/Users/buildslave/jenkins/sharedspace/clang-stage2-coverage-R@2/llvm/tools/clang/lib/CodeGen/CGBuiltin.cpp
Line | Count | Source (jump to first uncovered line) |
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 "CGCXXABI.h" |
15 | | #include "CGObjCRuntime.h" |
16 | | #include "CGOpenCLRuntime.h" |
17 | | #include "CodeGenFunction.h" |
18 | | #include "CodeGenModule.h" |
19 | | #include "ConstantEmitter.h" |
20 | | #include "TargetInfo.h" |
21 | | #include "clang/AST/ASTContext.h" |
22 | | #include "clang/AST/Decl.h" |
23 | | #include "clang/Analysis/Analyses/OSLog.h" |
24 | | #include "clang/Basic/TargetBuiltins.h" |
25 | | #include "clang/Basic/TargetInfo.h" |
26 | | #include "clang/CodeGen/CGFunctionInfo.h" |
27 | | #include "llvm/ADT/StringExtras.h" |
28 | | #include "llvm/IR/CallSite.h" |
29 | | #include "llvm/IR/DataLayout.h" |
30 | | #include "llvm/IR/InlineAsm.h" |
31 | | #include "llvm/IR/Intrinsics.h" |
32 | | #include "llvm/IR/MDBuilder.h" |
33 | | #include "llvm/Support/ConvertUTF.h" |
34 | | #include <sstream> |
35 | | |
36 | | using namespace clang; |
37 | | using namespace CodeGen; |
38 | | using namespace llvm; |
39 | | |
40 | | static |
41 | 8 | int64_t clamp(int64_t Value, int64_t Low, int64_t High) { |
42 | 8 | return std::min(High, std::max(Low, Value)); |
43 | 8 | } |
44 | | |
45 | | /// getBuiltinLibFunction - Given a builtin id for a function like |
46 | | /// "__builtin_fabsf", return a Function* for "fabsf". |
47 | | llvm::Constant *CodeGenModule::getBuiltinLibFunction(const FunctionDecl *FD, |
48 | 5.97k | unsigned BuiltinID) { |
49 | 5.97k | assert(Context.BuiltinInfo.isLibFunction(BuiltinID)); |
50 | 5.97k | |
51 | 5.97k | // Get the name, skip over the __builtin_ prefix (if necessary). |
52 | 5.97k | StringRef Name; |
53 | 5.97k | GlobalDecl D(FD); |
54 | 5.97k | |
55 | 5.97k | // If the builtin has been declared explicitly with an assembler label, |
56 | 5.97k | // use the mangled name. This differs from the plain label on platforms |
57 | 5.97k | // that prefix labels. |
58 | 5.97k | if (FD->hasAttr<AsmLabelAttr>()) |
59 | 0 | Name = getMangledName(D); |
60 | 5.97k | else |
61 | 5.97k | Name = Context.BuiltinInfo.getName(BuiltinID) + 10; |
62 | 5.97k | |
63 | 5.97k | llvm::FunctionType *Ty = |
64 | 5.97k | cast<llvm::FunctionType>(getTypes().ConvertType(FD->getType())); |
65 | 5.97k | |
66 | 5.97k | return GetOrCreateLLVMFunction(Name, Ty, D, /*ForVTable=*/false); |
67 | 5.97k | } |
68 | | |
69 | | /// Emit the conversions required to turn the given value into an |
70 | | /// integer of the given size. |
71 | | static Value *EmitToInt(CodeGenFunction &CGF, llvm::Value *V, |
72 | 723 | QualType T, llvm::IntegerType *IntType) { |
73 | 723 | V = CGF.EmitToMemory(V, T); |
74 | 723 | |
75 | 723 | if (V->getType()->isPointerTy()) |
76 | 69 | return CGF.Builder.CreatePtrToInt(V, IntType); |
77 | 654 | |
78 | 723 | assert(V->getType() == IntType); |
79 | 654 | return V; |
80 | 654 | } |
81 | | |
82 | | static Value *EmitFromInt(CodeGenFunction &CGF, llvm::Value *V, |
83 | 593 | QualType T, llvm::Type *ResultType) { |
84 | 593 | V = CGF.EmitFromMemory(V, T); |
85 | 593 | |
86 | 593 | if (ResultType->isPointerTy()) |
87 | 9 | return CGF.Builder.CreateIntToPtr(V, ResultType); |
88 | 584 | |
89 | 593 | assert(V->getType() == ResultType); |
90 | 584 | return V; |
91 | 584 | } |
92 | | |
93 | | /// Utility to insert an atomic instruction based on Instrinsic::ID |
94 | | /// and the expression node. |
95 | | static Value *MakeBinaryAtomicValue(CodeGenFunction &CGF, |
96 | | llvm::AtomicRMWInst::BinOp Kind, |
97 | 482 | const CallExpr *E) { |
98 | 482 | QualType T = E->getType(); |
99 | 482 | assert(E->getArg(0)->getType()->isPointerType()); |
100 | 482 | assert(CGF.getContext().hasSameUnqualifiedType(T, |
101 | 482 | E->getArg(0)->getType()->getPointeeType())); |
102 | 482 | assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType())); |
103 | 482 | |
104 | 482 | llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); |
105 | 482 | unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); |
106 | 482 | |
107 | 482 | llvm::IntegerType *IntType = |
108 | 482 | llvm::IntegerType::get(CGF.getLLVMContext(), |
109 | 482 | CGF.getContext().getTypeSize(T)); |
110 | 482 | llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); |
111 | 482 | |
112 | 482 | llvm::Value *Args[2]; |
113 | 482 | Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); |
114 | 482 | Args[1] = CGF.EmitScalarExpr(E->getArg(1)); |
115 | 482 | llvm::Type *ValueType = Args[1]->getType(); |
116 | 482 | Args[1] = EmitToInt(CGF, Args[1], T, IntType); |
117 | 482 | |
118 | 482 | llvm::Value *Result = CGF.Builder.CreateAtomicRMW( |
119 | 482 | Kind, Args[0], Args[1], llvm::AtomicOrdering::SequentiallyConsistent); |
120 | 482 | return EmitFromInt(CGF, Result, T, ValueType); |
121 | 482 | } |
122 | | |
123 | 64 | static Value *EmitNontemporalStore(CodeGenFunction &CGF, const CallExpr *E) { |
124 | 64 | Value *Val = CGF.EmitScalarExpr(E->getArg(0)); |
125 | 64 | Value *Address = CGF.EmitScalarExpr(E->getArg(1)); |
126 | 64 | |
127 | 64 | // Convert the type of the pointer to a pointer to the stored type. |
128 | 64 | Val = CGF.EmitToMemory(Val, E->getArg(0)->getType()); |
129 | 64 | Value *BC = CGF.Builder.CreateBitCast( |
130 | 64 | Address, llvm::PointerType::getUnqual(Val->getType()), "cast"); |
131 | 64 | LValue LV = CGF.MakeNaturalAlignAddrLValue(BC, E->getArg(0)->getType()); |
132 | 64 | LV.setNontemporal(true); |
133 | 64 | CGF.EmitStoreOfScalar(Val, LV, false); |
134 | 64 | return nullptr; |
135 | 64 | } |
136 | | |
137 | 30 | static Value *EmitNontemporalLoad(CodeGenFunction &CGF, const CallExpr *E) { |
138 | 30 | Value *Address = CGF.EmitScalarExpr(E->getArg(0)); |
139 | 30 | |
140 | 30 | LValue LV = CGF.MakeNaturalAlignAddrLValue(Address, E->getType()); |
141 | 30 | LV.setNontemporal(true); |
142 | 30 | return CGF.EmitLoadOfScalar(LV, E->getExprLoc()); |
143 | 30 | } |
144 | | |
145 | | static RValue EmitBinaryAtomic(CodeGenFunction &CGF, |
146 | | llvm::AtomicRMWInst::BinOp Kind, |
147 | 311 | const CallExpr *E) { |
148 | 311 | return RValue::get(MakeBinaryAtomicValue(CGF, Kind, E)); |
149 | 311 | } |
150 | | |
151 | | /// Utility to insert an atomic instruction based Instrinsic::ID and |
152 | | /// the expression node, where the return value is the result of the |
153 | | /// operation. |
154 | | static RValue EmitBinaryAtomicPost(CodeGenFunction &CGF, |
155 | | llvm::AtomicRMWInst::BinOp Kind, |
156 | | const CallExpr *E, |
157 | | Instruction::BinaryOps Op, |
158 | 59 | bool Invert = false) { |
159 | 59 | QualType T = E->getType(); |
160 | 59 | assert(E->getArg(0)->getType()->isPointerType()); |
161 | 59 | assert(CGF.getContext().hasSameUnqualifiedType(T, |
162 | 59 | E->getArg(0)->getType()->getPointeeType())); |
163 | 59 | assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType())); |
164 | 59 | |
165 | 59 | llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); |
166 | 59 | unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); |
167 | 59 | |
168 | 59 | llvm::IntegerType *IntType = |
169 | 59 | llvm::IntegerType::get(CGF.getLLVMContext(), |
170 | 59 | CGF.getContext().getTypeSize(T)); |
171 | 59 | llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); |
172 | 59 | |
173 | 59 | llvm::Value *Args[2]; |
174 | 59 | Args[1] = CGF.EmitScalarExpr(E->getArg(1)); |
175 | 59 | llvm::Type *ValueType = Args[1]->getType(); |
176 | 59 | Args[1] = EmitToInt(CGF, Args[1], T, IntType); |
177 | 59 | Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); |
178 | 59 | |
179 | 59 | llvm::Value *Result = CGF.Builder.CreateAtomicRMW( |
180 | 59 | Kind, Args[0], Args[1], llvm::AtomicOrdering::SequentiallyConsistent); |
181 | 59 | Result = CGF.Builder.CreateBinOp(Op, Result, Args[1]); |
182 | 59 | if (Invert) |
183 | 9 | Result = CGF.Builder.CreateBinOp(llvm::Instruction::Xor, Result, |
184 | 9 | llvm::ConstantInt::get(IntType, -1)); |
185 | 59 | Result = EmitFromInt(CGF, Result, T, ValueType); |
186 | 59 | return RValue::get(Result); |
187 | 59 | } |
188 | | |
189 | | /// @brief Utility to insert an atomic cmpxchg instruction. |
190 | | /// |
191 | | /// @param CGF The current codegen function. |
192 | | /// @param E Builtin call expression to convert to cmpxchg. |
193 | | /// arg0 - address to operate on |
194 | | /// arg1 - value to compare with |
195 | | /// arg2 - new value |
196 | | /// @param ReturnBool Specifies whether to return success flag of |
197 | | /// cmpxchg result or the old value. |
198 | | /// |
199 | | /// @returns result of cmpxchg, according to ReturnBool |
200 | | static Value *MakeAtomicCmpXchgValue(CodeGenFunction &CGF, const CallExpr *E, |
201 | 91 | bool ReturnBool) { |
202 | 91 | QualType T = ReturnBool ? E->getArg(1)->getType()39 : E->getType()52 ; |
203 | 91 | llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); |
204 | 91 | unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); |
205 | 91 | |
206 | 91 | llvm::IntegerType *IntType = llvm::IntegerType::get( |
207 | 91 | CGF.getLLVMContext(), CGF.getContext().getTypeSize(T)); |
208 | 91 | llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); |
209 | 91 | |
210 | 91 | Value *Args[3]; |
211 | 91 | Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); |
212 | 91 | Args[1] = CGF.EmitScalarExpr(E->getArg(1)); |
213 | 91 | llvm::Type *ValueType = Args[1]->getType(); |
214 | 91 | Args[1] = EmitToInt(CGF, Args[1], T, IntType); |
215 | 91 | Args[2] = EmitToInt(CGF, CGF.EmitScalarExpr(E->getArg(2)), T, IntType); |
216 | 91 | |
217 | 91 | Value *Pair = CGF.Builder.CreateAtomicCmpXchg( |
218 | 91 | Args[0], Args[1], Args[2], llvm::AtomicOrdering::SequentiallyConsistent, |
219 | 91 | llvm::AtomicOrdering::SequentiallyConsistent); |
220 | 91 | if (ReturnBool) |
221 | 91 | // Extract boolean success flag and zext it to int. |
222 | 39 | return CGF.Builder.CreateZExt(CGF.Builder.CreateExtractValue(Pair, 1), |
223 | 39 | CGF.ConvertType(E->getType())); |
224 | 91 | else |
225 | 91 | // Extract old value and emit it using the same type as compare value. |
226 | 52 | return EmitFromInt(CGF, CGF.Builder.CreateExtractValue(Pair, 0), T, |
227 | 52 | ValueType); |
228 | 0 | } |
229 | | |
230 | | // Emit a simple mangled intrinsic that has 1 argument and a return type |
231 | | // matching the argument type. |
232 | | static Value *emitUnaryBuiltin(CodeGenFunction &CGF, |
233 | | const CallExpr *E, |
234 | 326 | unsigned IntrinsicID) { |
235 | 326 | llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); |
236 | 326 | |
237 | 326 | Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); |
238 | 326 | return CGF.Builder.CreateCall(F, Src0); |
239 | 326 | } |
240 | | |
241 | | // Emit an intrinsic that has 2 operands of the same type as its result. |
242 | | static Value *emitBinaryBuiltin(CodeGenFunction &CGF, |
243 | | const CallExpr *E, |
244 | 1.61k | unsigned IntrinsicID) { |
245 | 1.61k | llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); |
246 | 1.61k | llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); |
247 | 1.61k | |
248 | 1.61k | Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); |
249 | 1.61k | return CGF.Builder.CreateCall(F, { Src0, Src1 }); |
250 | 1.61k | } |
251 | | |
252 | | // Emit an intrinsic that has 3 operands of the same type as its result. |
253 | | static Value *emitTernaryBuiltin(CodeGenFunction &CGF, |
254 | | const CallExpr *E, |
255 | 23 | unsigned IntrinsicID) { |
256 | 23 | llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); |
257 | 23 | llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); |
258 | 23 | llvm::Value *Src2 = CGF.EmitScalarExpr(E->getArg(2)); |
259 | 23 | |
260 | 23 | Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); |
261 | 23 | return CGF.Builder.CreateCall(F, { Src0, Src1, Src2 }); |
262 | 23 | } |
263 | | |
264 | | // Emit an intrinsic that has 1 float or double operand, and 1 integer. |
265 | | static Value *emitFPIntBuiltin(CodeGenFunction &CGF, |
266 | | const CallExpr *E, |
267 | 16 | unsigned IntrinsicID) { |
268 | 16 | llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); |
269 | 16 | llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); |
270 | 16 | |
271 | 16 | Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); |
272 | 16 | return CGF.Builder.CreateCall(F, {Src0, Src1}); |
273 | 16 | } |
274 | | |
275 | | /// EmitFAbs - Emit a call to @llvm.fabs(). |
276 | 2.08k | static Value *EmitFAbs(CodeGenFunction &CGF, Value *V) { |
277 | 2.08k | Value *F = CGF.CGM.getIntrinsic(Intrinsic::fabs, V->getType()); |
278 | 2.08k | llvm::CallInst *Call = CGF.Builder.CreateCall(F, V); |
279 | 2.08k | Call->setDoesNotAccessMemory(); |
280 | 2.08k | return Call; |
281 | 2.08k | } |
282 | | |
283 | | /// Emit the computation of the sign bit for a floating point value. Returns |
284 | | /// the i1 sign bit value. |
285 | 28 | static Value *EmitSignBit(CodeGenFunction &CGF, Value *V) { |
286 | 28 | LLVMContext &C = CGF.CGM.getLLVMContext(); |
287 | 28 | |
288 | 28 | llvm::Type *Ty = V->getType(); |
289 | 28 | int Width = Ty->getPrimitiveSizeInBits(); |
290 | 28 | llvm::Type *IntTy = llvm::IntegerType::get(C, Width); |
291 | 28 | V = CGF.Builder.CreateBitCast(V, IntTy); |
292 | 28 | if (Ty->isPPC_FP128Ty()28 ) { |
293 | 14 | // We want the sign bit of the higher-order double. The bitcast we just |
294 | 14 | // did works as if the double-double was stored to memory and then |
295 | 14 | // read as an i128. The "store" will put the higher-order double in the |
296 | 14 | // lower address in both little- and big-Endian modes, but the "load" |
297 | 14 | // will treat those bits as a different part of the i128: the low bits in |
298 | 14 | // little-Endian, the high bits in big-Endian. Therefore, on big-Endian |
299 | 14 | // we need to shift the high bits down to the low before truncating. |
300 | 14 | Width >>= 1; |
301 | 14 | if (CGF.getTarget().isBigEndian()14 ) { |
302 | 9 | Value *ShiftCst = llvm::ConstantInt::get(IntTy, Width); |
303 | 9 | V = CGF.Builder.CreateLShr(V, ShiftCst); |
304 | 9 | } |
305 | 14 | // We are truncating value in order to extract the higher-order |
306 | 14 | // double, which we will be using to extract the sign from. |
307 | 14 | IntTy = llvm::IntegerType::get(C, Width); |
308 | 14 | V = CGF.Builder.CreateTrunc(V, IntTy); |
309 | 14 | } |
310 | 28 | Value *Zero = llvm::Constant::getNullValue(IntTy); |
311 | 28 | return CGF.Builder.CreateICmpSLT(V, Zero); |
312 | 28 | } |
313 | | |
314 | | static RValue emitLibraryCall(CodeGenFunction &CGF, const FunctionDecl *FD, |
315 | 98.5k | const CallExpr *E, llvm::Constant *calleeValue) { |
316 | 98.5k | CGCallee callee = CGCallee::forDirect(calleeValue, FD); |
317 | 98.5k | return CGF.EmitCall(E->getCallee()->getType(), callee, E, ReturnValueSlot()); |
318 | 98.5k | } |
319 | | |
320 | | /// \brief Emit a call to llvm.{sadd,uadd,ssub,usub,smul,umul}.with.overflow.* |
321 | | /// depending on IntrinsicID. |
322 | | /// |
323 | | /// \arg CGF The current codegen function. |
324 | | /// \arg IntrinsicID The ID for the Intrinsic we wish to generate. |
325 | | /// \arg X The first argument to the llvm.*.with.overflow.*. |
326 | | /// \arg Y The second argument to the llvm.*.with.overflow.*. |
327 | | /// \arg Carry The carry returned by the llvm.*.with.overflow.*. |
328 | | /// \returns The result (i.e. sum/product) returned by the intrinsic. |
329 | | static llvm::Value *EmitOverflowIntrinsic(CodeGenFunction &CGF, |
330 | | const llvm::Intrinsic::ID IntrinsicID, |
331 | | llvm::Value *X, llvm::Value *Y, |
332 | 147 | llvm::Value *&Carry) { |
333 | 147 | // Make sure we have integers of the same width. |
334 | 147 | assert(X->getType() == Y->getType() && |
335 | 147 | "Arguments must be the same type. (Did you forget to make sure both " |
336 | 147 | "arguments have the same integer width?)"); |
337 | 147 | |
338 | 147 | llvm::Value *Callee = CGF.CGM.getIntrinsic(IntrinsicID, X->getType()); |
339 | 147 | llvm::Value *Tmp = CGF.Builder.CreateCall(Callee, {X, Y}); |
340 | 147 | Carry = CGF.Builder.CreateExtractValue(Tmp, 1); |
341 | 147 | return CGF.Builder.CreateExtractValue(Tmp, 0); |
342 | 147 | } |
343 | | |
344 | | static Value *emitRangedBuiltin(CodeGenFunction &CGF, |
345 | | unsigned IntrinsicID, |
346 | 9 | int low, int high) { |
347 | 9 | llvm::MDBuilder MDHelper(CGF.getLLVMContext()); |
348 | 9 | llvm::MDNode *RNode = MDHelper.createRange(APInt(32, low), APInt(32, high)); |
349 | 9 | Value *F = CGF.CGM.getIntrinsic(IntrinsicID, {}); |
350 | 9 | llvm::Instruction *Call = CGF.Builder.CreateCall(F); |
351 | 9 | Call->setMetadata(llvm::LLVMContext::MD_range, RNode); |
352 | 9 | return Call; |
353 | 9 | } |
354 | | |
355 | | namespace { |
356 | | struct WidthAndSignedness { |
357 | | unsigned Width; |
358 | | bool Signed; |
359 | | }; |
360 | | } |
361 | | |
362 | | static WidthAndSignedness |
363 | | getIntegerWidthAndSignedness(const clang::ASTContext &context, |
364 | 99 | const clang::QualType Type) { |
365 | 99 | assert(Type->isIntegerType() && "Given type is not an integer."); |
366 | 99 | unsigned Width = Type->isBooleanType() ? 118 : context.getTypeInfo(Type).Width81 ; |
367 | 99 | bool Signed = Type->isSignedIntegerType(); |
368 | 99 | return {Width, Signed}; |
369 | 99 | } |
370 | | |
371 | | // Given one or more integer types, this function produces an integer type that |
372 | | // encompasses them: any value in one of the given types could be expressed in |
373 | | // the encompassing type. |
374 | | static struct WidthAndSignedness |
375 | 33 | EncompassingIntegerType(ArrayRef<struct WidthAndSignedness> Types) { |
376 | 33 | assert(Types.size() > 0 && "Empty list of types."); |
377 | 33 | |
378 | 33 | // If any of the given types is signed, we must return a signed type. |
379 | 33 | bool Signed = false; |
380 | 99 | for (const auto &Type : Types) { |
381 | 99 | Signed |= Type.Signed; |
382 | 99 | } |
383 | 33 | |
384 | 33 | // The encompassing type must have a width greater than or equal to the width |
385 | 33 | // of the specified types. Aditionally, if the encompassing type is signed, |
386 | 33 | // its width must be strictly greater than the width of any unsigned types |
387 | 33 | // given. |
388 | 33 | unsigned Width = 0; |
389 | 99 | for (const auto &Type : Types) { |
390 | 45 | unsigned MinWidth = Type.Width + (Signed && !Type.Signed); |
391 | 99 | if (Width < MinWidth99 ) { |
392 | 36 | Width = MinWidth; |
393 | 36 | } |
394 | 99 | } |
395 | 33 | |
396 | 33 | return {Width, Signed}; |
397 | 33 | } |
398 | | |
399 | 650 | Value *CodeGenFunction::EmitVAStartEnd(Value *ArgValue, bool IsStart) { |
400 | 650 | llvm::Type *DestType = Int8PtrTy; |
401 | 650 | if (ArgValue->getType() != DestType) |
402 | 650 | ArgValue = |
403 | 650 | Builder.CreateBitCast(ArgValue, DestType, ArgValue->getName().data()); |
404 | 650 | |
405 | 650 | Intrinsic::ID inst = IsStart ? Intrinsic::vastart313 : Intrinsic::vaend337 ; |
406 | 650 | return Builder.CreateCall(CGM.getIntrinsic(inst), ArgValue); |
407 | 650 | } |
408 | | |
409 | | /// Checks if using the result of __builtin_object_size(p, @p From) in place of |
410 | | /// __builtin_object_size(p, @p To) is correct |
411 | 30 | static bool areBOSTypesCompatible(int From, int To) { |
412 | 30 | // Note: Our __builtin_object_size implementation currently treats Type=0 and |
413 | 30 | // Type=2 identically. Encoding this implementation detail here may make |
414 | 30 | // improving __builtin_object_size difficult in the future, so it's omitted. |
415 | 30 | return From == To || (From == 0 && 12 To == 13 ) || (From == 3 && 11 To == 23 ); |
416 | 30 | } |
417 | | |
418 | | static llvm::Value * |
419 | 20 | getDefaultBuiltinObjectSizeResult(unsigned Type, llvm::IntegerType *ResType) { |
420 | 20 | return ConstantInt::get(ResType, (Type & 2) ? 020 : -10 , /*isSigned=*/true); |
421 | 20 | } |
422 | | |
423 | | llvm::Value * |
424 | | CodeGenFunction::evaluateOrEmitBuiltinObjectSize(const Expr *E, unsigned Type, |
425 | | llvm::IntegerType *ResType, |
426 | 102 | llvm::Value *EmittedE) { |
427 | 102 | uint64_t ObjectSize; |
428 | 102 | if (!E->tryEvaluateObjectSize(ObjectSize, getContext(), Type)) |
429 | 51 | return emitBuiltinObjectSize(E, Type, ResType, EmittedE); |
430 | 51 | return ConstantInt::get(ResType, ObjectSize, /*isSigned=*/true); |
431 | 51 | } |
432 | | |
433 | | /// Returns a Value corresponding to the size of the given expression. |
434 | | /// This Value may be either of the following: |
435 | | /// - A llvm::Argument (if E is a param with the pass_object_size attribute on |
436 | | /// it) |
437 | | /// - A call to the @llvm.objectsize intrinsic |
438 | | /// |
439 | | /// EmittedE is the result of emitting `E` as a scalar expr. If it's non-null |
440 | | /// and we wouldn't otherwise try to reference a pass_object_size parameter, |
441 | | /// we'll call @llvm.objectsize on EmittedE, rather than emitting E. |
442 | | llvm::Value * |
443 | | CodeGenFunction::emitBuiltinObjectSize(const Expr *E, unsigned Type, |
444 | | llvm::IntegerType *ResType, |
445 | 4.04k | llvm::Value *EmittedE) { |
446 | 4.04k | // We need to reference an argument if the pointer is a parameter with the |
447 | 4.04k | // pass_object_size attribute. |
448 | 4.04k | if (auto *D4.04k = dyn_cast<DeclRefExpr>(E->IgnoreParenImpCasts())) { |
449 | 1.59k | auto *Param = dyn_cast<ParmVarDecl>(D->getDecl()); |
450 | 1.59k | auto *PS = D->getDecl()->getAttr<PassObjectSizeAttr>(); |
451 | 1.59k | if (Param != nullptr && 1.59k PS != nullptr565 && |
452 | 1.59k | areBOSTypesCompatible(PS->getType(), Type)30 ) { |
453 | 20 | auto Iter = SizeArguments.find(Param); |
454 | 20 | assert(Iter != SizeArguments.end()); |
455 | 20 | |
456 | 20 | const ImplicitParamDecl *D = Iter->second; |
457 | 20 | auto DIter = LocalDeclMap.find(D); |
458 | 20 | assert(DIter != LocalDeclMap.end()); |
459 | 20 | |
460 | 20 | return EmitLoadOfScalar(DIter->second, /*volatile=*/false, |
461 | 20 | getContext().getSizeType(), E->getLocStart()); |
462 | 20 | } |
463 | 4.02k | } |
464 | 4.02k | |
465 | 4.02k | // LLVM can't handle Type=3 appropriately, and __builtin_object_size shouldn't |
466 | 4.02k | // evaluate E for side-effects. In either case, we shouldn't lower to |
467 | 4.02k | // @llvm.objectsize. |
468 | 4.02k | if (4.02k Type == 3 || 4.02k (!EmittedE && 4.00k E->HasSideEffects(getContext())3.96k )) |
469 | 20 | return getDefaultBuiltinObjectSizeResult(Type, ResType); |
470 | 4.00k | |
471 | 4.00k | Value *Ptr = EmittedE ? 4.00k EmittedE37 : EmitScalarExpr(E)3.96k ; |
472 | 4.04k | assert(Ptr->getType()->isPointerTy() && |
473 | 4.04k | "Non-pointer passed to __builtin_object_size?"); |
474 | 4.04k | |
475 | 4.04k | Value *F = CGM.getIntrinsic(Intrinsic::objectsize, {ResType, Ptr->getType()}); |
476 | 4.04k | |
477 | 4.04k | // LLVM only supports 0 and 2, make sure that we pass along that as a boolean. |
478 | 4.04k | Value *Min = Builder.getInt1((Type & 2) != 0); |
479 | 4.04k | // For GCC compatability, __builtin_object_size treat NULL as unknown size. |
480 | 4.04k | Value *NullIsUnknown = Builder.getTrue(); |
481 | 4.04k | return Builder.CreateCall(F, {Ptr, Min, NullIsUnknown}); |
482 | 4.04k | } |
483 | | |
484 | | // Many of MSVC builtins are on both x64 and ARM; to avoid repeating code, we |
485 | | // handle them here. |
486 | | enum class CodeGenFunction::MSVCIntrin { |
487 | | _BitScanForward, |
488 | | _BitScanReverse, |
489 | | _InterlockedAnd, |
490 | | _InterlockedDecrement, |
491 | | _InterlockedExchange, |
492 | | _InterlockedExchangeAdd, |
493 | | _InterlockedExchangeSub, |
494 | | _InterlockedIncrement, |
495 | | _InterlockedOr, |
496 | | _InterlockedXor, |
497 | | _interlockedbittestandset, |
498 | | __fastfail, |
499 | | }; |
500 | | |
501 | | Value *CodeGenFunction::EmitMSVCBuiltinExpr(MSVCIntrin BuiltinID, |
502 | 127 | const CallExpr *E) { |
503 | 127 | switch (BuiltinID) { |
504 | 18 | case MSVCIntrin::_BitScanForward: |
505 | 18 | case MSVCIntrin::_BitScanReverse: { |
506 | 18 | Value *ArgValue = EmitScalarExpr(E->getArg(1)); |
507 | 18 | |
508 | 18 | llvm::Type *ArgType = ArgValue->getType(); |
509 | 18 | llvm::Type *IndexType = |
510 | 18 | EmitScalarExpr(E->getArg(0))->getType()->getPointerElementType(); |
511 | 18 | llvm::Type *ResultType = ConvertType(E->getType()); |
512 | 18 | |
513 | 18 | Value *ArgZero = llvm::Constant::getNullValue(ArgType); |
514 | 18 | Value *ResZero = llvm::Constant::getNullValue(ResultType); |
515 | 18 | Value *ResOne = llvm::ConstantInt::get(ResultType, 1); |
516 | 18 | |
517 | 18 | BasicBlock *Begin = Builder.GetInsertBlock(); |
518 | 18 | BasicBlock *End = createBasicBlock("bitscan_end", this->CurFn); |
519 | 18 | Builder.SetInsertPoint(End); |
520 | 18 | PHINode *Result = Builder.CreatePHI(ResultType, 2, "bitscan_result"); |
521 | 18 | |
522 | 18 | Builder.SetInsertPoint(Begin); |
523 | 18 | Value *IsZero = Builder.CreateICmpEQ(ArgValue, ArgZero); |
524 | 18 | BasicBlock *NotZero = createBasicBlock("bitscan_not_zero", this->CurFn); |
525 | 18 | Builder.CreateCondBr(IsZero, End, NotZero); |
526 | 18 | Result->addIncoming(ResZero, Begin); |
527 | 18 | |
528 | 18 | Builder.SetInsertPoint(NotZero); |
529 | 18 | Address IndexAddress = EmitPointerWithAlignment(E->getArg(0)); |
530 | 18 | |
531 | 18 | if (BuiltinID == MSVCIntrin::_BitScanForward18 ) { |
532 | 9 | Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); |
533 | 9 | Value *ZeroCount = Builder.CreateCall(F, {ArgValue, Builder.getTrue()}); |
534 | 9 | ZeroCount = Builder.CreateIntCast(ZeroCount, IndexType, false); |
535 | 9 | Builder.CreateStore(ZeroCount, IndexAddress, false); |
536 | 18 | } else { |
537 | 9 | unsigned ArgWidth = cast<llvm::IntegerType>(ArgType)->getBitWidth(); |
538 | 9 | Value *ArgTypeLastIndex = llvm::ConstantInt::get(IndexType, ArgWidth - 1); |
539 | 9 | |
540 | 9 | Value *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType); |
541 | 9 | Value *ZeroCount = Builder.CreateCall(F, {ArgValue, Builder.getTrue()}); |
542 | 9 | ZeroCount = Builder.CreateIntCast(ZeroCount, IndexType, false); |
543 | 9 | Value *Index = Builder.CreateNSWSub(ArgTypeLastIndex, ZeroCount); |
544 | 9 | Builder.CreateStore(Index, IndexAddress, false); |
545 | 9 | } |
546 | 18 | Builder.CreateBr(End); |
547 | 18 | Result->addIncoming(ResOne, NotZero); |
548 | 18 | |
549 | 18 | Builder.SetInsertPoint(End); |
550 | 18 | return Result; |
551 | 18 | } |
552 | 13 | case MSVCIntrin::_InterlockedAnd: |
553 | 13 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::And, E); |
554 | 16 | case MSVCIntrin::_InterlockedExchange: |
555 | 16 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xchg, E); |
556 | 13 | case MSVCIntrin::_InterlockedExchangeAdd: |
557 | 13 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Add, E); |
558 | 13 | case MSVCIntrin::_InterlockedExchangeSub: |
559 | 13 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Sub, E); |
560 | 13 | case MSVCIntrin::_InterlockedOr: |
561 | 13 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Or, E); |
562 | 13 | case MSVCIntrin::_InterlockedXor: |
563 | 13 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xor, E); |
564 | 18 | |
565 | 5 | case MSVCIntrin::_interlockedbittestandset: { |
566 | 5 | llvm::Value *Addr = EmitScalarExpr(E->getArg(0)); |
567 | 5 | llvm::Value *Bit = EmitScalarExpr(E->getArg(1)); |
568 | 5 | AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( |
569 | 5 | AtomicRMWInst::Or, Addr, |
570 | 5 | Builder.CreateShl(ConstantInt::get(Bit->getType(), 1), Bit), |
571 | 5 | llvm::AtomicOrdering::SequentiallyConsistent); |
572 | 5 | // Shift the relevant bit to the least significant position, truncate to |
573 | 5 | // the result type, and test the low bit. |
574 | 5 | llvm::Value *Shifted = Builder.CreateLShr(RMWI, Bit); |
575 | 5 | llvm::Value *Truncated = |
576 | 5 | Builder.CreateTrunc(Shifted, ConvertType(E->getType())); |
577 | 5 | return Builder.CreateAnd(Truncated, |
578 | 5 | ConstantInt::get(Truncated->getType(), 1)); |
579 | 18 | } |
580 | 18 | |
581 | 10 | case MSVCIntrin::_InterlockedDecrement: { |
582 | 10 | llvm::Type *IntTy = ConvertType(E->getType()); |
583 | 10 | AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( |
584 | 10 | AtomicRMWInst::Sub, |
585 | 10 | EmitScalarExpr(E->getArg(0)), |
586 | 10 | ConstantInt::get(IntTy, 1), |
587 | 10 | llvm::AtomicOrdering::SequentiallyConsistent); |
588 | 10 | return Builder.CreateSub(RMWI, ConstantInt::get(IntTy, 1)); |
589 | 18 | } |
590 | 10 | case MSVCIntrin::_InterlockedIncrement: { |
591 | 10 | llvm::Type *IntTy = ConvertType(E->getType()); |
592 | 10 | AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( |
593 | 10 | AtomicRMWInst::Add, |
594 | 10 | EmitScalarExpr(E->getArg(0)), |
595 | 10 | ConstantInt::get(IntTy, 1), |
596 | 10 | llvm::AtomicOrdering::SequentiallyConsistent); |
597 | 10 | return Builder.CreateAdd(RMWI, ConstantInt::get(IntTy, 1)); |
598 | 18 | } |
599 | 18 | |
600 | 3 | case MSVCIntrin::__fastfail: { |
601 | 3 | // Request immediate process termination from the kernel. The instruction |
602 | 3 | // sequences to do this are documented on MSDN: |
603 | 3 | // https://msdn.microsoft.com/en-us/library/dn774154.aspx |
604 | 3 | llvm::Triple::ArchType ISA = getTarget().getTriple().getArch(); |
605 | 3 | StringRef Asm, Constraints; |
606 | 3 | switch (ISA) { |
607 | 0 | default: |
608 | 0 | ErrorUnsupported(E, "__fastfail call for this architecture"); |
609 | 0 | break; |
610 | 2 | case llvm::Triple::x86: |
611 | 2 | case llvm::Triple::x86_64: |
612 | 2 | Asm = "int $$0x29"; |
613 | 2 | Constraints = "{cx}"; |
614 | 2 | break; |
615 | 1 | case llvm::Triple::thumb: |
616 | 1 | Asm = "udf #251"; |
617 | 1 | Constraints = "{r0}"; |
618 | 1 | break; |
619 | 3 | } |
620 | 3 | llvm::FunctionType *FTy = llvm::FunctionType::get(VoidTy, {Int32Ty}, false); |
621 | 3 | llvm::InlineAsm *IA = |
622 | 3 | llvm::InlineAsm::get(FTy, Asm, Constraints, /*SideEffects=*/true); |
623 | 3 | llvm::AttributeList NoReturnAttr = llvm::AttributeList::get( |
624 | 3 | getLLVMContext(), llvm::AttributeList::FunctionIndex, |
625 | 3 | llvm::Attribute::NoReturn); |
626 | 3 | CallSite CS = Builder.CreateCall(IA, EmitScalarExpr(E->getArg(0))); |
627 | 3 | CS.setAttributes(NoReturnAttr); |
628 | 3 | return CS.getInstruction(); |
629 | 3 | } |
630 | 0 | } |
631 | 0 | llvm_unreachable0 ("Incorrect MSVC intrinsic!"); |
632 | 0 | } |
633 | | |
634 | | namespace { |
635 | | // ARC cleanup for __builtin_os_log_format |
636 | | struct CallObjCArcUse final : EHScopeStack::Cleanup { |
637 | 1 | CallObjCArcUse(llvm::Value *object) : object(object) {} |
638 | | llvm::Value *object; |
639 | | |
640 | 1 | void Emit(CodeGenFunction &CGF, Flags flags) override { |
641 | 1 | CGF.EmitARCIntrinsicUse(object); |
642 | 1 | } |
643 | | }; |
644 | | } |
645 | | |
646 | | Value *CodeGenFunction::EmitCheckedArgForBuiltin(const Expr *E, |
647 | 1.16k | BuiltinCheckKind Kind) { |
648 | 1.16k | assert((Kind == BCK_CLZPassedZero || Kind == BCK_CTZPassedZero) |
649 | 1.16k | && "Unsupported builtin check kind"); |
650 | 1.16k | |
651 | 1.16k | Value *ArgValue = EmitScalarExpr(E); |
652 | 1.16k | if (!SanOpts.has(SanitizerKind::Builtin) || 1.16k !getTarget().isCLZForZeroUndef()12 ) |
653 | 1.15k | return ArgValue; |
654 | 6 | |
655 | 6 | SanitizerScope SanScope(this); |
656 | 6 | Value *Cond = Builder.CreateICmpNE( |
657 | 6 | ArgValue, llvm::Constant::getNullValue(ArgValue->getType())); |
658 | 6 | EmitCheck(std::make_pair(Cond, SanitizerKind::Builtin), |
659 | 6 | SanitizerHandler::InvalidBuiltin, |
660 | 6 | {EmitCheckSourceLocation(E->getExprLoc()), |
661 | 6 | llvm::ConstantInt::get(Builder.getInt8Ty(), Kind)}, |
662 | 6 | None); |
663 | 6 | return ArgValue; |
664 | 6 | } |
665 | | |
666 | | RValue CodeGenFunction::EmitBuiltinExpr(const FunctionDecl *FD, |
667 | | unsigned BuiltinID, const CallExpr *E, |
668 | 155k | ReturnValueSlot ReturnValue) { |
669 | 155k | // See if we can constant fold this builtin. If so, don't emit it at all. |
670 | 155k | Expr::EvalResult Result; |
671 | 155k | if (E->EvaluateAsRValue(Result, CGM.getContext()) && |
672 | 155k | !Result.hasSideEffects()2.61k ) { |
673 | 2.61k | if (Result.Val.isInt()) |
674 | 1.99k | return RValue::get(llvm::ConstantInt::get(getLLVMContext(), |
675 | 1.99k | Result.Val.getInt())); |
676 | 619 | if (619 Result.Val.isFloat()619 ) |
677 | 614 | return RValue::get(llvm::ConstantFP::get(getLLVMContext(), |
678 | 614 | Result.Val.getFloat())); |
679 | 152k | } |
680 | 152k | |
681 | 152k | switch (BuiltinID) { |
682 | 76.0k | default: break; // Handle intrinsics and libm functions below. |
683 | 5 | case Builtin::BI__builtin___CFStringMakeConstantString: |
684 | 5 | case Builtin::BI__builtin___NSStringMakeConstantString: |
685 | 5 | return RValue::get(ConstantEmitter(*this).emitAbstract(E, E->getType())); |
686 | 642 | case Builtin::BI__builtin_stdarg_start: |
687 | 642 | case Builtin::BI__builtin_va_start: |
688 | 642 | case Builtin::BI__va_start: |
689 | 642 | case Builtin::BI__builtin_va_end: |
690 | 642 | return RValue::get( |
691 | 642 | EmitVAStartEnd(BuiltinID == Builtin::BI__va_start |
692 | 0 | ? EmitScalarExpr(E->getArg(0)) |
693 | 642 | : EmitVAListRef(E->getArg(0)).getPointer(), |
694 | 642 | BuiltinID != Builtin::BI__builtin_va_end)); |
695 | 51 | case Builtin::BI__builtin_va_copy: { |
696 | 51 | Value *DstPtr = EmitVAListRef(E->getArg(0)).getPointer(); |
697 | 51 | Value *SrcPtr = EmitVAListRef(E->getArg(1)).getPointer(); |
698 | 51 | |
699 | 51 | llvm::Type *Type = Int8PtrTy; |
700 | 51 | |
701 | 51 | DstPtr = Builder.CreateBitCast(DstPtr, Type); |
702 | 51 | SrcPtr = Builder.CreateBitCast(SrcPtr, Type); |
703 | 51 | return RValue::get(Builder.CreateCall(CGM.getIntrinsic(Intrinsic::vacopy), |
704 | 51 | {DstPtr, SrcPtr})); |
705 | 642 | } |
706 | 3 | case Builtin::BI__builtin_abs: |
707 | 3 | case Builtin::BI__builtin_labs: |
708 | 3 | case Builtin::BI__builtin_llabs: { |
709 | 3 | Value *ArgValue = EmitScalarExpr(E->getArg(0)); |
710 | 3 | |
711 | 3 | Value *NegOp = Builder.CreateNeg(ArgValue, "neg"); |
712 | 3 | Value *CmpResult = |
713 | 3 | Builder.CreateICmpSGE(ArgValue, |
714 | 3 | llvm::Constant::getNullValue(ArgValue->getType()), |
715 | 3 | "abscond"); |
716 | 3 | Value *Result = |
717 | 3 | Builder.CreateSelect(CmpResult, ArgValue, NegOp, "abs"); |
718 | 3 | |
719 | 3 | return RValue::get(Result); |
720 | 3 | } |
721 | 207 | case Builtin::BI__builtin_fabs: |
722 | 207 | case Builtin::BI__builtin_fabsf: |
723 | 207 | case Builtin::BI__builtin_fabsl: { |
724 | 207 | return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::fabs)); |
725 | 207 | } |
726 | 6 | case Builtin::BI__builtin_fmod: |
727 | 6 | case Builtin::BI__builtin_fmodf: |
728 | 6 | case Builtin::BI__builtin_fmodl: { |
729 | 6 | Value *Arg1 = EmitScalarExpr(E->getArg(0)); |
730 | 6 | Value *Arg2 = EmitScalarExpr(E->getArg(1)); |
731 | 6 | Value *Result = Builder.CreateFRem(Arg1, Arg2, "fmod"); |
732 | 6 | return RValue::get(Result); |
733 | 6 | } |
734 | 1.50k | case Builtin::BI__builtin_copysign: |
735 | 1.50k | case Builtin::BI__builtin_copysignf: |
736 | 1.50k | case Builtin::BI__builtin_copysignl: { |
737 | 1.50k | return RValue::get(emitBinaryBuiltin(*this, E, Intrinsic::copysign)); |
738 | 1.50k | } |
739 | 6 | case Builtin::BI__builtin_ceil: |
740 | 6 | case Builtin::BI__builtin_ceilf: |
741 | 6 | case Builtin::BI__builtin_ceill: { |
742 | 6 | return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::ceil)); |
743 | 6 | } |
744 | 6 | case Builtin::BI__builtin_floor: |
745 | 6 | case Builtin::BI__builtin_floorf: |
746 | 6 | case Builtin::BI__builtin_floorl: { |
747 | 6 | return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::floor)); |
748 | 6 | } |
749 | 6 | case Builtin::BI__builtin_trunc: |
750 | 6 | case Builtin::BI__builtin_truncf: |
751 | 6 | case Builtin::BI__builtin_truncl: { |
752 | 6 | return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::trunc)); |
753 | 6 | } |
754 | 6 | case Builtin::BI__builtin_rint: |
755 | 6 | case Builtin::BI__builtin_rintf: |
756 | 6 | case Builtin::BI__builtin_rintl: { |
757 | 6 | return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::rint)); |
758 | 6 | } |
759 | 6 | case Builtin::BI__builtin_nearbyint: |
760 | 6 | case Builtin::BI__builtin_nearbyintf: |
761 | 6 | case Builtin::BI__builtin_nearbyintl: { |
762 | 6 | return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::nearbyint)); |
763 | 6 | } |
764 | 6 | case Builtin::BI__builtin_round: |
765 | 6 | case Builtin::BI__builtin_roundf: |
766 | 6 | case Builtin::BI__builtin_roundl: { |
767 | 6 | return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::round)); |
768 | 6 | } |
769 | 6 | case Builtin::BI__builtin_fmin: |
770 | 6 | case Builtin::BI__builtin_fminf: |
771 | 6 | case Builtin::BI__builtin_fminl: { |
772 | 6 | return RValue::get(emitBinaryBuiltin(*this, E, Intrinsic::minnum)); |
773 | 6 | } |
774 | 100 | case Builtin::BI__builtin_fmax: |
775 | 100 | case Builtin::BI__builtin_fmaxf: |
776 | 100 | case Builtin::BI__builtin_fmaxl: { |
777 | 100 | return RValue::get(emitBinaryBuiltin(*this, E, Intrinsic::maxnum)); |
778 | 100 | } |
779 | 6 | case Builtin::BI__builtin_conj: |
780 | 6 | case Builtin::BI__builtin_conjf: |
781 | 6 | case Builtin::BI__builtin_conjl: { |
782 | 6 | ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); |
783 | 6 | Value *Real = ComplexVal.first; |
784 | 6 | Value *Imag = ComplexVal.second; |
785 | 6 | Value *Zero = |
786 | 6 | Imag->getType()->isFPOrFPVectorTy() |
787 | 6 | ? llvm::ConstantFP::getZeroValueForNegation(Imag->getType()) |
788 | 0 | : llvm::Constant::getNullValue(Imag->getType()); |
789 | 6 | |
790 | 6 | Imag = Builder.CreateFSub(Zero, Imag, "sub"); |
791 | 6 | return RValue::getComplex(std::make_pair(Real, Imag)); |
792 | 6 | } |
793 | 30 | case Builtin::BI__builtin_creal: |
794 | 30 | case Builtin::BI__builtin_crealf: |
795 | 30 | case Builtin::BI__builtin_creall: |
796 | 30 | case Builtin::BIcreal: |
797 | 30 | case Builtin::BIcrealf: |
798 | 30 | case Builtin::BIcreall: { |
799 | 30 | ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); |
800 | 30 | return RValue::get(ComplexVal.first); |
801 | 30 | } |
802 | 30 | |
803 | 11 | case Builtin::BI__builtin_cimag: |
804 | 11 | case Builtin::BI__builtin_cimagf: |
805 | 11 | case Builtin::BI__builtin_cimagl: |
806 | 11 | case Builtin::BIcimag: |
807 | 11 | case Builtin::BIcimagf: |
808 | 11 | case Builtin::BIcimagl: { |
809 | 11 | ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); |
810 | 11 | return RValue::get(ComplexVal.second); |
811 | 11 | } |
812 | 11 | |
813 | 267 | case Builtin::BI__builtin_ctzs: |
814 | 267 | case Builtin::BI__builtin_ctz: |
815 | 267 | case Builtin::BI__builtin_ctzl: |
816 | 267 | case Builtin::BI__builtin_ctzll: { |
817 | 267 | Value *ArgValue = EmitCheckedArgForBuiltin(E->getArg(0), BCK_CTZPassedZero); |
818 | 267 | |
819 | 267 | llvm::Type *ArgType = ArgValue->getType(); |
820 | 267 | Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); |
821 | 267 | |
822 | 267 | llvm::Type *ResultType = ConvertType(E->getType()); |
823 | 267 | Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef()); |
824 | 267 | Value *Result = Builder.CreateCall(F, {ArgValue, ZeroUndef}); |
825 | 267 | if (Result->getType() != ResultType) |
826 | 57 | Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, |
827 | 57 | "cast"); |
828 | 267 | return RValue::get(Result); |
829 | 267 | } |
830 | 895 | case Builtin::BI__builtin_clzs: |
831 | 895 | case Builtin::BI__builtin_clz: |
832 | 895 | case Builtin::BI__builtin_clzl: |
833 | 895 | case Builtin::BI__builtin_clzll: { |
834 | 895 | Value *ArgValue = EmitCheckedArgForBuiltin(E->getArg(0), BCK_CLZPassedZero); |
835 | 895 | |
836 | 895 | llvm::Type *ArgType = ArgValue->getType(); |
837 | 895 | Value *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType); |
838 | 895 | |
839 | 895 | llvm::Type *ResultType = ConvertType(E->getType()); |
840 | 895 | Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef()); |
841 | 895 | Value *Result = Builder.CreateCall(F, {ArgValue, ZeroUndef}); |
842 | 895 | if (Result->getType() != ResultType) |
843 | 196 | Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, |
844 | 196 | "cast"); |
845 | 895 | return RValue::get(Result); |
846 | 895 | } |
847 | 10 | case Builtin::BI__builtin_ffs: |
848 | 10 | case Builtin::BI__builtin_ffsl: |
849 | 10 | case Builtin::BI__builtin_ffsll: { |
850 | 10 | // ffs(x) -> x ? cttz(x) + 1 : 0 |
851 | 10 | Value *ArgValue = EmitScalarExpr(E->getArg(0)); |
852 | 10 | |
853 | 10 | llvm::Type *ArgType = ArgValue->getType(); |
854 | 10 | Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); |
855 | 10 | |
856 | 10 | llvm::Type *ResultType = ConvertType(E->getType()); |
857 | 10 | Value *Tmp = |
858 | 10 | Builder.CreateAdd(Builder.CreateCall(F, {ArgValue, Builder.getTrue()}), |
859 | 10 | llvm::ConstantInt::get(ArgType, 1)); |
860 | 10 | Value *Zero = llvm::Constant::getNullValue(ArgType); |
861 | 10 | Value *IsZero = Builder.CreateICmpEQ(ArgValue, Zero, "iszero"); |
862 | 10 | Value *Result = Builder.CreateSelect(IsZero, Zero, Tmp, "ffs"); |
863 | 10 | if (Result->getType() != ResultType) |
864 | 6 | Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, |
865 | 6 | "cast"); |
866 | 10 | return RValue::get(Result); |
867 | 10 | } |
868 | 6 | case Builtin::BI__builtin_parity: |
869 | 6 | case Builtin::BI__builtin_parityl: |
870 | 6 | case Builtin::BI__builtin_parityll: { |
871 | 6 | // parity(x) -> ctpop(x) & 1 |
872 | 6 | Value *ArgValue = EmitScalarExpr(E->getArg(0)); |
873 | 6 | |
874 | 6 | llvm::Type *ArgType = ArgValue->getType(); |
875 | 6 | Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); |
876 | 6 | |
877 | 6 | llvm::Type *ResultType = ConvertType(E->getType()); |
878 | 6 | Value *Tmp = Builder.CreateCall(F, ArgValue); |
879 | 6 | Value *Result = Builder.CreateAnd(Tmp, llvm::ConstantInt::get(ArgType, 1)); |
880 | 6 | if (Result->getType() != ResultType) |
881 | 4 | Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, |
882 | 4 | "cast"); |
883 | 6 | return RValue::get(Result); |
884 | 6 | } |
885 | 16 | case Builtin::BI__popcnt16: |
886 | 16 | case Builtin::BI__popcnt: |
887 | 16 | case Builtin::BI__popcnt64: |
888 | 16 | case Builtin::BI__builtin_popcount: |
889 | 16 | case Builtin::BI__builtin_popcountl: |
890 | 16 | case Builtin::BI__builtin_popcountll: { |
891 | 16 | Value *ArgValue = EmitScalarExpr(E->getArg(0)); |
892 | 16 | |
893 | 16 | llvm::Type *ArgType = ArgValue->getType(); |
894 | 16 | Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); |
895 | 16 | |
896 | 16 | llvm::Type *ResultType = ConvertType(E->getType()); |
897 | 16 | Value *Result = Builder.CreateCall(F, ArgValue); |
898 | 16 | if (Result->getType() != ResultType) |
899 | 9 | Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, |
900 | 9 | "cast"); |
901 | 16 | return RValue::get(Result); |
902 | 16 | } |
903 | 30 | case Builtin::BI_rotr8: |
904 | 30 | case Builtin::BI_rotr16: |
905 | 30 | case Builtin::BI_rotr: |
906 | 30 | case Builtin::BI_lrotr: |
907 | 30 | case Builtin::BI_rotr64: { |
908 | 30 | Value *Val = EmitScalarExpr(E->getArg(0)); |
909 | 30 | Value *Shift = EmitScalarExpr(E->getArg(1)); |
910 | 30 | |
911 | 30 | llvm::Type *ArgType = Val->getType(); |
912 | 30 | Shift = Builder.CreateIntCast(Shift, ArgType, false); |
913 | 30 | unsigned ArgWidth = cast<llvm::IntegerType>(ArgType)->getBitWidth(); |
914 | 30 | Value *ArgTypeSize = llvm::ConstantInt::get(ArgType, ArgWidth); |
915 | 30 | Value *ArgZero = llvm::Constant::getNullValue(ArgType); |
916 | 30 | |
917 | 30 | Value *Mask = llvm::ConstantInt::get(ArgType, ArgWidth - 1); |
918 | 30 | Shift = Builder.CreateAnd(Shift, Mask); |
919 | 30 | Value *LeftShift = Builder.CreateSub(ArgTypeSize, Shift); |
920 | 30 | |
921 | 30 | Value *RightShifted = Builder.CreateLShr(Val, Shift); |
922 | 30 | Value *LeftShifted = Builder.CreateShl(Val, LeftShift); |
923 | 30 | Value *Rotated = Builder.CreateOr(LeftShifted, RightShifted); |
924 | 30 | |
925 | 30 | Value *ShiftIsZero = Builder.CreateICmpEQ(Shift, ArgZero); |
926 | 30 | Value *Result = Builder.CreateSelect(ShiftIsZero, Val, Rotated); |
927 | 30 | return RValue::get(Result); |
928 | 30 | } |
929 | 30 | case Builtin::BI_rotl8: |
930 | 30 | case Builtin::BI_rotl16: |
931 | 30 | case Builtin::BI_rotl: |
932 | 30 | case Builtin::BI_lrotl: |
933 | 30 | case Builtin::BI_rotl64: { |
934 | 30 | Value *Val = EmitScalarExpr(E->getArg(0)); |
935 | 30 | Value *Shift = EmitScalarExpr(E->getArg(1)); |
936 | 30 | |
937 | 30 | llvm::Type *ArgType = Val->getType(); |
938 | 30 | Shift = Builder.CreateIntCast(Shift, ArgType, false); |
939 | 30 | unsigned ArgWidth = cast<llvm::IntegerType>(ArgType)->getBitWidth(); |
940 | 30 | Value *ArgTypeSize = llvm::ConstantInt::get(ArgType, ArgWidth); |
941 | 30 | Value *ArgZero = llvm::Constant::getNullValue(ArgType); |
942 | 30 | |
943 | 30 | Value *Mask = llvm::ConstantInt::get(ArgType, ArgWidth - 1); |
944 | 30 | Shift = Builder.CreateAnd(Shift, Mask); |
945 | 30 | Value *RightShift = Builder.CreateSub(ArgTypeSize, Shift); |
946 | 30 | |
947 | 30 | Value *LeftShifted = Builder.CreateShl(Val, Shift); |
948 | 30 | Value *RightShifted = Builder.CreateLShr(Val, RightShift); |
949 | 30 | Value *Rotated = Builder.CreateOr(LeftShifted, RightShifted); |
950 | 30 | |
951 | 30 | Value *ShiftIsZero = Builder.CreateICmpEQ(Shift, ArgZero); |
952 | 30 | Value *Result = Builder.CreateSelect(ShiftIsZero, Val, Rotated); |
953 | 30 | return RValue::get(Result); |
954 | 30 | } |
955 | 4 | case Builtin::BI__builtin_unpredictable: { |
956 | 4 | // Always return the argument of __builtin_unpredictable. LLVM does not |
957 | 4 | // handle this builtin. Metadata for this builtin should be added directly |
958 | 4 | // to instructions such as branches or switches that use it. |
959 | 4 | return RValue::get(EmitScalarExpr(E->getArg(0))); |
960 | 30 | } |
961 | 5.06k | case Builtin::BI__builtin_expect: { |
962 | 5.06k | Value *ArgValue = EmitScalarExpr(E->getArg(0)); |
963 | 5.06k | llvm::Type *ArgType = ArgValue->getType(); |
964 | 5.06k | |
965 | 5.06k | Value *ExpectedValue = EmitScalarExpr(E->getArg(1)); |
966 | 5.06k | // Don't generate llvm.expect on -O0 as the backend won't use it for |
967 | 5.06k | // anything. |
968 | 5.06k | // Note, we still IRGen ExpectedValue because it could have side-effects. |
969 | 5.06k | if (CGM.getCodeGenOpts().OptimizationLevel == 0) |
970 | 7 | return RValue::get(ArgValue); |
971 | 5.05k | |
972 | 5.05k | Value *FnExpect = CGM.getIntrinsic(Intrinsic::expect, ArgType); |
973 | 5.05k | Value *Result = |
974 | 5.05k | Builder.CreateCall(FnExpect, {ArgValue, ExpectedValue}, "expval"); |
975 | 5.05k | return RValue::get(Result); |
976 | 5.05k | } |
977 | 4 | case Builtin::BI__builtin_assume_aligned: { |
978 | 4 | Value *PtrValue = EmitScalarExpr(E->getArg(0)); |
979 | 4 | Value *OffsetValue = |
980 | 4 | (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2))3 : nullptr1 ; |
981 | 4 | |
982 | 4 | Value *AlignmentValue = EmitScalarExpr(E->getArg(1)); |
983 | 4 | ConstantInt *AlignmentCI = cast<ConstantInt>(AlignmentValue); |
984 | 4 | unsigned Alignment = (unsigned) AlignmentCI->getZExtValue(); |
985 | 4 | |
986 | 4 | EmitAlignmentAssumption(PtrValue, Alignment, OffsetValue); |
987 | 4 | return RValue::get(PtrValue); |
988 | 5.05k | } |
989 | 28 | case Builtin::BI__assume: |
990 | 28 | case Builtin::BI__builtin_assume: { |
991 | 28 | if (E->getArg(0)->HasSideEffects(getContext())) |
992 | 4 | return RValue::get(nullptr); |
993 | 24 | |
994 | 24 | Value *ArgValue = EmitScalarExpr(E->getArg(0)); |
995 | 24 | Value *FnAssume = CGM.getIntrinsic(Intrinsic::assume); |
996 | 24 | return RValue::get(Builder.CreateCall(FnAssume, ArgValue)); |
997 | 24 | } |
998 | 26 | case Builtin::BI__builtin_bswap16: |
999 | 26 | case Builtin::BI__builtin_bswap32: |
1000 | 26 | case Builtin::BI__builtin_bswap64: { |
1001 | 26 | return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::bswap)); |
1002 | 26 | } |
1003 | 12 | case Builtin::BI__builtin_bitreverse8: |
1004 | 12 | case Builtin::BI__builtin_bitreverse16: |
1005 | 12 | case Builtin::BI__builtin_bitreverse32: |
1006 | 12 | case Builtin::BI__builtin_bitreverse64: { |
1007 | 12 | return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::bitreverse)); |
1008 | 12 | } |
1009 | 3.99k | case Builtin::BI__builtin_object_size: { |
1010 | 3.99k | unsigned Type = |
1011 | 3.99k | E->getArg(1)->EvaluateKnownConstInt(getContext()).getZExtValue(); |
1012 | 3.99k | auto *ResType = cast<llvm::IntegerType>(ConvertType(E->getType())); |
1013 | 3.99k | |
1014 | 3.99k | // We pass this builtin onto the optimizer so that it can figure out the |
1015 | 3.99k | // object size in more complex cases. |
1016 | 3.99k | return RValue::get(emitBuiltinObjectSize(E->getArg(0), Type, ResType, |
1017 | 3.99k | /*EmittedE=*/nullptr)); |
1018 | 12 | } |
1019 | 30 | case Builtin::BI__builtin_prefetch: { |
1020 | 30 | Value *Locality, *RW, *Address = EmitScalarExpr(E->getArg(0)); |
1021 | 30 | // FIXME: Technically these constants should of type 'int', yes? |
1022 | 19 | RW = (E->getNumArgs() > 1) ? EmitScalarExpr(E->getArg(1)) : |
1023 | 11 | llvm::ConstantInt::get(Int32Ty, 0); |
1024 | 16 | Locality = (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : |
1025 | 14 | llvm::ConstantInt::get(Int32Ty, 3); |
1026 | 30 | Value *Data = llvm::ConstantInt::get(Int32Ty, 1); |
1027 | 30 | Value *F = CGM.getIntrinsic(Intrinsic::prefetch); |
1028 | 30 | return RValue::get(Builder.CreateCall(F, {Address, RW, Locality, Data})); |
1029 | 12 | } |
1030 | 2 | case Builtin::BI__builtin_readcyclecounter: { |
1031 | 2 | Value *F = CGM.getIntrinsic(Intrinsic::readcyclecounter); |
1032 | 2 | return RValue::get(Builder.CreateCall(F)); |
1033 | 12 | } |
1034 | 1 | case Builtin::BI__builtin___clear_cache: { |
1035 | 1 | Value *Begin = EmitScalarExpr(E->getArg(0)); |
1036 | 1 | Value *End = EmitScalarExpr(E->getArg(1)); |
1037 | 1 | Value *F = CGM.getIntrinsic(Intrinsic::clear_cache); |
1038 | 1 | return RValue::get(Builder.CreateCall(F, {Begin, End})); |
1039 | 12 | } |
1040 | 18 | case Builtin::BI__builtin_trap: |
1041 | 18 | return RValue::get(EmitTrapCall(Intrinsic::trap)); |
1042 | 0 | case Builtin::BI__debugbreak: |
1043 | 0 | return RValue::get(EmitTrapCall(Intrinsic::debugtrap)); |
1044 | 23 | case Builtin::BI__builtin_unreachable: { |
1045 | 23 | if (SanOpts.has(SanitizerKind::Unreachable)23 ) { |
1046 | 7 | SanitizerScope SanScope(this); |
1047 | 7 | EmitCheck(std::make_pair(static_cast<llvm::Value *>(Builder.getFalse()), |
1048 | 7 | SanitizerKind::Unreachable), |
1049 | 7 | SanitizerHandler::BuiltinUnreachable, |
1050 | 7 | EmitCheckSourceLocation(E->getExprLoc()), None); |
1051 | 7 | } else |
1052 | 16 | Builder.CreateUnreachable(); |
1053 | 23 | |
1054 | 23 | // We do need to preserve an insertion point. |
1055 | 23 | EmitBlock(createBasicBlock("unreachable.cont")); |
1056 | 23 | |
1057 | 23 | return RValue::get(nullptr); |
1058 | 12 | } |
1059 | 12 | |
1060 | 6 | case Builtin::BI__builtin_powi: |
1061 | 6 | case Builtin::BI__builtin_powif: |
1062 | 6 | case Builtin::BI__builtin_powil: { |
1063 | 6 | Value *Base = EmitScalarExpr(E->getArg(0)); |
1064 | 6 | Value *Exponent = EmitScalarExpr(E->getArg(1)); |
1065 | 6 | llvm::Type *ArgType = Base->getType(); |
1066 | 6 | Value *F = CGM.getIntrinsic(Intrinsic::powi, ArgType); |
1067 | 6 | return RValue::get(Builder.CreateCall(F, {Base, Exponent})); |
1068 | 6 | } |
1069 | 6 | |
1070 | 62 | case Builtin::BI__builtin_isgreater: |
1071 | 62 | case Builtin::BI__builtin_isgreaterequal: |
1072 | 62 | case Builtin::BI__builtin_isless: |
1073 | 62 | case Builtin::BI__builtin_islessequal: |
1074 | 62 | case Builtin::BI__builtin_islessgreater: |
1075 | 62 | case Builtin::BI__builtin_isunordered: { |
1076 | 62 | // Ordered comparisons: we know the arguments to these are matching scalar |
1077 | 62 | // floating point values. |
1078 | 62 | Value *LHS = EmitScalarExpr(E->getArg(0)); |
1079 | 62 | Value *RHS = EmitScalarExpr(E->getArg(1)); |
1080 | 62 | |
1081 | 62 | switch (BuiltinID) { |
1082 | 0 | default: 0 llvm_unreachable0 ("Unknown ordered comparison"); |
1083 | 6 | case Builtin::BI__builtin_isgreater: |
1084 | 6 | LHS = Builder.CreateFCmpOGT(LHS, RHS, "cmp"); |
1085 | 6 | break; |
1086 | 6 | case Builtin::BI__builtin_isgreaterequal: |
1087 | 6 | LHS = Builder.CreateFCmpOGE(LHS, RHS, "cmp"); |
1088 | 6 | break; |
1089 | 7 | case Builtin::BI__builtin_isless: |
1090 | 7 | LHS = Builder.CreateFCmpOLT(LHS, RHS, "cmp"); |
1091 | 7 | break; |
1092 | 6 | case Builtin::BI__builtin_islessequal: |
1093 | 6 | LHS = Builder.CreateFCmpOLE(LHS, RHS, "cmp"); |
1094 | 6 | break; |
1095 | 6 | case Builtin::BI__builtin_islessgreater: |
1096 | 6 | LHS = Builder.CreateFCmpONE(LHS, RHS, "cmp"); |
1097 | 6 | break; |
1098 | 31 | case Builtin::BI__builtin_isunordered: |
1099 | 31 | LHS = Builder.CreateFCmpUNO(LHS, RHS, "cmp"); |
1100 | 31 | break; |
1101 | 62 | } |
1102 | 62 | // ZExt bool to int type. |
1103 | 62 | return RValue::get(Builder.CreateZExt(LHS, ConvertType(E->getType()))); |
1104 | 62 | } |
1105 | 1.16k | case Builtin::BI__builtin_isnan: { |
1106 | 1.16k | Value *V = EmitScalarExpr(E->getArg(0)); |
1107 | 1.16k | V = Builder.CreateFCmpUNO(V, V, "cmp"); |
1108 | 1.16k | return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); |
1109 | 62 | } |
1110 | 62 | |
1111 | 2.07k | case Builtin::BIfinite: |
1112 | 2.07k | case Builtin::BI__finite: |
1113 | 2.07k | case Builtin::BIfinitef: |
1114 | 2.07k | case Builtin::BI__finitef: |
1115 | 2.07k | case Builtin::BIfinitel: |
1116 | 2.07k | case Builtin::BI__finitel: |
1117 | 2.07k | case Builtin::BI__builtin_isinf: |
1118 | 2.07k | case Builtin::BI__builtin_isfinite: { |
1119 | 2.07k | // isinf(x) --> fabs(x) == infinity |
1120 | 2.07k | // isfinite(x) --> fabs(x) != infinity |
1121 | 2.07k | // x != NaN via the ordered compare in either case. |
1122 | 2.07k | Value *V = EmitScalarExpr(E->getArg(0)); |
1123 | 2.07k | Value *Fabs = EmitFAbs(*this, V); |
1124 | 2.07k | Constant *Infinity = ConstantFP::getInfinity(V->getType()); |
1125 | 2.07k | CmpInst::Predicate Pred = (BuiltinID == Builtin::BI__builtin_isinf) |
1126 | 1.60k | ? CmpInst::FCMP_OEQ |
1127 | 474 | : CmpInst::FCMP_ONE; |
1128 | 2.07k | Value *FCmp = Builder.CreateFCmp(Pred, Fabs, Infinity, "cmpinf"); |
1129 | 2.07k | return RValue::get(Builder.CreateZExt(FCmp, ConvertType(E->getType()))); |
1130 | 2.07k | } |
1131 | 2.07k | |
1132 | 6 | case Builtin::BI__builtin_isinf_sign: { |
1133 | 6 | // isinf_sign(x) -> fabs(x) == infinity ? (signbit(x) ? -1 : 1) : 0 |
1134 | 6 | Value *Arg = EmitScalarExpr(E->getArg(0)); |
1135 | 6 | Value *AbsArg = EmitFAbs(*this, Arg); |
1136 | 6 | Value *IsInf = Builder.CreateFCmpOEQ( |
1137 | 6 | AbsArg, ConstantFP::getInfinity(Arg->getType()), "isinf"); |
1138 | 6 | Value *IsNeg = EmitSignBit(*this, Arg); |
1139 | 6 | |
1140 | 6 | llvm::Type *IntTy = ConvertType(E->getType()); |
1141 | 6 | Value *Zero = Constant::getNullValue(IntTy); |
1142 | 6 | Value *One = ConstantInt::get(IntTy, 1); |
1143 | 6 | Value *NegativeOne = ConstantInt::get(IntTy, -1); |
1144 | 6 | Value *SignResult = Builder.CreateSelect(IsNeg, NegativeOne, One); |
1145 | 6 | Value *Result = Builder.CreateSelect(IsInf, SignResult, Zero); |
1146 | 6 | return RValue::get(Result); |
1147 | 2.07k | } |
1148 | 2.07k | |
1149 | 2 | case Builtin::BI__builtin_isnormal: { |
1150 | 2 | // isnormal(x) --> x == x && fabsf(x) < infinity && fabsf(x) >= float_min |
1151 | 2 | Value *V = EmitScalarExpr(E->getArg(0)); |
1152 | 2 | Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); |
1153 | 2 | |
1154 | 2 | Value *Abs = EmitFAbs(*this, V); |
1155 | 2 | Value *IsLessThanInf = |
1156 | 2 | Builder.CreateFCmpULT(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); |
1157 | 2 | APFloat Smallest = APFloat::getSmallestNormalized( |
1158 | 2 | getContext().getFloatTypeSemantics(E->getArg(0)->getType())); |
1159 | 2 | Value *IsNormal = |
1160 | 2 | Builder.CreateFCmpUGE(Abs, ConstantFP::get(V->getContext(), Smallest), |
1161 | 2 | "isnormal"); |
1162 | 2 | V = Builder.CreateAnd(Eq, IsLessThanInf, "and"); |
1163 | 2 | V = Builder.CreateAnd(V, IsNormal, "and"); |
1164 | 2 | return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); |
1165 | 2.07k | } |
1166 | 2.07k | |
1167 | 0 | case Builtin::BI__builtin_fpclassify: { |
1168 | 0 | Value *V = EmitScalarExpr(E->getArg(5)); |
1169 | 0 | llvm::Type *Ty = ConvertType(E->getArg(5)->getType()); |
1170 | 0 |
|
1171 | 0 | // Create Result |
1172 | 0 | BasicBlock *Begin = Builder.GetInsertBlock(); |
1173 | 0 | BasicBlock *End = createBasicBlock("fpclassify_end", this->CurFn); |
1174 | 0 | Builder.SetInsertPoint(End); |
1175 | 0 | PHINode *Result = |
1176 | 0 | Builder.CreatePHI(ConvertType(E->getArg(0)->getType()), 4, |
1177 | 0 | "fpclassify_result"); |
1178 | 0 |
|
1179 | 0 | // if (V==0) return FP_ZERO |
1180 | 0 | Builder.SetInsertPoint(Begin); |
1181 | 0 | Value *IsZero = Builder.CreateFCmpOEQ(V, Constant::getNullValue(Ty), |
1182 | 0 | "iszero"); |
1183 | 0 | Value *ZeroLiteral = EmitScalarExpr(E->getArg(4)); |
1184 | 0 | BasicBlock *NotZero = createBasicBlock("fpclassify_not_zero", this->CurFn); |
1185 | 0 | Builder.CreateCondBr(IsZero, End, NotZero); |
1186 | 0 | Result->addIncoming(ZeroLiteral, Begin); |
1187 | 0 |
|
1188 | 0 | // if (V != V) return FP_NAN |
1189 | 0 | Builder.SetInsertPoint(NotZero); |
1190 | 0 | Value *IsNan = Builder.CreateFCmpUNO(V, V, "cmp"); |
1191 | 0 | Value *NanLiteral = EmitScalarExpr(E->getArg(0)); |
1192 | 0 | BasicBlock *NotNan = createBasicBlock("fpclassify_not_nan", this->CurFn); |
1193 | 0 | Builder.CreateCondBr(IsNan, End, NotNan); |
1194 | 0 | Result->addIncoming(NanLiteral, NotZero); |
1195 | 0 |
|
1196 | 0 | // if (fabs(V) == infinity) return FP_INFINITY |
1197 | 0 | Builder.SetInsertPoint(NotNan); |
1198 | 0 | Value *VAbs = EmitFAbs(*this, V); |
1199 | 0 | Value *IsInf = |
1200 | 0 | Builder.CreateFCmpOEQ(VAbs, ConstantFP::getInfinity(V->getType()), |
1201 | 0 | "isinf"); |
1202 | 0 | Value *InfLiteral = EmitScalarExpr(E->getArg(1)); |
1203 | 0 | BasicBlock *NotInf = createBasicBlock("fpclassify_not_inf", this->CurFn); |
1204 | 0 | Builder.CreateCondBr(IsInf, End, NotInf); |
1205 | 0 | Result->addIncoming(InfLiteral, NotNan); |
1206 | 0 |
|
1207 | 0 | // if (fabs(V) >= MIN_NORMAL) return FP_NORMAL else FP_SUBNORMAL |
1208 | 0 | Builder.SetInsertPoint(NotInf); |
1209 | 0 | APFloat Smallest = APFloat::getSmallestNormalized( |
1210 | 0 | getContext().getFloatTypeSemantics(E->getArg(5)->getType())); |
1211 | 0 | Value *IsNormal = |
1212 | 0 | Builder.CreateFCmpUGE(VAbs, ConstantFP::get(V->getContext(), Smallest), |
1213 | 0 | "isnormal"); |
1214 | 0 | Value *NormalResult = |
1215 | 0 | Builder.CreateSelect(IsNormal, EmitScalarExpr(E->getArg(2)), |
1216 | 0 | EmitScalarExpr(E->getArg(3))); |
1217 | 0 | Builder.CreateBr(End); |
1218 | 0 | Result->addIncoming(NormalResult, NotInf); |
1219 | 0 |
|
1220 | 0 | // return Result |
1221 | 0 | Builder.SetInsertPoint(End); |
1222 | 0 | return RValue::get(Result); |
1223 | 2.07k | } |
1224 | 2.07k | |
1225 | 175 | case Builtin::BIalloca: |
1226 | 175 | case Builtin::BI_alloca: |
1227 | 175 | case Builtin::BI__builtin_alloca: { |
1228 | 175 | Value *Size = EmitScalarExpr(E->getArg(0)); |
1229 | 175 | const TargetInfo &TI = getContext().getTargetInfo(); |
1230 | 175 | // The alignment of the alloca should correspond to __BIGGEST_ALIGNMENT__. |
1231 | 175 | unsigned SuitableAlignmentInBytes = |
1232 | 175 | CGM.getContext() |
1233 | 175 | .toCharUnitsFromBits(TI.getSuitableAlign()) |
1234 | 175 | .getQuantity(); |
1235 | 175 | AllocaInst *AI = Builder.CreateAlloca(Builder.getInt8Ty(), Size); |
1236 | 175 | AI->setAlignment(SuitableAlignmentInBytes); |
1237 | 175 | return RValue::get(AI); |
1238 | 175 | } |
1239 | 175 | |
1240 | 1 | case Builtin::BI__builtin_alloca_with_align: { |
1241 | 1 | Value *Size = EmitScalarExpr(E->getArg(0)); |
1242 | 1 | Value *AlignmentInBitsValue = EmitScalarExpr(E->getArg(1)); |
1243 | 1 | auto *AlignmentInBitsCI = cast<ConstantInt>(AlignmentInBitsValue); |
1244 | 1 | unsigned AlignmentInBits = AlignmentInBitsCI->getZExtValue(); |
1245 | 1 | unsigned AlignmentInBytes = |
1246 | 1 | CGM.getContext().toCharUnitsFromBits(AlignmentInBits).getQuantity(); |
1247 | 1 | AllocaInst *AI = Builder.CreateAlloca(Builder.getInt8Ty(), Size); |
1248 | 1 | AI->setAlignment(AlignmentInBytes); |
1249 | 1 | return RValue::get(AI); |
1250 | 175 | } |
1251 | 175 | |
1252 | 42 | case Builtin::BIbzero: |
1253 | 42 | case Builtin::BI__builtin_bzero: { |
1254 | 42 | Address Dest = EmitPointerWithAlignment(E->getArg(0)); |
1255 | 42 | Value *SizeVal = EmitScalarExpr(E->getArg(1)); |
1256 | 42 | EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), |
1257 | 42 | E->getArg(0)->getExprLoc(), FD, 0); |
1258 | 42 | Builder.CreateMemSet(Dest, Builder.getInt8(0), SizeVal, false); |
1259 | 42 | return RValue::get(Dest.getPointer()); |
1260 | 42 | } |
1261 | 138 | case Builtin::BImemcpy: |
1262 | 138 | case Builtin::BI__builtin_memcpy: { |
1263 | 138 | Address Dest = EmitPointerWithAlignment(E->getArg(0)); |
1264 | 138 | Address Src = EmitPointerWithAlignment(E->getArg(1)); |
1265 | 138 | Value *SizeVal = EmitScalarExpr(E->getArg(2)); |
1266 | 138 | EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), |
1267 | 138 | E->getArg(0)->getExprLoc(), FD, 0); |
1268 | 138 | EmitNonNullArgCheck(RValue::get(Src.getPointer()), E->getArg(1)->getType(), |
1269 | 138 | E->getArg(1)->getExprLoc(), FD, 1); |
1270 | 138 | Builder.CreateMemCpy(Dest, Src, SizeVal, false); |
1271 | 138 | return RValue::get(Dest.getPointer()); |
1272 | 138 | } |
1273 | 138 | |
1274 | 1 | case Builtin::BI__builtin_char_memchr: |
1275 | 1 | BuiltinID = Builtin::BI__builtin_memchr; |
1276 | 1 | break; |
1277 | 138 | |
1278 | 1.76k | case Builtin::BI__builtin___memcpy_chk: { |
1279 | 1.76k | // fold __builtin_memcpy_chk(x, y, cst1, cst2) to memcpy iff cst1<=cst2. |
1280 | 1.76k | llvm::APSInt Size, DstSize; |
1281 | 1.76k | if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || |
1282 | 760 | !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) |
1283 | 1.63k | break; |
1284 | 130 | if (130 Size.ugt(DstSize)130 ) |
1285 | 0 | break; |
1286 | 130 | Address Dest = EmitPointerWithAlignment(E->getArg(0)); |
1287 | 130 | Address Src = EmitPointerWithAlignment(E->getArg(1)); |
1288 | 130 | Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); |
1289 | 130 | Builder.CreateMemCpy(Dest, Src, SizeVal, false); |
1290 | 130 | return RValue::get(Dest.getPointer()); |
1291 | 130 | } |
1292 | 130 | |
1293 | 1 | case Builtin::BI__builtin_objc_memmove_collectable: { |
1294 | 1 | Address DestAddr = EmitPointerWithAlignment(E->getArg(0)); |
1295 | 1 | Address SrcAddr = EmitPointerWithAlignment(E->getArg(1)); |
1296 | 1 | Value *SizeVal = EmitScalarExpr(E->getArg(2)); |
1297 | 1 | CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this, |
1298 | 1 | DestAddr, SrcAddr, SizeVal); |
1299 | 1 | return RValue::get(DestAddr.getPointer()); |
1300 | 130 | } |
1301 | 130 | |
1302 | 143 | case Builtin::BI__builtin___memmove_chk: { |
1303 | 143 | // fold __builtin_memmove_chk(x, y, cst1, cst2) to memmove iff cst1<=cst2. |
1304 | 143 | llvm::APSInt Size, DstSize; |
1305 | 143 | if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || |
1306 | 39 | !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) |
1307 | 134 | break; |
1308 | 9 | if (9 Size.ugt(DstSize)9 ) |
1309 | 0 | break; |
1310 | 9 | Address Dest = EmitPointerWithAlignment(E->getArg(0)); |
1311 | 9 | Address Src = EmitPointerWithAlignment(E->getArg(1)); |
1312 | 9 | Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); |
1313 | 9 | Builder.CreateMemMove(Dest, Src, SizeVal, false); |
1314 | 9 | return RValue::get(Dest.getPointer()); |
1315 | 9 | } |
1316 | 9 | |
1317 | 109 | case Builtin::BImemmove: |
1318 | 109 | case Builtin::BI__builtin_memmove: { |
1319 | 109 | Address Dest = EmitPointerWithAlignment(E->getArg(0)); |
1320 | 109 | Address Src = EmitPointerWithAlignment(E->getArg(1)); |
1321 | 109 | Value *SizeVal = EmitScalarExpr(E->getArg(2)); |
1322 | 109 | EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), |
1323 | 109 | E->getArg(0)->getExprLoc(), FD, 0); |
1324 | 109 | EmitNonNullArgCheck(RValue::get(Src.getPointer()), E->getArg(1)->getType(), |
1325 | 109 | E->getArg(1)->getExprLoc(), FD, 1); |
1326 | 109 | Builder.CreateMemMove(Dest, Src, SizeVal, false); |
1327 | 109 | return RValue::get(Dest.getPointer()); |
1328 | 109 | } |
1329 | 59 | case Builtin::BImemset: |
1330 | 59 | case Builtin::BI__builtin_memset: { |
1331 | 59 | Address Dest = EmitPointerWithAlignment(E->getArg(0)); |
1332 | 59 | Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), |
1333 | 59 | Builder.getInt8Ty()); |
1334 | 59 | Value *SizeVal = EmitScalarExpr(E->getArg(2)); |
1335 | 59 | EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), |
1336 | 59 | E->getArg(0)->getExprLoc(), FD, 0); |
1337 | 59 | Builder.CreateMemSet(Dest, ByteVal, SizeVal, false); |
1338 | 59 | return RValue::get(Dest.getPointer()); |
1339 | 59 | } |
1340 | 1.21k | case Builtin::BI__builtin___memset_chk: { |
1341 | 1.21k | // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2. |
1342 | 1.21k | llvm::APSInt Size, DstSize; |
1343 | 1.21k | if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || |
1344 | 831 | !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) |
1345 | 938 | break; |
1346 | 273 | if (273 Size.ugt(DstSize)273 ) |
1347 | 0 | break; |
1348 | 273 | Address Dest = EmitPointerWithAlignment(E->getArg(0)); |
1349 | 273 | Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), |
1350 | 273 | Builder.getInt8Ty()); |
1351 | 273 | Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); |
1352 | 273 | Builder.CreateMemSet(Dest, ByteVal, SizeVal, false); |
1353 | 273 | return RValue::get(Dest.getPointer()); |
1354 | 273 | } |
1355 | 0 | case Builtin::BI__builtin_dwarf_cfa: { |
1356 | 0 | // The offset in bytes from the first argument to the CFA. |
1357 | 0 | // |
1358 | 0 | // Why on earth is this in the frontend? Is there any reason at |
1359 | 0 | // all that the backend can't reasonably determine this while |
1360 | 0 | // lowering llvm.eh.dwarf.cfa()? |
1361 | 0 | // |
1362 | 0 | // TODO: If there's a satisfactory reason, add a target hook for |
1363 | 0 | // this instead of hard-coding 0, which is correct for most targets. |
1364 | 0 | int32_t Offset = 0; |
1365 | 0 |
|
1366 | 0 | Value *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa); |
1367 | 0 | return RValue::get(Builder.CreateCall(F, |
1368 | 0 | llvm::ConstantInt::get(Int32Ty, Offset))); |
1369 | 273 | } |
1370 | 30 | case Builtin::BI__builtin_return_address: { |
1371 | 30 | Value *Depth = ConstantEmitter(*this).emitAbstract(E->getArg(0), |
1372 | 30 | getContext().UnsignedIntTy); |
1373 | 30 | Value *F = CGM.getIntrinsic(Intrinsic::returnaddress); |
1374 | 30 | return RValue::get(Builder.CreateCall(F, Depth)); |
1375 | 273 | } |
1376 | 3 | case Builtin::BI_ReturnAddress: { |
1377 | 3 | Value *F = CGM.getIntrinsic(Intrinsic::returnaddress); |
1378 | 3 | return RValue::get(Builder.CreateCall(F, Builder.getInt32(0))); |
1379 | 273 | } |
1380 | 8 | case Builtin::BI__builtin_frame_address: { |
1381 | 8 | Value *Depth = ConstantEmitter(*this).emitAbstract(E->getArg(0), |
1382 | 8 | getContext().UnsignedIntTy); |
1383 | 8 | Value *F = CGM.getIntrinsic(Intrinsic::frameaddress); |
1384 | 8 | return RValue::get(Builder.CreateCall(F, Depth)); |
1385 | 273 | } |
1386 | 2 | case Builtin::BI__builtin_extract_return_addr: { |
1387 | 2 | Value *Address = EmitScalarExpr(E->getArg(0)); |
1388 | 2 | Value *Result = getTargetHooks().decodeReturnAddress(*this, Address); |
1389 | 2 | return RValue::get(Result); |
1390 | 273 | } |
1391 | 0 | case Builtin::BI__builtin_frob_return_addr: { |
1392 | 0 | Value *Address = EmitScalarExpr(E->getArg(0)); |
1393 | 0 | Value *Result = getTargetHooks().encodeReturnAddress(*this, Address); |
1394 | 0 | return RValue::get(Result); |
1395 | 273 | } |
1396 | 2 | case Builtin::BI__builtin_dwarf_sp_column: { |
1397 | 2 | llvm::IntegerType *Ty |
1398 | 2 | = cast<llvm::IntegerType>(ConvertType(E->getType())); |
1399 | 2 | int Column = getTargetHooks().getDwarfEHStackPointer(CGM); |
1400 | 2 | if (Column == -12 ) { |
1401 | 0 | CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column"); |
1402 | 0 | return RValue::get(llvm::UndefValue::get(Ty)); |
1403 | 0 | } |
1404 | 2 | return RValue::get(llvm::ConstantInt::get(Ty, Column, true)); |
1405 | 2 | } |
1406 | 2 | case Builtin::BI__builtin_init_dwarf_reg_size_table: { |
1407 | 2 | Value *Address = EmitScalarExpr(E->getArg(0)); |
1408 | 2 | if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address)) |
1409 | 0 | CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table"); |
1410 | 2 | return RValue::get(llvm::UndefValue::get(ConvertType(E->getType()))); |
1411 | 2 | } |
1412 | 0 | case Builtin::BI__builtin_eh_return: { |
1413 | 0 | Value *Int = EmitScalarExpr(E->getArg(0)); |
1414 | 0 | Value *Ptr = EmitScalarExpr(E->getArg(1)); |
1415 | 0 |
|
1416 | 0 | llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType()); |
1417 | 0 | assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) && |
1418 | 0 | "LLVM's __builtin_eh_return only supports 32- and 64-bit variants"); |
1419 | 0 | Value *F = CGM.getIntrinsic(IntTy->getBitWidth() == 32 |
1420 | 0 | ? Intrinsic::eh_return_i32 |
1421 | 0 | : Intrinsic::eh_return_i64); |
1422 | 0 | Builder.CreateCall(F, {Int, Ptr}); |
1423 | 0 | Builder.CreateUnreachable(); |
1424 | 0 |
|
1425 | 0 | // We do need to preserve an insertion point. |
1426 | 0 | EmitBlock(createBasicBlock("builtin_eh_return.cont")); |
1427 | 0 |
|
1428 | 0 | return RValue::get(nullptr); |
1429 | 2 | } |
1430 | 1 | case Builtin::BI__builtin_unwind_init: { |
1431 | 1 | Value *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init); |
1432 | 1 | return RValue::get(Builder.CreateCall(F)); |
1433 | 2 | } |
1434 | 0 | case Builtin::BI__builtin_extend_pointer: { |
1435 | 0 | // Extends a pointer to the size of an _Unwind_Word, which is |
1436 | 0 | // uint64_t on all platforms. Generally this gets poked into a |
1437 | 0 | // register and eventually used as an address, so if the |
1438 | 0 | // addressing registers are wider than pointers and the platform |
1439 | 0 | // doesn't implicitly ignore high-order bits when doing |
1440 | 0 | // addressing, we need to make sure we zext / sext based on |
1441 | 0 | // the platform's expectations. |
1442 | 0 | // |
1443 | 0 | // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html |
1444 | 0 |
|
1445 | 0 | // Cast the pointer to intptr_t. |
1446 | 0 | Value *Ptr = EmitScalarExpr(E->getArg(0)); |
1447 | 0 | Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast"); |
1448 | 0 |
|
1449 | 0 | // If that's 64 bits, we're done. |
1450 | 0 | if (IntPtrTy->getBitWidth() == 64) |
1451 | 0 | return RValue::get(Result); |
1452 | 0 |
|
1453 | 0 | // Otherwise, ask the codegen data what to do. |
1454 | 0 | if (0 getTargetHooks().extendPointerWithSExt()0 ) |
1455 | 0 | return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext")); |
1456 | 0 | else |
1457 | 0 | return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext")); |
1458 | 0 | } |
1459 | 6 | case Builtin::BI__builtin_setjmp: { |
1460 | 6 | // Buffer is a void**. |
1461 | 6 | Address Buf = EmitPointerWithAlignment(E->getArg(0)); |
1462 | 6 | |
1463 | 6 | // Store the frame pointer to the setjmp buffer. |
1464 | 6 | Value *FrameAddr = |
1465 | 6 | Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), |
1466 | 6 | ConstantInt::get(Int32Ty, 0)); |
1467 | 6 | Builder.CreateStore(FrameAddr, Buf); |
1468 | 6 | |
1469 | 6 | // Store the stack pointer to the setjmp buffer. |
1470 | 6 | Value *StackAddr = |
1471 | 6 | Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave)); |
1472 | 6 | Address StackSaveSlot = |
1473 | 6 | Builder.CreateConstInBoundsGEP(Buf, 2, getPointerSize()); |
1474 | 6 | Builder.CreateStore(StackAddr, StackSaveSlot); |
1475 | 6 | |
1476 | 6 | // Call LLVM's EH setjmp, which is lightweight. |
1477 | 6 | Value *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp); |
1478 | 6 | Buf = Builder.CreateBitCast(Buf, Int8PtrTy); |
1479 | 6 | return RValue::get(Builder.CreateCall(F, Buf.getPointer())); |
1480 | 0 | } |
1481 | 8 | case Builtin::BI__builtin_longjmp: { |
1482 | 8 | Value *Buf = EmitScalarExpr(E->getArg(0)); |
1483 | 8 | Buf = Builder.CreateBitCast(Buf, Int8PtrTy); |
1484 | 8 | |
1485 | 8 | // Call LLVM's EH longjmp, which is lightweight. |
1486 | 8 | Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf); |
1487 | 8 | |
1488 | 8 | // longjmp doesn't return; mark this as unreachable. |
1489 | 8 | Builder.CreateUnreachable(); |
1490 | 8 | |
1491 | 8 | // We do need to preserve an insertion point. |
1492 | 8 | EmitBlock(createBasicBlock("longjmp.cont")); |
1493 | 8 | |
1494 | 8 | return RValue::get(nullptr); |
1495 | 0 | } |
1496 | 0 | case Builtin::BI__sync_fetch_and_add: |
1497 | 0 | case Builtin::BI__sync_fetch_and_sub: |
1498 | 0 | case Builtin::BI__sync_fetch_and_or: |
1499 | 0 | case Builtin::BI__sync_fetch_and_and: |
1500 | 0 | case Builtin::BI__sync_fetch_and_xor: |
1501 | 0 | case Builtin::BI__sync_fetch_and_nand: |
1502 | 0 | case Builtin::BI__sync_add_and_fetch: |
1503 | 0 | case Builtin::BI__sync_sub_and_fetch: |
1504 | 0 | case Builtin::BI__sync_and_and_fetch: |
1505 | 0 | case Builtin::BI__sync_or_and_fetch: |
1506 | 0 | case Builtin::BI__sync_xor_and_fetch: |
1507 | 0 | case Builtin::BI__sync_nand_and_fetch: |
1508 | 0 | case Builtin::BI__sync_val_compare_and_swap: |
1509 | 0 | case Builtin::BI__sync_bool_compare_and_swap: |
1510 | 0 | case Builtin::BI__sync_lock_test_and_set: |
1511 | 0 | case Builtin::BI__sync_lock_release: |
1512 | 0 | case Builtin::BI__sync_swap: |
1513 | 0 | llvm_unreachable("Shouldn't make it through sema"); |
1514 | 94 | case Builtin::BI__sync_fetch_and_add_1: |
1515 | 94 | case Builtin::BI__sync_fetch_and_add_2: |
1516 | 94 | case Builtin::BI__sync_fetch_and_add_4: |
1517 | 94 | case Builtin::BI__sync_fetch_and_add_8: |
1518 | 94 | case Builtin::BI__sync_fetch_and_add_16: |
1519 | 94 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E); |
1520 | 22 | case Builtin::BI__sync_fetch_and_sub_1: |
1521 | 22 | case Builtin::BI__sync_fetch_and_sub_2: |
1522 | 22 | case Builtin::BI__sync_fetch_and_sub_4: |
1523 | 22 | case Builtin::BI__sync_fetch_and_sub_8: |
1524 | 22 | case Builtin::BI__sync_fetch_and_sub_16: |
1525 | 22 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E); |
1526 | 17 | case Builtin::BI__sync_fetch_and_or_1: |
1527 | 17 | case Builtin::BI__sync_fetch_and_or_2: |
1528 | 17 | case Builtin::BI__sync_fetch_and_or_4: |
1529 | 17 | case Builtin::BI__sync_fetch_and_or_8: |
1530 | 17 | case Builtin::BI__sync_fetch_and_or_16: |
1531 | 17 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E); |
1532 | 17 | case Builtin::BI__sync_fetch_and_and_1: |
1533 | 17 | case Builtin::BI__sync_fetch_and_and_2: |
1534 | 17 | case Builtin::BI__sync_fetch_and_and_4: |
1535 | 17 | case Builtin::BI__sync_fetch_and_and_8: |
1536 | 17 | case Builtin::BI__sync_fetch_and_and_16: |
1537 | 17 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E); |
1538 | 17 | case Builtin::BI__sync_fetch_and_xor_1: |
1539 | 17 | case Builtin::BI__sync_fetch_and_xor_2: |
1540 | 17 | case Builtin::BI__sync_fetch_and_xor_4: |
1541 | 17 | case Builtin::BI__sync_fetch_and_xor_8: |
1542 | 17 | case Builtin::BI__sync_fetch_and_xor_16: |
1543 | 17 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E); |
1544 | 17 | case Builtin::BI__sync_fetch_and_nand_1: |
1545 | 17 | case Builtin::BI__sync_fetch_and_nand_2: |
1546 | 17 | case Builtin::BI__sync_fetch_and_nand_4: |
1547 | 17 | case Builtin::BI__sync_fetch_and_nand_8: |
1548 | 17 | case Builtin::BI__sync_fetch_and_nand_16: |
1549 | 17 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Nand, E); |
1550 | 17 | |
1551 | 17 | // Clang extensions: not overloaded yet. |
1552 | 1 | case Builtin::BI__sync_fetch_and_min: |
1553 | 1 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E); |
1554 | 1 | case Builtin::BI__sync_fetch_and_max: |
1555 | 1 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E); |
1556 | 1 | case Builtin::BI__sync_fetch_and_umin: |
1557 | 1 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E); |
1558 | 1 | case Builtin::BI__sync_fetch_and_umax: |
1559 | 1 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E); |
1560 | 17 | |
1561 | 12 | case Builtin::BI__sync_add_and_fetch_1: |
1562 | 12 | case Builtin::BI__sync_add_and_fetch_2: |
1563 | 12 | case Builtin::BI__sync_add_and_fetch_4: |
1564 | 12 | case Builtin::BI__sync_add_and_fetch_8: |
1565 | 12 | case Builtin::BI__sync_add_and_fetch_16: |
1566 | 12 | return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E, |
1567 | 12 | llvm::Instruction::Add); |
1568 | 10 | case Builtin::BI__sync_sub_and_fetch_1: |
1569 | 10 | case Builtin::BI__sync_sub_and_fetch_2: |
1570 | 10 | case Builtin::BI__sync_sub_and_fetch_4: |
1571 | 10 | case Builtin::BI__sync_sub_and_fetch_8: |
1572 | 10 | case Builtin::BI__sync_sub_and_fetch_16: |
1573 | 10 | return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E, |
1574 | 10 | llvm::Instruction::Sub); |
1575 | 9 | case Builtin::BI__sync_and_and_fetch_1: |
1576 | 9 | case Builtin::BI__sync_and_and_fetch_2: |
1577 | 9 | case Builtin::BI__sync_and_and_fetch_4: |
1578 | 9 | case Builtin::BI__sync_and_and_fetch_8: |
1579 | 9 | case Builtin::BI__sync_and_and_fetch_16: |
1580 | 9 | return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::And, E, |
1581 | 9 | llvm::Instruction::And); |
1582 | 9 | case Builtin::BI__sync_or_and_fetch_1: |
1583 | 9 | case Builtin::BI__sync_or_and_fetch_2: |
1584 | 9 | case Builtin::BI__sync_or_and_fetch_4: |
1585 | 9 | case Builtin::BI__sync_or_and_fetch_8: |
1586 | 9 | case Builtin::BI__sync_or_and_fetch_16: |
1587 | 9 | return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E, |
1588 | 9 | llvm::Instruction::Or); |
1589 | 10 | case Builtin::BI__sync_xor_and_fetch_1: |
1590 | 10 | case Builtin::BI__sync_xor_and_fetch_2: |
1591 | 10 | case Builtin::BI__sync_xor_and_fetch_4: |
1592 | 10 | case Builtin::BI__sync_xor_and_fetch_8: |
1593 | 10 | case Builtin::BI__sync_xor_and_fetch_16: |
1594 | 10 | return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E, |
1595 | 10 | llvm::Instruction::Xor); |
1596 | 9 | case Builtin::BI__sync_nand_and_fetch_1: |
1597 | 9 | case Builtin::BI__sync_nand_and_fetch_2: |
1598 | 9 | case Builtin::BI__sync_nand_and_fetch_4: |
1599 | 9 | case Builtin::BI__sync_nand_and_fetch_8: |
1600 | 9 | case Builtin::BI__sync_nand_and_fetch_16: |
1601 | 9 | return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Nand, E, |
1602 | 9 | llvm::Instruction::And, true); |
1603 | 9 | |
1604 | 43 | case Builtin::BI__sync_val_compare_and_swap_1: |
1605 | 43 | case Builtin::BI__sync_val_compare_and_swap_2: |
1606 | 43 | case Builtin::BI__sync_val_compare_and_swap_4: |
1607 | 43 | case Builtin::BI__sync_val_compare_and_swap_8: |
1608 | 43 | case Builtin::BI__sync_val_compare_and_swap_16: |
1609 | 43 | return RValue::get(MakeAtomicCmpXchgValue(*this, E, false)); |
1610 | 43 | |
1611 | 39 | case Builtin::BI__sync_bool_compare_and_swap_1: |
1612 | 39 | case Builtin::BI__sync_bool_compare_and_swap_2: |
1613 | 39 | case Builtin::BI__sync_bool_compare_and_swap_4: |
1614 | 39 | case Builtin::BI__sync_bool_compare_and_swap_8: |
1615 | 39 | case Builtin::BI__sync_bool_compare_and_swap_16: |
1616 | 39 | return RValue::get(MakeAtomicCmpXchgValue(*this, E, true)); |
1617 | 39 | |
1618 | 1 | case Builtin::BI__sync_swap_1: |
1619 | 1 | case Builtin::BI__sync_swap_2: |
1620 | 1 | case Builtin::BI__sync_swap_4: |
1621 | 1 | case Builtin::BI__sync_swap_8: |
1622 | 1 | case Builtin::BI__sync_swap_16: |
1623 | 1 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); |
1624 | 1 | |
1625 | 122 | case Builtin::BI__sync_lock_test_and_set_1: |
1626 | 122 | case Builtin::BI__sync_lock_test_and_set_2: |
1627 | 122 | case Builtin::BI__sync_lock_test_and_set_4: |
1628 | 122 | case Builtin::BI__sync_lock_test_and_set_8: |
1629 | 122 | case Builtin::BI__sync_lock_test_and_set_16: |
1630 | 122 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); |
1631 | 122 | |
1632 | 11 | case Builtin::BI__sync_lock_release_1: |
1633 | 11 | case Builtin::BI__sync_lock_release_2: |
1634 | 11 | case Builtin::BI__sync_lock_release_4: |
1635 | 11 | case Builtin::BI__sync_lock_release_8: |
1636 | 11 | case Builtin::BI__sync_lock_release_16: { |
1637 | 11 | Value *Ptr = EmitScalarExpr(E->getArg(0)); |
1638 | 11 | QualType ElTy = E->getArg(0)->getType()->getPointeeType(); |
1639 | 11 | CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy); |
1640 | 11 | llvm::Type *ITy = llvm::IntegerType::get(getLLVMContext(), |
1641 | 11 | StoreSize.getQuantity() * 8); |
1642 | 11 | Ptr = Builder.CreateBitCast(Ptr, ITy->getPointerTo()); |
1643 | 11 | llvm::StoreInst *Store = |
1644 | 11 | Builder.CreateAlignedStore(llvm::Constant::getNullValue(ITy), Ptr, |
1645 | 11 | StoreSize); |
1646 | 11 | Store->setAtomic(llvm::AtomicOrdering::Release); |
1647 | 11 | return RValue::get(nullptr); |
1648 | 11 | } |
1649 | 11 | |
1650 | 1.01k | case Builtin::BI__sync_synchronize: { |
1651 | 1.01k | // We assume this is supposed to correspond to a C++0x-style |
1652 | 1.01k | // sequentially-consistent fence (i.e. this is only usable for |
1653 | 1.01k | // synchonization, not device I/O or anything like that). This intrinsic |
1654 | 1.01k | // is really badly designed in the sense that in theory, there isn't |
1655 | 1.01k | // any way to safely use it... but in practice, it mostly works |
1656 | 1.01k | // to use it with non-atomic loads and stores to get acquire/release |
1657 | 1.01k | // semantics. |
1658 | 1.01k | Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent); |
1659 | 1.01k | return RValue::get(nullptr); |
1660 | 11 | } |
1661 | 11 | |
1662 | 30 | case Builtin::BI__builtin_nontemporal_load: |
1663 | 30 | return RValue::get(EmitNontemporalLoad(*this, E)); |
1664 | 64 | case Builtin::BI__builtin_nontemporal_store: |
1665 | 64 | return RValue::get(EmitNontemporalStore(*this, E)); |
1666 | 20 | case Builtin::BI__c11_atomic_is_lock_free: |
1667 | 20 | case Builtin::BI__atomic_is_lock_free: { |
1668 | 20 | // Call "bool __atomic_is_lock_free(size_t size, void *ptr)". For the |
1669 | 20 | // __c11 builtin, ptr is 0 (indicating a properly-aligned object), since |
1670 | 20 | // _Atomic(T) is always properly-aligned. |
1671 | 20 | const char *LibCallName = "__atomic_is_lock_free"; |
1672 | 20 | CallArgList Args; |
1673 | 20 | Args.add(RValue::get(EmitScalarExpr(E->getArg(0))), |
1674 | 20 | getContext().getSizeType()); |
1675 | 20 | if (BuiltinID == Builtin::BI__atomic_is_lock_free) |
1676 | 16 | Args.add(RValue::get(EmitScalarExpr(E->getArg(1))), |
1677 | 16 | getContext().VoidPtrTy); |
1678 | 20 | else |
1679 | 4 | Args.add(RValue::get(llvm::Constant::getNullValue(VoidPtrTy)), |
1680 | 4 | getContext().VoidPtrTy); |
1681 | 20 | const CGFunctionInfo &FuncInfo = |
1682 | 20 | CGM.getTypes().arrangeBuiltinFunctionCall(E->getType(), Args); |
1683 | 20 | llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FuncInfo); |
1684 | 20 | llvm::Constant *Func = CGM.CreateRuntimeFunction(FTy, LibCallName); |
1685 | 20 | return EmitCall(FuncInfo, CGCallee::forDirect(Func), |
1686 | 20 | ReturnValueSlot(), Args); |
1687 | 20 | } |
1688 | 20 | |
1689 | 8 | case Builtin::BI__atomic_test_and_set: { |
1690 | 8 | // Look at the argument type to determine whether this is a volatile |
1691 | 8 | // operation. The parameter type is always volatile. |
1692 | 8 | QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); |
1693 | 8 | bool Volatile = |
1694 | 8 | PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); |
1695 | 8 | |
1696 | 8 | Value *Ptr = EmitScalarExpr(E->getArg(0)); |
1697 | 8 | unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); |
1698 | 8 | Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); |
1699 | 8 | Value *NewVal = Builder.getInt8(1); |
1700 | 8 | Value *Order = EmitScalarExpr(E->getArg(1)); |
1701 | 8 | if (isa<llvm::ConstantInt>(Order)8 ) { |
1702 | 8 | int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); |
1703 | 8 | AtomicRMWInst *Result = nullptr; |
1704 | 8 | switch (ord) { |
1705 | 0 | case 0: // memory_order_relaxed |
1706 | 0 | default: // invalid order |
1707 | 0 | Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal, |
1708 | 0 | llvm::AtomicOrdering::Monotonic); |
1709 | 0 | break; |
1710 | 4 | case 1: // memory_order_consume |
1711 | 4 | case 2: // memory_order_acquire |
1712 | 4 | Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal, |
1713 | 4 | llvm::AtomicOrdering::Acquire); |
1714 | 4 | break; |
1715 | 0 | case 3: // memory_order_release |
1716 | 0 | Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal, |
1717 | 0 | llvm::AtomicOrdering::Release); |
1718 | 0 | break; |
1719 | 0 | case 4: // memory_order_acq_rel |
1720 | 0 |
|
1721 | 0 | Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal, |
1722 | 0 | llvm::AtomicOrdering::AcquireRelease); |
1723 | 0 | break; |
1724 | 4 | case 5: // memory_order_seq_cst |
1725 | 4 | Result = Builder.CreateAtomicRMW( |
1726 | 4 | llvm::AtomicRMWInst::Xchg, Ptr, NewVal, |
1727 | 4 | llvm::AtomicOrdering::SequentiallyConsistent); |
1728 | 4 | break; |
1729 | 8 | } |
1730 | 8 | Result->setVolatile(Volatile); |
1731 | 8 | return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); |
1732 | 8 | } |
1733 | 0 |
|
1734 | 0 | llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); |
1735 | 0 |
|
1736 | 0 | llvm::BasicBlock *BBs[5] = { |
1737 | 0 | createBasicBlock("monotonic", CurFn), |
1738 | 0 | createBasicBlock("acquire", CurFn), |
1739 | 0 | createBasicBlock("release", CurFn), |
1740 | 0 | createBasicBlock("acqrel", CurFn), |
1741 | 0 | createBasicBlock("seqcst", CurFn) |
1742 | 0 | }; |
1743 | 0 | llvm::AtomicOrdering Orders[5] = { |
1744 | 0 | llvm::AtomicOrdering::Monotonic, llvm::AtomicOrdering::Acquire, |
1745 | 0 | llvm::AtomicOrdering::Release, llvm::AtomicOrdering::AcquireRelease, |
1746 | 0 | llvm::AtomicOrdering::SequentiallyConsistent}; |
1747 | 0 |
|
1748 | 0 | Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); |
1749 | 0 | llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); |
1750 | 0 |
|
1751 | 0 | Builder.SetInsertPoint(ContBB); |
1752 | 0 | PHINode *Result = Builder.CreatePHI(Int8Ty, 5, "was_set"); |
1753 | 0 |
|
1754 | 0 | for (unsigned i = 0; i < 50 ; ++i0 ) { |
1755 | 0 | Builder.SetInsertPoint(BBs[i]); |
1756 | 0 | AtomicRMWInst *RMW = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, |
1757 | 0 | Ptr, NewVal, Orders[i]); |
1758 | 0 | RMW->setVolatile(Volatile); |
1759 | 0 | Result->addIncoming(RMW, BBs[i]); |
1760 | 0 | Builder.CreateBr(ContBB); |
1761 | 0 | } |
1762 | 0 |
|
1763 | 0 | SI->addCase(Builder.getInt32(0), BBs[0]); |
1764 | 0 | SI->addCase(Builder.getInt32(1), BBs[1]); |
1765 | 0 | SI->addCase(Builder.getInt32(2), BBs[1]); |
1766 | 0 | SI->addCase(Builder.getInt32(3), BBs[2]); |
1767 | 0 | SI->addCase(Builder.getInt32(4), BBs[3]); |
1768 | 0 | SI->addCase(Builder.getInt32(5), BBs[4]); |
1769 | 0 |
|
1770 | 0 | Builder.SetInsertPoint(ContBB); |
1771 | 0 | return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); |
1772 | 0 | } |
1773 | 0 |
|
1774 | 8 | case Builtin::BI__atomic_clear: { |
1775 | 8 | QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); |
1776 | 8 | bool Volatile = |
1777 | 8 | PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); |
1778 | 8 | |
1779 | 8 | Address Ptr = EmitPointerWithAlignment(E->getArg(0)); |
1780 | 8 | unsigned AddrSpace = Ptr.getPointer()->getType()->getPointerAddressSpace(); |
1781 | 8 | Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); |
1782 | 8 | Value *NewVal = Builder.getInt8(0); |
1783 | 8 | Value *Order = EmitScalarExpr(E->getArg(1)); |
1784 | 8 | if (isa<llvm::ConstantInt>(Order)8 ) { |
1785 | 8 | int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); |
1786 | 8 | StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); |
1787 | 8 | switch (ord) { |
1788 | 0 | case 0: // memory_order_relaxed |
1789 | 0 | default: // invalid order |
1790 | 0 | Store->setOrdering(llvm::AtomicOrdering::Monotonic); |
1791 | 0 | break; |
1792 | 4 | case 3: // memory_order_release |
1793 | 4 | Store->setOrdering(llvm::AtomicOrdering::Release); |
1794 | 4 | break; |
1795 | 4 | case 5: // memory_order_seq_cst |
1796 | 4 | Store->setOrdering(llvm::AtomicOrdering::SequentiallyConsistent); |
1797 | 4 | break; |
1798 | 8 | } |
1799 | 8 | return RValue::get(nullptr); |
1800 | 8 | } |
1801 | 0 |
|
1802 | 0 | llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); |
1803 | 0 |
|
1804 | 0 | llvm::BasicBlock *BBs[3] = { |
1805 | 0 | createBasicBlock("monotonic", CurFn), |
1806 | 0 | createBasicBlock("release", CurFn), |
1807 | 0 | createBasicBlock("seqcst", CurFn) |
1808 | 0 | }; |
1809 | 0 | llvm::AtomicOrdering Orders[3] = { |
1810 | 0 | llvm::AtomicOrdering::Monotonic, llvm::AtomicOrdering::Release, |
1811 | 0 | llvm::AtomicOrdering::SequentiallyConsistent}; |
1812 | 0 |
|
1813 | 0 | Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); |
1814 | 0 | llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); |
1815 | 0 |
|
1816 | 0 | for (unsigned i = 0; i < 30 ; ++i0 ) { |
1817 | 0 | Builder.SetInsertPoint(BBs[i]); |
1818 | 0 | StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); |
1819 | 0 | Store->setOrdering(Orders[i]); |
1820 | 0 | Builder.CreateBr(ContBB); |
1821 | 0 | } |
1822 | 0 |
|
1823 | 0 | SI->addCase(Builder.getInt32(0), BBs[0]); |
1824 | 0 | SI->addCase(Builder.getInt32(3), BBs[1]); |
1825 | 0 | SI->addCase(Builder.getInt32(5), BBs[2]); |
1826 | 0 |
|
1827 | 0 | Builder.SetInsertPoint(ContBB); |
1828 | 0 | return RValue::get(nullptr); |
1829 | 0 | } |
1830 | 0 |
|
1831 | 64 | case Builtin::BI__atomic_thread_fence: |
1832 | 64 | case Builtin::BI__atomic_signal_fence: |
1833 | 64 | case Builtin::BI__c11_atomic_thread_fence: |
1834 | 64 | case Builtin::BI__c11_atomic_signal_fence: { |
1835 | 64 | llvm::SyncScope::ID SSID; |
1836 | 64 | if (BuiltinID == Builtin::BI__atomic_signal_fence || |
1837 | 64 | BuiltinID == Builtin::BI__c11_atomic_signal_fence) |
1838 | 32 | SSID = llvm::SyncScope::SingleThread; |
1839 | 64 | else |
1840 | 32 | SSID = llvm::SyncScope::System; |
1841 | 64 | Value *Order = EmitScalarExpr(E->getArg(0)); |
1842 | 64 | if (isa<llvm::ConstantInt>(Order)64 ) { |
1843 | 0 | int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); |
1844 | 0 | switch (ord) { |
1845 | 0 | case 0: // memory_order_relaxed |
1846 | 0 | default: // invalid order |
1847 | 0 | break; |
1848 | 0 | case 1: // memory_order_consume |
1849 | 0 | case 2: // memory_order_acquire |
1850 | 0 | Builder.CreateFence(llvm::AtomicOrdering::Acquire, SSID); |
1851 | 0 | break; |
1852 | 0 | case 3: // memory_order_release |
1853 | 0 | Builder.CreateFence(llvm::AtomicOrdering::Release, SSID); |
1854 | 0 | break; |
1855 | 0 | case 4: // memory_order_acq_rel |
1856 | 0 | Builder.CreateFence(llvm::AtomicOrdering::AcquireRelease, SSID); |
1857 | 0 | break; |
1858 | 0 | case 5: // memory_order_seq_cst |
1859 | 0 | Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent, SSID); |
1860 | 0 | break; |
1861 | 0 | } |
1862 | 0 | return RValue::get(nullptr); |
1863 | 0 | } |
1864 | 64 | |
1865 | 64 | llvm::BasicBlock *AcquireBB, *ReleaseBB, *AcqRelBB, *SeqCstBB; |
1866 | 64 | AcquireBB = createBasicBlock("acquire", CurFn); |
1867 | 64 | ReleaseBB = createBasicBlock("release", CurFn); |
1868 | 64 | AcqRelBB = createBasicBlock("acqrel", CurFn); |
1869 | 64 | SeqCstBB = createBasicBlock("seqcst", CurFn); |
1870 | 64 | llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); |
1871 | 64 | |
1872 | 64 | Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); |
1873 | 64 | llvm::SwitchInst *SI = Builder.CreateSwitch(Order, ContBB); |
1874 | 64 | |
1875 | 64 | Builder.SetInsertPoint(AcquireBB); |
1876 | 64 | Builder.CreateFence(llvm::AtomicOrdering::Acquire, SSID); |
1877 | 64 | Builder.CreateBr(ContBB); |
1878 | 64 | SI->addCase(Builder.getInt32(1), AcquireBB); |
1879 | 64 | SI->addCase(Builder.getInt32(2), AcquireBB); |
1880 | 64 | |
1881 | 64 | Builder.SetInsertPoint(ReleaseBB); |
1882 | 64 | Builder.CreateFence(llvm::AtomicOrdering::Release, SSID); |
1883 | 64 | Builder.CreateBr(ContBB); |
1884 | 64 | SI->addCase(Builder.getInt32(3), ReleaseBB); |
1885 | 64 | |
1886 | 64 | Builder.SetInsertPoint(AcqRelBB); |
1887 | 64 | Builder.CreateFence(llvm::AtomicOrdering::AcquireRelease, SSID); |
1888 | 64 | Builder.CreateBr(ContBB); |
1889 | 64 | SI->addCase(Builder.getInt32(4), AcqRelBB); |
1890 | 64 | |
1891 | 64 | Builder.SetInsertPoint(SeqCstBB); |
1892 | 64 | Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent, SSID); |
1893 | 64 | Builder.CreateBr(ContBB); |
1894 | 64 | SI->addCase(Builder.getInt32(5), SeqCstBB); |
1895 | 64 | |
1896 | 64 | Builder.SetInsertPoint(ContBB); |
1897 | 64 | return RValue::get(nullptr); |
1898 | 64 | } |
1899 | 64 | |
1900 | 64 | // Library functions with special handling. |
1901 | 478 | case Builtin::BIsqrt: |
1902 | 478 | case Builtin::BIsqrtf: |
1903 | 478 | case Builtin::BIsqrtl: { |
1904 | 478 | // Transform a call to sqrt* into a @llvm.sqrt.* intrinsic call, but only |
1905 | 478 | // in finite- or unsafe-math mode (the intrinsic has different semantics |
1906 | 478 | // for handling negative numbers compared to the library function, so |
1907 | 478 | // -fmath-errno=0 is not enough). |
1908 | 478 | if (!FD->hasAttr<ConstAttr>()) |
1909 | 4 | break; |
1910 | 474 | if (474 !(CGM.getCodeGenOpts().UnsafeFPMath || |
1911 | 471 | CGM.getCodeGenOpts().NoNaNsFPMath)) |
1912 | 471 | break; |
1913 | 3 | Value *Arg0 = EmitScalarExpr(E->getArg(0)); |
1914 | 3 | llvm::Type *ArgType = Arg0->getType(); |
1915 | 3 | Value *F = CGM.getIntrinsic(Intrinsic::sqrt, ArgType); |
1916 | 3 | return RValue::get(Builder.CreateCall(F, Arg0)); |
1917 | 3 | } |
1918 | 3 | |
1919 | 289 | case Builtin::BI__builtin_pow: |
1920 | 289 | case Builtin::BI__builtin_powf: |
1921 | 289 | case Builtin::BI__builtin_powl: |
1922 | 289 | case Builtin::BIpow: |
1923 | 289 | case Builtin::BIpowf: |
1924 | 289 | case Builtin::BIpowl: { |
1925 | 289 | // Transform a call to pow* into a @llvm.pow.* intrinsic call. |
1926 | 289 | if (!FD->hasAttr<ConstAttr>()) |
1927 | 3 | break; |
1928 | 286 | Value *Base = EmitScalarExpr(E->getArg(0)); |
1929 | 286 | Value *Exponent = EmitScalarExpr(E->getArg(1)); |
1930 | 286 | llvm::Type *ArgType = Base->getType(); |
1931 | 286 | Value *F = CGM.getIntrinsic(Intrinsic::pow, ArgType); |
1932 | 286 | return RValue::get(Builder.CreateCall(F, {Base, Exponent})); |
1933 | 286 | } |
1934 | 286 | |
1935 | 9 | case Builtin::BIfma: |
1936 | 9 | case Builtin::BIfmaf: |
1937 | 9 | case Builtin::BIfmal: |
1938 | 9 | case Builtin::BI__builtin_fma: |
1939 | 9 | case Builtin::BI__builtin_fmaf: |
1940 | 9 | case Builtin::BI__builtin_fmal: { |
1941 | 9 | // Rewrite fma to intrinsic. |
1942 | 9 | Value *FirstArg = EmitScalarExpr(E->getArg(0)); |
1943 | 9 | llvm::Type *ArgType = FirstArg->getType(); |
1944 | 9 | Value *F = CGM.getIntrinsic(Intrinsic::fma, ArgType); |
1945 | 9 | return RValue::get( |
1946 | 9 | Builder.CreateCall(F, {FirstArg, EmitScalarExpr(E->getArg(1)), |
1947 | 9 | EmitScalarExpr(E->getArg(2))})); |
1948 | 9 | } |
1949 | 9 | |
1950 | 22 | case Builtin::BI__builtin_signbit: |
1951 | 22 | case Builtin::BI__builtin_signbitf: |
1952 | 22 | case Builtin::BI__builtin_signbitl: { |
1953 | 22 | return RValue::get( |
1954 | 22 | Builder.CreateZExt(EmitSignBit(*this, EmitScalarExpr(E->getArg(0))), |
1955 | 22 | ConvertType(E->getType()))); |
1956 | 22 | } |
1957 | 10 | case Builtin::BI__annotation: { |
1958 | 10 | // Re-encode each wide string to UTF8 and make an MDString. |
1959 | 10 | SmallVector<Metadata *, 1> Strings; |
1960 | 16 | for (const Expr *Arg : E->arguments()) { |
1961 | 16 | const auto *Str = cast<StringLiteral>(Arg->IgnoreParenCasts()); |
1962 | 16 | assert(Str->getCharByteWidth() == 2); |
1963 | 16 | StringRef WideBytes = Str->getBytes(); |
1964 | 16 | std::string StrUtf8; |
1965 | 16 | if (!convertUTF16ToUTF8String( |
1966 | 16 | makeArrayRef(WideBytes.data(), WideBytes.size()), StrUtf8)) { |
1967 | 0 | CGM.ErrorUnsupported(E, "non-UTF16 __annotation argument"); |
1968 | 0 | continue; |
1969 | 0 | } |
1970 | 16 | Strings.push_back(llvm::MDString::get(getLLVMContext(), StrUtf8)); |
1971 | 16 | } |
1972 | 10 | |
1973 | 10 | // Build and MDTuple of MDStrings and emit the intrinsic call. |
1974 | 10 | llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::codeview_annotation, {}); |
1975 | 10 | MDTuple *StrTuple = MDTuple::get(getLLVMContext(), Strings); |
1976 | 10 | Builder.CreateCall(F, MetadataAsValue::get(getLLVMContext(), StrTuple)); |
1977 | 10 | return RValue::getIgnored(); |
1978 | 22 | } |
1979 | 7 | case Builtin::BI__builtin_annotation: { |
1980 | 7 | llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0)); |
1981 | 7 | llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::annotation, |
1982 | 7 | AnnVal->getType()); |
1983 | 7 | |
1984 | 7 | // Get the annotation string, go through casts. Sema requires this to be a |
1985 | 7 | // non-wide string literal, potentially casted, so the cast<> is safe. |
1986 | 7 | const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts(); |
1987 | 7 | StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString(); |
1988 | 7 | return RValue::get(EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc())); |
1989 | 22 | } |
1990 | 30 | case Builtin::BI__builtin_addcb: |
1991 | 30 | case Builtin::BI__builtin_addcs: |
1992 | 30 | case Builtin::BI__builtin_addc: |
1993 | 30 | case Builtin::BI__builtin_addcl: |
1994 | 30 | case Builtin::BI__builtin_addcll: |
1995 | 30 | case Builtin::BI__builtin_subcb: |
1996 | 30 | case Builtin::BI__builtin_subcs: |
1997 | 30 | case Builtin::BI__builtin_subc: |
1998 | 30 | case Builtin::BI__builtin_subcl: |
1999 | 30 | case Builtin::BI__builtin_subcll: { |
2000 | 30 | |
2001 | 30 | // We translate all of these builtins from expressions of the form: |
2002 | 30 | // int x = ..., y = ..., carryin = ..., carryout, result; |
2003 | 30 | // result = __builtin_addc(x, y, carryin, &carryout); |
2004 | 30 | // |
2005 | 30 | // to LLVM IR of the form: |
2006 | 30 | // |
2007 | 30 | // %tmp1 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %x, i32 %y) |
2008 | 30 | // %tmpsum1 = extractvalue {i32, i1} %tmp1, 0 |
2009 | 30 | // %carry1 = extractvalue {i32, i1} %tmp1, 1 |
2010 | 30 | // %tmp2 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %tmpsum1, |
2011 | 30 | // i32 %carryin) |
2012 | 30 | // %result = extractvalue {i32, i1} %tmp2, 0 |
2013 | 30 | // %carry2 = extractvalue {i32, i1} %tmp2, 1 |
2014 | 30 | // %tmp3 = or i1 %carry1, %carry2 |
2015 | 30 | // %tmp4 = zext i1 %tmp3 to i32 |
2016 | 30 | // store i32 %tmp4, i32* %carryout |
2017 | 30 | |
2018 | 30 | // Scalarize our inputs. |
2019 | 30 | llvm::Value *X = EmitScalarExpr(E->getArg(0)); |
2020 | 30 | llvm::Value *Y = EmitScalarExpr(E->getArg(1)); |
2021 | 30 | llvm::Value *Carryin = EmitScalarExpr(E->getArg(2)); |
2022 | 30 | Address CarryOutPtr = EmitPointerWithAlignment(E->getArg(3)); |
2023 | 30 | |
2024 | 30 | // Decide if we are lowering to a uadd.with.overflow or usub.with.overflow. |
2025 | 30 | llvm::Intrinsic::ID IntrinsicId; |
2026 | 30 | switch (BuiltinID) { |
2027 | 0 | default: 0 llvm_unreachable0 ("Unknown multiprecision builtin id."); |
2028 | 15 | case Builtin::BI__builtin_addcb: |
2029 | 15 | case Builtin::BI__builtin_addcs: |
2030 | 15 | case Builtin::BI__builtin_addc: |
2031 | 15 | case Builtin::BI__builtin_addcl: |
2032 | 15 | case Builtin::BI__builtin_addcll: |
2033 | 15 | IntrinsicId = llvm::Intrinsic::uadd_with_overflow; |
2034 | 15 | break; |
2035 | 15 | case Builtin::BI__builtin_subcb: |
2036 | 15 | case Builtin::BI__builtin_subcs: |
2037 | 15 | case Builtin::BI__builtin_subc: |
2038 | 15 | case Builtin::BI__builtin_subcl: |
2039 | 15 | case Builtin::BI__builtin_subcll: |
2040 | 15 | IntrinsicId = llvm::Intrinsic::usub_with_overflow; |
2041 | 15 | break; |
2042 | 30 | } |
2043 | 30 | |
2044 | 30 | // Construct our resulting LLVM IR expression. |
2045 | 30 | llvm::Value *Carry1; |
2046 | 30 | llvm::Value *Sum1 = EmitOverflowIntrinsic(*this, IntrinsicId, |
2047 | 30 | X, Y, Carry1); |
2048 | 30 | llvm::Value *Carry2; |
2049 | 30 | llvm::Value *Sum2 = EmitOverflowIntrinsic(*this, IntrinsicId, |
2050 | 30 | Sum1, Carryin, Carry2); |
2051 | 30 | llvm::Value *CarryOut = Builder.CreateZExt(Builder.CreateOr(Carry1, Carry2), |
2052 | 30 | X->getType()); |
2053 | 30 | Builder.CreateStore(CarryOut, CarryOutPtr); |
2054 | 30 | return RValue::get(Sum2); |
2055 | 30 | } |
2056 | 30 | |
2057 | 33 | case Builtin::BI__builtin_add_overflow: |
2058 | 33 | case Builtin::BI__builtin_sub_overflow: |
2059 | 33 | case Builtin::BI__builtin_mul_overflow: { |
2060 | 33 | const clang::Expr *LeftArg = E->getArg(0); |
2061 | 33 | const clang::Expr *RightArg = E->getArg(1); |
2062 | 33 | const clang::Expr *ResultArg = E->getArg(2); |
2063 | 33 | |
2064 | 33 | clang::QualType ResultQTy = |
2065 | 33 | ResultArg->getType()->castAs<PointerType>()->getPointeeType(); |
2066 | 33 | |
2067 | 33 | WidthAndSignedness LeftInfo = |
2068 | 33 | getIntegerWidthAndSignedness(CGM.getContext(), LeftArg->getType()); |
2069 | 33 | WidthAndSignedness RightInfo = |
2070 | 33 | getIntegerWidthAndSignedness(CGM.getContext(), RightArg->getType()); |
2071 | 33 | WidthAndSignedness ResultInfo = |
2072 | 33 | getIntegerWidthAndSignedness(CGM.getContext(), ResultQTy); |
2073 | 33 | WidthAndSignedness EncompassingInfo = |
2074 | 33 | EncompassingIntegerType({LeftInfo, RightInfo, ResultInfo}); |
2075 | 33 | |
2076 | 33 | llvm::Type *EncompassingLLVMTy = |
2077 | 33 | llvm::IntegerType::get(CGM.getLLVMContext(), EncompassingInfo.Width); |
2078 | 33 | |
2079 | 33 | llvm::Type *ResultLLVMTy = CGM.getTypes().ConvertType(ResultQTy); |
2080 | 33 | |
2081 | 33 | llvm::Intrinsic::ID IntrinsicId; |
2082 | 33 | switch (BuiltinID) { |
2083 | 0 | default: |
2084 | 0 | llvm_unreachable("Unknown overflow builtin id."); |
2085 | 21 | case Builtin::BI__builtin_add_overflow: |
2086 | 21 | IntrinsicId = EncompassingInfo.Signed |
2087 | 9 | ? llvm::Intrinsic::sadd_with_overflow |
2088 | 12 | : llvm::Intrinsic::uadd_with_overflow; |
2089 | 21 | break; |
2090 | 6 | case Builtin::BI__builtin_sub_overflow: |
2091 | 6 | IntrinsicId = EncompassingInfo.Signed |
2092 | 3 | ? llvm::Intrinsic::ssub_with_overflow |
2093 | 3 | : llvm::Intrinsic::usub_with_overflow; |
2094 | 6 | break; |
2095 | 6 | case Builtin::BI__builtin_mul_overflow: |
2096 | 6 | IntrinsicId = EncompassingInfo.Signed |
2097 | 3 | ? llvm::Intrinsic::smul_with_overflow |
2098 | 3 | : llvm::Intrinsic::umul_with_overflow; |
2099 | 6 | break; |
2100 | 33 | } |
2101 | 33 | |
2102 | 33 | llvm::Value *Left = EmitScalarExpr(LeftArg); |
2103 | 33 | llvm::Value *Right = EmitScalarExpr(RightArg); |
2104 | 33 | Address ResultPtr = EmitPointerWithAlignment(ResultArg); |
2105 | 33 | |
2106 | 33 | // Extend each operand to the encompassing type. |
2107 | 33 | Left = Builder.CreateIntCast(Left, EncompassingLLVMTy, LeftInfo.Signed); |
2108 | 33 | Right = Builder.CreateIntCast(Right, EncompassingLLVMTy, RightInfo.Signed); |
2109 | 33 | |
2110 | 33 | // Perform the operation on the extended values. |
2111 | 33 | llvm::Value *Overflow, *Result; |
2112 | 33 | Result = EmitOverflowIntrinsic(*this, IntrinsicId, Left, Right, Overflow); |
2113 | 33 | |
2114 | 33 | if (EncompassingInfo.Width > ResultInfo.Width33 ) { |
2115 | 6 | // The encompassing type is wider than the result type, so we need to |
2116 | 6 | // truncate it. |
2117 | 6 | llvm::Value *ResultTrunc = Builder.CreateTrunc(Result, ResultLLVMTy); |
2118 | 6 | |
2119 | 6 | // To see if the truncation caused an overflow, we will extend |
2120 | 6 | // the result and then compare it to the original result. |
2121 | 6 | llvm::Value *ResultTruncExt = Builder.CreateIntCast( |
2122 | 6 | ResultTrunc, EncompassingLLVMTy, ResultInfo.Signed); |
2123 | 6 | llvm::Value *TruncationOverflow = |
2124 | 6 | Builder.CreateICmpNE(Result, ResultTruncExt); |
2125 | 6 | |
2126 | 6 | Overflow = Builder.CreateOr(Overflow, TruncationOverflow); |
2127 | 6 | Result = ResultTrunc; |
2128 | 6 | } |
2129 | 33 | |
2130 | 33 | // Finally, store the result using the pointer. |
2131 | 33 | bool isVolatile = |
2132 | 33 | ResultArg->getType()->getPointeeType().isVolatileQualified(); |
2133 | 33 | Builder.CreateStore(EmitToMemory(Result, ResultQTy), ResultPtr, isVolatile); |
2134 | 33 | |
2135 | 33 | return RValue::get(Overflow); |
2136 | 33 | } |
2137 | 33 | |
2138 | 54 | case Builtin::BI__builtin_uadd_overflow: |
2139 | 54 | case Builtin::BI__builtin_uaddl_overflow: |
2140 | 54 | case Builtin::BI__builtin_uaddll_overflow: |
2141 | 54 | case Builtin::BI__builtin_usub_overflow: |
2142 | 54 | case Builtin::BI__builtin_usubl_overflow: |
2143 | 54 | case Builtin::BI__builtin_usubll_overflow: |
2144 | 54 | case Builtin::BI__builtin_umul_overflow: |
2145 | 54 | case Builtin::BI__builtin_umull_overflow: |
2146 | 54 | case Builtin::BI__builtin_umulll_overflow: |
2147 | 54 | case Builtin::BI__builtin_sadd_overflow: |
2148 | 54 | case Builtin::BI__builtin_saddl_overflow: |
2149 | 54 | case Builtin::BI__builtin_saddll_overflow: |
2150 | 54 | case Builtin::BI__builtin_ssub_overflow: |
2151 | 54 | case Builtin::BI__builtin_ssubl_overflow: |
2152 | 54 | case Builtin::BI__builtin_ssubll_overflow: |
2153 | 54 | case Builtin::BI__builtin_smul_overflow: |
2154 | 54 | case Builtin::BI__builtin_smull_overflow: |
2155 | 54 | case Builtin::BI__builtin_smulll_overflow: { |
2156 | 54 | |
2157 | 54 | // We translate all of these builtins directly to the relevant llvm IR node. |
2158 | 54 | |
2159 | 54 | // Scalarize our inputs. |
2160 | 54 | llvm::Value *X = EmitScalarExpr(E->getArg(0)); |
2161 | 54 | llvm::Value *Y = EmitScalarExpr(E->getArg(1)); |
2162 | 54 | Address SumOutPtr = EmitPointerWithAlignment(E->getArg(2)); |
2163 | 54 | |
2164 | 54 | // Decide which of the overflow intrinsics we are lowering to: |
2165 | 54 | llvm::Intrinsic::ID IntrinsicId; |
2166 | 54 | switch (BuiltinID) { |
2167 | 0 | default: 0 llvm_unreachable0 ("Unknown overflow builtin id."); |
2168 | 9 | case Builtin::BI__builtin_uadd_overflow: |
2169 | 9 | case Builtin::BI__builtin_uaddl_overflow: |
2170 | 9 | case Builtin::BI__builtin_uaddll_overflow: |
2171 | 9 | IntrinsicId = llvm::Intrinsic::uadd_with_overflow; |
2172 | 9 | break; |
2173 | 9 | case Builtin::BI__builtin_usub_overflow: |
2174 | 9 | case Builtin::BI__builtin_usubl_overflow: |
2175 | 9 | case Builtin::BI__builtin_usubll_overflow: |
2176 | 9 | IntrinsicId = llvm::Intrinsic::usub_with_overflow; |
2177 | 9 | break; |
2178 | 9 | case Builtin::BI__builtin_umul_overflow: |
2179 | 9 | case Builtin::BI__builtin_umull_overflow: |
2180 | 9 | case Builtin::BI__builtin_umulll_overflow: |
2181 | 9 | IntrinsicId = llvm::Intrinsic::umul_with_overflow; |
2182 | 9 | break; |
2183 | 9 | case Builtin::BI__builtin_sadd_overflow: |
2184 | 9 | case Builtin::BI__builtin_saddl_overflow: |
2185 | 9 | case Builtin::BI__builtin_saddll_overflow: |
2186 | 9 | IntrinsicId = llvm::Intrinsic::sadd_with_overflow; |
2187 | 9 | break; |
2188 | 9 | case Builtin::BI__builtin_ssub_overflow: |
2189 | 9 | case Builtin::BI__builtin_ssubl_overflow: |
2190 | 9 | case Builtin::BI__builtin_ssubll_overflow: |
2191 | 9 | IntrinsicId = llvm::Intrinsic::ssub_with_overflow; |
2192 | 9 | break; |
2193 | 9 | case Builtin::BI__builtin_smul_overflow: |
2194 | 9 | case Builtin::BI__builtin_smull_overflow: |
2195 | 9 | case Builtin::BI__builtin_smulll_overflow: |
2196 | 9 | IntrinsicId = llvm::Intrinsic::smul_with_overflow; |
2197 | 9 | break; |
2198 | 54 | } |
2199 | 54 | |
2200 | 54 | |
2201 | 54 | llvm::Value *Carry; |
2202 | 54 | llvm::Value *Sum = EmitOverflowIntrinsic(*this, IntrinsicId, X, Y, Carry); |
2203 | 54 | Builder.CreateStore(Sum, SumOutPtr); |
2204 | 54 | |
2205 | 54 | return RValue::get(Carry); |
2206 | 54 | } |
2207 | 1 | case Builtin::BI__builtin_addressof: |
2208 | 1 | return RValue::get(EmitLValue(E->getArg(0)).getPointer()); |
2209 | 2 | case Builtin::BI__builtin_operator_new: |
2210 | 2 | return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(), |
2211 | 2 | E->getArg(0), false); |
2212 | 2 | case Builtin::BI__builtin_operator_delete: |
2213 | 2 | return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(), |
2214 | 2 | E->getArg(0), true); |
2215 | 6 | case Builtin::BI__noop: |
2216 | 6 | // __noop always evaluates to an integer literal zero. |
2217 | 6 | return RValue::get(ConstantInt::get(IntTy, 0)); |
2218 | 8 | case Builtin::BI__builtin_call_with_static_chain: { |
2219 | 8 | const CallExpr *Call = cast<CallExpr>(E->getArg(0)); |
2220 | 8 | const Expr *Chain = E->getArg(1); |
2221 | 8 | return EmitCall(Call->getCallee()->getType(), |
2222 | 8 | EmitCallee(Call->getCallee()), Call, ReturnValue, |
2223 | 8 | EmitScalarExpr(Chain)); |
2224 | 54 | } |
2225 | 14 | case Builtin::BI_InterlockedExchange8: |
2226 | 14 | case Builtin::BI_InterlockedExchange16: |
2227 | 14 | case Builtin::BI_InterlockedExchange: |
2228 | 14 | case Builtin::BI_InterlockedExchangePointer: |
2229 | 14 | return RValue::get( |
2230 | 14 | EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchange, E)); |
2231 | 3 | case Builtin::BI_InterlockedCompareExchangePointer: { |
2232 | 3 | llvm::Type *RTy; |
2233 | 3 | llvm::IntegerType *IntType = |
2234 | 3 | IntegerType::get(getLLVMContext(), |
2235 | 3 | getContext().getTypeSize(E->getType())); |
2236 | 3 | llvm::Type *IntPtrType = IntType->getPointerTo(); |
2237 | 3 | |
2238 | 3 | llvm::Value *Destination = |
2239 | 3 | Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), IntPtrType); |
2240 | 3 | |
2241 | 3 | llvm::Value *Exchange = EmitScalarExpr(E->getArg(1)); |
2242 | 3 | RTy = Exchange->getType(); |
2243 | 3 | Exchange = Builder.CreatePtrToInt(Exchange, IntType); |
2244 | 3 | |
2245 | 3 | llvm::Value *Comparand = |
2246 | 3 | Builder.CreatePtrToInt(EmitScalarExpr(E->getArg(2)), IntType); |
2247 | 3 | |
2248 | 3 | auto Result = |
2249 | 3 | Builder.CreateAtomicCmpXchg(Destination, Comparand, Exchange, |
2250 | 3 | AtomicOrdering::SequentiallyConsistent, |
2251 | 3 | AtomicOrdering::SequentiallyConsistent); |
2252 | 3 | Result->setVolatile(true); |
2253 | 3 | |
2254 | 3 | return RValue::get(Builder.CreateIntToPtr(Builder.CreateExtractValue(Result, |
2255 | 3 | 0), |
2256 | 3 | RTy)); |
2257 | 14 | } |
2258 | 14 | case Builtin::BI_InterlockedCompareExchange8: |
2259 | 14 | case Builtin::BI_InterlockedCompareExchange16: |
2260 | 14 | case Builtin::BI_InterlockedCompareExchange: |
2261 | 14 | case Builtin::BI_InterlockedCompareExchange64: { |
2262 | 14 | AtomicCmpXchgInst *CXI = Builder.CreateAtomicCmpXchg( |
2263 | 14 | EmitScalarExpr(E->getArg(0)), |
2264 | 14 | EmitScalarExpr(E->getArg(2)), |
2265 | 14 | EmitScalarExpr(E->getArg(1)), |
2266 | 14 | AtomicOrdering::SequentiallyConsistent, |
2267 | 14 | AtomicOrdering::SequentiallyConsistent); |
2268 | 14 | CXI->setVolatile(true); |
2269 | 14 | return RValue::get(Builder.CreateExtractValue(CXI, 0)); |
2270 | 14 | } |
2271 | 8 | case Builtin::BI_InterlockedIncrement16: |
2272 | 8 | case Builtin::BI_InterlockedIncrement: |
2273 | 8 | return RValue::get( |
2274 | 8 | EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedIncrement, E)); |
2275 | 8 | case Builtin::BI_InterlockedDecrement16: |
2276 | 8 | case Builtin::BI_InterlockedDecrement: |
2277 | 8 | return RValue::get( |
2278 | 8 | EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedDecrement, E)); |
2279 | 11 | case Builtin::BI_InterlockedAnd8: |
2280 | 11 | case Builtin::BI_InterlockedAnd16: |
2281 | 11 | case Builtin::BI_InterlockedAnd: |
2282 | 11 | return RValue::get(EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedAnd, E)); |
2283 | 11 | case Builtin::BI_InterlockedExchangeAdd8: |
2284 | 11 | case Builtin::BI_InterlockedExchangeAdd16: |
2285 | 11 | case Builtin::BI_InterlockedExchangeAdd: |
2286 | 11 | return RValue::get( |
2287 | 11 | EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeAdd, E)); |
2288 | 11 | case Builtin::BI_InterlockedExchangeSub8: |
2289 | 11 | case Builtin::BI_InterlockedExchangeSub16: |
2290 | 11 | case Builtin::BI_InterlockedExchangeSub: |
2291 | 11 | return RValue::get( |
2292 | 11 | EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeSub, E)); |
2293 | 11 | case Builtin::BI_InterlockedOr8: |
2294 | 11 | case Builtin::BI_InterlockedOr16: |
2295 | 11 | case Builtin::BI_InterlockedOr: |
2296 | 11 | return RValue::get(EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedOr, E)); |
2297 | 11 | case Builtin::BI_InterlockedXor8: |
2298 | 11 | case Builtin::BI_InterlockedXor16: |
2299 | 11 | case Builtin::BI_InterlockedXor: |
2300 | 11 | return RValue::get(EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedXor, E)); |
2301 | 5 | case Builtin::BI_interlockedbittestandset: |
2302 | 5 | return RValue::get( |
2303 | 5 | EmitMSVCBuiltinExpr(MSVCIntrin::_interlockedbittestandset, E)); |
2304 | 11 | |
2305 | 14 | case Builtin::BI__exception_code: |
2306 | 14 | case Builtin::BI_exception_code: |
2307 | 14 | return RValue::get(EmitSEHExceptionCode()); |
2308 | 0 | case Builtin::BI__exception_info: |
2309 | 0 | case Builtin::BI_exception_info: |
2310 | 0 | return RValue::get(EmitSEHExceptionInfo()); |
2311 | 2 | case Builtin::BI__abnormal_termination: |
2312 | 2 | case Builtin::BI_abnormal_termination: |
2313 | 2 | return RValue::get(EmitSEHAbnormalTermination()); |
2314 | 4 | case Builtin::BI_setjmpex: { |
2315 | 4 | if (getTarget().getTriple().isOSMSVCRT()4 ) { |
2316 | 4 | llvm::Type *ArgTypes[] = {Int8PtrTy, Int8PtrTy}; |
2317 | 4 | llvm::AttributeList ReturnsTwiceAttr = llvm::AttributeList::get( |
2318 | 4 | getLLVMContext(), llvm::AttributeList::FunctionIndex, |
2319 | 4 | llvm::Attribute::ReturnsTwice); |
2320 | 4 | llvm::Constant *SetJmpEx = CGM.CreateRuntimeFunction( |
2321 | 4 | llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/false), |
2322 | 4 | "_setjmpex", ReturnsTwiceAttr, /*Local=*/true); |
2323 | 4 | llvm::Value *Buf = Builder.CreateBitOrPointerCast( |
2324 | 4 | EmitScalarExpr(E->getArg(0)), Int8PtrTy); |
2325 | 4 | llvm::Value *FrameAddr = |
2326 | 4 | Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), |
2327 | 4 | ConstantInt::get(Int32Ty, 0)); |
2328 | 4 | llvm::Value *Args[] = {Buf, FrameAddr}; |
2329 | 4 | llvm::CallSite CS = EmitRuntimeCallOrInvoke(SetJmpEx, Args); |
2330 | 4 | CS.setAttributes(ReturnsTwiceAttr); |
2331 | 4 | return RValue::get(CS.getInstruction()); |
2332 | 4 | } |
2333 | 0 | break; |
2334 | 0 | } |
2335 | 9 | case Builtin::BI_setjmp: { |
2336 | 9 | if (getTarget().getTriple().isOSMSVCRT()9 ) { |
2337 | 4 | llvm::AttributeList ReturnsTwiceAttr = llvm::AttributeList::get( |
2338 | 4 | getLLVMContext(), llvm::AttributeList::FunctionIndex, |
2339 | 4 | llvm::Attribute::ReturnsTwice); |
2340 | 4 | llvm::Value *Buf = Builder.CreateBitOrPointerCast( |
2341 | 4 | EmitScalarExpr(E->getArg(0)), Int8PtrTy); |
2342 | 4 | llvm::CallSite CS; |
2343 | 4 | if (getTarget().getTriple().getArch() == llvm::Triple::x864 ) { |
2344 | 2 | llvm::Type *ArgTypes[] = {Int8PtrTy, IntTy}; |
2345 | 2 | llvm::Constant *SetJmp3 = CGM.CreateRuntimeFunction( |
2346 | 2 | llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/true), |
2347 | 2 | "_setjmp3", ReturnsTwiceAttr, /*Local=*/true); |
2348 | 2 | llvm::Value *Count = ConstantInt::get(IntTy, 0); |
2349 | 2 | llvm::Value *Args[] = {Buf, Count}; |
2350 | 2 | CS = EmitRuntimeCallOrInvoke(SetJmp3, Args); |
2351 | 4 | } else { |
2352 | 2 | llvm::Type *ArgTypes[] = {Int8PtrTy, Int8PtrTy}; |
2353 | 2 | llvm::Constant *SetJmp = CGM.CreateRuntimeFunction( |
2354 | 2 | llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/false), |
2355 | 2 | "_setjmp", ReturnsTwiceAttr, /*Local=*/true); |
2356 | 2 | llvm::Value *FrameAddr = |
2357 | 2 | Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), |
2358 | 2 | ConstantInt::get(Int32Ty, 0)); |
2359 | 2 | llvm::Value *Args[] = {Buf, FrameAddr}; |
2360 | 2 | CS = EmitRuntimeCallOrInvoke(SetJmp, Args); |
2361 | 2 | } |
2362 | 4 | CS.setAttributes(ReturnsTwiceAttr); |
2363 | 4 | return RValue::get(CS.getInstruction()); |
2364 | 4 | } |
2365 | 5 | break; |
2366 | 5 | } |
2367 | 5 | |
2368 | 4 | case Builtin::BI__GetExceptionInfo: { |
2369 | 4 | if (llvm::GlobalVariable *GV = |
2370 | 4 | CGM.getCXXABI().getThrowInfo(FD->getParamDecl(0)->getType())) |
2371 | 4 | return RValue::get(llvm::ConstantExpr::getBitCast(GV, CGM.Int8PtrTy)); |
2372 | 0 | break; |
2373 | 0 | } |
2374 | 0 |
|
2375 | 3 | case Builtin::BI__fastfail: |
2376 | 3 | return RValue::get(EmitMSVCBuiltinExpr(MSVCIntrin::__fastfail, E)); |
2377 | 0 |
|
2378 | 34 | case Builtin::BI__builtin_coro_size: { |
2379 | 34 | auto & Context = getContext(); |
2380 | 34 | auto SizeTy = Context.getSizeType(); |
2381 | 34 | auto T = Builder.getIntNTy(Context.getTypeSize(SizeTy)); |
2382 | 34 | Value *F = CGM.getIntrinsic(Intrinsic::coro_size, T); |
2383 | 34 | return RValue::get(Builder.CreateCall(F)); |
2384 | 0 | } |
2385 | 0 |
|
2386 | 3 | case Builtin::BI__builtin_coro_id: |
2387 | 3 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_id); |
2388 | 1 | case Builtin::BI__builtin_coro_promise: |
2389 | 1 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_promise); |
2390 | 2 | case Builtin::BI__builtin_coro_resume: |
2391 | 2 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_resume); |
2392 | 133 | case Builtin::BI__builtin_coro_frame: |
2393 | 133 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_frame); |
2394 | 44 | case Builtin::BI__builtin_coro_free: |
2395 | 44 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_free); |
2396 | 1 | case Builtin::BI__builtin_coro_destroy: |
2397 | 1 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_destroy); |
2398 | 1 | case Builtin::BI__builtin_coro_done: |
2399 | 1 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_done); |
2400 | 2 | case Builtin::BI__builtin_coro_alloc: |
2401 | 2 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_alloc); |
2402 | 2 | case Builtin::BI__builtin_coro_begin: |
2403 | 2 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_begin); |
2404 | 1 | case Builtin::BI__builtin_coro_end: |
2405 | 1 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_end); |
2406 | 1 | case Builtin::BI__builtin_coro_suspend: |
2407 | 1 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_suspend); |
2408 | 1 | case Builtin::BI__builtin_coro_param: |
2409 | 1 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_param); |
2410 | 0 |
|
2411 | 0 | // OpenCL v2.0 s6.13.16.2, Built-in pipe read and write functions |
2412 | 8 | case Builtin::BIread_pipe: |
2413 | 8 | case Builtin::BIwrite_pipe: { |
2414 | 8 | Value *Arg0 = EmitScalarExpr(E->getArg(0)), |
2415 | 8 | *Arg1 = EmitScalarExpr(E->getArg(1)); |
2416 | 8 | CGOpenCLRuntime OpenCLRT(CGM); |
2417 | 8 | Value *PacketSize = OpenCLRT.getPipeElemSize(E->getArg(0)); |
2418 | 8 | Value *PacketAlign = OpenCLRT.getPipeElemAlign(E->getArg(0)); |
2419 | 8 | |
2420 | 8 | // Type of the generic packet parameter. |
2421 | 8 | unsigned GenericAS = |
2422 | 8 | getContext().getTargetAddressSpace(LangAS::opencl_generic); |
2423 | 8 | llvm::Type *I8PTy = llvm::PointerType::get( |
2424 | 8 | llvm::Type::getInt8Ty(getLLVMContext()), GenericAS); |
2425 | 8 | |
2426 | 8 | // Testing which overloaded version we should generate the call for. |
2427 | 8 | if (2U == E->getNumArgs()8 ) { |
2428 | 4 | const char *Name = (BuiltinID == Builtin::BIread_pipe) ? "__read_pipe_2" |
2429 | 2 | : "__write_pipe_2"; |
2430 | 6 | // Creating a generic function type to be able to call with any builtin or |
2431 | 6 | // user defined type. |
2432 | 6 | llvm::Type *ArgTys[] = {Arg0->getType(), I8PTy, Int32Ty, Int32Ty}; |
2433 | 6 | llvm::FunctionType *FTy = llvm::FunctionType::get( |
2434 | 6 | Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); |
2435 | 6 | Value *BCast = Builder.CreatePointerCast(Arg1, I8PTy); |
2436 | 6 | return RValue::get( |
2437 | 6 | Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), |
2438 | 6 | {Arg0, BCast, PacketSize, PacketAlign})); |
2439 | 0 | } else { |
2440 | 2 | assert(4 == E->getNumArgs() && |
2441 | 2 | "Illegal number of parameters to pipe function"); |
2442 | 1 | const char *Name = (BuiltinID == Builtin::BIread_pipe) ? "__read_pipe_4" |
2443 | 1 | : "__write_pipe_4"; |
2444 | 2 | |
2445 | 2 | llvm::Type *ArgTys[] = {Arg0->getType(), Arg1->getType(), Int32Ty, I8PTy, |
2446 | 2 | Int32Ty, Int32Ty}; |
2447 | 2 | Value *Arg2 = EmitScalarExpr(E->getArg(2)), |
2448 | 2 | *Arg3 = EmitScalarExpr(E->getArg(3)); |
2449 | 2 | llvm::FunctionType *FTy = llvm::FunctionType::get( |
2450 | 2 | Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); |
2451 | 2 | Value *BCast = Builder.CreatePointerCast(Arg3, I8PTy); |
2452 | 2 | // We know the third argument is an integer type, but we may need to cast |
2453 | 2 | // it to i32. |
2454 | 2 | if (Arg2->getType() != Int32Ty) |
2455 | 0 | Arg2 = Builder.CreateZExtOrTrunc(Arg2, Int32Ty); |
2456 | 2 | return RValue::get(Builder.CreateCall( |
2457 | 2 | CGM.CreateRuntimeFunction(FTy, Name), |
2458 | 2 | {Arg0, Arg1, Arg2, BCast, PacketSize, PacketAlign})); |
2459 | 2 | } |
2460 | 0 | } |
2461 | 0 | // OpenCL v2.0 s6.13.16 ,s9.17.3.5 - Built-in pipe reserve read and write |
2462 | 0 | // functions |
2463 | 6 | case Builtin::BIreserve_read_pipe: |
2464 | 6 | case Builtin::BIreserve_write_pipe: |
2465 | 6 | case Builtin::BIwork_group_reserve_read_pipe: |
2466 | 6 | case Builtin::BIwork_group_reserve_write_pipe: |
2467 | 6 | case Builtin::BIsub_group_reserve_read_pipe: |
2468 | 6 | case Builtin::BIsub_group_reserve_write_pipe: { |
2469 | 6 | // Composing the mangled name for the function. |
2470 | 6 | const char *Name; |
2471 | 6 | if (BuiltinID == Builtin::BIreserve_read_pipe) |
2472 | 1 | Name = "__reserve_read_pipe"; |
2473 | 5 | else if (5 BuiltinID == Builtin::BIreserve_write_pipe5 ) |
2474 | 1 | Name = "__reserve_write_pipe"; |
2475 | 4 | else if (4 BuiltinID == Builtin::BIwork_group_reserve_read_pipe4 ) |
2476 | 1 | Name = "__work_group_reserve_read_pipe"; |
2477 | 3 | else if (3 BuiltinID == Builtin::BIwork_group_reserve_write_pipe3 ) |
2478 | 1 | Name = "__work_group_reserve_write_pipe"; |
2479 | 2 | else if (2 BuiltinID == Builtin::BIsub_group_reserve_read_pipe2 ) |
2480 | 1 | Name = "__sub_group_reserve_read_pipe"; |
2481 | 2 | else |
2482 | 1 | Name = "__sub_group_reserve_write_pipe"; |
2483 | 6 | |
2484 | 6 | Value *Arg0 = EmitScalarExpr(E->getArg(0)), |
2485 | 6 | *Arg1 = EmitScalarExpr(E->getArg(1)); |
2486 | 6 | llvm::Type *ReservedIDTy = ConvertType(getContext().OCLReserveIDTy); |
2487 | 6 | CGOpenCLRuntime OpenCLRT(CGM); |
2488 | 6 | Value *PacketSize = OpenCLRT.getPipeElemSize(E->getArg(0)); |
2489 | 6 | Value *PacketAlign = OpenCLRT.getPipeElemAlign(E->getArg(0)); |
2490 | 6 | |
2491 | 6 | // Building the generic function prototype. |
2492 | 6 | llvm::Type *ArgTys[] = {Arg0->getType(), Int32Ty, Int32Ty, Int32Ty}; |
2493 | 6 | llvm::FunctionType *FTy = llvm::FunctionType::get( |
2494 | 6 | ReservedIDTy, llvm::ArrayRef<llvm::Type *>(ArgTys), false); |
2495 | 6 | // We know the second argument is an integer type, but we may need to cast |
2496 | 6 | // it to i32. |
2497 | 6 | if (Arg1->getType() != Int32Ty) |
2498 | 0 | Arg1 = Builder.CreateZExtOrTrunc(Arg1, Int32Ty); |
2499 | 6 | return RValue::get( |
2500 | 6 | Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), |
2501 | 6 | {Arg0, Arg1, PacketSize, PacketAlign})); |
2502 | 6 | } |
2503 | 6 | // OpenCL v2.0 s6.13.16, s9.17.3.5 - Built-in pipe commit read and write |
2504 | 6 | // functions |
2505 | 6 | case Builtin::BIcommit_read_pipe: |
2506 | 6 | case Builtin::BIcommit_write_pipe: |
2507 | 6 | case Builtin::BIwork_group_commit_read_pipe: |
2508 | 6 | case Builtin::BIwork_group_commit_write_pipe: |
2509 | 6 | case Builtin::BIsub_group_commit_read_pipe: |
2510 | 6 | case Builtin::BIsub_group_commit_write_pipe: { |
2511 | 6 | const char *Name; |
2512 | 6 | if (BuiltinID == Builtin::BIcommit_read_pipe) |
2513 | 1 | Name = "__commit_read_pipe"; |
2514 | 5 | else if (5 BuiltinID == Builtin::BIcommit_write_pipe5 ) |
2515 | 1 | Name = "__commit_write_pipe"; |
2516 | 4 | else if (4 BuiltinID == Builtin::BIwork_group_commit_read_pipe4 ) |
2517 | 1 | Name = "__work_group_commit_read_pipe"; |
2518 | 3 | else if (3 BuiltinID == Builtin::BIwork_group_commit_write_pipe3 ) |
2519 | 1 | Name = "__work_group_commit_write_pipe"; |
2520 | 2 | else if (2 BuiltinID == Builtin::BIsub_group_commit_read_pipe2 ) |
2521 | 1 | Name = "__sub_group_commit_read_pipe"; |
2522 | 2 | else |
2523 | 1 | Name = "__sub_group_commit_write_pipe"; |
2524 | 6 | |
2525 | 6 | Value *Arg0 = EmitScalarExpr(E->getArg(0)), |
2526 | 6 | *Arg1 = EmitScalarExpr(E->getArg(1)); |
2527 | 6 | CGOpenCLRuntime OpenCLRT(CGM); |
2528 | 6 | Value *PacketSize = OpenCLRT.getPipeElemSize(E->getArg(0)); |
2529 | 6 | Value *PacketAlign = OpenCLRT.getPipeElemAlign(E->getArg(0)); |
2530 | 6 | |
2531 | 6 | // Building the generic function prototype. |
2532 | 6 | llvm::Type *ArgTys[] = {Arg0->getType(), Arg1->getType(), Int32Ty, Int32Ty}; |
2533 | 6 | llvm::FunctionType *FTy = |
2534 | 6 | llvm::FunctionType::get(llvm::Type::getVoidTy(getLLVMContext()), |
2535 | 6 | llvm::ArrayRef<llvm::Type *>(ArgTys), false); |
2536 | 6 | |
2537 | 6 | return RValue::get( |
2538 | 6 | Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), |
2539 | 6 | {Arg0, Arg1, PacketSize, PacketAlign})); |
2540 | 6 | } |
2541 | 6 | // OpenCL v2.0 s6.13.16.4 Built-in pipe query functions |
2542 | 4 | case Builtin::BIget_pipe_num_packets: |
2543 | 4 | case Builtin::BIget_pipe_max_packets: { |
2544 | 4 | const char *Name; |
2545 | 4 | if (BuiltinID == Builtin::BIget_pipe_num_packets) |
2546 | 2 | Name = "__get_pipe_num_packets"; |
2547 | 4 | else |
2548 | 2 | Name = "__get_pipe_max_packets"; |
2549 | 4 | |
2550 | 4 | // Building the generic function prototype. |
2551 | 4 | Value *Arg0 = EmitScalarExpr(E->getArg(0)); |
2552 | 4 | CGOpenCLRuntime OpenCLRT(CGM); |
2553 | 4 | Value *PacketSize = OpenCLRT.getPipeElemSize(E->getArg(0)); |
2554 | 4 | Value *PacketAlign = OpenCLRT.getPipeElemAlign(E->getArg(0)); |
2555 | 4 | llvm::Type *ArgTys[] = {Arg0->getType(), Int32Ty, Int32Ty}; |
2556 | 4 | llvm::FunctionType *FTy = llvm::FunctionType::get( |
2557 | 4 | Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); |
2558 | 4 | |
2559 | 4 | return RValue::get(Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), |
2560 | 4 | {Arg0, PacketSize, PacketAlign})); |
2561 | 4 | } |
2562 | 4 | |
2563 | 4 | // OpenCL v2.0 s6.13.9 - Address space qualifier functions. |
2564 | 14 | case Builtin::BIto_global: |
2565 | 14 | case Builtin::BIto_local: |
2566 | 14 | case Builtin::BIto_private: { |
2567 | 14 | auto Arg0 = EmitScalarExpr(E->getArg(0)); |
2568 | 14 | auto NewArgT = llvm::PointerType::get(Int8Ty, |
2569 | 14 | CGM.getContext().getTargetAddressSpace(LangAS::opencl_generic)); |
2570 | 14 | auto NewRetT = llvm::PointerType::get(Int8Ty, |
2571 | 14 | CGM.getContext().getTargetAddressSpace( |
2572 | 14 | E->getType()->getPointeeType().getAddressSpace())); |
2573 | 14 | auto FTy = llvm::FunctionType::get(NewRetT, {NewArgT}, false); |
2574 | 14 | llvm::Value *NewArg; |
2575 | 14 | if (Arg0->getType()->getPointerAddressSpace() != |
2576 | 14 | NewArgT->getPointerAddressSpace()) |
2577 | 10 | NewArg = Builder.CreateAddrSpaceCast(Arg0, NewArgT); |
2578 | 14 | else |
2579 | 4 | NewArg = Builder.CreateBitOrPointerCast(Arg0, NewArgT); |
2580 | 14 | auto NewName = std::string("__") + E->getDirectCallee()->getName().str(); |
2581 | 14 | auto NewCall = |
2582 | 14 | Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, NewName), {NewArg}); |
2583 | 14 | return RValue::get(Builder.CreateBitOrPointerCast(NewCall, |
2584 | 14 | ConvertType(E->getType()))); |
2585 | 14 | } |
2586 | 14 | |
2587 | 14 | // OpenCL v2.0, s6.13.17 - Enqueue kernel function. |
2588 | 14 | // It contains four different overload formats specified in Table 6.13.17.1. |
2589 | 18 | case Builtin::BIenqueue_kernel: { |
2590 | 18 | StringRef Name; // Generated function call name |
2591 | 18 | unsigned NumArgs = E->getNumArgs(); |
2592 | 18 | |
2593 | 18 | llvm::Type *QueueTy = ConvertType(getContext().OCLQueueTy); |
2594 | 18 | llvm::Type *GenericVoidPtrTy = Builder.getInt8PtrTy( |
2595 | 18 | getContext().getTargetAddressSpace(LangAS::opencl_generic)); |
2596 | 18 | |
2597 | 18 | llvm::Value *Queue = EmitScalarExpr(E->getArg(0)); |
2598 | 18 | llvm::Value *Flags = EmitScalarExpr(E->getArg(1)); |
2599 | 18 | LValue NDRangeL = EmitAggExprToLValue(E->getArg(2)); |
2600 | 18 | llvm::Value *Range = NDRangeL.getAddress().getPointer(); |
2601 | 18 | llvm::Type *RangeTy = NDRangeL.getAddress().getType(); |
2602 | 18 | |
2603 | 18 | if (NumArgs == 418 ) { |
2604 | 2 | // The most basic form of the call with parameters: |
2605 | 2 | // queue_t, kernel_enqueue_flags_t, ndrange_t, block(void) |
2606 | 2 | Name = "__enqueue_kernel_basic"; |
2607 | 2 | llvm::Type *ArgTys[] = {QueueTy, Int32Ty, RangeTy, GenericVoidPtrTy}; |
2608 | 2 | llvm::FunctionType *FTy = llvm::FunctionType::get( |
2609 | 2 | Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys, 4), false); |
2610 | 2 | |
2611 | 2 | llvm::Value *Block = Builder.CreatePointerCast( |
2612 | 2 | EmitScalarExpr(E->getArg(3)), GenericVoidPtrTy); |
2613 | 2 | |
2614 | 2 | AttrBuilder B; |
2615 | 2 | B.addAttribute(Attribute::ByVal); |
2616 | 2 | llvm::AttributeList ByValAttrSet = |
2617 | 2 | llvm::AttributeList::get(CGM.getModule().getContext(), 3U, B); |
2618 | 2 | |
2619 | 2 | auto RTCall = |
2620 | 2 | Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name, ByValAttrSet), |
2621 | 2 | {Queue, Flags, Range, Block}); |
2622 | 2 | RTCall->setAttributes(ByValAttrSet); |
2623 | 2 | return RValue::get(RTCall); |
2624 | 2 | } |
2625 | 18 | assert(NumArgs >= 5 && "Invalid enqueue_kernel signature"); |
2626 | 16 | |
2627 | 16 | // Create a temporary array to hold the sizes of local pointer arguments |
2628 | 16 | // for the block. \p First is the position of the first size argument. |
2629 | 14 | auto CreateArrayForSizeVar = [=](unsigned First) { |
2630 | 14 | auto *AT = llvm::ArrayType::get(SizeTy, NumArgs - First); |
2631 | 14 | auto *Arr = Builder.CreateAlloca(AT); |
2632 | 14 | llvm::Value *Ptr; |
2633 | 14 | // Each of the following arguments specifies the size of the corresponding |
2634 | 14 | // argument passed to the enqueued block. |
2635 | 14 | auto *Zero = llvm::ConstantInt::get(IntTy, 0); |
2636 | 32 | for (unsigned I = First; I < NumArgs32 ; ++I18 ) { |
2637 | 18 | auto *Index = llvm::ConstantInt::get(IntTy, I - First); |
2638 | 18 | auto *GEP = Builder.CreateGEP(Arr, {Zero, Index}); |
2639 | 18 | if (I == First) |
2640 | 14 | Ptr = GEP; |
2641 | 18 | auto *V = |
2642 | 18 | Builder.CreateZExtOrTrunc(EmitScalarExpr(E->getArg(I)), SizeTy); |
2643 | 18 | Builder.CreateAlignedStore( |
2644 | 18 | V, GEP, CGM.getDataLayout().getPrefTypeAlignment(SizeTy)); |
2645 | 18 | } |
2646 | 14 | return Ptr; |
2647 | 14 | }; |
2648 | 16 | |
2649 | 16 | // Could have events and/or vaargs. |
2650 | 16 | if (E->getArg(3)->getType()->isBlockPointerType()16 ) { |
2651 | 10 | // No events passed, but has variadic arguments. |
2652 | 10 | Name = "__enqueue_kernel_vaargs"; |
2653 | 10 | auto *Block = Builder.CreatePointerCast(EmitScalarExpr(E->getArg(3)), |
2654 | 10 | GenericVoidPtrTy); |
2655 | 10 | auto *PtrToSizeArray = CreateArrayForSizeVar(4); |
2656 | 10 | |
2657 | 10 | // Create a vector of the arguments, as well as a constant value to |
2658 | 10 | // express to the runtime the number of variadic arguments. |
2659 | 10 | std::vector<llvm::Value *> Args = {Queue, |
2660 | 10 | Flags, |
2661 | 10 | Range, |
2662 | 10 | Block, |
2663 | 10 | ConstantInt::get(IntTy, NumArgs - 4), |
2664 | 10 | PtrToSizeArray}; |
2665 | 10 | std::vector<llvm::Type *> ArgTys = {QueueTy, IntTy, |
2666 | 10 | RangeTy, GenericVoidPtrTy, |
2667 | 10 | IntTy, PtrToSizeArray->getType()}; |
2668 | 10 | |
2669 | 10 | llvm::FunctionType *FTy = llvm::FunctionType::get( |
2670 | 10 | Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); |
2671 | 10 | return RValue::get( |
2672 | 10 | Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), |
2673 | 10 | llvm::ArrayRef<llvm::Value *>(Args))); |
2674 | 10 | } |
2675 | 6 | // Any calls now have event arguments passed. |
2676 | 6 | if (6 NumArgs >= 76 ) { |
2677 | 6 | llvm::Type *EventTy = ConvertType(getContext().OCLClkEventTy); |
2678 | 6 | llvm::Type *EventPtrTy = EventTy->getPointerTo( |
2679 | 6 | CGM.getContext().getTargetAddressSpace(LangAS::opencl_generic)); |
2680 | 6 | |
2681 | 6 | llvm::Value *NumEvents = |
2682 | 6 | Builder.CreateZExtOrTrunc(EmitScalarExpr(E->getArg(3)), Int32Ty); |
2683 | 6 | llvm::Value *EventList = |
2684 | 6 | E->getArg(4)->getType()->isArrayType() |
2685 | 4 | ? EmitArrayToPointerDecay(E->getArg(4)).getPointer() |
2686 | 2 | : EmitScalarExpr(E->getArg(4)); |
2687 | 6 | llvm::Value *ClkEvent = EmitScalarExpr(E->getArg(5)); |
2688 | 6 | // Convert to generic address space. |
2689 | 6 | EventList = Builder.CreatePointerCast(EventList, EventPtrTy); |
2690 | 6 | ClkEvent = Builder.CreatePointerCast(ClkEvent, EventPtrTy); |
2691 | 6 | llvm::Value *Block = Builder.CreatePointerCast( |
2692 | 6 | EmitScalarExpr(E->getArg(6)), GenericVoidPtrTy); |
2693 | 6 | |
2694 | 6 | std::vector<llvm::Type *> ArgTys = { |
2695 | 6 | QueueTy, Int32Ty, RangeTy, Int32Ty, |
2696 | 6 | EventPtrTy, EventPtrTy, GenericVoidPtrTy}; |
2697 | 6 | |
2698 | 6 | std::vector<llvm::Value *> Args = {Queue, Flags, Range, NumEvents, |
2699 | 6 | EventList, ClkEvent, Block}; |
2700 | 6 | |
2701 | 6 | if (NumArgs == 76 ) { |
2702 | 2 | // Has events but no variadics. |
2703 | 2 | Name = "__enqueue_kernel_basic_events"; |
2704 | 2 | llvm::FunctionType *FTy = llvm::FunctionType::get( |
2705 | 2 | Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); |
2706 | 2 | return RValue::get( |
2707 | 2 | Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), |
2708 | 2 | llvm::ArrayRef<llvm::Value *>(Args))); |
2709 | 2 | } |
2710 | 4 | // Has event info and variadics |
2711 | 4 | // Pass the number of variadics to the runtime function too. |
2712 | 4 | Args.push_back(ConstantInt::get(Int32Ty, NumArgs - 7)); |
2713 | 4 | ArgTys.push_back(Int32Ty); |
2714 | 4 | Name = "__enqueue_kernel_events_vaargs"; |
2715 | 4 | |
2716 | 4 | auto *PtrToSizeArray = CreateArrayForSizeVar(7); |
2717 | 4 | Args.push_back(PtrToSizeArray); |
2718 | 4 | ArgTys.push_back(PtrToSizeArray->getType()); |
2719 | 4 | |
2720 | 4 | llvm::FunctionType *FTy = llvm::FunctionType::get( |
2721 | 4 | Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); |
2722 | 4 | return RValue::get( |
2723 | 4 | Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), |
2724 | 4 | llvm::ArrayRef<llvm::Value *>(Args))); |
2725 | 4 | } |
2726 | 0 | LLVM_FALLTHROUGH0 ; |
2727 | 0 | } |
2728 | 0 | // OpenCL v2.0 s6.13.17.6 - Kernel query functions need bitcast of block |
2729 | 0 | // parameter. |
2730 | 4 | case Builtin::BIget_kernel_work_group_size: { |
2731 | 4 | llvm::Type *GenericVoidPtrTy = Builder.getInt8PtrTy( |
2732 | 4 | getContext().getTargetAddressSpace(LangAS::opencl_generic)); |
2733 | 4 | Value *Arg = EmitScalarExpr(E->getArg(0)); |
2734 | 4 | Arg = Builder.CreatePointerCast(Arg, GenericVoidPtrTy); |
2735 | 4 | return RValue::get(Builder.CreateCall( |
2736 | 4 | CGM.CreateRuntimeFunction( |
2737 | 4 | llvm::FunctionType::get(IntTy, GenericVoidPtrTy, false), |
2738 | 4 | "__get_kernel_work_group_size_impl"), |
2739 | 4 | Arg)); |
2740 | 0 | } |
2741 | 4 | case Builtin::BIget_kernel_preferred_work_group_size_multiple: { |
2742 | 4 | llvm::Type *GenericVoidPtrTy = Builder.getInt8PtrTy( |
2743 | 4 | getContext().getTargetAddressSpace(LangAS::opencl_generic)); |
2744 | 4 | Value *Arg = EmitScalarExpr(E->getArg(0)); |
2745 | 4 | Arg = Builder.CreatePointerCast(Arg, GenericVoidPtrTy); |
2746 | 4 | return RValue::get(Builder.CreateCall( |
2747 | 4 | CGM.CreateRuntimeFunction( |
2748 | 4 | llvm::FunctionType::get(IntTy, GenericVoidPtrTy, false), |
2749 | 4 | "__get_kernel_preferred_work_group_multiple_impl"), |
2750 | 4 | Arg)); |
2751 | 0 | } |
2752 | 4 | case Builtin::BIget_kernel_max_sub_group_size_for_ndrange: |
2753 | 4 | case Builtin::BIget_kernel_sub_group_count_for_ndrange: { |
2754 | 4 | llvm::Type *GenericVoidPtrTy = Builder.getInt8PtrTy( |
2755 | 4 | getContext().getTargetAddressSpace(LangAS::opencl_generic)); |
2756 | 4 | LValue NDRangeL = EmitAggExprToLValue(E->getArg(0)); |
2757 | 4 | llvm::Value *NDRange = NDRangeL.getAddress().getPointer(); |
2758 | 4 | Value *Block = EmitScalarExpr(E->getArg(1)); |
2759 | 4 | Block = Builder.CreatePointerCast(Block, GenericVoidPtrTy); |
2760 | 4 | const char *Name = |
2761 | 4 | BuiltinID == Builtin::BIget_kernel_max_sub_group_size_for_ndrange |
2762 | 2 | ? "__get_kernel_max_sub_group_size_for_ndrange_impl" |
2763 | 2 | : "__get_kernel_sub_group_count_for_ndrange_impl"; |
2764 | 4 | return RValue::get(Builder.CreateCall( |
2765 | 4 | CGM.CreateRuntimeFunction( |
2766 | 4 | llvm::FunctionType::get( |
2767 | 4 | IntTy, {NDRange->getType(), GenericVoidPtrTy}, false), |
2768 | 4 | Name), |
2769 | 4 | {NDRange, Block})); |
2770 | 4 | } |
2771 | 4 | |
2772 | 6 | case Builtin::BI__builtin_store_half: |
2773 | 6 | case Builtin::BI__builtin_store_halff: { |
2774 | 6 | Value *Val = EmitScalarExpr(E->getArg(0)); |
2775 | 6 | Address Address = EmitPointerWithAlignment(E->getArg(1)); |
2776 | 6 | Value *HalfVal = Builder.CreateFPTrunc(Val, Builder.getHalfTy()); |
2777 | 6 | return RValue::get(Builder.CreateStore(HalfVal, Address)); |
2778 | 6 | } |
2779 | 3 | case Builtin::BI__builtin_load_half: { |
2780 | 3 | Address Address = EmitPointerWithAlignment(E->getArg(0)); |
2781 | 3 | Value *HalfVal = Builder.CreateLoad(Address); |
2782 | 3 | return RValue::get(Builder.CreateFPExt(HalfVal, Builder.getDoubleTy())); |
2783 | 6 | } |
2784 | 3 | case Builtin::BI__builtin_load_halff: { |
2785 | 3 | Address Address = EmitPointerWithAlignment(E->getArg(0)); |
2786 | 3 | Value *HalfVal = Builder.CreateLoad(Address); |
2787 | 3 | return RValue::get(Builder.CreateFPExt(HalfVal, Builder.getFloatTy())); |
2788 | 6 | } |
2789 | 53.3k | case Builtin::BIprintf: |
2790 | 53.3k | if (getTarget().getTriple().isNVPTX()) |
2791 | 10 | return EmitNVPTXDevicePrintfCallExpr(E, ReturnValue); |
2792 | 53.3k | break; |
2793 | 6 | case Builtin::BI__builtin_canonicalize: |
2794 | 6 | case Builtin::BI__builtin_canonicalizef: |
2795 | 6 | case Builtin::BI__builtin_canonicalizel: |
2796 | 6 | return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::canonicalize)); |
2797 | 6 | |
2798 | 2 | case Builtin::BI__builtin_thread_pointer: { |
2799 | 2 | if (!getContext().getTargetInfo().isTLSSupported()) |
2800 | 0 | CGM.ErrorUnsupported(E, "__builtin_thread_pointer"); |
2801 | 2 | // Fall through - it's already mapped to the intrinsic by GCCBuiltin. |
2802 | 2 | break; |
2803 | 6 | } |
2804 | 13 | case Builtin::BI__builtin_os_log_format: { |
2805 | 13 | assert(E->getNumArgs() >= 2 && |
2806 | 13 | "__builtin_os_log_format takes at least 2 arguments"); |
2807 | 13 | analyze_os_log::OSLogBufferLayout Layout; |
2808 | 13 | analyze_os_log::computeOSLogBufferLayout(CGM.getContext(), E, Layout); |
2809 | 13 | Address BufAddr = EmitPointerWithAlignment(E->getArg(0)); |
2810 | 13 | // Ignore argument 1, the format string. It is not currently used. |
2811 | 13 | CharUnits Offset; |
2812 | 13 | Builder.CreateStore( |
2813 | 13 | Builder.getInt8(Layout.getSummaryByte()), |
2814 | 13 | Builder.CreateConstByteGEP(BufAddr, Offset++, "summary")); |
2815 | 13 | Builder.CreateStore( |
2816 | 13 | Builder.getInt8(Layout.getNumArgsByte()), |
2817 | 13 | Builder.CreateConstByteGEP(BufAddr, Offset++, "numArgs")); |
2818 | 13 | |
2819 | 13 | llvm::SmallVector<llvm::Value *, 4> RetainableOperands; |
2820 | 25 | for (const auto &Item : Layout.Items) { |
2821 | 25 | Builder.CreateStore( |
2822 | 25 | Builder.getInt8(Item.getDescriptorByte()), |
2823 | 25 | Builder.CreateConstByteGEP(BufAddr, Offset++, "argDescriptor")); |
2824 | 25 | Builder.CreateStore( |
2825 | 25 | Builder.getInt8(Item.getSizeByte()), |
2826 | 25 | Builder.CreateConstByteGEP(BufAddr, Offset++, "argSize")); |
2827 | 25 | Address Addr = Builder.CreateConstByteGEP(BufAddr, Offset); |
2828 | 25 | if (const Expr *TheExpr25 = Item.getExpr()) { |
2829 | 21 | Addr = Builder.CreateElementBitCast( |
2830 | 21 | Addr, ConvertTypeForMem(TheExpr->getType())); |
2831 | 21 | // Check if this is a retainable type. |
2832 | 21 | if (TheExpr->getType()->isObjCRetainableType()21 ) { |
2833 | 1 | assert(getEvaluationKind(TheExpr->getType()) == TEK_Scalar && |
2834 | 1 | "Only scalar can be a ObjC retainable type"); |
2835 | 1 | llvm::Value *SV = EmitScalarExpr(TheExpr, /*Ignore*/ false); |
2836 | 1 | RValue RV = RValue::get(SV); |
2837 | 1 | LValue LV = MakeAddrLValue(Addr, TheExpr->getType()); |
2838 | 1 | EmitStoreThroughLValue(RV, LV); |
2839 | 1 | // Check if the object is constant, if not, save it in |
2840 | 1 | // RetainableOperands. |
2841 | 1 | if (!isa<Constant>(SV)) |
2842 | 1 | RetainableOperands.push_back(SV); |
2843 | 21 | } else { |
2844 | 20 | EmitAnyExprToMem(TheExpr, Addr, Qualifiers(), /*isInit*/ true); |
2845 | 20 | } |
2846 | 25 | } else { |
2847 | 4 | Addr = Builder.CreateElementBitCast(Addr, Int32Ty); |
2848 | 4 | Builder.CreateStore( |
2849 | 4 | Builder.getInt32(Item.getConstValue().getQuantity()), Addr); |
2850 | 4 | } |
2851 | 25 | Offset += Item.size(); |
2852 | 25 | } |
2853 | 13 | |
2854 | 13 | // Push a clang.arc.use cleanup for each object in RetainableOperands. The |
2855 | 13 | // cleanup will cause the use to appear after the final log call, keeping |
2856 | 13 | // the object valid while it's held in the log buffer. Note that if there's |
2857 | 13 | // a release cleanup on the object, it will already be active; since |
2858 | 13 | // cleanups are emitted in reverse order, the use will occur before the |
2859 | 13 | // object is released. |
2860 | 13 | if (!RetainableOperands.empty() && 13 getLangOpts().ObjCAutoRefCount1 && |
2861 | 1 | CGM.getCodeGenOpts().OptimizationLevel != 0) |
2862 | 1 | for (llvm::Value *object : RetainableOperands) |
2863 | 1 | pushFullExprCleanup<CallObjCArcUse>(getARCCleanupKind(), object); |
2864 | 13 | |
2865 | 13 | return RValue::get(BufAddr.getPointer()); |
2866 | 6 | } |
2867 | 6 | |
2868 | 12 | case Builtin::BI__builtin_os_log_format_buffer_size: { |
2869 | 12 | analyze_os_log::OSLogBufferLayout Layout; |
2870 | 12 | analyze_os_log::computeOSLogBufferLayout(CGM.getContext(), E, Layout); |
2871 | 12 | return RValue::get(ConstantInt::get(ConvertType(E->getType()), |
2872 | 12 | Layout.size().getQuantity())); |
2873 | 6 | } |
2874 | 6 | |
2875 | 4 | case Builtin::BI__xray_customevent: { |
2876 | 4 | if (!ShouldXRayInstrumentFunction()) |
2877 | 0 | return RValue::getIgnored(); |
2878 | 4 | if (const auto *4 XRayAttr4 = CurFuncDecl->getAttr<XRayInstrumentAttr>()) { |
2879 | 4 | if (XRayAttr->neverXRayInstrument()) |
2880 | 1 | return RValue::getIgnored(); |
2881 | 3 | } |
2882 | 3 | Function *F = CGM.getIntrinsic(Intrinsic::xray_customevent); |
2883 | 3 | auto FTy = F->getFunctionType(); |
2884 | 3 | auto Arg0 = E->getArg(0); |
2885 | 3 | auto Arg0Val = EmitScalarExpr(Arg0); |
2886 | 3 | auto Arg0Ty = Arg0->getType(); |
2887 | 3 | auto PTy0 = FTy->getParamType(0); |
2888 | 3 | if (PTy0 != Arg0Val->getType()3 ) { |
2889 | 0 | if (Arg0Ty->isArrayType()) |
2890 | 0 | Arg0Val = EmitArrayToPointerDecay(Arg0).getPointer(); |
2891 | 0 | else |
2892 | 0 | Arg0Val = Builder.CreatePointerCast(Arg0Val, PTy0); |
2893 | 0 | } |
2894 | 3 | auto Arg1 = EmitScalarExpr(E->getArg(1)); |
2895 | 3 | auto PTy1 = FTy->getParamType(1); |
2896 | 3 | if (PTy1 != Arg1->getType()) |
2897 | 3 | Arg1 = Builder.CreateTruncOrBitCast(Arg1, PTy1); |
2898 | 3 | return RValue::get(Builder.CreateCall(F, {Arg0Val, Arg1})); |
2899 | 3 | } |
2900 | 3 | |
2901 | 8 | case Builtin::BI__builtin_ms_va_start: |
2902 | 8 | case Builtin::BI__builtin_ms_va_end: |
2903 | 8 | return RValue::get( |
2904 | 8 | EmitVAStartEnd(EmitMSVAListRef(E->getArg(0)).getPointer(), |
2905 | 8 | BuiltinID == Builtin::BI__builtin_ms_va_start)); |
2906 | 8 | |
2907 | 6 | case Builtin::BI__builtin_ms_va_copy: { |
2908 | 6 | // Lower this manually. We can't reliably determine whether or not any |
2909 | 6 | // given va_copy() is for a Win64 va_list from the calling convention |
2910 | 6 | // alone, because it's legal to do this from a System V ABI function. |
2911 | 6 | // With opaque pointer types, we won't have enough information in LLVM |
2912 | 6 | // IR to determine this from the argument types, either. Best to do it |
2913 | 6 | // now, while we have enough information. |
2914 | 6 | Address DestAddr = EmitMSVAListRef(E->getArg(0)); |
2915 | 6 | Address SrcAddr = EmitMSVAListRef(E->getArg(1)); |
2916 | 6 | |
2917 | 6 | llvm::Type *BPP = Int8PtrPtrTy; |
2918 | 6 | |
2919 | 6 | DestAddr = Address(Builder.CreateBitCast(DestAddr.getPointer(), BPP, "cp"), |
2920 | 6 | DestAddr.getAlignment()); |
2921 | 6 | SrcAddr = Address(Builder.CreateBitCast(SrcAddr.getPointer(), BPP, "ap"), |
2922 | 6 | SrcAddr.getAlignment()); |
2923 | 6 | |
2924 | 6 | Value *ArgPtr = Builder.CreateLoad(SrcAddr, "ap.val"); |
2925 | 6 | return RValue::get(Builder.CreateStore(ArgPtr, DestAddr)); |
2926 | 132k | } |
2927 | 132k | } |
2928 | 132k | |
2929 | 132k | // If this is an alias for a lib function (e.g. __builtin_sin), emit |
2930 | 132k | // the call using the normal call path, but using the unmangled |
2931 | 132k | // version of the function name. |
2932 | 132k | if (132k getContext().BuiltinInfo.isLibFunction(BuiltinID)132k ) |
2933 | 5.97k | return emitLibraryCall(*this, FD, E, |
2934 | 5.97k | CGM.getBuiltinLibFunction(FD, BuiltinID)); |
2935 | 126k | |
2936 | 126k | // If this is a predefined lib function (e.g. malloc), emit the call |
2937 | 126k | // using exactly the normal call path. |
2938 | 126k | if (126k getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID)126k ) |
2939 | 92.5k | return emitLibraryCall(*this, FD, E, |
2940 | 92.5k | cast<llvm::Constant>(EmitScalarExpr(E->getCallee()))); |
2941 | 34.0k | |
2942 | 34.0k | // Check that a call to a target specific builtin has the correct target |
2943 | 34.0k | // features. |
2944 | 34.0k | // This is down here to avoid non-target specific builtins, however, if |
2945 | 34.0k | // generic builtins start to require generic target features then we |
2946 | 34.0k | // can move this up to the beginning of the function. |
2947 | 34.0k | checkTargetFeatures(E, FD); |
2948 | 34.0k | |
2949 | 34.0k | // See if we have a target specific intrinsic. |
2950 | 34.0k | const char *Name = getContext().BuiltinInfo.getName(BuiltinID); |
2951 | 34.0k | Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic; |
2952 | 34.0k | StringRef Prefix = |
2953 | 34.0k | llvm::Triple::getArchTypePrefix(getTarget().getTriple().getArch()); |
2954 | 34.0k | if (!Prefix.empty()34.0k ) { |
2955 | 34.0k | IntrinsicID = Intrinsic::getIntrinsicForGCCBuiltin(Prefix.data(), Name); |
2956 | 34.0k | // NOTE we dont need to perform a compatibility flag check here since the |
2957 | 34.0k | // intrinsics are declared in Builtins*.def via LANGBUILTIN which filter the |
2958 | 34.0k | // MS builtins via ALL_MS_LANGUAGES and are filtered earlier. |
2959 | 34.0k | if (IntrinsicID == Intrinsic::not_intrinsic) |
2960 | 23.7k | IntrinsicID = Intrinsic::getIntrinsicForMSBuiltin(Prefix.data(), Name); |
2961 | 34.0k | } |
2962 | 34.0k | |
2963 | 34.0k | if (IntrinsicID != Intrinsic::not_intrinsic34.0k ) { |
2964 | 10.3k | SmallVector<Value*, 16> Args; |
2965 | 10.3k | |
2966 | 10.3k | // Find out if any arguments are required to be integer constant |
2967 | 10.3k | // expressions. |
2968 | 10.3k | unsigned ICEArguments = 0; |
2969 | 10.3k | ASTContext::GetBuiltinTypeError Error; |
2970 | 10.3k | getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); |
2971 | 10.3k | assert(Error == ASTContext::GE_None && "Should not codegen an error"); |
2972 | 10.3k | |
2973 | 10.3k | Function *F = CGM.getIntrinsic(IntrinsicID); |
2974 | 10.3k | llvm::FunctionType *FTy = F->getFunctionType(); |
2975 | 10.3k | |
2976 | 36.9k | for (unsigned i = 0, e = E->getNumArgs(); i != e36.9k ; ++i26.5k ) { |
2977 | 26.5k | Value *ArgValue; |
2978 | 26.5k | // If this is a normal argument, just emit it as a scalar. |
2979 | 26.5k | if ((ICEArguments & (1 << i)) == 026.5k ) { |
2980 | 24.4k | ArgValue = EmitScalarExpr(E->getArg(i)); |
2981 | 26.5k | } else { |
2982 | 2.12k | // If this is required to be a constant, constant fold it so that we |
2983 | 2.12k | // know that the generated intrinsic gets a ConstantInt. |
2984 | 2.12k | llvm::APSInt Result; |
2985 | 2.12k | bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result,getContext()); |
2986 | 2.12k | assert(IsConst && "Constant arg isn't actually constant?"); |
2987 | 2.12k | (void)IsConst; |
2988 | 2.12k | ArgValue = llvm::ConstantInt::get(getLLVMContext(), Result); |
2989 | 2.12k | } |
2990 | 26.5k | |
2991 | 26.5k | // If the intrinsic arg type is different from the builtin arg type |
2992 | 26.5k | // we need to do a bit cast. |
2993 | 26.5k | llvm::Type *PTy = FTy->getParamType(i); |
2994 | 26.5k | if (PTy != ArgValue->getType()26.5k ) { |
2995 | 980 | assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) && |
2996 | 980 | "Must be able to losslessly bit cast to param"); |
2997 | 980 | ArgValue = Builder.CreateBitCast(ArgValue, PTy); |
2998 | 980 | } |
2999 | 26.5k | |
3000 | 26.5k | Args.push_back(ArgValue); |
3001 | 26.5k | } |
3002 | 10.3k | |
3003 | 10.3k | Value *V = Builder.CreateCall(F, Args); |
3004 | 10.3k | QualType BuiltinRetType = E->getType(); |
3005 | 10.3k | |
3006 | 10.3k | llvm::Type *RetTy = VoidTy; |
3007 | 10.3k | if (!BuiltinRetType->isVoidType()) |
3008 | 10.0k | RetTy = ConvertType(BuiltinRetType); |
3009 | 10.3k | |
3010 | 10.3k | if (RetTy != V->getType()10.3k ) { |
3011 | 422 | assert(V->getType()->canLosslesslyBitCastTo(RetTy) && |
3012 | 422 | "Must be able to losslessly bit cast result type"); |
3013 | 422 | V = Builder.CreateBitCast(V, RetTy); |
3014 | 422 | } |
3015 | 10.3k | |
3016 | 10.3k | return RValue::get(V); |
3017 | 10.3k | } |
3018 | 23.7k | |
3019 | 23.7k | // See if we have a target specific builtin that needs to be lowered. |
3020 | 23.7k | if (Value *23.7k V23.7k = EmitTargetBuiltinExpr(BuiltinID, E)) |
3021 | 23.7k | return RValue::get(V); |
3022 | 0 |
|
3023 | 0 | ErrorUnsupported(E, "builtin function"); |
3024 | 0 |
|
3025 | 0 | // Unknown builtin, for now just dump it out and return undef. |
3026 | 0 | return GetUndefRValue(E->getType()); |
3027 | 0 | } |
3028 | | |
3029 | | static Value *EmitTargetArchBuiltinExpr(CodeGenFunction *CGF, |
3030 | | unsigned BuiltinID, const CallExpr *E, |
3031 | 23.7k | llvm::Triple::ArchType Arch) { |
3032 | 23.7k | switch (Arch) { |
3033 | 1.34k | case llvm::Triple::arm: |
3034 | 1.34k | case llvm::Triple::armeb: |
3035 | 1.34k | case llvm::Triple::thumb: |
3036 | 1.34k | case llvm::Triple::thumbeb: |
3037 | 1.34k | return CGF->EmitARMBuiltinExpr(BuiltinID, E); |
3038 | 17.3k | case llvm::Triple::aarch64: |
3039 | 17.3k | case llvm::Triple::aarch64_be: |
3040 | 17.3k | return CGF->EmitAArch64BuiltinExpr(BuiltinID, E); |
3041 | 2.70k | case llvm::Triple::x86: |
3042 | 2.70k | case llvm::Triple::x86_64: |
3043 | 2.70k | return CGF->EmitX86BuiltinExpr(BuiltinID, E); |
3044 | 1.01k | case llvm::Triple::ppc: |
3045 | 1.01k | case llvm::Triple::ppc64: |
3046 | 1.01k | case llvm::Triple::ppc64le: |
3047 | 1.01k | return CGF->EmitPPCBuiltinExpr(BuiltinID, E); |
3048 | 105 | case llvm::Triple::r600: |
3049 | 105 | case llvm::Triple::amdgcn: |
3050 | 105 | return CGF->EmitAMDGPUBuiltinExpr(BuiltinID, E); |
3051 | 815 | case llvm::Triple::systemz: |
3052 | 815 | return CGF->EmitSystemZBuiltinExpr(BuiltinID, E); |
3053 | 370 | case llvm::Triple::nvptx: |
3054 | 370 | case llvm::Triple::nvptx64: |
3055 | 370 | return CGF->EmitNVPTXBuiltinExpr(BuiltinID, E); |
3056 | 8 | case llvm::Triple::wasm32: |
3057 | 8 | case llvm::Triple::wasm64: |
3058 | 8 | return CGF->EmitWebAssemblyBuiltinExpr(BuiltinID, E); |
3059 | 0 | default: |
3060 | 0 | return nullptr; |
3061 | 0 | } |
3062 | 0 | } |
3063 | | |
3064 | | Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID, |
3065 | 23.7k | const CallExpr *E) { |
3066 | 23.7k | if (getContext().BuiltinInfo.isAuxBuiltinID(BuiltinID)23.7k ) { |
3067 | 0 | assert(getContext().getAuxTargetInfo() && "Missing aux target info"); |
3068 | 0 | return EmitTargetArchBuiltinExpr( |
3069 | 0 | this, getContext().BuiltinInfo.getAuxBuiltinID(BuiltinID), E, |
3070 | 0 | getContext().getAuxTargetInfo()->getTriple().getArch()); |
3071 | 0 | } |
3072 | 23.7k | |
3073 | 23.7k | return EmitTargetArchBuiltinExpr(this, BuiltinID, E, |
3074 | 23.7k | getTarget().getTriple().getArch()); |
3075 | 23.7k | } |
3076 | | |
3077 | | static llvm::VectorType *GetNeonType(CodeGenFunction *CGF, |
3078 | | NeonTypeFlags TypeFlags, |
3079 | 9.25k | bool V1Ty=false) { |
3080 | 9.25k | int IsQuad = TypeFlags.isQuad(); |
3081 | 9.25k | switch (TypeFlags.getEltType()) { |
3082 | 2.46k | case NeonTypeFlags::Int8: |
3083 | 2.46k | case NeonTypeFlags::Poly8: |
3084 | 2.46k | return llvm::VectorType::get(CGF->Int8Ty, V1Ty ? 10 : (8 << IsQuad)2.46k ); |
3085 | 2.15k | case NeonTypeFlags::Int16: |
3086 | 2.15k | case NeonTypeFlags::Poly16: |
3087 | 2.15k | case NeonTypeFlags::Float16: |
3088 | 2.15k | return llvm::VectorType::get(CGF->Int16Ty, V1Ty ? 10 : (4 << IsQuad)2.15k ); |
3089 | 2.17k | case NeonTypeFlags::Int32: |
3090 | 2.17k | return llvm::VectorType::get(CGF->Int32Ty, V1Ty ? 10 : (2 << IsQuad)2.17k ); |
3091 | 1.45k | case NeonTypeFlags::Int64: |
3092 | 1.45k | case NeonTypeFlags::Poly64: |
3093 | 1.45k | return llvm::VectorType::get(CGF->Int64Ty, V1Ty ? 10 : (1 << IsQuad)1.45k ); |
3094 | 0 | case NeonTypeFlags::Poly128: |
3095 | 0 | // FIXME: i128 and f128 doesn't get fully support in Clang and llvm. |
3096 | 0 | // There is a lot of i128 and f128 API missing. |
3097 | 0 | // so we use v16i8 to represent poly128 and get pattern matched. |
3098 | 0 | return llvm::VectorType::get(CGF->Int8Ty, 16); |
3099 | 605 | case NeonTypeFlags::Float32: |
3100 | 605 | return llvm::VectorType::get(CGF->FloatTy, V1Ty ? 10 : (2 << IsQuad)605 ); |
3101 | 403 | case NeonTypeFlags::Float64: |
3102 | 2.46k | return llvm::VectorType::get(CGF->DoubleTy, V1Ty ? 10 : (1 << IsQuad)403 ); |
3103 | 0 | } |
3104 | 0 | llvm_unreachable0 ("Unknown vector element type!"); |
3105 | 0 | } |
3106 | | |
3107 | | static llvm::VectorType *GetFloatNeonType(CodeGenFunction *CGF, |
3108 | 196 | NeonTypeFlags IntTypeFlags) { |
3109 | 196 | int IsQuad = IntTypeFlags.isQuad(); |
3110 | 196 | switch (IntTypeFlags.getEltType()) { |
3111 | 112 | case NeonTypeFlags::Int32: |
3112 | 112 | return llvm::VectorType::get(CGF->FloatTy, (2 << IsQuad)); |
3113 | 84 | case NeonTypeFlags::Int64: |
3114 | 84 | return llvm::VectorType::get(CGF->DoubleTy, (1 << IsQuad)); |
3115 | 0 | default: |
3116 | 0 | llvm_unreachable("Type can't be converted to floating-point!"); |
3117 | 0 | } |
3118 | 0 | } |
3119 | | |
3120 | 218 | Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) { |
3121 | 218 | unsigned nElts = V->getType()->getVectorNumElements(); |
3122 | 218 | Value* SV = llvm::ConstantVector::getSplat(nElts, C); |
3123 | 218 | return Builder.CreateShuffleVector(V, V, SV, "lane"); |
3124 | 218 | } |
3125 | | |
3126 | | Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops, |
3127 | | const char *name, |
3128 | 4.44k | unsigned shift, bool rightshift) { |
3129 | 4.44k | unsigned j = 0; |
3130 | 4.44k | for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); |
3131 | 14.8k | ai != ae14.8k ; ++ai, ++j10.3k ) |
3132 | 10.3k | if (10.3k shift > 0 && 10.3k shift == j466 ) |
3133 | 233 | Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift); |
3134 | 10.3k | else |
3135 | 10.1k | Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name); |
3136 | 4.44k | |
3137 | 4.44k | return Builder.CreateCall(F, Ops, name); |
3138 | 4.44k | } |
3139 | | |
3140 | | Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty, |
3141 | 603 | bool neg) { |
3142 | 603 | int SV = cast<ConstantInt>(V)->getSExtValue(); |
3143 | 603 | return ConstantInt::get(Ty, neg ? -SV173 : SV430 ); |
3144 | 603 | } |
3145 | | |
3146 | | // \brief Right-shift a vector by a constant. |
3147 | | Value *CodeGenFunction::EmitNeonRShiftImm(Value *Vec, Value *Shift, |
3148 | | llvm::Type *Ty, bool usgn, |
3149 | 140 | const char *name) { |
3150 | 140 | llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); |
3151 | 140 | |
3152 | 140 | int ShiftAmt = cast<ConstantInt>(Shift)->getSExtValue(); |
3153 | 140 | int EltSize = VTy->getScalarSizeInBits(); |
3154 | 140 | |
3155 | 140 | Vec = Builder.CreateBitCast(Vec, Ty); |
3156 | 140 | |
3157 | 140 | // lshr/ashr are undefined when the shift amount is equal to the vector |
3158 | 140 | // element size. |
3159 | 140 | if (ShiftAmt == EltSize140 ) { |
3160 | 10 | if (usgn10 ) { |
3161 | 4 | // Right-shifting an unsigned value by its size yields 0. |
3162 | 4 | return llvm::ConstantAggregateZero::get(VTy); |
3163 | 0 | } else { |
3164 | 6 | // Right-shifting a signed value by its size is equivalent |
3165 | 6 | // to a shift of size-1. |
3166 | 6 | --ShiftAmt; |
3167 | 6 | Shift = ConstantInt::get(VTy->getElementType(), ShiftAmt); |
3168 | 6 | } |
3169 | 10 | } |
3170 | 140 | |
3171 | 136 | Shift = EmitNeonShiftVector(Shift, Ty, false); |
3172 | 136 | if (usgn) |
3173 | 68 | return Builder.CreateLShr(Vec, Shift, name); |
3174 | 136 | else |
3175 | 68 | return Builder.CreateAShr(Vec, Shift, name); |
3176 | 0 | } |
3177 | | |
3178 | | enum { |
3179 | | AddRetType = (1 << 0), |
3180 | | Add1ArgType = (1 << 1), |
3181 | | Add2ArgTypes = (1 << 2), |
3182 | | |
3183 | | VectorizeRetType = (1 << 3), |
3184 | | VectorizeArgTypes = (1 << 4), |
3185 | | |
3186 | | InventFloatType = (1 << 5), |
3187 | | UnsignedAlts = (1 << 6), |
3188 | | |
3189 | | Use64BitVectors = (1 << 7), |
3190 | | Use128BitVectors = (1 << 8), |
3191 | | |
3192 | | Vectorize1ArgType = Add1ArgType | VectorizeArgTypes, |
3193 | | VectorRet = AddRetType | VectorizeRetType, |
3194 | | VectorRetGetArgs01 = |
3195 | | AddRetType | Add2ArgTypes | VectorizeRetType | VectorizeArgTypes, |
3196 | | FpCmpzModifiers = |
3197 | | AddRetType | VectorizeRetType | Add1ArgType | InventFloatType |
3198 | | }; |
3199 | | |
3200 | | namespace { |
3201 | | struct NeonIntrinsicInfo { |
3202 | | const char *NameHint; |
3203 | | unsigned BuiltinID; |
3204 | | unsigned LLVMIntrinsic; |
3205 | | unsigned AltLLVMIntrinsic; |
3206 | | unsigned TypeModifier; |
3207 | | |
3208 | 181k | bool operator<(unsigned RHSBuiltinID) const { |
3209 | 181k | return BuiltinID < RHSBuiltinID; |
3210 | 181k | } |
3211 | 0 | bool operator<(const NeonIntrinsicInfo &TE) const { |
3212 | 0 | return BuiltinID < TE.BuiltinID; |
3213 | 0 | } |
3214 | | }; |
3215 | | } // end anonymous namespace |
3216 | | |
3217 | | #define NEONMAP0(NameBase) \ |
3218 | | { #NameBase, NEON::BI__builtin_neon_ ## NameBase, 0, 0, 0 } |
3219 | | |
3220 | | #define NEONMAP1(NameBase, LLVMIntrinsic, TypeModifier) \ |
3221 | | { #NameBase, NEON:: BI__builtin_neon_ ## NameBase, \ |
3222 | | Intrinsic::LLVMIntrinsic, 0, TypeModifier } |
3223 | | |
3224 | | #define NEONMAP2(NameBase, LLVMIntrinsic, AltLLVMIntrinsic, TypeModifier) \ |
3225 | | { #NameBase, NEON:: BI__builtin_neon_ ## NameBase, \ |
3226 | | Intrinsic::LLVMIntrinsic, Intrinsic::AltLLVMIntrinsic, \ |
3227 | | TypeModifier } |
3228 | | |
3229 | | static const NeonIntrinsicInfo ARMSIMDIntrinsicMap [] = { |
3230 | | NEONMAP2(vabd_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), |
3231 | | NEONMAP2(vabdq_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), |
3232 | | NEONMAP1(vabs_v, arm_neon_vabs, 0), |
3233 | | NEONMAP1(vabsq_v, arm_neon_vabs, 0), |
3234 | | NEONMAP0(vaddhn_v), |
3235 | | NEONMAP1(vaesdq_v, arm_neon_aesd, 0), |
3236 | | NEONMAP1(vaeseq_v, arm_neon_aese, 0), |
3237 | | NEONMAP1(vaesimcq_v, arm_neon_aesimc, 0), |
3238 | | NEONMAP1(vaesmcq_v, arm_neon_aesmc, 0), |
3239 | | NEONMAP1(vbsl_v, arm_neon_vbsl, AddRetType), |
3240 | | NEONMAP1(vbslq_v, arm_neon_vbsl, AddRetType), |
3241 | | NEONMAP1(vcage_v, arm_neon_vacge, 0), |
3242 | | NEONMAP1(vcageq_v, arm_neon_vacge, 0), |
3243 | | NEONMAP1(vcagt_v, arm_neon_vacgt, 0), |
3244 | | NEONMAP1(vcagtq_v, arm_neon_vacgt, 0), |
3245 | | NEONMAP1(vcale_v, arm_neon_vacge, 0), |
3246 | | NEONMAP1(vcaleq_v, arm_neon_vacge, 0), |
3247 | | NEONMAP1(vcalt_v, arm_neon_vacgt, 0), |
3248 | | NEONMAP1(vcaltq_v, arm_neon_vacgt, 0), |
3249 | | NEONMAP1(vcls_v, arm_neon_vcls, Add1ArgType), |
3250 | | NEONMAP1(vclsq_v, arm_neon_vcls, Add1ArgType), |
3251 | | NEONMAP1(vclz_v, ctlz, Add1ArgType), |
3252 | | NEONMAP1(vclzq_v, ctlz, Add1ArgType), |
3253 | | NEONMAP1(vcnt_v, ctpop, Add1ArgType), |
3254 | | NEONMAP1(vcntq_v, ctpop, Add1ArgType), |
3255 | | NEONMAP1(vcvt_f16_f32, arm_neon_vcvtfp2hf, 0), |
3256 | | NEONMAP1(vcvt_f32_f16, arm_neon_vcvthf2fp, 0), |
3257 | | NEONMAP0(vcvt_f32_v), |
3258 | | NEONMAP2(vcvt_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), |
3259 | | NEONMAP1(vcvt_n_s32_v, arm_neon_vcvtfp2fxs, 0), |
3260 | | NEONMAP1(vcvt_n_s64_v, arm_neon_vcvtfp2fxs, 0), |
3261 | | NEONMAP1(vcvt_n_u32_v, arm_neon_vcvtfp2fxu, 0), |
3262 | | NEONMAP1(vcvt_n_u64_v, arm_neon_vcvtfp2fxu, 0), |
3263 | | NEONMAP0(vcvt_s32_v), |
3264 | | NEONMAP0(vcvt_s64_v), |
3265 | | NEONMAP0(vcvt_u32_v), |
3266 | | NEONMAP0(vcvt_u64_v), |
3267 | | NEONMAP1(vcvta_s32_v, arm_neon_vcvtas, 0), |
3268 | | NEONMAP1(vcvta_s64_v, arm_neon_vcvtas, 0), |
3269 | | NEONMAP1(vcvta_u32_v, arm_neon_vcvtau, 0), |
3270 | | NEONMAP1(vcvta_u64_v, arm_neon_vcvtau, 0), |
3271 | | NEONMAP1(vcvtaq_s32_v, arm_neon_vcvtas, 0), |
3272 | | NEONMAP1(vcvtaq_s64_v, arm_neon_vcvtas, 0), |
3273 | | NEONMAP1(vcvtaq_u32_v, arm_neon_vcvtau, 0), |
3274 | | NEONMAP1(vcvtaq_u64_v, arm_neon_vcvtau, 0), |
3275 | | NEONMAP1(vcvtm_s32_v, arm_neon_vcvtms, 0), |
3276 | | NEONMAP1(vcvtm_s64_v, arm_neon_vcvtms, 0), |
3277 | | NEONMAP1(vcvtm_u32_v, arm_neon_vcvtmu, 0), |
3278 | | NEONMAP1(vcvtm_u64_v, arm_neon_vcvtmu, 0), |
3279 | | NEONMAP1(vcvtmq_s32_v, arm_neon_vcvtms, 0), |
3280 | | NEONMAP1(vcvtmq_s64_v, arm_neon_vcvtms, 0), |
3281 | | NEONMAP1(vcvtmq_u32_v, arm_neon_vcvtmu, 0), |
3282 | | NEONMAP1(vcvtmq_u64_v, arm_neon_vcvtmu, 0), |
3283 | | NEONMAP1(vcvtn_s32_v, arm_neon_vcvtns, 0), |
3284 | | NEONMAP1(vcvtn_s64_v, arm_neon_vcvtns, 0), |
3285 | | NEONMAP1(vcvtn_u32_v, arm_neon_vcvtnu, 0), |
3286 | | NEONMAP1(vcvtn_u64_v, arm_neon_vcvtnu, 0), |
3287 | | NEONMAP1(vcvtnq_s32_v, arm_neon_vcvtns, 0), |
3288 | | NEONMAP1(vcvtnq_s64_v, arm_neon_vcvtns, 0), |
3289 | | NEONMAP1(vcvtnq_u32_v, arm_neon_vcvtnu, 0), |
3290 | | NEONMAP1(vcvtnq_u64_v, arm_neon_vcvtnu, 0), |
3291 | | NEONMAP1(vcvtp_s32_v, arm_neon_vcvtps, 0), |
3292 | | NEONMAP1(vcvtp_s64_v, arm_neon_vcvtps, 0), |
3293 | | NEONMAP1(vcvtp_u32_v, arm_neon_vcvtpu, 0), |
3294 | | NEONMAP1(vcvtp_u64_v, arm_neon_vcvtpu, 0), |
3295 | | NEONMAP1(vcvtpq_s32_v, arm_neon_vcvtps, 0), |
3296 | | NEONMAP1(vcvtpq_s64_v, arm_neon_vcvtps, 0), |
3297 | | NEONMAP1(vcvtpq_u32_v, arm_neon_vcvtpu, 0), |
3298 | | NEONMAP1(vcvtpq_u64_v, arm_neon_vcvtpu, 0), |
3299 | | NEONMAP0(vcvtq_f32_v), |
3300 | | NEONMAP2(vcvtq_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), |
3301 | | NEONMAP1(vcvtq_n_s32_v, arm_neon_vcvtfp2fxs, 0), |
3302 | | NEONMAP1(vcvtq_n_s64_v, arm_neon_vcvtfp2fxs, 0), |
3303 | | NEONMAP1(vcvtq_n_u32_v, arm_neon_vcvtfp2fxu, 0), |
3304 | | NEONMAP1(vcvtq_n_u64_v, arm_neon_vcvtfp2fxu, 0), |
3305 | | NEONMAP0(vcvtq_s32_v), |
3306 | | NEONMAP0(vcvtq_s64_v), |
3307 | | NEONMAP0(vcvtq_u32_v), |
3308 | | NEONMAP0(vcvtq_u64_v), |
3309 | | NEONMAP0(vext_v), |
3310 | | NEONMAP0(vextq_v), |
3311 | | NEONMAP0(vfma_v), |
3312 | | NEONMAP0(vfmaq_v), |
3313 | | NEONMAP2(vhadd_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), |
3314 | | NEONMAP2(vhaddq_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), |
3315 | | NEONMAP2(vhsub_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), |
3316 | | NEONMAP2(vhsubq_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), |
3317 | | NEONMAP0(vld1_dup_v), |
3318 | | NEONMAP1(vld1_v, arm_neon_vld1, 0), |
3319 | | NEONMAP0(vld1q_dup_v), |
3320 | | NEONMAP1(vld1q_v, arm_neon_vld1, 0), |
3321 | | NEONMAP1(vld2_lane_v, arm_neon_vld2lane, 0), |
3322 | | NEONMAP1(vld2_v, arm_neon_vld2, 0), |
3323 | | NEONMAP1(vld2q_lane_v, arm_neon_vld2lane, 0), |
3324 | | NEONMAP1(vld2q_v, arm_neon_vld2, 0), |
3325 | | NEONMAP1(vld3_lane_v, arm_neon_vld3lane, 0), |
3326 | | NEONMAP1(vld3_v, arm_neon_vld3, 0), |
3327 | | NEONMAP1(vld3q_lane_v, arm_neon_vld3lane, 0), |
3328 | | NEONMAP1(vld3q_v, arm_neon_vld3, 0), |
3329 | | NEONMAP1(vld4_lane_v, arm_neon_vld4lane, 0), |
3330 | | NEONMAP1(vld4_v, arm_neon_vld4, 0), |
3331 | | NEONMAP1(vld4q_lane_v, arm_neon_vld4lane, 0), |
3332 | | NEONMAP1(vld4q_v, arm_neon_vld4, 0), |
3333 | | NEONMAP2(vmax_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), |
3334 | | NEONMAP1(vmaxnm_v, arm_neon_vmaxnm, Add1ArgType), |
3335 | | NEONMAP1(vmaxnmq_v, arm_neon_vmaxnm, Add1ArgType), |
3336 | | NEONMAP2(vmaxq_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), |
3337 | | NEONMAP2(vmin_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), |
3338 | | NEONMAP1(vminnm_v, arm_neon_vminnm, Add1ArgType), |
3339 | | NEONMAP1(vminnmq_v, arm_neon_vminnm, Add1ArgType), |
3340 | | NEONMAP2(vminq_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), |
3341 | | NEONMAP0(vmovl_v), |
3342 | | NEONMAP0(vmovn_v), |
3343 | | NEONMAP1(vmul_v, arm_neon_vmulp, Add1ArgType), |
3344 | | NEONMAP0(vmull_v), |
3345 | | NEONMAP1(vmulq_v, arm_neon_vmulp, Add1ArgType), |
3346 | | NEONMAP2(vpadal_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), |
3347 | | NEONMAP2(vpadalq_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), |
3348 | | NEONMAP1(vpadd_v, arm_neon_vpadd, Add1ArgType), |
3349 | | NEONMAP2(vpaddl_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), |
3350 | | NEONMAP2(vpaddlq_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), |
3351 | | NEONMAP1(vpaddq_v, arm_neon_vpadd, Add1ArgType), |
3352 | | NEONMAP2(vpmax_v, arm_neon_vpmaxu, arm_neon_vpmaxs, Add1ArgType | UnsignedAlts), |
3353 | | NEONMAP2(vpmin_v, arm_neon_vpminu, arm_neon_vpmins, Add1ArgType | UnsignedAlts), |
3354 | | NEONMAP1(vqabs_v, arm_neon_vqabs, Add1ArgType), |
3355 | | NEONMAP1(vqabsq_v, arm_neon_vqabs, Add1ArgType), |
3356 | | NEONMAP2(vqadd_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), |
3357 | | NEONMAP2(vqaddq_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), |
3358 | | NEONMAP2(vqdmlal_v, arm_neon_vqdmull, arm_neon_vqadds, 0), |
3359 | | NEONMAP2(vqdmlsl_v, arm_neon_vqdmull, arm_neon_vqsubs, 0), |
3360 | | NEONMAP1(vqdmulh_v, arm_neon_vqdmulh, Add1ArgType), |
3361 | | NEONMAP1(vqdmulhq_v, arm_neon_vqdmulh, Add1ArgType), |
3362 | | NEONMAP1(vqdmull_v, arm_neon_vqdmull, Add1ArgType), |
3363 | | NEONMAP2(vqmovn_v, arm_neon_vqmovnu, arm_neon_vqmovns, Add1ArgType | UnsignedAlts), |
3364 | | NEONMAP1(vqmovun_v, arm_neon_vqmovnsu, Add1ArgType), |
3365 | | NEONMAP1(vqneg_v, arm_neon_vqneg, Add1ArgType), |
3366 | | NEONMAP1(vqnegq_v, arm_neon_vqneg, Add1ArgType), |
3367 | | NEONMAP1(vqrdmulh_v, arm_neon_vqrdmulh, Add1ArgType), |
3368 | | NEONMAP1(vqrdmulhq_v, arm_neon_vqrdmulh, Add1ArgType), |
3369 | | NEONMAP2(vqrshl_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), |
3370 | | NEONMAP2(vqrshlq_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), |
3371 | | NEONMAP2(vqshl_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), |
3372 | | NEONMAP2(vqshl_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), |
3373 | | NEONMAP2(vqshlq_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), |
3374 | | NEONMAP2(vqshlq_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), |
3375 | | NEONMAP1(vqshlu_n_v, arm_neon_vqshiftsu, 0), |
3376 | | NEONMAP1(vqshluq_n_v, arm_neon_vqshiftsu, 0), |
3377 | | NEONMAP2(vqsub_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), |
3378 | | NEONMAP2(vqsubq_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), |
3379 | | NEONMAP1(vraddhn_v, arm_neon_vraddhn, Add1ArgType), |
3380 | | NEONMAP2(vrecpe_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), |
3381 | | NEONMAP2(vrecpeq_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), |
3382 | | NEONMAP1(vrecps_v, arm_neon_vrecps, Add1ArgType), |
3383 | | NEONMAP1(vrecpsq_v, arm_neon_vrecps, Add1ArgType), |
3384 | | NEONMAP2(vrhadd_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), |
3385 | | NEONMAP2(vrhaddq_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), |
3386 | | NEONMAP1(vrnd_v, arm_neon_vrintz, Add1ArgType), |
3387 | | NEONMAP1(vrnda_v, arm_neon_vrinta, Add1ArgType), |
3388 | | NEONMAP1(vrndaq_v, arm_neon_vrinta, Add1ArgType), |
3389 | | NEONMAP1(vrndm_v, arm_neon_vrintm, Add1ArgType), |
3390 | | NEONMAP1(vrndmq_v, arm_neon_vrintm, Add1ArgType), |
3391 | | NEONMAP1(vrndn_v, arm_neon_vrintn, Add1ArgType), |
3392 | | NEONMAP1(vrndnq_v, arm_neon_vrintn, Add1ArgType), |
3393 | | NEONMAP1(vrndp_v, arm_neon_vrintp, Add1ArgType), |
3394 | | NEONMAP1(vrndpq_v, arm_neon_vrintp, Add1ArgType), |
3395 | | NEONMAP1(vrndq_v, arm_neon_vrintz, Add1ArgType), |
3396 | | NEONMAP1(vrndx_v, arm_neon_vrintx, Add1ArgType), |
3397 | | NEONMAP1(vrndxq_v, arm_neon_vrintx, Add1ArgType), |
3398 | | NEONMAP2(vrshl_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), |
3399 | | NEONMAP2(vrshlq_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), |
3400 | | NEONMAP2(vrshr_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), |
3401 | | NEONMAP2(vrshrq_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), |
3402 | | NEONMAP2(vrsqrte_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), |
3403 | | NEONMAP2(vrsqrteq_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), |
3404 | | NEONMAP1(vrsqrts_v, arm_neon_vrsqrts, Add1ArgType), |
3405 | | NEONMAP1(vrsqrtsq_v, arm_neon_vrsqrts, Add1ArgType), |
3406 | | NEONMAP1(vrsubhn_v, arm_neon_vrsubhn, Add1ArgType), |
3407 | | NEONMAP1(vsha1su0q_v, arm_neon_sha1su0, 0), |
3408 | | NEONMAP1(vsha1su1q_v, arm_neon_sha1su1, 0), |
3409 | | NEONMAP1(vsha256h2q_v, arm_neon_sha256h2, 0), |
3410 | | NEONMAP1(vsha256hq_v, arm_neon_sha256h, 0), |
3411 | | NEONMAP1(vsha256su0q_v, arm_neon_sha256su0, 0), |
3412 | | NEONMAP1(vsha256su1q_v, arm_neon_sha256su1, 0), |
3413 | | NEONMAP0(vshl_n_v), |
3414 | | NEONMAP2(vshl_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), |
3415 | | NEONMAP0(vshll_n_v), |
3416 | | NEONMAP0(vshlq_n_v), |
3417 | | NEONMAP2(vshlq_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), |
3418 | | NEONMAP0(vshr_n_v), |
3419 | | NEONMAP0(vshrn_n_v), |
3420 | | NEONMAP0(vshrq_n_v), |
3421 | | NEONMAP1(vst1_v, arm_neon_vst1, 0), |
3422 | | NEONMAP1(vst1q_v, arm_neon_vst1, 0), |
3423 | | NEONMAP1(vst2_lane_v, arm_neon_vst2lane, 0), |
3424 | | NEONMAP1(vst2_v, arm_neon_vst2, 0), |
3425 | | NEONMAP1(vst2q_lane_v, arm_neon_vst2lane, 0), |
3426 | | NEONMAP1(vst2q_v, arm_neon_vst2, 0), |
3427 | | NEONMAP1(vst3_lane_v, arm_neon_vst3lane, 0), |
3428 | | NEONMAP1(vst3_v, arm_neon_vst3, 0), |
3429 | | NEONMAP1(vst3q_lane_v, arm_neon_vst3lane, 0), |
3430 | | NEONMAP1(vst3q_v, arm_neon_vst3, 0), |
3431 | | NEONMAP1(vst4_lane_v, arm_neon_vst4lane, 0), |
3432 | | NEONMAP1(vst4_v, arm_neon_vst4, 0), |
3433 | | NEONMAP1(vst4q_lane_v, arm_neon_vst4lane, 0), |
3434 | | NEONMAP1(vst4q_v, arm_neon_vst4, 0), |
3435 | | NEONMAP0(vsubhn_v), |
3436 | | NEONMAP0(vtrn_v), |
3437 | | NEONMAP0(vtrnq_v), |
3438 | | NEONMAP0(vtst_v), |
3439 | | NEONMAP0(vtstq_v), |
3440 | | NEONMAP0(vuzp_v), |
3441 | | NEONMAP0(vuzpq_v), |
3442 | | NEONMAP0(vzip_v), |
3443 | | NEONMAP0(vzipq_v) |
3444 | | }; |
3445 | | |
3446 | | static const NeonIntrinsicInfo AArch64SIMDIntrinsicMap[] = { |
3447 | | NEONMAP1(vabs_v, aarch64_neon_abs, 0), |
3448 | | NEONMAP1(vabsq_v, aarch64_neon_abs, 0), |
3449 | | NEONMAP0(vaddhn_v), |
3450 | | NEONMAP1(vaesdq_v, aarch64_crypto_aesd, 0), |
3451 | | NEONMAP1(vaeseq_v, aarch64_crypto_aese, 0), |
3452 | | NEONMAP1(vaesimcq_v, aarch64_crypto_aesimc, 0), |
3453 | | NEONMAP1(vaesmcq_v, aarch64_crypto_aesmc, 0), |
3454 | | NEONMAP1(vcage_v, aarch64_neon_facge, 0), |
3455 | | NEONMAP1(vcageq_v, aarch64_neon_facge, 0), |
3456 | | NEONMAP1(vcagt_v, aarch64_neon_facgt, 0), |
3457 | | NEONMAP1(vcagtq_v, aarch64_neon_facgt, 0), |
3458 | | NEONMAP1(vcale_v, aarch64_neon_facge, 0), |
3459 | | NEONMAP1(vcaleq_v, aarch64_neon_facge, 0), |
3460 | | NEONMAP1(vcalt_v, aarch64_neon_facgt, 0), |
3461 | | NEONMAP1(vcaltq_v, aarch64_neon_facgt, 0), |
3462 | | NEONMAP1(vcls_v, aarch64_neon_cls, Add1ArgType), |
3463 | | NEONMAP1(vclsq_v, aarch64_neon_cls, Add1ArgType), |
3464 | | NEONMAP1(vclz_v, ctlz, Add1ArgType), |
3465 | | NEONMAP1(vclzq_v, ctlz, Add1ArgType), |
3466 | | NEONMAP1(vcnt_v, ctpop, Add1ArgType), |
3467 | | NEONMAP1(vcntq_v, ctpop, Add1ArgType), |
3468 | | NEONMAP1(vcvt_f16_f32, aarch64_neon_vcvtfp2hf, 0), |
3469 | | NEONMAP1(vcvt_f32_f16, aarch64_neon_vcvthf2fp, 0), |
3470 | | NEONMAP0(vcvt_f32_v), |
3471 | | NEONMAP2(vcvt_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), |
3472 | | NEONMAP2(vcvt_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), |
3473 | | NEONMAP1(vcvt_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), |
3474 | | NEONMAP1(vcvt_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), |
3475 | | NEONMAP1(vcvt_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), |
3476 | | NEONMAP1(vcvt_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), |
3477 | | NEONMAP0(vcvtq_f32_v), |
3478 | | NEONMAP2(vcvtq_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), |
3479 | | NEONMAP2(vcvtq_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), |
3480 | | NEONMAP1(vcvtq_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), |
3481 | | NEONMAP1(vcvtq_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), |
3482 | | NEONMAP1(vcvtq_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), |
3483 | | NEONMAP1(vcvtq_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), |
3484 | | NEONMAP1(vcvtx_f32_v, aarch64_neon_fcvtxn, AddRetType | Add1ArgType), |
3485 | | NEONMAP0(vext_v), |
3486 | | NEONMAP0(vextq_v), |
3487 | | NEONMAP0(vfma_v), |
3488 | | NEONMAP0(vfmaq_v), |
3489 | | NEONMAP2(vhadd_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), |
3490 | | NEONMAP2(vhaddq_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), |
3491 | | NEONMAP2(vhsub_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), |
3492 | | NEONMAP2(vhsubq_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), |
3493 | | NEONMAP0(vmovl_v), |
3494 | | NEONMAP0(vmovn_v), |
3495 | | NEONMAP1(vmul_v, aarch64_neon_pmul, Add1ArgType), |
3496 | | NEONMAP1(vmulq_v, aarch64_neon_pmul, Add1ArgType), |
3497 | | NEONMAP1(vpadd_v, aarch64_neon_addp, Add1ArgType), |
3498 | | NEONMAP2(vpaddl_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), |
3499 | | NEONMAP2(vpaddlq_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), |
3500 | | NEONMAP1(vpaddq_v, aarch64_neon_addp, Add1ArgType), |
3501 | | NEONMAP1(vqabs_v, aarch64_neon_sqabs, Add1ArgType), |
3502 | | NEONMAP1(vqabsq_v, aarch64_neon_sqabs, Add1ArgType), |
3503 | | NEONMAP2(vqadd_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), |
3504 | | NEONMAP2(vqaddq_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), |
3505 | | NEONMAP2(vqdmlal_v, aarch64_neon_sqdmull, aarch64_neon_sqadd, 0), |
3506 | | NEONMAP2(vqdmlsl_v, aarch64_neon_sqdmull, aarch64_neon_sqsub, 0), |
3507 | | NEONMAP1(vqdmulh_v, aarch64_neon_sqdmulh, Add1ArgType), |
3508 | | NEONMAP1(vqdmulhq_v, aarch64_neon_sqdmulh, Add1ArgType), |
3509 | | NEONMAP1(vqdmull_v, aarch64_neon_sqdmull, Add1ArgType), |
3510 | | NEONMAP2(vqmovn_v, aarch64_neon_uqxtn, aarch64_neon_sqxtn, Add1ArgType | UnsignedAlts), |
3511 | | NEONMAP1(vqmovun_v, aarch64_neon_sqxtun, Add1ArgType), |
3512 | | NEONMAP1(vqneg_v, aarch64_neon_sqneg, Add1ArgType), |
3513 | | NEONMAP1(vqnegq_v, aarch64_neon_sqneg, Add1ArgType), |
3514 | | NEONMAP1(vqrdmulh_v, aarch64_neon_sqrdmulh, Add1ArgType), |
3515 | | NEONMAP1(vqrdmulhq_v, aarch64_neon_sqrdmulh, Add1ArgType), |
3516 | | NEONMAP2(vqrshl_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), |
3517 | | NEONMAP2(vqrshlq_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), |
3518 | | NEONMAP2(vqshl_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl, UnsignedAlts), |
3519 | | NEONMAP2(vqshl_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), |
3520 | | NEONMAP2(vqshlq_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl,UnsignedAlts), |
3521 | | NEONMAP2(vqshlq_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), |
3522 | | NEONMAP1(vqshlu_n_v, aarch64_neon_sqshlu, 0), |
3523 | | NEONMAP1(vqshluq_n_v, aarch64_neon_sqshlu, 0), |
3524 | | NEONMAP2(vqsub_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), |
3525 | | NEONMAP2(vqsubq_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), |
3526 | | NEONMAP1(vraddhn_v, aarch64_neon_raddhn, Add1ArgType), |
3527 | | NEONMAP2(vrecpe_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), |
3528 | | NEONMAP2(vrecpeq_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), |
3529 | | NEONMAP1(vrecps_v, aarch64_neon_frecps, Add1ArgType), |
3530 | | NEONMAP1(vrecpsq_v, aarch64_neon_frecps, Add1ArgType), |
3531 | | NEONMAP2(vrhadd_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), |
3532 | | NEONMAP2(vrhaddq_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), |
3533 | | NEONMAP2(vrshl_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), |
3534 | | NEONMAP2(vrshlq_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), |
3535 | | NEONMAP2(vrshr_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), |
3536 | | NEONMAP2(vrshrq_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), |
3537 | | NEONMAP2(vrsqrte_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), |
3538 | | NEONMAP2(vrsqrteq_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), |
3539 | | NEONMAP1(vrsqrts_v, aarch64_neon_frsqrts, Add1ArgType), |
3540 | | NEONMAP1(vrsqrtsq_v, aarch64_neon_frsqrts, Add1ArgType), |
3541 | | NEONMAP1(vrsubhn_v, aarch64_neon_rsubhn, Add1ArgType), |
3542 | | NEONMAP1(vsha1su0q_v, aarch64_crypto_sha1su0, 0), |
3543 | | NEONMAP1(vsha1su1q_v, aarch64_crypto_sha1su1, 0), |
3544 | | NEONMAP1(vsha256h2q_v, aarch64_crypto_sha256h2, 0), |
3545 | | NEONMAP1(vsha256hq_v, aarch64_crypto_sha256h, 0), |
3546 | | NEONMAP1(vsha256su0q_v, aarch64_crypto_sha256su0, 0), |
3547 | | NEONMAP1(vsha256su1q_v, aarch64_crypto_sha256su1, 0), |
3548 | | NEONMAP0(vshl_n_v), |
3549 | | NEONMAP2(vshl_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), |
3550 | | NEONMAP0(vshll_n_v), |
3551 | | NEONMAP0(vshlq_n_v), |
3552 | | NEONMAP2(vshlq_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), |
3553 | | NEONMAP0(vshr_n_v), |
3554 | | NEONMAP0(vshrn_n_v), |
3555 | | NEONMAP0(vshrq_n_v), |
3556 | | NEONMAP0(vsubhn_v), |
3557 | | NEONMAP0(vtst_v), |
3558 | | NEONMAP0(vtstq_v), |
3559 | | }; |
3560 | | |
3561 | | static const NeonIntrinsicInfo AArch64SISDIntrinsicMap[] = { |
3562 | | NEONMAP1(vabdd_f64, aarch64_sisd_fabd, Add1ArgType), |
3563 | | NEONMAP1(vabds_f32, aarch64_sisd_fabd, Add1ArgType), |
3564 | | NEONMAP1(vabsd_s64, aarch64_neon_abs, Add1ArgType), |
3565 | | NEONMAP1(vaddlv_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), |
3566 | | NEONMAP1(vaddlv_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), |
3567 | | NEONMAP1(vaddlvq_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), |
3568 | | NEONMAP1(vaddlvq_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), |
3569 | | NEONMAP1(vaddv_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), |
3570 | | NEONMAP1(vaddv_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), |
3571 | | NEONMAP1(vaddv_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), |
3572 | | NEONMAP1(vaddvq_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), |
3573 | | NEONMAP1(vaddvq_f64, aarch64_neon_faddv, AddRetType | Add1ArgType), |
3574 | | NEONMAP1(vaddvq_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), |
3575 | | NEONMAP1(vaddvq_s64, aarch64_neon_saddv, AddRetType | Add1ArgType), |
3576 | | NEONMAP1(vaddvq_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), |
3577 | | NEONMAP1(vaddvq_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), |
3578 | | NEONMAP1(vcaged_f64, aarch64_neon_facge, AddRetType | Add1ArgType), |
3579 | | NEONMAP1(vcages_f32, aarch64_neon_facge, AddRetType | Add1ArgType), |
3580 | | NEONMAP1(vcagtd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), |
3581 | | NEONMAP1(vcagts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), |
3582 | | NEONMAP1(vcaled_f64, aarch64_neon_facge, AddRetType | Add1ArgType), |
3583 | | NEONMAP1(vcales_f32, aarch64_neon_facge, AddRetType | Add1ArgType), |
3584 | | NEONMAP1(vcaltd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), |
3585 | | NEONMAP1(vcalts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), |
3586 | | NEONMAP1(vcvtad_s64_f64, aarch64_neon_fcvtas, AddRetType | Add1ArgType), |
3587 | | NEONMAP1(vcvtad_u64_f64, aarch64_neon_fcvtau, AddRetType | Add1ArgType), |
3588 | | NEONMAP1(vcvtas_s32_f32, aarch64_neon_fcvtas, AddRetType | Add1ArgType), |
3589 | | NEONMAP1(vcvtas_u32_f32, aarch64_neon_fcvtau, AddRetType | Add1ArgType), |
3590 | | NEONMAP1(vcvtd_n_f64_s64, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), |
3591 | | NEONMAP1(vcvtd_n_f64_u64, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), |
3592 | | NEONMAP1(vcvtd_n_s64_f64, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), |
3593 | | NEONMAP1(vcvtd_n_u64_f64, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), |
3594 | | NEONMAP1(vcvtmd_s64_f64, aarch64_neon_fcvtms, AddRetType | Add1ArgType), |
3595 | | NEONMAP1(vcvtmd_u64_f64, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), |
3596 | | NEONMAP1(vcvtms_s32_f32, aarch64_neon_fcvtms, AddRetType | Add1ArgType), |
3597 | | NEONMAP1(vcvtms_u32_f32, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), |
3598 | | NEONMAP1(vcvtnd_s64_f64, aarch64_neon_fcvtns, AddRetType | Add1ArgType), |
3599 | | NEONMAP1(vcvtnd_u64_f64, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), |
3600 | | NEONMAP1(vcvtns_s32_f32, aarch64_neon_fcvtns, AddRetType | Add1ArgType), |
3601 | | NEONMAP1(vcvtns_u32_f32, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), |
3602 | | NEONMAP1(vcvtpd_s64_f64, aarch64_neon_fcvtps, AddRetType | Add1ArgType), |
3603 | | NEONMAP1(vcvtpd_u64_f64, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), |
3604 | | NEONMAP1(vcvtps_s32_f32, aarch64_neon_fcvtps, AddRetType | Add1ArgType), |
3605 | | NEONMAP1(vcvtps_u32_f32, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), |
3606 | | NEONMAP1(vcvts_n_f32_s32, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), |
3607 | | NEONMAP1(vcvts_n_f32_u32, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), |
3608 | | NEONMAP1(vcvts_n_s32_f32, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), |
3609 | | NEONMAP1(vcvts_n_u32_f32, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), |
3610 | | NEONMAP1(vcvtxd_f32_f64, aarch64_sisd_fcvtxn, 0), |
3611 | | NEONMAP1(vmaxnmv_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), |
3612 | | NEONMAP1(vmaxnmvq_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), |
3613 | | NEONMAP1(vmaxnmvq_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), |
3614 | | NEONMAP1(vmaxv_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), |
3615 | | NEONMAP1(vmaxv_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), |
3616 | | NEONMAP1(vmaxv_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), |
3617 | | NEONMAP1(vmaxvq_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), |
3618 | | NEONMAP1(vmaxvq_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), |
3619 | | NEONMAP1(vmaxvq_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), |
3620 | | NEONMAP1(vmaxvq_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), |
3621 | | NEONMAP1(vminnmv_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), |
3622 | | NEONMAP1(vminnmvq_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), |
3623 | | NEONMAP1(vminnmvq_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), |
3624 | | NEONMAP1(vminv_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), |
3625 | | NEONMAP1(vminv_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), |
3626 | | NEONMAP1(vminv_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), |
3627 | | NEONMAP1(vminvq_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), |
3628 | | NEONMAP1(vminvq_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), |
3629 | | NEONMAP1(vminvq_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), |
3630 | | NEONMAP1(vminvq_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), |
3631 | | NEONMAP1(vmull_p64, aarch64_neon_pmull64, 0), |
3632 | | NEONMAP1(vmulxd_f64, aarch64_neon_fmulx, Add1ArgType), |
3633 | | NEONMAP1(vmulxs_f32, aarch64_neon_fmulx, Add1ArgType), |
3634 | | NEONMAP1(vpaddd_s64, aarch64_neon_uaddv, AddRetType | Add1ArgType), |
3635 | | NEONMAP1(vpaddd_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), |
3636 | | NEONMAP1(vpmaxnmqd_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), |
3637 | | NEONMAP1(vpmaxnms_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), |
3638 | | NEONMAP1(vpmaxqd_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), |
3639 | | NEONMAP1(vpmaxs_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), |
3640 | | NEONMAP1(vpminnmqd_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), |
3641 | | NEONMAP1(vpminnms_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), |
3642 | | NEONMAP1(vpminqd_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), |
3643 | | NEONMAP1(vpmins_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), |
3644 | | NEONMAP1(vqabsb_s8, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), |
3645 | | NEONMAP1(vqabsd_s64, aarch64_neon_sqabs, Add1ArgType), |
3646 | | NEONMAP1(vqabsh_s16, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), |
3647 | | NEONMAP1(vqabss_s32, aarch64_neon_sqabs, Add1ArgType), |
3648 | | NEONMAP1(vqaddb_s8, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), |
3649 | | NEONMAP1(vqaddb_u8, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), |
3650 | | NEONMAP1(vqaddd_s64, aarch64_neon_sqadd, Add1ArgType), |
3651 | | NEONMAP1(vqaddd_u64, aarch64_neon_uqadd, Add1ArgType), |
3652 | | NEONMAP1(vqaddh_s16, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), |
3653 | | NEONMAP1(vqaddh_u16, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), |
3654 | | NEONMAP1(vqadds_s32, aarch64_neon_sqadd, Add1ArgType), |
3655 | | NEONMAP1(vqadds_u32, aarch64_neon_uqadd, Add1ArgType), |
3656 | | NEONMAP1(vqdmulhh_s16, aarch64_neon_sqdmulh, Vectorize1ArgType | Use64BitVectors), |
3657 | | NEONMAP1(vqdmulhs_s32, aarch64_neon_sqdmulh, Add1ArgType), |
3658 | | NEONMAP1(vqdmullh_s16, aarch64_neon_sqdmull, VectorRet | Use128BitVectors), |
3659 | | NEONMAP1(vqdmulls_s32, aarch64_neon_sqdmulls_scalar, 0), |
3660 | | NEONMAP1(vqmovnd_s64, aarch64_neon_scalar_sqxtn, AddRetType | Add1ArgType), |
3661 | | NEONMAP1(vqmovnd_u64, aarch64_neon_scalar_uqxtn, AddRetType | Add1ArgType), |
3662 | | NEONMAP1(vqmovnh_s16, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), |
3663 | | NEONMAP1(vqmovnh_u16, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), |
3664 | | NEONMAP1(vqmovns_s32, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), |
3665 | | NEONMAP1(vqmovns_u32, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), |
3666 | | NEONMAP1(vqmovund_s64, aarch64_neon_scalar_sqxtun, AddRetType | Add1ArgType), |
3667 | | NEONMAP1(vqmovunh_s16, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), |
3668 | | NEONMAP1(vqmovuns_s32, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), |
3669 | | NEONMAP1(vqnegb_s8, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), |
3670 | | NEONMAP1(vqnegd_s64, aarch64_neon_sqneg, Add1ArgType), |
3671 | | NEONMAP1(vqnegh_s16, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), |
3672 | | NEONMAP1(vqnegs_s32, aarch64_neon_sqneg, Add1ArgType), |
3673 | | NEONMAP1(vqrdmulhh_s16, aarch64_neon_sqrdmulh, Vectorize1ArgType | Use64BitVectors), |
3674 | | NEONMAP1(vqrdmulhs_s32, aarch64_neon_sqrdmulh, Add1ArgType), |
3675 | | NEONMAP1(vqrshlb_s8, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), |
3676 | | NEONMAP1(vqrshlb_u8, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), |
3677 | | NEONMAP1(vqrshld_s64, aarch64_neon_sqrshl, Add1ArgType), |
3678 | | NEONMAP1(vqrshld_u64, aarch64_neon_uqrshl, Add1ArgType), |
3679 | | NEONMAP1(vqrshlh_s16, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), |
3680 | | NEONMAP1(vqrshlh_u16, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), |
3681 | | NEONMAP1(vqrshls_s32, aarch64_neon_sqrshl, Add1ArgType), |
3682 | | NEONMAP1(vqrshls_u32, aarch64_neon_uqrshl, Add1ArgType), |
3683 | | NEONMAP1(vqrshrnd_n_s64, aarch64_neon_sqrshrn, AddRetType), |
3684 | | NEONMAP1(vqrshrnd_n_u64, aarch64_neon_uqrshrn, AddRetType), |
3685 | | NEONMAP1(vqrshrnh_n_s16, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), |
3686 | | NEONMAP1(vqrshrnh_n_u16, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), |
3687 | | NEONMAP1(vqrshrns_n_s32, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), |
3688 | | NEONMAP1(vqrshrns_n_u32, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), |
3689 | | NEONMAP1(vqrshrund_n_s64, aarch64_neon_sqrshrun, AddRetType), |
3690 | | NEONMAP1(vqrshrunh_n_s16, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), |
3691 | | NEONMAP1(vqrshruns_n_s32, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), |
3692 | | NEONMAP1(vqshlb_n_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), |
3693 | | NEONMAP1(vqshlb_n_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), |
3694 | | NEONMAP1(vqshlb_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), |
3695 | | NEONMAP1(vqshlb_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), |
3696 | | NEONMAP1(vqshld_s64, aarch64_neon_sqshl, Add1ArgType), |
3697 | | NEONMAP1(vqshld_u64, aarch64_neon_uqshl, Add1ArgType), |
3698 | | NEONMAP1(vqshlh_n_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), |
3699 | | NEONMAP1(vqshlh_n_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), |
3700 | | NEONMAP1(vqshlh_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), |
3701 | | NEONMAP1(vqshlh_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), |
3702 | | NEONMAP1(vqshls_n_s32, aarch64_neon_sqshl, Add1ArgType), |
3703 | | NEONMAP1(vqshls_n_u32, aarch64_neon_uqshl, Add1ArgType), |
3704 | | NEONMAP1(vqshls_s32, aarch64_neon_sqshl, Add1ArgType), |
3705 | | NEONMAP1(vqshls_u32, aarch64_neon_uqshl, Add1ArgType), |
3706 | | NEONMAP1(vqshlub_n_s8, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), |
3707 | | NEONMAP1(vqshluh_n_s16, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), |
3708 | | NEONMAP1(vqshlus_n_s32, aarch64_neon_sqshlu, Add1ArgType), |
3709 | | NEONMAP1(vqshrnd_n_s64, aarch64_neon_sqshrn, AddRetType), |
3710 | | NEONMAP1(vqshrnd_n_u64, aarch64_neon_uqshrn, AddRetType), |
3711 | | NEONMAP1(vqshrnh_n_s16, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), |
3712 | | NEONMAP1(vqshrnh_n_u16, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), |
3713 | | NEONMAP1(vqshrns_n_s32, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), |
3714 | | NEONMAP1(vqshrns_n_u32, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), |
3715 | | NEONMAP1(vqshrund_n_s64, aarch64_neon_sqshrun, AddRetType), |
3716 | | NEONMAP1(vqshrunh_n_s16, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), |
3717 | | NEONMAP1(vqshruns_n_s32, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), |
3718 | | NEONMAP1(vqsubb_s8, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), |
3719 | | NEONMAP1(vqsubb_u8, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), |
3720 | | NEONMAP1(vqsubd_s64, aarch64_neon_sqsub, Add1ArgType), |
3721 | | NEONMAP1(vqsubd_u64, aarch64_neon_uqsub, Add1ArgType), |
3722 | | NEONMAP1(vqsubh_s16, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), |
3723 | | NEONMAP1(vqsubh_u16, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), |
3724 | | NEONMAP1(vqsubs_s32, aarch64_neon_sqsub, Add1ArgType), |
3725 | | NEONMAP1(vqsubs_u32, aarch64_neon_uqsub, Add1ArgType), |
3726 | | NEONMAP1(vrecped_f64, aarch64_neon_frecpe, Add1ArgType), |
3727 | | NEONMAP1(vrecpes_f32, aarch64_neon_frecpe, Add1ArgType), |
3728 | | NEONMAP1(vrecpxd_f64, aarch64_neon_frecpx, Add1ArgType), |
3729 | | NEONMAP1(vrecpxs_f32, aarch64_neon_frecpx, Add1ArgType), |
3730 | | NEONMAP1(vrshld_s64, aarch64_neon_srshl, Add1ArgType), |
3731 | | NEONMAP1(vrshld_u64, aarch64_neon_urshl, Add1ArgType), |
3732 | | NEONMAP1(vrsqrted_f64, aarch64_neon_frsqrte, Add1ArgType), |
3733 | | NEONMAP1(vrsqrtes_f32, aarch64_neon_frsqrte, Add1ArgType), |
3734 | | NEONMAP1(vrsqrtsd_f64, aarch64_neon_frsqrts, Add1ArgType), |
3735 | | NEONMAP1(vrsqrtss_f32, aarch64_neon_frsqrts, Add1ArgType), |
3736 | | NEONMAP1(vsha1cq_u32, aarch64_crypto_sha1c, 0), |
3737 | | NEONMAP1(vsha1h_u32, aarch64_crypto_sha1h, 0), |
3738 | | NEONMAP1(vsha1mq_u32, aarch64_crypto_sha1m, 0), |
3739 | | NEONMAP1(vsha1pq_u32, aarch64_crypto_sha1p, 0), |
3740 | | NEONMAP1(vshld_s64, aarch64_neon_sshl, Add1ArgType), |
3741 | | NEONMAP1(vshld_u64, aarch64_neon_ushl, Add1ArgType), |
3742 | | NEONMAP1(vslid_n_s64, aarch64_neon_vsli, Vectorize1ArgType), |
3743 | | NEONMAP1(vslid_n_u64, aarch64_neon_vsli, Vectorize1ArgType), |
3744 | | NEONMAP1(vsqaddb_u8, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), |
3745 | | NEONMAP1(vsqaddd_u64, aarch64_neon_usqadd, Add1ArgType), |
3746 | | NEONMAP1(vsqaddh_u16, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), |
3747 | | NEONMAP1(vsqadds_u32, aarch64_neon_usqadd, Add1ArgType), |
3748 | | NEONMAP1(vsrid_n_s64, aarch64_neon_vsri, Vectorize1ArgType), |
3749 | | NEONMAP1(vsrid_n_u64, aarch64_neon_vsri, Vectorize1ArgType), |
3750 | | NEONMAP1(vuqaddb_s8, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), |
3751 | | NEONMAP1(vuqaddd_s64, aarch64_neon_suqadd, Add1ArgType), |
3752 | | NEONMAP1(vuqaddh_s16, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), |
3753 | | NEONMAP1(vuqadds_s32, aarch64_neon_suqadd, Add1ArgType), |
3754 | | }; |
3755 | | |
3756 | | #undef NEONMAP0 |
3757 | | #undef NEONMAP1 |
3758 | | #undef NEONMAP2 |
3759 | | |
3760 | | static bool NEONSIMDIntrinsicsProvenSorted = false; |
3761 | | |
3762 | | static bool AArch64SIMDIntrinsicsProvenSorted = false; |
3763 | | static bool AArch64SISDIntrinsicsProvenSorted = false; |
3764 | | |
3765 | | |
3766 | | static const NeonIntrinsicInfo * |
3767 | | findNeonIntrinsicInMap(ArrayRef<NeonIntrinsicInfo> IntrinsicMap, |
3768 | 23.7k | unsigned BuiltinID, bool &MapProvenSorted) { |
3769 | 23.7k | |
3770 | | #ifndef NDEBUG |
3771 | | if (!MapProvenSorted) { |
3772 | | assert(std::is_sorted(std::begin(IntrinsicMap), std::end(IntrinsicMap))); |
3773 | | MapProvenSorted = true; |
3774 | | } |
3775 | | #endif |
3776 | | |
3777 | 23.7k | const NeonIntrinsicInfo *Builtin = |
3778 | 23.7k | std::lower_bound(IntrinsicMap.begin(), IntrinsicMap.end(), BuiltinID); |
3779 | 23.7k | |
3780 | 23.7k | if (Builtin != IntrinsicMap.end() && 23.7k Builtin->BuiltinID == BuiltinID23.4k ) |
3781 | 3.11k | return Builtin; |
3782 | 20.5k | |
3783 | 20.5k | return nullptr; |
3784 | 20.5k | } |
3785 | | |
3786 | | Function *CodeGenFunction::LookupNeonLLVMIntrinsic(unsigned IntrinsicID, |
3787 | | unsigned Modifier, |
3788 | | llvm::Type *ArgType, |
3789 | 1.71k | const CallExpr *E) { |
3790 | 1.71k | int VectorSize = 0; |
3791 | 1.71k | if (Modifier & Use64BitVectors) |
3792 | 155 | VectorSize = 64; |
3793 | 1.55k | else if (1.55k Modifier & Use128BitVectors1.55k ) |
3794 | 4 | VectorSize = 128; |
3795 | 1.71k | |
3796 | 1.71k | // Return type. |
3797 | 1.71k | SmallVector<llvm::Type *, 3> Tys; |
3798 | 1.71k | if (Modifier & AddRetType1.71k ) { |
3799 | 335 | llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext())); |
3800 | 335 | if (Modifier & VectorizeRetType) |
3801 | 58 | Ty = llvm::VectorType::get( |
3802 | 58 | Ty, VectorSize ? VectorSize / Ty->getPrimitiveSizeInBits()58 : 10 ); |
3803 | 335 | |
3804 | 335 | Tys.push_back(Ty); |
3805 | 335 | } |
3806 | 1.71k | |
3807 | 1.71k | // Arguments. |
3808 | 1.71k | if (Modifier & VectorizeArgTypes1.71k ) { |
3809 | 113 | int Elts = VectorSize ? VectorSize / ArgType->getPrimitiveSizeInBits()101 : 112 ; |
3810 | 113 | ArgType = llvm::VectorType::get(ArgType, Elts); |
3811 | 113 | } |
3812 | 1.71k | |
3813 | 1.71k | if (Modifier & (Add1ArgType | Add2ArgTypes)) |
3814 | 1.54k | Tys.push_back(ArgType); |
3815 | 1.71k | |
3816 | 1.71k | if (Modifier & Add2ArgTypes) |
3817 | 0 | Tys.push_back(ArgType); |
3818 | 1.71k | |
3819 | 1.71k | if (Modifier & InventFloatType) |
3820 | 0 | Tys.push_back(FloatTy); |
3821 | 1.71k | |
3822 | 1.71k | return CGM.getIntrinsic(IntrinsicID, Tys); |
3823 | 1.71k | } |
3824 | | |
3825 | | static Value *EmitCommonNeonSISDBuiltinExpr(CodeGenFunction &CGF, |
3826 | | const NeonIntrinsicInfo &SISDInfo, |
3827 | | SmallVectorImpl<Value *> &Ops, |
3828 | 582 | const CallExpr *E) { |
3829 | 582 | unsigned BuiltinID = SISDInfo.BuiltinID; |
3830 | 582 | unsigned int Int = SISDInfo.LLVMIntrinsic; |
3831 | 582 | unsigned Modifier = SISDInfo.TypeModifier; |
3832 | 582 | const char *s = SISDInfo.NameHint; |
3833 | 582 | |
3834 | 582 | switch (BuiltinID) { |
3835 | 12 | case NEON::BI__builtin_neon_vcled_s64: |
3836 | 12 | case NEON::BI__builtin_neon_vcled_u64: |
3837 | 12 | case NEON::BI__builtin_neon_vcles_f32: |
3838 | 12 | case NEON::BI__builtin_neon_vcled_f64: |
3839 | 12 | case NEON::BI__builtin_neon_vcltd_s64: |
3840 | 12 | case NEON::BI__builtin_neon_vcltd_u64: |
3841 | 12 | case NEON::BI__builtin_neon_vclts_f32: |
3842 | 12 | case NEON::BI__builtin_neon_vcltd_f64: |
3843 | 12 | case NEON::BI__builtin_neon_vcales_f32: |
3844 | 12 | case NEON::BI__builtin_neon_vcaled_f64: |
3845 | 12 | case NEON::BI__builtin_neon_vcalts_f32: |
3846 | 12 | case NEON::BI__builtin_neon_vcaltd_f64: |
3847 | 12 | // Only one direction of comparisons actually exist, cmle is actually a cmge |
3848 | 12 | // with swapped operands. The table gives us the right intrinsic but we |
3849 | 12 | // still need to do the swap. |
3850 | 12 | std::swap(Ops[0], Ops[1]); |
3851 | 12 | break; |
3852 | 582 | } |
3853 | 582 | |
3854 | 582 | assert(Int && "Generic code assumes a valid intrinsic"); |
3855 | 582 | |
3856 | 582 | // Determine the type(s) of this overloaded AArch64 intrinsic. |
3857 | 582 | const Expr *Arg = E->getArg(0); |
3858 | 582 | llvm::Type *ArgTy = CGF.ConvertType(Arg->getType()); |
3859 | 582 | Function *F = CGF.LookupNeonLLVMIntrinsic(Int, Modifier, ArgTy, E); |
3860 | 582 | |
3861 | 582 | int j = 0; |
3862 | 582 | ConstantInt *C0 = ConstantInt::get(CGF.SizeTy, 0); |
3863 | 582 | for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); |
3864 | 1.51k | ai != ae1.51k ; ++ai, ++j936 ) { |
3865 | 936 | llvm::Type *ArgTy = ai->getType(); |
3866 | 936 | if (Ops[j]->getType()->getPrimitiveSizeInBits() == |
3867 | 936 | ArgTy->getPrimitiveSizeInBits()) |
3868 | 684 | continue; |
3869 | 252 | |
3870 | 936 | assert(ArgTy->isVectorTy() && !Ops[j]->getType()->isVectorTy()); |
3871 | 252 | // The constant argument to an _n_ intrinsic always has Int32Ty, so truncate |
3872 | 252 | // it before inserting. |
3873 | 252 | Ops[j] = |
3874 | 252 | CGF.Builder.CreateTruncOrBitCast(Ops[j], ArgTy->getVectorElementType()); |
3875 | 252 | Ops[j] = |
3876 | 252 | CGF.Builder.CreateInsertElement(UndefValue::get(ArgTy), Ops[j], C0); |
3877 | 252 | } |
3878 | 582 | |
3879 | 582 | Value *Result = CGF.EmitNeonCall(F, Ops, s); |
3880 | 582 | llvm::Type *ResultType = CGF.ConvertType(E->getType()); |
3881 | 582 | if (ResultType->getPrimitiveSizeInBits() < |
3882 | 582 | Result->getType()->getPrimitiveSizeInBits()) |
3883 | 159 | return CGF.Builder.CreateExtractElement(Result, C0); |
3884 | 423 | |
3885 | 423 | return CGF.Builder.CreateBitCast(Result, ResultType, s); |
3886 | 423 | } |
3887 | | |
3888 | | Value *CodeGenFunction::EmitCommonNeonBuiltinExpr( |
3889 | | unsigned BuiltinID, unsigned LLVMIntrinsic, unsigned AltLLVMIntrinsic, |
3890 | | const char *NameHint, unsigned Modifier, const CallExpr *E, |
3891 | 2.53k | SmallVectorImpl<llvm::Value *> &Ops, Address PtrOp0, Address PtrOp1) { |
3892 | 2.53k | // Get the last argument, which specifies the vector type. |
3893 | 2.53k | llvm::APSInt NeonTypeConst; |
3894 | 2.53k | const Expr *Arg = E->getArg(E->getNumArgs() - 1); |
3895 | 2.53k | if (!Arg->isIntegerConstantExpr(NeonTypeConst, getContext())) |
3896 | 0 | return nullptr; |
3897 | 2.53k | |
3898 | 2.53k | // Determine the type of this overloaded NEON intrinsic. |
3899 | 2.53k | NeonTypeFlags Type(NeonTypeConst.getZExtValue()); |
3900 | 2.53k | bool Usgn = Type.isUnsigned(); |
3901 | 2.53k | bool Quad = Type.isQuad(); |
3902 | 2.53k | |
3903 | 2.53k | llvm::VectorType *VTy = GetNeonType(this, Type); |
3904 | 2.53k | llvm::Type *Ty = VTy; |
3905 | 2.53k | if (!Ty) |
3906 | 0 | return nullptr; |
3907 | 2.53k | |
3908 | 2.53k | auto getAlignmentValue32 = [&](Address addr) -> Value* 2.53k { |
3909 | 286 | return Builder.getInt32(addr.getAlignment().getQuantity()); |
3910 | 286 | }; |
3911 | 2.53k | |
3912 | 2.53k | unsigned Int = LLVMIntrinsic; |
3913 | 2.53k | if ((Modifier & UnsignedAlts) && 2.53k !Usgn820 ) |
3914 | 426 | Int = AltLLVMIntrinsic; |
3915 | 2.53k | |
3916 | 2.53k | switch (BuiltinID) { |
3917 | 1.08k | default: break; |
3918 | 44 | case NEON::BI__builtin_neon_vabs_v: |
3919 | 44 | case NEON::BI__builtin_neon_vabsq_v: |
3920 | 44 | if (VTy->getElementType()->isFloatingPointTy()) |
3921 | 14 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::fabs, Ty), Ops, "vabs"); |
3922 | 30 | return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vabs"); |
3923 | 24 | case NEON::BI__builtin_neon_vaddhn_v: { |
3924 | 24 | llvm::VectorType *SrcTy = |
3925 | 24 | llvm::VectorType::getExtendedElementVectorType(VTy); |
3926 | 24 | |
3927 | 24 | // %sum = add <4 x i32> %lhs, %rhs |
3928 | 24 | Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); |
3929 | 24 | Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); |
3930 | 24 | Ops[0] = Builder.CreateAdd(Ops[0], Ops[1], "vaddhn"); |
3931 | 24 | |
3932 | 24 | // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> |
3933 | 24 | Constant *ShiftAmt = |
3934 | 24 | ConstantInt::get(SrcTy, SrcTy->getScalarSizeInBits() / 2); |
3935 | 24 | Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vaddhn"); |
3936 | 24 | |
3937 | 24 | // %res = trunc <4 x i32> %high to <4 x i16> |
3938 | 24 | return Builder.CreateTrunc(Ops[0], VTy, "vaddhn"); |
3939 | 30 | } |
3940 | 28 | case NEON::BI__builtin_neon_vcale_v: |
3941 | 28 | case NEON::BI__builtin_neon_vcaleq_v: |
3942 | 28 | case NEON::BI__builtin_neon_vcalt_v: |
3943 | 28 | case NEON::BI__builtin_neon_vcaltq_v: |
3944 | 28 | std::swap(Ops[0], Ops[1]); |
3945 | 28 | LLVM_FALLTHROUGH; |
3946 | 56 | case NEON::BI__builtin_neon_vcage_v: |
3947 | 56 | case NEON::BI__builtin_neon_vcageq_v: |
3948 | 56 | case NEON::BI__builtin_neon_vcagt_v: |
3949 | 56 | case NEON::BI__builtin_neon_vcagtq_v: { |
3950 | 56 | llvm::Type *VecFlt = llvm::VectorType::get( |
3951 | 56 | VTy->getScalarSizeInBits() == 32 ? FloatTy32 : DoubleTy24 , |
3952 | 56 | VTy->getNumElements()); |
3953 | 56 | llvm::Type *Tys[] = { VTy, VecFlt }; |
3954 | 56 | Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); |
3955 | 56 | return EmitNeonCall(F, Ops, NameHint); |
3956 | 56 | } |
3957 | 48 | case NEON::BI__builtin_neon_vclz_v: |
3958 | 48 | case NEON::BI__builtin_neon_vclzq_v: |
3959 | 48 | // We generate target-independent intrinsic, which needs a second argument |
3960 | 48 | // for whether or not clz of zero is undefined; on ARM it isn't. |
3961 | 48 | Ops.push_back(Builder.getInt1(getTarget().isCLZForZeroUndef())); |
3962 | 48 | break; |
3963 | 16 | case NEON::BI__builtin_neon_vcvt_f32_v: |
3964 | 16 | case NEON::BI__builtin_neon_vcvtq_f32_v: |
3965 | 16 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
3966 | 16 | Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, Quad)); |
3967 | 8 | return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") |
3968 | 8 | : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); |
3969 | 28 | case NEON::BI__builtin_neon_vcvt_n_f32_v: |
3970 | 28 | case NEON::BI__builtin_neon_vcvt_n_f64_v: |
3971 | 28 | case NEON::BI__builtin_neon_vcvtq_n_f32_v: |
3972 | 28 | case NEON::BI__builtin_neon_vcvtq_n_f64_v: { |
3973 | 28 | llvm::Type *Tys[2] = { GetFloatNeonType(this, Type), Ty }; |
3974 | 28 | Int = Usgn ? LLVMIntrinsic14 : AltLLVMIntrinsic14 ; |
3975 | 28 | Function *F = CGM.getIntrinsic(Int, Tys); |
3976 | 28 | return EmitNeonCall(F, Ops, "vcvt_n"); |
3977 | 28 | } |
3978 | 28 | case NEON::BI__builtin_neon_vcvt_n_s32_v: |
3979 | 28 | case NEON::BI__builtin_neon_vcvt_n_u32_v: |
3980 | 28 | case NEON::BI__builtin_neon_vcvt_n_s64_v: |
3981 | 28 | case NEON::BI__builtin_neon_vcvt_n_u64_v: |
3982 | 28 | case NEON::BI__builtin_neon_vcvtq_n_s32_v: |
3983 | 28 | case NEON::BI__builtin_neon_vcvtq_n_u32_v: |
3984 | 28 | case NEON::BI__builtin_neon_vcvtq_n_s64_v: |
3985 | 28 | case NEON::BI__builtin_neon_vcvtq_n_u64_v: { |
3986 | 28 | llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; |
3987 | 28 | Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); |
3988 | 28 | return EmitNeonCall(F, Ops, "vcvt_n"); |
3989 | 28 | } |
3990 | 4 | case NEON::BI__builtin_neon_vcvt_s32_v: |
3991 | 4 | case NEON::BI__builtin_neon_vcvt_u32_v: |
3992 | 4 | case NEON::BI__builtin_neon_vcvt_s64_v: |
3993 | 4 | case NEON::BI__builtin_neon_vcvt_u64_v: |
3994 | 4 | case NEON::BI__builtin_neon_vcvtq_s32_v: |
3995 | 4 | case NEON::BI__builtin_neon_vcvtq_u32_v: |
3996 | 4 | case NEON::BI__builtin_neon_vcvtq_s64_v: |
3997 | 4 | case NEON::BI__builtin_neon_vcvtq_u64_v: { |
3998 | 4 | Ops[0] = Builder.CreateBitCast(Ops[0], GetFloatNeonType(this, Type)); |
3999 | 2 | return Usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt") |
4000 | 2 | : Builder.CreateFPToSI(Ops[0], Ty, "vcvt"); |
4001 | 4 | } |
4002 | 16 | case NEON::BI__builtin_neon_vcvta_s32_v: |
4003 | 16 | case NEON::BI__builtin_neon_vcvta_s64_v: |
4004 | 16 | case NEON::BI__builtin_neon_vcvta_u32_v: |
4005 | 16 | case NEON::BI__builtin_neon_vcvta_u64_v: |
4006 | 16 | case NEON::BI__builtin_neon_vcvtaq_s32_v: |
4007 | 16 | case NEON::BI__builtin_neon_vcvtaq_s64_v: |
4008 | 16 | case NEON::BI__builtin_neon_vcvtaq_u32_v: |
4009 | 16 | case NEON::BI__builtin_neon_vcvtaq_u64_v: |
4010 | 16 | case NEON::BI__builtin_neon_vcvtn_s32_v: |
4011 | 16 | case NEON::BI__builtin_neon_vcvtn_s64_v: |
4012 | 16 | case NEON::BI__builtin_neon_vcvtn_u32_v: |
4013 | 16 | case NEON::BI__builtin_neon_vcvtn_u64_v: |
4014 | 16 | case NEON::BI__builtin_neon_vcvtnq_s32_v: |
4015 | 16 | case NEON::BI__builtin_neon_vcvtnq_s64_v: |
4016 | 16 | case NEON::BI__builtin_neon_vcvtnq_u32_v: |
4017 | 16 | case NEON::BI__builtin_neon_vcvtnq_u64_v: |
4018 | 16 | case NEON::BI__builtin_neon_vcvtp_s32_v: |
4019 | 16 | case NEON::BI__builtin_neon_vcvtp_s64_v: |
4020 | 16 | case NEON::BI__builtin_neon_vcvtp_u32_v: |
4021 | 16 | case NEON::BI__builtin_neon_vcvtp_u64_v: |
4022 | 16 | case NEON::BI__builtin_neon_vcvtpq_s32_v: |
4023 | 16 | case NEON::BI__builtin_neon_vcvtpq_s64_v: |
4024 | 16 | case NEON::BI__builtin_neon_vcvtpq_u32_v: |
4025 | 16 | case NEON::BI__builtin_neon_vcvtpq_u64_v: |
4026 | 16 | case NEON::BI__builtin_neon_vcvtm_s32_v: |
4027 | 16 | case NEON::BI__builtin_neon_vcvtm_s64_v: |
4028 | 16 | case NEON::BI__builtin_neon_vcvtm_u32_v: |
4029 | 16 | case NEON::BI__builtin_neon_vcvtm_u64_v: |
4030 | 16 | case NEON::BI__builtin_neon_vcvtmq_s32_v: |
4031 | 16 | case NEON::BI__builtin_neon_vcvtmq_s64_v: |
4032 | 16 | case NEON::BI__builtin_neon_vcvtmq_u32_v: |
4033 | 16 | case NEON::BI__builtin_neon_vcvtmq_u64_v: { |
4034 | 16 | llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; |
4035 | 16 | return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, NameHint); |
4036 | 16 | } |
4037 | 96 | case NEON::BI__builtin_neon_vext_v: |
4038 | 96 | case NEON::BI__builtin_neon_vextq_v: { |
4039 | 96 | int CV = cast<ConstantInt>(Ops[2])->getSExtValue(); |
4040 | 96 | SmallVector<uint32_t, 16> Indices; |
4041 | 636 | for (unsigned i = 0, e = VTy->getNumElements(); i != e636 ; ++i540 ) |
4042 | 540 | Indices.push_back(i+CV); |
4043 | 96 | |
4044 | 96 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
4045 | 96 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
4046 | 96 | return Builder.CreateShuffleVector(Ops[0], Ops[1], Indices, "vext"); |
4047 | 96 | } |
4048 | 19 | case NEON::BI__builtin_neon_vfma_v: |
4049 | 19 | case NEON::BI__builtin_neon_vfmaq_v: { |
4050 | 19 | Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); |
4051 | 19 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
4052 | 19 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
4053 | 19 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
4054 | 19 | |
4055 | 19 | // NEON intrinsic puts accumulator first, unlike the LLVM fma. |
4056 | 19 | return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); |
4057 | 19 | } |
4058 | 27 | case NEON::BI__builtin_neon_vld1_v: |
4059 | 27 | case NEON::BI__builtin_neon_vld1q_v: { |
4060 | 27 | llvm::Type *Tys[] = {Ty, Int8PtrTy}; |
4061 | 27 | Ops.push_back(getAlignmentValue32(PtrOp0)); |
4062 | 27 | return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, "vld1"); |
4063 | 27 | } |
4064 | 66 | case NEON::BI__builtin_neon_vld2_v: |
4065 | 66 | case NEON::BI__builtin_neon_vld2q_v: |
4066 | 66 | case NEON::BI__builtin_neon_vld3_v: |
4067 | 66 | case NEON::BI__builtin_neon_vld3q_v: |
4068 | 66 | case NEON::BI__builtin_neon_vld4_v: |
4069 | 66 | case NEON::BI__builtin_neon_vld4q_v: { |
4070 | 66 | llvm::Type *Tys[] = {Ty, Int8PtrTy}; |
4071 | 66 | Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); |
4072 | 66 | Value *Align = getAlignmentValue32(PtrOp1); |
4073 | 66 | Ops[1] = Builder.CreateCall(F, {Ops[1], Align}, NameHint); |
4074 | 66 | Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); |
4075 | 66 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
4076 | 66 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
4077 | 66 | } |
4078 | 25 | case NEON::BI__builtin_neon_vld1_dup_v: |
4079 | 25 | case NEON::BI__builtin_neon_vld1q_dup_v: { |
4080 | 25 | Value *V = UndefValue::get(Ty); |
4081 | 25 | Ty = llvm::PointerType::getUnqual(VTy->getElementType()); |
4082 | 25 | PtrOp0 = Builder.CreateBitCast(PtrOp0, Ty); |
4083 | 25 | LoadInst *Ld = Builder.CreateLoad(PtrOp0); |
4084 | 25 | llvm::Constant *CI = ConstantInt::get(SizeTy, 0); |
4085 | 25 | Ops[0] = Builder.CreateInsertElement(V, Ld, CI); |
4086 | 25 | return EmitNeonSplat(Ops[0], CI); |
4087 | 25 | } |
4088 | 51 | case NEON::BI__builtin_neon_vld2_lane_v: |
4089 | 51 | case NEON::BI__builtin_neon_vld2q_lane_v: |
4090 | 51 | case NEON::BI__builtin_neon_vld3_lane_v: |
4091 | 51 | case NEON::BI__builtin_neon_vld3q_lane_v: |
4092 | 51 | case NEON::BI__builtin_neon_vld4_lane_v: |
4093 | 51 | case NEON::BI__builtin_neon_vld4q_lane_v: { |
4094 | 51 | llvm::Type *Tys[] = {Ty, Int8PtrTy}; |
4095 | 51 | Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); |
4096 | 204 | for (unsigned I = 2; I < Ops.size() - 1204 ; ++I153 ) |
4097 | 153 | Ops[I] = Builder.CreateBitCast(Ops[I], Ty); |
4098 | 51 | Ops.push_back(getAlignmentValue32(PtrOp1)); |
4099 | 51 | Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), NameHint); |
4100 | 51 | Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); |
4101 | 51 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
4102 | 51 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
4103 | 51 | } |
4104 | 24 | case NEON::BI__builtin_neon_vmovl_v: { |
4105 | 24 | llvm::Type *DTy =llvm::VectorType::getTruncatedElementVectorType(VTy); |
4106 | 24 | Ops[0] = Builder.CreateBitCast(Ops[0], DTy); |
4107 | 24 | if (Usgn) |
4108 | 12 | return Builder.CreateZExt(Ops[0], Ty, "vmovl"); |
4109 | 12 | return Builder.CreateSExt(Ops[0], Ty, "vmovl"); |
4110 | 12 | } |
4111 | 24 | case NEON::BI__builtin_neon_vmovn_v: { |
4112 | 24 | llvm::Type *QTy = llvm::VectorType::getExtendedElementVectorType(VTy); |
4113 | 24 | Ops[0] = Builder.CreateBitCast(Ops[0], QTy); |
4114 | 24 | return Builder.CreateTrunc(Ops[0], Ty, "vmovn"); |
4115 | 12 | } |
4116 | 11 | case NEON::BI__builtin_neon_vmull_v: |
4117 | 11 | // FIXME: the integer vmull operations could be emitted in terms of pure |
4118 | 11 | // LLVM IR (2 exts followed by a mul). Unfortunately LLVM has a habit of |
4119 | 11 | // hoisting the exts outside loops. Until global ISel comes along that can |
4120 | 11 | // see through such movement this leads to bad CodeGen. So we need an |
4121 | 11 | // intrinsic for now. |
4122 | 11 | Int = Usgn ? Intrinsic::arm_neon_vmullu5 : Intrinsic::arm_neon_vmulls6 ; |
4123 | 11 | Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp1 : Int10 ; |
4124 | 11 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); |
4125 | 12 | case NEON::BI__builtin_neon_vpadal_v: |
4126 | 12 | case NEON::BI__builtin_neon_vpadalq_v: { |
4127 | 12 | // The source operand type has twice as many elements of half the size. |
4128 | 12 | unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); |
4129 | 12 | llvm::Type *EltTy = |
4130 | 12 | llvm::IntegerType::get(getLLVMContext(), EltBits / 2); |
4131 | 12 | llvm::Type *NarrowTy = |
4132 | 12 | llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); |
4133 | 12 | llvm::Type *Tys[2] = { Ty, NarrowTy }; |
4134 | 12 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, NameHint); |
4135 | 12 | } |
4136 | 48 | case NEON::BI__builtin_neon_vpaddl_v: |
4137 | 48 | case NEON::BI__builtin_neon_vpaddlq_v: { |
4138 | 48 | // The source operand type has twice as many elements of half the size. |
4139 | 48 | unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); |
4140 | 48 | llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2); |
4141 | 48 | llvm::Type *NarrowTy = |
4142 | 48 | llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); |
4143 | 48 | llvm::Type *Tys[2] = { Ty, NarrowTy }; |
4144 | 48 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl"); |
4145 | 48 | } |
4146 | 36 | case NEON::BI__builtin_neon_vqdmlal_v: |
4147 | 36 | case NEON::BI__builtin_neon_vqdmlsl_v: { |
4148 | 36 | SmallVector<Value *, 2> MulOps(Ops.begin() + 1, Ops.end()); |
4149 | 36 | Ops[1] = |
4150 | 36 | EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), MulOps, "vqdmlal"); |
4151 | 36 | Ops.resize(2); |
4152 | 36 | return EmitNeonCall(CGM.getIntrinsic(AltLLVMIntrinsic, Ty), Ops, NameHint); |
4153 | 36 | } |
4154 | 64 | case NEON::BI__builtin_neon_vqshl_n_v: |
4155 | 64 | case NEON::BI__builtin_neon_vqshlq_n_v: |
4156 | 64 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n", |
4157 | 64 | 1, false); |
4158 | 32 | case NEON::BI__builtin_neon_vqshlu_n_v: |
4159 | 32 | case NEON::BI__builtin_neon_vqshluq_n_v: |
4160 | 32 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshlu_n", |
4161 | 32 | 1, false); |
4162 | 44 | case NEON::BI__builtin_neon_vrecpe_v: |
4163 | 44 | case NEON::BI__builtin_neon_vrecpeq_v: |
4164 | 44 | case NEON::BI__builtin_neon_vrsqrte_v: |
4165 | 44 | case NEON::BI__builtin_neon_vrsqrteq_v: |
4166 | 44 | Int = Ty->isFPOrFPVectorTy() ? LLVMIntrinsic28 : AltLLVMIntrinsic16 ; |
4167 | 44 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, NameHint); |
4168 | 44 | |
4169 | 64 | case NEON::BI__builtin_neon_vrshr_n_v: |
4170 | 64 | case NEON::BI__builtin_neon_vrshrq_n_v: |
4171 | 64 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", |
4172 | 64 | 1, true); |
4173 | 64 | case NEON::BI__builtin_neon_vshl_n_v: |
4174 | 64 | case NEON::BI__builtin_neon_vshlq_n_v: |
4175 | 64 | Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); |
4176 | 64 | return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1], |
4177 | 64 | "vshl_n"); |
4178 | 72 | case NEON::BI__builtin_neon_vshll_n_v: { |
4179 | 72 | llvm::Type *SrcTy = llvm::VectorType::getTruncatedElementVectorType(VTy); |
4180 | 72 | Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); |
4181 | 72 | if (Usgn) |
4182 | 36 | Ops[0] = Builder.CreateZExt(Ops[0], VTy); |
4183 | 72 | else |
4184 | 36 | Ops[0] = Builder.CreateSExt(Ops[0], VTy); |
4185 | 72 | Ops[1] = EmitNeonShiftVector(Ops[1], VTy, false); |
4186 | 72 | return Builder.CreateShl(Ops[0], Ops[1], "vshll_n"); |
4187 | 64 | } |
4188 | 42 | case NEON::BI__builtin_neon_vshrn_n_v: { |
4189 | 42 | llvm::Type *SrcTy = llvm::VectorType::getExtendedElementVectorType(VTy); |
4190 | 42 | Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); |
4191 | 42 | Ops[1] = EmitNeonShiftVector(Ops[1], SrcTy, false); |
4192 | 42 | if (Usgn) |
4193 | 21 | Ops[0] = Builder.CreateLShr(Ops[0], Ops[1]); |
4194 | 42 | else |
4195 | 21 | Ops[0] = Builder.CreateAShr(Ops[0], Ops[1]); |
4196 | 42 | return Builder.CreateTrunc(Ops[0], Ty, "vshrn_n"); |
4197 | 64 | } |
4198 | 70 | case NEON::BI__builtin_neon_vshr_n_v: |
4199 | 70 | case NEON::BI__builtin_neon_vshrq_n_v: |
4200 | 70 | return EmitNeonRShiftImm(Ops[0], Ops[1], Ty, Usgn, "vshr_n"); |
4201 | 142 | case NEON::BI__builtin_neon_vst1_v: |
4202 | 142 | case NEON::BI__builtin_neon_vst1q_v: |
4203 | 142 | case NEON::BI__builtin_neon_vst2_v: |
4204 | 142 | case NEON::BI__builtin_neon_vst2q_v: |
4205 | 142 | case NEON::BI__builtin_neon_vst3_v: |
4206 | 142 | case NEON::BI__builtin_neon_vst3q_v: |
4207 | 142 | case NEON::BI__builtin_neon_vst4_v: |
4208 | 142 | case NEON::BI__builtin_neon_vst4q_v: |
4209 | 142 | case NEON::BI__builtin_neon_vst2_lane_v: |
4210 | 142 | case NEON::BI__builtin_neon_vst2q_lane_v: |
4211 | 142 | case NEON::BI__builtin_neon_vst3_lane_v: |
4212 | 142 | case NEON::BI__builtin_neon_vst3q_lane_v: |
4213 | 142 | case NEON::BI__builtin_neon_vst4_lane_v: |
4214 | 142 | case NEON::BI__builtin_neon_vst4q_lane_v: { |
4215 | 142 | llvm::Type *Tys[] = {Int8PtrTy, Ty}; |
4216 | 142 | Ops.push_back(getAlignmentValue32(PtrOp0)); |
4217 | 142 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, ""); |
4218 | 142 | } |
4219 | 24 | case NEON::BI__builtin_neon_vsubhn_v: { |
4220 | 24 | llvm::VectorType *SrcTy = |
4221 | 24 | llvm::VectorType::getExtendedElementVectorType(VTy); |
4222 | 24 | |
4223 | 24 | // %sum = add <4 x i32> %lhs, %rhs |
4224 | 24 | Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); |
4225 | 24 | Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); |
4226 | 24 | Ops[0] = Builder.CreateSub(Ops[0], Ops[1], "vsubhn"); |
4227 | 24 | |
4228 | 24 | // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> |
4229 | 24 | Constant *ShiftAmt = |
4230 | 24 | ConstantInt::get(SrcTy, SrcTy->getScalarSizeInBits() / 2); |
4231 | 24 | Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vsubhn"); |
4232 | 24 | |
4233 | 24 | // %res = trunc <4 x i32> %high to <4 x i16> |
4234 | 24 | return Builder.CreateTrunc(Ops[0], VTy, "vsubhn"); |
4235 | 142 | } |
4236 | 18 | case NEON::BI__builtin_neon_vtrn_v: |
4237 | 18 | case NEON::BI__builtin_neon_vtrnq_v: { |
4238 | 18 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); |
4239 | 18 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
4240 | 18 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
4241 | 18 | Value *SV = nullptr; |
4242 | 18 | |
4243 | 54 | for (unsigned vi = 0; vi != 254 ; ++vi36 ) { |
4244 | 36 | SmallVector<uint32_t, 16> Indices; |
4245 | 162 | for (unsigned i = 0, e = VTy->getNumElements(); i != e162 ; i += 2126 ) { |
4246 | 126 | Indices.push_back(i+vi); |
4247 | 126 | Indices.push_back(i+e+vi); |
4248 | 126 | } |
4249 | 36 | Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); |
4250 | 36 | SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vtrn"); |
4251 | 36 | SV = Builder.CreateDefaultAlignedStore(SV, Addr); |
4252 | 36 | } |
4253 | 18 | return SV; |
4254 | 18 | } |
4255 | 78 | case NEON::BI__builtin_neon_vtst_v: |
4256 | 78 | case NEON::BI__builtin_neon_vtstq_v: { |
4257 | 78 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
4258 | 78 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
4259 | 78 | Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); |
4260 | 78 | Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], |
4261 | 78 | ConstantAggregateZero::get(Ty)); |
4262 | 78 | return Builder.CreateSExt(Ops[0], Ty, "vtst"); |
4263 | 78 | } |
4264 | 18 | case NEON::BI__builtin_neon_vuzp_v: |
4265 | 18 | case NEON::BI__builtin_neon_vuzpq_v: { |
4266 | 18 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); |
4267 | 18 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
4268 | 18 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
4269 | 18 | Value *SV = nullptr; |
4270 | 18 | |
4271 | 54 | for (unsigned vi = 0; vi != 254 ; ++vi36 ) { |
4272 | 36 | SmallVector<uint32_t, 16> Indices; |
4273 | 288 | for (unsigned i = 0, e = VTy->getNumElements(); i != e288 ; ++i252 ) |
4274 | 252 | Indices.push_back(2*i+vi); |
4275 | 36 | |
4276 | 36 | Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); |
4277 | 36 | SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vuzp"); |
4278 | 36 | SV = Builder.CreateDefaultAlignedStore(SV, Addr); |
4279 | 36 | } |
4280 | 18 | return SV; |
4281 | 18 | } |
4282 | 19 | case NEON::BI__builtin_neon_vzip_v: |
4283 | 19 | case NEON::BI__builtin_neon_vzipq_v: { |
4284 | 19 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); |
4285 | 19 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
4286 | 19 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
4287 | 19 | Value *SV = nullptr; |
4288 | 19 | |
4289 | 57 | for (unsigned vi = 0; vi != 257 ; ++vi38 ) { |
4290 | 38 | SmallVector<uint32_t, 16> Indices; |
4291 | 180 | for (unsigned i = 0, e = VTy->getNumElements(); i != e180 ; i += 2142 ) { |
4292 | 142 | Indices.push_back((i + vi*e) >> 1); |
4293 | 142 | Indices.push_back(((i + vi*e) >> 1)+e); |
4294 | 142 | } |
4295 | 38 | Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); |
4296 | 38 | SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vzip"); |
4297 | 38 | SV = Builder.CreateDefaultAlignedStore(SV, Addr); |
4298 | 38 | } |
4299 | 19 | return SV; |
4300 | 1.12k | } |
4301 | 1.12k | } |
4302 | 1.12k | |
4303 | 2.53k | assert(Int && "Expected valid intrinsic number"); |
4304 | 1.12k | |
4305 | 1.12k | // Determine the type(s) of this overloaded AArch64 intrinsic. |
4306 | 1.12k | Function *F = LookupNeonLLVMIntrinsic(Int, Modifier, Ty, E); |
4307 | 1.12k | |
4308 | 1.12k | Value *Result = EmitNeonCall(F, Ops, NameHint); |
4309 | 1.12k | llvm::Type *ResultType = ConvertType(E->getType()); |
4310 | 1.12k | // AArch64 intrinsic one-element vector type cast to |
4311 | 1.12k | // scalar type expected by the builtin |
4312 | 1.12k | return Builder.CreateBitCast(Result, ResultType, NameHint); |
4313 | 1.12k | } |
4314 | | |
4315 | | Value *CodeGenFunction::EmitAArch64CompareBuiltinExpr( |
4316 | | Value *Op, llvm::Type *Ty, const CmpInst::Predicate Fp, |
4317 | 259 | const CmpInst::Predicate Ip, const Twine &Name) { |
4318 | 259 | llvm::Type *OTy = Op->getType(); |
4319 | 259 | |
4320 | 259 | // FIXME: this is utterly horrific. We should not be looking at previous |
4321 | 259 | // codegen context to find out what needs doing. Unfortunately TableGen |
4322 | 259 | // currently gives us exactly the same calls for vceqz_f32 and vceqz_s32 |
4323 | 259 | // (etc). |
4324 | 259 | if (BitCastInst *BI = dyn_cast<BitCastInst>(Op)) |
4325 | 172 | OTy = BI->getOperand(0)->getType(); |
4326 | 259 | |
4327 | 259 | Op = Builder.CreateBitCast(Op, OTy); |
4328 | 259 | if (OTy->getScalarType()->isFloatingPointTy()259 ) { |
4329 | 90 | Op = Builder.CreateFCmp(Fp, Op, Constant::getNullValue(OTy)); |
4330 | 259 | } else { |
4331 | 169 | Op = Builder.CreateICmp(Ip, Op, Constant::getNullValue(OTy)); |
4332 | 169 | } |
4333 | 259 | return Builder.CreateSExt(Op, Ty, Name); |
4334 | 259 | } |
4335 | | |
4336 | | static Value *packTBLDVectorList(CodeGenFunction &CGF, ArrayRef<Value *> Ops, |
4337 | | Value *ExtOp, Value *IndexOp, |
4338 | | llvm::Type *ResTy, unsigned IntID, |
4339 | 72 | const char *Name) { |
4340 | 72 | SmallVector<Value *, 2> TblOps; |
4341 | 72 | if (ExtOp) |
4342 | 18 | TblOps.push_back(ExtOp); |
4343 | 72 | |
4344 | 72 | // Build a vector containing sequential number like (0, 1, 2, ..., 15) |
4345 | 72 | SmallVector<uint32_t, 16> Indices; |
4346 | 72 | llvm::VectorType *TblTy = cast<llvm::VectorType>(Ops[0]->getType()); |
4347 | 648 | for (unsigned i = 0, e = TblTy->getNumElements(); i != e648 ; ++i576 ) { |
4348 | 576 | Indices.push_back(2*i); |
4349 | 576 | Indices.push_back(2*i+1); |
4350 | 576 | } |
4351 | 72 | |
4352 | 72 | int PairPos = 0, End = Ops.size() - 1; |
4353 | 144 | while (PairPos < End144 ) { |
4354 | 72 | TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], |
4355 | 72 | Ops[PairPos+1], Indices, |
4356 | 72 | Name)); |
4357 | 72 | PairPos += 2; |
4358 | 72 | } |
4359 | 72 | |
4360 | 72 | // If there's an odd number of 64-bit lookup table, fill the high 64-bit |
4361 | 72 | // of the 128-bit lookup table with zero. |
4362 | 72 | if (PairPos == End72 ) { |
4363 | 36 | Value *ZeroTbl = ConstantAggregateZero::get(TblTy); |
4364 | 36 | TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], |
4365 | 36 | ZeroTbl, Indices, Name)); |
4366 | 36 | } |
4367 | 72 | |
4368 | 72 | Function *TblF; |
4369 | 72 | TblOps.push_back(IndexOp); |
4370 | 72 | TblF = CGF.CGM.getIntrinsic(IntID, ResTy); |
4371 | 72 | |
4372 | 72 | return CGF.EmitNeonCall(TblF, TblOps, Name); |
4373 | 72 | } |
4374 | | |
4375 | 1.34k | Value *CodeGenFunction::GetValueForARMHint(unsigned BuiltinID) { |
4376 | 1.34k | unsigned Value; |
4377 | 1.34k | switch (BuiltinID) { |
4378 | 1.32k | default: |
4379 | 1.32k | return nullptr; |
4380 | 2 | case ARM::BI__builtin_arm_nop: |
4381 | 2 | Value = 0; |
4382 | 2 | break; |
4383 | 3 | case ARM::BI__builtin_arm_yield: |
4384 | 3 | case ARM::BI__yield: |
4385 | 3 | Value = 1; |
4386 | 3 | break; |
4387 | 3 | case ARM::BI__builtin_arm_wfe: |
4388 | 3 | case ARM::BI__wfe: |
4389 | 3 | Value = 2; |
4390 | 3 | break; |
4391 | 3 | case ARM::BI__builtin_arm_wfi: |
4392 | 3 | case ARM::BI__wfi: |
4393 | 3 | Value = 3; |
4394 | 3 | break; |
4395 | 3 | case ARM::BI__builtin_arm_sev: |
4396 | 3 | case ARM::BI__sev: |
4397 | 3 | Value = 4; |
4398 | 3 | break; |
4399 | 3 | case ARM::BI__builtin_arm_sevl: |
4400 | 3 | case ARM::BI__sevl: |
4401 | 3 | Value = 5; |
4402 | 3 | break; |
4403 | 17 | } |
4404 | 17 | |
4405 | 17 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), |
4406 | 17 | llvm::ConstantInt::get(Int32Ty, Value)); |
4407 | 17 | } |
4408 | | |
4409 | | // Generates the IR for the read/write special register builtin, |
4410 | | // ValueType is the type of the value that is to be written or read, |
4411 | | // RegisterType is the type of the register being written to or read from. |
4412 | | static Value *EmitSpecialRegisterBuiltin(CodeGenFunction &CGF, |
4413 | | const CallExpr *E, |
4414 | | llvm::Type *RegisterType, |
4415 | | llvm::Type *ValueType, |
4416 | | bool IsRead, |
4417 | 26 | StringRef SysReg = "") { |
4418 | 26 | // write and register intrinsics only support 32 and 64 bit operations. |
4419 | 26 | assert((RegisterType->isIntegerTy(32) || RegisterType->isIntegerTy(64)) |
4420 | 26 | && "Unsupported size for register."); |
4421 | 26 | |
4422 | 26 | CodeGen::CGBuilderTy &Builder = CGF.Builder; |
4423 | 26 | CodeGen::CodeGenModule &CGM = CGF.CGM; |
4424 | 26 | LLVMContext &Context = CGM.getLLVMContext(); |
4425 | 26 | |
4426 | 26 | if (SysReg.empty()26 ) { |
4427 | 24 | const Expr *SysRegStrExpr = E->getArg(0)->IgnoreParenCasts(); |
4428 | 24 | SysReg = cast<clang::StringLiteral>(SysRegStrExpr)->getString(); |
4429 | 24 | } |
4430 | 26 | |
4431 | 26 | llvm::Metadata *Ops[] = { llvm::MDString::get(Context, SysReg) }; |
4432 | 26 | llvm::MDNode *RegName = llvm::MDNode::get(Context, Ops); |
4433 | 26 | llvm::Value *Metadata = llvm::MetadataAsValue::get(Context, RegName); |
4434 | 26 | |
4435 | 26 | llvm::Type *Types[] = { RegisterType }; |
4436 | 26 | |
4437 | 18 | bool MixedTypes = RegisterType->isIntegerTy(64) && ValueType->isIntegerTy(32); |
4438 | 26 | assert(!(RegisterType->isIntegerTy(32) && ValueType->isIntegerTy(64)) |
4439 | 26 | && "Can't fit 64-bit value in 32-bit register"); |
4440 | 26 | |
4441 | 26 | if (IsRead26 ) { |
4442 | 14 | llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::read_register, Types); |
4443 | 14 | llvm::Value *Call = Builder.CreateCall(F, Metadata); |
4444 | 14 | |
4445 | 14 | if (MixedTypes) |
4446 | 14 | // Read into 64 bit register and then truncate result to 32 bit. |
4447 | 2 | return Builder.CreateTrunc(Call, ValueType); |
4448 | 12 | |
4449 | 12 | if (12 ValueType->isPointerTy()12 ) |
4450 | 12 | // Have i32/i64 result (Call) but want to return a VoidPtrTy (i8*). |
4451 | 4 | return Builder.CreateIntToPtr(Call, ValueType); |
4452 | 8 | |
4453 | 8 | return Call; |
4454 | 8 | } |
4455 | 12 | |
4456 | 12 | llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::write_register, Types); |
4457 | 12 | llvm::Value *ArgValue = CGF.EmitScalarExpr(E->getArg(1)); |
4458 | 12 | if (MixedTypes12 ) { |
4459 | 2 | // Extend 32 bit write value to 64 bit to pass to write. |
4460 | 2 | ArgValue = Builder.CreateZExt(ArgValue, RegisterType); |
4461 | 2 | return Builder.CreateCall(F, { Metadata, ArgValue }); |
4462 | 2 | } |
4463 | 10 | |
4464 | 10 | if (10 ValueType->isPointerTy()10 ) { |
4465 | 4 | // Have VoidPtrTy ArgValue but want to return an i32/i64. |
4466 | 4 | ArgValue = Builder.CreatePtrToInt(ArgValue, RegisterType); |
4467 | 4 | return Builder.CreateCall(F, { Metadata, ArgValue }); |
4468 | 4 | } |
4469 | 6 | |
4470 | 6 | return Builder.CreateCall(F, { Metadata, ArgValue }); |
4471 | 6 | } |
4472 | | |
4473 | | /// Return true if BuiltinID is an overloaded Neon intrinsic with an extra |
4474 | | /// argument that specifies the vector type. |
4475 | 1.23k | static bool HasExtraNeonArgument(unsigned BuiltinID) { |
4476 | 1.23k | switch (BuiltinID) { |
4477 | 1.18k | default: break; |
4478 | 54 | case NEON::BI__builtin_neon_vget_lane_i8: |
4479 | 54 | case NEON::BI__builtin_neon_vget_lane_i16: |
4480 | 54 | case NEON::BI__builtin_neon_vget_lane_i32: |
4481 | 54 | case NEON::BI__builtin_neon_vget_lane_i64: |
4482 | 54 | case NEON::BI__builtin_neon_vget_lane_f32: |
4483 | 54 | case NEON::BI__builtin_neon_vgetq_lane_i8: |
4484 | 54 | case NEON::BI__builtin_neon_vgetq_lane_i16: |
4485 | 54 | case NEON::BI__builtin_neon_vgetq_lane_i32: |
4486 | 54 | case NEON::BI__builtin_neon_vgetq_lane_i64: |
4487 | 54 | case NEON::BI__builtin_neon_vgetq_lane_f32: |
4488 | 54 | case NEON::BI__builtin_neon_vset_lane_i8: |
4489 | 54 | case NEON::BI__builtin_neon_vset_lane_i16: |
4490 | 54 | case NEON::BI__builtin_neon_vset_lane_i32: |
4491 | 54 | case NEON::BI__builtin_neon_vset_lane_i64: |
4492 | 54 | case NEON::BI__builtin_neon_vset_lane_f32: |
4493 | 54 | case NEON::BI__builtin_neon_vsetq_lane_i8: |
4494 | 54 | case NEON::BI__builtin_neon_vsetq_lane_i16: |
4495 | 54 | case NEON::BI__builtin_neon_vsetq_lane_i32: |
4496 | 54 | case NEON::BI__builtin_neon_vsetq_lane_i64: |
4497 | 54 | case NEON::BI__builtin_neon_vsetq_lane_f32: |
4498 | 54 | case NEON::BI__builtin_neon_vsha1h_u32: |
4499 | 54 | case NEON::BI__builtin_neon_vsha1cq_u32: |
4500 | 54 | case NEON::BI__builtin_neon_vsha1pq_u32: |
4501 | 54 | case NEON::BI__builtin_neon_vsha1mq_u32: |
4502 | 54 | case ARM::BI_MoveToCoprocessor: |
4503 | 54 | case ARM::BI_MoveToCoprocessor2: |
4504 | 54 | return false; |
4505 | 1.18k | } |
4506 | 1.18k | return true; |
4507 | 1.18k | } |
4508 | | |
4509 | | Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID, |
4510 | 1.34k | const CallExpr *E) { |
4511 | 1.34k | if (auto Hint = GetValueForARMHint(BuiltinID)) |
4512 | 17 | return Hint; |
4513 | 1.32k | |
4514 | 1.32k | if (1.32k BuiltinID == ARM::BI__emit1.32k ) { |
4515 | 2 | bool IsThumb = getTarget().getTriple().getArch() == llvm::Triple::thumb; |
4516 | 2 | llvm::FunctionType *FTy = |
4517 | 2 | llvm::FunctionType::get(VoidTy, /*Variadic=*/false); |
4518 | 2 | |
4519 | 2 | APSInt Value; |
4520 | 2 | if (!E->getArg(0)->EvaluateAsInt(Value, CGM.getContext())) |
4521 | 0 | llvm_unreachable("Sema will ensure that the parameter is constant"); |
4522 | 2 | |
4523 | 2 | uint64_t ZExtValue = Value.zextOrTrunc(IsThumb ? 2 162 : 320 ).getZExtValue(); |
4524 | 2 | |
4525 | 2 | llvm::InlineAsm *Emit = |
4526 | 2 | IsThumb ? InlineAsm::get(FTy, ".inst.n 0x" + utohexstr(ZExtValue), "", |
4527 | 2 | /*SideEffects=*/true) |
4528 | 0 | : InlineAsm::get(FTy, ".inst 0x" + utohexstr(ZExtValue), "", |
4529 | 0 | /*SideEffects=*/true); |
4530 | 2 | |
4531 | 2 | return Builder.CreateCall(Emit); |
4532 | 1.32k | } |
4533 | 1.32k | |
4534 | 1.32k | if (1.32k BuiltinID == ARM::BI__builtin_arm_dbg1.32k ) { |
4535 | 2 | Value *Option = EmitScalarExpr(E->getArg(0)); |
4536 | 2 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_dbg), Option); |
4537 | 2 | } |
4538 | 1.32k | |
4539 | 1.32k | if (1.32k BuiltinID == ARM::BI__builtin_arm_prefetch1.32k ) { |
4540 | 7 | Value *Address = EmitScalarExpr(E->getArg(0)); |
4541 | 7 | Value *RW = EmitScalarExpr(E->getArg(1)); |
4542 | 7 | Value *IsData = EmitScalarExpr(E->getArg(2)); |
4543 | 7 | |
4544 | 7 | // Locality is not supported on ARM target |
4545 | 7 | Value *Locality = llvm::ConstantInt::get(Int32Ty, 3); |
4546 | 7 | |
4547 | 7 | Value *F = CGM.getIntrinsic(Intrinsic::prefetch); |
4548 | 7 | return Builder.CreateCall(F, {Address, RW, Locality, IsData}); |
4549 | 7 | } |
4550 | 1.31k | |
4551 | 1.31k | if (1.31k BuiltinID == ARM::BI__builtin_arm_rbit1.31k ) { |
4552 | 4 | llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); |
4553 | 4 | return Builder.CreateCall( |
4554 | 4 | CGM.getIntrinsic(Intrinsic::bitreverse, Arg->getType()), Arg, "rbit"); |
4555 | 4 | } |
4556 | 1.31k | |
4557 | 1.31k | if (1.31k BuiltinID == ARM::BI__clear_cache1.31k ) { |
4558 | 2 | assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); |
4559 | 2 | const FunctionDecl *FD = E->getDirectCallee(); |
4560 | 2 | Value *Ops[2]; |
4561 | 6 | for (unsigned i = 0; i < 26 ; i++4 ) |
4562 | 4 | Ops[i] = EmitScalarExpr(E->getArg(i)); |
4563 | 2 | llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); |
4564 | 2 | llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); |
4565 | 2 | StringRef Name = FD->getName(); |
4566 | 2 | return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); |
4567 | 2 | } |
4568 | 1.31k | |
4569 | 1.31k | if (1.31k BuiltinID == ARM::BI__builtin_arm_mcrr || |
4570 | 1.31k | BuiltinID == ARM::BI__builtin_arm_mcrr21.31k ) { |
4571 | 2 | Function *F; |
4572 | 2 | |
4573 | 2 | switch (BuiltinID) { |
4574 | 0 | default: 0 llvm_unreachable0 ("unexpected builtin"); |
4575 | 1 | case ARM::BI__builtin_arm_mcrr: |
4576 | 1 | F = CGM.getIntrinsic(Intrinsic::arm_mcrr); |
4577 | 1 | break; |
4578 | 1 | case ARM::BI__builtin_arm_mcrr2: |
4579 | 1 | F = CGM.getIntrinsic(Intrinsic::arm_mcrr2); |
4580 | 1 | break; |
4581 | 2 | } |
4582 | 2 | |
4583 | 2 | // MCRR{2} instruction has 5 operands but |
4584 | 2 | // the intrinsic has 4 because Rt and Rt2 |
4585 | 2 | // are represented as a single unsigned 64 |
4586 | 2 | // bit integer in the intrinsic definition |
4587 | 2 | // but internally it's represented as 2 32 |
4588 | 2 | // bit integers. |
4589 | 2 | |
4590 | 2 | Value *Coproc = EmitScalarExpr(E->getArg(0)); |
4591 | 2 | Value *Opc1 = EmitScalarExpr(E->getArg(1)); |
4592 | 2 | Value *RtAndRt2 = EmitScalarExpr(E->getArg(2)); |
4593 | 2 | Value *CRm = EmitScalarExpr(E->getArg(3)); |
4594 | 2 | |
4595 | 2 | Value *C1 = llvm::ConstantInt::get(Int64Ty, 32); |
4596 | 2 | Value *Rt = Builder.CreateTruncOrBitCast(RtAndRt2, Int32Ty); |
4597 | 2 | Value *Rt2 = Builder.CreateLShr(RtAndRt2, C1); |
4598 | 2 | Rt2 = Builder.CreateTruncOrBitCast(Rt2, Int32Ty); |
4599 | 2 | |
4600 | 2 | return Builder.CreateCall(F, {Coproc, Opc1, Rt, Rt2, CRm}); |
4601 | 2 | } |
4602 | 1.30k | |
4603 | 1.30k | if (1.30k BuiltinID == ARM::BI__builtin_arm_mrrc || |
4604 | 1.30k | BuiltinID == ARM::BI__builtin_arm_mrrc21.30k ) { |
4605 | 2 | Function *F; |
4606 | 2 | |
4607 | 2 | switch (BuiltinID) { |
4608 | 0 | default: 0 llvm_unreachable0 ("unexpected builtin"); |
4609 | 1 | case ARM::BI__builtin_arm_mrrc: |
4610 | 1 | F = CGM.getIntrinsic(Intrinsic::arm_mrrc); |
4611 | 1 | break; |
4612 | 1 | case ARM::BI__builtin_arm_mrrc2: |
4613 | 1 | F = CGM.getIntrinsic(Intrinsic::arm_mrrc2); |
4614 | 1 | break; |
4615 | 2 | } |
4616 | 2 | |
4617 | 2 | Value *Coproc = EmitScalarExpr(E->getArg(0)); |
4618 | 2 | Value *Opc1 = EmitScalarExpr(E->getArg(1)); |
4619 | 2 | Value *CRm = EmitScalarExpr(E->getArg(2)); |
4620 | 2 | Value *RtAndRt2 = Builder.CreateCall(F, {Coproc, Opc1, CRm}); |
4621 | 2 | |
4622 | 2 | // Returns an unsigned 64 bit integer, represented |
4623 | 2 | // as two 32 bit integers. |
4624 | 2 | |
4625 | 2 | Value *Rt = Builder.CreateExtractValue(RtAndRt2, 1); |
4626 | 2 | Value *Rt1 = Builder.CreateExtractValue(RtAndRt2, 0); |
4627 | 2 | Rt = Builder.CreateZExt(Rt, Int64Ty); |
4628 | 2 | Rt1 = Builder.CreateZExt(Rt1, Int64Ty); |
4629 | 2 | |
4630 | 2 | Value *ShiftCast = llvm::ConstantInt::get(Int64Ty, 32); |
4631 | 2 | RtAndRt2 = Builder.CreateShl(Rt, ShiftCast, "shl", true); |
4632 | 2 | RtAndRt2 = Builder.CreateOr(RtAndRt2, Rt1); |
4633 | 2 | |
4634 | 2 | return Builder.CreateBitCast(RtAndRt2, ConvertType(E->getType())); |
4635 | 2 | } |
4636 | 1.30k | |
4637 | 1.30k | if (1.30k BuiltinID == ARM::BI__builtin_arm_ldrexd || |
4638 | 1.30k | ((BuiltinID == ARM::BI__builtin_arm_ldrex || |
4639 | 1.29k | BuiltinID == ARM::BI__builtin_arm_ldaex) && |
4640 | 20 | getContext().getTypeSize(E->getType()) == 64) || |
4641 | 1.30k | BuiltinID == ARM::BI__ldrexd1.30k ) { |
4642 | 7 | Function *F; |
4643 | 7 | |
4644 | 7 | switch (BuiltinID) { |
4645 | 0 | default: 0 llvm_unreachable0 ("unexpected builtin"); |
4646 | 3 | case ARM::BI__builtin_arm_ldaex: |
4647 | 3 | F = CGM.getIntrinsic(Intrinsic::arm_ldaexd); |
4648 | 3 | break; |
4649 | 4 | case ARM::BI__builtin_arm_ldrexd: |
4650 | 4 | case ARM::BI__builtin_arm_ldrex: |
4651 | 4 | case ARM::BI__ldrexd: |
4652 | 4 | F = CGM.getIntrinsic(Intrinsic::arm_ldrexd); |
4653 | 4 | break; |
4654 | 7 | } |
4655 | 7 | |
4656 | 7 | Value *LdPtr = EmitScalarExpr(E->getArg(0)); |
4657 | 7 | Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), |
4658 | 7 | "ldrexd"); |
4659 | 7 | |
4660 | 7 | Value *Val0 = Builder.CreateExtractValue(Val, 1); |
4661 | 7 | Value *Val1 = Builder.CreateExtractValue(Val, 0); |
4662 | 7 | Val0 = Builder.CreateZExt(Val0, Int64Ty); |
4663 | 7 | Val1 = Builder.CreateZExt(Val1, Int64Ty); |
4664 | 7 | |
4665 | 7 | Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32); |
4666 | 7 | Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); |
4667 | 7 | Val = Builder.CreateOr(Val, Val1); |
4668 | 7 | return Builder.CreateBitCast(Val, ConvertType(E->getType())); |
4669 | 7 | } |
4670 | 1.30k | |
4671 | 1.30k | if (1.30k BuiltinID == ARM::BI__builtin_arm_ldrex || |
4672 | 1.30k | BuiltinID == ARM::BI__builtin_arm_ldaex1.29k ) { |
4673 | 14 | Value *LoadAddr = EmitScalarExpr(E->getArg(0)); |
4674 | 14 | |
4675 | 14 | QualType Ty = E->getType(); |
4676 | 14 | llvm::Type *RealResTy = ConvertType(Ty); |
4677 | 14 | llvm::Type *PtrTy = llvm::IntegerType::get( |
4678 | 14 | getLLVMContext(), getContext().getTypeSize(Ty))->getPointerTo(); |
4679 | 14 | LoadAddr = Builder.CreateBitCast(LoadAddr, PtrTy); |
4680 | 14 | |
4681 | 14 | Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_ldaex |
4682 | 6 | ? Intrinsic::arm_ldaex |
4683 | 8 | : Intrinsic::arm_ldrex, |
4684 | 14 | PtrTy); |
4685 | 14 | Value *Val = Builder.CreateCall(F, LoadAddr, "ldrex"); |
4686 | 14 | |
4687 | 14 | if (RealResTy->isPointerTy()) |
4688 | 4 | return Builder.CreateIntToPtr(Val, RealResTy); |
4689 | 10 | else { |
4690 | 10 | llvm::Type *IntResTy = llvm::IntegerType::get( |
4691 | 10 | getLLVMContext(), CGM.getDataLayout().getTypeSizeInBits(RealResTy)); |
4692 | 10 | Val = Builder.CreateTruncOrBitCast(Val, IntResTy); |
4693 | 10 | return Builder.CreateBitCast(Val, RealResTy); |
4694 | 10 | } |
4695 | 1.28k | } |
4696 | 1.28k | |
4697 | 1.28k | if (1.28k BuiltinID == ARM::BI__builtin_arm_strexd || |
4698 | 1.28k | ((BuiltinID == ARM::BI__builtin_arm_stlex || |
4699 | 1.27k | BuiltinID == ARM::BI__builtin_arm_strex) && |
4700 | 1.28k | getContext().getTypeSize(E->getArg(0)->getType()) == 6416 )) { |
4701 | 4 | Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex |
4702 | 2 | ? Intrinsic::arm_stlexd |
4703 | 2 | : Intrinsic::arm_strexd); |
4704 | 4 | llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty); |
4705 | 4 | |
4706 | 4 | Address Tmp = CreateMemTemp(E->getArg(0)->getType()); |
4707 | 4 | Value *Val = EmitScalarExpr(E->getArg(0)); |
4708 | 4 | Builder.CreateStore(Val, Tmp); |
4709 | 4 | |
4710 | 4 | Address LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); |
4711 | 4 | Val = Builder.CreateLoad(LdPtr); |
4712 | 4 | |
4713 | 4 | Value *Arg0 = Builder.CreateExtractValue(Val, 0); |
4714 | 4 | Value *Arg1 = Builder.CreateExtractValue(Val, 1); |
4715 | 4 | Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), Int8PtrTy); |
4716 | 4 | return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "strexd"); |
4717 | 4 | } |
4718 | 1.28k | |
4719 | 1.28k | if (1.28k BuiltinID == ARM::BI__builtin_arm_strex || |
4720 | 1.28k | BuiltinID == ARM::BI__builtin_arm_stlex1.27k ) { |
4721 | 12 | Value *StoreVal = EmitScalarExpr(E->getArg(0)); |
4722 | 12 | Value *StoreAddr = EmitScalarExpr(E->getArg(1)); |
4723 | 12 | |
4724 | 12 | QualType Ty = E->getArg(0)->getType(); |
4725 | 12 | llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), |
4726 | 12 | getContext().getTypeSize(Ty)); |
4727 | 12 | StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); |
4728 | 12 | |
4729 | 12 | if (StoreVal->getType()->isPointerTy()) |
4730 | 2 | StoreVal = Builder.CreatePtrToInt(StoreVal, Int32Ty); |
4731 | 10 | else { |
4732 | 10 | llvm::Type *IntTy = llvm::IntegerType::get( |
4733 | 10 | getLLVMContext(), |
4734 | 10 | CGM.getDataLayout().getTypeSizeInBits(StoreVal->getType())); |
4735 | 10 | StoreVal = Builder.CreateBitCast(StoreVal, IntTy); |
4736 | 10 | StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int32Ty); |
4737 | 10 | } |
4738 | 12 | |
4739 | 12 | Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex |
4740 | 5 | ? Intrinsic::arm_stlex |
4741 | 7 | : Intrinsic::arm_strex, |
4742 | 12 | StoreAddr->getType()); |
4743 | 12 | return Builder.CreateCall(F, {StoreVal, StoreAddr}, "strex"); |
4744 | 12 | } |
4745 | 1.27k | |
4746 | 1.27k | switch (BuiltinID) { |
4747 | 1 | case ARM::BI__iso_volatile_load8: |
4748 | 1 | case ARM::BI__iso_volatile_load16: |
4749 | 1 | case ARM::BI__iso_volatile_load32: |
4750 | 1 | case ARM::BI__iso_volatile_load64: { |
4751 | 1 | Value *Ptr = EmitScalarExpr(E->getArg(0)); |
4752 | 1 | QualType ElTy = E->getArg(0)->getType()->getPointeeType(); |
4753 | 1 | CharUnits LoadSize = getContext().getTypeSizeInChars(ElTy); |
4754 | 1 | llvm::Type *ITy = llvm::IntegerType::get(getLLVMContext(), |
4755 | 1 | LoadSize.getQuantity() * 8); |
4756 | 1 | Ptr = Builder.CreateBitCast(Ptr, ITy->getPointerTo()); |
4757 | 1 | llvm::LoadInst *Load = |
4758 | 1 | Builder.CreateAlignedLoad(Ptr, LoadSize); |
4759 | 1 | Load->setVolatile(true); |
4760 | 1 | return Load; |
4761 | 1 | } |
4762 | 1 | case ARM::BI__iso_volatile_store8: |
4763 | 1 | case ARM::BI__iso_volatile_store16: |
4764 | 1 | case ARM::BI__iso_volatile_store32: |
4765 | 1 | case ARM::BI__iso_volatile_store64: { |
4766 | 1 | Value *Ptr = EmitScalarExpr(E->getArg(0)); |
4767 | 1 | Value *Value = EmitScalarExpr(E->getArg(1)); |
4768 | 1 | QualType ElTy = E->getArg(0)->getType()->getPointeeType(); |
4769 | 1 | CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy); |
4770 | 1 | llvm::Type *ITy = llvm::IntegerType::get(getLLVMContext(), |
4771 | 1 | StoreSize.getQuantity() * 8); |
4772 | 1 | Ptr = Builder.CreateBitCast(Ptr, ITy->getPointerTo()); |
4773 | 1 | llvm::StoreInst *Store = |
4774 | 1 | Builder.CreateAlignedStore(Value, Ptr, |
4775 | 1 | StoreSize); |
4776 | 1 | Store->setVolatile(true); |
4777 | 1 | return Store; |
4778 | 1.26k | } |
4779 | 1.26k | } |
4780 | 1.26k | |
4781 | 1.26k | if (1.26k BuiltinID == ARM::BI__builtin_arm_clrex1.26k ) { |
4782 | 1 | Function *F = CGM.getIntrinsic(Intrinsic::arm_clrex); |
4783 | 1 | return Builder.CreateCall(F); |
4784 | 1 | } |
4785 | 1.26k | |
4786 | 1.26k | // CRC32 |
4787 | 1.26k | Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; |
4788 | 1.26k | switch (BuiltinID) { |
4789 | 2 | case ARM::BI__builtin_arm_crc32b: |
4790 | 2 | CRCIntrinsicID = Intrinsic::arm_crc32b; break; |
4791 | 2 | case ARM::BI__builtin_arm_crc32cb: |
4792 | 2 | CRCIntrinsicID = Intrinsic::arm_crc32cb; break; |
4793 | 2 | case ARM::BI__builtin_arm_crc32h: |
4794 | 2 | CRCIntrinsicID = Intrinsic::arm_crc32h; break; |
4795 | 2 | case ARM::BI__builtin_arm_crc32ch: |
4796 | 2 | CRCIntrinsicID = Intrinsic::arm_crc32ch; break; |
4797 | 4 | case ARM::BI__builtin_arm_crc32w: |
4798 | 4 | case ARM::BI__builtin_arm_crc32d: |
4799 | 4 | CRCIntrinsicID = Intrinsic::arm_crc32w; break; |
4800 | 4 | case ARM::BI__builtin_arm_crc32cw: |
4801 | 4 | case ARM::BI__builtin_arm_crc32cd: |
4802 | 4 | CRCIntrinsicID = Intrinsic::arm_crc32cw; break; |
4803 | 1.26k | } |
4804 | 1.26k | |
4805 | 1.26k | if (1.26k CRCIntrinsicID != Intrinsic::not_intrinsic1.26k ) { |
4806 | 16 | Value *Arg0 = EmitScalarExpr(E->getArg(0)); |
4807 | 16 | Value *Arg1 = EmitScalarExpr(E->getArg(1)); |
4808 | 16 | |
4809 | 16 | // crc32{c,}d intrinsics are implemnted as two calls to crc32{c,}w |
4810 | 16 | // intrinsics, hence we need different codegen for these cases. |
4811 | 16 | if (BuiltinID == ARM::BI__builtin_arm_crc32d || |
4812 | 16 | BuiltinID == ARM::BI__builtin_arm_crc32cd14 ) { |
4813 | 4 | Value *C1 = llvm::ConstantInt::get(Int64Ty, 32); |
4814 | 4 | Value *Arg1a = Builder.CreateTruncOrBitCast(Arg1, Int32Ty); |
4815 | 4 | Value *Arg1b = Builder.CreateLShr(Arg1, C1); |
4816 | 4 | Arg1b = Builder.CreateTruncOrBitCast(Arg1b, Int32Ty); |
4817 | 4 | |
4818 | 4 | Function *F = CGM.getIntrinsic(CRCIntrinsicID); |
4819 | 4 | Value *Res = Builder.CreateCall(F, {Arg0, Arg1a}); |
4820 | 4 | return Builder.CreateCall(F, {Res, Arg1b}); |
4821 | 0 | } else { |
4822 | 12 | Arg1 = Builder.CreateZExtOrBitCast(Arg1, Int32Ty); |
4823 | 12 | |
4824 | 12 | Function *F = CGM.getIntrinsic(CRCIntrinsicID); |
4825 | 12 | return Builder.CreateCall(F, {Arg0, Arg1}); |
4826 | 12 | } |
4827 | 1.25k | } |
4828 | 1.25k | |
4829 | 1.25k | if (1.25k BuiltinID == ARM::BI__builtin_arm_rsr || |
4830 | 1.24k | BuiltinID == ARM::BI__builtin_arm_rsr64 || |
4831 | 1.24k | BuiltinID == ARM::BI__builtin_arm_rsrp || |
4832 | 1.24k | BuiltinID == ARM::BI__builtin_arm_wsr || |
4833 | 1.24k | BuiltinID == ARM::BI__builtin_arm_wsr64 || |
4834 | 1.25k | BuiltinID == ARM::BI__builtin_arm_wsrp1.24k ) { |
4835 | 12 | |
4836 | 12 | bool IsRead = BuiltinID == ARM::BI__builtin_arm_rsr || |
4837 | 10 | BuiltinID == ARM::BI__builtin_arm_rsr64 || |
4838 | 8 | BuiltinID == ARM::BI__builtin_arm_rsrp; |
4839 | 12 | |
4840 | 12 | bool IsPointerBuiltin = BuiltinID == ARM::BI__builtin_arm_rsrp || |
4841 | 10 | BuiltinID == ARM::BI__builtin_arm_wsrp; |
4842 | 12 | |
4843 | 12 | bool Is64Bit = BuiltinID == ARM::BI__builtin_arm_rsr64 || |
4844 | 10 | BuiltinID == ARM::BI__builtin_arm_wsr64; |
4845 | 12 | |
4846 | 12 | llvm::Type *ValueType; |
4847 | 12 | llvm::Type *RegisterType; |
4848 | 12 | if (IsPointerBuiltin12 ) { |
4849 | 4 | ValueType = VoidPtrTy; |
4850 | 4 | RegisterType = Int32Ty; |
4851 | 12 | } else if (8 Is64Bit8 ) { |
4852 | 4 | ValueType = RegisterType = Int64Ty; |
4853 | 8 | } else { |
4854 | 4 | ValueType = RegisterType = Int32Ty; |
4855 | 4 | } |
4856 | 12 | |
4857 | 12 | return EmitSpecialRegisterBuiltin(*this, E, RegisterType, ValueType, IsRead); |
4858 | 12 | } |
4859 | 1.23k | |
4860 | 1.23k | // Find out if any arguments are required to be integer constant |
4861 | 1.23k | // expressions. |
4862 | 1.23k | unsigned ICEArguments = 0; |
4863 | 1.23k | ASTContext::GetBuiltinTypeError Error; |
4864 | 1.23k | getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); |
4865 | 1.23k | assert(Error == ASTContext::GE_None && "Should not codegen an error"); |
4866 | 1.23k | |
4867 | 46 | auto getAlignmentValue32 = [&](Address addr) -> Value* { |
4868 | 46 | return Builder.getInt32(addr.getAlignment().getQuantity()); |
4869 | 46 | }; |
4870 | 1.23k | |
4871 | 1.23k | Address PtrOp0 = Address::invalid(); |
4872 | 1.23k | Address PtrOp1 = Address::invalid(); |
4873 | 1.23k | SmallVector<Value*, 4> Ops; |
4874 | 1.23k | bool HasExtraArg = HasExtraNeonArgument(BuiltinID); |
4875 | 1.23k | unsigned NumArgs = E->getNumArgs() - (HasExtraArg ? 11.18k : 054 ); |
4876 | 4.33k | for (unsigned i = 0, e = NumArgs; i != e4.33k ; i++3.09k ) { |
4877 | 3.09k | if (i == 03.09k ) { |
4878 | 1.23k | switch (BuiltinID) { |
4879 | 245 | case NEON::BI__builtin_neon_vld1_v: |
4880 | 245 | case NEON::BI__builtin_neon_vld1q_v: |
4881 | 245 | case NEON::BI__builtin_neon_vld1q_lane_v: |
4882 | 245 | case NEON::BI__builtin_neon_vld1_lane_v: |
4883 | 245 | case NEON::BI__builtin_neon_vld1_dup_v: |
4884 | 245 | case NEON::BI__builtin_neon_vld1q_dup_v: |
4885 | 245 | case NEON::BI__builtin_neon_vst1_v: |
4886 | 245 | case NEON::BI__builtin_neon_vst1q_v: |
4887 | 245 | case NEON::BI__builtin_neon_vst1q_lane_v: |
4888 | 245 | case NEON::BI__builtin_neon_vst1_lane_v: |
4889 | 245 | case NEON::BI__builtin_neon_vst2_v: |
4890 | 245 | case NEON::BI__builtin_neon_vst2q_v: |
4891 | 245 | case NEON::BI__builtin_neon_vst2_lane_v: |
4892 | 245 | case NEON::BI__builtin_neon_vst2q_lane_v: |
4893 | 245 | case NEON::BI__builtin_neon_vst3_v: |
4894 | 245 | case NEON::BI__builtin_neon_vst3q_v: |
4895 | 245 | case NEON::BI__builtin_neon_vst3_lane_v: |
4896 | 245 | case NEON::BI__builtin_neon_vst3q_lane_v: |
4897 | 245 | case NEON::BI__builtin_neon_vst4_v: |
4898 | 245 | case NEON::BI__builtin_neon_vst4q_v: |
4899 | 245 | case NEON::BI__builtin_neon_vst4_lane_v: |
4900 | 245 | case NEON::BI__builtin_neon_vst4q_lane_v: |
4901 | 245 | // Get the alignment for the argument in addition to the value; |
4902 | 245 | // we'll use it later. |
4903 | 245 | PtrOp0 = EmitPointerWithAlignment(E->getArg(0)); |
4904 | 245 | Ops.push_back(PtrOp0.getPointer()); |
4905 | 245 | continue; |
4906 | 2.85k | } |
4907 | 2.85k | } |
4908 | 2.85k | if (2.85k i == 12.85k ) { |
4909 | 1.04k | switch (BuiltinID) { |
4910 | 157 | case NEON::BI__builtin_neon_vld2_v: |
4911 | 157 | case NEON::BI__builtin_neon_vld2q_v: |
4912 | 157 | case NEON::BI__builtin_neon_vld3_v: |
4913 | 157 | case NEON::BI__builtin_neon_vld3q_v: |
4914 | 157 | case NEON::BI__builtin_neon_vld4_v: |
4915 | 157 | case NEON::BI__builtin_neon_vld4q_v: |
4916 | 157 | case NEON::BI__builtin_neon_vld2_lane_v: |
4917 | 157 | case NEON::BI__builtin_neon_vld2q_lane_v: |
4918 | 157 | case NEON::BI__builtin_neon_vld3_lane_v: |
4919 | 157 | case NEON::BI__builtin_neon_vld3q_lane_v: |
4920 | 157 | case NEON::BI__builtin_neon_vld4_lane_v: |
4921 | 157 | case NEON::BI__builtin_neon_vld4q_lane_v: |
4922 | 157 | case NEON::BI__builtin_neon_vld2_dup_v: |
4923 | 157 | case NEON::BI__builtin_neon_vld3_dup_v: |
4924 | 157 | case NEON::BI__builtin_neon_vld4_dup_v: |
4925 | 157 | // Get the alignment for the argument in addition to the value; |
4926 | 157 | // we'll use it later. |
4927 | 157 | PtrOp1 = EmitPointerWithAlignment(E->getArg(1)); |
4928 | 157 | Ops.push_back(PtrOp1.getPointer()); |
4929 | 157 | continue; |
4930 | 2.69k | } |
4931 | 2.69k | } |
4932 | 2.69k | |
4933 | 2.69k | if (2.69k (ICEArguments & (1 << i)) == 02.69k ) { |
4934 | 2.26k | Ops.push_back(EmitScalarExpr(E->getArg(i))); |
4935 | 2.69k | } else { |
4936 | 428 | // If this is required to be a constant, constant fold it so that we know |
4937 | 428 | // that the generated intrinsic gets a ConstantInt. |
4938 | 428 | llvm::APSInt Result; |
4939 | 428 | bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); |
4940 | 428 | assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; |
4941 | 428 | Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); |
4942 | 428 | } |
4943 | 3.09k | } |
4944 | 1.23k | |
4945 | 1.23k | switch (BuiltinID) { |
4946 | 1.17k | default: break; |
4947 | 1.23k | |
4948 | 24 | case NEON::BI__builtin_neon_vget_lane_i8: |
4949 | 24 | case NEON::BI__builtin_neon_vget_lane_i16: |
4950 | 24 | case NEON::BI__builtin_neon_vget_lane_i32: |
4951 | 24 | case NEON::BI__builtin_neon_vget_lane_i64: |
4952 | 24 | case NEON::BI__builtin_neon_vget_lane_f32: |
4953 | 24 | case NEON::BI__builtin_neon_vgetq_lane_i8: |
4954 | 24 | case NEON::BI__builtin_neon_vgetq_lane_i16: |
4955 | 24 | case NEON::BI__builtin_neon_vgetq_lane_i32: |
4956 | 24 | case NEON::BI__builtin_neon_vgetq_lane_i64: |
4957 | 24 | case NEON::BI__builtin_neon_vgetq_lane_f32: |
4958 | 24 | return Builder.CreateExtractElement(Ops[0], Ops[1], "vget_lane"); |
4959 | 24 | |
4960 | 24 | case NEON::BI__builtin_neon_vset_lane_i8: |
4961 | 24 | case NEON::BI__builtin_neon_vset_lane_i16: |
4962 | 24 | case NEON::BI__builtin_neon_vset_lane_i32: |
4963 | 24 | case NEON::BI__builtin_neon_vset_lane_i64: |
4964 | 24 | case NEON::BI__builtin_neon_vset_lane_f32: |
4965 | 24 | case NEON::BI__builtin_neon_vsetq_lane_i8: |
4966 | 24 | case NEON::BI__builtin_neon_vsetq_lane_i16: |
4967 | 24 | case NEON::BI__builtin_neon_vsetq_lane_i32: |
4968 | 24 | case NEON::BI__builtin_neon_vsetq_lane_i64: |
4969 | 24 | case NEON::BI__builtin_neon_vsetq_lane_f32: |
4970 | 24 | return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); |
4971 | 24 | |
4972 | 1 | case NEON::BI__builtin_neon_vsha1h_u32: |
4973 | 1 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1h), Ops, |
4974 | 1 | "vsha1h"); |
4975 | 1 | case NEON::BI__builtin_neon_vsha1cq_u32: |
4976 | 1 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1c), Ops, |
4977 | 1 | "vsha1h"); |
4978 | 1 | case NEON::BI__builtin_neon_vsha1pq_u32: |
4979 | 1 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1p), Ops, |
4980 | 1 | "vsha1h"); |
4981 | 1 | case NEON::BI__builtin_neon_vsha1mq_u32: |
4982 | 1 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1m), Ops, |
4983 | 1 | "vsha1h"); |
4984 | 24 | |
4985 | 24 | // The ARM _MoveToCoprocessor builtins put the input register value as |
4986 | 24 | // the first argument, but the LLVM intrinsic expects it as the third one. |
4987 | 2 | case ARM::BI_MoveToCoprocessor: |
4988 | 2 | case ARM::BI_MoveToCoprocessor2: { |
4989 | 2 | Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI_MoveToCoprocessor ? |
4990 | 2 | Intrinsic::arm_mcr1 : Intrinsic::arm_mcr21 ); |
4991 | 2 | return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0], |
4992 | 2 | Ops[3], Ops[4], Ops[5]}); |
4993 | 2 | } |
4994 | 2 | case ARM::BI_BitScanForward: |
4995 | 2 | case ARM::BI_BitScanForward64: |
4996 | 2 | return EmitMSVCBuiltinExpr(MSVCIntrin::_BitScanForward, E); |
4997 | 2 | case ARM::BI_BitScanReverse: |
4998 | 2 | case ARM::BI_BitScanReverse64: |
4999 | 2 | return EmitMSVCBuiltinExpr(MSVCIntrin::_BitScanReverse, E); |
5000 | 2 | |
5001 | 1 | case ARM::BI_InterlockedAnd64: |
5002 | 1 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedAnd, E); |
5003 | 1 | case ARM::BI_InterlockedExchange64: |
5004 | 1 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchange, E); |
5005 | 1 | case ARM::BI_InterlockedExchangeAdd64: |
5006 | 1 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeAdd, E); |
5007 | 1 | case ARM::BI_InterlockedExchangeSub64: |
5008 | 1 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeSub, E); |
5009 | 1 | case ARM::BI_InterlockedOr64: |
5010 | 1 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedOr, E); |
5011 | 1 | case ARM::BI_InterlockedXor64: |
5012 | 1 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedXor, E); |
5013 | 1 | case ARM::BI_InterlockedDecrement64: |
5014 | 1 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedDecrement, E); |
5015 | 1 | case ARM::BI_InterlockedIncrement64: |
5016 | 1 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedIncrement, E); |
5017 | 1.17k | } |
5018 | 1.17k | |
5019 | 1.17k | // Get the last argument, which specifies the vector type. |
5020 | 1.23k | assert(HasExtraArg); |
5021 | 1.17k | llvm::APSInt Result; |
5022 | 1.17k | const Expr *Arg = E->getArg(E->getNumArgs()-1); |
5023 | 1.17k | if (!Arg->isIntegerConstantExpr(Result, getContext())) |
5024 | 0 | return nullptr; |
5025 | 1.17k | |
5026 | 1.17k | if (1.17k BuiltinID == ARM::BI__builtin_arm_vcvtr_f || |
5027 | 1.17k | BuiltinID == ARM::BI__builtin_arm_vcvtr_d1.17k ) { |
5028 | 0 | // Determine the overloaded type of this builtin. |
5029 | 0 | llvm::Type *Ty; |
5030 | 0 | if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f) |
5031 | 0 | Ty = FloatTy; |
5032 | 0 | else |
5033 | 0 | Ty = DoubleTy; |
5034 | 0 |
|
5035 | 0 | // Determine whether this is an unsigned conversion or not. |
5036 | 0 | bool usgn = Result.getZExtValue() == 1; |
5037 | 0 | unsigned Int = usgn ? Intrinsic::arm_vcvtru0 : Intrinsic::arm_vcvtr0 ; |
5038 | 0 |
|
5039 | 0 | // Call the appropriate intrinsic. |
5040 | 0 | Function *F = CGM.getIntrinsic(Int, Ty); |
5041 | 0 | return Builder.CreateCall(F, Ops, "vcvtr"); |
5042 | 0 | } |
5043 | 1.17k | |
5044 | 1.17k | // Determine the type of this overloaded NEON intrinsic. |
5045 | 1.17k | NeonTypeFlags Type(Result.getZExtValue()); |
5046 | 1.17k | bool usgn = Type.isUnsigned(); |
5047 | 1.17k | bool rightShift = false; |
5048 | 1.17k | |
5049 | 1.17k | llvm::VectorType *VTy = GetNeonType(this, Type); |
5050 | 1.17k | llvm::Type *Ty = VTy; |
5051 | 1.17k | if (!Ty) |
5052 | 0 | return nullptr; |
5053 | 1.17k | |
5054 | 1.17k | // Many NEON builtins have identical semantics and uses in ARM and |
5055 | 1.17k | // AArch64. Emit these in a single function. |
5056 | 1.17k | auto IntrinsicMap = makeArrayRef(ARMSIMDIntrinsicMap); |
5057 | 1.17k | const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( |
5058 | 1.17k | IntrinsicMap, BuiltinID, NEONSIMDIntrinsicsProvenSorted); |
5059 | 1.17k | if (Builtin) |
5060 | 958 | return EmitCommonNeonBuiltinExpr( |
5061 | 958 | Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, |
5062 | 958 | Builtin->NameHint, Builtin->TypeModifier, E, Ops, PtrOp0, PtrOp1); |
5063 | 215 | |
5064 | 215 | unsigned Int; |
5065 | 215 | switch (BuiltinID) { |
5066 | 0 | default: return nullptr; |
5067 | 13 | case NEON::BI__builtin_neon_vld1q_lane_v: |
5068 | 13 | // Handle 64-bit integer elements as a special case. Use shuffles of |
5069 | 13 | // one-element vectors to avoid poor code for i64 in the backend. |
5070 | 13 | if (VTy->getElementType()->isIntegerTy(64)13 ) { |
5071 | 3 | // Extract the other lane. |
5072 | 3 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
5073 | 3 | uint32_t Lane = cast<ConstantInt>(Ops[2])->getZExtValue(); |
5074 | 3 | Value *SV = llvm::ConstantVector::get(ConstantInt::get(Int32Ty, 1-Lane)); |
5075 | 3 | Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); |
5076 | 3 | // Load the value as a one-element vector. |
5077 | 3 | Ty = llvm::VectorType::get(VTy->getElementType(), 1); |
5078 | 3 | llvm::Type *Tys[] = {Ty, Int8PtrTy}; |
5079 | 3 | Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Tys); |
5080 | 3 | Value *Align = getAlignmentValue32(PtrOp0); |
5081 | 3 | Value *Ld = Builder.CreateCall(F, {Ops[0], Align}); |
5082 | 3 | // Combine them. |
5083 | 3 | uint32_t Indices[] = {1 - Lane, Lane}; |
5084 | 3 | SV = llvm::ConstantDataVector::get(getLLVMContext(), Indices); |
5085 | 3 | return Builder.CreateShuffleVector(Ops[1], Ld, SV, "vld1q_lane"); |
5086 | 3 | } |
5087 | 10 | // fall through |
5088 | 22 | case NEON::BI__builtin_neon_vld1_lane_v: { |
5089 | 22 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
5090 | 22 | PtrOp0 = Builder.CreateElementBitCast(PtrOp0, VTy->getElementType()); |
5091 | 22 | Value *Ld = Builder.CreateLoad(PtrOp0); |
5092 | 22 | return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane"); |
5093 | 10 | } |
5094 | 40 | case NEON::BI__builtin_neon_vld2_dup_v: |
5095 | 40 | case NEON::BI__builtin_neon_vld3_dup_v: |
5096 | 40 | case NEON::BI__builtin_neon_vld4_dup_v: { |
5097 | 40 | // Handle 64-bit elements as a special-case. There is no "dup" needed. |
5098 | 40 | if (VTy->getElementType()->getPrimitiveSizeInBits() == 6440 ) { |
5099 | 8 | switch (BuiltinID) { |
5100 | 2 | case NEON::BI__builtin_neon_vld2_dup_v: |
5101 | 2 | Int = Intrinsic::arm_neon_vld2; |
5102 | 2 | break; |
5103 | 3 | case NEON::BI__builtin_neon_vld3_dup_v: |
5104 | 3 | Int = Intrinsic::arm_neon_vld3; |
5105 | 3 | break; |
5106 | 3 | case NEON::BI__builtin_neon_vld4_dup_v: |
5107 | 3 | Int = Intrinsic::arm_neon_vld4; |
5108 | 3 | break; |
5109 | 0 | default: 0 llvm_unreachable0 ("unknown vld_dup intrinsic?"); |
5110 | 8 | } |
5111 | 8 | llvm::Type *Tys[] = {Ty, Int8PtrTy}; |
5112 | 8 | Function *F = CGM.getIntrinsic(Int, Tys); |
5113 | 8 | llvm::Value *Align = getAlignmentValue32(PtrOp1); |
5114 | 8 | Ops[1] = Builder.CreateCall(F, {Ops[1], Align}, "vld_dup"); |
5115 | 8 | Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); |
5116 | 8 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
5117 | 8 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
5118 | 8 | } |
5119 | 32 | switch (BuiltinID) { |
5120 | 10 | case NEON::BI__builtin_neon_vld2_dup_v: |
5121 | 10 | Int = Intrinsic::arm_neon_vld2lane; |
5122 | 10 | break; |
5123 | 11 | case NEON::BI__builtin_neon_vld3_dup_v: |
5124 | 11 | Int = Intrinsic::arm_neon_vld3lane; |
5125 | 11 | break; |
5126 | 11 | case NEON::BI__builtin_neon_vld4_dup_v: |
5127 | 11 | Int = Intrinsic::arm_neon_vld4lane; |
5128 | 11 | break; |
5129 | 0 | default: 0 llvm_unreachable0 ("unknown vld_dup intrinsic?"); |
5130 | 32 | } |
5131 | 32 | llvm::Type *Tys[] = {Ty, Int8PtrTy}; |
5132 | 32 | Function *F = CGM.getIntrinsic(Int, Tys); |
5133 | 32 | llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType()); |
5134 | 32 | |
5135 | 32 | SmallVector<Value*, 6> Args; |
5136 | 32 | Args.push_back(Ops[1]); |
5137 | 32 | Args.append(STy->getNumElements(), UndefValue::get(Ty)); |
5138 | 32 | |
5139 | 32 | llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); |
5140 | 32 | Args.push_back(CI); |
5141 | 32 | Args.push_back(getAlignmentValue32(PtrOp1)); |
5142 | 32 | |
5143 | 32 | Ops[1] = Builder.CreateCall(F, Args, "vld_dup"); |
5144 | 32 | // splat lane 0 to all elts in each vector of the result. |
5145 | 129 | for (unsigned i = 0, e = STy->getNumElements(); i != e129 ; ++i97 ) { |
5146 | 97 | Value *Val = Builder.CreateExtractValue(Ops[1], i); |
5147 | 97 | Value *Elt = Builder.CreateBitCast(Val, Ty); |
5148 | 97 | Elt = EmitNeonSplat(Elt, CI); |
5149 | 97 | Elt = Builder.CreateBitCast(Elt, Val->getType()); |
5150 | 97 | Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i); |
5151 | 97 | } |
5152 | 32 | Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); |
5153 | 32 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
5154 | 32 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
5155 | 32 | } |
5156 | 7 | case NEON::BI__builtin_neon_vqrshrn_n_v: |
5157 | 7 | Int = |
5158 | 7 | usgn ? Intrinsic::arm_neon_vqrshiftnu3 : Intrinsic::arm_neon_vqrshiftns4 ; |
5159 | 7 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n", |
5160 | 7 | 1, true); |
5161 | 3 | case NEON::BI__builtin_neon_vqrshrun_n_v: |
5162 | 3 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty), |
5163 | 3 | Ops, "vqrshrun_n", 1, true); |
5164 | 6 | case NEON::BI__builtin_neon_vqshrn_n_v: |
5165 | 6 | Int = usgn ? Intrinsic::arm_neon_vqshiftnu3 : Intrinsic::arm_neon_vqshiftns3 ; |
5166 | 6 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n", |
5167 | 6 | 1, true); |
5168 | 3 | case NEON::BI__builtin_neon_vqshrun_n_v: |
5169 | 3 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty), |
5170 | 3 | Ops, "vqshrun_n", 1, true); |
5171 | 0 | case NEON::BI__builtin_neon_vrecpe_v: |
5172 | 0 | case NEON::BI__builtin_neon_vrecpeq_v: |
5173 | 0 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty), |
5174 | 0 | Ops, "vrecpe"); |
5175 | 6 | case NEON::BI__builtin_neon_vrshrn_n_v: |
5176 | 6 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty), |
5177 | 6 | Ops, "vrshrn_n", 1, true); |
5178 | 16 | case NEON::BI__builtin_neon_vrsra_n_v: |
5179 | 16 | case NEON::BI__builtin_neon_vrsraq_n_v: |
5180 | 16 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
5181 | 16 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
5182 | 16 | Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true); |
5183 | 16 | Int = usgn ? Intrinsic::arm_neon_vrshiftu8 : Intrinsic::arm_neon_vrshifts8 ; |
5184 | 16 | Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Ty), {Ops[1], Ops[2]}); |
5185 | 16 | return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n"); |
5186 | 20 | case NEON::BI__builtin_neon_vsri_n_v: |
5187 | 20 | case NEON::BI__builtin_neon_vsriq_n_v: |
5188 | 20 | rightShift = true; |
5189 | 20 | LLVM_FALLTHROUGH; |
5190 | 40 | case NEON::BI__builtin_neon_vsli_n_v: |
5191 | 40 | case NEON::BI__builtin_neon_vsliq_n_v: |
5192 | 40 | Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift); |
5193 | 40 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty), |
5194 | 40 | Ops, "vsli_n"); |
5195 | 19 | case NEON::BI__builtin_neon_vsra_n_v: |
5196 | 19 | case NEON::BI__builtin_neon_vsraq_n_v: |
5197 | 19 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
5198 | 19 | Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); |
5199 | 19 | return Builder.CreateAdd(Ops[0], Ops[1]); |
5200 | 13 | case NEON::BI__builtin_neon_vst1q_lane_v: |
5201 | 13 | // Handle 64-bit integer elements as a special case. Use a shuffle to get |
5202 | 13 | // a one-element vector and avoid poor code for i64 in the backend. |
5203 | 13 | if (VTy->getElementType()->isIntegerTy(64)13 ) { |
5204 | 3 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
5205 | 3 | Value *SV = llvm::ConstantVector::get(cast<llvm::Constant>(Ops[2])); |
5206 | 3 | Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); |
5207 | 3 | Ops[2] = getAlignmentValue32(PtrOp0); |
5208 | 3 | llvm::Type *Tys[] = {Int8PtrTy, Ops[1]->getType()}; |
5209 | 3 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, |
5210 | 3 | Tys), Ops); |
5211 | 3 | } |
5212 | 10 | // fall through |
5213 | 23 | case NEON::BI__builtin_neon_vst1_lane_v: { |
5214 | 23 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
5215 | 23 | Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); |
5216 | 23 | Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); |
5217 | 23 | auto St = Builder.CreateStore(Ops[1], Builder.CreateBitCast(PtrOp0, Ty)); |
5218 | 23 | return St; |
5219 | 10 | } |
5220 | 3 | case NEON::BI__builtin_neon_vtbl1_v: |
5221 | 3 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1), |
5222 | 3 | Ops, "vtbl1"); |
5223 | 3 | case NEON::BI__builtin_neon_vtbl2_v: |
5224 | 3 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2), |
5225 | 3 | Ops, "vtbl2"); |
5226 | 3 | case NEON::BI__builtin_neon_vtbl3_v: |
5227 | 3 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3), |
5228 | 3 | Ops, "vtbl3"); |
5229 | 3 | case NEON::BI__builtin_neon_vtbl4_v: |
5230 | 3 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4), |
5231 | 3 | Ops, "vtbl4"); |
5232 | 3 | case NEON::BI__builtin_neon_vtbx1_v: |
5233 | 3 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1), |
5234 | 3 | Ops, "vtbx1"); |
5235 | 3 | case NEON::BI__builtin_neon_vtbx2_v: |
5236 | 3 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2), |
5237 | 3 | Ops, "vtbx2"); |
5238 | 3 | case NEON::BI__builtin_neon_vtbx3_v: |
5239 | 3 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3), |
5240 | 3 | Ops, "vtbx3"); |
5241 | 3 | case NEON::BI__builtin_neon_vtbx4_v: |
5242 | 3 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4), |
5243 | 3 | Ops, "vtbx4"); |
5244 | 0 | } |
5245 | 0 | } |
5246 | | |
5247 | | static Value *EmitAArch64TblBuiltinExpr(CodeGenFunction &CGF, unsigned BuiltinID, |
5248 | | const CallExpr *E, |
5249 | 3.70k | SmallVectorImpl<Value *> &Ops) { |
5250 | 3.70k | unsigned int Int = 0; |
5251 | 3.70k | const char *s = nullptr; |
5252 | 3.70k | |
5253 | 3.70k | switch (BuiltinID) { |
5254 | 3.48k | default: |
5255 | 3.48k | return nullptr; |
5256 | 108 | case NEON::BI__builtin_neon_vtbl1_v: |
5257 | 108 | case NEON::BI__builtin_neon_vqtbl1_v: |
5258 | 108 | case NEON::BI__builtin_neon_vqtbl1q_v: |
5259 | 108 | case NEON::BI__builtin_neon_vtbl2_v: |
5260 | 108 | case NEON::BI__builtin_neon_vqtbl2_v: |
5261 | 108 | case NEON::BI__builtin_neon_vqtbl2q_v: |
5262 | 108 | case NEON::BI__builtin_neon_vtbl3_v: |
5263 | 108 | case NEON::BI__builtin_neon_vqtbl3_v: |
5264 | 108 | case NEON::BI__builtin_neon_vqtbl3q_v: |
5265 | 108 | case NEON::BI__builtin_neon_vtbl4_v: |
5266 | 108 | case NEON::BI__builtin_neon_vqtbl4_v: |
5267 | 108 | case NEON::BI__builtin_neon_vqtbl4q_v: |
5268 | 108 | break; |
5269 | 108 | case NEON::BI__builtin_neon_vtbx1_v: |
5270 | 108 | case NEON::BI__builtin_neon_vqtbx1_v: |
5271 | 108 | case NEON::BI__builtin_neon_vqtbx1q_v: |
5272 | 108 | case NEON::BI__builtin_neon_vtbx2_v: |
5273 | 108 | case NEON::BI__builtin_neon_vqtbx2_v: |
5274 | 108 | case NEON::BI__builtin_neon_vqtbx2q_v: |
5275 | 108 | case NEON::BI__builtin_neon_vtbx3_v: |
5276 | 108 | case NEON::BI__builtin_neon_vqtbx3_v: |
5277 | 108 | case NEON::BI__builtin_neon_vqtbx3q_v: |
5278 | 108 | case NEON::BI__builtin_neon_vtbx4_v: |
5279 | 108 | case NEON::BI__builtin_neon_vqtbx4_v: |
5280 | 108 | case NEON::BI__builtin_neon_vqtbx4q_v: |
5281 | 108 | break; |
5282 | 216 | } |
5283 | 216 | |
5284 | 3.70k | assert(E->getNumArgs() >= 3); |
5285 | 216 | |
5286 | 216 | // Get the last argument, which specifies the vector type. |
5287 | 216 | llvm::APSInt Result; |
5288 | 216 | const Expr *Arg = E->getArg(E->getNumArgs() - 1); |
5289 | 216 | if (!Arg->isIntegerConstantExpr(Result, CGF.getContext())) |
5290 | 0 | return nullptr; |
5291 | 216 | |
5292 | 216 | // Determine the type of this overloaded NEON intrinsic. |
5293 | 216 | NeonTypeFlags Type(Result.getZExtValue()); |
5294 | 216 | llvm::VectorType *Ty = GetNeonType(&CGF, Type); |
5295 | 216 | if (!Ty) |
5296 | 0 | return nullptr; |
5297 | 216 | |
5298 | 216 | CodeGen::CGBuilderTy &Builder = CGF.Builder; |
5299 | 216 | |
5300 | 216 | // AArch64 scalar builtins are not overloaded, they do not have an extra |
5301 | 216 | // argument that specifies the vector type, need to handle each case. |
5302 | 216 | switch (BuiltinID) { |
5303 | 9 | case NEON::BI__builtin_neon_vtbl1_v: { |
5304 | 9 | return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 1), nullptr, |
5305 | 9 | Ops[1], Ty, Intrinsic::aarch64_neon_tbl1, |
5306 | 9 | "vtbl1"); |
5307 | 216 | } |
5308 | 9 | case NEON::BI__builtin_neon_vtbl2_v: { |
5309 | 9 | return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 2), nullptr, |
5310 | 9 | Ops[2], Ty, Intrinsic::aarch64_neon_tbl1, |
5311 | 9 | "vtbl1"); |
5312 | 216 | } |
5313 | 9 | case NEON::BI__builtin_neon_vtbl3_v: { |
5314 | 9 | return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 3), nullptr, |
5315 | 9 | Ops[3], Ty, Intrinsic::aarch64_neon_tbl2, |
5316 | 9 | "vtbl2"); |
5317 | 216 | } |
5318 | 9 | case NEON::BI__builtin_neon_vtbl4_v: { |
5319 | 9 | return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 4), nullptr, |
5320 | 9 | Ops[4], Ty, Intrinsic::aarch64_neon_tbl2, |
5321 | 9 | "vtbl2"); |
5322 | 216 | } |
5323 | 9 | case NEON::BI__builtin_neon_vtbx1_v: { |
5324 | 9 | Value *TblRes = |
5325 | 9 | packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 1), nullptr, Ops[2], |
5326 | 9 | Ty, Intrinsic::aarch64_neon_tbl1, "vtbl1"); |
5327 | 9 | |
5328 | 9 | llvm::Constant *EightV = ConstantInt::get(Ty, 8); |
5329 | 9 | Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[2], EightV); |
5330 | 9 | CmpRes = Builder.CreateSExt(CmpRes, Ty); |
5331 | 9 | |
5332 | 9 | Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); |
5333 | 9 | Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); |
5334 | 9 | return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); |
5335 | 216 | } |
5336 | 9 | case NEON::BI__builtin_neon_vtbx2_v: { |
5337 | 9 | return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 2), Ops[0], |
5338 | 9 | Ops[3], Ty, Intrinsic::aarch64_neon_tbx1, |
5339 | 9 | "vtbx1"); |
5340 | 216 | } |
5341 | 9 | case NEON::BI__builtin_neon_vtbx3_v: { |
5342 | 9 | Value *TblRes = |
5343 | 9 | packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 3), nullptr, Ops[4], |
5344 | 9 | Ty, Intrinsic::aarch64_neon_tbl2, "vtbl2"); |
5345 | 9 | |
5346 | 9 | llvm::Constant *TwentyFourV = ConstantInt::get(Ty, 24); |
5347 | 9 | Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[4], |
5348 | 9 | TwentyFourV); |
5349 | 9 | CmpRes = Builder.CreateSExt(CmpRes, Ty); |
5350 | 9 | |
5351 | 9 | Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); |
5352 | 9 | Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); |
5353 | 9 | return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); |
5354 | 216 | } |
5355 | 9 | case NEON::BI__builtin_neon_vtbx4_v: { |
5356 | 9 | return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 4), Ops[0], |
5357 | 9 | Ops[5], Ty, Intrinsic::aarch64_neon_tbx2, |
5358 | 9 | "vtbx2"); |
5359 | 216 | } |
5360 | 18 | case NEON::BI__builtin_neon_vqtbl1_v: |
5361 | 18 | case NEON::BI__builtin_neon_vqtbl1q_v: |
5362 | 18 | Int = Intrinsic::aarch64_neon_tbl1; s = "vtbl1"; break; |
5363 | 18 | case NEON::BI__builtin_neon_vqtbl2_v: |
5364 | 18 | case NEON::BI__builtin_neon_vqtbl2q_v: { |
5365 | 18 | Int = Intrinsic::aarch64_neon_tbl2; s = "vtbl2"; break; |
5366 | 18 | case NEON::BI__builtin_neon_vqtbl3_v: |
5367 | 18 | case NEON::BI__builtin_neon_vqtbl3q_v: |
5368 | 18 | Int = Intrinsic::aarch64_neon_tbl3; s = "vtbl3"; break; |
5369 | 18 | case NEON::BI__builtin_neon_vqtbl4_v: |
5370 | 18 | case NEON::BI__builtin_neon_vqtbl4q_v: |
5371 | 18 | Int = Intrinsic::aarch64_neon_tbl4; s = "vtbl4"; break; |
5372 | 18 | case NEON::BI__builtin_neon_vqtbx1_v: |
5373 | 18 | case NEON::BI__builtin_neon_vqtbx1q_v: |
5374 | 18 | Int = Intrinsic::aarch64_neon_tbx1; s = "vtbx1"; break; |
5375 | 18 | case NEON::BI__builtin_neon_vqtbx2_v: |
5376 | 18 | case NEON::BI__builtin_neon_vqtbx2q_v: |
5377 | 18 | Int = Intrinsic::aarch64_neon_tbx2; s = "vtbx2"; break; |
5378 | 18 | case NEON::BI__builtin_neon_vqtbx3_v: |
5379 | 18 | case NEON::BI__builtin_neon_vqtbx3q_v: |
5380 | 18 | Int = Intrinsic::aarch64_neon_tbx3; s = "vtbx3"; break; |
5381 | 18 | case NEON::BI__builtin_neon_vqtbx4_v: |
5382 | 18 | case NEON::BI__builtin_neon_vqtbx4q_v: |
5383 | 18 | Int = Intrinsic::aarch64_neon_tbx4; s = "vtbx4"; break; |
5384 | 144 | } |
5385 | 144 | } |
5386 | 144 | |
5387 | 144 | if (144 !Int144 ) |
5388 | 0 | return nullptr; |
5389 | 144 | |
5390 | 144 | Function *F = CGF.CGM.getIntrinsic(Int, Ty); |
5391 | 144 | return CGF.EmitNeonCall(F, Ops, s); |
5392 | 144 | } |
5393 | | |
5394 | 36 | Value *CodeGenFunction::vectorWrapScalar16(Value *Op) { |
5395 | 36 | llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4); |
5396 | 36 | Op = Builder.CreateBitCast(Op, Int16Ty); |
5397 | 36 | Value *V = UndefValue::get(VTy); |
5398 | 36 | llvm::Constant *CI = ConstantInt::get(SizeTy, 0); |
5399 | 36 | Op = Builder.CreateInsertElement(V, Op, CI); |
5400 | 36 | return Op; |
5401 | 36 | } |
5402 | | |
5403 | | Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID, |
5404 | 17.3k | const CallExpr *E) { |
5405 | 17.3k | unsigned HintID = static_cast<unsigned>(-1); |
5406 | 17.3k | switch (BuiltinID) { |
5407 | 17.3k | default: break; |
5408 | 2 | case AArch64::BI__builtin_arm_nop: |
5409 | 2 | HintID = 0; |
5410 | 2 | break; |
5411 | 2 | case AArch64::BI__builtin_arm_yield: |
5412 | 2 | HintID = 1; |
5413 | 2 | break; |
5414 | 2 | case AArch64::BI__builtin_arm_wfe: |
5415 | 2 | HintID = 2; |
5416 | 2 | break; |
5417 | 2 | case AArch64::BI__builtin_arm_wfi: |
5418 | 2 | HintID = 3; |
5419 | 2 | break; |
5420 | 2 | case AArch64::BI__builtin_arm_sev: |
5421 | 2 | HintID = 4; |
5422 | 2 | break; |
5423 | 2 | case AArch64::BI__builtin_arm_sevl: |
5424 | 2 | HintID = 5; |
5425 | 2 | break; |
5426 | 17.3k | } |
5427 | 17.3k | |
5428 | 17.3k | if (17.3k HintID != static_cast<unsigned>(-1)17.3k ) { |
5429 | 12 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_hint); |
5430 | 12 | return Builder.CreateCall(F, llvm::ConstantInt::get(Int32Ty, HintID)); |
5431 | 12 | } |
5432 | 17.3k | |
5433 | 17.3k | if (17.3k BuiltinID == AArch64::BI__builtin_arm_prefetch17.3k ) { |
5434 | 8 | Value *Address = EmitScalarExpr(E->getArg(0)); |
5435 | 8 | Value *RW = EmitScalarExpr(E->getArg(1)); |
5436 | 8 | Value *CacheLevel = EmitScalarExpr(E->getArg(2)); |
5437 | 8 | Value *RetentionPolicy = EmitScalarExpr(E->getArg(3)); |
5438 | 8 | Value *IsData = EmitScalarExpr(E->getArg(4)); |
5439 | 8 | |
5440 | 8 | Value *Locality = nullptr; |
5441 | 8 | if (cast<llvm::ConstantInt>(RetentionPolicy)->isZero()8 ) { |
5442 | 6 | // Temporal fetch, needs to convert cache level to locality. |
5443 | 6 | Locality = llvm::ConstantInt::get(Int32Ty, |
5444 | 6 | -cast<llvm::ConstantInt>(CacheLevel)->getValue() + 3); |
5445 | 8 | } else { |
5446 | 2 | // Streaming fetch. |
5447 | 2 | Locality = llvm::ConstantInt::get(Int32Ty, 0); |
5448 | 2 | } |
5449 | 8 | |
5450 | 8 | // FIXME: We need AArch64 specific LLVM intrinsic if we want to specify |
5451 | 8 | // PLDL3STRM or PLDL2STRM. |
5452 | 8 | Value *F = CGM.getIntrinsic(Intrinsic::prefetch); |
5453 | 8 | return Builder.CreateCall(F, {Address, RW, Locality, IsData}); |
5454 | 8 | } |
5455 | 17.3k | |
5456 | 17.3k | if (17.3k BuiltinID == AArch64::BI__builtin_arm_rbit17.3k ) { |
5457 | 2 | assert((getContext().getTypeSize(E->getType()) == 32) && |
5458 | 2 | "rbit of unusual size!"); |
5459 | 2 | llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); |
5460 | 2 | return Builder.CreateCall( |
5461 | 2 | CGM.getIntrinsic(Intrinsic::bitreverse, Arg->getType()), Arg, "rbit"); |
5462 | 2 | } |
5463 | 17.3k | if (17.3k BuiltinID == AArch64::BI__builtin_arm_rbit6417.3k ) { |
5464 | 2 | assert((getContext().getTypeSize(E->getType()) == 64) && |
5465 | 2 | "rbit of unusual size!"); |
5466 | 2 | llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); |
5467 | 2 | return Builder.CreateCall( |
5468 | 2 | CGM.getIntrinsic(Intrinsic::bitreverse, Arg->getType()), Arg, "rbit"); |
5469 | 2 | } |
5470 | 17.3k | |
5471 | 17.3k | if (17.3k BuiltinID == AArch64::BI__clear_cache17.3k ) { |
5472 | 1 | assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); |
5473 | 1 | const FunctionDecl *FD = E->getDirectCallee(); |
5474 | 1 | Value *Ops[2]; |
5475 | 3 | for (unsigned i = 0; i < 23 ; i++2 ) |
5476 | 2 | Ops[i] = EmitScalarExpr(E->getArg(i)); |
5477 | 1 | llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); |
5478 | 1 | llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); |
5479 | 1 | StringRef Name = FD->getName(); |
5480 | 1 | return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); |
5481 | 1 | } |
5482 | 17.3k | |
5483 | 17.3k | if (17.3k (BuiltinID == AArch64::BI__builtin_arm_ldrex || |
5484 | 17.3k | BuiltinID == AArch64::BI__builtin_arm_ldaex) && |
5485 | 17.3k | getContext().getTypeSize(E->getType()) == 12822 ) { |
5486 | 2 | Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex |
5487 | 1 | ? Intrinsic::aarch64_ldaxp |
5488 | 1 | : Intrinsic::aarch64_ldxp); |
5489 | 2 | |
5490 | 2 | Value *LdPtr = EmitScalarExpr(E->getArg(0)); |
5491 | 2 | Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), |
5492 | 2 | "ldxp"); |
5493 | 2 | |
5494 | 2 | Value *Val0 = Builder.CreateExtractValue(Val, 1); |
5495 | 2 | Value *Val1 = Builder.CreateExtractValue(Val, 0); |
5496 | 2 | llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128); |
5497 | 2 | Val0 = Builder.CreateZExt(Val0, Int128Ty); |
5498 | 2 | Val1 = Builder.CreateZExt(Val1, Int128Ty); |
5499 | 2 | |
5500 | 2 | Value *ShiftCst = llvm::ConstantInt::get(Int128Ty, 64); |
5501 | 2 | Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); |
5502 | 2 | Val = Builder.CreateOr(Val, Val1); |
5503 | 2 | return Builder.CreateBitCast(Val, ConvertType(E->getType())); |
5504 | 17.3k | } else if (17.3k BuiltinID == AArch64::BI__builtin_arm_ldrex || |
5505 | 17.3k | BuiltinID == AArch64::BI__builtin_arm_ldaex17.3k ) { |
5506 | 20 | Value *LoadAddr = EmitScalarExpr(E->getArg(0)); |
5507 | 20 | |
5508 | 20 | QualType Ty = E->getType(); |
5509 | 20 | llvm::Type *RealResTy = ConvertType(Ty); |
5510 | 20 | llvm::Type *PtrTy = llvm::IntegerType::get( |
5511 | 20 | getLLVMContext(), getContext().getTypeSize(Ty))->getPointerTo(); |
5512 | 20 | LoadAddr = Builder.CreateBitCast(LoadAddr, PtrTy); |
5513 | 20 | |
5514 | 20 | Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex |
5515 | 9 | ? Intrinsic::aarch64_ldaxr |
5516 | 11 | : Intrinsic::aarch64_ldxr, |
5517 | 20 | PtrTy); |
5518 | 20 | Value *Val = Builder.CreateCall(F, LoadAddr, "ldxr"); |
5519 | 20 | |
5520 | 20 | if (RealResTy->isPointerTy()) |
5521 | 4 | return Builder.CreateIntToPtr(Val, RealResTy); |
5522 | 16 | |
5523 | 16 | llvm::Type *IntResTy = llvm::IntegerType::get( |
5524 | 16 | getLLVMContext(), CGM.getDataLayout().getTypeSizeInBits(RealResTy)); |
5525 | 16 | Val = Builder.CreateTruncOrBitCast(Val, IntResTy); |
5526 | 16 | return Builder.CreateBitCast(Val, RealResTy); |
5527 | 16 | } |
5528 | 17.2k | |
5529 | 17.2k | if (17.2k (BuiltinID == AArch64::BI__builtin_arm_strex || |
5530 | 17.2k | BuiltinID == AArch64::BI__builtin_arm_stlex) && |
5531 | 17.2k | getContext().getTypeSize(E->getArg(0)->getType()) == 12818 ) { |
5532 | 2 | Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex |
5533 | 1 | ? Intrinsic::aarch64_stlxp |
5534 | 1 | : Intrinsic::aarch64_stxp); |
5535 | 2 | llvm::Type *STy = llvm::StructType::get(Int64Ty, Int64Ty); |
5536 | 2 | |
5537 | 2 | Address Tmp = CreateMemTemp(E->getArg(0)->getType()); |
5538 | 2 | EmitAnyExprToMem(E->getArg(0), Tmp, Qualifiers(), /*init*/ true); |
5539 | 2 | |
5540 | 2 | Tmp = Builder.CreateBitCast(Tmp, llvm::PointerType::getUnqual(STy)); |
5541 | 2 | llvm::Value *Val = Builder.CreateLoad(Tmp); |
5542 | 2 | |
5543 | 2 | Value *Arg0 = Builder.CreateExtractValue(Val, 0); |
5544 | 2 | Value *Arg1 = Builder.CreateExtractValue(Val, 1); |
5545 | 2 | Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), |
5546 | 2 | Int8PtrTy); |
5547 | 2 | return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "stxp"); |
5548 | 2 | } |
5549 | 17.2k | |
5550 | 17.2k | if (17.2k BuiltinID == AArch64::BI__builtin_arm_strex || |
5551 | 17.2k | BuiltinID == AArch64::BI__builtin_arm_stlex17.2k ) { |
5552 | 16 | Value *StoreVal = EmitScalarExpr(E->getArg(0)); |
5553 | 16 | Value *StoreAddr = EmitScalarExpr(E->getArg(1)); |
5554 | 16 | |
5555 | 16 | QualType Ty = E->getArg(0)->getType(); |
5556 | 16 | llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), |
5557 | 16 | getContext().getTypeSize(Ty)); |
5558 | 16 | StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); |
5559 | 16 | |
5560 | 16 | if (StoreVal->getType()->isPointerTy()) |
5561 | 2 | StoreVal = Builder.CreatePtrToInt(StoreVal, Int64Ty); |
5562 | 14 | else { |
5563 | 14 | llvm::Type *IntTy = llvm::IntegerType::get( |
5564 | 14 | getLLVMContext(), |
5565 | 14 | CGM.getDataLayout().getTypeSizeInBits(StoreVal->getType())); |
5566 | 14 | StoreVal = Builder.CreateBitCast(StoreVal, IntTy); |
5567 | 14 | StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int64Ty); |
5568 | 14 | } |
5569 | 16 | |
5570 | 16 | Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex |
5571 | 7 | ? Intrinsic::aarch64_stlxr |
5572 | 9 | : Intrinsic::aarch64_stxr, |
5573 | 16 | StoreAddr->getType()); |
5574 | 16 | return Builder.CreateCall(F, {StoreVal, StoreAddr}, "stxr"); |
5575 | 16 | } |
5576 | 17.2k | |
5577 | 17.2k | if (17.2k BuiltinID == AArch64::BI__builtin_arm_clrex17.2k ) { |
5578 | 1 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_clrex); |
5579 | 1 | return Builder.CreateCall(F); |
5580 | 1 | } |
5581 | 17.2k | |
5582 | 17.2k | // CRC32 |
5583 | 17.2k | Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; |
5584 | 17.2k | switch (BuiltinID) { |
5585 | 2 | case AArch64::BI__builtin_arm_crc32b: |
5586 | 2 | CRCIntrinsicID = Intrinsic::aarch64_crc32b; break; |
5587 | 2 | case AArch64::BI__builtin_arm_crc32cb: |
5588 | 2 | CRCIntrinsicID = Intrinsic::aarch64_crc32cb; break; |
5589 | 2 | case AArch64::BI__builtin_arm_crc32h: |
5590 | 2 | CRCIntrinsicID = Intrinsic::aarch64_crc32h; break; |
5591 | 2 | case AArch64::BI__builtin_arm_crc32ch: |
5592 | 2 | CRCIntrinsicID = Intrinsic::aarch64_crc32ch; break; |
5593 | 2 | case AArch64::BI__builtin_arm_crc32w: |
5594 | 2 | CRCIntrinsicID = Intrinsic::aarch64_crc32w; break; |
5595 | 2 | case AArch64::BI__builtin_arm_crc32cw: |
5596 | 2 | CRCIntrinsicID = Intrinsic::aarch64_crc32cw; break; |
5597 | 2 | case AArch64::BI__builtin_arm_crc32d: |
5598 | 2 | CRCIntrinsicID = Intrinsic::aarch64_crc32x; break; |
5599 | 2 | case AArch64::BI__builtin_arm_crc32cd: |
5600 | 2 | CRCIntrinsicID = Intrinsic::aarch64_crc32cx; break; |
5601 | 17.2k | } |
5602 | 17.2k | |
5603 | 17.2k | if (17.2k CRCIntrinsicID != Intrinsic::not_intrinsic17.2k ) { |
5604 | 16 | Value *Arg0 = EmitScalarExpr(E->getArg(0)); |
5605 | 16 | Value *Arg1 = EmitScalarExpr(E->getArg(1)); |
5606 | 16 | Function *F = CGM.getIntrinsic(CRCIntrinsicID); |
5607 | 16 | |
5608 | 16 | llvm::Type *DataTy = F->getFunctionType()->getParamType(1); |
5609 | 16 | Arg1 = Builder.CreateZExtOrBitCast(Arg1, DataTy); |
5610 | 16 | |
5611 | 16 | return Builder.CreateCall(F, {Arg0, Arg1}); |
5612 | 16 | } |
5613 | 17.2k | |
5614 | 17.2k | if (17.2k BuiltinID == AArch64::BI__builtin_arm_rsr || |
5615 | 17.2k | BuiltinID == AArch64::BI__builtin_arm_rsr64 || |
5616 | 17.2k | BuiltinID == AArch64::BI__builtin_arm_rsrp || |
5617 | 17.2k | BuiltinID == AArch64::BI__builtin_arm_wsr || |
5618 | 17.2k | BuiltinID == AArch64::BI__builtin_arm_wsr64 || |
5619 | 17.2k | BuiltinID == AArch64::BI__builtin_arm_wsrp17.2k ) { |
5620 | 12 | |
5621 | 12 | bool IsRead = BuiltinID == AArch64::BI__builtin_arm_rsr || |
5622 | 10 | BuiltinID == AArch64::BI__builtin_arm_rsr64 || |
5623 | 8 | BuiltinID == AArch64::BI__builtin_arm_rsrp; |
5624 | 12 | |
5625 | 12 | bool IsPointerBuiltin = BuiltinID == AArch64::BI__builtin_arm_rsrp || |
5626 | 10 | BuiltinID == AArch64::BI__builtin_arm_wsrp; |
5627 | 12 | |
5628 | 12 | bool Is64Bit = BuiltinID != AArch64::BI__builtin_arm_rsr && |
5629 | 10 | BuiltinID != AArch64::BI__builtin_arm_wsr; |
5630 | 12 | |
5631 | 12 | llvm::Type *ValueType; |
5632 | 12 | llvm::Type *RegisterType = Int64Ty; |
5633 | 12 | if (IsPointerBuiltin12 ) { |
5634 | 4 | ValueType = VoidPtrTy; |
5635 | 12 | } else if (8 Is64Bit8 ) { |
5636 | 4 | ValueType = Int64Ty; |
5637 | 8 | } else { |
5638 | 4 | ValueType = Int32Ty; |
5639 | 4 | } |
5640 | 12 | |
5641 | 12 | return EmitSpecialRegisterBuiltin(*this, E, RegisterType, ValueType, IsRead); |
5642 | 12 | } |
5643 | 17.2k | |
5644 | 17.2k | // Find out if any arguments are required to be integer constant |
5645 | 17.2k | // expressions. |
5646 | 17.2k | unsigned ICEArguments = 0; |
5647 | 17.2k | ASTContext::GetBuiltinTypeError Error; |
5648 | 17.2k | getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); |
5649 | 17.2k | assert(Error == ASTContext::GE_None && "Should not codegen an error"); |
5650 | 17.2k | |
5651 | 17.2k | llvm::SmallVector<Value*, 4> Ops; |
5652 | 42.2k | for (unsigned i = 0, e = E->getNumArgs() - 1; i != e42.2k ; i++25.0k ) { |
5653 | 25.0k | if ((ICEArguments & (1 << i)) == 025.0k ) { |
5654 | 23.4k | Ops.push_back(EmitScalarExpr(E->getArg(i))); |
5655 | 25.0k | } else { |
5656 | 1.51k | // If this is required to be a constant, constant fold it so that we know |
5657 | 1.51k | // that the generated intrinsic gets a ConstantInt. |
5658 | 1.51k | llvm::APSInt Result; |
5659 | 1.51k | bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); |
5660 | 1.51k | assert(IsConst && "Constant arg isn't actually constant?"); |
5661 | 1.51k | (void)IsConst; |
5662 | 1.51k | Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); |
5663 | 1.51k | } |
5664 | 25.0k | } |
5665 | 17.2k | |
5666 | 17.2k | auto SISDMap = makeArrayRef(AArch64SISDIntrinsicMap); |
5667 | 17.2k | const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( |
5668 | 17.2k | SISDMap, BuiltinID, AArch64SISDIntrinsicsProvenSorted); |
5669 | 17.2k | |
5670 | 17.2k | if (Builtin17.2k ) { |
5671 | 582 | Ops.push_back(EmitScalarExpr(E->getArg(E->getNumArgs() - 1))); |
5672 | 582 | Value *Result = EmitCommonNeonSISDBuiltinExpr(*this, *Builtin, Ops, E); |
5673 | 582 | assert(Result && "SISD intrinsic should have been handled"); |
5674 | 582 | return Result; |
5675 | 582 | } |
5676 | 16.6k | |
5677 | 16.6k | llvm::APSInt Result; |
5678 | 16.6k | const Expr *Arg = E->getArg(E->getNumArgs()-1); |
5679 | 16.6k | NeonTypeFlags Type(0); |
5680 | 16.6k | if (Arg->isIntegerConstantExpr(Result, getContext())) |
5681 | 16.6k | // Determine the type of this overloaded NEON intrinsic. |
5682 | 16.3k | Type = NeonTypeFlags(Result.getZExtValue()); |
5683 | 16.6k | |
5684 | 16.6k | bool usgn = Type.isUnsigned(); |
5685 | 16.6k | bool quad = Type.isQuad(); |
5686 | 16.6k | |
5687 | 16.6k | // Handle non-overloaded intrinsics first. |
5688 | 16.6k | switch (BuiltinID) { |
5689 | 5.28k | default: break; |
5690 | 2 | case NEON::BI__builtin_neon_vldrq_p128: { |
5691 | 2 | llvm::Type *Int128Ty = llvm::Type::getIntNTy(getLLVMContext(), 128); |
5692 | 2 | llvm::Type *Int128PTy = llvm::PointerType::get(Int128Ty, 0); |
5693 | 2 | Value *Ptr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int128PTy); |
5694 | 2 | return Builder.CreateAlignedLoad(Int128Ty, Ptr, |
5695 | 2 | CharUnits::fromQuantity(16)); |
5696 | 16.6k | } |
5697 | 2 | case NEON::BI__builtin_neon_vstrq_p128: { |
5698 | 2 | llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); |
5699 | 2 | Value *Ptr = Builder.CreateBitCast(Ops[0], Int128PTy); |
5700 | 2 | return Builder.CreateDefaultAlignedStore(EmitScalarExpr(E->getArg(1)), Ptr); |
5701 | 16.6k | } |
5702 | 6 | case NEON::BI__builtin_neon_vcvts_u32_f32: |
5703 | 6 | case NEON::BI__builtin_neon_vcvtd_u64_f64: |
5704 | 6 | usgn = true; |
5705 | 6 | // FALL THROUGH |
5706 | 12 | case NEON::BI__builtin_neon_vcvts_s32_f32: |
5707 | 12 | case NEON::BI__builtin_neon_vcvtd_s64_f64: { |
5708 | 12 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
5709 | 12 | bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; |
5710 | 12 | llvm::Type *InTy = Is64 ? Int64Ty6 : Int32Ty6 ; |
5711 | 12 | llvm::Type *FTy = Is64 ? DoubleTy6 : FloatTy6 ; |
5712 | 12 | Ops[0] = Builder.CreateBitCast(Ops[0], FTy); |
5713 | 12 | if (usgn) |
5714 | 6 | return Builder.CreateFPToUI(Ops[0], InTy); |
5715 | 6 | return Builder.CreateFPToSI(Ops[0], InTy); |
5716 | 6 | } |
5717 | 6 | case NEON::BI__builtin_neon_vcvts_f32_u32: |
5718 | 6 | case NEON::BI__builtin_neon_vcvtd_f64_u64: |
5719 | 6 | usgn = true; |
5720 | 6 | // FALL THROUGH |
5721 | 12 | case NEON::BI__builtin_neon_vcvts_f32_s32: |
5722 | 12 | case NEON::BI__builtin_neon_vcvtd_f64_s64: { |
5723 | 12 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
5724 | 12 | bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; |
5725 | 12 | llvm::Type *InTy = Is64 ? Int64Ty6 : Int32Ty6 ; |
5726 | 12 | llvm::Type *FTy = Is64 ? DoubleTy6 : FloatTy6 ; |
5727 | 12 | Ops[0] = Builder.CreateBitCast(Ops[0], InTy); |
5728 | 12 | if (usgn) |
5729 | 6 | return Builder.CreateUIToFP(Ops[0], FTy); |
5730 | 6 | return Builder.CreateSIToFP(Ops[0], FTy); |
5731 | 6 | } |
5732 | 0 | case NEON::BI__builtin_neon_vpaddd_s64: { |
5733 | 0 | llvm::Type *Ty = llvm::VectorType::get(Int64Ty, 2); |
5734 | 0 | Value *Vec = EmitScalarExpr(E->getArg(0)); |
5735 | 0 | // The vector is v2f64, so make sure it's bitcast to that. |
5736 | 0 | Vec = Builder.CreateBitCast(Vec, Ty, "v2i64"); |
5737 | 0 | llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); |
5738 | 0 | llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); |
5739 | 0 | Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); |
5740 | 0 | Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); |
5741 | 0 | // Pairwise addition of a v2f64 into a scalar f64. |
5742 | 0 | return Builder.CreateAdd(Op0, Op1, "vpaddd"); |
5743 | 6 | } |
5744 | 3 | case NEON::BI__builtin_neon_vpaddd_f64: { |
5745 | 3 | llvm::Type *Ty = |
5746 | 3 | llvm::VectorType::get(DoubleTy, 2); |
5747 | 3 | Value *Vec = EmitScalarExpr(E->getArg(0)); |
5748 | 3 | // The vector is v2f64, so make sure it's bitcast to that. |
5749 | 3 | Vec = Builder.CreateBitCast(Vec, Ty, "v2f64"); |
5750 | 3 | llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); |
5751 | 3 | llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); |
5752 | 3 | Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); |
5753 | 3 | Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); |
5754 | 3 | // Pairwise addition of a v2f64 into a scalar f64. |
5755 | 3 | return Builder.CreateFAdd(Op0, Op1, "vpaddd"); |
5756 | 6 | } |
5757 | 3 | case NEON::BI__builtin_neon_vpadds_f32: { |
5758 | 3 | llvm::Type *Ty = |
5759 | 3 | llvm::VectorType::get(FloatTy, 2); |
5760 | 3 | Value *Vec = EmitScalarExpr(E->getArg(0)); |
5761 | 3 | // The vector is v2f32, so make sure it's bitcast to that. |
5762 | 3 | Vec = Builder.CreateBitCast(Vec, Ty, "v2f32"); |
5763 | 3 | llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); |
5764 | 3 | llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); |
5765 | 3 | Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); |
5766 | 3 | Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); |
5767 | 3 | // Pairwise addition of a v2f32 into a scalar f32. |
5768 | 3 | return Builder.CreateFAdd(Op0, Op1, "vpaddd"); |
5769 | 6 | } |
5770 | 9 | case NEON::BI__builtin_neon_vceqzd_s64: |
5771 | 9 | case NEON::BI__builtin_neon_vceqzd_f64: |
5772 | 9 | case NEON::BI__builtin_neon_vceqzs_f32: |
5773 | 9 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
5774 | 9 | return EmitAArch64CompareBuiltinExpr( |
5775 | 9 | Ops[0], ConvertType(E->getCallReturnType(getContext())), |
5776 | 9 | ICmpInst::FCMP_OEQ, ICmpInst::ICMP_EQ, "vceqz"); |
5777 | 9 | case NEON::BI__builtin_neon_vcgezd_s64: |
5778 | 9 | case NEON::BI__builtin_neon_vcgezd_f64: |
5779 | 9 | case NEON::BI__builtin_neon_vcgezs_f32: |
5780 | 9 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
5781 | 9 | return EmitAArch64CompareBuiltinExpr( |
5782 | 9 | Ops[0], ConvertType(E->getCallReturnType(getContext())), |
5783 | 9 | ICmpInst::FCMP_OGE, ICmpInst::ICMP_SGE, "vcgez"); |
5784 | 9 | case NEON::BI__builtin_neon_vclezd_s64: |
5785 | 9 | case NEON::BI__builtin_neon_vclezd_f64: |
5786 | 9 | case NEON::BI__builtin_neon_vclezs_f32: |
5787 | 9 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
5788 | 9 | return EmitAArch64CompareBuiltinExpr( |
5789 | 9 | Ops[0], ConvertType(E->getCallReturnType(getContext())), |
5790 | 9 | ICmpInst::FCMP_OLE, ICmpInst::ICMP_SLE, "vclez"); |
5791 | 9 | case NEON::BI__builtin_neon_vcgtzd_s64: |
5792 | 9 | case NEON::BI__builtin_neon_vcgtzd_f64: |
5793 | 9 | case NEON::BI__builtin_neon_vcgtzs_f32: |
5794 | 9 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
5795 | 9 | return EmitAArch64CompareBuiltinExpr( |
5796 | 9 | Ops[0], ConvertType(E->getCallReturnType(getContext())), |
5797 | 9 | ICmpInst::FCMP_OGT, ICmpInst::ICMP_SGT, "vcgtz"); |
5798 | 9 | case NEON::BI__builtin_neon_vcltzd_s64: |
5799 | 9 | case NEON::BI__builtin_neon_vcltzd_f64: |
5800 | 9 | case NEON::BI__builtin_neon_vcltzs_f32: |
5801 | 9 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
5802 | 9 | return EmitAArch64CompareBuiltinExpr( |
5803 | 9 | Ops[0], ConvertType(E->getCallReturnType(getContext())), |
5804 | 9 | ICmpInst::FCMP_OLT, ICmpInst::ICMP_SLT, "vcltz"); |
5805 | 9 | |
5806 | 3 | case NEON::BI__builtin_neon_vceqzd_u64: { |
5807 | 3 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
5808 | 3 | Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); |
5809 | 3 | Ops[0] = |
5810 | 3 | Builder.CreateICmpEQ(Ops[0], llvm::Constant::getNullValue(Int64Ty)); |
5811 | 3 | return Builder.CreateSExt(Ops[0], Int64Ty, "vceqzd"); |
5812 | 9 | } |
5813 | 15 | case NEON::BI__builtin_neon_vceqd_f64: |
5814 | 15 | case NEON::BI__builtin_neon_vcled_f64: |
5815 | 15 | case NEON::BI__builtin_neon_vcltd_f64: |
5816 | 15 | case NEON::BI__builtin_neon_vcged_f64: |
5817 | 15 | case NEON::BI__builtin_neon_vcgtd_f64: { |
5818 | 15 | llvm::CmpInst::Predicate P; |
5819 | 15 | switch (BuiltinID) { |
5820 | 0 | default: 0 llvm_unreachable0 ("missing builtin ID in switch!"); |
5821 | 3 | case NEON::BI__builtin_neon_vceqd_f64: P = llvm::FCmpInst::FCMP_OEQ; break; |
5822 | 3 | case NEON::BI__builtin_neon_vcled_f64: P = llvm::FCmpInst::FCMP_OLE; break; |
5823 | 3 | case NEON::BI__builtin_neon_vcltd_f64: P = llvm::FCmpInst::FCMP_OLT; break; |
5824 | 3 | case NEON::BI__builtin_neon_vcged_f64: P = llvm::FCmpInst::FCMP_OGE; break; |
5825 | 3 | case NEON::BI__builtin_neon_vcgtd_f64: P = llvm::FCmpInst::FCMP_OGT; break; |
5826 | 15 | } |
5827 | 15 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
5828 | 15 | Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); |
5829 | 15 | Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); |
5830 | 15 | Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); |
5831 | 15 | return Builder.CreateSExt(Ops[0], Int64Ty, "vcmpd"); |
5832 | 15 | } |
5833 | 15 | case NEON::BI__builtin_neon_vceqs_f32: |
5834 | 15 | case NEON::BI__builtin_neon_vcles_f32: |
5835 | 15 | case NEON::BI__builtin_neon_vclts_f32: |
5836 | 15 | case NEON::BI__builtin_neon_vcges_f32: |
5837 | 15 | case NEON::BI__builtin_neon_vcgts_f32: { |
5838 | 15 | llvm::CmpInst::Predicate P; |
5839 | 15 | switch (BuiltinID) { |
5840 | 0 | default: 0 llvm_unreachable0 ("missing builtin ID in switch!"); |
5841 | 3 | case NEON::BI__builtin_neon_vceqs_f32: P = llvm::FCmpInst::FCMP_OEQ; break; |
5842 | 3 | case NEON::BI__builtin_neon_vcles_f32: P = llvm::FCmpInst::FCMP_OLE; break; |
5843 | 3 | case NEON::BI__builtin_neon_vclts_f32: P = llvm::FCmpInst::FCMP_OLT; break; |
5844 | 3 | case NEON::BI__builtin_neon_vcges_f32: P = llvm::FCmpInst::FCMP_OGE; break; |
5845 | 3 | case NEON::BI__builtin_neon_vcgts_f32: P = llvm::FCmpInst::FCMP_OGT; break; |
5846 | 15 | } |
5847 | 15 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
5848 | 15 | Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); |
5849 | 15 | Ops[1] = Builder.CreateBitCast(Ops[1], FloatTy); |
5850 | 15 | Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); |
5851 | 15 | return Builder.CreateSExt(Ops[0], Int32Ty, "vcmpd"); |
5852 | 15 | } |
5853 | 30 | case NEON::BI__builtin_neon_vceqd_s64: |
5854 | 30 | case NEON::BI__builtin_neon_vceqd_u64: |
5855 | 30 | case NEON::BI__builtin_neon_vcgtd_s64: |
5856 | 30 | case NEON::BI__builtin_neon_vcgtd_u64: |
5857 | 30 | case NEON::BI__builtin_neon_vcltd_s64: |
5858 | 30 | case NEON::BI__builtin_neon_vcltd_u64: |
5859 | 30 | case NEON::BI__builtin_neon_vcged_u64: |
5860 | 30 | case NEON::BI__builtin_neon_vcged_s64: |
5861 | 30 | case NEON::BI__builtin_neon_vcled_u64: |
5862 | 30 | case NEON::BI__builtin_neon_vcled_s64: { |
5863 | 30 | llvm::CmpInst::Predicate P; |
5864 | 30 | switch (BuiltinID) { |
5865 | 0 | default: 0 llvm_unreachable0 ("missing builtin ID in switch!"); |
5866 | 6 | case NEON::BI__builtin_neon_vceqd_s64: |
5867 | 6 | case NEON::BI__builtin_neon_vceqd_u64:P = llvm::ICmpInst::ICMP_EQ;break; |
5868 | 3 | case NEON::BI__builtin_neon_vcgtd_s64:P = llvm::ICmpInst::ICMP_SGT;break; |
5869 | 3 | case NEON::BI__builtin_neon_vcgtd_u64:P = llvm::ICmpInst::ICMP_UGT;break; |
5870 | 3 | case NEON::BI__builtin_neon_vcltd_s64:P = llvm::ICmpInst::ICMP_SLT;break; |
5871 | 3 | case NEON::BI__builtin_neon_vcltd_u64:P = llvm::ICmpInst::ICMP_ULT;break; |
5872 | 3 | case NEON::BI__builtin_neon_vcged_u64:P = llvm::ICmpInst::ICMP_UGE;break; |
5873 | 3 | case NEON::BI__builtin_neon_vcged_s64:P = llvm::ICmpInst::ICMP_SGE;break; |
5874 | 3 | case NEON::BI__builtin_neon_vcled_u64:P = llvm::ICmpInst::ICMP_ULE;break; |
5875 | 3 | case NEON::BI__builtin_neon_vcled_s64:P = llvm::ICmpInst::ICMP_SLE;break; |
5876 | 30 | } |
5877 | 30 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
5878 | 30 | Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); |
5879 | 30 | Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); |
5880 | 30 | Ops[0] = Builder.CreateICmp(P, Ops[0], Ops[1]); |
5881 | 30 | return Builder.CreateSExt(Ops[0], Int64Ty, "vceqd"); |
5882 | 30 | } |
5883 | 6 | case NEON::BI__builtin_neon_vtstd_s64: |
5884 | 6 | case NEON::BI__builtin_neon_vtstd_u64: { |
5885 | 6 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
5886 | 6 | Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); |
5887 | 6 | Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); |
5888 | 6 | Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); |
5889 | 6 | Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], |
5890 | 6 | llvm::Constant::getNullValue(Int64Ty)); |
5891 | 6 | return Builder.CreateSExt(Ops[0], Int64Ty, "vtstd"); |
5892 | 6 | } |
5893 | 182 | case NEON::BI__builtin_neon_vset_lane_i8: |
5894 | 182 | case NEON::BI__builtin_neon_vset_lane_i16: |
5895 | 182 | case NEON::BI__builtin_neon_vset_lane_i32: |
5896 | 182 | case NEON::BI__builtin_neon_vset_lane_i64: |
5897 | 182 | case NEON::BI__builtin_neon_vset_lane_f32: |
5898 | 182 | case NEON::BI__builtin_neon_vsetq_lane_i8: |
5899 | 182 | case NEON::BI__builtin_neon_vsetq_lane_i16: |
5900 | 182 | case NEON::BI__builtin_neon_vsetq_lane_i32: |
5901 | 182 | case NEON::BI__builtin_neon_vsetq_lane_i64: |
5902 | 182 | case NEON::BI__builtin_neon_vsetq_lane_f32: |
5903 | 182 | Ops.push_back(EmitScalarExpr(E->getArg(2))); |
5904 | 182 | return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); |
5905 | 15 | case NEON::BI__builtin_neon_vset_lane_f64: |
5906 | 15 | // The vector type needs a cast for the v1f64 variant. |
5907 | 15 | Ops[1] = Builder.CreateBitCast(Ops[1], |
5908 | 15 | llvm::VectorType::get(DoubleTy, 1)); |
5909 | 15 | Ops.push_back(EmitScalarExpr(E->getArg(2))); |
5910 | 15 | return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); |
5911 | 7 | case NEON::BI__builtin_neon_vsetq_lane_f64: |
5912 | 7 | // The vector type needs a cast for the v2f64 variant. |
5913 | 7 | Ops[1] = Builder.CreateBitCast(Ops[1], |
5914 | 7 | llvm::VectorType::get(DoubleTy, 2)); |
5915 | 7 | Ops.push_back(EmitScalarExpr(E->getArg(2))); |
5916 | 7 | return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); |
5917 | 182 | |
5918 | 34 | case NEON::BI__builtin_neon_vget_lane_i8: |
5919 | 34 | case NEON::BI__builtin_neon_vdupb_lane_i8: |
5920 | 34 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int8Ty, 8)); |
5921 | 34 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
5922 | 34 | "vget_lane"); |
5923 | 32 | case NEON::BI__builtin_neon_vgetq_lane_i8: |
5924 | 32 | case NEON::BI__builtin_neon_vdupb_laneq_i8: |
5925 | 32 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int8Ty, 16)); |
5926 | 32 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
5927 | 32 | "vgetq_lane"); |
5928 | 46 | case NEON::BI__builtin_neon_vget_lane_i16: |
5929 | 46 | case NEON::BI__builtin_neon_vduph_lane_i16: |
5930 | 46 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int16Ty, 4)); |
5931 | 46 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
5932 | 46 | "vget_lane"); |
5933 | 44 | case NEON::BI__builtin_neon_vgetq_lane_i16: |
5934 | 44 | case NEON::BI__builtin_neon_vduph_laneq_i16: |
5935 | 44 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int16Ty, 8)); |
5936 | 44 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
5937 | 44 | "vgetq_lane"); |
5938 | 35 | case NEON::BI__builtin_neon_vget_lane_i32: |
5939 | 35 | case NEON::BI__builtin_neon_vdups_lane_i32: |
5940 | 35 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int32Ty, 2)); |
5941 | 35 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
5942 | 35 | "vget_lane"); |
5943 | 5 | case NEON::BI__builtin_neon_vdups_lane_f32: |
5944 | 5 | Ops[0] = Builder.CreateBitCast(Ops[0], |
5945 | 5 | llvm::VectorType::get(FloatTy, 2)); |
5946 | 5 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
5947 | 5 | "vdups_lane"); |
5948 | 33 | case NEON::BI__builtin_neon_vgetq_lane_i32: |
5949 | 33 | case NEON::BI__builtin_neon_vdups_laneq_i32: |
5950 | 33 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int32Ty, 4)); |
5951 | 33 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
5952 | 33 | "vgetq_lane"); |
5953 | 2.60k | case NEON::BI__builtin_neon_vget_lane_i64: |
5954 | 2.60k | case NEON::BI__builtin_neon_vdupd_lane_i64: |
5955 | 2.60k | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int64Ty, 1)); |
5956 | 2.60k | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
5957 | 2.60k | "vget_lane"); |
5958 | 5 | case NEON::BI__builtin_neon_vdupd_lane_f64: |
5959 | 5 | Ops[0] = Builder.CreateBitCast(Ops[0], |
5960 | 5 | llvm::VectorType::get(DoubleTy, 1)); |
5961 | 5 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
5962 | 5 | "vdupd_lane"); |
5963 | 6.05k | case NEON::BI__builtin_neon_vgetq_lane_i64: |
5964 | 6.05k | case NEON::BI__builtin_neon_vdupd_laneq_i64: |
5965 | 6.05k | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int64Ty, 2)); |
5966 | 6.05k | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
5967 | 6.05k | "vgetq_lane"); |
5968 | 497 | case NEON::BI__builtin_neon_vget_lane_f32: |
5969 | 497 | Ops[0] = Builder.CreateBitCast(Ops[0], |
5970 | 497 | llvm::VectorType::get(FloatTy, 2)); |
5971 | 497 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
5972 | 497 | "vget_lane"); |
5973 | 191 | case NEON::BI__builtin_neon_vget_lane_f64: |
5974 | 191 | Ops[0] = Builder.CreateBitCast(Ops[0], |
5975 | 191 | llvm::VectorType::get(DoubleTy, 1)); |
5976 | 191 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
5977 | 191 | "vget_lane"); |
5978 | 969 | case NEON::BI__builtin_neon_vgetq_lane_f32: |
5979 | 969 | case NEON::BI__builtin_neon_vdups_laneq_f32: |
5980 | 969 | Ops[0] = Builder.CreateBitCast(Ops[0], |
5981 | 969 | llvm::VectorType::get(FloatTy, 4)); |
5982 | 969 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
5983 | 969 | "vgetq_lane"); |
5984 | 397 | case NEON::BI__builtin_neon_vgetq_lane_f64: |
5985 | 397 | case NEON::BI__builtin_neon_vdupd_laneq_f64: |
5986 | 397 | Ops[0] = Builder.CreateBitCast(Ops[0], |
5987 | 397 | llvm::VectorType::get(DoubleTy, 2)); |
5988 | 397 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
5989 | 397 | "vgetq_lane"); |
5990 | 6 | case NEON::BI__builtin_neon_vaddd_s64: |
5991 | 6 | case NEON::BI__builtin_neon_vaddd_u64: |
5992 | 6 | return Builder.CreateAdd(Ops[0], EmitScalarExpr(E->getArg(1)), "vaddd"); |
5993 | 6 | case NEON::BI__builtin_neon_vsubd_s64: |
5994 | 6 | case NEON::BI__builtin_neon_vsubd_u64: |
5995 | 6 | return Builder.CreateSub(Ops[0], EmitScalarExpr(E->getArg(1)), "vsubd"); |
5996 | 6 | case NEON::BI__builtin_neon_vqdmlalh_s16: |
5997 | 6 | case NEON::BI__builtin_neon_vqdmlslh_s16: { |
5998 | 6 | SmallVector<Value *, 2> ProductOps; |
5999 | 6 | ProductOps.push_back(vectorWrapScalar16(Ops[1])); |
6000 | 6 | ProductOps.push_back(vectorWrapScalar16(EmitScalarExpr(E->getArg(2)))); |
6001 | 6 | llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); |
6002 | 6 | Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), |
6003 | 6 | ProductOps, "vqdmlXl"); |
6004 | 6 | Constant *CI = ConstantInt::get(SizeTy, 0); |
6005 | 6 | Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); |
6006 | 6 | |
6007 | 6 | unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlalh_s16 |
6008 | 3 | ? Intrinsic::aarch64_neon_sqadd |
6009 | 3 | : Intrinsic::aarch64_neon_sqsub; |
6010 | 6 | return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int32Ty), Ops, "vqdmlXl"); |
6011 | 6 | } |
6012 | 3 | case NEON::BI__builtin_neon_vqshlud_n_s64: { |
6013 | 3 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
6014 | 3 | Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); |
6015 | 3 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqshlu, Int64Ty), |
6016 | 3 | Ops, "vqshlu_n"); |
6017 | 6 | } |
6018 | 6 | case NEON::BI__builtin_neon_vqshld_n_u64: |
6019 | 6 | case NEON::BI__builtin_neon_vqshld_n_s64: { |
6020 | 6 | unsigned Int = BuiltinID == NEON::BI__builtin_neon_vqshld_n_u64 |
6021 | 3 | ? Intrinsic::aarch64_neon_uqshl |
6022 | 3 | : Intrinsic::aarch64_neon_sqshl; |
6023 | 6 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
6024 | 6 | Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); |
6025 | 6 | return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vqshl_n"); |
6026 | 6 | } |
6027 | 6 | case NEON::BI__builtin_neon_vrshrd_n_u64: |
6028 | 6 | case NEON::BI__builtin_neon_vrshrd_n_s64: { |
6029 | 6 | unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrshrd_n_u64 |
6030 | 3 | ? Intrinsic::aarch64_neon_urshl |
6031 | 3 | : Intrinsic::aarch64_neon_srshl; |
6032 | 6 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
6033 | 6 | int SV = cast<ConstantInt>(Ops[1])->getSExtValue(); |
6034 | 6 | Ops[1] = ConstantInt::get(Int64Ty, -SV); |
6035 | 6 | return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vrshr_n"); |
6036 | 6 | } |
6037 | 6 | case NEON::BI__builtin_neon_vrsrad_n_u64: |
6038 | 6 | case NEON::BI__builtin_neon_vrsrad_n_s64: { |
6039 | 6 | unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrsrad_n_u64 |
6040 | 3 | ? Intrinsic::aarch64_neon_urshl |
6041 | 3 | : Intrinsic::aarch64_neon_srshl; |
6042 | 6 | Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); |
6043 | 6 | Ops.push_back(Builder.CreateNeg(EmitScalarExpr(E->getArg(2)))); |
6044 | 6 | Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Int64Ty), |
6045 | 6 | {Ops[1], Builder.CreateSExt(Ops[2], Int64Ty)}); |
6046 | 6 | return Builder.CreateAdd(Ops[0], Builder.CreateBitCast(Ops[1], Int64Ty)); |
6047 | 6 | } |
6048 | 6 | case NEON::BI__builtin_neon_vshld_n_s64: |
6049 | 6 | case NEON::BI__builtin_neon_vshld_n_u64: { |
6050 | 6 | llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); |
6051 | 6 | return Builder.CreateShl( |
6052 | 6 | Ops[0], ConstantInt::get(Int64Ty, Amt->getZExtValue()), "shld_n"); |
6053 | 6 | } |
6054 | 3 | case NEON::BI__builtin_neon_vshrd_n_s64: { |
6055 | 3 | llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); |
6056 | 3 | return Builder.CreateAShr( |
6057 | 3 | Ops[0], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), |
6058 | 3 | Amt->getZExtValue())), |
6059 | 3 | "shrd_n"); |
6060 | 6 | } |
6061 | 4 | case NEON::BI__builtin_neon_vshrd_n_u64: { |
6062 | 4 | llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); |
6063 | 4 | uint64_t ShiftAmt = Amt->getZExtValue(); |
6064 | 4 | // Right-shifting an unsigned value by its size yields 0. |
6065 | 4 | if (ShiftAmt == 64) |
6066 | 2 | return ConstantInt::get(Int64Ty, 0); |
6067 | 2 | return Builder.CreateLShr(Ops[0], ConstantInt::get(Int64Ty, ShiftAmt), |
6068 | 2 | "shrd_n"); |
6069 | 2 | } |
6070 | 3 | case NEON::BI__builtin_neon_vsrad_n_s64: { |
6071 | 3 | llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); |
6072 | 3 | Ops[1] = Builder.CreateAShr( |
6073 | 3 | Ops[1], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), |
6074 | 3 | Amt->getZExtValue())), |
6075 | 3 | "shrd_n"); |
6076 | 3 | return Builder.CreateAdd(Ops[0], Ops[1]); |
6077 | 2 | } |
6078 | 4 | case NEON::BI__builtin_neon_vsrad_n_u64: { |
6079 | 4 | llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); |
6080 | 4 | uint64_t ShiftAmt = Amt->getZExtValue(); |
6081 | 4 | // Right-shifting an unsigned value by its size yields 0. |
6082 | 4 | // As Op + 0 = Op, return Ops[0] directly. |
6083 | 4 | if (ShiftAmt == 64) |
6084 | 1 | return Ops[0]; |
6085 | 3 | Ops[1] = Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, ShiftAmt), |
6086 | 3 | "shrd_n"); |
6087 | 3 | return Builder.CreateAdd(Ops[0], Ops[1]); |
6088 | 3 | } |
6089 | 12 | case NEON::BI__builtin_neon_vqdmlalh_lane_s16: |
6090 | 12 | case NEON::BI__builtin_neon_vqdmlalh_laneq_s16: |
6091 | 12 | case NEON::BI__builtin_neon_vqdmlslh_lane_s16: |
6092 | 12 | case NEON::BI__builtin_neon_vqdmlslh_laneq_s16: { |
6093 | 12 | Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), |
6094 | 12 | "lane"); |
6095 | 12 | SmallVector<Value *, 2> ProductOps; |
6096 | 12 | ProductOps.push_back(vectorWrapScalar16(Ops[1])); |
6097 | 12 | ProductOps.push_back(vectorWrapScalar16(Ops[2])); |
6098 | 12 | llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); |
6099 | 12 | Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), |
6100 | 12 | ProductOps, "vqdmlXl"); |
6101 | 12 | Constant *CI = ConstantInt::get(SizeTy, 0); |
6102 | 12 | Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); |
6103 | 12 | Ops.pop_back(); |
6104 | 12 | |
6105 | 12 | unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlalh_lane_s16 || |
6106 | 9 | BuiltinID == NEON::BI__builtin_neon_vqdmlalh_laneq_s16) |
6107 | 6 | ? Intrinsic::aarch64_neon_sqadd |
6108 | 6 | : Intrinsic::aarch64_neon_sqsub; |
6109 | 12 | return EmitNeonCall(CGM.getIntrinsic(AccInt, Int32Ty), Ops, "vqdmlXl"); |
6110 | 12 | } |
6111 | 6 | case NEON::BI__builtin_neon_vqdmlals_s32: |
6112 | 6 | case NEON::BI__builtin_neon_vqdmlsls_s32: { |
6113 | 6 | SmallVector<Value *, 2> ProductOps; |
6114 | 6 | ProductOps.push_back(Ops[1]); |
6115 | 6 | ProductOps.push_back(EmitScalarExpr(E->getArg(2))); |
6116 | 6 | Ops[1] = |
6117 | 6 | EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), |
6118 | 6 | ProductOps, "vqdmlXl"); |
6119 | 6 | |
6120 | 6 | unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlals_s32 |
6121 | 3 | ? Intrinsic::aarch64_neon_sqadd |
6122 | 3 | : Intrinsic::aarch64_neon_sqsub; |
6123 | 6 | return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int64Ty), Ops, "vqdmlXl"); |
6124 | 6 | } |
6125 | 12 | case NEON::BI__builtin_neon_vqdmlals_lane_s32: |
6126 | 12 | case NEON::BI__builtin_neon_vqdmlals_laneq_s32: |
6127 | 12 | case NEON::BI__builtin_neon_vqdmlsls_lane_s32: |
6128 | 12 | case NEON::BI__builtin_neon_vqdmlsls_laneq_s32: { |
6129 | 12 | Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), |
6130 | 12 | "lane"); |
6131 | 12 | SmallVector<Value *, 2> ProductOps; |
6132 | 12 | ProductOps.push_back(Ops[1]); |
6133 | 12 | ProductOps.push_back(Ops[2]); |
6134 | 12 | Ops[1] = |
6135 | 12 | EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), |
6136 | 12 | ProductOps, "vqdmlXl"); |
6137 | 12 | Ops.pop_back(); |
6138 | 12 | |
6139 | 12 | unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlals_lane_s32 || |
6140 | 9 | BuiltinID == NEON::BI__builtin_neon_vqdmlals_laneq_s32) |
6141 | 6 | ? Intrinsic::aarch64_neon_sqadd |
6142 | 6 | : Intrinsic::aarch64_neon_sqsub; |
6143 | 12 | return EmitNeonCall(CGM.getIntrinsic(AccInt, Int64Ty), Ops, "vqdmlXl"); |
6144 | 5.28k | } |
6145 | 5.28k | } |
6146 | 5.28k | |
6147 | 5.28k | llvm::VectorType *VTy = GetNeonType(this, Type); |
6148 | 5.28k | llvm::Type *Ty = VTy; |
6149 | 5.28k | if (!Ty) |
6150 | 0 | return nullptr; |
6151 | 5.28k | |
6152 | 5.28k | // Not all intrinsics handled by the common case work for AArch64 yet, so only |
6153 | 5.28k | // defer to common code if it's been added to our special map. |
6154 | 5.28k | Builtin = findNeonIntrinsicInMap(AArch64SIMDIntrinsicMap, BuiltinID, |
6155 | 5.28k | AArch64SIMDIntrinsicsProvenSorted); |
6156 | 5.28k | |
6157 | 5.28k | if (Builtin) |
6158 | 1.57k | return EmitCommonNeonBuiltinExpr( |
6159 | 1.57k | Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, |
6160 | 1.57k | Builtin->NameHint, Builtin->TypeModifier, E, Ops, |
6161 | 1.57k | /*never use addresses*/ Address::invalid(), Address::invalid()); |
6162 | 3.70k | |
6163 | 3.70k | if (Value *3.70k V3.70k = EmitAArch64TblBuiltinExpr(*this, BuiltinID, E, Ops)) |
6164 | 216 | return V; |
6165 | 3.48k | |
6166 | 3.48k | unsigned Int; |
6167 | 3.48k | switch (BuiltinID) { |
6168 | 0 | default: return nullptr; |
6169 | 73 | case NEON::BI__builtin_neon_vbsl_v: |
6170 | 73 | case NEON::BI__builtin_neon_vbslq_v: { |
6171 | 73 | llvm::Type *BitTy = llvm::VectorType::getInteger(VTy); |
6172 | 73 | Ops[0] = Builder.CreateBitCast(Ops[0], BitTy, "vbsl"); |
6173 | 73 | Ops[1] = Builder.CreateBitCast(Ops[1], BitTy, "vbsl"); |
6174 | 73 | Ops[2] = Builder.CreateBitCast(Ops[2], BitTy, "vbsl"); |
6175 | 73 | |
6176 | 73 | Ops[1] = Builder.CreateAnd(Ops[0], Ops[1], "vbsl"); |
6177 | 73 | Ops[2] = Builder.CreateAnd(Builder.CreateNot(Ops[0]), Ops[2], "vbsl"); |
6178 | 73 | Ops[0] = Builder.CreateOr(Ops[1], Ops[2], "vbsl"); |
6179 | 73 | return Builder.CreateBitCast(Ops[0], Ty); |
6180 | 73 | } |
6181 | 28 | case NEON::BI__builtin_neon_vfma_lane_v: |
6182 | 28 | case NEON::BI__builtin_neon_vfmaq_lane_v: { // Only used for FP types |
6183 | 28 | // The ARM builtins (and instructions) have the addend as the first |
6184 | 28 | // operand, but the 'fma' intrinsics have it last. Swap it around here. |
6185 | 28 | Value *Addend = Ops[0]; |
6186 | 28 | Value *Multiplicand = Ops[1]; |
6187 | 28 | Value *LaneSource = Ops[2]; |
6188 | 28 | Ops[0] = Multiplicand; |
6189 | 28 | Ops[1] = LaneSource; |
6190 | 28 | Ops[2] = Addend; |
6191 | 28 | |
6192 | 28 | // Now adjust things to handle the lane access. |
6193 | 28 | llvm::Type *SourceTy = BuiltinID == NEON::BI__builtin_neon_vfmaq_lane_v ? |
6194 | 14 | llvm::VectorType::get(VTy->getElementType(), VTy->getNumElements() / 2) : |
6195 | 14 | VTy; |
6196 | 28 | llvm::Constant *cst = cast<Constant>(Ops[3]); |
6197 | 28 | Value *SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), cst); |
6198 | 28 | Ops[1] = Builder.CreateBitCast(Ops[1], SourceTy); |
6199 | 28 | Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV, "lane"); |
6200 | 28 | |
6201 | 28 | Ops.pop_back(); |
6202 | 28 | Int = Intrinsic::fma; |
6203 | 28 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmla"); |
6204 | 28 | } |
6205 | 14 | case NEON::BI__builtin_neon_vfma_laneq_v: { |
6206 | 14 | llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); |
6207 | 14 | // v1f64 fma should be mapped to Neon scalar f64 fma |
6208 | 14 | if (VTy && 14 VTy->getElementType() == DoubleTy14 ) { |
6209 | 6 | Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); |
6210 | 6 | Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); |
6211 | 6 | llvm::Type *VTy = GetNeonType(this, |
6212 | 6 | NeonTypeFlags(NeonTypeFlags::Float64, false, true)); |
6213 | 6 | Ops[2] = Builder.CreateBitCast(Ops[2], VTy); |
6214 | 6 | Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); |
6215 | 6 | Value *F = CGM.getIntrinsic(Intrinsic::fma, DoubleTy); |
6216 | 6 | Value *Result = Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); |
6217 | 6 | return Builder.CreateBitCast(Result, Ty); |
6218 | 6 | } |
6219 | 8 | Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); |
6220 | 8 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
6221 | 8 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
6222 | 8 | |
6223 | 8 | llvm::Type *STy = llvm::VectorType::get(VTy->getElementType(), |
6224 | 8 | VTy->getNumElements() * 2); |
6225 | 8 | Ops[2] = Builder.CreateBitCast(Ops[2], STy); |
6226 | 8 | Value* SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), |
6227 | 8 | cast<ConstantInt>(Ops[3])); |
6228 | 8 | Ops[2] = Builder.CreateShuffleVector(Ops[2], Ops[2], SV, "lane"); |
6229 | 8 | |
6230 | 8 | return Builder.CreateCall(F, {Ops[2], Ops[1], Ops[0]}); |
6231 | 8 | } |
6232 | 16 | case NEON::BI__builtin_neon_vfmaq_laneq_v: { |
6233 | 16 | Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); |
6234 | 16 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
6235 | 16 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
6236 | 16 | |
6237 | 16 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
6238 | 16 | Ops[2] = EmitNeonSplat(Ops[2], cast<ConstantInt>(Ops[3])); |
6239 | 16 | return Builder.CreateCall(F, {Ops[2], Ops[1], Ops[0]}); |
6240 | 8 | } |
6241 | 24 | case NEON::BI__builtin_neon_vfmas_lane_f32: |
6242 | 24 | case NEON::BI__builtin_neon_vfmas_laneq_f32: |
6243 | 24 | case NEON::BI__builtin_neon_vfmad_lane_f64: |
6244 | 24 | case NEON::BI__builtin_neon_vfmad_laneq_f64: { |
6245 | 24 | Ops.push_back(EmitScalarExpr(E->getArg(3))); |
6246 | 24 | llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext())); |
6247 | 24 | Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); |
6248 | 24 | Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); |
6249 | 24 | return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); |
6250 | 24 | } |
6251 | 37 | case NEON::BI__builtin_neon_vmull_v: |
6252 | 37 | // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. |
6253 | 37 | Int = usgn ? Intrinsic::aarch64_neon_umull17 : Intrinsic::aarch64_neon_smull20 ; |
6254 | 37 | if (Type.isPoly()37 ) Int = Intrinsic::aarch64_neon_pmull3 ; |
6255 | 37 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); |
6256 | 48 | case NEON::BI__builtin_neon_vmax_v: |
6257 | 48 | case NEON::BI__builtin_neon_vmaxq_v: |
6258 | 48 | // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. |
6259 | 48 | Int = usgn ? Intrinsic::aarch64_neon_umax18 : Intrinsic::aarch64_neon_smax30 ; |
6260 | 48 | if (Ty->isFPOrFPVectorTy()48 ) Int = Intrinsic::aarch64_neon_fmax12 ; |
6261 | 48 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax"); |
6262 | 48 | case NEON::BI__builtin_neon_vmin_v: |
6263 | 48 | case NEON::BI__builtin_neon_vminq_v: |
6264 | 48 | // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. |
6265 | 48 | Int = usgn ? Intrinsic::aarch64_neon_umin18 : Intrinsic::aarch64_neon_smin30 ; |
6266 | 48 | if (Ty->isFPOrFPVectorTy()48 ) Int = Intrinsic::aarch64_neon_fmin12 ; |
6267 | 48 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin"); |
6268 | 48 | case NEON::BI__builtin_neon_vabd_v: |
6269 | 48 | case NEON::BI__builtin_neon_vabdq_v: |
6270 | 48 | // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. |
6271 | 48 | Int = usgn ? Intrinsic::aarch64_neon_uabd18 : Intrinsic::aarch64_neon_sabd30 ; |
6272 | 48 | if (Ty->isFPOrFPVectorTy()48 ) Int = Intrinsic::aarch64_neon_fabd12 ; |
6273 | 48 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd"); |
6274 | 36 | case NEON::BI__builtin_neon_vpadal_v: |
6275 | 36 | case NEON::BI__builtin_neon_vpadalq_v: { |
6276 | 36 | unsigned ArgElts = VTy->getNumElements(); |
6277 | 36 | llvm::IntegerType *EltTy = cast<IntegerType>(VTy->getElementType()); |
6278 | 36 | unsigned BitWidth = EltTy->getBitWidth(); |
6279 | 36 | llvm::Type *ArgTy = llvm::VectorType::get( |
6280 | 36 | llvm::IntegerType::get(getLLVMContext(), BitWidth/2), 2*ArgElts); |
6281 | 36 | llvm::Type* Tys[2] = { VTy, ArgTy }; |
6282 | 36 | Int = usgn ? Intrinsic::aarch64_neon_uaddlp18 : Intrinsic::aarch64_neon_saddlp18 ; |
6283 | 36 | SmallVector<llvm::Value*, 1> TmpOps; |
6284 | 36 | TmpOps.push_back(Ops[1]); |
6285 | 36 | Function *F = CGM.getIntrinsic(Int, Tys); |
6286 | 36 | llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vpadal"); |
6287 | 36 | llvm::Value *addend = Builder.CreateBitCast(Ops[0], tmp->getType()); |
6288 | 36 | return Builder.CreateAdd(tmp, addend); |
6289 | 36 | } |
6290 | 45 | case NEON::BI__builtin_neon_vpmin_v: |
6291 | 45 | case NEON::BI__builtin_neon_vpminq_v: |
6292 | 45 | // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. |
6293 | 45 | Int = usgn ? Intrinsic::aarch64_neon_uminp18 : Intrinsic::aarch64_neon_sminp27 ; |
6294 | 45 | if (Ty->isFPOrFPVectorTy()45 ) Int = Intrinsic::aarch64_neon_fminp9 ; |
6295 | 45 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin"); |
6296 | 45 | case NEON::BI__builtin_neon_vpmax_v: |
6297 | 45 | case NEON::BI__builtin_neon_vpmaxq_v: |
6298 | 45 | // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. |
6299 | 45 | Int = usgn ? Intrinsic::aarch64_neon_umaxp18 : Intrinsic::aarch64_neon_smaxp27 ; |
6300 | 45 | if (Ty->isFPOrFPVectorTy()45 ) Int = Intrinsic::aarch64_neon_fmaxp9 ; |
6301 | 45 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax"); |
6302 | 12 | case NEON::BI__builtin_neon_vminnm_v: |
6303 | 12 | case NEON::BI__builtin_neon_vminnmq_v: |
6304 | 12 | Int = Intrinsic::aarch64_neon_fminnm; |
6305 | 12 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vminnm"); |
6306 | 12 | case NEON::BI__builtin_neon_vmaxnm_v: |
6307 | 12 | case NEON::BI__builtin_neon_vmaxnmq_v: |
6308 | 12 | Int = Intrinsic::aarch64_neon_fmaxnm; |
6309 | 12 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmaxnm"); |
6310 | 3 | case NEON::BI__builtin_neon_vrecpss_f32: { |
6311 | 3 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
6312 | 3 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, FloatTy), |
6313 | 3 | Ops, "vrecps"); |
6314 | 12 | } |
6315 | 3 | case NEON::BI__builtin_neon_vrecpsd_f64: { |
6316 | 3 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
6317 | 3 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, DoubleTy), |
6318 | 3 | Ops, "vrecps"); |
6319 | 12 | } |
6320 | 18 | case NEON::BI__builtin_neon_vqshrun_n_v: |
6321 | 18 | Int = Intrinsic::aarch64_neon_sqshrun; |
6322 | 18 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrun_n"); |
6323 | 18 | case NEON::BI__builtin_neon_vqrshrun_n_v: |
6324 | 18 | Int = Intrinsic::aarch64_neon_sqrshrun; |
6325 | 18 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrun_n"); |
6326 | 36 | case NEON::BI__builtin_neon_vqshrn_n_v: |
6327 | 36 | Int = usgn ? Intrinsic::aarch64_neon_uqshrn18 : Intrinsic::aarch64_neon_sqshrn18 ; |
6328 | 36 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n"); |
6329 | 36 | case NEON::BI__builtin_neon_vrshrn_n_v: |
6330 | 36 | Int = Intrinsic::aarch64_neon_rshrn; |
6331 | 36 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshrn_n"); |
6332 | 37 | case NEON::BI__builtin_neon_vqrshrn_n_v: |
6333 | 37 | Int = usgn ? Intrinsic::aarch64_neon_uqrshrn18 : Intrinsic::aarch64_neon_sqrshrn19 ; |
6334 | 37 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n"); |
6335 | 15 | case NEON::BI__builtin_neon_vrnda_v: |
6336 | 15 | case NEON::BI__builtin_neon_vrndaq_v: { |
6337 | 15 | Int = Intrinsic::round; |
6338 | 15 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnda"); |
6339 | 15 | } |
6340 | 12 | case NEON::BI__builtin_neon_vrndi_v: |
6341 | 12 | case NEON::BI__builtin_neon_vrndiq_v: { |
6342 | 12 | Int = Intrinsic::nearbyint; |
6343 | 12 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndi"); |
6344 | 12 | } |
6345 | 15 | case NEON::BI__builtin_neon_vrndm_v: |
6346 | 15 | case NEON::BI__builtin_neon_vrndmq_v: { |
6347 | 15 | Int = Intrinsic::floor; |
6348 | 15 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndm"); |
6349 | 15 | } |
6350 | 15 | case NEON::BI__builtin_neon_vrndn_v: |
6351 | 15 | case NEON::BI__builtin_neon_vrndnq_v: { |
6352 | 15 | Int = Intrinsic::aarch64_neon_frintn; |
6353 | 15 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndn"); |
6354 | 15 | } |
6355 | 15 | case NEON::BI__builtin_neon_vrndp_v: |
6356 | 15 | case NEON::BI__builtin_neon_vrndpq_v: { |
6357 | 15 | Int = Intrinsic::ceil; |
6358 | 15 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndp"); |
6359 | 15 | } |
6360 | 15 | case NEON::BI__builtin_neon_vrndx_v: |
6361 | 15 | case NEON::BI__builtin_neon_vrndxq_v: { |
6362 | 15 | Int = Intrinsic::rint; |
6363 | 15 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndx"); |
6364 | 15 | } |
6365 | 15 | case NEON::BI__builtin_neon_vrnd_v: |
6366 | 15 | case NEON::BI__builtin_neon_vrndq_v: { |
6367 | 15 | Int = Intrinsic::trunc; |
6368 | 15 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndz"); |
6369 | 15 | } |
6370 | 70 | case NEON::BI__builtin_neon_vceqz_v: |
6371 | 70 | case NEON::BI__builtin_neon_vceqzq_v: |
6372 | 70 | return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OEQ, |
6373 | 70 | ICmpInst::ICMP_EQ, "vceqz"); |
6374 | 36 | case NEON::BI__builtin_neon_vcgez_v: |
6375 | 36 | case NEON::BI__builtin_neon_vcgezq_v: |
6376 | 36 | return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGE, |
6377 | 36 | ICmpInst::ICMP_SGE, "vcgez"); |
6378 | 36 | case NEON::BI__builtin_neon_vclez_v: |
6379 | 36 | case NEON::BI__builtin_neon_vclezq_v: |
6380 | 36 | return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLE, |
6381 | 36 | ICmpInst::ICMP_SLE, "vclez"); |
6382 | 36 | case NEON::BI__builtin_neon_vcgtz_v: |
6383 | 36 | case NEON::BI__builtin_neon_vcgtzq_v: |
6384 | 36 | return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGT, |
6385 | 36 | ICmpInst::ICMP_SGT, "vcgtz"); |
6386 | 36 | case NEON::BI__builtin_neon_vcltz_v: |
6387 | 36 | case NEON::BI__builtin_neon_vcltzq_v: |
6388 | 36 | return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLT, |
6389 | 36 | ICmpInst::ICMP_SLT, "vcltz"); |
6390 | 12 | case NEON::BI__builtin_neon_vcvt_f64_v: |
6391 | 12 | case NEON::BI__builtin_neon_vcvtq_f64_v: |
6392 | 12 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
6393 | 12 | Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, quad)); |
6394 | 6 | return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") |
6395 | 6 | : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); |
6396 | 3 | case NEON::BI__builtin_neon_vcvt_f64_f32: { |
6397 | 3 | assert(Type.getEltType() == NeonTypeFlags::Float64 && quad && |
6398 | 3 | "unexpected vcvt_f64_f32 builtin"); |
6399 | 3 | NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float32, false, false); |
6400 | 3 | Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); |
6401 | 3 | |
6402 | 3 | return Builder.CreateFPExt(Ops[0], Ty, "vcvt"); |
6403 | 12 | } |
6404 | 3 | case NEON::BI__builtin_neon_vcvt_f32_f64: { |
6405 | 3 | assert(Type.getEltType() == NeonTypeFlags::Float32 && |
6406 | 3 | "unexpected vcvt_f32_f64 builtin"); |
6407 | 3 | NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float64, false, true); |
6408 | 3 | Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); |
6409 | 3 | |
6410 | 3 | return Builder.CreateFPTrunc(Ops[0], Ty, "vcvt"); |
6411 | 12 | } |
6412 | 24 | case NEON::BI__builtin_neon_vcvt_s32_v: |
6413 | 24 | case NEON::BI__builtin_neon_vcvt_u32_v: |
6414 | 24 | case NEON::BI__builtin_neon_vcvt_s64_v: |
6415 | 24 | case NEON::BI__builtin_neon_vcvt_u64_v: |
6416 | 24 | case NEON::BI__builtin_neon_vcvtq_s32_v: |
6417 | 24 | case NEON::BI__builtin_neon_vcvtq_u32_v: |
6418 | 24 | case NEON::BI__builtin_neon_vcvtq_s64_v: |
6419 | 24 | case NEON::BI__builtin_neon_vcvtq_u64_v: { |
6420 | 24 | Ops[0] = Builder.CreateBitCast(Ops[0], GetFloatNeonType(this, Type)); |
6421 | 24 | if (usgn) |
6422 | 12 | return Builder.CreateFPToUI(Ops[0], Ty); |
6423 | 12 | return Builder.CreateFPToSI(Ops[0], Ty); |
6424 | 12 | } |
6425 | 24 | case NEON::BI__builtin_neon_vcvta_s32_v: |
6426 | 24 | case NEON::BI__builtin_neon_vcvtaq_s32_v: |
6427 | 24 | case NEON::BI__builtin_neon_vcvta_u32_v: |
6428 | 24 | case NEON::BI__builtin_neon_vcvtaq_u32_v: |
6429 | 24 | case NEON::BI__builtin_neon_vcvta_s64_v: |
6430 | 24 | case NEON::BI__builtin_neon_vcvtaq_s64_v: |
6431 | 24 | case NEON::BI__builtin_neon_vcvta_u64_v: |
6432 | 24 | case NEON::BI__builtin_neon_vcvtaq_u64_v: { |
6433 | 24 | Int = usgn ? Intrinsic::aarch64_neon_fcvtau12 : Intrinsic::aarch64_neon_fcvtas12 ; |
6434 | 24 | llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; |
6435 | 24 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvta"); |
6436 | 24 | } |
6437 | 24 | case NEON::BI__builtin_neon_vcvtm_s32_v: |
6438 | 24 | case NEON::BI__builtin_neon_vcvtmq_s32_v: |
6439 | 24 | case NEON::BI__builtin_neon_vcvtm_u32_v: |
6440 | 24 | case NEON::BI__builtin_neon_vcvtmq_u32_v: |
6441 | 24 | case NEON::BI__builtin_neon_vcvtm_s64_v: |
6442 | 24 | case NEON::BI__builtin_neon_vcvtmq_s64_v: |
6443 | 24 | case NEON::BI__builtin_neon_vcvtm_u64_v: |
6444 | 24 | case NEON::BI__builtin_neon_vcvtmq_u64_v: { |
6445 | 24 | Int = usgn ? Intrinsic::aarch64_neon_fcvtmu12 : Intrinsic::aarch64_neon_fcvtms12 ; |
6446 | 24 | llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; |
6447 | 24 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtm"); |
6448 | 24 | } |
6449 | 24 | case NEON::BI__builtin_neon_vcvtn_s32_v: |
6450 | 24 | case NEON::BI__builtin_neon_vcvtnq_s32_v: |
6451 | 24 | case NEON::BI__builtin_neon_vcvtn_u32_v: |
6452 | 24 | case NEON::BI__builtin_neon_vcvtnq_u32_v: |
6453 | 24 | case NEON::BI__builtin_neon_vcvtn_s64_v: |
6454 | 24 | case NEON::BI__builtin_neon_vcvtnq_s64_v: |
6455 | 24 | case NEON::BI__builtin_neon_vcvtn_u64_v: |
6456 | 24 | case NEON::BI__builtin_neon_vcvtnq_u64_v: { |
6457 | 24 | Int = usgn ? Intrinsic::aarch64_neon_fcvtnu12 : Intrinsic::aarch64_neon_fcvtns12 ; |
6458 | 24 | llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; |
6459 | 24 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtn"); |
6460 | 24 | } |
6461 | 24 | case NEON::BI__builtin_neon_vcvtp_s32_v: |
6462 | 24 | case NEON::BI__builtin_neon_vcvtpq_s32_v: |
6463 | 24 | case NEON::BI__builtin_neon_vcvtp_u32_v: |
6464 | 24 | case NEON::BI__builtin_neon_vcvtpq_u32_v: |
6465 | 24 | case NEON::BI__builtin_neon_vcvtp_s64_v: |
6466 | 24 | case NEON::BI__builtin_neon_vcvtpq_s64_v: |
6467 | 24 | case NEON::BI__builtin_neon_vcvtp_u64_v: |
6468 | 24 | case NEON::BI__builtin_neon_vcvtpq_u64_v: { |
6469 | 24 | Int = usgn ? Intrinsic::aarch64_neon_fcvtpu12 : Intrinsic::aarch64_neon_fcvtps12 ; |
6470 | 24 | llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; |
6471 | 24 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtp"); |
6472 | 24 | } |
6473 | 15 | case NEON::BI__builtin_neon_vmulx_v: |
6474 | 15 | case NEON::BI__builtin_neon_vmulxq_v: { |
6475 | 15 | Int = Intrinsic::aarch64_neon_fmulx; |
6476 | 15 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmulx"); |
6477 | 15 | } |
6478 | 7 | case NEON::BI__builtin_neon_vmul_lane_v: |
6479 | 7 | case NEON::BI__builtin_neon_vmul_laneq_v: { |
6480 | 7 | // v1f64 vmul_lane should be mapped to Neon scalar mul lane |
6481 | 7 | bool Quad = false; |
6482 | 7 | if (BuiltinID == NEON::BI__builtin_neon_vmul_laneq_v) |
6483 | 4 | Quad = true; |
6484 | 7 | Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); |
6485 | 7 | llvm::Type *VTy = GetNeonType(this, |
6486 | 7 | NeonTypeFlags(NeonTypeFlags::Float64, false, Quad)); |
6487 | 7 | Ops[1] = Builder.CreateBitCast(Ops[1], VTy); |
6488 | 7 | Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract"); |
6489 | 7 | Value *Result = Builder.CreateFMul(Ops[0], Ops[1]); |
6490 | 7 | return Builder.CreateBitCast(Result, Ty); |
6491 | 7 | } |
6492 | 1 | case NEON::BI__builtin_neon_vnegd_s64: |
6493 | 1 | return Builder.CreateNeg(EmitScalarExpr(E->getArg(0)), "vnegd"); |
6494 | 9 | case NEON::BI__builtin_neon_vpmaxnm_v: |
6495 | 9 | case NEON::BI__builtin_neon_vpmaxnmq_v: { |
6496 | 9 | Int = Intrinsic::aarch64_neon_fmaxnmp; |
6497 | 9 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmaxnm"); |
6498 | 9 | } |
6499 | 9 | case NEON::BI__builtin_neon_vpminnm_v: |
6500 | 9 | case NEON::BI__builtin_neon_vpminnmq_v: { |
6501 | 9 | Int = Intrinsic::aarch64_neon_fminnmp; |
6502 | 9 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpminnm"); |
6503 | 9 | } |
6504 | 12 | case NEON::BI__builtin_neon_vsqrt_v: |
6505 | 12 | case NEON::BI__builtin_neon_vsqrtq_v: { |
6506 | 12 | Int = Intrinsic::sqrt; |
6507 | 12 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
6508 | 12 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqrt"); |
6509 | 12 | } |
6510 | 18 | case NEON::BI__builtin_neon_vrbit_v: |
6511 | 18 | case NEON::BI__builtin_neon_vrbitq_v: { |
6512 | 18 | Int = Intrinsic::aarch64_neon_rbit; |
6513 | 18 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrbit"); |
6514 | 18 | } |
6515 | 3 | case NEON::BI__builtin_neon_vaddv_u8: |
6516 | 3 | // FIXME: These are handled by the AArch64 scalar code. |
6517 | 3 | usgn = true; |
6518 | 3 | // FALLTHROUGH |
6519 | 6 | case NEON::BI__builtin_neon_vaddv_s8: { |
6520 | 6 | Int = usgn ? Intrinsic::aarch64_neon_uaddv3 : Intrinsic::aarch64_neon_saddv3 ; |
6521 | 6 | Ty = Int32Ty; |
6522 | 6 | VTy = llvm::VectorType::get(Int8Ty, 8); |
6523 | 6 | llvm::Type *Tys[2] = { Ty, VTy }; |
6524 | 6 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
6525 | 6 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); |
6526 | 6 | return Builder.CreateTrunc(Ops[0], Int8Ty); |
6527 | 3 | } |
6528 | 3 | case NEON::BI__builtin_neon_vaddv_u16: |
6529 | 3 | usgn = true; |
6530 | 3 | // FALLTHROUGH |
6531 | 6 | case NEON::BI__builtin_neon_vaddv_s16: { |
6532 | 6 | Int = usgn ? Intrinsic::aarch64_neon_uaddv3 : Intrinsic::aarch64_neon_saddv3 ; |
6533 | 6 | Ty = Int32Ty; |
6534 | 6 | VTy = llvm::VectorType::get(Int16Ty, 4); |
6535 | 6 | llvm::Type *Tys[2] = { Ty, VTy }; |
6536 | 6 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
6537 | 6 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); |
6538 | 6 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
6539 | 3 | } |
6540 | 3 | case NEON::BI__builtin_neon_vaddvq_u8: |
6541 | 3 | usgn = true; |
6542 | 3 | // FALLTHROUGH |
6543 | 6 | case NEON::BI__builtin_neon_vaddvq_s8: { |
6544 | 6 | Int = usgn ? Intrinsic::aarch64_neon_uaddv3 : Intrinsic::aarch64_neon_saddv3 ; |
6545 | 6 | Ty = Int32Ty; |
6546 | 6 | VTy = llvm::VectorType::get(Int8Ty, 16); |
6547 | 6 | llvm::Type *Tys[2] = { Ty, VTy }; |
6548 | 6 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
6549 | 6 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); |
6550 | 6 | return Builder.CreateTrunc(Ops[0], Int8Ty); |
6551 | 3 | } |
6552 | 3 | case NEON::BI__builtin_neon_vaddvq_u16: |
6553 | 3 | usgn = true; |
6554 | 3 | // FALLTHROUGH |
6555 | 6 | case NEON::BI__builtin_neon_vaddvq_s16: { |
6556 | 6 | Int = usgn ? Intrinsic::aarch64_neon_uaddv3 : Intrinsic::aarch64_neon_saddv3 ; |
6557 | 6 | Ty = Int32Ty; |
6558 | 6 | VTy = llvm::VectorType::get(Int16Ty, 8); |
6559 | 6 | llvm::Type *Tys[2] = { Ty, VTy }; |
6560 | 6 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
6561 | 6 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); |
6562 | 6 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
6563 | 3 | } |
6564 | 3 | case NEON::BI__builtin_neon_vmaxv_u8: { |
6565 | 3 | Int = Intrinsic::aarch64_neon_umaxv; |
6566 | 3 | Ty = Int32Ty; |
6567 | 3 | VTy = llvm::VectorType::get(Int8Ty, 8); |
6568 | 3 | llvm::Type *Tys[2] = { Ty, VTy }; |
6569 | 3 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
6570 | 3 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); |
6571 | 3 | return Builder.CreateTrunc(Ops[0], Int8Ty); |
6572 | 3 | } |
6573 | 3 | case NEON::BI__builtin_neon_vmaxv_u16: { |
6574 | 3 | Int = Intrinsic::aarch64_neon_umaxv; |
6575 | 3 | Ty = Int32Ty; |
6576 | 3 | VTy = llvm::VectorType::get(Int16Ty, 4); |
6577 | 3 | llvm::Type *Tys[2] = { Ty, VTy }; |
6578 | 3 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
6579 | 3 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); |
6580 | 3 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
6581 | 3 | } |
6582 | 3 | case NEON::BI__builtin_neon_vmaxvq_u8: { |
6583 | 3 | Int = Intrinsic::aarch64_neon_umaxv; |
6584 | 3 | Ty = Int32Ty; |
6585 | 3 | VTy = llvm::VectorType::get(Int8Ty, 16); |
6586 | 3 | llvm::Type *Tys[2] = { Ty, VTy }; |
6587 | 3 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
6588 | 3 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); |
6589 | 3 | return Builder.CreateTrunc(Ops[0], Int8Ty); |
6590 | 3 | } |
6591 | 3 | case NEON::BI__builtin_neon_vmaxvq_u16: { |
6592 | 3 | Int = Intrinsic::aarch64_neon_umaxv; |
6593 | 3 | Ty = Int32Ty; |
6594 | 3 | VTy = llvm::VectorType::get(Int16Ty, 8); |
6595 | 3 | llvm::Type *Tys[2] = { Ty, VTy }; |
6596 | 3 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
6597 | 3 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); |
6598 | 3 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
6599 | 3 | } |
6600 | 3 | case NEON::BI__builtin_neon_vmaxv_s8: { |
6601 | 3 | Int = Intrinsic::aarch64_neon_smaxv; |
6602 | 3 | Ty = Int32Ty; |
6603 | 3 | VTy = llvm::VectorType::get(Int8Ty, 8); |
6604 | 3 | llvm::Type *Tys[2] = { Ty, VTy }; |
6605 | 3 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
6606 | 3 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); |
6607 | 3 | return Builder.CreateTrunc(Ops[0], Int8Ty); |
6608 | 3 | } |
6609 | 3 | case NEON::BI__builtin_neon_vmaxv_s16: { |
6610 | 3 | Int = Intrinsic::aarch64_neon_smaxv; |
6611 | 3 | Ty = Int32Ty; |
6612 | 3 | VTy = llvm::VectorType::get(Int16Ty, 4); |
6613 | 3 | llvm::Type *Tys[2] = { Ty, VTy }; |
6614 | 3 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
6615 | 3 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); |
6616 | 3 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
6617 | 3 | } |
6618 | 3 | case NEON::BI__builtin_neon_vmaxvq_s8: { |
6619 | 3 | Int = Intrinsic::aarch64_neon_smaxv; |
6620 | 3 | Ty = Int32Ty; |
6621 | 3 | VTy = llvm::VectorType::get(Int8Ty, 16); |
6622 | 3 | llvm::Type *Tys[2] = { Ty, VTy }; |
6623 | 3 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
6624 | 3 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); |
6625 | 3 | return Builder.CreateTrunc(Ops[0], Int8Ty); |
6626 | 3 | } |
6627 | 3 | case NEON::BI__builtin_neon_vmaxvq_s16: { |
6628 | 3 | Int = Intrinsic::aarch64_neon_smaxv; |
6629 | 3 | Ty = Int32Ty; |
6630 | 3 | VTy = llvm::VectorType::get(Int16Ty, 8); |
6631 | 3 | llvm::Type *Tys[2] = { Ty, VTy }; |
6632 | 3 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
6633 | 3 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); |
6634 | 3 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
6635 | 3 | } |
6636 | 3 | case NEON::BI__builtin_neon_vminv_u8: { |
6637 | 3 | Int = Intrinsic::aarch64_neon_uminv; |
6638 | 3 | Ty = Int32Ty; |
6639 | 3 | VTy = llvm::VectorType::get(Int8Ty, 8); |
6640 | 3 | llvm::Type *Tys[2] = { Ty, VTy }; |
6641 | 3 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
6642 | 3 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); |
6643 | 3 | return Builder.CreateTrunc(Ops[0], Int8Ty); |
6644 | 3 | } |
6645 | 3 | case NEON::BI__builtin_neon_vminv_u16: { |
6646 | 3 | Int = Intrinsic::aarch64_neon_uminv; |
6647 | 3 | Ty = Int32Ty; |
6648 | 3 | VTy = llvm::VectorType::get(Int16Ty, 4); |
6649 | 3 | llvm::Type *Tys[2] = { Ty, VTy }; |
6650 | 3 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
6651 | 3 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); |
6652 | 3 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
6653 | 3 | } |
6654 | 3 | case NEON::BI__builtin_neon_vminvq_u8: { |
6655 | 3 | Int = Intrinsic::aarch64_neon_uminv; |
6656 | 3 | Ty = Int32Ty; |
6657 | 3 | VTy = llvm::VectorType::get(Int8Ty, 16); |
6658 | 3 | llvm::Type *Tys[2] = { Ty, VTy }; |
6659 | 3 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
6660 | 3 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); |
6661 | 3 | return Builder.CreateTrunc(Ops[0], Int8Ty); |
6662 | 3 | } |
6663 | 3 | case NEON::BI__builtin_neon_vminvq_u16: { |
6664 | 3 | Int = Intrinsic::aarch64_neon_uminv; |
6665 | 3 | Ty = Int32Ty; |
6666 | 3 | VTy = llvm::VectorType::get(Int16Ty, 8); |
6667 | 3 | llvm::Type *Tys[2] = { Ty, VTy }; |
6668 | 3 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
6669 | 3 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); |
6670 | 3 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
6671 | 3 | } |
6672 | 3 | case NEON::BI__builtin_neon_vminv_s8: { |
6673 | 3 | Int = Intrinsic::aarch64_neon_sminv; |
6674 | 3 | Ty = Int32Ty; |
6675 | 3 | VTy = llvm::VectorType::get(Int8Ty, 8); |
6676 | 3 | llvm::Type *Tys[2] = { Ty, VTy }; |
6677 | 3 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
6678 | 3 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); |
6679 | 3 | return Builder.CreateTrunc(Ops[0], Int8Ty); |
6680 | 3 | } |
6681 | 3 | case NEON::BI__builtin_neon_vminv_s16: { |
6682 | 3 | Int = Intrinsic::aarch64_neon_sminv; |
6683 | 3 | Ty = Int32Ty; |
6684 | 3 | VTy = llvm::VectorType::get(Int16Ty, 4); |
6685 | 3 | llvm::Type *Tys[2] = { Ty, VTy }; |
6686 | 3 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
6687 | 3 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); |
6688 | 3 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
6689 | 3 | } |
6690 | 3 | case NEON::BI__builtin_neon_vminvq_s8: { |
6691 | 3 | Int = Intrinsic::aarch64_neon_sminv; |
6692 | 3 | Ty = Int32Ty; |
6693 | 3 | VTy = llvm::VectorType::get(Int8Ty, 16); |
6694 | 3 | llvm::Type *Tys[2] = { Ty, VTy }; |
6695 | 3 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
6696 | 3 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); |
6697 | 3 | return Builder.CreateTrunc(Ops[0], Int8Ty); |
6698 | 3 | } |
6699 | 3 | case NEON::BI__builtin_neon_vminvq_s16: { |
6700 | 3 | Int = Intrinsic::aarch64_neon_sminv; |
6701 | 3 | Ty = Int32Ty; |
6702 | 3 | VTy = llvm::VectorType::get(Int16Ty, 8); |
6703 | 3 | llvm::Type *Tys[2] = { Ty, VTy }; |
6704 | 3 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
6705 | 3 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); |
6706 | 3 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
6707 | 3 | } |
6708 | 3 | case NEON::BI__builtin_neon_vmul_n_f64: { |
6709 | 3 | Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); |
6710 | 3 | Value *RHS = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), DoubleTy); |
6711 | 3 | return Builder.CreateFMul(Ops[0], RHS); |
6712 | 3 | } |
6713 | 3 | case NEON::BI__builtin_neon_vaddlv_u8: { |
6714 | 3 | Int = Intrinsic::aarch64_neon_uaddlv; |
6715 | 3 | Ty = Int32Ty; |
6716 | 3 | VTy = llvm::VectorType::get(Int8Ty, 8); |
6717 | 3 | llvm::Type *Tys[2] = { Ty, VTy }; |
6718 | 3 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
6719 | 3 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); |
6720 | 3 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
6721 | 3 | } |
6722 | 3 | case NEON::BI__builtin_neon_vaddlv_u16: { |
6723 | 3 | Int = Intrinsic::aarch64_neon_uaddlv; |
6724 | 3 | Ty = Int32Ty; |
6725 | 3 | VTy = llvm::VectorType::get(Int16Ty, 4); |
6726 | 3 | llvm::Type *Tys[2] = { Ty, VTy }; |
6727 | 3 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
6728 | 3 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); |
6729 | 3 | } |
6730 | 3 | case NEON::BI__builtin_neon_vaddlvq_u8: { |
6731 | 3 | Int = Intrinsic::aarch64_neon_uaddlv; |
6732 | 3 | Ty = Int32Ty; |
6733 | 3 | VTy = llvm::VectorType::get(Int8Ty, 16); |
6734 | 3 | llvm::Type *Tys[2] = { Ty, VTy }; |
6735 | 3 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
6736 | 3 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); |
6737 | 3 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
6738 | 3 | } |
6739 | 3 | case NEON::BI__builtin_neon_vaddlvq_u16: { |
6740 | 3 | Int = Intrinsic::aarch64_neon_uaddlv; |
6741 | 3 | Ty = Int32Ty; |
6742 | 3 | VTy = llvm::VectorType::get(Int16Ty, 8); |
6743 | 3 | llvm::Type *Tys[2] = { Ty, VTy }; |
6744 | 3 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
6745 | 3 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); |
6746 | 3 | } |
6747 | 3 | case NEON::BI__builtin_neon_vaddlv_s8: { |
6748 | 3 | Int = Intrinsic::aarch64_neon_saddlv; |
6749 | 3 | Ty = Int32Ty; |
6750 | 3 | VTy = llvm::VectorType::get(Int8Ty, 8); |
6751 | 3 | llvm::Type *Tys[2] = { Ty, VTy }; |
6752 | 3 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
6753 | 3 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); |
6754 | 3 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
6755 | 3 | } |
6756 | 3 | case NEON::BI__builtin_neon_vaddlv_s16: { |
6757 | 3 | Int = Intrinsic::aarch64_neon_saddlv; |
6758 | 3 | Ty = Int32Ty; |
6759 | 3 | VTy = llvm::VectorType::get(Int16Ty, 4); |
6760 | 3 | llvm::Type *Tys[2] = { Ty, VTy }; |
6761 | 3 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
6762 | 3 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); |
6763 | 3 | } |
6764 | 3 | case NEON::BI__builtin_neon_vaddlvq_s8: { |
6765 | 3 | Int = Intrinsic::aarch64_neon_saddlv; |
6766 | 3 | Ty = Int32Ty; |
6767 | 3 | VTy = llvm::VectorType::get(Int8Ty, 16); |
6768 | 3 | llvm::Type *Tys[2] = { Ty, VTy }; |
6769 | 3 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
6770 | 3 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); |
6771 | 3 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
6772 | 3 | } |
6773 | 3 | case NEON::BI__builtin_neon_vaddlvq_s16: { |
6774 | 3 | Int = Intrinsic::aarch64_neon_saddlv; |
6775 | 3 | Ty = Int32Ty; |
6776 | 3 | VTy = llvm::VectorType::get(Int16Ty, 8); |
6777 | 3 | llvm::Type *Tys[2] = { Ty, VTy }; |
6778 | 3 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
6779 | 3 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); |
6780 | 3 | } |
6781 | 62 | case NEON::BI__builtin_neon_vsri_n_v: |
6782 | 62 | case NEON::BI__builtin_neon_vsriq_n_v: { |
6783 | 62 | Int = Intrinsic::aarch64_neon_vsri; |
6784 | 62 | llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); |
6785 | 62 | return EmitNeonCall(Intrin, Ops, "vsri_n"); |
6786 | 62 | } |
6787 | 62 | case NEON::BI__builtin_neon_vsli_n_v: |
6788 | 62 | case NEON::BI__builtin_neon_vsliq_n_v: { |
6789 | 62 | Int = Intrinsic::aarch64_neon_vsli; |
6790 | 62 | llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); |
6791 | 62 | return EmitNeonCall(Intrin, Ops, "vsli_n"); |
6792 | 62 | } |
6793 | 51 | case NEON::BI__builtin_neon_vsra_n_v: |
6794 | 51 | case NEON::BI__builtin_neon_vsraq_n_v: |
6795 | 51 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
6796 | 51 | Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); |
6797 | 51 | return Builder.CreateAdd(Ops[0], Ops[1]); |
6798 | 48 | case NEON::BI__builtin_neon_vrsra_n_v: |
6799 | 48 | case NEON::BI__builtin_neon_vrsraq_n_v: { |
6800 | 48 | Int = usgn ? Intrinsic::aarch64_neon_urshl24 : Intrinsic::aarch64_neon_srshl24 ; |
6801 | 48 | SmallVector<llvm::Value*,2> TmpOps; |
6802 | 48 | TmpOps.push_back(Ops[1]); |
6803 | 48 | TmpOps.push_back(Ops[2]); |
6804 | 48 | Function* F = CGM.getIntrinsic(Int, Ty); |
6805 | 48 | llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vrshr_n", 1, true); |
6806 | 48 | Ops[0] = Builder.CreateBitCast(Ops[0], VTy); |
6807 | 48 | return Builder.CreateAdd(Ops[0], tmp); |
6808 | 48 | } |
6809 | 48 | // FIXME: Sharing loads & stores with 32-bit is complicated by the absence |
6810 | 48 | // of an Align parameter here. |
6811 | 84 | case NEON::BI__builtin_neon_vld1_x2_v: |
6812 | 84 | case NEON::BI__builtin_neon_vld1q_x2_v: |
6813 | 84 | case NEON::BI__builtin_neon_vld1_x3_v: |
6814 | 84 | case NEON::BI__builtin_neon_vld1q_x3_v: |
6815 | 84 | case NEON::BI__builtin_neon_vld1_x4_v: |
6816 | 84 | case NEON::BI__builtin_neon_vld1q_x4_v: { |
6817 | 84 | llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); |
6818 | 84 | Ops[1] = Builder.CreateBitCast(Ops[1], PTy); |
6819 | 84 | llvm::Type *Tys[2] = { VTy, PTy }; |
6820 | 84 | unsigned Int; |
6821 | 84 | switch (BuiltinID) { |
6822 | 28 | case NEON::BI__builtin_neon_vld1_x2_v: |
6823 | 28 | case NEON::BI__builtin_neon_vld1q_x2_v: |
6824 | 28 | Int = Intrinsic::aarch64_neon_ld1x2; |
6825 | 28 | break; |
6826 | 28 | case NEON::BI__builtin_neon_vld1_x3_v: |
6827 | 28 | case NEON::BI__builtin_neon_vld1q_x3_v: |
6828 | 28 | Int = Intrinsic::aarch64_neon_ld1x3; |
6829 | 28 | break; |
6830 | 28 | case NEON::BI__builtin_neon_vld1_x4_v: |
6831 | 28 | case NEON::BI__builtin_neon_vld1q_x4_v: |
6832 | 28 | Int = Intrinsic::aarch64_neon_ld1x4; |
6833 | 28 | break; |
6834 | 84 | } |
6835 | 84 | Function *F = CGM.getIntrinsic(Int, Tys); |
6836 | 84 | Ops[1] = Builder.CreateCall(F, Ops[1], "vld1xN"); |
6837 | 84 | Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); |
6838 | 84 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
6839 | 84 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
6840 | 84 | } |
6841 | 84 | case NEON::BI__builtin_neon_vst1_x2_v: |
6842 | 84 | case NEON::BI__builtin_neon_vst1q_x2_v: |
6843 | 84 | case NEON::BI__builtin_neon_vst1_x3_v: |
6844 | 84 | case NEON::BI__builtin_neon_vst1q_x3_v: |
6845 | 84 | case NEON::BI__builtin_neon_vst1_x4_v: |
6846 | 84 | case NEON::BI__builtin_neon_vst1q_x4_v: { |
6847 | 84 | llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); |
6848 | 84 | llvm::Type *Tys[2] = { VTy, PTy }; |
6849 | 84 | unsigned Int; |
6850 | 84 | switch (BuiltinID) { |
6851 | 28 | case NEON::BI__builtin_neon_vst1_x2_v: |
6852 | 28 | case NEON::BI__builtin_neon_vst1q_x2_v: |
6853 | 28 | Int = Intrinsic::aarch64_neon_st1x2; |
6854 | 28 | break; |
6855 | 28 | case NEON::BI__builtin_neon_vst1_x3_v: |
6856 | 28 | case NEON::BI__builtin_neon_vst1q_x3_v: |
6857 | 28 | Int = Intrinsic::aarch64_neon_st1x3; |
6858 | 28 | break; |
6859 | 28 | case NEON::BI__builtin_neon_vst1_x4_v: |
6860 | 28 | case NEON::BI__builtin_neon_vst1q_x4_v: |
6861 | 28 | Int = Intrinsic::aarch64_neon_st1x4; |
6862 | 28 | break; |
6863 | 84 | } |
6864 | 84 | std::rotate(Ops.begin(), Ops.begin() + 1, Ops.end()); |
6865 | 84 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, ""); |
6866 | 84 | } |
6867 | 88 | case NEON::BI__builtin_neon_vld1_v: |
6868 | 88 | case NEON::BI__builtin_neon_vld1q_v: { |
6869 | 88 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); |
6870 | 88 | auto Alignment = CharUnits::fromQuantity( |
6871 | 88 | BuiltinID == NEON::BI__builtin_neon_vld1_v ? 842 : 1646 ); |
6872 | 88 | return Builder.CreateAlignedLoad(VTy, Ops[0], Alignment); |
6873 | 88 | } |
6874 | 86 | case NEON::BI__builtin_neon_vst1_v: |
6875 | 86 | case NEON::BI__builtin_neon_vst1q_v: |
6876 | 86 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); |
6877 | 86 | Ops[1] = Builder.CreateBitCast(Ops[1], VTy); |
6878 | 86 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
6879 | 80 | case NEON::BI__builtin_neon_vld1_lane_v: |
6880 | 80 | case NEON::BI__builtin_neon_vld1q_lane_v: { |
6881 | 80 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
6882 | 80 | Ty = llvm::PointerType::getUnqual(VTy->getElementType()); |
6883 | 80 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
6884 | 80 | auto Alignment = CharUnits::fromQuantity( |
6885 | 80 | BuiltinID == NEON::BI__builtin_neon_vld1_lane_v ? 840 : 1640 ); |
6886 | 80 | Ops[0] = |
6887 | 80 | Builder.CreateAlignedLoad(VTy->getElementType(), Ops[0], Alignment); |
6888 | 80 | return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vld1_lane"); |
6889 | 80 | } |
6890 | 80 | case NEON::BI__builtin_neon_vld1_dup_v: |
6891 | 80 | case NEON::BI__builtin_neon_vld1q_dup_v: { |
6892 | 80 | Value *V = UndefValue::get(Ty); |
6893 | 80 | Ty = llvm::PointerType::getUnqual(VTy->getElementType()); |
6894 | 80 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
6895 | 80 | auto Alignment = CharUnits::fromQuantity( |
6896 | 80 | BuiltinID == NEON::BI__builtin_neon_vld1_dup_v ? 840 : 1640 ); |
6897 | 80 | Ops[0] = |
6898 | 80 | Builder.CreateAlignedLoad(VTy->getElementType(), Ops[0], Alignment); |
6899 | 80 | llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); |
6900 | 80 | Ops[0] = Builder.CreateInsertElement(V, Ops[0], CI); |
6901 | 80 | return EmitNeonSplat(Ops[0], CI); |
6902 | 80 | } |
6903 | 80 | case NEON::BI__builtin_neon_vst1_lane_v: |
6904 | 80 | case NEON::BI__builtin_neon_vst1q_lane_v: |
6905 | 80 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
6906 | 80 | Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); |
6907 | 80 | Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); |
6908 | 80 | return Builder.CreateDefaultAlignedStore(Ops[1], |
6909 | 80 | Builder.CreateBitCast(Ops[0], Ty)); |
6910 | 80 | case NEON::BI__builtin_neon_vld2_v: |
6911 | 80 | case NEON::BI__builtin_neon_vld2q_v: { |
6912 | 80 | llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); |
6913 | 80 | Ops[1] = Builder.CreateBitCast(Ops[1], PTy); |
6914 | 80 | llvm::Type *Tys[2] = { VTy, PTy }; |
6915 | 80 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2, Tys); |
6916 | 80 | Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); |
6917 | 80 | Ops[0] = Builder.CreateBitCast(Ops[0], |
6918 | 80 | llvm::PointerType::getUnqual(Ops[1]->getType())); |
6919 | 80 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
6920 | 80 | } |
6921 | 80 | case NEON::BI__builtin_neon_vld3_v: |
6922 | 80 | case NEON::BI__builtin_neon_vld3q_v: { |
6923 | 80 | llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); |
6924 | 80 | Ops[1] = Builder.CreateBitCast(Ops[1], PTy); |
6925 | 80 | llvm::Type *Tys[2] = { VTy, PTy }; |
6926 | 80 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3, Tys); |
6927 | 80 | Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); |
6928 | 80 | Ops[0] = Builder.CreateBitCast(Ops[0], |
6929 | 80 | llvm::PointerType::getUnqual(Ops[1]->getType())); |
6930 | 80 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
6931 | 80 | } |
6932 | 80 | case NEON::BI__builtin_neon_vld4_v: |
6933 | 80 | case NEON::BI__builtin_neon_vld4q_v: { |
6934 | 80 | llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); |
6935 | 80 | Ops[1] = Builder.CreateBitCast(Ops[1], PTy); |
6936 | 80 | llvm::Type *Tys[2] = { VTy, PTy }; |
6937 | 80 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4, Tys); |
6938 | 80 | Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); |
6939 | 80 | Ops[0] = Builder.CreateBitCast(Ops[0], |
6940 | 80 | llvm::PointerType::getUnqual(Ops[1]->getType())); |
6941 | 80 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
6942 | 80 | } |
6943 | 80 | case NEON::BI__builtin_neon_vld2_dup_v: |
6944 | 80 | case NEON::BI__builtin_neon_vld2q_dup_v: { |
6945 | 80 | llvm::Type *PTy = |
6946 | 80 | llvm::PointerType::getUnqual(VTy->getElementType()); |
6947 | 80 | Ops[1] = Builder.CreateBitCast(Ops[1], PTy); |
6948 | 80 | llvm::Type *Tys[2] = { VTy, PTy }; |
6949 | 80 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2r, Tys); |
6950 | 80 | Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); |
6951 | 80 | Ops[0] = Builder.CreateBitCast(Ops[0], |
6952 | 80 | llvm::PointerType::getUnqual(Ops[1]->getType())); |
6953 | 80 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
6954 | 80 | } |
6955 | 80 | case NEON::BI__builtin_neon_vld3_dup_v: |
6956 | 80 | case NEON::BI__builtin_neon_vld3q_dup_v: { |
6957 | 80 | llvm::Type *PTy = |
6958 | 80 | llvm::PointerType::getUnqual(VTy->getElementType()); |
6959 | 80 | Ops[1] = Builder.CreateBitCast(Ops[1], PTy); |
6960 | 80 | llvm::Type *Tys[2] = { VTy, PTy }; |
6961 | 80 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3r, Tys); |
6962 | 80 | Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); |
6963 | 80 | Ops[0] = Builder.CreateBitCast(Ops[0], |
6964 | 80 | llvm::PointerType::getUnqual(Ops[1]->getType())); |
6965 | 80 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
6966 | 80 | } |
6967 | 80 | case NEON::BI__builtin_neon_vld4_dup_v: |
6968 | 80 | case NEON::BI__builtin_neon_vld4q_dup_v: { |
6969 | 80 | llvm::Type *PTy = |
6970 | 80 | llvm::PointerType::getUnqual(VTy->getElementType()); |
6971 | 80 | Ops[1] = Builder.CreateBitCast(Ops[1], PTy); |
6972 | 80 | llvm::Type *Tys[2] = { VTy, PTy }; |
6973 | 80 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4r, Tys); |
6974 | 80 | Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); |
6975 | 80 | Ops[0] = Builder.CreateBitCast(Ops[0], |
6976 | 80 | llvm::PointerType::getUnqual(Ops[1]->getType())); |
6977 | 80 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
6978 | 80 | } |
6979 | 80 | case NEON::BI__builtin_neon_vld2_lane_v: |
6980 | 80 | case NEON::BI__builtin_neon_vld2q_lane_v: { |
6981 | 80 | llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; |
6982 | 80 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2lane, Tys); |
6983 | 80 | Ops.push_back(Ops[1]); |
6984 | 80 | Ops.erase(Ops.begin()+1); |
6985 | 80 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
6986 | 80 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
6987 | 80 | Ops[3] = Builder.CreateZExt(Ops[3], Int64Ty); |
6988 | 80 | Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld2_lane"); |
6989 | 80 | Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); |
6990 | 80 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
6991 | 80 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
6992 | 80 | } |
6993 | 80 | case NEON::BI__builtin_neon_vld3_lane_v: |
6994 | 80 | case NEON::BI__builtin_neon_vld3q_lane_v: { |
6995 | 80 | llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; |
6996 | 80 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3lane, Tys); |
6997 | 80 | Ops.push_back(Ops[1]); |
6998 | 80 | Ops.erase(Ops.begin()+1); |
6999 | 80 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
7000 | 80 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
7001 | 80 | Ops[3] = Builder.CreateBitCast(Ops[3], Ty); |
7002 | 80 | Ops[4] = Builder.CreateZExt(Ops[4], Int64Ty); |
7003 | 80 | Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane"); |
7004 | 80 | Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); |
7005 | 80 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
7006 | 80 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
7007 | 80 | } |
7008 | 80 | case NEON::BI__builtin_neon_vld4_lane_v: |
7009 | 80 | case NEON::BI__builtin_neon_vld4q_lane_v: { |
7010 | 80 | llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; |
7011 | 80 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4lane, Tys); |
7012 | 80 | Ops.push_back(Ops[1]); |
7013 | 80 | Ops.erase(Ops.begin()+1); |
7014 | 80 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
7015 | 80 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
7016 | 80 | Ops[3] = Builder.CreateBitCast(Ops[3], Ty); |
7017 | 80 | Ops[4] = Builder.CreateBitCast(Ops[4], Ty); |
7018 | 80 | Ops[5] = Builder.CreateZExt(Ops[5], Int64Ty); |
7019 | 80 | Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld4_lane"); |
7020 | 80 | Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); |
7021 | 80 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
7022 | 80 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
7023 | 80 | } |
7024 | 80 | case NEON::BI__builtin_neon_vst2_v: |
7025 | 80 | case NEON::BI__builtin_neon_vst2q_v: { |
7026 | 80 | Ops.push_back(Ops[0]); |
7027 | 80 | Ops.erase(Ops.begin()); |
7028 | 80 | llvm::Type *Tys[2] = { VTy, Ops[2]->getType() }; |
7029 | 80 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2, Tys), |
7030 | 80 | Ops, ""); |
7031 | 80 | } |
7032 | 80 | case NEON::BI__builtin_neon_vst2_lane_v: |
7033 | 80 | case NEON::BI__builtin_neon_vst2q_lane_v: { |
7034 | 80 | Ops.push_back(Ops[0]); |
7035 | 80 | Ops.erase(Ops.begin()); |
7036 | 80 | Ops[2] = Builder.CreateZExt(Ops[2], Int64Ty); |
7037 | 80 | llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; |
7038 | 80 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2lane, Tys), |
7039 | 80 | Ops, ""); |
7040 | 80 | } |
7041 | 80 | case NEON::BI__builtin_neon_vst3_v: |
7042 | 80 | case NEON::BI__builtin_neon_vst3q_v: { |
7043 | 80 | Ops.push_back(Ops[0]); |
7044 | 80 | Ops.erase(Ops.begin()); |
7045 | 80 | llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; |
7046 | 80 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3, Tys), |
7047 | 80 | Ops, ""); |
7048 | 80 | } |
7049 | 80 | case NEON::BI__builtin_neon_vst3_lane_v: |
7050 | 80 | case NEON::BI__builtin_neon_vst3q_lane_v: { |
7051 | 80 | Ops.push_back(Ops[0]); |
7052 | 80 | Ops.erase(Ops.begin()); |
7053 | 80 | Ops[3] = Builder.CreateZExt(Ops[3], Int64Ty); |
7054 | 80 | llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; |
7055 | 80 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3lane, Tys), |
7056 | 80 | Ops, ""); |
7057 | 80 | } |
7058 | 80 | case NEON::BI__builtin_neon_vst4_v: |
7059 | 80 | case NEON::BI__builtin_neon_vst4q_v: { |
7060 | 80 | Ops.push_back(Ops[0]); |
7061 | 80 | Ops.erase(Ops.begin()); |
7062 | 80 | llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; |
7063 | 80 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4, Tys), |
7064 | 80 | Ops, ""); |
7065 | 80 | } |
7066 | 80 | case NEON::BI__builtin_neon_vst4_lane_v: |
7067 | 80 | case NEON::BI__builtin_neon_vst4q_lane_v: { |
7068 | 80 | Ops.push_back(Ops[0]); |
7069 | 80 | Ops.erase(Ops.begin()); |
7070 | 80 | Ops[4] = Builder.CreateZExt(Ops[4], Int64Ty); |
7071 | 80 | llvm::Type *Tys[2] = { VTy, Ops[5]->getType() }; |
7072 | 80 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4lane, Tys), |
7073 | 80 | Ops, ""); |
7074 | 80 | } |
7075 | 54 | case NEON::BI__builtin_neon_vtrn_v: |
7076 | 54 | case NEON::BI__builtin_neon_vtrnq_v: { |
7077 | 54 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); |
7078 | 54 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
7079 | 54 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
7080 | 54 | Value *SV = nullptr; |
7081 | 54 | |
7082 | 162 | for (unsigned vi = 0; vi != 2162 ; ++vi108 ) { |
7083 | 108 | SmallVector<uint32_t, 16> Indices; |
7084 | 486 | for (unsigned i = 0, e = VTy->getNumElements(); i != e486 ; i += 2378 ) { |
7085 | 378 | Indices.push_back(i+vi); |
7086 | 378 | Indices.push_back(i+e+vi); |
7087 | 378 | } |
7088 | 108 | Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); |
7089 | 108 | SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vtrn"); |
7090 | 108 | SV = Builder.CreateDefaultAlignedStore(SV, Addr); |
7091 | 108 | } |
7092 | 54 | return SV; |
7093 | 54 | } |
7094 | 60 | case NEON::BI__builtin_neon_vuzp_v: |
7095 | 60 | case NEON::BI__builtin_neon_vuzpq_v: { |
7096 | 60 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); |
7097 | 60 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
7098 | 60 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
7099 | 60 | Value *SV = nullptr; |
7100 | 60 | |
7101 | 180 | for (unsigned vi = 0; vi != 2180 ; ++vi120 ) { |
7102 | 120 | SmallVector<uint32_t, 16> Indices; |
7103 | 988 | for (unsigned i = 0, e = VTy->getNumElements(); i != e988 ; ++i868 ) |
7104 | 868 | Indices.push_back(2*i+vi); |
7105 | 120 | |
7106 | 120 | Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); |
7107 | 120 | SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vuzp"); |
7108 | 120 | SV = Builder.CreateDefaultAlignedStore(SV, Addr); |
7109 | 120 | } |
7110 | 60 | return SV; |
7111 | 60 | } |
7112 | 61 | case NEON::BI__builtin_neon_vzip_v: |
7113 | 61 | case NEON::BI__builtin_neon_vzipq_v: { |
7114 | 61 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); |
7115 | 61 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
7116 | 61 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
7117 | 61 | Value *SV = nullptr; |
7118 | 61 | |
7119 | 183 | for (unsigned vi = 0; vi != 2183 ; ++vi122 ) { |
7120 | 122 | SmallVector<uint32_t, 16> Indices; |
7121 | 572 | for (unsigned i = 0, e = VTy->getNumElements(); i != e572 ; i += 2450 ) { |
7122 | 450 | Indices.push_back((i + vi*e) >> 1); |
7123 | 450 | Indices.push_back(((i + vi*e) >> 1)+e); |
7124 | 450 | } |
7125 | 122 | Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); |
7126 | 122 | SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vzip"); |
7127 | 122 | SV = Builder.CreateDefaultAlignedStore(SV, Addr); |
7128 | 122 | } |
7129 | 61 | return SV; |
7130 | 61 | } |
7131 | 0 | case NEON::BI__builtin_neon_vqtbl1q_v: { |
7132 | 0 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl1, Ty), |
7133 | 0 | Ops, "vtbl1"); |
7134 | 61 | } |
7135 | 0 | case NEON::BI__builtin_neon_vqtbl2q_v: { |
7136 | 0 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl2, Ty), |
7137 | 0 | Ops, "vtbl2"); |
7138 | 61 | } |
7139 | 0 | case NEON::BI__builtin_neon_vqtbl3q_v: { |
7140 | 0 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl3, Ty), |
7141 | 0 | Ops, "vtbl3"); |
7142 | 61 | } |
7143 | 0 | case NEON::BI__builtin_neon_vqtbl4q_v: { |
7144 | 0 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl4, Ty), |
7145 | 0 | Ops, "vtbl4"); |
7146 | 61 | } |
7147 | 0 | case NEON::BI__builtin_neon_vqtbx1q_v: { |
7148 | 0 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx1, Ty), |
7149 | 0 | Ops, "vtbx1"); |
7150 | 61 | } |
7151 | 0 | case NEON::BI__builtin_neon_vqtbx2q_v: { |
7152 | 0 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx2, Ty), |
7153 | 0 | Ops, "vtbx2"); |
7154 | 61 | } |
7155 | 0 | case NEON::BI__builtin_neon_vqtbx3q_v: { |
7156 | 0 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx3, Ty), |
7157 | 0 | Ops, "vtbx3"); |
7158 | 61 | } |
7159 | 0 | case NEON::BI__builtin_neon_vqtbx4q_v: { |
7160 | 0 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx4, Ty), |
7161 | 0 | Ops, "vtbx4"); |
7162 | 61 | } |
7163 | 24 | case NEON::BI__builtin_neon_vsqadd_v: |
7164 | 24 | case NEON::BI__builtin_neon_vsqaddq_v: { |
7165 | 24 | Int = Intrinsic::aarch64_neon_usqadd; |
7166 | 24 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqadd"); |
7167 | 24 | } |
7168 | 24 | case NEON::BI__builtin_neon_vuqadd_v: |
7169 | 24 | case NEON::BI__builtin_neon_vuqaddq_v: { |
7170 | 24 | Int = Intrinsic::aarch64_neon_suqadd; |
7171 | 24 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vuqadd"); |
7172 | 0 | } |
7173 | 17.3k | } |
7174 | 17.3k | } |
7175 | | |
7176 | | llvm::Value *CodeGenFunction:: |
7177 | 8 | BuildVector(ArrayRef<llvm::Value*> Ops) { |
7178 | 8 | assert((Ops.size() & (Ops.size() - 1)) == 0 && |
7179 | 8 | "Not a power-of-two sized vector!"); |
7180 | 8 | bool AllConstants = true; |
7181 | 16 | for (unsigned i = 0, e = Ops.size(); i != e && 16 AllConstants16 ; ++i8 ) |
7182 | 8 | AllConstants &= isa<Constant>(Ops[i]); |
7183 | 8 | |
7184 | 8 | // If this is a constant vector, create a ConstantVector. |
7185 | 8 | if (AllConstants8 ) { |
7186 | 0 | SmallVector<llvm::Constant*, 16> CstOps; |
7187 | 0 | for (unsigned i = 0, e = Ops.size(); i != e0 ; ++i0 ) |
7188 | 0 | CstOps.push_back(cast<Constant>(Ops[i])); |
7189 | 0 | return llvm::ConstantVector::get(CstOps); |
7190 | 0 | } |
7191 | 8 | |
7192 | 8 | // Otherwise, insertelement the values to build the vector. |
7193 | 8 | Value *Result = |
7194 | 8 | llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size())); |
7195 | 8 | |
7196 | 40 | for (unsigned i = 0, e = Ops.size(); i != e40 ; ++i32 ) |
7197 | 32 | Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt32(i)); |
7198 | 8 | |
7199 | 8 | return Result; |
7200 | 8 | } |
7201 | | |
7202 | | // Convert the mask from an integer type to a vector of i1. |
7203 | | static Value *getMaskVecValue(CodeGenFunction &CGF, Value *Mask, |
7204 | 1.78k | unsigned NumElts) { |
7205 | 1.78k | |
7206 | 1.78k | llvm::VectorType *MaskTy = llvm::VectorType::get(CGF.Builder.getInt1Ty(), |
7207 | 1.78k | cast<IntegerType>(Mask->getType())->getBitWidth()); |
7208 | 1.78k | Value *MaskVec = CGF.Builder.CreateBitCast(Mask, MaskTy); |
7209 | 1.78k | |
7210 | 1.78k | // If we have less than 8 elements, then the starting mask was an i8 and |
7211 | 1.78k | // we need to extract down to the right number of elements. |
7212 | 1.78k | if (NumElts < 81.78k ) { |
7213 | 444 | uint32_t Indices[4]; |
7214 | 1.93k | for (unsigned i = 0; i != NumElts1.93k ; ++i1.48k ) |
7215 | 1.48k | Indices[i] = i; |
7216 | 444 | MaskVec = CGF.Builder.CreateShuffleVector(MaskVec, MaskVec, |
7217 | 444 | makeArrayRef(Indices, NumElts), |
7218 | 444 | "extract"); |
7219 | 444 | } |
7220 | 1.78k | return MaskVec; |
7221 | 1.78k | } |
7222 | | |
7223 | | static Value *EmitX86MaskedStore(CodeGenFunction &CGF, |
7224 | | SmallVectorImpl<Value *> &Ops, |
7225 | 39 | unsigned Align) { |
7226 | 39 | // Cast the pointer to right type. |
7227 | 39 | Ops[0] = CGF.Builder.CreateBitCast(Ops[0], |
7228 | 39 | llvm::PointerType::getUnqual(Ops[1]->getType())); |
7229 | 39 | |
7230 | 39 | // If the mask is all ones just emit a regular store. |
7231 | 39 | if (const auto *C = dyn_cast<Constant>(Ops[2])) |
7232 | 3 | if (3 C->isAllOnesValue()3 ) |
7233 | 3 | return CGF.Builder.CreateAlignedStore(Ops[1], Ops[0], Align); |
7234 | 36 | |
7235 | 36 | Value *MaskVec = getMaskVecValue(CGF, Ops[2], |
7236 | 36 | Ops[1]->getType()->getVectorNumElements()); |
7237 | 36 | |
7238 | 36 | return CGF.Builder.CreateMaskedStore(Ops[1], Ops[0], Align, MaskVec); |
7239 | 36 | } |
7240 | | |
7241 | | static Value *EmitX86MaskedLoad(CodeGenFunction &CGF, |
7242 | 75 | SmallVectorImpl<Value *> &Ops, unsigned Align) { |
7243 | 75 | // Cast the pointer to right type. |
7244 | 75 | Ops[0] = CGF.Builder.CreateBitCast(Ops[0], |
7245 | 75 | llvm::PointerType::getUnqual(Ops[1]->getType())); |
7246 | 75 | |
7247 | 75 | // If the mask is all ones just emit a regular store. |
7248 | 75 | if (const auto *C = dyn_cast<Constant>(Ops[2])) |
7249 | 3 | if (3 C->isAllOnesValue()3 ) |
7250 | 3 | return CGF.Builder.CreateAlignedLoad(Ops[0], Align); |
7251 | 72 | |
7252 | 72 | Value *MaskVec = getMaskVecValue(CGF, Ops[2], |
7253 | 72 | Ops[1]->getType()->getVectorNumElements()); |
7254 | 72 | |
7255 | 72 | return CGF.Builder.CreateMaskedLoad(Ops[0], Align, MaskVec, Ops[1]); |
7256 | 72 | } |
7257 | | |
7258 | | static Value *EmitX86SubVectorBroadcast(CodeGenFunction &CGF, |
7259 | | SmallVectorImpl<Value *> &Ops, |
7260 | | llvm::Type *DstTy, |
7261 | | unsigned SrcSizeInBits, |
7262 | 6 | unsigned Align) { |
7263 | 6 | // Load the subvector. |
7264 | 6 | Ops[0] = CGF.Builder.CreateAlignedLoad(Ops[0], Align); |
7265 | 6 | |
7266 | 6 | // Create broadcast mask. |
7267 | 6 | unsigned NumDstElts = DstTy->getVectorNumElements(); |
7268 | 6 | unsigned NumSrcElts = SrcSizeInBits / DstTy->getScalarSizeInBits(); |
7269 | 6 | |
7270 | 6 | SmallVector<uint32_t, 8> Mask; |
7271 | 18 | for (unsigned i = 0; i != NumDstElts18 ; i += NumSrcElts12 ) |
7272 | 48 | for (unsigned j = 0; 12 j != NumSrcElts48 ; ++j36 ) |
7273 | 36 | Mask.push_back(j); |
7274 | 6 | |
7275 | 6 | return CGF.Builder.CreateShuffleVector(Ops[0], Ops[0], Mask, "subvecbcst"); |
7276 | 6 | } |
7277 | | |
7278 | | static Value *EmitX86Select(CodeGenFunction &CGF, |
7279 | 1.46k | Value *Mask, Value *Op0, Value *Op1) { |
7280 | 1.46k | |
7281 | 1.46k | // If the mask is all ones just return first argument. |
7282 | 1.46k | if (const auto *C = dyn_cast<Constant>(Mask)) |
7283 | 53 | if (53 C->isAllOnesValue()53 ) |
7284 | 53 | return Op0; |
7285 | 1.41k | |
7286 | 1.41k | Mask = getMaskVecValue(CGF, Mask, Op0->getType()->getVectorNumElements()); |
7287 | 1.41k | |
7288 | 1.41k | return CGF.Builder.CreateSelect(Mask, Op0, Op1); |
7289 | 1.41k | } |
7290 | | |
7291 | | static Value *EmitX86MaskedCompare(CodeGenFunction &CGF, unsigned CC, |
7292 | 500 | bool Signed, SmallVectorImpl<Value *> &Ops) { |
7293 | 500 | unsigned NumElts = Ops[0]->getType()->getVectorNumElements(); |
7294 | 500 | Value *Cmp; |
7295 | 500 | |
7296 | 500 | if (CC == 3500 ) { |
7297 | 0 | Cmp = Constant::getNullValue( |
7298 | 0 | llvm::VectorType::get(CGF.Builder.getInt1Ty(), NumElts)); |
7299 | 500 | } else if (500 CC == 7500 ) { |
7300 | 0 | Cmp = Constant::getAllOnesValue( |
7301 | 0 | llvm::VectorType::get(CGF.Builder.getInt1Ty(), NumElts)); |
7302 | 500 | } else { |
7303 | 500 | ICmpInst::Predicate Pred; |
7304 | 500 | switch (CC) { |
7305 | 0 | default: 0 llvm_unreachable0 ("Unknown condition code"); |
7306 | 137 | case 0: Pred = ICmpInst::ICMP_EQ; break; |
7307 | 74 | case 1: Pred = Signed ? 74 ICmpInst::ICMP_SLT38 : ICmpInst::ICMP_ULT36 ; break; |
7308 | 73 | case 2: Pred = Signed ? 73 ICmpInst::ICMP_SLE37 : ICmpInst::ICMP_ULE36 ; break; |
7309 | 72 | case 4: Pred = ICmpInst::ICMP_NE; break; |
7310 | 72 | case 5: Pred = Signed ? 72 ICmpInst::ICMP_SGE36 : ICmpInst::ICMP_UGE36 ; break; |
7311 | 72 | case 6: Pred = Signed ? 72 ICmpInst::ICMP_SGT36 : ICmpInst::ICMP_UGT36 ; break; |
7312 | 500 | } |
7313 | 500 | Cmp = CGF.Builder.CreateICmp(Pred, Ops[0], Ops[1]); |
7314 | 500 | } |
7315 | 500 | |
7316 | 500 | const auto *C = dyn_cast<Constant>(Ops.back()); |
7317 | 500 | if (!C || 500 !C->isAllOnesValue()250 ) |
7318 | 250 | Cmp = CGF.Builder.CreateAnd(Cmp, getMaskVecValue(CGF, Ops.back(), NumElts)); |
7319 | 500 | |
7320 | 500 | if (NumElts < 8500 ) { |
7321 | 82 | uint32_t Indices[8]; |
7322 | 354 | for (unsigned i = 0; i != NumElts354 ; ++i272 ) |
7323 | 272 | Indices[i] = i; |
7324 | 466 | for (unsigned i = NumElts; i != 8466 ; ++i384 ) |
7325 | 384 | Indices[i] = i % NumElts + NumElts; |
7326 | 82 | Cmp = CGF.Builder.CreateShuffleVector( |
7327 | 82 | Cmp, llvm::Constant::getNullValue(Cmp->getType()), Indices); |
7328 | 82 | } |
7329 | 500 | return CGF.Builder.CreateBitCast(Cmp, |
7330 | 500 | IntegerType::get(CGF.getLLVMContext(), |
7331 | 500 | std::max(NumElts, 8U))); |
7332 | 500 | } |
7333 | | |
7334 | 44 | static Value *EmitX86Abs(CodeGenFunction &CGF, ArrayRef<Value *> Ops) { |
7335 | 44 | |
7336 | 44 | llvm::Type *Ty = Ops[0]->getType(); |
7337 | 44 | Value *Zero = llvm::Constant::getNullValue(Ty); |
7338 | 44 | Value *Sub = CGF.Builder.CreateSub(Zero, Ops[0]); |
7339 | 44 | Value *Cmp = CGF.Builder.CreateICmp(ICmpInst::ICMP_SGT, Ops[0], Zero); |
7340 | 44 | Value *Res = CGF.Builder.CreateSelect(Cmp, Ops[0], Sub); |
7341 | 44 | if (Ops.size() == 1) |
7342 | 22 | return Res; |
7343 | 22 | return EmitX86Select(CGF, Ops[2], Res, Ops[1]); |
7344 | 22 | } |
7345 | | |
7346 | | static Value *EmitX86MinMax(CodeGenFunction &CGF, ICmpInst::Predicate Pred, |
7347 | 195 | ArrayRef<Value *> Ops) { |
7348 | 195 | Value *Cmp = CGF.Builder.CreateICmp(Pred, Ops[0], Ops[1]); |
7349 | 195 | Value *Res = CGF.Builder.CreateSelect(Cmp, Ops[0], Ops[1]); |
7350 | 195 | |
7351 | 195 | if (Ops.size() == 2) |
7352 | 92 | return Res; |
7353 | 103 | |
7354 | 195 | assert(Ops.size() == 4); |
7355 | 103 | return EmitX86Select(CGF, Ops[3], Res, Ops[2]); |
7356 | 103 | } |
7357 | | |
7358 | | static Value *EmitX86SExtMask(CodeGenFunction &CGF, Value *Op, |
7359 | 18 | llvm::Type *DstTy) { |
7360 | 18 | unsigned NumberOfElements = DstTy->getVectorNumElements(); |
7361 | 18 | Value *Mask = getMaskVecValue(CGF, Op, NumberOfElements); |
7362 | 18 | return CGF.Builder.CreateSExt(Mask, DstTy, "vpmovm2"); |
7363 | 18 | } |
7364 | | |
7365 | 35 | Value *CodeGenFunction::EmitX86CpuIs(const CallExpr *E) { |
7366 | 35 | const Expr *CPUExpr = E->getArg(0)->IgnoreParenCasts(); |
7367 | 35 | StringRef CPUStr = cast<clang::StringLiteral>(CPUExpr)->getString(); |
7368 | 35 | return EmitX86CpuIs(CPUStr); |
7369 | 35 | } |
7370 | | |
7371 | 35 | Value *CodeGenFunction::EmitX86CpuIs(StringRef CPUStr) { |
7372 | 35 | |
7373 | 35 | // This enum contains the vendor, type, and subtype enums from the |
7374 | 35 | // runtime library concatenated together. The _START labels mark |
7375 | 35 | // the start and are used to adjust the value into the correct |
7376 | 35 | // encoding space. |
7377 | 35 | enum X86CPUs { |
7378 | 35 | INTEL = 1, |
7379 | 35 | AMD, |
7380 | 35 | CPU_TYPE_START, |
7381 | 35 | INTEL_BONNELL, |
7382 | 35 | INTEL_CORE2, |
7383 | 35 | INTEL_COREI7, |
7384 | 35 | AMDFAM10H, |
7385 | 35 | AMDFAM15H, |
7386 | 35 | INTEL_SILVERMONT, |
7387 | 35 | INTEL_KNL, |
7388 | 35 | AMD_BTVER1, |
7389 | 35 | AMD_BTVER2, |
7390 | 35 | CPU_SUBTYPE_START, |
7391 | 35 | INTEL_COREI7_NEHALEM, |
7392 | 35 | INTEL_COREI7_WESTMERE, |
7393 | 35 | INTEL_COREI7_SANDYBRIDGE, |
7394 | 35 | AMDFAM10H_BARCELONA, |
7395 | 35 | AMDFAM10H_SHANGHAI, |
7396 | 35 | AMDFAM10H_ISTANBUL, |
7397 | 35 | AMDFAM15H_BDVER1, |
7398 | 35 | AMDFAM15H_BDVER2, |
7399 | 35 | AMDFAM15H_BDVER3, |
7400 | 35 | AMDFAM15H_BDVER4, |
7401 | 35 | AMDFAM17H_ZNVER1, |
7402 | 35 | INTEL_COREI7_IVYBRIDGE, |
7403 | 35 | INTEL_COREI7_HASWELL, |
7404 | 35 | INTEL_COREI7_BROADWELL, |
7405 | 35 | INTEL_COREI7_SKYLAKE, |
7406 | 35 | INTEL_COREI7_SKYLAKE_AVX512, |
7407 | 35 | }; |
7408 | 35 | |
7409 | 35 | X86CPUs CPU = |
7410 | 35 | StringSwitch<X86CPUs>(CPUStr) |
7411 | 35 | .Case("amd", AMD) |
7412 | 35 | .Case("amdfam10h", AMDFAM10H) |
7413 | 35 | .Case("amdfam10", AMDFAM10H) |
7414 | 35 | .Case("amdfam15h", AMDFAM15H) |
7415 | 35 | .Case("amdfam15", AMDFAM15H) |
7416 | 35 | .Case("atom", INTEL_BONNELL) |
7417 | 35 | .Case("barcelona", AMDFAM10H_BARCELONA) |
7418 | 35 | .Case("bdver1", AMDFAM15H_BDVER1) |
7419 | 35 | .Case("bdver2", AMDFAM15H_BDVER2) |
7420 | 35 | .Case("bdver3", AMDFAM15H_BDVER3) |
7421 | 35 | .Case("bdver4", AMDFAM15H_BDVER4) |
7422 | 35 | .Case("bonnell", INTEL_BONNELL) |
7423 | 35 | .Case("broadwell", INTEL_COREI7_BROADWELL) |
7424 | 35 | .Case("btver1", AMD_BTVER1) |
7425 | 35 | .Case("btver2", AMD_BTVER2) |
7426 | 35 | .Case("core2", INTEL_CORE2) |
7427 | 35 | .Case("corei7", INTEL_COREI7) |
7428 | 35 | .Case("haswell", INTEL_COREI7_HASWELL) |
7429 | 35 | .Case("intel", INTEL) |
7430 | 35 | .Case("istanbul", AMDFAM10H_ISTANBUL) |
7431 | 35 | .Case("ivybridge", INTEL_COREI7_IVYBRIDGE) |
7432 | 35 | .Case("knl", INTEL_KNL) |
7433 | 35 | .Case("nehalem", INTEL_COREI7_NEHALEM) |
7434 | 35 | .Case("sandybridge", INTEL_COREI7_SANDYBRIDGE) |
7435 | 35 | .Case("shanghai", AMDFAM10H_SHANGHAI) |
7436 | 35 | .Case("silvermont", INTEL_SILVERMONT) |
7437 | 35 | .Case("skylake", INTEL_COREI7_SKYLAKE) |
7438 | 35 | .Case("skylake-avx512", INTEL_COREI7_SKYLAKE_AVX512) |
7439 | 35 | .Case("slm", INTEL_SILVERMONT) |
7440 | 35 | .Case("westmere", INTEL_COREI7_WESTMERE) |
7441 | 35 | .Case("znver1", AMDFAM17H_ZNVER1); |
7442 | 35 | |
7443 | 35 | llvm::Type *Int32Ty = Builder.getInt32Ty(); |
7444 | 35 | |
7445 | 35 | // Matching the struct layout from the compiler-rt/libgcc structure that is |
7446 | 35 | // filled in: |
7447 | 35 | // unsigned int __cpu_vendor; |
7448 | 35 | // unsigned int __cpu_type; |
7449 | 35 | // unsigned int __cpu_subtype; |
7450 | 35 | // unsigned int __cpu_features[1]; |
7451 | 35 | llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, Int32Ty, |
7452 | 35 | llvm::ArrayType::get(Int32Ty, 1)); |
7453 | 35 | |
7454 | 35 | // Grab the global __cpu_model. |
7455 | 35 | llvm::Constant *CpuModel = CGM.CreateRuntimeVariable(STy, "__cpu_model"); |
7456 | 35 | |
7457 | 35 | // Calculate the index needed to access the correct field based on the |
7458 | 35 | // range. Also adjust the expected value. |
7459 | 35 | unsigned Index; |
7460 | 35 | unsigned Value; |
7461 | 35 | if (CPU > CPU_SUBTYPE_START35 ) { |
7462 | 18 | Index = 2; |
7463 | 18 | Value = CPU - CPU_SUBTYPE_START; |
7464 | 35 | } else if (17 CPU > CPU_TYPE_START17 ) { |
7465 | 13 | Index = 1; |
7466 | 13 | Value = CPU - CPU_TYPE_START; |
7467 | 17 | } else { |
7468 | 4 | Index = 0; |
7469 | 4 | Value = CPU; |
7470 | 4 | } |
7471 | 35 | |
7472 | 35 | // Grab the appropriate field from __cpu_model. |
7473 | 35 | llvm::Value *Idxs[] = { |
7474 | 35 | ConstantInt::get(Int32Ty, 0), |
7475 | 35 | ConstantInt::get(Int32Ty, Index) |
7476 | 35 | }; |
7477 | 35 | llvm::Value *CpuValue = Builder.CreateGEP(STy, CpuModel, Idxs); |
7478 | 35 | CpuValue = Builder.CreateAlignedLoad(CpuValue, CharUnits::fromQuantity(4)); |
7479 | 35 | |
7480 | 35 | // Check the value of the field against the requested value. |
7481 | 35 | return Builder.CreateICmpEQ(CpuValue, |
7482 | 35 | llvm::ConstantInt::get(Int32Ty, Value)); |
7483 | 35 | } |
7484 | | |
7485 | 34 | Value *CodeGenFunction::EmitX86CpuSupports(const CallExpr *E) { |
7486 | 34 | const Expr *FeatureExpr = E->getArg(0)->IgnoreParenCasts(); |
7487 | 34 | StringRef FeatureStr = cast<StringLiteral>(FeatureExpr)->getString(); |
7488 | 34 | return EmitX86CpuSupports(FeatureStr); |
7489 | 34 | } |
7490 | | |
7491 | 34 | Value *CodeGenFunction::EmitX86CpuSupports(ArrayRef<StringRef> FeatureStrs) { |
7492 | 34 | // TODO: When/if this becomes more than x86 specific then use a TargetInfo |
7493 | 34 | // based mapping. |
7494 | 34 | // Processor features and mapping to processor feature value. |
7495 | 34 | enum X86Features { |
7496 | 34 | CMOV = 0, |
7497 | 34 | MMX, |
7498 | 34 | POPCNT, |
7499 | 34 | SSE, |
7500 | 34 | SSE2, |
7501 | 34 | SSE3, |
7502 | 34 | SSSE3, |
7503 | 34 | SSE4_1, |
7504 | 34 | SSE4_2, |
7505 | 34 | AVX, |
7506 | 34 | AVX2, |
7507 | 34 | SSE4_A, |
7508 | 34 | FMA4, |
7509 | 34 | XOP, |
7510 | 34 | FMA, |
7511 | 34 | AVX512F, |
7512 | 34 | BMI, |
7513 | 34 | BMI2, |
7514 | 34 | AES, |
7515 | 34 | PCLMUL, |
7516 | 34 | AVX512VL, |
7517 | 34 | AVX512BW, |
7518 | 34 | AVX512DQ, |
7519 | 34 | AVX512CD, |
7520 | 34 | AVX512ER, |
7521 | 34 | AVX512PF, |
7522 | 34 | AVX512VBMI, |
7523 | 34 | AVX512IFMA, |
7524 | 34 | AVX5124VNNIW, |
7525 | 34 | AVX5124FMAPS, |
7526 | 34 | AVX512VPOPCNTDQ, |
7527 | 34 | MAX |
7528 | 34 | }; |
7529 | 34 | |
7530 | 34 | uint32_t FeaturesMask = 0; |
7531 | 34 | |
7532 | 34 | for (const StringRef &FeatureStr : FeatureStrs) { |
7533 | 34 | X86Features Feature = |
7534 | 34 | StringSwitch<X86Features>(FeatureStr) |
7535 | 34 | .Case("cmov", X86Features::CMOV) |
7536 | 34 | .Case("mmx", X86Features::MMX) |
7537 | 34 | .Case("popcnt", X86Features::POPCNT) |
7538 | 34 | .Case("sse", X86Features::SSE) |
7539 | 34 | .Case("sse2", X86Features::SSE2) |
7540 | 34 | .Case("sse3", X86Features::SSE3) |
7541 | 34 | .Case("ssse3", X86Features::SSSE3) |
7542 | 34 | .Case("sse4.1", X86Features::SSE4_1) |
7543 | 34 | .Case("sse4.2", X86Features::SSE4_2) |
7544 | 34 | .Case("avx", X86Features::AVX) |
7545 | 34 | .Case("avx2", X86Features::AVX2) |
7546 | 34 | .Case("sse4a", X86Features::SSE4_A) |
7547 | 34 | .Case("fma4", X86Features::FMA4) |
7548 | 34 | .Case("xop", X86Features::XOP) |
7549 | 34 | .Case("fma", X86Features::FMA) |
7550 | 34 | .Case("avx512f", X86Features::AVX512F) |
7551 | 34 | .Case("bmi", X86Features::BMI) |
7552 | 34 | .Case("bmi2", X86Features::BMI2) |
7553 | 34 | .Case("aes", X86Features::AES) |
7554 | 34 | .Case("pclmul", X86Features::PCLMUL) |
7555 | 34 | .Case("avx512vl", X86Features::AVX512VL) |
7556 | 34 | .Case("avx512bw", X86Features::AVX512BW) |
7557 | 34 | .Case("avx512dq", X86Features::AVX512DQ) |
7558 | 34 | .Case("avx512cd", X86Features::AVX512CD) |
7559 | 34 | .Case("avx512er", X86Features::AVX512ER) |
7560 | 34 | .Case("avx512pf", X86Features::AVX512PF) |
7561 | 34 | .Case("avx512vbmi", X86Features::AVX512VBMI) |
7562 | 34 | .Case("avx512ifma", X86Features::AVX512IFMA) |
7563 | 34 | .Case("avx5124vnniw", X86Features::AVX5124VNNIW) |
7564 | 34 | .Case("avx5124fmaps", X86Features::AVX5124FMAPS) |
7565 | 34 | .Case("avx512vpopcntdq", X86Features::AVX512VPOPCNTDQ) |
7566 | 34 | .Default(X86Features::MAX); |
7567 | 34 | assert(Feature != X86Features::MAX && "Invalid feature!"); |
7568 | 34 | FeaturesMask |= (1U << Feature); |
7569 | 34 | } |
7570 | 34 | |
7571 | 34 | // Matching the struct layout from the compiler-rt/libgcc structure that is |
7572 | 34 | // filled in: |
7573 | 34 | // unsigned int __cpu_vendor; |
7574 | 34 | // unsigned int __cpu_type; |
7575 | 34 | // unsigned int __cpu_subtype; |
7576 | 34 | // unsigned int __cpu_features[1]; |
7577 | 34 | llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, Int32Ty, |
7578 | 34 | llvm::ArrayType::get(Int32Ty, 1)); |
7579 | 34 | |
7580 | 34 | // Grab the global __cpu_model. |
7581 | 34 | llvm::Constant *CpuModel = CGM.CreateRuntimeVariable(STy, "__cpu_model"); |
7582 | 34 | |
7583 | 34 | // Grab the first (0th) element from the field __cpu_features off of the |
7584 | 34 | // global in the struct STy. |
7585 | 34 | Value *Idxs[] = {ConstantInt::get(Int32Ty, 0), ConstantInt::get(Int32Ty, 3), |
7586 | 34 | ConstantInt::get(Int32Ty, 0)}; |
7587 | 34 | Value *CpuFeatures = Builder.CreateGEP(STy, CpuModel, Idxs); |
7588 | 34 | Value *Features = |
7589 | 34 | Builder.CreateAlignedLoad(CpuFeatures, CharUnits::fromQuantity(4)); |
7590 | 34 | |
7591 | 34 | // Check the value of the bit corresponding to the feature requested. |
7592 | 34 | Value *Bitset = Builder.CreateAnd( |
7593 | 34 | Features, llvm::ConstantInt::get(Int32Ty, FeaturesMask)); |
7594 | 34 | return Builder.CreateICmpNE(Bitset, llvm::ConstantInt::get(Int32Ty, 0)); |
7595 | 34 | } |
7596 | | |
7597 | | Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID, |
7598 | 2.70k | const CallExpr *E) { |
7599 | 2.70k | if (BuiltinID == X86::BI__builtin_cpu_is) |
7600 | 35 | return EmitX86CpuIs(E); |
7601 | 2.67k | if (2.67k BuiltinID == X86::BI__builtin_cpu_supports2.67k ) |
7602 | 34 | return EmitX86CpuSupports(E); |
7603 | 2.63k | |
7604 | 2.63k | SmallVector<Value*, 4> Ops; |
7605 | 2.63k | |
7606 | 2.63k | // Find out if any arguments are required to be integer constant expressions. |
7607 | 2.63k | unsigned ICEArguments = 0; |
7608 | 2.63k | ASTContext::GetBuiltinTypeError Error; |
7609 | 2.63k | getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); |
7610 | 2.63k | assert(Error == ASTContext::GE_None && "Should not codegen an error"); |
7611 | 2.63k | |
7612 | 10.3k | for (unsigned i = 0, e = E->getNumArgs(); i != e10.3k ; i++7.69k ) { |
7613 | 7.69k | // If this is a normal argument, just emit it as a scalar. |
7614 | 7.69k | if ((ICEArguments & (1 << i)) == 07.69k ) { |
7615 | 7.17k | Ops.push_back(EmitScalarExpr(E->getArg(i))); |
7616 | 7.17k | continue; |
7617 | 7.17k | } |
7618 | 521 | |
7619 | 521 | // If this is required to be a constant, constant fold it so that we know |
7620 | 521 | // that the generated intrinsic gets a ConstantInt. |
7621 | 521 | llvm::APSInt Result; |
7622 | 521 | bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); |
7623 | 521 | assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; |
7624 | 521 | Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); |
7625 | 521 | } |
7626 | 2.63k | |
7627 | 2.63k | // These exist so that the builtin that takes an immediate can be bounds |
7628 | 2.63k | // checked by clang to avoid passing bad immediates to the backend. Since |
7629 | 2.63k | // AVX has a larger immediate than SSE we would need separate builtins to |
7630 | 2.63k | // do the different bounds checking. Rather than create a clang specific |
7631 | 2.63k | // SSE only builtin, this implements eight separate builtins to match gcc |
7632 | 2.63k | // implementation. |
7633 | 44 | auto getCmpIntrinsicCall = [this, &Ops](Intrinsic::ID ID, unsigned Imm) { |
7634 | 44 | Ops.push_back(llvm::ConstantInt::get(Int8Ty, Imm)); |
7635 | 44 | llvm::Function *F = CGM.getIntrinsic(ID); |
7636 | 44 | return Builder.CreateCall(F, Ops); |
7637 | 44 | }; |
7638 | 2.63k | |
7639 | 2.63k | // For the vector forms of FP comparisons, translate the builtins directly to |
7640 | 2.63k | // IR. |
7641 | 2.63k | // TODO: The builtins could be removed if the SSE header files used vector |
7642 | 2.63k | // extension comparisons directly (vector ordered/unordered may need |
7643 | 2.63k | // additional support via __builtin_isnan()). |
7644 | 66 | auto getVectorFCmpIR = [this, &Ops](CmpInst::Predicate Pred) { |
7645 | 66 | Value *Cmp = Builder.CreateFCmp(Pred, Ops[0], Ops[1]); |
7646 | 66 | llvm::VectorType *FPVecTy = cast<llvm::VectorType>(Ops[0]->getType()); |
7647 | 66 | llvm::VectorType *IntVecTy = llvm::VectorType::getInteger(FPVecTy); |
7648 | 66 | Value *Sext = Builder.CreateSExt(Cmp, IntVecTy); |
7649 | 66 | return Builder.CreateBitCast(Sext, FPVecTy); |
7650 | 66 | }; |
7651 | 2.63k | |
7652 | 2.63k | switch (BuiltinID) { |
7653 | 0 | default: return nullptr; |
7654 | 1 | case X86::BI__builtin_cpu_init: { |
7655 | 1 | llvm::FunctionType *FTy = llvm::FunctionType::get(VoidTy, |
7656 | 1 | /*Variadic*/false); |
7657 | 1 | llvm::Constant *Func = CGM.CreateRuntimeFunction(FTy, |
7658 | 1 | "__cpu_indicator_init"); |
7659 | 1 | return Builder.CreateCall(Func); |
7660 | 2.63k | } |
7661 | 0 | case X86::BI_mm_prefetch: { |
7662 | 0 | Value *Address = Ops[0]; |
7663 | 0 | Value *RW = ConstantInt::get(Int32Ty, 0); |
7664 | 0 | Value *Locality = Ops[1]; |
7665 | 0 | Value *Data = ConstantInt::get(Int32Ty, 1); |
7666 | 0 | Value *F = CGM.getIntrinsic(Intrinsic::prefetch); |
7667 | 0 | return Builder.CreateCall(F, {Address, RW, Locality, Data}); |
7668 | 2.63k | } |
7669 | 3 | case X86::BI_mm_clflush: { |
7670 | 3 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse2_clflush), |
7671 | 3 | Ops[0]); |
7672 | 2.63k | } |
7673 | 3 | case X86::BI_mm_lfence: { |
7674 | 3 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse2_lfence)); |
7675 | 2.63k | } |
7676 | 3 | case X86::BI_mm_mfence: { |
7677 | 3 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse2_mfence)); |
7678 | 2.63k | } |
7679 | 2 | case X86::BI_mm_sfence: { |
7680 | 2 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_sfence)); |
7681 | 2.63k | } |
7682 | 5 | case X86::BI_mm_pause: { |
7683 | 5 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse2_pause)); |
7684 | 2.63k | } |
7685 | 2 | case X86::BI__rdtsc: { |
7686 | 2 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_rdtsc)); |
7687 | 2.63k | } |
7688 | 73 | case X86::BI__builtin_ia32_undef128: |
7689 | 73 | case X86::BI__builtin_ia32_undef256: |
7690 | 73 | case X86::BI__builtin_ia32_undef512: |
7691 | 73 | // The x86 definition of "undef" is not the same as the LLVM definition |
7692 | 73 | // (PR32176). We leave optimizing away an unnecessary zero constant to the |
7693 | 73 | // IR optimizer and backend. |
7694 | 73 | // TODO: If we had a "freeze" IR instruction to generate a fixed undef |
7695 | 73 | // value, we should use that here instead of a zero. |
7696 | 73 | return llvm::Constant::getNullValue(ConvertType(E->getType())); |
7697 | 8 | case X86::BI__builtin_ia32_vec_init_v8qi: |
7698 | 8 | case X86::BI__builtin_ia32_vec_init_v4hi: |
7699 | 8 | case X86::BI__builtin_ia32_vec_init_v2si: |
7700 | 8 | return Builder.CreateBitCast(BuildVector(Ops), |
7701 | 8 | llvm::Type::getX86_MMXTy(getLLVMContext())); |
7702 | 4 | case X86::BI__builtin_ia32_vec_ext_v2si: |
7703 | 4 | return Builder.CreateExtractElement(Ops[0], |
7704 | 4 | llvm::ConstantInt::get(Ops[1]->getType(), 0)); |
7705 | 7 | case X86::BI_mm_setcsr: |
7706 | 7 | case X86::BI__builtin_ia32_ldmxcsr: { |
7707 | 7 | Address Tmp = CreateMemTemp(E->getArg(0)->getType()); |
7708 | 7 | Builder.CreateStore(Ops[0], Tmp); |
7709 | 7 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr), |
7710 | 7 | Builder.CreateBitCast(Tmp.getPointer(), Int8PtrTy)); |
7711 | 7 | } |
7712 | 11 | case X86::BI_mm_getcsr: |
7713 | 11 | case X86::BI__builtin_ia32_stmxcsr: { |
7714 | 11 | Address Tmp = CreateMemTemp(E->getType()); |
7715 | 11 | Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr), |
7716 | 11 | Builder.CreateBitCast(Tmp.getPointer(), Int8PtrTy)); |
7717 | 11 | return Builder.CreateLoad(Tmp, "stmxcsr"); |
7718 | 11 | } |
7719 | 48 | case X86::BI__builtin_ia32_xsave: |
7720 | 48 | case X86::BI__builtin_ia32_xsave64: |
7721 | 48 | case X86::BI__builtin_ia32_xrstor: |
7722 | 48 | case X86::BI__builtin_ia32_xrstor64: |
7723 | 48 | case X86::BI__builtin_ia32_xsaveopt: |
7724 | 48 | case X86::BI__builtin_ia32_xsaveopt64: |
7725 | 48 | case X86::BI__builtin_ia32_xrstors: |
7726 | 48 | case X86::BI__builtin_ia32_xrstors64: |
7727 | 48 | case X86::BI__builtin_ia32_xsavec: |
7728 | 48 | case X86::BI__builtin_ia32_xsavec64: |
7729 | 48 | case X86::BI__builtin_ia32_xsaves: |
7730 | 48 | case X86::BI__builtin_ia32_xsaves64: { |
7731 | 48 | Intrinsic::ID ID; |
7732 | 48 | #define INTRINSIC_X86_XSAVE_ID(NAME) \ |
7733 | 48 | case X86::BI__builtin_ia32_##NAME: \ |
7734 | 48 | ID = Intrinsic::x86_##NAME; \ |
7735 | 48 | break |
7736 | 48 | switch (BuiltinID) { |
7737 | 0 | default: 0 llvm_unreachable0 ("Unsupported intrinsic!"); |
7738 | 5 | INTRINSIC_X86_XSAVE_ID5 (xsave); |
7739 | 3 | INTRINSIC_X86_XSAVE_ID3 (xsave64); |
7740 | 5 | INTRINSIC_X86_XSAVE_ID5 (xrstor); |
7741 | 3 | INTRINSIC_X86_XSAVE_ID3 (xrstor64); |
7742 | 5 | INTRINSIC_X86_XSAVE_ID5 (xsaveopt); |
7743 | 3 | INTRINSIC_X86_XSAVE_ID3 (xsaveopt64); |
7744 | 5 | INTRINSIC_X86_XSAVE_ID5 (xrstors); |
7745 | 3 | INTRINSIC_X86_XSAVE_ID3 (xrstors64); |
7746 | 5 | INTRINSIC_X86_XSAVE_ID5 (xsavec); |
7747 | 3 | INTRINSIC_X86_XSAVE_ID3 (xsavec64); |
7748 | 5 | INTRINSIC_X86_XSAVE_ID5 (xsaves); |
7749 | 3 | INTRINSIC_X86_XSAVE_ID(xsaves64); |
7750 | 48 | } |
7751 | 48 | #undef INTRINSIC_X86_XSAVE_ID |
7752 | 48 | Value *Mhi = Builder.CreateTrunc( |
7753 | 48 | Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, 32)), Int32Ty); |
7754 | 48 | Value *Mlo = Builder.CreateTrunc(Ops[1], Int32Ty); |
7755 | 48 | Ops[1] = Mhi; |
7756 | 48 | Ops.push_back(Mlo); |
7757 | 48 | return Builder.CreateCall(CGM.getIntrinsic(ID), Ops); |
7758 | 48 | } |
7759 | 25 | case X86::BI__builtin_ia32_storedqudi128_mask: |
7760 | 25 | case X86::BI__builtin_ia32_storedqusi128_mask: |
7761 | 25 | case X86::BI__builtin_ia32_storedquhi128_mask: |
7762 | 25 | case X86::BI__builtin_ia32_storedquqi128_mask: |
7763 | 25 | case X86::BI__builtin_ia32_storeupd128_mask: |
7764 | 25 | case X86::BI__builtin_ia32_storeups128_mask: |
7765 | 25 | case X86::BI__builtin_ia32_storedqudi256_mask: |
7766 | 25 | case X86::BI__builtin_ia32_storedqusi256_mask: |
7767 | 25 | case X86::BI__builtin_ia32_storedquhi256_mask: |
7768 | 25 | case X86::BI__builtin_ia32_storedquqi256_mask: |
7769 | 25 | case X86::BI__builtin_ia32_storeupd256_mask: |
7770 | 25 | case X86::BI__builtin_ia32_storeups256_mask: |
7771 | 25 | case X86::BI__builtin_ia32_storedqudi512_mask: |
7772 | 25 | case X86::BI__builtin_ia32_storedqusi512_mask: |
7773 | 25 | case X86::BI__builtin_ia32_storedquhi512_mask: |
7774 | 25 | case X86::BI__builtin_ia32_storedquqi512_mask: |
7775 | 25 | case X86::BI__builtin_ia32_storeupd512_mask: |
7776 | 25 | case X86::BI__builtin_ia32_storeups512_mask: |
7777 | 25 | return EmitX86MaskedStore(*this, Ops, 1); |
7778 | 25 | |
7779 | 2 | case X86::BI__builtin_ia32_storess128_mask: |
7780 | 2 | case X86::BI__builtin_ia32_storesd128_mask: { |
7781 | 2 | return EmitX86MaskedStore(*this, Ops, 16); |
7782 | 2 | } |
7783 | 2 | case X86::BI__builtin_ia32_vpopcntd_512: |
7784 | 2 | case X86::BI__builtin_ia32_vpopcntq_512: { |
7785 | 2 | llvm::Type *ResultType = ConvertType(E->getType()); |
7786 | 2 | llvm::Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ResultType); |
7787 | 2 | return Builder.CreateCall(F, Ops); |
7788 | 2 | } |
7789 | 18 | case X86::BI__builtin_ia32_cvtmask2b128: |
7790 | 18 | case X86::BI__builtin_ia32_cvtmask2b256: |
7791 | 18 | case X86::BI__builtin_ia32_cvtmask2b512: |
7792 | 18 | case X86::BI__builtin_ia32_cvtmask2w128: |
7793 | 18 | case X86::BI__builtin_ia32_cvtmask2w256: |
7794 | 18 | case X86::BI__builtin_ia32_cvtmask2w512: |
7795 | 18 | case X86::BI__builtin_ia32_cvtmask2d128: |
7796 | 18 | case X86::BI__builtin_ia32_cvtmask2d256: |
7797 | 18 | case X86::BI__builtin_ia32_cvtmask2d512: |
7798 | 18 | case X86::BI__builtin_ia32_cvtmask2q128: |
7799 | 18 | case X86::BI__builtin_ia32_cvtmask2q256: |
7800 | 18 | case X86::BI__builtin_ia32_cvtmask2q512: |
7801 | 18 | return EmitX86SExtMask(*this, Ops[0], ConvertType(E->getType())); |
7802 | 18 | |
7803 | 12 | case X86::BI__builtin_ia32_movdqa32store128_mask: |
7804 | 12 | case X86::BI__builtin_ia32_movdqa64store128_mask: |
7805 | 12 | case X86::BI__builtin_ia32_storeaps128_mask: |
7806 | 12 | case X86::BI__builtin_ia32_storeapd128_mask: |
7807 | 12 | case X86::BI__builtin_ia32_movdqa32store256_mask: |
7808 | 12 | case X86::BI__builtin_ia32_movdqa64store256_mask: |
7809 | 12 | case X86::BI__builtin_ia32_storeaps256_mask: |
7810 | 12 | case X86::BI__builtin_ia32_storeapd256_mask: |
7811 | 12 | case X86::BI__builtin_ia32_movdqa32store512_mask: |
7812 | 12 | case X86::BI__builtin_ia32_movdqa64store512_mask: |
7813 | 12 | case X86::BI__builtin_ia32_storeaps512_mask: |
7814 | 12 | case X86::BI__builtin_ia32_storeapd512_mask: { |
7815 | 12 | unsigned Align = |
7816 | 12 | getContext().getTypeAlignInChars(E->getArg(1)->getType()).getQuantity(); |
7817 | 12 | return EmitX86MaskedStore(*this, Ops, Align); |
7818 | 12 | } |
7819 | 45 | case X86::BI__builtin_ia32_loadups128_mask: |
7820 | 45 | case X86::BI__builtin_ia32_loadups256_mask: |
7821 | 45 | case X86::BI__builtin_ia32_loadups512_mask: |
7822 | 45 | case X86::BI__builtin_ia32_loadupd128_mask: |
7823 | 45 | case X86::BI__builtin_ia32_loadupd256_mask: |
7824 | 45 | case X86::BI__builtin_ia32_loadupd512_mask: |
7825 | 45 | case X86::BI__builtin_ia32_loaddquqi128_mask: |
7826 | 45 | case X86::BI__builtin_ia32_loaddquqi256_mask: |
7827 | 45 | case X86::BI__builtin_ia32_loaddquqi512_mask: |
7828 | 45 | case X86::BI__builtin_ia32_loaddquhi128_mask: |
7829 | 45 | case X86::BI__builtin_ia32_loaddquhi256_mask: |
7830 | 45 | case X86::BI__builtin_ia32_loaddquhi512_mask: |
7831 | 45 | case X86::BI__builtin_ia32_loaddqusi128_mask: |
7832 | 45 | case X86::BI__builtin_ia32_loaddqusi256_mask: |
7833 | 45 | case X86::BI__builtin_ia32_loaddqusi512_mask: |
7834 | 45 | case X86::BI__builtin_ia32_loaddqudi128_mask: |
7835 | 45 | case X86::BI__builtin_ia32_loaddqudi256_mask: |
7836 | 45 | case X86::BI__builtin_ia32_loaddqudi512_mask: |
7837 | 45 | return EmitX86MaskedLoad(*this, Ops, 1); |
7838 | 45 | |
7839 | 4 | case X86::BI__builtin_ia32_loadss128_mask: |
7840 | 4 | case X86::BI__builtin_ia32_loadsd128_mask: |
7841 | 4 | return EmitX86MaskedLoad(*this, Ops, 16); |
7842 | 4 | |
7843 | 26 | case X86::BI__builtin_ia32_loadaps128_mask: |
7844 | 26 | case X86::BI__builtin_ia32_loadaps256_mask: |
7845 | 26 | case X86::BI__builtin_ia32_loadaps512_mask: |
7846 | 26 | case X86::BI__builtin_ia32_loadapd128_mask: |
7847 | 26 | case X86::BI__builtin_ia32_loadapd256_mask: |
7848 | 26 | case X86::BI__builtin_ia32_loadapd512_mask: |
7849 | 26 | case X86::BI__builtin_ia32_movdqa32load128_mask: |
7850 | 26 | case X86::BI__builtin_ia32_movdqa32load256_mask: |
7851 | 26 | case X86::BI__builtin_ia32_movdqa32load512_mask: |
7852 | 26 | case X86::BI__builtin_ia32_movdqa64load128_mask: |
7853 | 26 | case X86::BI__builtin_ia32_movdqa64load256_mask: |
7854 | 26 | case X86::BI__builtin_ia32_movdqa64load512_mask: { |
7855 | 26 | unsigned Align = |
7856 | 26 | getContext().getTypeAlignInChars(E->getArg(1)->getType()).getQuantity(); |
7857 | 26 | return EmitX86MaskedLoad(*this, Ops, Align); |
7858 | 26 | } |
7859 | 26 | |
7860 | 6 | case X86::BI__builtin_ia32_vbroadcastf128_pd256: |
7861 | 6 | case X86::BI__builtin_ia32_vbroadcastf128_ps256: { |
7862 | 6 | llvm::Type *DstTy = ConvertType(E->getType()); |
7863 | 6 | return EmitX86SubVectorBroadcast(*this, Ops, DstTy, 128, 1); |
7864 | 6 | } |
7865 | 6 | |
7866 | 4 | case X86::BI__builtin_ia32_storehps: |
7867 | 4 | case X86::BI__builtin_ia32_storelps: { |
7868 | 4 | llvm::Type *PtrTy = llvm::PointerType::getUnqual(Int64Ty); |
7869 | 4 | llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); |
7870 | 4 | |
7871 | 4 | // cast val v2i64 |
7872 | 4 | Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast"); |
7873 | 4 | |
7874 | 4 | // extract (0, 1) |
7875 | 4 | unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 02 : 12 ; |
7876 | 4 | llvm::Value *Idx = llvm::ConstantInt::get(SizeTy, Index); |
7877 | 4 | Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract"); |
7878 | 4 | |
7879 | 4 | // cast pointer to i64 & store |
7880 | 4 | Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy); |
7881 | 4 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
7882 | 4 | } |
7883 | 25 | case X86::BI__builtin_ia32_palignr128: |
7884 | 25 | case X86::BI__builtin_ia32_palignr256: |
7885 | 25 | case X86::BI__builtin_ia32_palignr512_mask: { |
7886 | 25 | unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); |
7887 | 25 | |
7888 | 25 | unsigned NumElts = Ops[0]->getType()->getVectorNumElements(); |
7889 | 25 | assert(NumElts % 16 == 0); |
7890 | 25 | |
7891 | 25 | // If palignr is shifting the pair of vectors more than the size of two |
7892 | 25 | // lanes, emit zero. |
7893 | 25 | if (ShiftVal >= 32) |
7894 | 2 | return llvm::Constant::getNullValue(ConvertType(E->getType())); |
7895 | 23 | |
7896 | 23 | // If palignr is shifting the pair of input vectors more than one lane, |
7897 | 23 | // but less than two lanes, convert to shifting in zeroes. |
7898 | 23 | if (23 ShiftVal > 1623 ) { |
7899 | 4 | ShiftVal -= 16; |
7900 | 4 | Ops[1] = Ops[0]; |
7901 | 4 | Ops[0] = llvm::Constant::getNullValue(Ops[0]->getType()); |
7902 | 4 | } |
7903 | 23 | |
7904 | 23 | uint32_t Indices[64]; |
7905 | 23 | // 256-bit palignr operates on 128-bit lanes so we need to handle that |
7906 | 72 | for (unsigned l = 0; l != NumElts72 ; l += 1649 ) { |
7907 | 833 | for (unsigned i = 0; i != 16833 ; ++i784 ) { |
7908 | 784 | unsigned Idx = ShiftVal + i; |
7909 | 784 | if (Idx >= 16) |
7910 | 119 | Idx += NumElts - 16; // End of lane, switch operand. |
7911 | 784 | Indices[l + i] = Idx + l; |
7912 | 784 | } |
7913 | 49 | } |
7914 | 23 | |
7915 | 23 | Value *Align = Builder.CreateShuffleVector(Ops[1], Ops[0], |
7916 | 23 | makeArrayRef(Indices, NumElts), |
7917 | 23 | "palignr"); |
7918 | 23 | |
7919 | 23 | // If this isn't a masked builtin, just return the align operation. |
7920 | 23 | if (Ops.size() == 3) |
7921 | 17 | return Align; |
7922 | 6 | |
7923 | 6 | return EmitX86Select(*this, Ops[4], Align, Ops[3]); |
7924 | 6 | } |
7925 | 6 | |
7926 | 14 | case X86::BI__builtin_ia32_vperm2f128_pd256: |
7927 | 14 | case X86::BI__builtin_ia32_vperm2f128_ps256: |
7928 | 14 | case X86::BI__builtin_ia32_vperm2f128_si256: |
7929 | 14 | case X86::BI__builtin_ia32_permti256: { |
7930 | 14 | unsigned Imm = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); |
7931 | 14 | unsigned NumElts = Ops[0]->getType()->getVectorNumElements(); |
7932 | 14 | |
7933 | 14 | // This takes a very simple approach since there are two lanes and a |
7934 | 14 | // shuffle can have 2 inputs. So we reserve the first input for the first |
7935 | 14 | // lane and the second input for the second lane. This may result in |
7936 | 14 | // duplicate sources, but this can be dealt with in the backend. |
7937 | 14 | |
7938 | 14 | Value *OutOps[2]; |
7939 | 14 | uint32_t Indices[8]; |
7940 | 42 | for (unsigned l = 0; l != 242 ; ++l28 ) { |
7941 | 28 | // Determine the source for this lane. |
7942 | 28 | if (Imm & (1 << ((l * 4) + 3))) |
7943 | 2 | OutOps[l] = llvm::ConstantAggregateZero::get(Ops[0]->getType()); |
7944 | 26 | else if (26 Imm & (1 << ((l * 4) + 1))26 ) |
7945 | 14 | OutOps[l] = Ops[1]; |
7946 | 26 | else |
7947 | 12 | OutOps[l] = Ops[0]; |
7948 | 28 | |
7949 | 116 | for (unsigned i = 0; i != NumElts/2116 ; ++i88 ) { |
7950 | 88 | // Start with ith element of the source for this lane. |
7951 | 88 | unsigned Idx = (l * NumElts) + i; |
7952 | 88 | // If bit 0 of the immediate half is set, switch to the high half of |
7953 | 88 | // the source. |
7954 | 88 | if (Imm & (1 << (l * 4))) |
7955 | 50 | Idx += NumElts/2; |
7956 | 88 | Indices[(l * (NumElts/2)) + i] = Idx; |
7957 | 88 | } |
7958 | 28 | } |
7959 | 14 | |
7960 | 14 | return Builder.CreateShuffleVector(OutOps[0], OutOps[1], |
7961 | 14 | makeArrayRef(Indices, NumElts), |
7962 | 14 | "vperm"); |
7963 | 14 | } |
7964 | 14 | |
7965 | 8 | case X86::BI__builtin_ia32_movnti: |
7966 | 8 | case X86::BI__builtin_ia32_movnti64: |
7967 | 8 | case X86::BI__builtin_ia32_movntsd: |
7968 | 8 | case X86::BI__builtin_ia32_movntss: { |
7969 | 8 | llvm::MDNode *Node = llvm::MDNode::get( |
7970 | 8 | getLLVMContext(), llvm::ConstantAsMetadata::get(Builder.getInt32(1))); |
7971 | 8 | |
7972 | 8 | Value *Ptr = Ops[0]; |
7973 | 8 | Value *Src = Ops[1]; |
7974 | 8 | |
7975 | 8 | // Extract the 0'th element of the source vector. |
7976 | 8 | if (BuiltinID == X86::BI__builtin_ia32_movntsd || |
7977 | 7 | BuiltinID == X86::BI__builtin_ia32_movntss) |
7978 | 2 | Src = Builder.CreateExtractElement(Src, (uint64_t)0, "extract"); |
7979 | 8 | |
7980 | 8 | // Convert the type of the pointer to a pointer to the stored type. |
7981 | 8 | Value *BC = Builder.CreateBitCast( |
7982 | 8 | Ptr, llvm::PointerType::getUnqual(Src->getType()), "cast"); |
7983 | 8 | |
7984 | 8 | // Unaligned nontemporal store of the scalar value. |
7985 | 8 | StoreInst *SI = Builder.CreateDefaultAlignedStore(Src, BC); |
7986 | 8 | SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node); |
7987 | 8 | SI->setAlignment(1); |
7988 | 8 | return SI; |
7989 | 8 | } |
7990 | 8 | |
7991 | 1.31k | case X86::BI__builtin_ia32_selectb_128: |
7992 | 1.31k | case X86::BI__builtin_ia32_selectb_256: |
7993 | 1.31k | case X86::BI__builtin_ia32_selectb_512: |
7994 | 1.31k | case X86::BI__builtin_ia32_selectw_128: |
7995 | 1.31k | case X86::BI__builtin_ia32_selectw_256: |
7996 | 1.31k | case X86::BI__builtin_ia32_selectw_512: |
7997 | 1.31k | case X86::BI__builtin_ia32_selectd_128: |
7998 | 1.31k | case X86::BI__builtin_ia32_selectd_256: |
7999 | 1.31k | case X86::BI__builtin_ia32_selectd_512: |
8000 | 1.31k | case X86::BI__builtin_ia32_selectq_128: |
8001 | 1.31k | case X86::BI__builtin_ia32_selectq_256: |
8002 | 1.31k | case X86::BI__builtin_ia32_selectq_512: |
8003 | 1.31k | case X86::BI__builtin_ia32_selectps_128: |
8004 | 1.31k | case X86::BI__builtin_ia32_selectps_256: |
8005 | 1.31k | case X86::BI__builtin_ia32_selectps_512: |
8006 | 1.31k | case X86::BI__builtin_ia32_selectpd_128: |
8007 | 1.31k | case X86::BI__builtin_ia32_selectpd_256: |
8008 | 1.31k | case X86::BI__builtin_ia32_selectpd_512: |
8009 | 1.31k | return EmitX86Select(*this, Ops[0], Ops[1], Ops[2]); |
8010 | 36 | case X86::BI__builtin_ia32_pcmpeqb128_mask: |
8011 | 36 | case X86::BI__builtin_ia32_pcmpeqb256_mask: |
8012 | 36 | case X86::BI__builtin_ia32_pcmpeqb512_mask: |
8013 | 36 | case X86::BI__builtin_ia32_pcmpeqw128_mask: |
8014 | 36 | case X86::BI__builtin_ia32_pcmpeqw256_mask: |
8015 | 36 | case X86::BI__builtin_ia32_pcmpeqw512_mask: |
8016 | 36 | case X86::BI__builtin_ia32_pcmpeqd128_mask: |
8017 | 36 | case X86::BI__builtin_ia32_pcmpeqd256_mask: |
8018 | 36 | case X86::BI__builtin_ia32_pcmpeqd512_mask: |
8019 | 36 | case X86::BI__builtin_ia32_pcmpeqq128_mask: |
8020 | 36 | case X86::BI__builtin_ia32_pcmpeqq256_mask: |
8021 | 36 | case X86::BI__builtin_ia32_pcmpeqq512_mask: |
8022 | 36 | return EmitX86MaskedCompare(*this, 0, false, Ops); |
8023 | 36 | case X86::BI__builtin_ia32_pcmpgtb128_mask: |
8024 | 36 | case X86::BI__builtin_ia32_pcmpgtb256_mask: |
8025 | 36 | case X86::BI__builtin_ia32_pcmpgtb512_mask: |
8026 | 36 | case X86::BI__builtin_ia32_pcmpgtw128_mask: |
8027 | 36 | case X86::BI__builtin_ia32_pcmpgtw256_mask: |
8028 | 36 | case X86::BI__builtin_ia32_pcmpgtw512_mask: |
8029 | 36 | case X86::BI__builtin_ia32_pcmpgtd128_mask: |
8030 | 36 | case X86::BI__builtin_ia32_pcmpgtd256_mask: |
8031 | 36 | case X86::BI__builtin_ia32_pcmpgtd512_mask: |
8032 | 36 | case X86::BI__builtin_ia32_pcmpgtq128_mask: |
8033 | 36 | case X86::BI__builtin_ia32_pcmpgtq256_mask: |
8034 | 36 | case X86::BI__builtin_ia32_pcmpgtq512_mask: |
8035 | 36 | return EmitX86MaskedCompare(*this, 6, true, Ops); |
8036 | 180 | case X86::BI__builtin_ia32_cmpb128_mask: |
8037 | 180 | case X86::BI__builtin_ia32_cmpb256_mask: |
8038 | 180 | case X86::BI__builtin_ia32_cmpb512_mask: |
8039 | 180 | case X86::BI__builtin_ia32_cmpw128_mask: |
8040 | 180 | case X86::BI__builtin_ia32_cmpw256_mask: |
8041 | 180 | case X86::BI__builtin_ia32_cmpw512_mask: |
8042 | 180 | case X86::BI__builtin_ia32_cmpd128_mask: |
8043 | 180 | case X86::BI__builtin_ia32_cmpd256_mask: |
8044 | 180 | case X86::BI__builtin_ia32_cmpd512_mask: |
8045 | 180 | case X86::BI__builtin_ia32_cmpq128_mask: |
8046 | 180 | case X86::BI__builtin_ia32_cmpq256_mask: |
8047 | 180 | case X86::BI__builtin_ia32_cmpq512_mask: { |
8048 | 180 | unsigned CC = cast<llvm::ConstantInt>(Ops[2])->getZExtValue() & 0x7; |
8049 | 180 | return EmitX86MaskedCompare(*this, CC, true, Ops); |
8050 | 180 | } |
8051 | 248 | case X86::BI__builtin_ia32_ucmpb128_mask: |
8052 | 248 | case X86::BI__builtin_ia32_ucmpb256_mask: |
8053 | 248 | case X86::BI__builtin_ia32_ucmpb512_mask: |
8054 | 248 | case X86::BI__builtin_ia32_ucmpw128_mask: |
8055 | 248 | case X86::BI__builtin_ia32_ucmpw256_mask: |
8056 | 248 | case X86::BI__builtin_ia32_ucmpw512_mask: |
8057 | 248 | case X86::BI__builtin_ia32_ucmpd128_mask: |
8058 | 248 | case X86::BI__builtin_ia32_ucmpd256_mask: |
8059 | 248 | case X86::BI__builtin_ia32_ucmpd512_mask: |
8060 | 248 | case X86::BI__builtin_ia32_ucmpq128_mask: |
8061 | 248 | case X86::BI__builtin_ia32_ucmpq256_mask: |
8062 | 248 | case X86::BI__builtin_ia32_ucmpq512_mask: { |
8063 | 248 | unsigned CC = cast<llvm::ConstantInt>(Ops[2])->getZExtValue() & 0x7; |
8064 | 248 | return EmitX86MaskedCompare(*this, CC, false, Ops); |
8065 | 248 | } |
8066 | 248 | |
8067 | 18 | case X86::BI__builtin_ia32_vplzcntd_128_mask: |
8068 | 18 | case X86::BI__builtin_ia32_vplzcntd_256_mask: |
8069 | 18 | case X86::BI__builtin_ia32_vplzcntd_512_mask: |
8070 | 18 | case X86::BI__builtin_ia32_vplzcntq_128_mask: |
8071 | 18 | case X86::BI__builtin_ia32_vplzcntq_256_mask: |
8072 | 18 | case X86::BI__builtin_ia32_vplzcntq_512_mask: { |
8073 | 18 | Function *F = CGM.getIntrinsic(Intrinsic::ctlz, Ops[0]->getType()); |
8074 | 18 | return EmitX86Select(*this, Ops[2], |
8075 | 18 | Builder.CreateCall(F, {Ops[0],Builder.getInt1(false)}), |
8076 | 18 | Ops[1]); |
8077 | 18 | } |
8078 | 18 | |
8079 | 44 | case X86::BI__builtin_ia32_pabsb128: |
8080 | 44 | case X86::BI__builtin_ia32_pabsw128: |
8081 | 44 | case X86::BI__builtin_ia32_pabsd128: |
8082 | 44 | case X86::BI__builtin_ia32_pabsb256: |
8083 | 44 | case X86::BI__builtin_ia32_pabsw256: |
8084 | 44 | case X86::BI__builtin_ia32_pabsd256: |
8085 | 44 | case X86::BI__builtin_ia32_pabsq128_mask: |
8086 | 44 | case X86::BI__builtin_ia32_pabsq256_mask: |
8087 | 44 | case X86::BI__builtin_ia32_pabsb512_mask: |
8088 | 44 | case X86::BI__builtin_ia32_pabsw512_mask: |
8089 | 44 | case X86::BI__builtin_ia32_pabsd512_mask: |
8090 | 44 | case X86::BI__builtin_ia32_pabsq512_mask: |
8091 | 44 | return EmitX86Abs(*this, Ops); |
8092 | 44 | |
8093 | 49 | case X86::BI__builtin_ia32_pmaxsb128: |
8094 | 49 | case X86::BI__builtin_ia32_pmaxsw128: |
8095 | 49 | case X86::BI__builtin_ia32_pmaxsd128: |
8096 | 49 | case X86::BI__builtin_ia32_pmaxsq128_mask: |
8097 | 49 | case X86::BI__builtin_ia32_pmaxsb256: |
8098 | 49 | case X86::BI__builtin_ia32_pmaxsw256: |
8099 | 49 | case X86::BI__builtin_ia32_pmaxsd256: |
8100 | 49 | case X86::BI__builtin_ia32_pmaxsq256_mask: |
8101 | 49 | case X86::BI__builtin_ia32_pmaxsb512_mask: |
8102 | 49 | case X86::BI__builtin_ia32_pmaxsw512_mask: |
8103 | 49 | case X86::BI__builtin_ia32_pmaxsd512_mask: |
8104 | 49 | case X86::BI__builtin_ia32_pmaxsq512_mask: |
8105 | 49 | return EmitX86MinMax(*this, ICmpInst::ICMP_SGT, Ops); |
8106 | 49 | case X86::BI__builtin_ia32_pmaxub128: |
8107 | 49 | case X86::BI__builtin_ia32_pmaxuw128: |
8108 | 49 | case X86::BI__builtin_ia32_pmaxud128: |
8109 | 49 | case X86::BI__builtin_ia32_pmaxuq128_mask: |
8110 | 49 | case X86::BI__builtin_ia32_pmaxub256: |
8111 | 49 | case X86::BI__builtin_ia32_pmaxuw256: |
8112 | 49 | case X86::BI__builtin_ia32_pmaxud256: |
8113 | 49 | case X86::BI__builtin_ia32_pmaxuq256_mask: |
8114 | 49 | case X86::BI__builtin_ia32_pmaxub512_mask: |
8115 | 49 | case X86::BI__builtin_ia32_pmaxuw512_mask: |
8116 | 49 | case X86::BI__builtin_ia32_pmaxud512_mask: |
8117 | 49 | case X86::BI__builtin_ia32_pmaxuq512_mask: |
8118 | 49 | return EmitX86MinMax(*this, ICmpInst::ICMP_UGT, Ops); |
8119 | 49 | case X86::BI__builtin_ia32_pminsb128: |
8120 | 49 | case X86::BI__builtin_ia32_pminsw128: |
8121 | 49 | case X86::BI__builtin_ia32_pminsd128: |
8122 | 49 | case X86::BI__builtin_ia32_pminsq128_mask: |
8123 | 49 | case X86::BI__builtin_ia32_pminsb256: |
8124 | 49 | case X86::BI__builtin_ia32_pminsw256: |
8125 | 49 | case X86::BI__builtin_ia32_pminsd256: |
8126 | 49 | case X86::BI__builtin_ia32_pminsq256_mask: |
8127 | 49 | case X86::BI__builtin_ia32_pminsb512_mask: |
8128 | 49 | case X86::BI__builtin_ia32_pminsw512_mask: |
8129 | 49 | case X86::BI__builtin_ia32_pminsd512_mask: |
8130 | 49 | case X86::BI__builtin_ia32_pminsq512_mask: |
8131 | 49 | return EmitX86MinMax(*this, ICmpInst::ICMP_SLT, Ops); |
8132 | 48 | case X86::BI__builtin_ia32_pminub128: |
8133 | 48 | case X86::BI__builtin_ia32_pminuw128: |
8134 | 48 | case X86::BI__builtin_ia32_pminud128: |
8135 | 48 | case X86::BI__builtin_ia32_pminuq128_mask: |
8136 | 48 | case X86::BI__builtin_ia32_pminub256: |
8137 | 48 | case X86::BI__builtin_ia32_pminuw256: |
8138 | 48 | case X86::BI__builtin_ia32_pminud256: |
8139 | 48 | case X86::BI__builtin_ia32_pminuq256_mask: |
8140 | 48 | case X86::BI__builtin_ia32_pminub512_mask: |
8141 | 48 | case X86::BI__builtin_ia32_pminuw512_mask: |
8142 | 48 | case X86::BI__builtin_ia32_pminud512_mask: |
8143 | 48 | case X86::BI__builtin_ia32_pminuq512_mask: |
8144 | 48 | return EmitX86MinMax(*this, ICmpInst::ICMP_ULT, Ops); |
8145 | 48 | |
8146 | 48 | // 3DNow! |
8147 | 4 | case X86::BI__builtin_ia32_pswapdsf: |
8148 | 4 | case X86::BI__builtin_ia32_pswapdsi: { |
8149 | 4 | llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext()); |
8150 | 4 | Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast"); |
8151 | 4 | llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_3dnowa_pswapd); |
8152 | 4 | return Builder.CreateCall(F, Ops, "pswapd"); |
8153 | 4 | } |
8154 | 6 | case X86::BI__builtin_ia32_rdrand16_step: |
8155 | 6 | case X86::BI__builtin_ia32_rdrand32_step: |
8156 | 6 | case X86::BI__builtin_ia32_rdrand64_step: |
8157 | 6 | case X86::BI__builtin_ia32_rdseed16_step: |
8158 | 6 | case X86::BI__builtin_ia32_rdseed32_step: |
8159 | 6 | case X86::BI__builtin_ia32_rdseed64_step: { |
8160 | 6 | Intrinsic::ID ID; |
8161 | 6 | switch (BuiltinID) { |
8162 | 0 | default: 0 llvm_unreachable0 ("Unsupported intrinsic!"); |
8163 | 1 | case X86::BI__builtin_ia32_rdrand16_step: |
8164 | 1 | ID = Intrinsic::x86_rdrand_16; |
8165 | 1 | break; |
8166 | 1 | case X86::BI__builtin_ia32_rdrand32_step: |
8167 | 1 | ID = Intrinsic::x86_rdrand_32; |
8168 | 1 | break; |
8169 | 1 | case X86::BI__builtin_ia32_rdrand64_step: |
8170 | 1 | ID = Intrinsic::x86_rdrand_64; |
8171 | 1 | break; |
8172 | 1 | case X86::BI__builtin_ia32_rdseed16_step: |
8173 | 1 | ID = Intrinsic::x86_rdseed_16; |
8174 | 1 | break; |
8175 | 1 | case X86::BI__builtin_ia32_rdseed32_step: |
8176 | 1 | ID = Intrinsic::x86_rdseed_32; |
8177 | 1 | break; |
8178 | 1 | case X86::BI__builtin_ia32_rdseed64_step: |
8179 | 1 | ID = Intrinsic::x86_rdseed_64; |
8180 | 1 | break; |
8181 | 6 | } |
8182 | 6 | |
8183 | 6 | Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID)); |
8184 | 6 | Builder.CreateDefaultAlignedStore(Builder.CreateExtractValue(Call, 0), |
8185 | 6 | Ops[0]); |
8186 | 6 | return Builder.CreateExtractValue(Call, 1); |
8187 | 6 | } |
8188 | 6 | |
8189 | 6 | // SSE packed comparison intrinsics |
8190 | 7 | case X86::BI__builtin_ia32_cmpeqps: |
8191 | 7 | case X86::BI__builtin_ia32_cmpeqpd: |
8192 | 7 | return getVectorFCmpIR(CmpInst::FCMP_OEQ); |
8193 | 6 | case X86::BI__builtin_ia32_cmpltps: |
8194 | 6 | case X86::BI__builtin_ia32_cmpltpd: |
8195 | 6 | return getVectorFCmpIR(CmpInst::FCMP_OLT); |
8196 | 6 | case X86::BI__builtin_ia32_cmpleps: |
8197 | 6 | case X86::BI__builtin_ia32_cmplepd: |
8198 | 6 | return getVectorFCmpIR(CmpInst::FCMP_OLE); |
8199 | 3 | case X86::BI__builtin_ia32_cmpunordps: |
8200 | 3 | case X86::BI__builtin_ia32_cmpunordpd: |
8201 | 3 | return getVectorFCmpIR(CmpInst::FCMP_UNO); |
8202 | 3 | case X86::BI__builtin_ia32_cmpneqps: |
8203 | 3 | case X86::BI__builtin_ia32_cmpneqpd: |
8204 | 3 | return getVectorFCmpIR(CmpInst::FCMP_UNE); |
8205 | 6 | case X86::BI__builtin_ia32_cmpnltps: |
8206 | 6 | case X86::BI__builtin_ia32_cmpnltpd: |
8207 | 6 | return getVectorFCmpIR(CmpInst::FCMP_UGE); |
8208 | 6 | case X86::BI__builtin_ia32_cmpnleps: |
8209 | 6 | case X86::BI__builtin_ia32_cmpnlepd: |
8210 | 6 | return getVectorFCmpIR(CmpInst::FCMP_UGT); |
8211 | 3 | case X86::BI__builtin_ia32_cmpordps: |
8212 | 3 | case X86::BI__builtin_ia32_cmpordpd: |
8213 | 3 | return getVectorFCmpIR(CmpInst::FCMP_ORD); |
8214 | 54 | case X86::BI__builtin_ia32_cmpps: |
8215 | 54 | case X86::BI__builtin_ia32_cmpps256: |
8216 | 54 | case X86::BI__builtin_ia32_cmppd: |
8217 | 54 | case X86::BI__builtin_ia32_cmppd256: { |
8218 | 54 | unsigned CC = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); |
8219 | 54 | // If this one of the SSE immediates, we can use native IR. |
8220 | 54 | if (CC < 854 ) { |
8221 | 26 | FCmpInst::Predicate Pred; |
8222 | 26 | switch (CC) { |
8223 | 12 | case 0: Pred = FCmpInst::FCMP_OEQ; break; |
8224 | 2 | case 1: Pred = FCmpInst::FCMP_OLT; break; |
8225 | 2 | case 2: Pred = FCmpInst::FCMP_OLE; break; |
8226 | 2 | case 3: Pred = FCmpInst::FCMP_UNO; break; |
8227 | 2 | case 4: Pred = FCmpInst::FCMP_UNE; break; |
8228 | 2 | case 5: Pred = FCmpInst::FCMP_UGE; break; |
8229 | 2 | case 6: Pred = FCmpInst::FCMP_UGT; break; |
8230 | 2 | case 7: Pred = FCmpInst::FCMP_ORD; break; |
8231 | 26 | } |
8232 | 26 | return getVectorFCmpIR(Pred); |
8233 | 26 | } |
8234 | 28 | |
8235 | 28 | // We can't handle 8-31 immediates with native IR, use the intrinsic. |
8236 | 28 | // Except for predicates that create constants. |
8237 | 28 | Intrinsic::ID ID; |
8238 | 28 | switch (BuiltinID) { |
8239 | 0 | default: 0 llvm_unreachable0 ("Unsupported intrinsic!"); |
8240 | 3 | case X86::BI__builtin_ia32_cmpps: |
8241 | 3 | ID = Intrinsic::x86_sse_cmp_ps; |
8242 | 3 | break; |
8243 | 11 | case X86::BI__builtin_ia32_cmpps256: |
8244 | 11 | // _CMP_TRUE_UQ, _CMP_TRUE_US produce -1,-1... vector |
8245 | 11 | // on any input and _CMP_FALSE_OQ, _CMP_FALSE_OS produce 0, 0... |
8246 | 11 | if (CC == 0xf || 11 CC == 0xb9 || CC == 0x1b7 || CC == 0x1f5 ) { |
8247 | 6 | Value *Constant = (CC == 0xf || CC == 0x1f) ? |
8248 | 4 | llvm::Constant::getAllOnesValue(Builder.getInt32Ty()) : |
8249 | 4 | llvm::Constant::getNullValue(Builder.getInt32Ty()); |
8250 | 8 | Value *Vec = Builder.CreateVectorSplat( |
8251 | 8 | Ops[0]->getType()->getVectorNumElements(), Constant); |
8252 | 8 | return Builder.CreateBitCast(Vec, Ops[0]->getType()); |
8253 | 8 | } |
8254 | 3 | ID = Intrinsic::x86_avx_cmp_ps_256; |
8255 | 3 | break; |
8256 | 3 | case X86::BI__builtin_ia32_cmppd: |
8257 | 3 | ID = Intrinsic::x86_sse2_cmp_pd; |
8258 | 3 | break; |
8259 | 11 | case X86::BI__builtin_ia32_cmppd256: |
8260 | 11 | // _CMP_TRUE_UQ, _CMP_TRUE_US produce -1,-1... vector |
8261 | 11 | // on any input and _CMP_FALSE_OQ, _CMP_FALSE_OS produce 0, 0... |
8262 | 11 | if (CC == 0xf || 11 CC == 0xb9 || CC == 0x1b7 || CC == 0x1f5 ) { |
8263 | 6 | Value *Constant = (CC == 0xf || CC == 0x1f) ? |
8264 | 4 | llvm::Constant::getAllOnesValue(Builder.getInt64Ty()) : |
8265 | 4 | llvm::Constant::getNullValue(Builder.getInt64Ty()); |
8266 | 8 | Value *Vec = Builder.CreateVectorSplat( |
8267 | 8 | Ops[0]->getType()->getVectorNumElements(), Constant); |
8268 | 8 | return Builder.CreateBitCast(Vec, Ops[0]->getType()); |
8269 | 8 | } |
8270 | 3 | ID = Intrinsic::x86_avx_cmp_pd_256; |
8271 | 3 | break; |
8272 | 12 | } |
8273 | 12 | |
8274 | 12 | return Builder.CreateCall(CGM.getIntrinsic(ID), Ops); |
8275 | 12 | } |
8276 | 12 | |
8277 | 12 | // SSE scalar comparison intrinsics |
8278 | 1 | case X86::BI__builtin_ia32_cmpeqss: |
8279 | 1 | return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 0); |
8280 | 3 | case X86::BI__builtin_ia32_cmpltss: |
8281 | 3 | return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 1); |
8282 | 3 | case X86::BI__builtin_ia32_cmpless: |
8283 | 3 | return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 2); |
8284 | 1 | case X86::BI__builtin_ia32_cmpunordss: |
8285 | 1 | return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 3); |
8286 | 1 | case X86::BI__builtin_ia32_cmpneqss: |
8287 | 1 | return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 4); |
8288 | 3 | case X86::BI__builtin_ia32_cmpnltss: |
8289 | 3 | return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 5); |
8290 | 3 | case X86::BI__builtin_ia32_cmpnless: |
8291 | 3 | return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 6); |
8292 | 1 | case X86::BI__builtin_ia32_cmpordss: |
8293 | 1 | return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 7); |
8294 | 2 | case X86::BI__builtin_ia32_cmpeqsd: |
8295 | 2 | return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 0); |
8296 | 5 | case X86::BI__builtin_ia32_cmpltsd: |
8297 | 5 | return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 1); |
8298 | 5 | case X86::BI__builtin_ia32_cmplesd: |
8299 | 5 | return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 2); |
8300 | 2 | case X86::BI__builtin_ia32_cmpunordsd: |
8301 | 2 | return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 3); |
8302 | 2 | case X86::BI__builtin_ia32_cmpneqsd: |
8303 | 2 | return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 4); |
8304 | 5 | case X86::BI__builtin_ia32_cmpnltsd: |
8305 | 5 | return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 5); |
8306 | 5 | case X86::BI__builtin_ia32_cmpnlesd: |
8307 | 5 | return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 6); |
8308 | 2 | case X86::BI__builtin_ia32_cmpordsd: |
8309 | 2 | return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 7); |
8310 | 12 | |
8311 | 4 | case X86::BI__emul: |
8312 | 4 | case X86::BI__emulu: { |
8313 | 4 | llvm::Type *Int64Ty = llvm::IntegerType::get(getLLVMContext(), 64); |
8314 | 4 | bool isSigned = (BuiltinID == X86::BI__emul); |
8315 | 4 | Value *LHS = Builder.CreateIntCast(Ops[0], Int64Ty, isSigned); |
8316 | 4 | Value *RHS = Builder.CreateIntCast(Ops[1], Int64Ty, isSigned); |
8317 | 4 | return Builder.CreateMul(LHS, RHS, "", !isSigned, isSigned); |
8318 | 4 | } |
8319 | 4 | case X86::BI__mulh: |
8320 | 4 | case X86::BI__umulh: |
8321 | 4 | case X86::BI_mul128: |
8322 | 4 | case X86::BI_umul128: { |
8323 | 4 | llvm::Type *ResType = ConvertType(E->getType()); |
8324 | 4 | llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128); |
8325 | 4 | |
8326 | 3 | bool IsSigned = (BuiltinID == X86::BI__mulh || BuiltinID == X86::BI_mul128); |
8327 | 4 | Value *LHS = Builder.CreateIntCast(Ops[0], Int128Ty, IsSigned); |
8328 | 4 | Value *RHS = Builder.CreateIntCast(Ops[1], Int128Ty, IsSigned); |
8329 | 4 | |
8330 | 4 | Value *MulResult, *HigherBits; |
8331 | 4 | if (IsSigned4 ) { |
8332 | 2 | MulResult = Builder.CreateNSWMul(LHS, RHS); |
8333 | 2 | HigherBits = Builder.CreateAShr(MulResult, 64); |
8334 | 4 | } else { |
8335 | 2 | MulResult = Builder.CreateNUWMul(LHS, RHS); |
8336 | 2 | HigherBits = Builder.CreateLShr(MulResult, 64); |
8337 | 2 | } |
8338 | 4 | HigherBits = Builder.CreateIntCast(HigherBits, ResType, IsSigned); |
8339 | 4 | |
8340 | 4 | if (BuiltinID == X86::BI__mulh || 4 BuiltinID == X86::BI__umulh3 ) |
8341 | 2 | return HigherBits; |
8342 | 2 | |
8343 | 2 | Address HighBitsAddress = EmitPointerWithAlignment(E->getArg(2)); |
8344 | 2 | Builder.CreateStore(HigherBits, HighBitsAddress); |
8345 | 2 | return Builder.CreateIntCast(MulResult, ResType, IsSigned); |
8346 | 2 | } |
8347 | 2 | |
8348 | 1 | case X86::BI__faststorefence: { |
8349 | 1 | return Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent, |
8350 | 1 | llvm::SyncScope::System); |
8351 | 2 | } |
8352 | 6 | case X86::BI_ReadWriteBarrier: |
8353 | 6 | case X86::BI_ReadBarrier: |
8354 | 6 | case X86::BI_WriteBarrier: { |
8355 | 6 | return Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent, |
8356 | 6 | llvm::SyncScope::SingleThread); |
8357 | 6 | } |
8358 | 7 | case X86::BI_BitScanForward: |
8359 | 7 | case X86::BI_BitScanForward64: |
8360 | 7 | return EmitMSVCBuiltinExpr(MSVCIntrin::_BitScanForward, E); |
8361 | 7 | case X86::BI_BitScanReverse: |
8362 | 7 | case X86::BI_BitScanReverse64: |
8363 | 7 | return EmitMSVCBuiltinExpr(MSVCIntrin::_BitScanReverse, E); |
8364 | 7 | |
8365 | 1 | case X86::BI_InterlockedAnd64: |
8366 | 1 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedAnd, E); |
8367 | 1 | case X86::BI_InterlockedExchange64: |
8368 | 1 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchange, E); |
8369 | 1 | case X86::BI_InterlockedExchangeAdd64: |
8370 | 1 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeAdd, E); |
8371 | 1 | case X86::BI_InterlockedExchangeSub64: |
8372 | 1 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeSub, E); |
8373 | 1 | case X86::BI_InterlockedOr64: |
8374 | 1 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedOr, E); |
8375 | 1 | case X86::BI_InterlockedXor64: |
8376 | 1 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedXor, E); |
8377 | 1 | case X86::BI_InterlockedDecrement64: |
8378 | 1 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedDecrement, E); |
8379 | 1 | case X86::BI_InterlockedIncrement64: |
8380 | 1 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedIncrement, E); |
8381 | 7 | |
8382 | 2 | case X86::BI_AddressOfReturnAddress: { |
8383 | 2 | Value *F = CGM.getIntrinsic(Intrinsic::addressofreturnaddress); |
8384 | 2 | return Builder.CreateCall(F); |
8385 | 7 | } |
8386 | 2 | case X86::BI__stosb: { |
8387 | 2 | // We treat __stosb as a volatile memset - it may not generate "rep stosb" |
8388 | 2 | // instruction, but it will create a memset that won't be optimized away. |
8389 | 2 | return Builder.CreateMemSet(Ops[0], Ops[1], Ops[2], 1, true); |
8390 | 7 | } |
8391 | 2 | case X86::BI__ud2: |
8392 | 2 | // llvm.trap makes a ud2a instruction on x86. |
8393 | 2 | return EmitTrapCall(Intrinsic::trap); |
8394 | 2 | case X86::BI__int2c: { |
8395 | 2 | // This syscall signals a driver assertion failure in x86 NT kernels. |
8396 | 2 | llvm::FunctionType *FTy = llvm::FunctionType::get(VoidTy, false); |
8397 | 2 | llvm::InlineAsm *IA = |
8398 | 2 | llvm::InlineAsm::get(FTy, "int $$0x2c", "", /*SideEffects=*/true); |
8399 | 2 | llvm::AttributeList NoReturnAttr = llvm::AttributeList::get( |
8400 | 2 | getLLVMContext(), llvm::AttributeList::FunctionIndex, |
8401 | 2 | llvm::Attribute::NoReturn); |
8402 | 2 | CallSite CS = Builder.CreateCall(IA); |
8403 | 2 | CS.setAttributes(NoReturnAttr); |
8404 | 2 | return CS.getInstruction(); |
8405 | 7 | } |
8406 | 4 | case X86::BI__readfsbyte: |
8407 | 4 | case X86::BI__readfsword: |
8408 | 4 | case X86::BI__readfsdword: |
8409 | 4 | case X86::BI__readfsqword: { |
8410 | 4 | llvm::Type *IntTy = ConvertType(E->getType()); |
8411 | 4 | Value *Ptr = Builder.CreateIntToPtr(EmitScalarExpr(E->getArg(0)), |
8412 | 4 | llvm::PointerType::get(IntTy, 257)); |
8413 | 4 | LoadInst *Load = Builder.CreateAlignedLoad( |
8414 | 4 | IntTy, Ptr, getContext().getTypeAlignInChars(E->getType())); |
8415 | 4 | Load->setVolatile(true); |
8416 | 4 | return Load; |
8417 | 4 | } |
8418 | 4 | case X86::BI__readgsbyte: |
8419 | 4 | case X86::BI__readgsword: |
8420 | 4 | case X86::BI__readgsdword: |
8421 | 4 | case X86::BI__readgsqword: { |
8422 | 4 | llvm::Type *IntTy = ConvertType(E->getType()); |
8423 | 4 | Value *Ptr = Builder.CreateIntToPtr(EmitScalarExpr(E->getArg(0)), |
8424 | 4 | llvm::PointerType::get(IntTy, 256)); |
8425 | 4 | LoadInst *Load = Builder.CreateAlignedLoad( |
8426 | 4 | IntTy, Ptr, getContext().getTypeAlignInChars(E->getType())); |
8427 | 4 | Load->setVolatile(true); |
8428 | 4 | return Load; |
8429 | 0 | } |
8430 | 2.70k | } |
8431 | 2.70k | } |
8432 | | |
8433 | | |
8434 | | Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID, |
8435 | 1.01k | const CallExpr *E) { |
8436 | 1.01k | SmallVector<Value*, 4> Ops; |
8437 | 1.01k | |
8438 | 3.50k | for (unsigned i = 0, e = E->getNumArgs(); i != e3.50k ; i++2.49k ) |
8439 | 2.49k | Ops.push_back(EmitScalarExpr(E->getArg(i))); |
8440 | 1.01k | |
8441 | 1.01k | Intrinsic::ID ID = Intrinsic::not_intrinsic; |
8442 | 1.01k | |
8443 | 1.01k | switch (BuiltinID) { |
8444 | 0 | default: return nullptr; |
8445 | 1.01k | |
8446 | 1.01k | // __builtin_ppc_get_timebase is GCC 4.8+'s PowerPC-specific name for what we |
8447 | 1.01k | // call __builtin_readcyclecounter. |
8448 | 1 | case PPC::BI__builtin_ppc_get_timebase: |
8449 | 1 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::readcyclecounter)); |
8450 | 1.01k | |
8451 | 1.01k | // vec_ld, vec_xl_be, vec_lvsl, vec_lvsr |
8452 | 374 | case PPC::BI__builtin_altivec_lvx: |
8453 | 374 | case PPC::BI__builtin_altivec_lvxl: |
8454 | 374 | case PPC::BI__builtin_altivec_lvebx: |
8455 | 374 | case PPC::BI__builtin_altivec_lvehx: |
8456 | 374 | case PPC::BI__builtin_altivec_lvewx: |
8457 | 374 | case PPC::BI__builtin_altivec_lvsl: |
8458 | 374 | case PPC::BI__builtin_altivec_lvsr: |
8459 | 374 | case PPC::BI__builtin_vsx_lxvd2x: |
8460 | 374 | case PPC::BI__builtin_vsx_lxvw4x: |
8461 | 374 | case PPC::BI__builtin_vsx_lxvd2x_be: |
8462 | 374 | case PPC::BI__builtin_vsx_lxvw4x_be: |
8463 | 374 | case PPC::BI__builtin_vsx_lxvl: |
8464 | 374 | case PPC::BI__builtin_vsx_lxvll: |
8465 | 374 | { |
8466 | 374 | if(BuiltinID == PPC::BI__builtin_vsx_lxvl || |
8467 | 374 | BuiltinID == PPC::BI__builtin_vsx_lxvll350 ){ |
8468 | 26 | Ops[0] = Builder.CreateBitCast(Ops[0], Int8PtrTy); |
8469 | 374 | }else { |
8470 | 348 | Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy); |
8471 | 348 | Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]); |
8472 | 348 | Ops.pop_back(); |
8473 | 348 | } |
8474 | 374 | |
8475 | 374 | switch (BuiltinID) { |
8476 | 0 | default: 0 llvm_unreachable0 ("Unsupported ld/lvsl/lvsr intrinsic!"); |
8477 | 108 | case PPC::BI__builtin_altivec_lvx: |
8478 | 108 | ID = Intrinsic::ppc_altivec_lvx; |
8479 | 108 | break; |
8480 | 108 | case PPC::BI__builtin_altivec_lvxl: |
8481 | 108 | ID = Intrinsic::ppc_altivec_lvxl; |
8482 | 108 | break; |
8483 | 12 | case PPC::BI__builtin_altivec_lvebx: |
8484 | 12 | ID = Intrinsic::ppc_altivec_lvebx; |
8485 | 12 | break; |
8486 | 12 | case PPC::BI__builtin_altivec_lvehx: |
8487 | 12 | ID = Intrinsic::ppc_altivec_lvehx; |
8488 | 12 | break; |
8489 | 18 | case PPC::BI__builtin_altivec_lvewx: |
8490 | 18 | ID = Intrinsic::ppc_altivec_lvewx; |
8491 | 18 | break; |
8492 | 19 | case PPC::BI__builtin_altivec_lvsl: |
8493 | 19 | ID = Intrinsic::ppc_altivec_lvsl; |
8494 | 19 | break; |
8495 | 19 | case PPC::BI__builtin_altivec_lvsr: |
8496 | 19 | ID = Intrinsic::ppc_altivec_lvsr; |
8497 | 19 | break; |
8498 | 8 | case PPC::BI__builtin_vsx_lxvd2x: |
8499 | 8 | ID = Intrinsic::ppc_vsx_lxvd2x; |
8500 | 8 | break; |
8501 | 34 | case PPC::BI__builtin_vsx_lxvw4x: |
8502 | 34 | ID = Intrinsic::ppc_vsx_lxvw4x; |
8503 | 34 | break; |
8504 | 7 | case PPC::BI__builtin_vsx_lxvd2x_be: |
8505 | 7 | ID = Intrinsic::ppc_vsx_lxvd2x_be; |
8506 | 7 | break; |
8507 | 3 | case PPC::BI__builtin_vsx_lxvw4x_be: |
8508 | 3 | ID = Intrinsic::ppc_vsx_lxvw4x_be; |
8509 | 3 | break; |
8510 | 24 | case PPC::BI__builtin_vsx_lxvl: |
8511 | 24 | ID = Intrinsic::ppc_vsx_lxvl; |
8512 | 24 | break; |
8513 | 2 | case PPC::BI__builtin_vsx_lxvll: |
8514 | 2 | ID = Intrinsic::ppc_vsx_lxvll; |
8515 | 2 | break; |
8516 | 374 | } |
8517 | 374 | llvm::Function *F = CGM.getIntrinsic(ID); |
8518 | 374 | return Builder.CreateCall(F, Ops, ""); |
8519 | 374 | } |
8520 | 374 | |
8521 | 374 | // vec_st, vec_xst_be |
8522 | 486 | case PPC::BI__builtin_altivec_stvx: |
8523 | 486 | case PPC::BI__builtin_altivec_stvxl: |
8524 | 486 | case PPC::BI__builtin_altivec_stvebx: |
8525 | 486 | case PPC::BI__builtin_altivec_stvehx: |
8526 | 486 | case PPC::BI__builtin_altivec_stvewx: |
8527 | 486 | case PPC::BI__builtin_vsx_stxvd2x: |
8528 | 486 | case PPC::BI__builtin_vsx_stxvw4x: |
8529 | 486 | case PPC::BI__builtin_vsx_stxvd2x_be: |
8530 | 486 | case PPC::BI__builtin_vsx_stxvw4x_be: |
8531 | 486 | case PPC::BI__builtin_vsx_stxvl: |
8532 | 486 | case PPC::BI__builtin_vsx_stxvll: |
8533 | 486 | { |
8534 | 486 | if(BuiltinID == PPC::BI__builtin_vsx_stxvl || |
8535 | 486 | BuiltinID == PPC::BI__builtin_vsx_stxvll462 ){ |
8536 | 26 | Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy); |
8537 | 486 | }else { |
8538 | 460 | Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy); |
8539 | 460 | Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]); |
8540 | 460 | Ops.pop_back(); |
8541 | 460 | } |
8542 | 486 | |
8543 | 486 | switch (BuiltinID) { |
8544 | 0 | default: 0 llvm_unreachable0 ("Unsupported st intrinsic!"); |
8545 | 150 | case PPC::BI__builtin_altivec_stvx: |
8546 | 150 | ID = Intrinsic::ppc_altivec_stvx; |
8547 | 150 | break; |
8548 | 156 | case PPC::BI__builtin_altivec_stvxl: |
8549 | 156 | ID = Intrinsic::ppc_altivec_stvxl; |
8550 | 156 | break; |
8551 | 24 | case PPC::BI__builtin_altivec_stvebx: |
8552 | 24 | ID = Intrinsic::ppc_altivec_stvebx; |
8553 | 24 | break; |
8554 | 36 | case PPC::BI__builtin_altivec_stvehx: |
8555 | 36 | ID = Intrinsic::ppc_altivec_stvehx; |
8556 | 36 | break; |
8557 | 30 | case PPC::BI__builtin_altivec_stvewx: |
8558 | 30 | ID = Intrinsic::ppc_altivec_stvewx; |
8559 | 30 | break; |
8560 | 8 | case PPC::BI__builtin_vsx_stxvd2x: |
8561 | 8 | ID = Intrinsic::ppc_vsx_stxvd2x; |
8562 | 8 | break; |
8563 | 46 | case PPC::BI__builtin_vsx_stxvw4x: |
8564 | 46 | ID = Intrinsic::ppc_vsx_stxvw4x; |
8565 | 46 | break; |
8566 | 7 | case PPC::BI__builtin_vsx_stxvd2x_be: |
8567 | 7 | ID = Intrinsic::ppc_vsx_stxvd2x_be; |
8568 | 7 | break; |
8569 | 3 | case PPC::BI__builtin_vsx_stxvw4x_be: |
8570 | 3 | ID = Intrinsic::ppc_vsx_stxvw4x_be; |
8571 | 3 | break; |
8572 | 24 | case PPC::BI__builtin_vsx_stxvl: |
8573 | 24 | ID = Intrinsic::ppc_vsx_stxvl; |
8574 | 24 | break; |
8575 | 2 | case PPC::BI__builtin_vsx_stxvll: |
8576 | 2 | ID = Intrinsic::ppc_vsx_stxvll; |
8577 | 2 | break; |
8578 | 486 | } |
8579 | 486 | llvm::Function *F = CGM.getIntrinsic(ID); |
8580 | 486 | return Builder.CreateCall(F, Ops, ""); |
8581 | 486 | } |
8582 | 486 | // Square root |
8583 | 4 | case PPC::BI__builtin_vsx_xvsqrtsp: |
8584 | 4 | case PPC::BI__builtin_vsx_xvsqrtdp: { |
8585 | 4 | llvm::Type *ResultType = ConvertType(E->getType()); |
8586 | 4 | Value *X = EmitScalarExpr(E->getArg(0)); |
8587 | 4 | ID = Intrinsic::sqrt; |
8588 | 4 | llvm::Function *F = CGM.getIntrinsic(ID, ResultType); |
8589 | 4 | return Builder.CreateCall(F, X); |
8590 | 4 | } |
8591 | 4 | // Count leading zeros |
8592 | 17 | case PPC::BI__builtin_altivec_vclzb: |
8593 | 17 | case PPC::BI__builtin_altivec_vclzh: |
8594 | 17 | case PPC::BI__builtin_altivec_vclzw: |
8595 | 17 | case PPC::BI__builtin_altivec_vclzd: { |
8596 | 17 | llvm::Type *ResultType = ConvertType(E->getType()); |
8597 | 17 | Value *X = EmitScalarExpr(E->getArg(0)); |
8598 | 17 | Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); |
8599 | 17 | Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ResultType); |
8600 | 17 | return Builder.CreateCall(F, {X, Undef}); |
8601 | 17 | } |
8602 | 16 | case PPC::BI__builtin_altivec_vctzb: |
8603 | 16 | case PPC::BI__builtin_altivec_vctzh: |
8604 | 16 | case PPC::BI__builtin_altivec_vctzw: |
8605 | 16 | case PPC::BI__builtin_altivec_vctzd: { |
8606 | 16 | llvm::Type *ResultType = ConvertType(E->getType()); |
8607 | 16 | Value *X = EmitScalarExpr(E->getArg(0)); |
8608 | 16 | Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); |
8609 | 16 | Function *F = CGM.getIntrinsic(Intrinsic::cttz, ResultType); |
8610 | 16 | return Builder.CreateCall(F, {X, Undef}); |
8611 | 16 | } |
8612 | 16 | case PPC::BI__builtin_altivec_vpopcntb: |
8613 | 16 | case PPC::BI__builtin_altivec_vpopcnth: |
8614 | 16 | case PPC::BI__builtin_altivec_vpopcntw: |
8615 | 16 | case PPC::BI__builtin_altivec_vpopcntd: { |
8616 | 16 | llvm::Type *ResultType = ConvertType(E->getType()); |
8617 | 16 | Value *X = EmitScalarExpr(E->getArg(0)); |
8618 | 16 | llvm::Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ResultType); |
8619 | 16 | return Builder.CreateCall(F, X); |
8620 | 16 | } |
8621 | 16 | // Copy sign |
8622 | 4 | case PPC::BI__builtin_vsx_xvcpsgnsp: |
8623 | 4 | case PPC::BI__builtin_vsx_xvcpsgndp: { |
8624 | 4 | llvm::Type *ResultType = ConvertType(E->getType()); |
8625 | 4 | Value *X = EmitScalarExpr(E->getArg(0)); |
8626 | 4 | Value *Y = EmitScalarExpr(E->getArg(1)); |
8627 | 4 | ID = Intrinsic::copysign; |
8628 | 4 | llvm::Function *F = CGM.getIntrinsic(ID, ResultType); |
8629 | 4 | return Builder.CreateCall(F, {X, Y}); |
8630 | 4 | } |
8631 | 4 | // Rounding/truncation |
8632 | 24 | case PPC::BI__builtin_vsx_xvrspip: |
8633 | 24 | case PPC::BI__builtin_vsx_xvrdpip: |
8634 | 24 | case PPC::BI__builtin_vsx_xvrdpim: |
8635 | 24 | case PPC::BI__builtin_vsx_xvrspim: |
8636 | 24 | case PPC::BI__builtin_vsx_xvrdpi: |
8637 | 24 | case PPC::BI__builtin_vsx_xvrspi: |
8638 | 24 | case PPC::BI__builtin_vsx_xvrdpic: |
8639 | 24 | case PPC::BI__builtin_vsx_xvrspic: |
8640 | 24 | case PPC::BI__builtin_vsx_xvrdpiz: |
8641 | 24 | case PPC::BI__builtin_vsx_xvrspiz: { |
8642 | 24 | llvm::Type *ResultType = ConvertType(E->getType()); |
8643 | 24 | Value *X = EmitScalarExpr(E->getArg(0)); |
8644 | 24 | if (BuiltinID == PPC::BI__builtin_vsx_xvrdpim || |
8645 | 22 | BuiltinID == PPC::BI__builtin_vsx_xvrspim) |
8646 | 4 | ID = Intrinsic::floor; |
8647 | 20 | else if (20 BuiltinID == PPC::BI__builtin_vsx_xvrdpi || |
8648 | 16 | BuiltinID == PPC::BI__builtin_vsx_xvrspi) |
8649 | 8 | ID = Intrinsic::round; |
8650 | 12 | else if (12 BuiltinID == PPC::BI__builtin_vsx_xvrdpic || |
8651 | 10 | BuiltinID == PPC::BI__builtin_vsx_xvrspic) |
8652 | 4 | ID = Intrinsic::nearbyint; |
8653 | 8 | else if (8 BuiltinID == PPC::BI__builtin_vsx_xvrdpip || |
8654 | 6 | BuiltinID == PPC::BI__builtin_vsx_xvrspip) |
8655 | 4 | ID = Intrinsic::ceil; |
8656 | 4 | else if (4 BuiltinID == PPC::BI__builtin_vsx_xvrdpiz || |
8657 | 2 | BuiltinID == PPC::BI__builtin_vsx_xvrspiz) |
8658 | 4 | ID = Intrinsic::trunc; |
8659 | 24 | llvm::Function *F = CGM.getIntrinsic(ID, ResultType); |
8660 | 24 | return Builder.CreateCall(F, X); |
8661 | 24 | } |
8662 | 24 | |
8663 | 24 | // Absolute value |
8664 | 4 | case PPC::BI__builtin_vsx_xvabsdp: |
8665 | 4 | case PPC::BI__builtin_vsx_xvabssp: { |
8666 | 4 | llvm::Type *ResultType = ConvertType(E->getType()); |
8667 | 4 | Value *X = EmitScalarExpr(E->getArg(0)); |
8668 | 4 | llvm::Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); |
8669 | 4 | return Builder.CreateCall(F, X); |
8670 | 4 | } |
8671 | 4 | |
8672 | 4 | // FMA variations |
8673 | 16 | case PPC::BI__builtin_vsx_xvmaddadp: |
8674 | 16 | case PPC::BI__builtin_vsx_xvmaddasp: |
8675 | 16 | case PPC::BI__builtin_vsx_xvnmaddadp: |
8676 | 16 | case PPC::BI__builtin_vsx_xvnmaddasp: |
8677 | 16 | case PPC::BI__builtin_vsx_xvmsubadp: |
8678 | 16 | case PPC::BI__builtin_vsx_xvmsubasp: |
8679 | 16 | case PPC::BI__builtin_vsx_xvnmsubadp: |
8680 | 16 | case PPC::BI__builtin_vsx_xvnmsubasp: { |
8681 | 16 | llvm::Type *ResultType = ConvertType(E->getType()); |
8682 | 16 | Value *X = EmitScalarExpr(E->getArg(0)); |
8683 | 16 | Value *Y = EmitScalarExpr(E->getArg(1)); |
8684 | 16 | Value *Z = EmitScalarExpr(E->getArg(2)); |
8685 | 16 | Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); |
8686 | 16 | llvm::Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); |
8687 | 16 | switch (BuiltinID) { |
8688 | 4 | case PPC::BI__builtin_vsx_xvmaddadp: |
8689 | 4 | case PPC::BI__builtin_vsx_xvmaddasp: |
8690 | 4 | return Builder.CreateCall(F, {X, Y, Z}); |
8691 | 4 | case PPC::BI__builtin_vsx_xvnmaddadp: |
8692 | 4 | case PPC::BI__builtin_vsx_xvnmaddasp: |
8693 | 4 | return Builder.CreateFSub(Zero, |
8694 | 4 | Builder.CreateCall(F, {X, Y, Z}), "sub"); |
8695 | 4 | case PPC::BI__builtin_vsx_xvmsubadp: |
8696 | 4 | case PPC::BI__builtin_vsx_xvmsubasp: |
8697 | 4 | return Builder.CreateCall(F, |
8698 | 4 | {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); |
8699 | 4 | case PPC::BI__builtin_vsx_xvnmsubadp: |
8700 | 4 | case PPC::BI__builtin_vsx_xvnmsubasp: |
8701 | 4 | Value *FsubRes = |
8702 | 4 | Builder.CreateCall(F, {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); |
8703 | 4 | return Builder.CreateFSub(Zero, FsubRes, "sub"); |
8704 | 0 | } |
8705 | 0 | llvm_unreachable0 ("Unknown FMA operation"); |
8706 | 0 | return nullptr; // Suppress no-return warning |
8707 | 0 | } |
8708 | 0 |
|
8709 | 4 | case PPC::BI__builtin_vsx_insertword: { |
8710 | 4 | llvm::Function *F = CGM.getIntrinsic(Intrinsic::ppc_vsx_xxinsertw); |
8711 | 4 | |
8712 | 4 | // Third argument is a compile time constant int. It must be clamped to |
8713 | 4 | // to the range [0, 12]. |
8714 | 4 | ConstantInt *ArgCI = dyn_cast<ConstantInt>(Ops[2]); |
8715 | 4 | assert(ArgCI && |
8716 | 4 | "Third arg to xxinsertw intrinsic must be constant integer"); |
8717 | 4 | const int64_t MaxIndex = 12; |
8718 | 4 | int64_t Index = clamp(ArgCI->getSExtValue(), 0, MaxIndex); |
8719 | 4 | |
8720 | 4 | // The builtin semantics don't exactly match the xxinsertw instructions |
8721 | 4 | // semantics (which ppc_vsx_xxinsertw follows). The builtin extracts the |
8722 | 4 | // word from the first argument, and inserts it in the second argument. The |
8723 | 4 | // instruction extracts the word from its second input register and inserts |
8724 | 4 | // it into its first input register, so swap the first and second arguments. |
8725 | 4 | std::swap(Ops[0], Ops[1]); |
8726 | 4 | |
8727 | 4 | // Need to cast the second argument from a vector of unsigned int to a |
8728 | 4 | // vector of long long. |
8729 | 4 | Ops[1] = Builder.CreateBitCast(Ops[1], llvm::VectorType::get(Int64Ty, 2)); |
8730 | 4 | |
8731 | 4 | if (getTarget().isLittleEndian()4 ) { |
8732 | 2 | // Create a shuffle mask of (1, 0) |
8733 | 2 | Constant *ShuffleElts[2] = { ConstantInt::get(Int32Ty, 1), |
8734 | 2 | ConstantInt::get(Int32Ty, 0) |
8735 | 2 | }; |
8736 | 2 | Constant *ShuffleMask = llvm::ConstantVector::get(ShuffleElts); |
8737 | 2 | |
8738 | 2 | // Reverse the double words in the vector we will extract from. |
8739 | 2 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int64Ty, 2)); |
8740 | 2 | Ops[0] = Builder.CreateShuffleVector(Ops[0], Ops[0], ShuffleMask); |
8741 | 2 | |
8742 | 2 | // Reverse the index. |
8743 | 2 | Index = MaxIndex - Index; |
8744 | 2 | } |
8745 | 4 | |
8746 | 4 | // Intrinsic expects the first arg to be a vector of int. |
8747 | 4 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int32Ty, 4)); |
8748 | 4 | Ops[2] = ConstantInt::getSigned(Int32Ty, Index); |
8749 | 4 | return Builder.CreateCall(F, Ops); |
8750 | 0 | } |
8751 | 0 |
|
8752 | 4 | case PPC::BI__builtin_vsx_extractuword: { |
8753 | 4 | llvm::Function *F = CGM.getIntrinsic(Intrinsic::ppc_vsx_xxextractuw); |
8754 | 4 | |
8755 | 4 | // Intrinsic expects the first argument to be a vector of doublewords. |
8756 | 4 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int64Ty, 2)); |
8757 | 4 | |
8758 | 4 | // The second argument is a compile time constant int that needs to |
8759 | 4 | // be clamped to the range [0, 12]. |
8760 | 4 | ConstantInt *ArgCI = dyn_cast<ConstantInt>(Ops[1]); |
8761 | 4 | assert(ArgCI && |
8762 | 4 | "Second Arg to xxextractuw intrinsic must be a constant integer!"); |
8763 | 4 | const int64_t MaxIndex = 12; |
8764 | 4 | int64_t Index = clamp(ArgCI->getSExtValue(), 0, MaxIndex); |
8765 | 4 | |
8766 | 4 | if (getTarget().isLittleEndian()4 ) { |
8767 | 2 | // Reverse the index. |
8768 | 2 | Index = MaxIndex - Index; |
8769 | 2 | Ops[1] = ConstantInt::getSigned(Int32Ty, Index); |
8770 | 2 | |
8771 | 2 | // Emit the call, then reverse the double words of the results vector. |
8772 | 2 | Value *Call = Builder.CreateCall(F, Ops); |
8773 | 2 | |
8774 | 2 | // Create a shuffle mask of (1, 0) |
8775 | 2 | Constant *ShuffleElts[2] = { ConstantInt::get(Int32Ty, 1), |
8776 | 2 | ConstantInt::get(Int32Ty, 0) |
8777 | 2 | }; |
8778 | 2 | Constant *ShuffleMask = llvm::ConstantVector::get(ShuffleElts); |
8779 | 2 | |
8780 | 2 | Value *ShuffleCall = Builder.CreateShuffleVector(Call, Call, ShuffleMask); |
8781 | 2 | return ShuffleCall; |
8782 | 0 | } else { |
8783 | 2 | Ops[1] = ConstantInt::getSigned(Int32Ty, Index); |
8784 | 2 | return Builder.CreateCall(F, Ops); |
8785 | 2 | } |
8786 | 0 | } |
8787 | 0 |
|
8788 | 22 | case PPC::BI__builtin_vsx_xxpermdi: { |
8789 | 22 | ConstantInt *ArgCI = dyn_cast<ConstantInt>(Ops[2]); |
8790 | 22 | assert(ArgCI && "Third arg must be constant integer!"); |
8791 | 22 | |
8792 | 22 | unsigned Index = ArgCI->getZExtValue(); |
8793 | 22 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int64Ty, 2)); |
8794 | 22 | Ops[1] = Builder.CreateBitCast(Ops[1], llvm::VectorType::get(Int64Ty, 2)); |
8795 | 22 | |
8796 | 22 | // Element zero comes from the first input vector and element one comes from |
8797 | 22 | // the second. The element indices within each vector are numbered in big |
8798 | 22 | // endian order so the shuffle mask must be adjusted for this on little |
8799 | 22 | // endian platforms (i.e. index is complemented and source vector reversed). |
8800 | 22 | unsigned ElemIdx0; |
8801 | 22 | unsigned ElemIdx1; |
8802 | 22 | if (getTarget().isLittleEndian()22 ) { |
8803 | 11 | ElemIdx0 = (~Index & 1) + 2; |
8804 | 11 | ElemIdx1 = (~Index & 2) >> 1; |
8805 | 22 | } else { // BigEndian |
8806 | 11 | ElemIdx0 = (Index & 2) >> 1; |
8807 | 11 | ElemIdx1 = 2 + (Index & 1); |
8808 | 11 | } |
8809 | 22 | |
8810 | 22 | Constant *ShuffleElts[2] = {ConstantInt::get(Int32Ty, ElemIdx0), |
8811 | 22 | ConstantInt::get(Int32Ty, ElemIdx1)}; |
8812 | 22 | Constant *ShuffleMask = llvm::ConstantVector::get(ShuffleElts); |
8813 | 22 | |
8814 | 22 | Value *ShuffleCall = |
8815 | 22 | Builder.CreateShuffleVector(Ops[0], Ops[1], ShuffleMask); |
8816 | 22 | QualType BIRetType = E->getType(); |
8817 | 22 | auto RetTy = ConvertType(BIRetType); |
8818 | 22 | return Builder.CreateBitCast(ShuffleCall, RetTy); |
8819 | 0 | } |
8820 | 0 |
|
8821 | 22 | case PPC::BI__builtin_vsx_xxsldwi: { |
8822 | 22 | ConstantInt *ArgCI = dyn_cast<ConstantInt>(Ops[2]); |
8823 | 22 | assert(ArgCI && "Third argument must be a compile time constant"); |
8824 | 22 | unsigned Index = ArgCI->getZExtValue() & 0x3; |
8825 | 22 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int32Ty, 4)); |
8826 | 22 | Ops[1] = Builder.CreateBitCast(Ops[1], llvm::VectorType::get(Int32Ty, 4)); |
8827 | 22 | |
8828 | 22 | // Create a shuffle mask |
8829 | 22 | unsigned ElemIdx0; |
8830 | 22 | unsigned ElemIdx1; |
8831 | 22 | unsigned ElemIdx2; |
8832 | 22 | unsigned ElemIdx3; |
8833 | 22 | if (getTarget().isLittleEndian()22 ) { |
8834 | 11 | // Little endian element N comes from element 8+N-Index of the |
8835 | 11 | // concatenated wide vector (of course, using modulo arithmetic on |
8836 | 11 | // the total number of elements). |
8837 | 11 | ElemIdx0 = (8 - Index) % 8; |
8838 | 11 | ElemIdx1 = (9 - Index) % 8; |
8839 | 11 | ElemIdx2 = (10 - Index) % 8; |
8840 | 11 | ElemIdx3 = (11 - Index) % 8; |
8841 | 22 | } else { |
8842 | 11 | // Big endian ElemIdx<N> = Index + N |
8843 | 11 | ElemIdx0 = Index; |
8844 | 11 | ElemIdx1 = Index + 1; |
8845 | 11 | ElemIdx2 = Index + 2; |
8846 | 11 | ElemIdx3 = Index + 3; |
8847 | 11 | } |
8848 | 22 | |
8849 | 22 | Constant *ShuffleElts[4] = {ConstantInt::get(Int32Ty, ElemIdx0), |
8850 | 22 | ConstantInt::get(Int32Ty, ElemIdx1), |
8851 | 22 | ConstantInt::get(Int32Ty, ElemIdx2), |
8852 | 22 | ConstantInt::get(Int32Ty, ElemIdx3)}; |
8853 | 22 | |
8854 | 22 | Constant *ShuffleMask = llvm::ConstantVector::get(ShuffleElts); |
8855 | 22 | Value *ShuffleCall = |
8856 | 22 | Builder.CreateShuffleVector(Ops[0], Ops[1], ShuffleMask); |
8857 | 22 | QualType BIRetType = E->getType(); |
8858 | 22 | auto RetTy = ConvertType(BIRetType); |
8859 | 22 | return Builder.CreateBitCast(ShuffleCall, RetTy); |
8860 | 0 | } |
8861 | 1.01k | } |
8862 | 1.01k | } |
8863 | | |
8864 | | Value *CodeGenFunction::EmitAMDGPUBuiltinExpr(unsigned BuiltinID, |
8865 | 105 | const CallExpr *E) { |
8866 | 105 | switch (BuiltinID) { |
8867 | 4 | case AMDGPU::BI__builtin_amdgcn_div_scale: |
8868 | 4 | case AMDGPU::BI__builtin_amdgcn_div_scalef: { |
8869 | 4 | // Translate from the intrinsics's struct return to the builtin's out |
8870 | 4 | // argument. |
8871 | 4 | |
8872 | 4 | Address FlagOutPtr = EmitPointerWithAlignment(E->getArg(3)); |
8873 | 4 | |
8874 | 4 | llvm::Value *X = EmitScalarExpr(E->getArg(0)); |
8875 | 4 | llvm::Value *Y = EmitScalarExpr(E->getArg(1)); |
8876 | 4 | llvm::Value *Z = EmitScalarExpr(E->getArg(2)); |
8877 | 4 | |
8878 | 4 | llvm::Value *Callee = CGM.getIntrinsic(Intrinsic::amdgcn_div_scale, |
8879 | 4 | X->getType()); |
8880 | 4 | |
8881 | 4 | llvm::Value *Tmp = Builder.CreateCall(Callee, {X, Y, Z}); |
8882 | 4 | |
8883 | 4 | llvm::Value *Result = Builder.CreateExtractValue(Tmp, 0); |
8884 | 4 | llvm::Value *Flag = Builder.CreateExtractValue(Tmp, 1); |
8885 | 4 | |
8886 | 4 | llvm::Type *RealFlagType |
8887 | 4 | = FlagOutPtr.getPointer()->getType()->getPointerElementType(); |
8888 | 4 | |
8889 | 4 | llvm::Value *FlagExt = Builder.CreateZExt(Flag, RealFlagType); |
8890 | 4 | Builder.CreateStore(FlagExt, FlagOutPtr); |
8891 | 4 | return Result; |
8892 | 4 | } |
8893 | 4 | case AMDGPU::BI__builtin_amdgcn_div_fmas: |
8894 | 4 | case AMDGPU::BI__builtin_amdgcn_div_fmasf: { |
8895 | 4 | llvm::Value *Src0 = EmitScalarExpr(E->getArg(0)); |
8896 | 4 | llvm::Value *Src1 = EmitScalarExpr(E->getArg(1)); |
8897 | 4 | llvm::Value *Src2 = EmitScalarExpr(E->getArg(2)); |
8898 | 4 | llvm::Value *Src3 = EmitScalarExpr(E->getArg(3)); |
8899 | 4 | |
8900 | 4 | llvm::Value *F = CGM.getIntrinsic(Intrinsic::amdgcn_div_fmas, |
8901 | 4 | Src0->getType()); |
8902 | 4 | llvm::Value *Src3ToBool = Builder.CreateIsNotNull(Src3); |
8903 | 4 | return Builder.CreateCall(F, {Src0, Src1, Src2, Src3ToBool}); |
8904 | 4 | } |
8905 | 4 | |
8906 | 0 | case AMDGPU::BI__builtin_amdgcn_ds_swizzle: |
8907 | 0 | return emitBinaryBuiltin(*this, E, Intrinsic::amdgcn_ds_swizzle); |
8908 | 2 | case AMDGPU::BI__builtin_amdgcn_mov_dpp: { |
8909 | 2 | llvm::SmallVector<llvm::Value *, 5> Args; |
8910 | 12 | for (unsigned I = 0; I != 512 ; ++I10 ) |
8911 | 10 | Args.push_back(EmitScalarExpr(E->getArg(I))); |
8912 | 2 | Value *F = CGM.getIntrinsic(Intrinsic::amdgcn_mov_dpp, |
8913 | 2 | Args[0]->getType()); |
8914 | 2 | return Builder.CreateCall(F, Args); |
8915 | 4 | } |
8916 | 6 | case AMDGPU::BI__builtin_amdgcn_div_fixup: |
8917 | 6 | case AMDGPU::BI__builtin_amdgcn_div_fixupf: |
8918 | 6 | case AMDGPU::BI__builtin_amdgcn_div_fixuph: |
8919 | 6 | return emitTernaryBuiltin(*this, E, Intrinsic::amdgcn_div_fixup); |
8920 | 4 | case AMDGPU::BI__builtin_amdgcn_trig_preop: |
8921 | 4 | case AMDGPU::BI__builtin_amdgcn_trig_preopf: |
8922 | 4 | return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_trig_preop); |
8923 | 6 | case AMDGPU::BI__builtin_amdgcn_rcp: |
8924 | 6 | case AMDGPU::BI__builtin_amdgcn_rcpf: |
8925 | 6 | case AMDGPU::BI__builtin_amdgcn_rcph: |
8926 | 6 | return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rcp); |
8927 | 6 | case AMDGPU::BI__builtin_amdgcn_rsq: |
8928 | 6 | case AMDGPU::BI__builtin_amdgcn_rsqf: |
8929 | 6 | case AMDGPU::BI__builtin_amdgcn_rsqh: |
8930 | 6 | return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rsq); |
8931 | 4 | case AMDGPU::BI__builtin_amdgcn_rsq_clamp: |
8932 | 4 | case AMDGPU::BI__builtin_amdgcn_rsq_clampf: |
8933 | 4 | return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rsq_clamp); |
8934 | 4 | case AMDGPU::BI__builtin_amdgcn_sinf: |
8935 | 4 | case AMDGPU::BI__builtin_amdgcn_sinh: |
8936 | 4 | return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_sin); |
8937 | 4 | case AMDGPU::BI__builtin_amdgcn_cosf: |
8938 | 4 | case AMDGPU::BI__builtin_amdgcn_cosh: |
8939 | 4 | return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_cos); |
8940 | 2 | case AMDGPU::BI__builtin_amdgcn_log_clampf: |
8941 | 2 | return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_log_clamp); |
8942 | 6 | case AMDGPU::BI__builtin_amdgcn_ldexp: |
8943 | 6 | case AMDGPU::BI__builtin_amdgcn_ldexpf: |
8944 | 6 | case AMDGPU::BI__builtin_amdgcn_ldexph: |
8945 | 6 | return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_ldexp); |
8946 | 6 | case AMDGPU::BI__builtin_amdgcn_frexp_mant: |
8947 | 6 | case AMDGPU::BI__builtin_amdgcn_frexp_mantf: |
8948 | 6 | case AMDGPU::BI__builtin_amdgcn_frexp_manth: |
8949 | 6 | return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_frexp_mant); |
8950 | 4 | case AMDGPU::BI__builtin_amdgcn_frexp_exp: |
8951 | 4 | case AMDGPU::BI__builtin_amdgcn_frexp_expf: { |
8952 | 4 | Value *Src0 = EmitScalarExpr(E->getArg(0)); |
8953 | 4 | Value *F = CGM.getIntrinsic(Intrinsic::amdgcn_frexp_exp, |
8954 | 4 | { Builder.getInt32Ty(), Src0->getType() }); |
8955 | 4 | return Builder.CreateCall(F, Src0); |
8956 | 4 | } |
8957 | 2 | case AMDGPU::BI__builtin_amdgcn_frexp_exph: { |
8958 | 2 | Value *Src0 = EmitScalarExpr(E->getArg(0)); |
8959 | 2 | Value *F = CGM.getIntrinsic(Intrinsic::amdgcn_frexp_exp, |
8960 | 2 | { Builder.getInt16Ty(), Src0->getType() }); |
8961 | 2 | return Builder.CreateCall(F, Src0); |
8962 | 4 | } |
8963 | 6 | case AMDGPU::BI__builtin_amdgcn_fract: |
8964 | 6 | case AMDGPU::BI__builtin_amdgcn_fractf: |
8965 | 6 | case AMDGPU::BI__builtin_amdgcn_fracth: |
8966 | 6 | return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_fract); |
8967 | 0 | case AMDGPU::BI__builtin_amdgcn_lerp: |
8968 | 0 | return emitTernaryBuiltin(*this, E, Intrinsic::amdgcn_lerp); |
8969 | 8 | case AMDGPU::BI__builtin_amdgcn_uicmp: |
8970 | 8 | case AMDGPU::BI__builtin_amdgcn_uicmpl: |
8971 | 8 | case AMDGPU::BI__builtin_amdgcn_sicmp: |
8972 | 8 | case AMDGPU::BI__builtin_amdgcn_sicmpl: |
8973 | 8 | return emitTernaryBuiltin(*this, E, Intrinsic::amdgcn_icmp); |
8974 | 4 | case AMDGPU::BI__builtin_amdgcn_fcmp: |
8975 | 4 | case AMDGPU::BI__builtin_amdgcn_fcmpf: |
8976 | 4 | return emitTernaryBuiltin(*this, E, Intrinsic::amdgcn_fcmp); |
8977 | 6 | case AMDGPU::BI__builtin_amdgcn_class: |
8978 | 6 | case AMDGPU::BI__builtin_amdgcn_classf: |
8979 | 6 | case AMDGPU::BI__builtin_amdgcn_classh: |
8980 | 6 | return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_class); |
8981 | 5 | case AMDGPU::BI__builtin_amdgcn_fmed3f: |
8982 | 5 | case AMDGPU::BI__builtin_amdgcn_fmed3h: |
8983 | 5 | return emitTernaryBuiltin(*this, E, Intrinsic::amdgcn_fmed3); |
8984 | 2 | case AMDGPU::BI__builtin_amdgcn_read_exec: { |
8985 | 2 | CallInst *CI = cast<CallInst>( |
8986 | 2 | EmitSpecialRegisterBuiltin(*this, E, Int64Ty, Int64Ty, true, "exec")); |
8987 | 2 | CI->setConvergent(); |
8988 | 2 | return CI; |
8989 | 5 | } |
8990 | 5 | |
8991 | 5 | // amdgcn workitem |
8992 | 2 | case AMDGPU::BI__builtin_amdgcn_workitem_id_x: |
8993 | 2 | return emitRangedBuiltin(*this, Intrinsic::amdgcn_workitem_id_x, 0, 1024); |
8994 | 2 | case AMDGPU::BI__builtin_amdgcn_workitem_id_y: |
8995 | 2 | return emitRangedBuiltin(*this, Intrinsic::amdgcn_workitem_id_y, 0, 1024); |
8996 | 2 | case AMDGPU::BI__builtin_amdgcn_workitem_id_z: |
8997 | 2 | return emitRangedBuiltin(*this, Intrinsic::amdgcn_workitem_id_z, 0, 1024); |
8998 | 5 | |
8999 | 5 | // r600 intrinsics |
9000 | 1 | case AMDGPU::BI__builtin_r600_recipsqrt_ieee: |
9001 | 1 | case AMDGPU::BI__builtin_r600_recipsqrt_ieeef: |
9002 | 1 | return emitUnaryBuiltin(*this, E, Intrinsic::r600_recipsqrt_ieee); |
9003 | 1 | case AMDGPU::BI__builtin_r600_read_tidig_x: |
9004 | 1 | return emitRangedBuiltin(*this, Intrinsic::r600_read_tidig_x, 0, 1024); |
9005 | 1 | case AMDGPU::BI__builtin_r600_read_tidig_y: |
9006 | 1 | return emitRangedBuiltin(*this, Intrinsic::r600_read_tidig_y, 0, 1024); |
9007 | 1 | case AMDGPU::BI__builtin_r600_read_tidig_z: |
9008 | 1 | return emitRangedBuiltin(*this, Intrinsic::r600_read_tidig_z, 0, 1024); |
9009 | 0 | default: |
9010 | 0 | return nullptr; |
9011 | 0 | } |
9012 | 0 | } |
9013 | | |
9014 | | /// Handle a SystemZ function in which the final argument is a pointer |
9015 | | /// to an int that receives the post-instruction CC value. At the LLVM level |
9016 | | /// this is represented as a function that returns a {result, cc} pair. |
9017 | | static Value *EmitSystemZIntrinsicWithCC(CodeGenFunction &CGF, |
9018 | | unsigned IntrinsicID, |
9019 | 626 | const CallExpr *E) { |
9020 | 626 | unsigned NumArgs = E->getNumArgs() - 1; |
9021 | 626 | SmallVector<Value *, 8> Args(NumArgs); |
9022 | 1.99k | for (unsigned I = 0; I < NumArgs1.99k ; ++I1.36k ) |
9023 | 1.36k | Args[I] = CGF.EmitScalarExpr(E->getArg(I)); |
9024 | 626 | Address CCPtr = CGF.EmitPointerWithAlignment(E->getArg(NumArgs)); |
9025 | 626 | Value *F = CGF.CGM.getIntrinsic(IntrinsicID); |
9026 | 626 | Value *Call = CGF.Builder.CreateCall(F, Args); |
9027 | 626 | Value *CC = CGF.Builder.CreateExtractValue(Call, 1); |
9028 | 626 | CGF.Builder.CreateStore(CC, CCPtr); |
9029 | 626 | return CGF.Builder.CreateExtractValue(Call, 0); |
9030 | 626 | } |
9031 | | |
9032 | | Value *CodeGenFunction::EmitSystemZBuiltinExpr(unsigned BuiltinID, |
9033 | 815 | const CallExpr *E) { |
9034 | 815 | switch (BuiltinID) { |
9035 | 13 | case SystemZ::BI__builtin_tbegin: { |
9036 | 13 | Value *TDB = EmitScalarExpr(E->getArg(0)); |
9037 | 13 | Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c); |
9038 | 13 | Value *F = CGM.getIntrinsic(Intrinsic::s390_tbegin); |
9039 | 13 | return Builder.CreateCall(F, {TDB, Control}); |
9040 | 815 | } |
9041 | 16 | case SystemZ::BI__builtin_tbegin_nofloat: { |
9042 | 16 | Value *TDB = EmitScalarExpr(E->getArg(0)); |
9043 | 16 | Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c); |
9044 | 16 | Value *F = CGM.getIntrinsic(Intrinsic::s390_tbegin_nofloat); |
9045 | 16 | return Builder.CreateCall(F, {TDB, Control}); |
9046 | 815 | } |
9047 | 2 | case SystemZ::BI__builtin_tbeginc: { |
9048 | 2 | Value *TDB = llvm::ConstantPointerNull::get(Int8PtrTy); |
9049 | 2 | Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff08); |
9050 | 2 | Value *F = CGM.getIntrinsic(Intrinsic::s390_tbeginc); |
9051 | 2 | return Builder.CreateCall(F, {TDB, Control}); |
9052 | 815 | } |
9053 | 10 | case SystemZ::BI__builtin_tabort: { |
9054 | 10 | Value *Data = EmitScalarExpr(E->getArg(0)); |
9055 | 10 | Value *F = CGM.getIntrinsic(Intrinsic::s390_tabort); |
9056 | 10 | return Builder.CreateCall(F, Builder.CreateSExt(Data, Int64Ty, "tabort")); |
9057 | 815 | } |
9058 | 22 | case SystemZ::BI__builtin_non_tx_store: { |
9059 | 22 | Value *Address = EmitScalarExpr(E->getArg(0)); |
9060 | 22 | Value *Data = EmitScalarExpr(E->getArg(1)); |
9061 | 22 | Value *F = CGM.getIntrinsic(Intrinsic::s390_ntstg); |
9062 | 22 | return Builder.CreateCall(F, {Data, Address}); |
9063 | 815 | } |
9064 | 815 | |
9065 | 815 | // Vector builtins. Note that most vector builtins are mapped automatically |
9066 | 815 | // to target-specific LLVM intrinsics. The ones handled specially here can |
9067 | 815 | // be represented via standard LLVM IR, which is preferable to enable common |
9068 | 815 | // LLVM optimizations. |
9069 | 815 | |
9070 | 12 | case SystemZ::BI__builtin_s390_vpopctb: |
9071 | 12 | case SystemZ::BI__builtin_s390_vpopcth: |
9072 | 12 | case SystemZ::BI__builtin_s390_vpopctf: |
9073 | 12 | case SystemZ::BI__builtin_s390_vpopctg: { |
9074 | 12 | llvm::Type *ResultType = ConvertType(E->getType()); |
9075 | 12 | Value *X = EmitScalarExpr(E->getArg(0)); |
9076 | 12 | Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ResultType); |
9077 | 12 | return Builder.CreateCall(F, X); |
9078 | 12 | } |
9079 | 12 | |
9080 | 12 | case SystemZ::BI__builtin_s390_vclzb: |
9081 | 12 | case SystemZ::BI__builtin_s390_vclzh: |
9082 | 12 | case SystemZ::BI__builtin_s390_vclzf: |
9083 | 12 | case SystemZ::BI__builtin_s390_vclzg: { |
9084 | 12 | llvm::Type *ResultType = ConvertType(E->getType()); |
9085 | 12 | Value *X = EmitScalarExpr(E->getArg(0)); |
9086 | 12 | Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); |
9087 | 12 | Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ResultType); |
9088 | 12 | return Builder.CreateCall(F, {X, Undef}); |
9089 | 12 | } |
9090 | 12 | |
9091 | 12 | case SystemZ::BI__builtin_s390_vctzb: |
9092 | 12 | case SystemZ::BI__builtin_s390_vctzh: |
9093 | 12 | case SystemZ::BI__builtin_s390_vctzf: |
9094 | 12 | case SystemZ::BI__builtin_s390_vctzg: { |
9095 | 12 | llvm::Type *ResultType = ConvertType(E->getType()); |
9096 | 12 | Value *X = EmitScalarExpr(E->getArg(0)); |
9097 | 12 | Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); |
9098 | 12 | Function *F = CGM.getIntrinsic(Intrinsic::cttz, ResultType); |
9099 | 12 | return Builder.CreateCall(F, {X, Undef}); |
9100 | 12 | } |
9101 | 12 | |
9102 | 5 | case SystemZ::BI__builtin_s390_vfsqsb: |
9103 | 5 | case SystemZ::BI__builtin_s390_vfsqdb: { |
9104 | 5 | llvm::Type *ResultType = ConvertType(E->getType()); |
9105 | 5 | Value *X = EmitScalarExpr(E->getArg(0)); |
9106 | 5 | Function *F = CGM.getIntrinsic(Intrinsic::sqrt, ResultType); |
9107 | 5 | return Builder.CreateCall(F, X); |
9108 | 5 | } |
9109 | 5 | case SystemZ::BI__builtin_s390_vfmasb: |
9110 | 5 | case SystemZ::BI__builtin_s390_vfmadb: { |
9111 | 5 | llvm::Type *ResultType = ConvertType(E->getType()); |
9112 | 5 | Value *X = EmitScalarExpr(E->getArg(0)); |
9113 | 5 | Value *Y = EmitScalarExpr(E->getArg(1)); |
9114 | 5 | Value *Z = EmitScalarExpr(E->getArg(2)); |
9115 | 5 | Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); |
9116 | 5 | return Builder.CreateCall(F, {X, Y, Z}); |
9117 | 5 | } |
9118 | 5 | case SystemZ::BI__builtin_s390_vfmssb: |
9119 | 5 | case SystemZ::BI__builtin_s390_vfmsdb: { |
9120 | 5 | llvm::Type *ResultType = ConvertType(E->getType()); |
9121 | 5 | Value *X = EmitScalarExpr(E->getArg(0)); |
9122 | 5 | Value *Y = EmitScalarExpr(E->getArg(1)); |
9123 | 5 | Value *Z = EmitScalarExpr(E->getArg(2)); |
9124 | 5 | Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); |
9125 | 5 | Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); |
9126 | 5 | return Builder.CreateCall(F, {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); |
9127 | 5 | } |
9128 | 4 | case SystemZ::BI__builtin_s390_vfnmasb: |
9129 | 4 | case SystemZ::BI__builtin_s390_vfnmadb: { |
9130 | 4 | llvm::Type *ResultType = ConvertType(E->getType()); |
9131 | 4 | Value *X = EmitScalarExpr(E->getArg(0)); |
9132 | 4 | Value *Y = EmitScalarExpr(E->getArg(1)); |
9133 | 4 | Value *Z = EmitScalarExpr(E->getArg(2)); |
9134 | 4 | Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); |
9135 | 4 | Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); |
9136 | 4 | return Builder.CreateFSub(Zero, Builder.CreateCall(F, {X, Y, Z}), "sub"); |
9137 | 4 | } |
9138 | 4 | case SystemZ::BI__builtin_s390_vfnmssb: |
9139 | 4 | case SystemZ::BI__builtin_s390_vfnmsdb: { |
9140 | 4 | llvm::Type *ResultType = ConvertType(E->getType()); |
9141 | 4 | Value *X = EmitScalarExpr(E->getArg(0)); |
9142 | 4 | Value *Y = EmitScalarExpr(E->getArg(1)); |
9143 | 4 | Value *Z = EmitScalarExpr(E->getArg(2)); |
9144 | 4 | Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); |
9145 | 4 | Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); |
9146 | 4 | Value *NegZ = Builder.CreateFSub(Zero, Z, "sub"); |
9147 | 4 | return Builder.CreateFSub(Zero, Builder.CreateCall(F, {X, Y, NegZ})); |
9148 | 4 | } |
9149 | 5 | case SystemZ::BI__builtin_s390_vflpsb: |
9150 | 5 | case SystemZ::BI__builtin_s390_vflpdb: { |
9151 | 5 | llvm::Type *ResultType = ConvertType(E->getType()); |
9152 | 5 | Value *X = EmitScalarExpr(E->getArg(0)); |
9153 | 5 | Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); |
9154 | 5 | return Builder.CreateCall(F, X); |
9155 | 5 | } |
9156 | 5 | case SystemZ::BI__builtin_s390_vflnsb: |
9157 | 5 | case SystemZ::BI__builtin_s390_vflndb: { |
9158 | 5 | llvm::Type *ResultType = ConvertType(E->getType()); |
9159 | 5 | Value *X = EmitScalarExpr(E->getArg(0)); |
9160 | 5 | Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); |
9161 | 5 | Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); |
9162 | 5 | return Builder.CreateFSub(Zero, Builder.CreateCall(F, X), "sub"); |
9163 | 5 | } |
9164 | 41 | case SystemZ::BI__builtin_s390_vfisb: |
9165 | 41 | case SystemZ::BI__builtin_s390_vfidb: { |
9166 | 41 | llvm::Type *ResultType = ConvertType(E->getType()); |
9167 | 41 | Value *X = EmitScalarExpr(E->getArg(0)); |
9168 | 41 | // Constant-fold the M4 and M5 mask arguments. |
9169 | 41 | llvm::APSInt M4, M5; |
9170 | 41 | bool IsConstM4 = E->getArg(1)->isIntegerConstantExpr(M4, getContext()); |
9171 | 41 | bool IsConstM5 = E->getArg(2)->isIntegerConstantExpr(M5, getContext()); |
9172 | 41 | assert(IsConstM4 && IsConstM5 && "Constant arg isn't actually constant?"); |
9173 | 41 | (void)IsConstM4; (void)IsConstM5; |
9174 | 41 | // Check whether this instance can be represented via a LLVM standard |
9175 | 41 | // intrinsic. We only support some combinations of M4 and M5. |
9176 | 41 | Intrinsic::ID ID = Intrinsic::not_intrinsic; |
9177 | 41 | switch (M4.getZExtValue()) { |
9178 | 0 | default: break; |
9179 | 5 | case 0: // IEEE-inexact exception allowed |
9180 | 5 | switch (M5.getZExtValue()) { |
9181 | 0 | default: break; |
9182 | 5 | case 0: ID = Intrinsic::rint; break; |
9183 | 5 | } |
9184 | 5 | break; |
9185 | 36 | case 4: // IEEE-inexact exception suppressed |
9186 | 36 | switch (M5.getZExtValue()) { |
9187 | 5 | default: break; |
9188 | 5 | case 0: ID = Intrinsic::nearbyint; break; |
9189 | 2 | case 1: ID = Intrinsic::round; break; |
9190 | 8 | case 5: ID = Intrinsic::trunc; break; |
9191 | 8 | case 6: ID = Intrinsic::ceil; break; |
9192 | 8 | case 7: ID = Intrinsic::floor; break; |
9193 | 36 | } |
9194 | 36 | break; |
9195 | 41 | } |
9196 | 41 | if (41 ID != Intrinsic::not_intrinsic41 ) { |
9197 | 36 | Function *F = CGM.getIntrinsic(ID, ResultType); |
9198 | 36 | return Builder.CreateCall(F, X); |
9199 | 36 | } |
9200 | 5 | switch (BuiltinID) { |
9201 | 2 | case SystemZ::BI__builtin_s390_vfisb: ID = Intrinsic::s390_vfisb; break; |
9202 | 3 | case SystemZ::BI__builtin_s390_vfidb: ID = Intrinsic::s390_vfidb; break; |
9203 | 0 | default: 0 llvm_unreachable0 ("Unknown BuiltinID"); |
9204 | 5 | } |
9205 | 5 | Function *F = CGM.getIntrinsic(ID); |
9206 | 5 | Value *M4Value = llvm::ConstantInt::get(getLLVMContext(), M4); |
9207 | 5 | Value *M5Value = llvm::ConstantInt::get(getLLVMContext(), M5); |
9208 | 5 | return Builder.CreateCall(F, {X, M4Value, M5Value}); |
9209 | 5 | } |
9210 | 8 | case SystemZ::BI__builtin_s390_vfmaxsb: |
9211 | 8 | case SystemZ::BI__builtin_s390_vfmaxdb: { |
9212 | 8 | llvm::Type *ResultType = ConvertType(E->getType()); |
9213 | 8 | Value *X = EmitScalarExpr(E->getArg(0)); |
9214 | 8 | Value *Y = EmitScalarExpr(E->getArg(1)); |
9215 | 8 | // Constant-fold the M4 mask argument. |
9216 | 8 | llvm::APSInt M4; |
9217 | 8 | bool IsConstM4 = E->getArg(2)->isIntegerConstantExpr(M4, getContext()); |
9218 | 8 | assert(IsConstM4 && "Constant arg isn't actually constant?"); |
9219 | 8 | (void)IsConstM4; |
9220 | 8 | // Check whether this instance can be represented via a LLVM standard |
9221 | 8 | // intrinsic. We only support some values of M4. |
9222 | 8 | Intrinsic::ID ID = Intrinsic::not_intrinsic; |
9223 | 8 | switch (M4.getZExtValue()) { |
9224 | 6 | default: break; |
9225 | 2 | case 4: ID = Intrinsic::maxnum; break; |
9226 | 8 | } |
9227 | 8 | if (8 ID != Intrinsic::not_intrinsic8 ) { |
9228 | 2 | Function *F = CGM.getIntrinsic(ID, ResultType); |
9229 | 2 | return Builder.CreateCall(F, {X, Y}); |
9230 | 2 | } |
9231 | 6 | switch (BuiltinID) { |
9232 | 3 | case SystemZ::BI__builtin_s390_vfmaxsb: ID = Intrinsic::s390_vfmaxsb; break; |
9233 | 3 | case SystemZ::BI__builtin_s390_vfmaxdb: ID = Intrinsic::s390_vfmaxdb; break; |
9234 | 0 | default: 0 llvm_unreachable0 ("Unknown BuiltinID"); |
9235 | 6 | } |
9236 | 6 | Function *F = CGM.getIntrinsic(ID); |
9237 | 6 | Value *M4Value = llvm::ConstantInt::get(getLLVMContext(), M4); |
9238 | 6 | return Builder.CreateCall(F, {X, Y, M4Value}); |
9239 | 6 | } |
9240 | 8 | case SystemZ::BI__builtin_s390_vfminsb: |
9241 | 8 | case SystemZ::BI__builtin_s390_vfmindb: { |
9242 | 8 | llvm::Type *ResultType = ConvertType(E->getType()); |
9243 | 8 | Value *X = EmitScalarExpr(E->getArg(0)); |
9244 | 8 | Value *Y = EmitScalarExpr(E->getArg(1)); |
9245 | 8 | // Constant-fold the M4 mask argument. |
9246 | 8 | llvm::APSInt M4; |
9247 | 8 | bool IsConstM4 = E->getArg(2)->isIntegerConstantExpr(M4, getContext()); |
9248 | 8 | assert(IsConstM4 && "Constant arg isn't actually constant?"); |
9249 | 8 | (void)IsConstM4; |
9250 | 8 | // Check whether this instance can be represented via a LLVM standard |
9251 | 8 | // intrinsic. We only support some values of M4. |
9252 | 8 | Intrinsic::ID ID = Intrinsic::not_intrinsic; |
9253 | 8 | switch (M4.getZExtValue()) { |
9254 | 6 | default: break; |
9255 | 2 | case 4: ID = Intrinsic::minnum; break; |
9256 | 8 | } |
9257 | 8 | if (8 ID != Intrinsic::not_intrinsic8 ) { |
9258 | 2 | Function *F = CGM.getIntrinsic(ID, ResultType); |
9259 | 2 | return Builder.CreateCall(F, {X, Y}); |
9260 | 2 | } |
9261 | 6 | switch (BuiltinID) { |
9262 | 3 | case SystemZ::BI__builtin_s390_vfminsb: ID = Intrinsic::s390_vfminsb; break; |
9263 | 3 | case SystemZ::BI__builtin_s390_vfmindb: ID = Intrinsic::s390_vfmindb; break; |
9264 | 0 | default: 0 llvm_unreachable0 ("Unknown BuiltinID"); |
9265 | 6 | } |
9266 | 6 | Function *F = CGM.getIntrinsic(ID); |
9267 | 6 | Value *M4Value = llvm::ConstantInt::get(getLLVMContext(), M4); |
9268 | 6 | return Builder.CreateCall(F, {X, Y, M4Value}); |
9269 | 6 | } |
9270 | 6 | |
9271 | 6 | // Vector intrisincs that output the post-instruction CC value. |
9272 | 6 | |
9273 | 6 | #define INTRINSIC_WITH_CC(NAME) \ |
9274 | 626 | case SystemZ::BI__builtin_##NAME: \ |
9275 | 626 | return EmitSystemZIntrinsicWithCC(*this, Intrinsic::NAME, E) |
9276 | 6 | |
9277 | 2 | INTRINSIC_WITH_CC2 (s390_vpkshs); |
9278 | 2 | INTRINSIC_WITH_CC2 (s390_vpksfs); |
9279 | 2 | INTRINSIC_WITH_CC2 (s390_vpksgs); |
9280 | 6 | |
9281 | 3 | INTRINSIC_WITH_CC3 (s390_vpklshs); |
9282 | 3 | INTRINSIC_WITH_CC3 (s390_vpklsfs); |
9283 | 3 | INTRINSIC_WITH_CC3 (s390_vpklsgs); |
9284 | 6 | |
9285 | 29 | INTRINSIC_WITH_CC29 (s390_vceqbs); |
9286 | 29 | INTRINSIC_WITH_CC29 (s390_vceqhs); |
9287 | 29 | INTRINSIC_WITH_CC29 (s390_vceqfs); |
9288 | 29 | INTRINSIC_WITH_CC29 (s390_vceqgs); |
9289 | 6 | |
9290 | 25 | INTRINSIC_WITH_CC25 (s390_vchbs); |
9291 | 25 | INTRINSIC_WITH_CC25 (s390_vchhs); |
9292 | 25 | INTRINSIC_WITH_CC25 (s390_vchfs); |
9293 | 25 | INTRINSIC_WITH_CC25 (s390_vchgs); |
9294 | 6 | |
9295 | 33 | INTRINSIC_WITH_CC33 (s390_vchlbs); |
9296 | 33 | INTRINSIC_WITH_CC33 (s390_vchlhs); |
9297 | 33 | INTRINSIC_WITH_CC33 (s390_vchlfs); |
9298 | 33 | INTRINSIC_WITH_CC33 (s390_vchlgs); |
9299 | 6 | |
9300 | 14 | INTRINSIC_WITH_CC14 (s390_vfaebs); |
9301 | 14 | INTRINSIC_WITH_CC14 (s390_vfaehs); |
9302 | 14 | INTRINSIC_WITH_CC14 (s390_vfaefs); |
9303 | 6 | |
9304 | 8 | INTRINSIC_WITH_CC8 (s390_vfaezbs); |
9305 | 8 | INTRINSIC_WITH_CC8 (s390_vfaezhs); |
9306 | 8 | INTRINSIC_WITH_CC8 (s390_vfaezfs); |
9307 | 6 | |
9308 | 4 | INTRINSIC_WITH_CC4 (s390_vfeebs); |
9309 | 4 | INTRINSIC_WITH_CC4 (s390_vfeehs); |
9310 | 4 | INTRINSIC_WITH_CC4 (s390_vfeefs); |
9311 | 6 | |
9312 | 4 | INTRINSIC_WITH_CC4 (s390_vfeezbs); |
9313 | 4 | INTRINSIC_WITH_CC4 (s390_vfeezhs); |
9314 | 4 | INTRINSIC_WITH_CC4 (s390_vfeezfs); |
9315 | 6 | |
9316 | 4 | INTRINSIC_WITH_CC4 (s390_vfenebs); |
9317 | 4 | INTRINSIC_WITH_CC4 (s390_vfenehs); |
9318 | 4 | INTRINSIC_WITH_CC4 (s390_vfenefs); |
9319 | 6 | |
9320 | 4 | INTRINSIC_WITH_CC4 (s390_vfenezbs); |
9321 | 4 | INTRINSIC_WITH_CC4 (s390_vfenezhs); |
9322 | 4 | INTRINSIC_WITH_CC4 (s390_vfenezfs); |
9323 | 6 | |
9324 | 4 | INTRINSIC_WITH_CC4 (s390_vistrbs); |
9325 | 4 | INTRINSIC_WITH_CC4 (s390_vistrhs); |
9326 | 4 | INTRINSIC_WITH_CC4 (s390_vistrfs); |
9327 | 6 | |
9328 | 6 | INTRINSIC_WITH_CC6 (s390_vstrcbs); |
9329 | 6 | INTRINSIC_WITH_CC6 (s390_vstrchs); |
9330 | 6 | INTRINSIC_WITH_CC6 (s390_vstrcfs); |
9331 | 6 | |
9332 | 4 | INTRINSIC_WITH_CC4 (s390_vstrczbs); |
9333 | 4 | INTRINSIC_WITH_CC4 (s390_vstrczhs); |
9334 | 4 | INTRINSIC_WITH_CC4 (s390_vstrczfs); |
9335 | 6 | |
9336 | 5 | INTRINSIC_WITH_CC5 (s390_vfcesbs); |
9337 | 9 | INTRINSIC_WITH_CC9 (s390_vfcedbs); |
9338 | 9 | INTRINSIC_WITH_CC9 (s390_vfchsbs); |
9339 | 17 | INTRINSIC_WITH_CC17 (s390_vfchdbs); |
9340 | 8 | INTRINSIC_WITH_CC8 (s390_vfchesbs); |
9341 | 17 | INTRINSIC_WITH_CC17 (s390_vfchedbs); |
9342 | 6 | |
9343 | 8 | INTRINSIC_WITH_CC8 (s390_vftcisb); |
9344 | 34 | INTRINSIC_WITH_CC34 (s390_vftcidb); |
9345 | 6 | |
9346 | 6 | #undef INTRINSIC_WITH_CC |
9347 | 6 | |
9348 | 0 | default: |
9349 | 0 | return nullptr; |
9350 | 0 | } |
9351 | 0 | } |
9352 | | |
9353 | | Value *CodeGenFunction::EmitNVPTXBuiltinExpr(unsigned BuiltinID, |
9354 | 370 | const CallExpr *E) { |
9355 | 54 | auto MakeLdg = [&](unsigned IntrinsicID) { |
9356 | 54 | Value *Ptr = EmitScalarExpr(E->getArg(0)); |
9357 | 54 | clang::CharUnits Align = |
9358 | 54 | getNaturalPointeeTypeAlignment(E->getArg(0)->getType()); |
9359 | 54 | return Builder.CreateCall( |
9360 | 54 | CGM.getIntrinsic(IntrinsicID, {Ptr->getType()->getPointerElementType(), |
9361 | 54 | Ptr->getType()}), |
9362 | 54 | {Ptr, ConstantInt::get(Builder.getInt32Ty(), Align.getQuantity())}); |
9363 | 54 | }; |
9364 | 186 | auto MakeScopedAtomic = [&](unsigned IntrinsicID) { |
9365 | 186 | Value *Ptr = EmitScalarExpr(E->getArg(0)); |
9366 | 186 | return Builder.CreateCall( |
9367 | 186 | CGM.getIntrinsic(IntrinsicID, {Ptr->getType()->getPointerElementType(), |
9368 | 186 | Ptr->getType()}), |
9369 | 186 | {Ptr, EmitScalarExpr(E->getArg(1))}); |
9370 | 186 | }; |
9371 | 370 | switch (BuiltinID) { |
9372 | 9 | case NVPTX::BI__nvvm_atom_add_gen_i: |
9373 | 9 | case NVPTX::BI__nvvm_atom_add_gen_l: |
9374 | 9 | case NVPTX::BI__nvvm_atom_add_gen_ll: |
9375 | 9 | return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Add, E); |
9376 | 9 | |
9377 | 9 | case NVPTX::BI__nvvm_atom_sub_gen_i: |
9378 | 9 | case NVPTX::BI__nvvm_atom_sub_gen_l: |
9379 | 9 | case NVPTX::BI__nvvm_atom_sub_gen_ll: |
9380 | 9 | return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Sub, E); |
9381 | 9 | |
9382 | 9 | case NVPTX::BI__nvvm_atom_and_gen_i: |
9383 | 9 | case NVPTX::BI__nvvm_atom_and_gen_l: |
9384 | 9 | case NVPTX::BI__nvvm_atom_and_gen_ll: |
9385 | 9 | return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::And, E); |
9386 | 9 | |
9387 | 9 | case NVPTX::BI__nvvm_atom_or_gen_i: |
9388 | 9 | case NVPTX::BI__nvvm_atom_or_gen_l: |
9389 | 9 | case NVPTX::BI__nvvm_atom_or_gen_ll: |
9390 | 9 | return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Or, E); |
9391 | 9 | |
9392 | 9 | case NVPTX::BI__nvvm_atom_xor_gen_i: |
9393 | 9 | case NVPTX::BI__nvvm_atom_xor_gen_l: |
9394 | 9 | case NVPTX::BI__nvvm_atom_xor_gen_ll: |
9395 | 9 | return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Xor, E); |
9396 | 9 | |
9397 | 9 | case NVPTX::BI__nvvm_atom_xchg_gen_i: |
9398 | 9 | case NVPTX::BI__nvvm_atom_xchg_gen_l: |
9399 | 9 | case NVPTX::BI__nvvm_atom_xchg_gen_ll: |
9400 | 9 | return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Xchg, E); |
9401 | 9 | |
9402 | 9 | case NVPTX::BI__nvvm_atom_max_gen_i: |
9403 | 9 | case NVPTX::BI__nvvm_atom_max_gen_l: |
9404 | 9 | case NVPTX::BI__nvvm_atom_max_gen_ll: |
9405 | 9 | return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Max, E); |
9406 | 9 | |
9407 | 9 | case NVPTX::BI__nvvm_atom_max_gen_ui: |
9408 | 9 | case NVPTX::BI__nvvm_atom_max_gen_ul: |
9409 | 9 | case NVPTX::BI__nvvm_atom_max_gen_ull: |
9410 | 9 | return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::UMax, E); |
9411 | 9 | |
9412 | 9 | case NVPTX::BI__nvvm_atom_min_gen_i: |
9413 | 9 | case NVPTX::BI__nvvm_atom_min_gen_l: |
9414 | 9 | case NVPTX::BI__nvvm_atom_min_gen_ll: |
9415 | 9 | return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Min, E); |
9416 | 9 | |
9417 | 9 | case NVPTX::BI__nvvm_atom_min_gen_ui: |
9418 | 9 | case NVPTX::BI__nvvm_atom_min_gen_ul: |
9419 | 9 | case NVPTX::BI__nvvm_atom_min_gen_ull: |
9420 | 9 | return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::UMin, E); |
9421 | 9 | |
9422 | 9 | case NVPTX::BI__nvvm_atom_cas_gen_i: |
9423 | 9 | case NVPTX::BI__nvvm_atom_cas_gen_l: |
9424 | 9 | case NVPTX::BI__nvvm_atom_cas_gen_ll: |
9425 | 9 | // __nvvm_atom_cas_gen_* should return the old value rather than the |
9426 | 9 | // success flag. |
9427 | 9 | return MakeAtomicCmpXchgValue(*this, E, /*ReturnBool=*/false); |
9428 | 9 | |
9429 | 3 | case NVPTX::BI__nvvm_atom_add_gen_f: { |
9430 | 3 | Value *Ptr = EmitScalarExpr(E->getArg(0)); |
9431 | 3 | Value *Val = EmitScalarExpr(E->getArg(1)); |
9432 | 3 | // atomicrmw only deals with integer arguments so we need to use |
9433 | 3 | // LLVM's nvvm_atomic_load_add_f32 intrinsic for that. |
9434 | 3 | Value *FnALAF32 = |
9435 | 3 | CGM.getIntrinsic(Intrinsic::nvvm_atomic_load_add_f32, Ptr->getType()); |
9436 | 3 | return Builder.CreateCall(FnALAF32, {Ptr, Val}); |
9437 | 9 | } |
9438 | 9 | |
9439 | 3 | case NVPTX::BI__nvvm_atom_inc_gen_ui: { |
9440 | 3 | Value *Ptr = EmitScalarExpr(E->getArg(0)); |
9441 | 3 | Value *Val = EmitScalarExpr(E->getArg(1)); |
9442 | 3 | Value *FnALI32 = |
9443 | 3 | CGM.getIntrinsic(Intrinsic::nvvm_atomic_load_inc_32, Ptr->getType()); |
9444 | 3 | return Builder.CreateCall(FnALI32, {Ptr, Val}); |
9445 | 9 | } |
9446 | 9 | |
9447 | 3 | case NVPTX::BI__nvvm_atom_dec_gen_ui: { |
9448 | 3 | Value *Ptr = EmitScalarExpr(E->getArg(0)); |
9449 | 3 | Value *Val = EmitScalarExpr(E->getArg(1)); |
9450 | 3 | Value *FnALD32 = |
9451 | 3 | CGM.getIntrinsic(Intrinsic::nvvm_atomic_load_dec_32, Ptr->getType()); |
9452 | 3 | return Builder.CreateCall(FnALD32, {Ptr, Val}); |
9453 | 9 | } |
9454 | 9 | |
9455 | 44 | case NVPTX::BI__nvvm_ldg_c: |
9456 | 44 | case NVPTX::BI__nvvm_ldg_c2: |
9457 | 44 | case NVPTX::BI__nvvm_ldg_c4: |
9458 | 44 | case NVPTX::BI__nvvm_ldg_s: |
9459 | 44 | case NVPTX::BI__nvvm_ldg_s2: |
9460 | 44 | case NVPTX::BI__nvvm_ldg_s4: |
9461 | 44 | case NVPTX::BI__nvvm_ldg_i: |
9462 | 44 | case NVPTX::BI__nvvm_ldg_i2: |
9463 | 44 | case NVPTX::BI__nvvm_ldg_i4: |
9464 | 44 | case NVPTX::BI__nvvm_ldg_l: |
9465 | 44 | case NVPTX::BI__nvvm_ldg_ll: |
9466 | 44 | case NVPTX::BI__nvvm_ldg_ll2: |
9467 | 44 | case NVPTX::BI__nvvm_ldg_uc: |
9468 | 44 | case NVPTX::BI__nvvm_ldg_uc2: |
9469 | 44 | case NVPTX::BI__nvvm_ldg_uc4: |
9470 | 44 | case NVPTX::BI__nvvm_ldg_us: |
9471 | 44 | case NVPTX::BI__nvvm_ldg_us2: |
9472 | 44 | case NVPTX::BI__nvvm_ldg_us4: |
9473 | 44 | case NVPTX::BI__nvvm_ldg_ui: |
9474 | 44 | case NVPTX::BI__nvvm_ldg_ui2: |
9475 | 44 | case NVPTX::BI__nvvm_ldg_ui4: |
9476 | 44 | case NVPTX::BI__nvvm_ldg_ul: |
9477 | 44 | case NVPTX::BI__nvvm_ldg_ull: |
9478 | 44 | case NVPTX::BI__nvvm_ldg_ull2: |
9479 | 44 | // PTX Interoperability section 2.2: "For a vector with an even number of |
9480 | 44 | // elements, its alignment is set to number of elements times the alignment |
9481 | 44 | // of its member: n*alignof(t)." |
9482 | 44 | return MakeLdg(Intrinsic::nvvm_ldg_global_i); |
9483 | 10 | case NVPTX::BI__nvvm_ldg_f: |
9484 | 10 | case NVPTX::BI__nvvm_ldg_f2: |
9485 | 10 | case NVPTX::BI__nvvm_ldg_f4: |
9486 | 10 | case NVPTX::BI__nvvm_ldg_d: |
9487 | 10 | case NVPTX::BI__nvvm_ldg_d2: |
9488 | 10 | return MakeLdg(Intrinsic::nvvm_ldg_global_f); |
9489 | 10 | |
9490 | 9 | case NVPTX::BI__nvvm_atom_cta_add_gen_i: |
9491 | 9 | case NVPTX::BI__nvvm_atom_cta_add_gen_l: |
9492 | 9 | case NVPTX::BI__nvvm_atom_cta_add_gen_ll: |
9493 | 9 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_add_gen_i_cta); |
9494 | 9 | case NVPTX::BI__nvvm_atom_sys_add_gen_i: |
9495 | 9 | case NVPTX::BI__nvvm_atom_sys_add_gen_l: |
9496 | 9 | case NVPTX::BI__nvvm_atom_sys_add_gen_ll: |
9497 | 9 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_add_gen_i_sys); |
9498 | 6 | case NVPTX::BI__nvvm_atom_cta_add_gen_f: |
9499 | 6 | case NVPTX::BI__nvvm_atom_cta_add_gen_d: |
9500 | 6 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_add_gen_f_cta); |
9501 | 6 | case NVPTX::BI__nvvm_atom_sys_add_gen_f: |
9502 | 6 | case NVPTX::BI__nvvm_atom_sys_add_gen_d: |
9503 | 6 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_add_gen_f_sys); |
9504 | 9 | case NVPTX::BI__nvvm_atom_cta_xchg_gen_i: |
9505 | 9 | case NVPTX::BI__nvvm_atom_cta_xchg_gen_l: |
9506 | 9 | case NVPTX::BI__nvvm_atom_cta_xchg_gen_ll: |
9507 | 9 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_exch_gen_i_cta); |
9508 | 9 | case NVPTX::BI__nvvm_atom_sys_xchg_gen_i: |
9509 | 9 | case NVPTX::BI__nvvm_atom_sys_xchg_gen_l: |
9510 | 9 | case NVPTX::BI__nvvm_atom_sys_xchg_gen_ll: |
9511 | 9 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_exch_gen_i_sys); |
9512 | 18 | case NVPTX::BI__nvvm_atom_cta_max_gen_i: |
9513 | 18 | case NVPTX::BI__nvvm_atom_cta_max_gen_ui: |
9514 | 18 | case NVPTX::BI__nvvm_atom_cta_max_gen_l: |
9515 | 18 | case NVPTX::BI__nvvm_atom_cta_max_gen_ul: |
9516 | 18 | case NVPTX::BI__nvvm_atom_cta_max_gen_ll: |
9517 | 18 | case NVPTX::BI__nvvm_atom_cta_max_gen_ull: |
9518 | 18 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_max_gen_i_cta); |
9519 | 18 | case NVPTX::BI__nvvm_atom_sys_max_gen_i: |
9520 | 18 | case NVPTX::BI__nvvm_atom_sys_max_gen_ui: |
9521 | 18 | case NVPTX::BI__nvvm_atom_sys_max_gen_l: |
9522 | 18 | case NVPTX::BI__nvvm_atom_sys_max_gen_ul: |
9523 | 18 | case NVPTX::BI__nvvm_atom_sys_max_gen_ll: |
9524 | 18 | case NVPTX::BI__nvvm_atom_sys_max_gen_ull: |
9525 | 18 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_max_gen_i_sys); |
9526 | 18 | case NVPTX::BI__nvvm_atom_cta_min_gen_i: |
9527 | 18 | case NVPTX::BI__nvvm_atom_cta_min_gen_ui: |
9528 | 18 | case NVPTX::BI__nvvm_atom_cta_min_gen_l: |
9529 | 18 | case NVPTX::BI__nvvm_atom_cta_min_gen_ul: |
9530 | 18 | case NVPTX::BI__nvvm_atom_cta_min_gen_ll: |
9531 | 18 | case NVPTX::BI__nvvm_atom_cta_min_gen_ull: |
9532 | 18 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_min_gen_i_cta); |
9533 | 18 | case NVPTX::BI__nvvm_atom_sys_min_gen_i: |
9534 | 18 | case NVPTX::BI__nvvm_atom_sys_min_gen_ui: |
9535 | 18 | case NVPTX::BI__nvvm_atom_sys_min_gen_l: |
9536 | 18 | case NVPTX::BI__nvvm_atom_sys_min_gen_ul: |
9537 | 18 | case NVPTX::BI__nvvm_atom_sys_min_gen_ll: |
9538 | 18 | case NVPTX::BI__nvvm_atom_sys_min_gen_ull: |
9539 | 18 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_min_gen_i_sys); |
9540 | 3 | case NVPTX::BI__nvvm_atom_cta_inc_gen_ui: |
9541 | 3 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_inc_gen_i_cta); |
9542 | 3 | case NVPTX::BI__nvvm_atom_cta_dec_gen_ui: |
9543 | 3 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_dec_gen_i_cta); |
9544 | 3 | case NVPTX::BI__nvvm_atom_sys_inc_gen_ui: |
9545 | 3 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_inc_gen_i_sys); |
9546 | 3 | case NVPTX::BI__nvvm_atom_sys_dec_gen_ui: |
9547 | 3 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_dec_gen_i_sys); |
9548 | 9 | case NVPTX::BI__nvvm_atom_cta_and_gen_i: |
9549 | 9 | case NVPTX::BI__nvvm_atom_cta_and_gen_l: |
9550 | 9 | case NVPTX::BI__nvvm_atom_cta_and_gen_ll: |
9551 | 9 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_and_gen_i_cta); |
9552 | 9 | case NVPTX::BI__nvvm_atom_sys_and_gen_i: |
9553 | 9 | case NVPTX::BI__nvvm_atom_sys_and_gen_l: |
9554 | 9 | case NVPTX::BI__nvvm_atom_sys_and_gen_ll: |
9555 | 9 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_and_gen_i_sys); |
9556 | 9 | case NVPTX::BI__nvvm_atom_cta_or_gen_i: |
9557 | 9 | case NVPTX::BI__nvvm_atom_cta_or_gen_l: |
9558 | 9 | case NVPTX::BI__nvvm_atom_cta_or_gen_ll: |
9559 | 9 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_or_gen_i_cta); |
9560 | 9 | case NVPTX::BI__nvvm_atom_sys_or_gen_i: |
9561 | 9 | case NVPTX::BI__nvvm_atom_sys_or_gen_l: |
9562 | 9 | case NVPTX::BI__nvvm_atom_sys_or_gen_ll: |
9563 | 9 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_or_gen_i_sys); |
9564 | 9 | case NVPTX::BI__nvvm_atom_cta_xor_gen_i: |
9565 | 9 | case NVPTX::BI__nvvm_atom_cta_xor_gen_l: |
9566 | 9 | case NVPTX::BI__nvvm_atom_cta_xor_gen_ll: |
9567 | 9 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_xor_gen_i_cta); |
9568 | 9 | case NVPTX::BI__nvvm_atom_sys_xor_gen_i: |
9569 | 9 | case NVPTX::BI__nvvm_atom_sys_xor_gen_l: |
9570 | 9 | case NVPTX::BI__nvvm_atom_sys_xor_gen_ll: |
9571 | 9 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_xor_gen_i_sys); |
9572 | 9 | case NVPTX::BI__nvvm_atom_cta_cas_gen_i: |
9573 | 9 | case NVPTX::BI__nvvm_atom_cta_cas_gen_l: |
9574 | 9 | case NVPTX::BI__nvvm_atom_cta_cas_gen_ll: { |
9575 | 9 | Value *Ptr = EmitScalarExpr(E->getArg(0)); |
9576 | 9 | return Builder.CreateCall( |
9577 | 9 | CGM.getIntrinsic( |
9578 | 9 | Intrinsic::nvvm_atomic_cas_gen_i_cta, |
9579 | 9 | {Ptr->getType()->getPointerElementType(), Ptr->getType()}), |
9580 | 9 | {Ptr, EmitScalarExpr(E->getArg(1)), EmitScalarExpr(E->getArg(2))}); |
9581 | 9 | } |
9582 | 9 | case NVPTX::BI__nvvm_atom_sys_cas_gen_i: |
9583 | 9 | case NVPTX::BI__nvvm_atom_sys_cas_gen_l: |
9584 | 9 | case NVPTX::BI__nvvm_atom_sys_cas_gen_ll: { |
9585 | 9 | Value *Ptr = EmitScalarExpr(E->getArg(0)); |
9586 | 9 | return Builder.CreateCall( |
9587 | 9 | CGM.getIntrinsic( |
9588 | 9 | Intrinsic::nvvm_atomic_cas_gen_i_sys, |
9589 | 9 | {Ptr->getType()->getPointerElementType(), Ptr->getType()}), |
9590 | 9 | {Ptr, EmitScalarExpr(E->getArg(1)), EmitScalarExpr(E->getArg(2))}); |
9591 | 9 | } |
9592 | 4 | case NVPTX::BI__nvvm_match_all_sync_i32p: |
9593 | 4 | case NVPTX::BI__nvvm_match_all_sync_i64p: { |
9594 | 4 | Value *Mask = EmitScalarExpr(E->getArg(0)); |
9595 | 4 | Value *Val = EmitScalarExpr(E->getArg(1)); |
9596 | 4 | Address PredOutPtr = EmitPointerWithAlignment(E->getArg(2)); |
9597 | 4 | Value *ResultPair = Builder.CreateCall( |
9598 | 4 | CGM.getIntrinsic(BuiltinID == NVPTX::BI__nvvm_match_all_sync_i32p |
9599 | 2 | ? Intrinsic::nvvm_match_all_sync_i32p |
9600 | 2 | : Intrinsic::nvvm_match_all_sync_i64p), |
9601 | 4 | {Mask, Val}); |
9602 | 4 | Value *Pred = Builder.CreateZExt(Builder.CreateExtractValue(ResultPair, 1), |
9603 | 4 | PredOutPtr.getElementType()); |
9604 | 4 | Builder.CreateStore(Pred, PredOutPtr); |
9605 | 4 | return Builder.CreateExtractValue(ResultPair, 0); |
9606 | 4 | } |
9607 | 0 | default: |
9608 | 0 | return nullptr; |
9609 | 0 | } |
9610 | 0 | } |
9611 | | |
9612 | | Value *CodeGenFunction::EmitWebAssemblyBuiltinExpr(unsigned BuiltinID, |
9613 | 8 | const CallExpr *E) { |
9614 | 8 | switch (BuiltinID) { |
9615 | 2 | case WebAssembly::BI__builtin_wasm_current_memory: { |
9616 | 2 | llvm::Type *ResultType = ConvertType(E->getType()); |
9617 | 2 | Value *Callee = CGM.getIntrinsic(Intrinsic::wasm_current_memory, ResultType); |
9618 | 2 | return Builder.CreateCall(Callee); |
9619 | 8 | } |
9620 | 2 | case WebAssembly::BI__builtin_wasm_grow_memory: { |
9621 | 2 | Value *X = EmitScalarExpr(E->getArg(0)); |
9622 | 2 | Value *Callee = CGM.getIntrinsic(Intrinsic::wasm_grow_memory, X->getType()); |
9623 | 2 | return Builder.CreateCall(Callee, X); |
9624 | 8 | } |
9625 | 2 | case WebAssembly::BI__builtin_wasm_throw: { |
9626 | 2 | Value *Tag = EmitScalarExpr(E->getArg(0)); |
9627 | 2 | Value *Obj = EmitScalarExpr(E->getArg(1)); |
9628 | 2 | Value *Callee = CGM.getIntrinsic(Intrinsic::wasm_throw); |
9629 | 2 | return Builder.CreateCall(Callee, {Tag, Obj}); |
9630 | 8 | } |
9631 | 2 | case WebAssembly::BI__builtin_wasm_rethrow: { |
9632 | 2 | Value *Callee = CGM.getIntrinsic(Intrinsic::wasm_rethrow); |
9633 | 2 | return Builder.CreateCall(Callee); |
9634 | 8 | } |
9635 | 8 | |
9636 | 0 | default: |
9637 | 0 | return nullptr; |
9638 | 0 | } |
9639 | 0 | } |