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy);
5614 Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
5615 llvm::Type *Tys[2] = { VTy, PTy };
5616 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4, Tys);
5617 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4");
5618 Ops[0] = Builder.CreateBitCast(Ops[0],
5619 llvm::PointerType::getUnqual(Ops[1]->getType()));
5620 return Builder.CreateStore(Ops[1], Ops[0]);
5621 }
5622 case NEON::BI__builtin_neon_vld2_dup_v:
5623 case NEON::BI__builtin_neon_vld2q_dup_v: {
5624 llvm::Type *PTy =
5625 llvm::PointerType::getUnqual(VTy->getElementType());
5626 Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
5627 llvm::Type *Tys[2] = { VTy, PTy };
5628 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2r, Tys);
5629 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2");
5630 Ops[0] = Builder.CreateBitCast(Ops[0],
5631 llvm::PointerType::getUnqual(Ops[1]->getType()));
5632 return Builder.CreateStore(Ops[1], Ops[0]);
5633 }
5634 case NEON::BI__builtin_neon_vld3_dup_v:
5635 case NEON::BI__builtin_neon_vld3q_dup_v: {
5636 llvm::Type *PTy =
5637 llvm::PointerType::getUnqual(VTy->getElementType());
5638 Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
5639 llvm::Type *Tys[2] = { VTy, PTy };
5640 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3r, Tys);
5641 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3");
5642 Ops[0] = Builder.CreateBitCast(Ops[0],
5643 llvm::PointerType::getUnqual(Ops[1]->getType()));
5644 return Builder.CreateStore(Ops[1], Ops[0]);
5645 }
5646 case NEON::BI__builtin_neon_vld4_dup_v:
5647 case NEON::BI__builtin_neon_vld4q_dup_v: {
5648 llvm::Type *PTy =
5649 llvm::PointerType::getUnqual(VTy->getElementType());
5650 Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
5651 llvm::Type *Tys[2] = { VTy, PTy };
5652 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4r, Tys);
5653 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4");
5654 Ops[0] = Builder.CreateBitCast(Ops[0],
5655 llvm::PointerType::getUnqual(Ops[1]->getType()));
5656 return Builder.CreateStore(Ops[1], Ops[0]);
5657 }
5658 case NEON::BI__builtin_neon_vld2_lane_v:
5659 case NEON::BI__builtin_neon_vld2q_lane_v: {
5660 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() };
5661 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2lane, Tys);
5662 Ops.push_back(Ops[1]);
5663 Ops.erase(Ops.begin()+1);
5664 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
5665 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
5666 Ops[3] = Builder.CreateZExt(Ops[3],
5667 llvm::IntegerType::get(getLLVMContext(), 64));
5668 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld2_lane");
5669 Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
5670 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
5671 return Builder.CreateStore(Ops[1], Ops[0]);
5672 }
5673 case NEON::BI__builtin_neon_vld3_lane_v:
5674 case NEON::BI__builtin_neon_vld3q_lane_v: {
5675 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() };
5676 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3lane, Tys);
5677 Ops.push_back(Ops[1]);
5678 Ops.erase(Ops.begin()+1);
5679 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
5680 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
5681 Ops[3] = Builder.CreateBitCast(Ops[3], Ty);
5682 Ops[4] = Builder.CreateZExt(Ops[4],
5683 llvm::IntegerType::get(getLLVMContext(), 64));
5684 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane");
5685 Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
5686 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
5687 return Builder.CreateStore(Ops[1], Ops[0]);
5688 }
5689 case NEON::BI__builtin_neon_vld4_lane_v:
5690 case NEON::BI__builtin_neon_vld4q_lane_v: {
5691 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() };
5692 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4lane, Tys);
5693 Ops.push_back(Ops[1]);
5694 Ops.erase(Ops.begin()+1);
5695 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
5696 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
5697 Ops[3] = Builder.CreateBitCast(Ops[3], Ty);
5698 Ops[4] = Builder.CreateBitCast(Ops[4], Ty);
5699 Ops[5] = Builder.CreateZExt(Ops[5],
5700 llvm::IntegerType::get(getLLVMContext(), 64));
5701 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld4_lane");
5702 Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
5703 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
5704 return Builder.CreateStore(Ops[1], Ops[0]);
5705 }
5706 case NEON::BI__builtin_neon_vst2_v:
5707 case NEON::BI__builtin_neon_vst2q_v: {
5708 Ops.push_back(Ops[0]);
5709 Ops.erase(Ops.begin());
5710 llvm::Type *Tys[2] = { VTy, Ops[2]->getType() };
5711 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2, Tys),
5712 Ops, "");
5713 }
5714 case NEON::BI__builtin_neon_vst2_lane_v:
5715 case NEON::BI__builtin_neon_vst2q_lane_v: {
5716 Ops.push_back(Ops[0]);
5717 Ops.erase(Ops.begin());
5718 Ops[2] = Builder.CreateZExt(Ops[2],
5719 llvm::IntegerType::get(getLLVMContext(), 64));
5720 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() };
5721 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2lane, Tys),
5722 Ops, "");
5723 }
5724 case NEON::BI__builtin_neon_vst3_v:
5725 case NEON::BI__builtin_neon_vst3q_v: {
5726 Ops.push_back(Ops[0]);
5727 Ops.erase(Ops.begin());
5728 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() };
5729 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3, Tys),
5730 Ops, "");
5731 }
5732 case NEON::BI__builtin_neon_vst3_lane_v:
5733 case NEON::BI__builtin_neon_vst3q_lane_v: {
5734 Ops.push_back(Ops[0]);
5735 Ops.erase(Ops.begin());
5736 Ops[3] = Builder.CreateZExt(Ops[3],
5737 llvm::IntegerType::get(getLLVMContext(), 64));
5738 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() };
5739 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3lane, Tys),
5740 Ops, "");
5741 }
5742 case NEON::BI__builtin_neon_vst4_v:
5743 case NEON::BI__builtin_neon_vst4q_v: {
5744 Ops.push_back(Ops[0]);
5745 Ops.erase(Ops.begin());
5746 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() };
5747 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4, Tys),
5748 Ops, "");
5749 }
5750 case NEON::BI__builtin_neon_vst4_lane_v:
5751 case NEON::BI__builtin_neon_vst4q_lane_v: {
5752 Ops.push_back(Ops[0]);
5753 Ops.erase(Ops.begin());
5754 Ops[4] = Builder.CreateZExt(Ops[4],
5755 llvm::IntegerType::get(getLLVMContext(), 64));
5756 llvm::Type *Tys[2] = { VTy, Ops[5]->getType() };
5757 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4lane, Tys),
5758 Ops, "");
5759 }
5760 case NEON::BI__builtin_neon_vtrn_v:
5761 case NEON::BI__builtin_neon_vtrnq_v: {
5762 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
5763 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
5764 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
5765 Value *SV = nullptr;
5766
5767 for (unsigned vi = 0; vi != 2; ++vi) {
5768 SmallVector<Constant*, 16> Indices;
5769 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) {
5770 Indices.push_back(ConstantInt::get(Int32Ty, i+vi));
5771 Indices.push_back(ConstantInt::get(Int32Ty, i+e+vi));
5772 }
5773 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi);
5774 SV = llvm::ConstantVector::get(Indices);
5775 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn");
5776 SV = Builder.CreateStore(SV, Addr);
5777 }
5778 return SV;
5779 }
5780 case NEON::BI__builtin_neon_vuzp_v:
5781 case NEON::BI__builtin_neon_vuzpq_v: {
5782 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
5783 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
5784 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
5785 Value *SV = nullptr;
5786
5787 for (unsigned vi = 0; vi != 2; ++vi) {
5788 SmallVector<Constant*, 16> Indices;
5789 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i)
5790 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi));
5791
5792 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi);
5793 SV = llvm::ConstantVector::get(Indices);
5794 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp");
5795 SV = Builder.CreateStore(SV, Addr);
5796 }
5797 return SV;
5798 }
5799 case NEON::BI__builtin_neon_vzip_v:
5800 case NEON::BI__builtin_neon_vzipq_v: {
5801 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
5802 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
5803 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
5804 Value *SV = nullptr;
5805
5806 for (unsigned vi = 0; vi != 2; ++vi) {
5807 SmallVector<Constant*, 16> Indices;
5808 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) {
5809 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1));
5810 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e));
5811 }
5812 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi);
5813 SV = llvm::ConstantVector::get(Indices);
5814 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip");
5815 SV = Builder.CreateStore(SV, Addr);
5816 }
5817 return SV;
5818 }
5819 case NEON::BI__builtin_neon_vqtbl1q_v: {
5820 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl1, Ty),
5821 Ops, "vtbl1");
5822 }
5823 case NEON::BI__builtin_neon_vqtbl2q_v: {
5824 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl2, Ty),
5825 Ops, "vtbl2");
5826 }
5827 case NEON::BI__builtin_neon_vqtbl3q_v: {
5828 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl3, Ty),
5829 Ops, "vtbl3");
5830 }
5831 case NEON::BI__builtin_neon_vqtbl4q_v: {
5832 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl4, Ty),
5833 Ops, "vtbl4");
5834 }
5835 case NEON::BI__builtin_neon_vqtbx1q_v: {
5836 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx1, Ty),
5837 Ops, "vtbx1");
5838 }
5839 case NEON::BI__builtin_neon_vqtbx2q_v: {
5840 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx2, Ty),
5841 Ops, "vtbx2");
5842 }
5843 case NEON::BI__builtin_neon_vqtbx3q_v: {
5844 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx3, Ty),
5845 Ops, "vtbx3");
5846 }
5847 case NEON::BI__builtin_neon_vqtbx4q_v: {
5848 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx4, Ty),
5849 Ops, "vtbx4");
5850 }
5851 case NEON::BI__builtin_neon_vsqadd_v:
5852 case NEON::BI__builtin_neon_vsqaddq_v: {
5853 Int = Intrinsic::aarch64_neon_usqadd;
5854 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqadd");
5855 }
5856 case NEON::BI__builtin_neon_vuqadd_v:
5857 case NEON::BI__builtin_neon_vuqaddq_v: {
5858 Int = Intrinsic::aarch64_neon_suqadd;
5859 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vuqadd");
5860 }
5861 }
5862 }
5863
5864 llvm::Value *CodeGenFunction::
BuildVector(ArrayRef<llvm::Value * > Ops)5865 BuildVector(ArrayRef<llvm::Value*> Ops) {
5866 assert((Ops.size() & (Ops.size() - 1)) == 0 &&
5867 "Not a power-of-two sized vector!");
5868 bool AllConstants = true;
5869 for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i)
5870 AllConstants &= isa<Constant>(Ops[i]);
5871
5872 // If this is a constant vector, create a ConstantVector.
5873 if (AllConstants) {
5874 SmallVector<llvm::Constant*, 16> CstOps;
5875 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
5876 CstOps.push_back(cast<Constant>(Ops[i]));
5877 return llvm::ConstantVector::get(CstOps);
5878 }
5879
5880 // Otherwise, insertelement the values to build the vector.
5881 Value *Result =
5882 llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size()));
5883
5884 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
5885 Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt32(i));
5886
5887 return Result;
5888 }
5889
EmitX86BuiltinExpr(unsigned BuiltinID,const CallExpr * E)5890 Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID,
5891 const CallExpr *E) {
5892 SmallVector<Value*, 4> Ops;
5893
5894 // Find out if any arguments are required to be integer constant expressions.
5895 unsigned ICEArguments = 0;
5896 ASTContext::GetBuiltinTypeError Error;
5897 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments);
5898 assert(Error == ASTContext::GE_None && "Should not codegen an error");
5899
5900 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) {
5901 // If this is a normal argument, just emit it as a scalar.
5902 if ((ICEArguments & (1 << i)) == 0) {
5903 Ops.push_back(EmitScalarExpr(E->getArg(i)));
5904 continue;
5905 }
5906
5907 // If this is required to be a constant, constant fold it so that we know
5908 // that the generated intrinsic gets a ConstantInt.
5909 llvm::APSInt Result;
5910 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext());
5911 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst;
5912 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result));
5913 }
5914
5915 switch (BuiltinID) {
5916 default: return nullptr;
5917 case X86::BI_mm_prefetch: {
5918 Value *Address = EmitScalarExpr(E->getArg(0));
5919 Value *RW = ConstantInt::get(Int32Ty, 0);
5920 Value *Locality = EmitScalarExpr(E->getArg(1));
5921 Value *Data = ConstantInt::get(Int32Ty, 1);
5922 Value *F = CGM.getIntrinsic(Intrinsic::prefetch);
5923 return Builder.CreateCall4(F, Address, RW, Locality, Data);
5924 }
5925 case X86::BI__builtin_ia32_vec_init_v8qi:
5926 case X86::BI__builtin_ia32_vec_init_v4hi:
5927 case X86::BI__builtin_ia32_vec_init_v2si:
5928 return Builder.CreateBitCast(BuildVector(Ops),
5929 llvm::Type::getX86_MMXTy(getLLVMContext()));
5930 case X86::BI__builtin_ia32_vec_ext_v2si:
5931 return Builder.CreateExtractElement(Ops[0],
5932 llvm::ConstantInt::get(Ops[1]->getType(), 0));
5933 case X86::BI__builtin_ia32_ldmxcsr: {
5934 Value *Tmp = CreateMemTemp(E->getArg(0)->getType());
5935 Builder.CreateStore(Ops[0], Tmp);
5936 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr),
5937 Builder.CreateBitCast(Tmp, Int8PtrTy));
5938 }
5939 case X86::BI__builtin_ia32_stmxcsr: {
5940 Value *Tmp = CreateMemTemp(E->getType());
5941 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr),
5942 Builder.CreateBitCast(Tmp, Int8PtrTy));
5943 return Builder.CreateLoad(Tmp, "stmxcsr");
5944 }
5945 case X86::BI__builtin_ia32_storehps:
5946 case X86::BI__builtin_ia32_storelps: {
5947 llvm::Type *PtrTy = llvm::PointerType::getUnqual(Int64Ty);
5948 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2);
5949
5950 // cast val v2i64
5951 Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast");
5952
5953 // extract (0, 1)
5954 unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1;
5955 llvm::Value *Idx = llvm::ConstantInt::get(SizeTy, Index);
5956 Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract");
5957
5958 // cast pointer to i64 & store
5959 Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy);
5960 return Builder.CreateStore(Ops[1], Ops[0]);
5961 }
5962 case X86::BI__builtin_ia32_palignr128:
5963 case X86::BI__builtin_ia32_palignr256: {
5964 unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue();
5965
5966 unsigned NumElts =
5967 cast<llvm::VectorType>(Ops[0]->getType())->getNumElements();
5968 assert(NumElts % 16 == 0);
5969 unsigned NumLanes = NumElts / 16;
5970 unsigned NumLaneElts = NumElts / NumLanes;
5971
5972 // If palignr is shifting the pair of vectors more than the size of two
5973 // lanes, emit zero.
5974 if (ShiftVal >= (2 * NumLaneElts))
5975 return llvm::Constant::getNullValue(ConvertType(E->getType()));
5976
5977 // If palignr is shifting the pair of input vectors more than one lane,
5978 // but less than two lanes, convert to shifting in zeroes.
5979 if (ShiftVal > NumLaneElts) {
5980 ShiftVal -= NumLaneElts;
5981 Ops[0] = llvm::Constant::getNullValue(Ops[0]->getType());
5982 }
5983
5984 SmallVector<llvm::Constant*, 32> Indices;
5985 // 256-bit palignr operates on 128-bit lanes so we need to handle that
5986 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
5987 for (unsigned i = 0; i != NumLaneElts; ++i) {
5988 unsigned Idx = ShiftVal + i;
5989 if (Idx >= NumLaneElts)
5990 Idx += NumElts - NumLaneElts; // End of lane, switch operand.
5991 Indices.push_back(llvm::ConstantInt::get(Int32Ty, Idx + l));
5992 }
5993 }
5994
5995 Value* SV = llvm::ConstantVector::get(Indices);
5996 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr");
5997 }
5998 case X86::BI__builtin_ia32_pslldqi256: {
5999 // Shift value is in bits so divide by 8.
6000 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() >> 3;
6001
6002 // If pslldq is shifting the vector more than 15 bytes, emit zero.
6003 if (shiftVal >= 16)
6004 return llvm::Constant::getNullValue(ConvertType(E->getType()));
6005
6006 SmallVector<llvm::Constant*, 32> Indices;
6007 // 256-bit pslldq operates on 128-bit lanes so we need to handle that
6008 for (unsigned l = 0; l != 32; l += 16) {
6009 for (unsigned i = 0; i != 16; ++i) {
6010 unsigned Idx = 32 + i - shiftVal;
6011 if (Idx < 32) Idx -= 16; // end of lane, switch operand.
6012 Indices.push_back(llvm::ConstantInt::get(Int32Ty, Idx + l));
6013 }
6014 }
6015
6016 llvm::Type *VecTy = llvm::VectorType::get(Int8Ty, 32);
6017 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast");
6018 Value *Zero = llvm::Constant::getNullValue(VecTy);
6019
6020 Value *SV = llvm::ConstantVector::get(Indices);
6021 SV = Builder.CreateShuffleVector(Zero, Ops[0], SV, "pslldq");
6022 llvm::Type *ResultType = ConvertType(E->getType());
6023 return Builder.CreateBitCast(SV, ResultType, "cast");
6024 }
6025 case X86::BI__builtin_ia32_psrldqi256: {
6026 // Shift value is in bits so divide by 8.
6027 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() >> 3;
6028
6029 // If psrldq is shifting the vector more than 15 bytes, emit zero.
6030 if (shiftVal >= 16)
6031 return llvm::Constant::getNullValue(ConvertType(E->getType()));
6032
6033 SmallVector<llvm::Constant*, 32> Indices;
6034 // 256-bit psrldq operates on 128-bit lanes so we need to handle that
6035 for (unsigned l = 0; l != 32; l += 16) {
6036 for (unsigned i = 0; i != 16; ++i) {
6037 unsigned Idx = i + shiftVal;
6038 if (Idx >= 16) Idx += 16; // end of lane, switch operand.
6039 Indices.push_back(llvm::ConstantInt::get(Int32Ty, Idx + l));
6040 }
6041 }
6042
6043 llvm::Type *VecTy = llvm::VectorType::get(Int8Ty, 32);
6044 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast");
6045 Value *Zero = llvm::Constant::getNullValue(VecTy);
6046
6047 Value *SV = llvm::ConstantVector::get(Indices);
6048 SV = Builder.CreateShuffleVector(Ops[0], Zero, SV, "psrldq");
6049 llvm::Type *ResultType = ConvertType(E->getType());
6050 return Builder.CreateBitCast(SV, ResultType, "cast");
6051 }
6052 case X86::BI__builtin_ia32_movntps:
6053 case X86::BI__builtin_ia32_movntps256:
6054 case X86::BI__builtin_ia32_movntpd:
6055 case X86::BI__builtin_ia32_movntpd256:
6056 case X86::BI__builtin_ia32_movntdq:
6057 case X86::BI__builtin_ia32_movntdq256:
6058 case X86::BI__builtin_ia32_movnti:
6059 case X86::BI__builtin_ia32_movnti64: {
6060 llvm::MDNode *Node = llvm::MDNode::get(
6061 getLLVMContext(), llvm::ConstantAsMetadata::get(Builder.getInt32(1)));
6062
6063 // Convert the type of the pointer to a pointer to the stored type.
6064 Value *BC = Builder.CreateBitCast(Ops[0],
6065 llvm::PointerType::getUnqual(Ops[1]->getType()),
6066 "cast");
6067 StoreInst *SI = Builder.CreateStore(Ops[1], BC);
6068 SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node);
6069
6070 // If the operand is an integer, we can't assume alignment. Otherwise,
6071 // assume natural alignment.
6072 QualType ArgTy = E->getArg(1)->getType();
6073 unsigned Align;
6074 if (ArgTy->isIntegerType())
6075 Align = 1;
6076 else
6077 Align = getContext().getTypeSizeInChars(ArgTy).getQuantity();
6078 SI->setAlignment(Align);
6079 return SI;
6080 }
6081 // 3DNow!
6082 case X86::BI__builtin_ia32_pswapdsf:
6083 case X86::BI__builtin_ia32_pswapdsi: {
6084 llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext());
6085 Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast");
6086 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_3dnowa_pswapd);
6087 return Builder.CreateCall(F, Ops, "pswapd");
6088 }
6089 case X86::BI__builtin_ia32_rdrand16_step:
6090 case X86::BI__builtin_ia32_rdrand32_step:
6091 case X86::BI__builtin_ia32_rdrand64_step:
6092 case X86::BI__builtin_ia32_rdseed16_step:
6093 case X86::BI__builtin_ia32_rdseed32_step:
6094 case X86::BI__builtin_ia32_rdseed64_step: {
6095 Intrinsic::ID ID;
6096 switch (BuiltinID) {
6097 default: llvm_unreachable("Unsupported intrinsic!");
6098 case X86::BI__builtin_ia32_rdrand16_step:
6099 ID = Intrinsic::x86_rdrand_16;
6100 break;
6101 case X86::BI__builtin_ia32_rdrand32_step:
6102 ID = Intrinsic::x86_rdrand_32;
6103 break;
6104 case X86::BI__builtin_ia32_rdrand64_step:
6105 ID = Intrinsic::x86_rdrand_64;
6106 break;
6107 case X86::BI__builtin_ia32_rdseed16_step:
6108 ID = Intrinsic::x86_rdseed_16;
6109 break;
6110 case X86::BI__builtin_ia32_rdseed32_step:
6111 ID = Intrinsic::x86_rdseed_32;
6112 break;
6113 case X86::BI__builtin_ia32_rdseed64_step:
6114 ID = Intrinsic::x86_rdseed_64;
6115 break;
6116 }
6117
6118 Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID));
6119 Builder.CreateStore(Builder.CreateExtractValue(Call, 0), Ops[0]);
6120 return Builder.CreateExtractValue(Call, 1);
6121 }
6122 // SSE comparison intrisics
6123 case X86::BI__builtin_ia32_cmpeqps:
6124 case X86::BI__builtin_ia32_cmpltps:
6125 case X86::BI__builtin_ia32_cmpleps:
6126 case X86::BI__builtin_ia32_cmpunordps:
6127 case X86::BI__builtin_ia32_cmpneqps:
6128 case X86::BI__builtin_ia32_cmpnltps:
6129 case X86::BI__builtin_ia32_cmpnleps:
6130 case X86::BI__builtin_ia32_cmpordps:
6131 case X86::BI__builtin_ia32_cmpeqss:
6132 case X86::BI__builtin_ia32_cmpltss:
6133 case X86::BI__builtin_ia32_cmpless:
6134 case X86::BI__builtin_ia32_cmpunordss:
6135 case X86::BI__builtin_ia32_cmpneqss:
6136 case X86::BI__builtin_ia32_cmpnltss:
6137 case X86::BI__builtin_ia32_cmpnless:
6138 case X86::BI__builtin_ia32_cmpordss:
6139 case X86::BI__builtin_ia32_cmpeqpd:
6140 case X86::BI__builtin_ia32_cmpltpd:
6141 case X86::BI__builtin_ia32_cmplepd:
6142 case X86::BI__builtin_ia32_cmpunordpd:
6143 case X86::BI__builtin_ia32_cmpneqpd:
6144 case X86::BI__builtin_ia32_cmpnltpd:
6145 case X86::BI__builtin_ia32_cmpnlepd:
6146 case X86::BI__builtin_ia32_cmpordpd:
6147 case X86::BI__builtin_ia32_cmpeqsd:
6148 case X86::BI__builtin_ia32_cmpltsd:
6149 case X86::BI__builtin_ia32_cmplesd:
6150 case X86::BI__builtin_ia32_cmpunordsd:
6151 case X86::BI__builtin_ia32_cmpneqsd:
6152 case X86::BI__builtin_ia32_cmpnltsd:
6153 case X86::BI__builtin_ia32_cmpnlesd:
6154 case X86::BI__builtin_ia32_cmpordsd:
6155 // These exist so that the builtin that takes an immediate can be bounds
6156 // checked by clang to avoid passing bad immediates to the backend. Since
6157 // AVX has a larger immediate than SSE we would need separate builtins to
6158 // do the different bounds checking. Rather than create a clang specific
6159 // SSE only builtin, this implements eight separate builtins to match gcc
6160 // implementation.
6161
6162 // Choose the immediate.
6163 unsigned Imm;
6164 switch (BuiltinID) {
6165 default: llvm_unreachable("Unsupported intrinsic!");
6166 case X86::BI__builtin_ia32_cmpeqps:
6167 case X86::BI__builtin_ia32_cmpeqss:
6168 case X86::BI__builtin_ia32_cmpeqpd:
6169 case X86::BI__builtin_ia32_cmpeqsd:
6170 Imm = 0;
6171 break;
6172 case X86::BI__builtin_ia32_cmpltps:
6173 case X86::BI__builtin_ia32_cmpltss:
6174 case X86::BI__builtin_ia32_cmpltpd:
6175 case X86::BI__builtin_ia32_cmpltsd:
6176 Imm = 1;
6177 break;
6178 case X86::BI__builtin_ia32_cmpleps:
6179 case X86::BI__builtin_ia32_cmpless:
6180 case X86::BI__builtin_ia32_cmplepd:
6181 case X86::BI__builtin_ia32_cmplesd:
6182 Imm = 2;
6183 break;
6184 case X86::BI__builtin_ia32_cmpunordps:
6185 case X86::BI__builtin_ia32_cmpunordss:
6186 case X86::BI__builtin_ia32_cmpunordpd:
6187 case X86::BI__builtin_ia32_cmpunordsd:
6188 Imm = 3;
6189 break;
6190 case X86::BI__builtin_ia32_cmpneqps:
6191 case X86::BI__builtin_ia32_cmpneqss:
6192 case X86::BI__builtin_ia32_cmpneqpd:
6193 case X86::BI__builtin_ia32_cmpneqsd:
6194 Imm = 4;
6195 break;
6196 case X86::BI__builtin_ia32_cmpnltps:
6197 case X86::BI__builtin_ia32_cmpnltss:
6198 case X86::BI__builtin_ia32_cmpnltpd:
6199 case X86::BI__builtin_ia32_cmpnltsd:
6200 Imm = 5;
6201 break;
6202 case X86::BI__builtin_ia32_cmpnleps:
6203 case X86::BI__builtin_ia32_cmpnless:
6204 case X86::BI__builtin_ia32_cmpnlepd:
6205 case X86::BI__builtin_ia32_cmpnlesd:
6206 Imm = 6;
6207 break;
6208 case X86::BI__builtin_ia32_cmpordps:
6209 case X86::BI__builtin_ia32_cmpordss:
6210 case X86::BI__builtin_ia32_cmpordpd:
6211 case X86::BI__builtin_ia32_cmpordsd:
6212 Imm = 7;
6213 break;
6214 }
6215
6216 // Choose the intrinsic ID.
6217 const char *name;
6218 Intrinsic::ID ID;
6219 switch (BuiltinID) {
6220 default: llvm_unreachable("Unsupported intrinsic!");
6221 case X86::BI__builtin_ia32_cmpeqps:
6222 case X86::BI__builtin_ia32_cmpltps:
6223 case X86::BI__builtin_ia32_cmpleps:
6224 case X86::BI__builtin_ia32_cmpunordps:
6225 case X86::BI__builtin_ia32_cmpneqps:
6226 case X86::BI__builtin_ia32_cmpnltps:
6227 case X86::BI__builtin_ia32_cmpnleps:
6228 case X86::BI__builtin_ia32_cmpordps:
6229 name = "cmpps";
6230 ID = Intrinsic::x86_sse_cmp_ps;
6231 break;
6232 case X86::BI__builtin_ia32_cmpeqss:
6233 case X86::BI__builtin_ia32_cmpltss:
6234 case X86::BI__builtin_ia32_cmpless:
6235 case X86::BI__builtin_ia32_cmpunordss:
6236 case X86::BI__builtin_ia32_cmpneqss:
6237 case X86::BI__builtin_ia32_cmpnltss:
6238 case X86::BI__builtin_ia32_cmpnless:
6239 case X86::BI__builtin_ia32_cmpordss:
6240 name = "cmpss";
6241 ID = Intrinsic::x86_sse_cmp_ss;
6242 break;
6243 case X86::BI__builtin_ia32_cmpeqpd:
6244 case X86::BI__builtin_ia32_cmpltpd:
6245 case X86::BI__builtin_ia32_cmplepd:
6246 case X86::BI__builtin_ia32_cmpunordpd:
6247 case X86::BI__builtin_ia32_cmpneqpd:
6248 case X86::BI__builtin_ia32_cmpnltpd:
6249 case X86::BI__builtin_ia32_cmpnlepd:
6250 case X86::BI__builtin_ia32_cmpordpd:
6251 name = "cmppd";
6252 ID = Intrinsic::x86_sse2_cmp_pd;
6253 break;
6254 case X86::BI__builtin_ia32_cmpeqsd:
6255 case X86::BI__builtin_ia32_cmpltsd:
6256 case X86::BI__builtin_ia32_cmplesd:
6257 case X86::BI__builtin_ia32_cmpunordsd:
6258 case X86::BI__builtin_ia32_cmpneqsd:
6259 case X86::BI__builtin_ia32_cmpnltsd:
6260 case X86::BI__builtin_ia32_cmpnlesd:
6261 case X86::BI__builtin_ia32_cmpordsd:
6262 name = "cmpsd";
6263 ID = Intrinsic::x86_sse2_cmp_sd;
6264 break;
6265 }
6266
6267 Ops.push_back(llvm::ConstantInt::get(Int8Ty, Imm));
6268 llvm::Function *F = CGM.getIntrinsic(ID);
6269 return Builder.CreateCall(F, Ops, name);
6270 }
6271 }
6272
6273
EmitPPCBuiltinExpr(unsigned BuiltinID,const CallExpr * E)6274 Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID,
6275 const CallExpr *E) {
6276 SmallVector<Value*, 4> Ops;
6277
6278 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++)
6279 Ops.push_back(EmitScalarExpr(E->getArg(i)));
6280
6281 Intrinsic::ID ID = Intrinsic::not_intrinsic;
6282
6283 switch (BuiltinID) {
6284 default: return nullptr;
6285
6286 // vec_ld, vec_lvsl, vec_lvsr
6287 case PPC::BI__builtin_altivec_lvx:
6288 case PPC::BI__builtin_altivec_lvxl:
6289 case PPC::BI__builtin_altivec_lvebx:
6290 case PPC::BI__builtin_altivec_lvehx:
6291 case PPC::BI__builtin_altivec_lvewx:
6292 case PPC::BI__builtin_altivec_lvsl:
6293 case PPC::BI__builtin_altivec_lvsr:
6294 case PPC::BI__builtin_vsx_lxvd2x:
6295 case PPC::BI__builtin_vsx_lxvw4x:
6296 {
6297 Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy);
6298
6299 Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]);
6300 Ops.pop_back();
6301
6302 switch (BuiltinID) {
6303 default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!");
6304 case PPC::BI__builtin_altivec_lvx:
6305 ID = Intrinsic::ppc_altivec_lvx;
6306 break;
6307 case PPC::BI__builtin_altivec_lvxl:
6308 ID = Intrinsic::ppc_altivec_lvxl;
6309 break;
6310 case PPC::BI__builtin_altivec_lvebx:
6311 ID = Intrinsic::ppc_altivec_lvebx;
6312 break;
6313 case PPC::BI__builtin_altivec_lvehx:
6314 ID = Intrinsic::ppc_altivec_lvehx;
6315 break;
6316 case PPC::BI__builtin_altivec_lvewx:
6317 ID = Intrinsic::ppc_altivec_lvewx;
6318 break;
6319 case PPC::BI__builtin_altivec_lvsl:
6320 ID = Intrinsic::ppc_altivec_lvsl;
6321 break;
6322 case PPC::BI__builtin_altivec_lvsr:
6323 ID = Intrinsic::ppc_altivec_lvsr;
6324 break;
6325 case PPC::BI__builtin_vsx_lxvd2x:
6326 ID = Intrinsic::ppc_vsx_lxvd2x;
6327 break;
6328 case PPC::BI__builtin_vsx_lxvw4x:
6329 ID = Intrinsic::ppc_vsx_lxvw4x;
6330 break;
6331 }
6332 llvm::Function *F = CGM.getIntrinsic(ID);
6333 return Builder.CreateCall(F, Ops, "");
6334 }
6335
6336 // vec_st
6337 case PPC::BI__builtin_altivec_stvx:
6338 case PPC::BI__builtin_altivec_stvxl:
6339 case PPC::BI__builtin_altivec_stvebx:
6340 case PPC::BI__builtin_altivec_stvehx:
6341 case PPC::BI__builtin_altivec_stvewx:
6342 case PPC::BI__builtin_vsx_stxvd2x:
6343 case PPC::BI__builtin_vsx_stxvw4x:
6344 {
6345 Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy);
6346 Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]);
6347 Ops.pop_back();
6348
6349 switch (BuiltinID) {
6350 default: llvm_unreachable("Unsupported st intrinsic!");
6351 case PPC::BI__builtin_altivec_stvx:
6352 ID = Intrinsic::ppc_altivec_stvx;
6353 break;
6354 case PPC::BI__builtin_altivec_stvxl:
6355 ID = Intrinsic::ppc_altivec_stvxl;
6356 break;
6357 case PPC::BI__builtin_altivec_stvebx:
6358 ID = Intrinsic::ppc_altivec_stvebx;
6359 break;
6360 case PPC::BI__builtin_altivec_stvehx:
6361 ID = Intrinsic::ppc_altivec_stvehx;
6362 break;
6363 case PPC::BI__builtin_altivec_stvewx:
6364 ID = Intrinsic::ppc_altivec_stvewx;
6365 break;
6366 case PPC::BI__builtin_vsx_stxvd2x:
6367 ID = Intrinsic::ppc_vsx_stxvd2x;
6368 break;
6369 case PPC::BI__builtin_vsx_stxvw4x:
6370 ID = Intrinsic::ppc_vsx_stxvw4x;
6371 break;
6372 }
6373 llvm::Function *F = CGM.getIntrinsic(ID);
6374 return Builder.CreateCall(F, Ops, "");
6375 }
6376 }
6377 }
6378
6379 // Emit an intrinsic that has 1 float or double.
emitUnaryFPBuiltin(CodeGenFunction & CGF,const CallExpr * E,unsigned IntrinsicID)6380 static Value *emitUnaryFPBuiltin(CodeGenFunction &CGF,
6381 const CallExpr *E,
6382 unsigned IntrinsicID) {
6383 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0));
6384
6385 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType());
6386 return CGF.Builder.CreateCall(F, Src0);
6387 }
6388
6389 // Emit an intrinsic that has 3 float or double operands.
emitTernaryFPBuiltin(CodeGenFunction & CGF,const CallExpr * E,unsigned IntrinsicID)6390 static Value *emitTernaryFPBuiltin(CodeGenFunction &CGF,
6391 const CallExpr *E,
6392 unsigned IntrinsicID) {
6393 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0));
6394 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1));
6395 llvm::Value *Src2 = CGF.EmitScalarExpr(E->getArg(2));
6396
6397 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType());
6398 return CGF.Builder.CreateCall3(F, Src0, Src1, Src2);
6399 }
6400
6401 // Emit an intrinsic that has 1 float or double operand, and 1 integer.
emitFPIntBuiltin(CodeGenFunction & CGF,const CallExpr * E,unsigned IntrinsicID)6402 static Value *emitFPIntBuiltin(CodeGenFunction &CGF,
6403 const CallExpr *E,
6404 unsigned IntrinsicID) {
6405 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0));
6406 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1));
6407
6408 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType());
6409 return CGF.Builder.CreateCall2(F, Src0, Src1);
6410 }
6411
EmitR600BuiltinExpr(unsigned BuiltinID,const CallExpr * E)6412 Value *CodeGenFunction::EmitR600BuiltinExpr(unsigned BuiltinID,
6413 const CallExpr *E) {
6414 switch (BuiltinID) {
6415 case R600::BI__builtin_amdgpu_div_scale:
6416 case R600::BI__builtin_amdgpu_div_scalef: {
6417 // Translate from the intrinsics's struct return to the builtin's out
6418 // argument.
6419
6420 std::pair<llvm::Value *, unsigned> FlagOutPtr
6421 = EmitPointerWithAlignment(E->getArg(3));
6422
6423 llvm::Value *X = EmitScalarExpr(E->getArg(0));
6424 llvm::Value *Y = EmitScalarExpr(E->getArg(1));
6425 llvm::Value *Z = EmitScalarExpr(E->getArg(2));
6426
6427 llvm::Value *Callee = CGM.getIntrinsic(Intrinsic::AMDGPU_div_scale,
6428 X->getType());
6429
6430 llvm::Value *Tmp = Builder.CreateCall3(Callee, X, Y, Z);
6431
6432 llvm::Value *Result = Builder.CreateExtractValue(Tmp, 0);
6433 llvm::Value *Flag = Builder.CreateExtractValue(Tmp, 1);
6434
6435 llvm::Type *RealFlagType
6436 = FlagOutPtr.first->getType()->getPointerElementType();
6437
6438 llvm::Value *FlagExt = Builder.CreateZExt(Flag, RealFlagType);
6439 llvm::StoreInst *FlagStore = Builder.CreateStore(FlagExt, FlagOutPtr.first);
6440 FlagStore->setAlignment(FlagOutPtr.second);
6441 return Result;
6442 }
6443 case R600::BI__builtin_amdgpu_div_fmas:
6444 case R600::BI__builtin_amdgpu_div_fmasf: {
6445 llvm::Value *Src0 = EmitScalarExpr(E->getArg(0));
6446 llvm::Value *Src1 = EmitScalarExpr(E->getArg(1));
6447 llvm::Value *Src2 = EmitScalarExpr(E->getArg(2));
6448 llvm::Value *Src3 = EmitScalarExpr(E->getArg(3));
6449
6450 llvm::Value *F = CGM.getIntrinsic(Intrinsic::AMDGPU_div_fmas,
6451 Src0->getType());
6452 llvm::Value *Src3ToBool = Builder.CreateIsNotNull(Src3);
6453 return Builder.CreateCall4(F, Src0, Src1, Src2, Src3ToBool);
6454 }
6455 case R600::BI__builtin_amdgpu_div_fixup:
6456 case R600::BI__builtin_amdgpu_div_fixupf:
6457 return emitTernaryFPBuiltin(*this, E, Intrinsic::AMDGPU_div_fixup);
6458 case R600::BI__builtin_amdgpu_trig_preop:
6459 case R600::BI__builtin_amdgpu_trig_preopf:
6460 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_trig_preop);
6461 case R600::BI__builtin_amdgpu_rcp:
6462 case R600::BI__builtin_amdgpu_rcpf:
6463 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rcp);
6464 case R600::BI__builtin_amdgpu_rsq:
6465 case R600::BI__builtin_amdgpu_rsqf:
6466 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rsq);
6467 case R600::BI__builtin_amdgpu_rsq_clamped:
6468 case R600::BI__builtin_amdgpu_rsq_clampedf:
6469 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rsq_clamped);
6470 case R600::BI__builtin_amdgpu_ldexp:
6471 case R600::BI__builtin_amdgpu_ldexpf:
6472 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_ldexp);
6473 case R600::BI__builtin_amdgpu_class:
6474 case R600::BI__builtin_amdgpu_classf:
6475 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_class);
6476 default:
6477 return nullptr;
6478 }
6479 }
6480
EmitSystemZBuiltinExpr(unsigned BuiltinID,const CallExpr * E)6481 Value *CodeGenFunction::EmitSystemZBuiltinExpr(unsigned BuiltinID,
6482 const CallExpr *E) {
6483 switch (BuiltinID) {
6484 case SystemZ::BI__builtin_tbegin: {
6485 Value *TDB = EmitScalarExpr(E->getArg(0));
6486 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c);
6487 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbegin);
6488 return Builder.CreateCall2(F, TDB, Control);
6489 }
6490 case SystemZ::BI__builtin_tbegin_nofloat: {
6491 Value *TDB = EmitScalarExpr(E->getArg(0));
6492 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c);
6493 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbegin_nofloat);
6494 return Builder.CreateCall2(F, TDB, Control);
6495 }
6496 case SystemZ::BI__builtin_tbeginc: {
6497 Value *TDB = llvm::ConstantPointerNull::get(Int8PtrTy);
6498 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff08);
6499 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbeginc);
6500 return Builder.CreateCall2(F, TDB, Control);
6501 }
6502 case SystemZ::BI__builtin_tabort: {
6503 Value *Data = EmitScalarExpr(E->getArg(0));
6504 Value *F = CGM.getIntrinsic(Intrinsic::s390_tabort);
6505 return Builder.CreateCall(F, Builder.CreateSExt(Data, Int64Ty, "tabort"));
6506 }
6507 case SystemZ::BI__builtin_non_tx_store: {
6508 Value *Address = EmitScalarExpr(E->getArg(0));
6509 Value *Data = EmitScalarExpr(E->getArg(1));
6510 Value *F = CGM.getIntrinsic(Intrinsic::s390_ntstg);
6511 return Builder.CreateCall2(F, Data, Address);
6512 }
6513
6514 default:
6515 return nullptr;
6516 }
6517 }
6518