File: | tools/clang/lib/CodeGen/CGBuiltin.cpp |
Warning: | line 1435, column 5 Value stored to 'Off' is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===---- CGBuiltin.cpp - Emit LLVM Code for builtins ---------------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This contains code to emit Builtin calls as LLVM code. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "CGCXXABI.h" |
14 | #include "CGObjCRuntime.h" |
15 | #include "CGOpenCLRuntime.h" |
16 | #include "CGRecordLayout.h" |
17 | #include "CodeGenFunction.h" |
18 | #include "CodeGenModule.h" |
19 | #include "ConstantEmitter.h" |
20 | #include "PatternInit.h" |
21 | #include "TargetInfo.h" |
22 | #include "clang/AST/ASTContext.h" |
23 | #include "clang/AST/Decl.h" |
24 | #include "clang/AST/OSLog.h" |
25 | #include "clang/Basic/TargetBuiltins.h" |
26 | #include "clang/Basic/TargetInfo.h" |
27 | #include "clang/CodeGen/CGFunctionInfo.h" |
28 | #include "llvm/ADT/SmallPtrSet.h" |
29 | #include "llvm/ADT/StringExtras.h" |
30 | #include "llvm/IR/DataLayout.h" |
31 | #include "llvm/IR/InlineAsm.h" |
32 | #include "llvm/IR/Intrinsics.h" |
33 | #include "llvm/IR/MDBuilder.h" |
34 | #include "llvm/Support/ConvertUTF.h" |
35 | #include "llvm/Support/ScopedPrinter.h" |
36 | #include "llvm/Support/TargetParser.h" |
37 | #include <sstream> |
38 | |
39 | using namespace clang; |
40 | using namespace CodeGen; |
41 | using namespace llvm; |
42 | |
43 | static |
44 | int64_t clamp(int64_t Value, int64_t Low, int64_t High) { |
45 | return std::min(High, std::max(Low, Value)); |
46 | } |
47 | |
48 | static void initializeAlloca(CodeGenFunction &CGF, AllocaInst *AI, Value *Size, unsigned AlignmentInBytes) { |
49 | ConstantInt *Byte; |
50 | switch (CGF.getLangOpts().getTrivialAutoVarInit()) { |
51 | case LangOptions::TrivialAutoVarInitKind::Uninitialized: |
52 | // Nothing to initialize. |
53 | return; |
54 | case LangOptions::TrivialAutoVarInitKind::Zero: |
55 | Byte = CGF.Builder.getInt8(0x00); |
56 | break; |
57 | case LangOptions::TrivialAutoVarInitKind::Pattern: { |
58 | llvm::Type *Int8 = llvm::IntegerType::getInt8Ty(CGF.CGM.getLLVMContext()); |
59 | Byte = llvm::dyn_cast<llvm::ConstantInt>( |
60 | initializationPatternFor(CGF.CGM, Int8)); |
61 | break; |
62 | } |
63 | } |
64 | CGF.Builder.CreateMemSet(AI, Byte, Size, AlignmentInBytes); |
65 | } |
66 | |
67 | /// getBuiltinLibFunction - Given a builtin id for a function like |
68 | /// "__builtin_fabsf", return a Function* for "fabsf". |
69 | llvm::Constant *CodeGenModule::getBuiltinLibFunction(const FunctionDecl *FD, |
70 | unsigned BuiltinID) { |
71 | assert(Context.BuiltinInfo.isLibFunction(BuiltinID))((Context.BuiltinInfo.isLibFunction(BuiltinID)) ? static_cast <void> (0) : __assert_fail ("Context.BuiltinInfo.isLibFunction(BuiltinID)" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 71, __PRETTY_FUNCTION__)); |
72 | |
73 | // Get the name, skip over the __builtin_ prefix (if necessary). |
74 | StringRef Name; |
75 | GlobalDecl D(FD); |
76 | |
77 | // If the builtin has been declared explicitly with an assembler label, |
78 | // use the mangled name. This differs from the plain label on platforms |
79 | // that prefix labels. |
80 | if (FD->hasAttr<AsmLabelAttr>()) |
81 | Name = getMangledName(D); |
82 | else |
83 | Name = Context.BuiltinInfo.getName(BuiltinID) + 10; |
84 | |
85 | llvm::FunctionType *Ty = |
86 | cast<llvm::FunctionType>(getTypes().ConvertType(FD->getType())); |
87 | |
88 | return GetOrCreateLLVMFunction(Name, Ty, D, /*ForVTable=*/false); |
89 | } |
90 | |
91 | /// Emit the conversions required to turn the given value into an |
92 | /// integer of the given size. |
93 | static Value *EmitToInt(CodeGenFunction &CGF, llvm::Value *V, |
94 | QualType T, llvm::IntegerType *IntType) { |
95 | V = CGF.EmitToMemory(V, T); |
96 | |
97 | if (V->getType()->isPointerTy()) |
98 | return CGF.Builder.CreatePtrToInt(V, IntType); |
99 | |
100 | assert(V->getType() == IntType)((V->getType() == IntType) ? static_cast<void> (0) : __assert_fail ("V->getType() == IntType", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 100, __PRETTY_FUNCTION__)); |
101 | return V; |
102 | } |
103 | |
104 | static Value *EmitFromInt(CodeGenFunction &CGF, llvm::Value *V, |
105 | QualType T, llvm::Type *ResultType) { |
106 | V = CGF.EmitFromMemory(V, T); |
107 | |
108 | if (ResultType->isPointerTy()) |
109 | return CGF.Builder.CreateIntToPtr(V, ResultType); |
110 | |
111 | assert(V->getType() == ResultType)((V->getType() == ResultType) ? static_cast<void> (0 ) : __assert_fail ("V->getType() == ResultType", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 111, __PRETTY_FUNCTION__)); |
112 | return V; |
113 | } |
114 | |
115 | /// Utility to insert an atomic instruction based on Intrinsic::ID |
116 | /// and the expression node. |
117 | static Value *MakeBinaryAtomicValue( |
118 | CodeGenFunction &CGF, llvm::AtomicRMWInst::BinOp Kind, const CallExpr *E, |
119 | AtomicOrdering Ordering = AtomicOrdering::SequentiallyConsistent) { |
120 | QualType T = E->getType(); |
121 | assert(E->getArg(0)->getType()->isPointerType())((E->getArg(0)->getType()->isPointerType()) ? static_cast <void> (0) : __assert_fail ("E->getArg(0)->getType()->isPointerType()" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 121, __PRETTY_FUNCTION__)); |
122 | assert(CGF.getContext().hasSameUnqualifiedType(T,((CGF.getContext().hasSameUnqualifiedType(T, E->getArg(0)-> getType()->getPointeeType())) ? static_cast<void> (0 ) : __assert_fail ("CGF.getContext().hasSameUnqualifiedType(T, E->getArg(0)->getType()->getPointeeType())" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 123, __PRETTY_FUNCTION__)) |
123 | E->getArg(0)->getType()->getPointeeType()))((CGF.getContext().hasSameUnqualifiedType(T, E->getArg(0)-> getType()->getPointeeType())) ? static_cast<void> (0 ) : __assert_fail ("CGF.getContext().hasSameUnqualifiedType(T, E->getArg(0)->getType()->getPointeeType())" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 123, __PRETTY_FUNCTION__)); |
124 | assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType()))((CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)-> getType())) ? static_cast<void> (0) : __assert_fail ("CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType())" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 124, __PRETTY_FUNCTION__)); |
125 | |
126 | llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); |
127 | unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); |
128 | |
129 | llvm::IntegerType *IntType = |
130 | llvm::IntegerType::get(CGF.getLLVMContext(), |
131 | CGF.getContext().getTypeSize(T)); |
132 | llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); |
133 | |
134 | llvm::Value *Args[2]; |
135 | Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); |
136 | Args[1] = CGF.EmitScalarExpr(E->getArg(1)); |
137 | llvm::Type *ValueType = Args[1]->getType(); |
138 | Args[1] = EmitToInt(CGF, Args[1], T, IntType); |
139 | |
140 | llvm::Value *Result = CGF.Builder.CreateAtomicRMW( |
141 | Kind, Args[0], Args[1], Ordering); |
142 | return EmitFromInt(CGF, Result, T, ValueType); |
143 | } |
144 | |
145 | static Value *EmitNontemporalStore(CodeGenFunction &CGF, const CallExpr *E) { |
146 | Value *Val = CGF.EmitScalarExpr(E->getArg(0)); |
147 | Value *Address = CGF.EmitScalarExpr(E->getArg(1)); |
148 | |
149 | // Convert the type of the pointer to a pointer to the stored type. |
150 | Val = CGF.EmitToMemory(Val, E->getArg(0)->getType()); |
151 | Value *BC = CGF.Builder.CreateBitCast( |
152 | Address, llvm::PointerType::getUnqual(Val->getType()), "cast"); |
153 | LValue LV = CGF.MakeNaturalAlignAddrLValue(BC, E->getArg(0)->getType()); |
154 | LV.setNontemporal(true); |
155 | CGF.EmitStoreOfScalar(Val, LV, false); |
156 | return nullptr; |
157 | } |
158 | |
159 | static Value *EmitNontemporalLoad(CodeGenFunction &CGF, const CallExpr *E) { |
160 | Value *Address = CGF.EmitScalarExpr(E->getArg(0)); |
161 | |
162 | LValue LV = CGF.MakeNaturalAlignAddrLValue(Address, E->getType()); |
163 | LV.setNontemporal(true); |
164 | return CGF.EmitLoadOfScalar(LV, E->getExprLoc()); |
165 | } |
166 | |
167 | static RValue EmitBinaryAtomic(CodeGenFunction &CGF, |
168 | llvm::AtomicRMWInst::BinOp Kind, |
169 | const CallExpr *E) { |
170 | return RValue::get(MakeBinaryAtomicValue(CGF, Kind, E)); |
171 | } |
172 | |
173 | /// Utility to insert an atomic instruction based Intrinsic::ID and |
174 | /// the expression node, where the return value is the result of the |
175 | /// operation. |
176 | static RValue EmitBinaryAtomicPost(CodeGenFunction &CGF, |
177 | llvm::AtomicRMWInst::BinOp Kind, |
178 | const CallExpr *E, |
179 | Instruction::BinaryOps Op, |
180 | bool Invert = false) { |
181 | QualType T = E->getType(); |
182 | assert(E->getArg(0)->getType()->isPointerType())((E->getArg(0)->getType()->isPointerType()) ? static_cast <void> (0) : __assert_fail ("E->getArg(0)->getType()->isPointerType()" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 182, __PRETTY_FUNCTION__)); |
183 | assert(CGF.getContext().hasSameUnqualifiedType(T,((CGF.getContext().hasSameUnqualifiedType(T, E->getArg(0)-> getType()->getPointeeType())) ? static_cast<void> (0 ) : __assert_fail ("CGF.getContext().hasSameUnqualifiedType(T, E->getArg(0)->getType()->getPointeeType())" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 184, __PRETTY_FUNCTION__)) |
184 | E->getArg(0)->getType()->getPointeeType()))((CGF.getContext().hasSameUnqualifiedType(T, E->getArg(0)-> getType()->getPointeeType())) ? static_cast<void> (0 ) : __assert_fail ("CGF.getContext().hasSameUnqualifiedType(T, E->getArg(0)->getType()->getPointeeType())" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 184, __PRETTY_FUNCTION__)); |
185 | assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType()))((CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)-> getType())) ? static_cast<void> (0) : __assert_fail ("CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType())" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 185, __PRETTY_FUNCTION__)); |
186 | |
187 | llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); |
188 | unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); |
189 | |
190 | llvm::IntegerType *IntType = |
191 | llvm::IntegerType::get(CGF.getLLVMContext(), |
192 | CGF.getContext().getTypeSize(T)); |
193 | llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); |
194 | |
195 | llvm::Value *Args[2]; |
196 | Args[1] = CGF.EmitScalarExpr(E->getArg(1)); |
197 | llvm::Type *ValueType = Args[1]->getType(); |
198 | Args[1] = EmitToInt(CGF, Args[1], T, IntType); |
199 | Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); |
200 | |
201 | llvm::Value *Result = CGF.Builder.CreateAtomicRMW( |
202 | Kind, Args[0], Args[1], llvm::AtomicOrdering::SequentiallyConsistent); |
203 | Result = CGF.Builder.CreateBinOp(Op, Result, Args[1]); |
204 | if (Invert) |
205 | Result = CGF.Builder.CreateBinOp(llvm::Instruction::Xor, Result, |
206 | llvm::ConstantInt::get(IntType, -1)); |
207 | Result = EmitFromInt(CGF, Result, T, ValueType); |
208 | return RValue::get(Result); |
209 | } |
210 | |
211 | /// Utility to insert an atomic cmpxchg instruction. |
212 | /// |
213 | /// @param CGF The current codegen function. |
214 | /// @param E Builtin call expression to convert to cmpxchg. |
215 | /// arg0 - address to operate on |
216 | /// arg1 - value to compare with |
217 | /// arg2 - new value |
218 | /// @param ReturnBool Specifies whether to return success flag of |
219 | /// cmpxchg result or the old value. |
220 | /// |
221 | /// @returns result of cmpxchg, according to ReturnBool |
222 | /// |
223 | /// Note: In order to lower Microsoft's _InterlockedCompareExchange* intrinsics |
224 | /// invoke the function EmitAtomicCmpXchgForMSIntrin. |
225 | static Value *MakeAtomicCmpXchgValue(CodeGenFunction &CGF, const CallExpr *E, |
226 | bool ReturnBool) { |
227 | QualType T = ReturnBool ? E->getArg(1)->getType() : E->getType(); |
228 | llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); |
229 | unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); |
230 | |
231 | llvm::IntegerType *IntType = llvm::IntegerType::get( |
232 | CGF.getLLVMContext(), CGF.getContext().getTypeSize(T)); |
233 | llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); |
234 | |
235 | Value *Args[3]; |
236 | Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); |
237 | Args[1] = CGF.EmitScalarExpr(E->getArg(1)); |
238 | llvm::Type *ValueType = Args[1]->getType(); |
239 | Args[1] = EmitToInt(CGF, Args[1], T, IntType); |
240 | Args[2] = EmitToInt(CGF, CGF.EmitScalarExpr(E->getArg(2)), T, IntType); |
241 | |
242 | Value *Pair = CGF.Builder.CreateAtomicCmpXchg( |
243 | Args[0], Args[1], Args[2], llvm::AtomicOrdering::SequentiallyConsistent, |
244 | llvm::AtomicOrdering::SequentiallyConsistent); |
245 | if (ReturnBool) |
246 | // Extract boolean success flag and zext it to int. |
247 | return CGF.Builder.CreateZExt(CGF.Builder.CreateExtractValue(Pair, 1), |
248 | CGF.ConvertType(E->getType())); |
249 | else |
250 | // Extract old value and emit it using the same type as compare value. |
251 | return EmitFromInt(CGF, CGF.Builder.CreateExtractValue(Pair, 0), T, |
252 | ValueType); |
253 | } |
254 | |
255 | /// This function should be invoked to emit atomic cmpxchg for Microsoft's |
256 | /// _InterlockedCompareExchange* intrinsics which have the following signature: |
257 | /// T _InterlockedCompareExchange(T volatile *Destination, |
258 | /// T Exchange, |
259 | /// T Comparand); |
260 | /// |
261 | /// Whereas the llvm 'cmpxchg' instruction has the following syntax: |
262 | /// cmpxchg *Destination, Comparand, Exchange. |
263 | /// So we need to swap Comparand and Exchange when invoking |
264 | /// CreateAtomicCmpXchg. That is the reason we could not use the above utility |
265 | /// function MakeAtomicCmpXchgValue since it expects the arguments to be |
266 | /// already swapped. |
267 | |
268 | static |
269 | Value *EmitAtomicCmpXchgForMSIntrin(CodeGenFunction &CGF, const CallExpr *E, |
270 | AtomicOrdering SuccessOrdering = AtomicOrdering::SequentiallyConsistent) { |
271 | assert(E->getArg(0)->getType()->isPointerType())((E->getArg(0)->getType()->isPointerType()) ? static_cast <void> (0) : __assert_fail ("E->getArg(0)->getType()->isPointerType()" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 271, __PRETTY_FUNCTION__)); |
272 | assert(CGF.getContext().hasSameUnqualifiedType(((CGF.getContext().hasSameUnqualifiedType( E->getType(), E ->getArg(0)->getType()->getPointeeType())) ? static_cast <void> (0) : __assert_fail ("CGF.getContext().hasSameUnqualifiedType( E->getType(), E->getArg(0)->getType()->getPointeeType())" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 273, __PRETTY_FUNCTION__)) |
273 | E->getType(), E->getArg(0)->getType()->getPointeeType()))((CGF.getContext().hasSameUnqualifiedType( E->getType(), E ->getArg(0)->getType()->getPointeeType())) ? static_cast <void> (0) : __assert_fail ("CGF.getContext().hasSameUnqualifiedType( E->getType(), E->getArg(0)->getType()->getPointeeType())" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 273, __PRETTY_FUNCTION__)); |
274 | assert(CGF.getContext().hasSameUnqualifiedType(E->getType(),((CGF.getContext().hasSameUnqualifiedType(E->getType(), E-> getArg(1)->getType())) ? static_cast<void> (0) : __assert_fail ("CGF.getContext().hasSameUnqualifiedType(E->getType(), E->getArg(1)->getType())" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 275, __PRETTY_FUNCTION__)) |
275 | E->getArg(1)->getType()))((CGF.getContext().hasSameUnqualifiedType(E->getType(), E-> getArg(1)->getType())) ? static_cast<void> (0) : __assert_fail ("CGF.getContext().hasSameUnqualifiedType(E->getType(), E->getArg(1)->getType())" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 275, __PRETTY_FUNCTION__)); |
276 | assert(CGF.getContext().hasSameUnqualifiedType(E->getType(),((CGF.getContext().hasSameUnqualifiedType(E->getType(), E-> getArg(2)->getType())) ? static_cast<void> (0) : __assert_fail ("CGF.getContext().hasSameUnqualifiedType(E->getType(), E->getArg(2)->getType())" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 277, __PRETTY_FUNCTION__)) |
277 | E->getArg(2)->getType()))((CGF.getContext().hasSameUnqualifiedType(E->getType(), E-> getArg(2)->getType())) ? static_cast<void> (0) : __assert_fail ("CGF.getContext().hasSameUnqualifiedType(E->getType(), E->getArg(2)->getType())" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 277, __PRETTY_FUNCTION__)); |
278 | |
279 | auto *Destination = CGF.EmitScalarExpr(E->getArg(0)); |
280 | auto *Comparand = CGF.EmitScalarExpr(E->getArg(2)); |
281 | auto *Exchange = CGF.EmitScalarExpr(E->getArg(1)); |
282 | |
283 | // For Release ordering, the failure ordering should be Monotonic. |
284 | auto FailureOrdering = SuccessOrdering == AtomicOrdering::Release ? |
285 | AtomicOrdering::Monotonic : |
286 | SuccessOrdering; |
287 | |
288 | auto *Result = CGF.Builder.CreateAtomicCmpXchg( |
289 | Destination, Comparand, Exchange, |
290 | SuccessOrdering, FailureOrdering); |
291 | Result->setVolatile(true); |
292 | return CGF.Builder.CreateExtractValue(Result, 0); |
293 | } |
294 | |
295 | static Value *EmitAtomicIncrementValue(CodeGenFunction &CGF, const CallExpr *E, |
296 | AtomicOrdering Ordering = AtomicOrdering::SequentiallyConsistent) { |
297 | assert(E->getArg(0)->getType()->isPointerType())((E->getArg(0)->getType()->isPointerType()) ? static_cast <void> (0) : __assert_fail ("E->getArg(0)->getType()->isPointerType()" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 297, __PRETTY_FUNCTION__)); |
298 | |
299 | auto *IntTy = CGF.ConvertType(E->getType()); |
300 | auto *Result = CGF.Builder.CreateAtomicRMW( |
301 | AtomicRMWInst::Add, |
302 | CGF.EmitScalarExpr(E->getArg(0)), |
303 | ConstantInt::get(IntTy, 1), |
304 | Ordering); |
305 | return CGF.Builder.CreateAdd(Result, ConstantInt::get(IntTy, 1)); |
306 | } |
307 | |
308 | static Value *EmitAtomicDecrementValue(CodeGenFunction &CGF, const CallExpr *E, |
309 | AtomicOrdering Ordering = AtomicOrdering::SequentiallyConsistent) { |
310 | assert(E->getArg(0)->getType()->isPointerType())((E->getArg(0)->getType()->isPointerType()) ? static_cast <void> (0) : __assert_fail ("E->getArg(0)->getType()->isPointerType()" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 310, __PRETTY_FUNCTION__)); |
311 | |
312 | auto *IntTy = CGF.ConvertType(E->getType()); |
313 | auto *Result = CGF.Builder.CreateAtomicRMW( |
314 | AtomicRMWInst::Sub, |
315 | CGF.EmitScalarExpr(E->getArg(0)), |
316 | ConstantInt::get(IntTy, 1), |
317 | Ordering); |
318 | return CGF.Builder.CreateSub(Result, ConstantInt::get(IntTy, 1)); |
319 | } |
320 | |
321 | // Build a plain volatile load. |
322 | static Value *EmitISOVolatileLoad(CodeGenFunction &CGF, const CallExpr *E) { |
323 | Value *Ptr = CGF.EmitScalarExpr(E->getArg(0)); |
324 | QualType ElTy = E->getArg(0)->getType()->getPointeeType(); |
325 | CharUnits LoadSize = CGF.getContext().getTypeSizeInChars(ElTy); |
326 | llvm::Type *ITy = |
327 | llvm::IntegerType::get(CGF.getLLVMContext(), LoadSize.getQuantity() * 8); |
328 | Ptr = CGF.Builder.CreateBitCast(Ptr, ITy->getPointerTo()); |
329 | llvm::LoadInst *Load = CGF.Builder.CreateAlignedLoad(Ptr, LoadSize); |
330 | Load->setVolatile(true); |
331 | return Load; |
332 | } |
333 | |
334 | // Build a plain volatile store. |
335 | static Value *EmitISOVolatileStore(CodeGenFunction &CGF, const CallExpr *E) { |
336 | Value *Ptr = CGF.EmitScalarExpr(E->getArg(0)); |
337 | Value *Value = CGF.EmitScalarExpr(E->getArg(1)); |
338 | QualType ElTy = E->getArg(0)->getType()->getPointeeType(); |
339 | CharUnits StoreSize = CGF.getContext().getTypeSizeInChars(ElTy); |
340 | llvm::Type *ITy = |
341 | llvm::IntegerType::get(CGF.getLLVMContext(), StoreSize.getQuantity() * 8); |
342 | Ptr = CGF.Builder.CreateBitCast(Ptr, ITy->getPointerTo()); |
343 | llvm::StoreInst *Store = |
344 | CGF.Builder.CreateAlignedStore(Value, Ptr, StoreSize); |
345 | Store->setVolatile(true); |
346 | return Store; |
347 | } |
348 | |
349 | // Emit a simple mangled intrinsic that has 1 argument and a return type |
350 | // matching the argument type. |
351 | static Value *emitUnaryBuiltin(CodeGenFunction &CGF, |
352 | const CallExpr *E, |
353 | unsigned IntrinsicID) { |
354 | llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); |
355 | |
356 | Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); |
357 | return CGF.Builder.CreateCall(F, Src0); |
358 | } |
359 | |
360 | // Emit an intrinsic that has 2 operands of the same type as its result. |
361 | static Value *emitBinaryBuiltin(CodeGenFunction &CGF, |
362 | const CallExpr *E, |
363 | unsigned IntrinsicID) { |
364 | llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); |
365 | llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); |
366 | |
367 | Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); |
368 | return CGF.Builder.CreateCall(F, { Src0, Src1 }); |
369 | } |
370 | |
371 | // Emit an intrinsic that has 3 operands of the same type as its result. |
372 | static Value *emitTernaryBuiltin(CodeGenFunction &CGF, |
373 | const CallExpr *E, |
374 | unsigned IntrinsicID) { |
375 | llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); |
376 | llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); |
377 | llvm::Value *Src2 = CGF.EmitScalarExpr(E->getArg(2)); |
378 | |
379 | Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); |
380 | return CGF.Builder.CreateCall(F, { Src0, Src1, Src2 }); |
381 | } |
382 | |
383 | // Emit an intrinsic that has 1 float or double operand, and 1 integer. |
384 | static Value *emitFPIntBuiltin(CodeGenFunction &CGF, |
385 | const CallExpr *E, |
386 | unsigned IntrinsicID) { |
387 | llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); |
388 | llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); |
389 | |
390 | Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); |
391 | return CGF.Builder.CreateCall(F, {Src0, Src1}); |
392 | } |
393 | |
394 | // Emit an intrinsic that has overloaded integer result and fp operand. |
395 | static Value *emitFPToIntRoundBuiltin(CodeGenFunction &CGF, |
396 | const CallExpr *E, |
397 | unsigned IntrinsicID) { |
398 | llvm::Type *ResultType = CGF.ConvertType(E->getType()); |
399 | llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); |
400 | |
401 | Function *F = CGF.CGM.getIntrinsic(IntrinsicID, |
402 | {ResultType, Src0->getType()}); |
403 | return CGF.Builder.CreateCall(F, Src0); |
404 | } |
405 | |
406 | /// EmitFAbs - Emit a call to @llvm.fabs(). |
407 | static Value *EmitFAbs(CodeGenFunction &CGF, Value *V) { |
408 | Function *F = CGF.CGM.getIntrinsic(Intrinsic::fabs, V->getType()); |
409 | llvm::CallInst *Call = CGF.Builder.CreateCall(F, V); |
410 | Call->setDoesNotAccessMemory(); |
411 | return Call; |
412 | } |
413 | |
414 | /// Emit the computation of the sign bit for a floating point value. Returns |
415 | /// the i1 sign bit value. |
416 | static Value *EmitSignBit(CodeGenFunction &CGF, Value *V) { |
417 | LLVMContext &C = CGF.CGM.getLLVMContext(); |
418 | |
419 | llvm::Type *Ty = V->getType(); |
420 | int Width = Ty->getPrimitiveSizeInBits(); |
421 | llvm::Type *IntTy = llvm::IntegerType::get(C, Width); |
422 | V = CGF.Builder.CreateBitCast(V, IntTy); |
423 | if (Ty->isPPC_FP128Ty()) { |
424 | // We want the sign bit of the higher-order double. The bitcast we just |
425 | // did works as if the double-double was stored to memory and then |
426 | // read as an i128. The "store" will put the higher-order double in the |
427 | // lower address in both little- and big-Endian modes, but the "load" |
428 | // will treat those bits as a different part of the i128: the low bits in |
429 | // little-Endian, the high bits in big-Endian. Therefore, on big-Endian |
430 | // we need to shift the high bits down to the low before truncating. |
431 | Width >>= 1; |
432 | if (CGF.getTarget().isBigEndian()) { |
433 | Value *ShiftCst = llvm::ConstantInt::get(IntTy, Width); |
434 | V = CGF.Builder.CreateLShr(V, ShiftCst); |
435 | } |
436 | // We are truncating value in order to extract the higher-order |
437 | // double, which we will be using to extract the sign from. |
438 | IntTy = llvm::IntegerType::get(C, Width); |
439 | V = CGF.Builder.CreateTrunc(V, IntTy); |
440 | } |
441 | Value *Zero = llvm::Constant::getNullValue(IntTy); |
442 | return CGF.Builder.CreateICmpSLT(V, Zero); |
443 | } |
444 | |
445 | static RValue emitLibraryCall(CodeGenFunction &CGF, const FunctionDecl *FD, |
446 | const CallExpr *E, llvm::Constant *calleeValue) { |
447 | CGCallee callee = CGCallee::forDirect(calleeValue, GlobalDecl(FD)); |
448 | return CGF.EmitCall(E->getCallee()->getType(), callee, E, ReturnValueSlot()); |
449 | } |
450 | |
451 | /// Emit a call to llvm.{sadd,uadd,ssub,usub,smul,umul}.with.overflow.* |
452 | /// depending on IntrinsicID. |
453 | /// |
454 | /// \arg CGF The current codegen function. |
455 | /// \arg IntrinsicID The ID for the Intrinsic we wish to generate. |
456 | /// \arg X The first argument to the llvm.*.with.overflow.*. |
457 | /// \arg Y The second argument to the llvm.*.with.overflow.*. |
458 | /// \arg Carry The carry returned by the llvm.*.with.overflow.*. |
459 | /// \returns The result (i.e. sum/product) returned by the intrinsic. |
460 | static llvm::Value *EmitOverflowIntrinsic(CodeGenFunction &CGF, |
461 | const llvm::Intrinsic::ID IntrinsicID, |
462 | llvm::Value *X, llvm::Value *Y, |
463 | llvm::Value *&Carry) { |
464 | // Make sure we have integers of the same width. |
465 | assert(X->getType() == Y->getType() &&((X->getType() == Y->getType() && "Arguments must be the same type. (Did you forget to make sure both " "arguments have the same integer width?)") ? static_cast< void> (0) : __assert_fail ("X->getType() == Y->getType() && \"Arguments must be the same type. (Did you forget to make sure both \" \"arguments have the same integer width?)\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 467, __PRETTY_FUNCTION__)) |
466 | "Arguments must be the same type. (Did you forget to make sure both "((X->getType() == Y->getType() && "Arguments must be the same type. (Did you forget to make sure both " "arguments have the same integer width?)") ? static_cast< void> (0) : __assert_fail ("X->getType() == Y->getType() && \"Arguments must be the same type. (Did you forget to make sure both \" \"arguments have the same integer width?)\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 467, __PRETTY_FUNCTION__)) |
467 | "arguments have the same integer width?)")((X->getType() == Y->getType() && "Arguments must be the same type. (Did you forget to make sure both " "arguments have the same integer width?)") ? static_cast< void> (0) : __assert_fail ("X->getType() == Y->getType() && \"Arguments must be the same type. (Did you forget to make sure both \" \"arguments have the same integer width?)\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 467, __PRETTY_FUNCTION__)); |
468 | |
469 | Function *Callee = CGF.CGM.getIntrinsic(IntrinsicID, X->getType()); |
470 | llvm::Value *Tmp = CGF.Builder.CreateCall(Callee, {X, Y}); |
471 | Carry = CGF.Builder.CreateExtractValue(Tmp, 1); |
472 | return CGF.Builder.CreateExtractValue(Tmp, 0); |
473 | } |
474 | |
475 | static Value *emitRangedBuiltin(CodeGenFunction &CGF, |
476 | unsigned IntrinsicID, |
477 | int low, int high) { |
478 | llvm::MDBuilder MDHelper(CGF.getLLVMContext()); |
479 | llvm::MDNode *RNode = MDHelper.createRange(APInt(32, low), APInt(32, high)); |
480 | Function *F = CGF.CGM.getIntrinsic(IntrinsicID, {}); |
481 | llvm::Instruction *Call = CGF.Builder.CreateCall(F); |
482 | Call->setMetadata(llvm::LLVMContext::MD_range, RNode); |
483 | return Call; |
484 | } |
485 | |
486 | namespace { |
487 | struct WidthAndSignedness { |
488 | unsigned Width; |
489 | bool Signed; |
490 | }; |
491 | } |
492 | |
493 | static WidthAndSignedness |
494 | getIntegerWidthAndSignedness(const clang::ASTContext &context, |
495 | const clang::QualType Type) { |
496 | assert(Type->isIntegerType() && "Given type is not an integer.")((Type->isIntegerType() && "Given type is not an integer." ) ? static_cast<void> (0) : __assert_fail ("Type->isIntegerType() && \"Given type is not an integer.\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 496, __PRETTY_FUNCTION__)); |
497 | unsigned Width = Type->isBooleanType() ? 1 : context.getTypeInfo(Type).Width; |
498 | bool Signed = Type->isSignedIntegerType(); |
499 | return {Width, Signed}; |
500 | } |
501 | |
502 | // Given one or more integer types, this function produces an integer type that |
503 | // encompasses them: any value in one of the given types could be expressed in |
504 | // the encompassing type. |
505 | static struct WidthAndSignedness |
506 | EncompassingIntegerType(ArrayRef<struct WidthAndSignedness> Types) { |
507 | assert(Types.size() > 0 && "Empty list of types.")((Types.size() > 0 && "Empty list of types.") ? static_cast <void> (0) : __assert_fail ("Types.size() > 0 && \"Empty list of types.\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 507, __PRETTY_FUNCTION__)); |
508 | |
509 | // If any of the given types is signed, we must return a signed type. |
510 | bool Signed = false; |
511 | for (const auto &Type : Types) { |
512 | Signed |= Type.Signed; |
513 | } |
514 | |
515 | // The encompassing type must have a width greater than or equal to the width |
516 | // of the specified types. Additionally, if the encompassing type is signed, |
517 | // its width must be strictly greater than the width of any unsigned types |
518 | // given. |
519 | unsigned Width = 0; |
520 | for (const auto &Type : Types) { |
521 | unsigned MinWidth = Type.Width + (Signed && !Type.Signed); |
522 | if (Width < MinWidth) { |
523 | Width = MinWidth; |
524 | } |
525 | } |
526 | |
527 | return {Width, Signed}; |
528 | } |
529 | |
530 | Value *CodeGenFunction::EmitVAStartEnd(Value *ArgValue, bool IsStart) { |
531 | llvm::Type *DestType = Int8PtrTy; |
532 | if (ArgValue->getType() != DestType) |
533 | ArgValue = |
534 | Builder.CreateBitCast(ArgValue, DestType, ArgValue->getName().data()); |
535 | |
536 | Intrinsic::ID inst = IsStart ? Intrinsic::vastart : Intrinsic::vaend; |
537 | return Builder.CreateCall(CGM.getIntrinsic(inst), ArgValue); |
538 | } |
539 | |
540 | /// Checks if using the result of __builtin_object_size(p, @p From) in place of |
541 | /// __builtin_object_size(p, @p To) is correct |
542 | static bool areBOSTypesCompatible(int From, int To) { |
543 | // Note: Our __builtin_object_size implementation currently treats Type=0 and |
544 | // Type=2 identically. Encoding this implementation detail here may make |
545 | // improving __builtin_object_size difficult in the future, so it's omitted. |
546 | return From == To || (From == 0 && To == 1) || (From == 3 && To == 2); |
547 | } |
548 | |
549 | static llvm::Value * |
550 | getDefaultBuiltinObjectSizeResult(unsigned Type, llvm::IntegerType *ResType) { |
551 | return ConstantInt::get(ResType, (Type & 2) ? 0 : -1, /*isSigned=*/true); |
552 | } |
553 | |
554 | llvm::Value * |
555 | CodeGenFunction::evaluateOrEmitBuiltinObjectSize(const Expr *E, unsigned Type, |
556 | llvm::IntegerType *ResType, |
557 | llvm::Value *EmittedE, |
558 | bool IsDynamic) { |
559 | uint64_t ObjectSize; |
560 | if (!E->tryEvaluateObjectSize(ObjectSize, getContext(), Type)) |
561 | return emitBuiltinObjectSize(E, Type, ResType, EmittedE, IsDynamic); |
562 | return ConstantInt::get(ResType, ObjectSize, /*isSigned=*/true); |
563 | } |
564 | |
565 | /// Returns a Value corresponding to the size of the given expression. |
566 | /// This Value may be either of the following: |
567 | /// - A llvm::Argument (if E is a param with the pass_object_size attribute on |
568 | /// it) |
569 | /// - A call to the @llvm.objectsize intrinsic |
570 | /// |
571 | /// EmittedE is the result of emitting `E` as a scalar expr. If it's non-null |
572 | /// and we wouldn't otherwise try to reference a pass_object_size parameter, |
573 | /// we'll call @llvm.objectsize on EmittedE, rather than emitting E. |
574 | llvm::Value * |
575 | CodeGenFunction::emitBuiltinObjectSize(const Expr *E, unsigned Type, |
576 | llvm::IntegerType *ResType, |
577 | llvm::Value *EmittedE, bool IsDynamic) { |
578 | // We need to reference an argument if the pointer is a parameter with the |
579 | // pass_object_size attribute. |
580 | if (auto *D = dyn_cast<DeclRefExpr>(E->IgnoreParenImpCasts())) { |
581 | auto *Param = dyn_cast<ParmVarDecl>(D->getDecl()); |
582 | auto *PS = D->getDecl()->getAttr<PassObjectSizeAttr>(); |
583 | if (Param != nullptr && PS != nullptr && |
584 | areBOSTypesCompatible(PS->getType(), Type)) { |
585 | auto Iter = SizeArguments.find(Param); |
586 | assert(Iter != SizeArguments.end())((Iter != SizeArguments.end()) ? static_cast<void> (0) : __assert_fail ("Iter != SizeArguments.end()", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 586, __PRETTY_FUNCTION__)); |
587 | |
588 | const ImplicitParamDecl *D = Iter->second; |
589 | auto DIter = LocalDeclMap.find(D); |
590 | assert(DIter != LocalDeclMap.end())((DIter != LocalDeclMap.end()) ? static_cast<void> (0) : __assert_fail ("DIter != LocalDeclMap.end()", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 590, __PRETTY_FUNCTION__)); |
591 | |
592 | return EmitLoadOfScalar(DIter->second, /*volatile=*/false, |
593 | getContext().getSizeType(), E->getBeginLoc()); |
594 | } |
595 | } |
596 | |
597 | // LLVM can't handle Type=3 appropriately, and __builtin_object_size shouldn't |
598 | // evaluate E for side-effects. In either case, we shouldn't lower to |
599 | // @llvm.objectsize. |
600 | if (Type == 3 || (!EmittedE && E->HasSideEffects(getContext()))) |
601 | return getDefaultBuiltinObjectSizeResult(Type, ResType); |
602 | |
603 | Value *Ptr = EmittedE ? EmittedE : EmitScalarExpr(E); |
604 | assert(Ptr->getType()->isPointerTy() &&((Ptr->getType()->isPointerTy() && "Non-pointer passed to __builtin_object_size?" ) ? static_cast<void> (0) : __assert_fail ("Ptr->getType()->isPointerTy() && \"Non-pointer passed to __builtin_object_size?\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 605, __PRETTY_FUNCTION__)) |
605 | "Non-pointer passed to __builtin_object_size?")((Ptr->getType()->isPointerTy() && "Non-pointer passed to __builtin_object_size?" ) ? static_cast<void> (0) : __assert_fail ("Ptr->getType()->isPointerTy() && \"Non-pointer passed to __builtin_object_size?\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 605, __PRETTY_FUNCTION__)); |
606 | |
607 | Function *F = |
608 | CGM.getIntrinsic(Intrinsic::objectsize, {ResType, Ptr->getType()}); |
609 | |
610 | // LLVM only supports 0 and 2, make sure that we pass along that as a boolean. |
611 | Value *Min = Builder.getInt1((Type & 2) != 0); |
612 | // For GCC compatibility, __builtin_object_size treat NULL as unknown size. |
613 | Value *NullIsUnknown = Builder.getTrue(); |
614 | Value *Dynamic = Builder.getInt1(IsDynamic); |
615 | return Builder.CreateCall(F, {Ptr, Min, NullIsUnknown, Dynamic}); |
616 | } |
617 | |
618 | namespace { |
619 | /// A struct to generically describe a bit test intrinsic. |
620 | struct BitTest { |
621 | enum ActionKind : uint8_t { TestOnly, Complement, Reset, Set }; |
622 | enum InterlockingKind : uint8_t { |
623 | Unlocked, |
624 | Sequential, |
625 | Acquire, |
626 | Release, |
627 | NoFence |
628 | }; |
629 | |
630 | ActionKind Action; |
631 | InterlockingKind Interlocking; |
632 | bool Is64Bit; |
633 | |
634 | static BitTest decodeBitTestBuiltin(unsigned BuiltinID); |
635 | }; |
636 | } // namespace |
637 | |
638 | BitTest BitTest::decodeBitTestBuiltin(unsigned BuiltinID) { |
639 | switch (BuiltinID) { |
640 | // Main portable variants. |
641 | case Builtin::BI_bittest: |
642 | return {TestOnly, Unlocked, false}; |
643 | case Builtin::BI_bittestandcomplement: |
644 | return {Complement, Unlocked, false}; |
645 | case Builtin::BI_bittestandreset: |
646 | return {Reset, Unlocked, false}; |
647 | case Builtin::BI_bittestandset: |
648 | return {Set, Unlocked, false}; |
649 | case Builtin::BI_interlockedbittestandreset: |
650 | return {Reset, Sequential, false}; |
651 | case Builtin::BI_interlockedbittestandset: |
652 | return {Set, Sequential, false}; |
653 | |
654 | // X86-specific 64-bit variants. |
655 | case Builtin::BI_bittest64: |
656 | return {TestOnly, Unlocked, true}; |
657 | case Builtin::BI_bittestandcomplement64: |
658 | return {Complement, Unlocked, true}; |
659 | case Builtin::BI_bittestandreset64: |
660 | return {Reset, Unlocked, true}; |
661 | case Builtin::BI_bittestandset64: |
662 | return {Set, Unlocked, true}; |
663 | case Builtin::BI_interlockedbittestandreset64: |
664 | return {Reset, Sequential, true}; |
665 | case Builtin::BI_interlockedbittestandset64: |
666 | return {Set, Sequential, true}; |
667 | |
668 | // ARM/AArch64-specific ordering variants. |
669 | case Builtin::BI_interlockedbittestandset_acq: |
670 | return {Set, Acquire, false}; |
671 | case Builtin::BI_interlockedbittestandset_rel: |
672 | return {Set, Release, false}; |
673 | case Builtin::BI_interlockedbittestandset_nf: |
674 | return {Set, NoFence, false}; |
675 | case Builtin::BI_interlockedbittestandreset_acq: |
676 | return {Reset, Acquire, false}; |
677 | case Builtin::BI_interlockedbittestandreset_rel: |
678 | return {Reset, Release, false}; |
679 | case Builtin::BI_interlockedbittestandreset_nf: |
680 | return {Reset, NoFence, false}; |
681 | } |
682 | llvm_unreachable("expected only bittest intrinsics")::llvm::llvm_unreachable_internal("expected only bittest intrinsics" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 682); |
683 | } |
684 | |
685 | static char bitActionToX86BTCode(BitTest::ActionKind A) { |
686 | switch (A) { |
687 | case BitTest::TestOnly: return '\0'; |
688 | case BitTest::Complement: return 'c'; |
689 | case BitTest::Reset: return 'r'; |
690 | case BitTest::Set: return 's'; |
691 | } |
692 | llvm_unreachable("invalid action")::llvm::llvm_unreachable_internal("invalid action", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 692); |
693 | } |
694 | |
695 | static llvm::Value *EmitX86BitTestIntrinsic(CodeGenFunction &CGF, |
696 | BitTest BT, |
697 | const CallExpr *E, Value *BitBase, |
698 | Value *BitPos) { |
699 | char Action = bitActionToX86BTCode(BT.Action); |
700 | char SizeSuffix = BT.Is64Bit ? 'q' : 'l'; |
701 | |
702 | // Build the assembly. |
703 | SmallString<64> Asm; |
704 | raw_svector_ostream AsmOS(Asm); |
705 | if (BT.Interlocking != BitTest::Unlocked) |
706 | AsmOS << "lock "; |
707 | AsmOS << "bt"; |
708 | if (Action) |
709 | AsmOS << Action; |
710 | AsmOS << SizeSuffix << " $2, ($1)\n\tsetc ${0:b}"; |
711 | |
712 | // Build the constraints. FIXME: We should support immediates when possible. |
713 | std::string Constraints = "=r,r,r,~{cc},~{flags},~{fpsr}"; |
714 | llvm::IntegerType *IntType = llvm::IntegerType::get( |
715 | CGF.getLLVMContext(), |
716 | CGF.getContext().getTypeSize(E->getArg(1)->getType())); |
717 | llvm::Type *IntPtrType = IntType->getPointerTo(); |
718 | llvm::FunctionType *FTy = |
719 | llvm::FunctionType::get(CGF.Int8Ty, {IntPtrType, IntType}, false); |
720 | |
721 | llvm::InlineAsm *IA = |
722 | llvm::InlineAsm::get(FTy, Asm, Constraints, /*SideEffects=*/true); |
723 | return CGF.Builder.CreateCall(IA, {BitBase, BitPos}); |
724 | } |
725 | |
726 | static llvm::AtomicOrdering |
727 | getBitTestAtomicOrdering(BitTest::InterlockingKind I) { |
728 | switch (I) { |
729 | case BitTest::Unlocked: return llvm::AtomicOrdering::NotAtomic; |
730 | case BitTest::Sequential: return llvm::AtomicOrdering::SequentiallyConsistent; |
731 | case BitTest::Acquire: return llvm::AtomicOrdering::Acquire; |
732 | case BitTest::Release: return llvm::AtomicOrdering::Release; |
733 | case BitTest::NoFence: return llvm::AtomicOrdering::Monotonic; |
734 | } |
735 | llvm_unreachable("invalid interlocking")::llvm::llvm_unreachable_internal("invalid interlocking", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 735); |
736 | } |
737 | |
738 | /// Emit a _bittest* intrinsic. These intrinsics take a pointer to an array of |
739 | /// bits and a bit position and read and optionally modify the bit at that |
740 | /// position. The position index can be arbitrarily large, i.e. it can be larger |
741 | /// than 31 or 63, so we need an indexed load in the general case. |
742 | static llvm::Value *EmitBitTestIntrinsic(CodeGenFunction &CGF, |
743 | unsigned BuiltinID, |
744 | const CallExpr *E) { |
745 | Value *BitBase = CGF.EmitScalarExpr(E->getArg(0)); |
746 | Value *BitPos = CGF.EmitScalarExpr(E->getArg(1)); |
747 | |
748 | BitTest BT = BitTest::decodeBitTestBuiltin(BuiltinID); |
749 | |
750 | // X86 has special BT, BTC, BTR, and BTS instructions that handle the array |
751 | // indexing operation internally. Use them if possible. |
752 | llvm::Triple::ArchType Arch = CGF.getTarget().getTriple().getArch(); |
753 | if (Arch == llvm::Triple::x86 || Arch == llvm::Triple::x86_64) |
754 | return EmitX86BitTestIntrinsic(CGF, BT, E, BitBase, BitPos); |
755 | |
756 | // Otherwise, use generic code to load one byte and test the bit. Use all but |
757 | // the bottom three bits as the array index, and the bottom three bits to form |
758 | // a mask. |
759 | // Bit = BitBaseI8[BitPos >> 3] & (1 << (BitPos & 0x7)) != 0; |
760 | Value *ByteIndex = CGF.Builder.CreateAShr( |
761 | BitPos, llvm::ConstantInt::get(BitPos->getType(), 3), "bittest.byteidx"); |
762 | Value *BitBaseI8 = CGF.Builder.CreatePointerCast(BitBase, CGF.Int8PtrTy); |
763 | Address ByteAddr(CGF.Builder.CreateInBoundsGEP(CGF.Int8Ty, BitBaseI8, |
764 | ByteIndex, "bittest.byteaddr"), |
765 | CharUnits::One()); |
766 | Value *PosLow = |
767 | CGF.Builder.CreateAnd(CGF.Builder.CreateTrunc(BitPos, CGF.Int8Ty), |
768 | llvm::ConstantInt::get(CGF.Int8Ty, 0x7)); |
769 | |
770 | // The updating instructions will need a mask. |
771 | Value *Mask = nullptr; |
772 | if (BT.Action != BitTest::TestOnly) { |
773 | Mask = CGF.Builder.CreateShl(llvm::ConstantInt::get(CGF.Int8Ty, 1), PosLow, |
774 | "bittest.mask"); |
775 | } |
776 | |
777 | // Check the action and ordering of the interlocked intrinsics. |
778 | llvm::AtomicOrdering Ordering = getBitTestAtomicOrdering(BT.Interlocking); |
779 | |
780 | Value *OldByte = nullptr; |
781 | if (Ordering != llvm::AtomicOrdering::NotAtomic) { |
782 | // Emit a combined atomicrmw load/store operation for the interlocked |
783 | // intrinsics. |
784 | llvm::AtomicRMWInst::BinOp RMWOp = llvm::AtomicRMWInst::Or; |
785 | if (BT.Action == BitTest::Reset) { |
786 | Mask = CGF.Builder.CreateNot(Mask); |
787 | RMWOp = llvm::AtomicRMWInst::And; |
788 | } |
789 | OldByte = CGF.Builder.CreateAtomicRMW(RMWOp, ByteAddr.getPointer(), Mask, |
790 | Ordering); |
791 | } else { |
792 | // Emit a plain load for the non-interlocked intrinsics. |
793 | OldByte = CGF.Builder.CreateLoad(ByteAddr, "bittest.byte"); |
794 | Value *NewByte = nullptr; |
795 | switch (BT.Action) { |
796 | case BitTest::TestOnly: |
797 | // Don't store anything. |
798 | break; |
799 | case BitTest::Complement: |
800 | NewByte = CGF.Builder.CreateXor(OldByte, Mask); |
801 | break; |
802 | case BitTest::Reset: |
803 | NewByte = CGF.Builder.CreateAnd(OldByte, CGF.Builder.CreateNot(Mask)); |
804 | break; |
805 | case BitTest::Set: |
806 | NewByte = CGF.Builder.CreateOr(OldByte, Mask); |
807 | break; |
808 | } |
809 | if (NewByte) |
810 | CGF.Builder.CreateStore(NewByte, ByteAddr); |
811 | } |
812 | |
813 | // However we loaded the old byte, either by plain load or atomicrmw, shift |
814 | // the bit into the low position and mask it to 0 or 1. |
815 | Value *ShiftedByte = CGF.Builder.CreateLShr(OldByte, PosLow, "bittest.shr"); |
816 | return CGF.Builder.CreateAnd( |
817 | ShiftedByte, llvm::ConstantInt::get(CGF.Int8Ty, 1), "bittest.res"); |
818 | } |
819 | |
820 | namespace { |
821 | enum class MSVCSetJmpKind { |
822 | _setjmpex, |
823 | _setjmp3, |
824 | _setjmp |
825 | }; |
826 | } |
827 | |
828 | /// MSVC handles setjmp a bit differently on different platforms. On every |
829 | /// architecture except 32-bit x86, the frame address is passed. On x86, extra |
830 | /// parameters can be passed as variadic arguments, but we always pass none. |
831 | static RValue EmitMSVCRTSetJmp(CodeGenFunction &CGF, MSVCSetJmpKind SJKind, |
832 | const CallExpr *E) { |
833 | llvm::Value *Arg1 = nullptr; |
834 | llvm::Type *Arg1Ty = nullptr; |
835 | StringRef Name; |
836 | bool IsVarArg = false; |
837 | if (SJKind == MSVCSetJmpKind::_setjmp3) { |
838 | Name = "_setjmp3"; |
839 | Arg1Ty = CGF.Int32Ty; |
840 | Arg1 = llvm::ConstantInt::get(CGF.IntTy, 0); |
841 | IsVarArg = true; |
842 | } else { |
843 | Name = SJKind == MSVCSetJmpKind::_setjmp ? "_setjmp" : "_setjmpex"; |
844 | Arg1Ty = CGF.Int8PtrTy; |
845 | if (CGF.getTarget().getTriple().getArch() == llvm::Triple::aarch64) { |
846 | Arg1 = CGF.Builder.CreateCall(CGF.CGM.getIntrinsic(Intrinsic::sponentry)); |
847 | } else |
848 | Arg1 = CGF.Builder.CreateCall(CGF.CGM.getIntrinsic(Intrinsic::frameaddress), |
849 | llvm::ConstantInt::get(CGF.Int32Ty, 0)); |
850 | } |
851 | |
852 | // Mark the call site and declaration with ReturnsTwice. |
853 | llvm::Type *ArgTypes[2] = {CGF.Int8PtrTy, Arg1Ty}; |
854 | llvm::AttributeList ReturnsTwiceAttr = llvm::AttributeList::get( |
855 | CGF.getLLVMContext(), llvm::AttributeList::FunctionIndex, |
856 | llvm::Attribute::ReturnsTwice); |
857 | llvm::FunctionCallee SetJmpFn = CGF.CGM.CreateRuntimeFunction( |
858 | llvm::FunctionType::get(CGF.IntTy, ArgTypes, IsVarArg), Name, |
859 | ReturnsTwiceAttr, /*Local=*/true); |
860 | |
861 | llvm::Value *Buf = CGF.Builder.CreateBitOrPointerCast( |
862 | CGF.EmitScalarExpr(E->getArg(0)), CGF.Int8PtrTy); |
863 | llvm::Value *Args[] = {Buf, Arg1}; |
864 | llvm::CallBase *CB = CGF.EmitRuntimeCallOrInvoke(SetJmpFn, Args); |
865 | CB->setAttributes(ReturnsTwiceAttr); |
866 | return RValue::get(CB); |
867 | } |
868 | |
869 | // Many of MSVC builtins are on x64, ARM and AArch64; to avoid repeating code, |
870 | // we handle them here. |
871 | enum class CodeGenFunction::MSVCIntrin { |
872 | _BitScanForward, |
873 | _BitScanReverse, |
874 | _InterlockedAnd, |
875 | _InterlockedDecrement, |
876 | _InterlockedExchange, |
877 | _InterlockedExchangeAdd, |
878 | _InterlockedExchangeSub, |
879 | _InterlockedIncrement, |
880 | _InterlockedOr, |
881 | _InterlockedXor, |
882 | _InterlockedExchangeAdd_acq, |
883 | _InterlockedExchangeAdd_rel, |
884 | _InterlockedExchangeAdd_nf, |
885 | _InterlockedExchange_acq, |
886 | _InterlockedExchange_rel, |
887 | _InterlockedExchange_nf, |
888 | _InterlockedCompareExchange_acq, |
889 | _InterlockedCompareExchange_rel, |
890 | _InterlockedCompareExchange_nf, |
891 | _InterlockedOr_acq, |
892 | _InterlockedOr_rel, |
893 | _InterlockedOr_nf, |
894 | _InterlockedXor_acq, |
895 | _InterlockedXor_rel, |
896 | _InterlockedXor_nf, |
897 | _InterlockedAnd_acq, |
898 | _InterlockedAnd_rel, |
899 | _InterlockedAnd_nf, |
900 | _InterlockedIncrement_acq, |
901 | _InterlockedIncrement_rel, |
902 | _InterlockedIncrement_nf, |
903 | _InterlockedDecrement_acq, |
904 | _InterlockedDecrement_rel, |
905 | _InterlockedDecrement_nf, |
906 | __fastfail, |
907 | }; |
908 | |
909 | Value *CodeGenFunction::EmitMSVCBuiltinExpr(MSVCIntrin BuiltinID, |
910 | const CallExpr *E) { |
911 | switch (BuiltinID) { |
912 | case MSVCIntrin::_BitScanForward: |
913 | case MSVCIntrin::_BitScanReverse: { |
914 | Value *ArgValue = EmitScalarExpr(E->getArg(1)); |
915 | |
916 | llvm::Type *ArgType = ArgValue->getType(); |
917 | llvm::Type *IndexType = |
918 | EmitScalarExpr(E->getArg(0))->getType()->getPointerElementType(); |
919 | llvm::Type *ResultType = ConvertType(E->getType()); |
920 | |
921 | Value *ArgZero = llvm::Constant::getNullValue(ArgType); |
922 | Value *ResZero = llvm::Constant::getNullValue(ResultType); |
923 | Value *ResOne = llvm::ConstantInt::get(ResultType, 1); |
924 | |
925 | BasicBlock *Begin = Builder.GetInsertBlock(); |
926 | BasicBlock *End = createBasicBlock("bitscan_end", this->CurFn); |
927 | Builder.SetInsertPoint(End); |
928 | PHINode *Result = Builder.CreatePHI(ResultType, 2, "bitscan_result"); |
929 | |
930 | Builder.SetInsertPoint(Begin); |
931 | Value *IsZero = Builder.CreateICmpEQ(ArgValue, ArgZero); |
932 | BasicBlock *NotZero = createBasicBlock("bitscan_not_zero", this->CurFn); |
933 | Builder.CreateCondBr(IsZero, End, NotZero); |
934 | Result->addIncoming(ResZero, Begin); |
935 | |
936 | Builder.SetInsertPoint(NotZero); |
937 | Address IndexAddress = EmitPointerWithAlignment(E->getArg(0)); |
938 | |
939 | if (BuiltinID == MSVCIntrin::_BitScanForward) { |
940 | Function *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); |
941 | Value *ZeroCount = Builder.CreateCall(F, {ArgValue, Builder.getTrue()}); |
942 | ZeroCount = Builder.CreateIntCast(ZeroCount, IndexType, false); |
943 | Builder.CreateStore(ZeroCount, IndexAddress, false); |
944 | } else { |
945 | unsigned ArgWidth = cast<llvm::IntegerType>(ArgType)->getBitWidth(); |
946 | Value *ArgTypeLastIndex = llvm::ConstantInt::get(IndexType, ArgWidth - 1); |
947 | |
948 | Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType); |
949 | Value *ZeroCount = Builder.CreateCall(F, {ArgValue, Builder.getTrue()}); |
950 | ZeroCount = Builder.CreateIntCast(ZeroCount, IndexType, false); |
951 | Value *Index = Builder.CreateNSWSub(ArgTypeLastIndex, ZeroCount); |
952 | Builder.CreateStore(Index, IndexAddress, false); |
953 | } |
954 | Builder.CreateBr(End); |
955 | Result->addIncoming(ResOne, NotZero); |
956 | |
957 | Builder.SetInsertPoint(End); |
958 | return Result; |
959 | } |
960 | case MSVCIntrin::_InterlockedAnd: |
961 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::And, E); |
962 | case MSVCIntrin::_InterlockedExchange: |
963 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xchg, E); |
964 | case MSVCIntrin::_InterlockedExchangeAdd: |
965 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Add, E); |
966 | case MSVCIntrin::_InterlockedExchangeSub: |
967 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Sub, E); |
968 | case MSVCIntrin::_InterlockedOr: |
969 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Or, E); |
970 | case MSVCIntrin::_InterlockedXor: |
971 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xor, E); |
972 | case MSVCIntrin::_InterlockedExchangeAdd_acq: |
973 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Add, E, |
974 | AtomicOrdering::Acquire); |
975 | case MSVCIntrin::_InterlockedExchangeAdd_rel: |
976 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Add, E, |
977 | AtomicOrdering::Release); |
978 | case MSVCIntrin::_InterlockedExchangeAdd_nf: |
979 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Add, E, |
980 | AtomicOrdering::Monotonic); |
981 | case MSVCIntrin::_InterlockedExchange_acq: |
982 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xchg, E, |
983 | AtomicOrdering::Acquire); |
984 | case MSVCIntrin::_InterlockedExchange_rel: |
985 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xchg, E, |
986 | AtomicOrdering::Release); |
987 | case MSVCIntrin::_InterlockedExchange_nf: |
988 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xchg, E, |
989 | AtomicOrdering::Monotonic); |
990 | case MSVCIntrin::_InterlockedCompareExchange_acq: |
991 | return EmitAtomicCmpXchgForMSIntrin(*this, E, AtomicOrdering::Acquire); |
992 | case MSVCIntrin::_InterlockedCompareExchange_rel: |
993 | return EmitAtomicCmpXchgForMSIntrin(*this, E, AtomicOrdering::Release); |
994 | case MSVCIntrin::_InterlockedCompareExchange_nf: |
995 | return EmitAtomicCmpXchgForMSIntrin(*this, E, AtomicOrdering::Monotonic); |
996 | case MSVCIntrin::_InterlockedOr_acq: |
997 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Or, E, |
998 | AtomicOrdering::Acquire); |
999 | case MSVCIntrin::_InterlockedOr_rel: |
1000 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Or, E, |
1001 | AtomicOrdering::Release); |
1002 | case MSVCIntrin::_InterlockedOr_nf: |
1003 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Or, E, |
1004 | AtomicOrdering::Monotonic); |
1005 | case MSVCIntrin::_InterlockedXor_acq: |
1006 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xor, E, |
1007 | AtomicOrdering::Acquire); |
1008 | case MSVCIntrin::_InterlockedXor_rel: |
1009 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xor, E, |
1010 | AtomicOrdering::Release); |
1011 | case MSVCIntrin::_InterlockedXor_nf: |
1012 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xor, E, |
1013 | AtomicOrdering::Monotonic); |
1014 | case MSVCIntrin::_InterlockedAnd_acq: |
1015 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::And, E, |
1016 | AtomicOrdering::Acquire); |
1017 | case MSVCIntrin::_InterlockedAnd_rel: |
1018 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::And, E, |
1019 | AtomicOrdering::Release); |
1020 | case MSVCIntrin::_InterlockedAnd_nf: |
1021 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::And, E, |
1022 | AtomicOrdering::Monotonic); |
1023 | case MSVCIntrin::_InterlockedIncrement_acq: |
1024 | return EmitAtomicIncrementValue(*this, E, AtomicOrdering::Acquire); |
1025 | case MSVCIntrin::_InterlockedIncrement_rel: |
1026 | return EmitAtomicIncrementValue(*this, E, AtomicOrdering::Release); |
1027 | case MSVCIntrin::_InterlockedIncrement_nf: |
1028 | return EmitAtomicIncrementValue(*this, E, AtomicOrdering::Monotonic); |
1029 | case MSVCIntrin::_InterlockedDecrement_acq: |
1030 | return EmitAtomicDecrementValue(*this, E, AtomicOrdering::Acquire); |
1031 | case MSVCIntrin::_InterlockedDecrement_rel: |
1032 | return EmitAtomicDecrementValue(*this, E, AtomicOrdering::Release); |
1033 | case MSVCIntrin::_InterlockedDecrement_nf: |
1034 | return EmitAtomicDecrementValue(*this, E, AtomicOrdering::Monotonic); |
1035 | |
1036 | case MSVCIntrin::_InterlockedDecrement: |
1037 | return EmitAtomicDecrementValue(*this, E); |
1038 | case MSVCIntrin::_InterlockedIncrement: |
1039 | return EmitAtomicIncrementValue(*this, E); |
1040 | |
1041 | case MSVCIntrin::__fastfail: { |
1042 | // Request immediate process termination from the kernel. The instruction |
1043 | // sequences to do this are documented on MSDN: |
1044 | // https://msdn.microsoft.com/en-us/library/dn774154.aspx |
1045 | llvm::Triple::ArchType ISA = getTarget().getTriple().getArch(); |
1046 | StringRef Asm, Constraints; |
1047 | switch (ISA) { |
1048 | default: |
1049 | ErrorUnsupported(E, "__fastfail call for this architecture"); |
1050 | break; |
1051 | case llvm::Triple::x86: |
1052 | case llvm::Triple::x86_64: |
1053 | Asm = "int $$0x29"; |
1054 | Constraints = "{cx}"; |
1055 | break; |
1056 | case llvm::Triple::thumb: |
1057 | Asm = "udf #251"; |
1058 | Constraints = "{r0}"; |
1059 | break; |
1060 | case llvm::Triple::aarch64: |
1061 | Asm = "brk #0xF003"; |
1062 | Constraints = "{w0}"; |
1063 | } |
1064 | llvm::FunctionType *FTy = llvm::FunctionType::get(VoidTy, {Int32Ty}, false); |
1065 | llvm::InlineAsm *IA = |
1066 | llvm::InlineAsm::get(FTy, Asm, Constraints, /*SideEffects=*/true); |
1067 | llvm::AttributeList NoReturnAttr = llvm::AttributeList::get( |
1068 | getLLVMContext(), llvm::AttributeList::FunctionIndex, |
1069 | llvm::Attribute::NoReturn); |
1070 | llvm::CallInst *CI = Builder.CreateCall(IA, EmitScalarExpr(E->getArg(0))); |
1071 | CI->setAttributes(NoReturnAttr); |
1072 | return CI; |
1073 | } |
1074 | } |
1075 | llvm_unreachable("Incorrect MSVC intrinsic!")::llvm::llvm_unreachable_internal("Incorrect MSVC intrinsic!" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 1075); |
1076 | } |
1077 | |
1078 | namespace { |
1079 | // ARC cleanup for __builtin_os_log_format |
1080 | struct CallObjCArcUse final : EHScopeStack::Cleanup { |
1081 | CallObjCArcUse(llvm::Value *object) : object(object) {} |
1082 | llvm::Value *object; |
1083 | |
1084 | void Emit(CodeGenFunction &CGF, Flags flags) override { |
1085 | CGF.EmitARCIntrinsicUse(object); |
1086 | } |
1087 | }; |
1088 | } |
1089 | |
1090 | Value *CodeGenFunction::EmitCheckedArgForBuiltin(const Expr *E, |
1091 | BuiltinCheckKind Kind) { |
1092 | assert((Kind == BCK_CLZPassedZero || Kind == BCK_CTZPassedZero)(((Kind == BCK_CLZPassedZero || Kind == BCK_CTZPassedZero) && "Unsupported builtin check kind") ? static_cast<void> ( 0) : __assert_fail ("(Kind == BCK_CLZPassedZero || Kind == BCK_CTZPassedZero) && \"Unsupported builtin check kind\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 1093, __PRETTY_FUNCTION__)) |
1093 | && "Unsupported builtin check kind")(((Kind == BCK_CLZPassedZero || Kind == BCK_CTZPassedZero) && "Unsupported builtin check kind") ? static_cast<void> ( 0) : __assert_fail ("(Kind == BCK_CLZPassedZero || Kind == BCK_CTZPassedZero) && \"Unsupported builtin check kind\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 1093, __PRETTY_FUNCTION__)); |
1094 | |
1095 | Value *ArgValue = EmitScalarExpr(E); |
1096 | if (!SanOpts.has(SanitizerKind::Builtin) || !getTarget().isCLZForZeroUndef()) |
1097 | return ArgValue; |
1098 | |
1099 | SanitizerScope SanScope(this); |
1100 | Value *Cond = Builder.CreateICmpNE( |
1101 | ArgValue, llvm::Constant::getNullValue(ArgValue->getType())); |
1102 | EmitCheck(std::make_pair(Cond, SanitizerKind::Builtin), |
1103 | SanitizerHandler::InvalidBuiltin, |
1104 | {EmitCheckSourceLocation(E->getExprLoc()), |
1105 | llvm::ConstantInt::get(Builder.getInt8Ty(), Kind)}, |
1106 | None); |
1107 | return ArgValue; |
1108 | } |
1109 | |
1110 | /// Get the argument type for arguments to os_log_helper. |
1111 | static CanQualType getOSLogArgType(ASTContext &C, int Size) { |
1112 | QualType UnsignedTy = C.getIntTypeForBitwidth(Size * 8, /*Signed=*/false); |
1113 | return C.getCanonicalType(UnsignedTy); |
1114 | } |
1115 | |
1116 | llvm::Function *CodeGenFunction::generateBuiltinOSLogHelperFunction( |
1117 | const analyze_os_log::OSLogBufferLayout &Layout, |
1118 | CharUnits BufferAlignment) { |
1119 | ASTContext &Ctx = getContext(); |
1120 | |
1121 | llvm::SmallString<64> Name; |
1122 | { |
1123 | raw_svector_ostream OS(Name); |
1124 | OS << "__os_log_helper"; |
1125 | OS << "_" << BufferAlignment.getQuantity(); |
1126 | OS << "_" << int(Layout.getSummaryByte()); |
1127 | OS << "_" << int(Layout.getNumArgsByte()); |
1128 | for (const auto &Item : Layout.Items) |
1129 | OS << "_" << int(Item.getSizeByte()) << "_" |
1130 | << int(Item.getDescriptorByte()); |
1131 | } |
1132 | |
1133 | if (llvm::Function *F = CGM.getModule().getFunction(Name)) |
1134 | return F; |
1135 | |
1136 | llvm::SmallVector<QualType, 4> ArgTys; |
1137 | llvm::SmallVector<ImplicitParamDecl, 4> Params; |
1138 | Params.emplace_back(Ctx, nullptr, SourceLocation(), &Ctx.Idents.get("buffer"), |
1139 | Ctx.VoidPtrTy, ImplicitParamDecl::Other); |
1140 | ArgTys.emplace_back(Ctx.VoidPtrTy); |
1141 | |
1142 | for (unsigned int I = 0, E = Layout.Items.size(); I < E; ++I) { |
1143 | char Size = Layout.Items[I].getSizeByte(); |
1144 | if (!Size) |
1145 | continue; |
1146 | |
1147 | QualType ArgTy = getOSLogArgType(Ctx, Size); |
1148 | Params.emplace_back( |
1149 | Ctx, nullptr, SourceLocation(), |
1150 | &Ctx.Idents.get(std::string("arg") + llvm::to_string(I)), ArgTy, |
1151 | ImplicitParamDecl::Other); |
1152 | ArgTys.emplace_back(ArgTy); |
1153 | } |
1154 | |
1155 | FunctionArgList Args; |
1156 | for (auto &P : Params) |
1157 | Args.push_back(&P); |
1158 | |
1159 | QualType ReturnTy = Ctx.VoidTy; |
1160 | QualType FuncionTy = Ctx.getFunctionType(ReturnTy, ArgTys, {}); |
1161 | |
1162 | // The helper function has linkonce_odr linkage to enable the linker to merge |
1163 | // identical functions. To ensure the merging always happens, 'noinline' is |
1164 | // attached to the function when compiling with -Oz. |
1165 | const CGFunctionInfo &FI = |
1166 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(ReturnTy, Args); |
1167 | llvm::FunctionType *FuncTy = CGM.getTypes().GetFunctionType(FI); |
1168 | llvm::Function *Fn = llvm::Function::Create( |
1169 | FuncTy, llvm::GlobalValue::LinkOnceODRLinkage, Name, &CGM.getModule()); |
1170 | Fn->setVisibility(llvm::GlobalValue::HiddenVisibility); |
1171 | CGM.SetLLVMFunctionAttributes(GlobalDecl(), FI, Fn); |
1172 | CGM.SetLLVMFunctionAttributesForDefinition(nullptr, Fn); |
1173 | Fn->setDoesNotThrow(); |
1174 | |
1175 | // Attach 'noinline' at -Oz. |
1176 | if (CGM.getCodeGenOpts().OptimizeSize == 2) |
1177 | Fn->addFnAttr(llvm::Attribute::NoInline); |
1178 | |
1179 | auto NL = ApplyDebugLocation::CreateEmpty(*this); |
1180 | IdentifierInfo *II = &Ctx.Idents.get(Name); |
1181 | FunctionDecl *FD = FunctionDecl::Create( |
1182 | Ctx, Ctx.getTranslationUnitDecl(), SourceLocation(), SourceLocation(), II, |
1183 | FuncionTy, nullptr, SC_PrivateExtern, false, false); |
1184 | |
1185 | StartFunction(FD, ReturnTy, Fn, FI, Args); |
1186 | |
1187 | // Create a scope with an artificial location for the body of this function. |
1188 | auto AL = ApplyDebugLocation::CreateArtificial(*this); |
1189 | |
1190 | CharUnits Offset; |
1191 | Address BufAddr(Builder.CreateLoad(GetAddrOfLocalVar(&Params[0]), "buf"), |
1192 | BufferAlignment); |
1193 | Builder.CreateStore(Builder.getInt8(Layout.getSummaryByte()), |
1194 | Builder.CreateConstByteGEP(BufAddr, Offset++, "summary")); |
1195 | Builder.CreateStore(Builder.getInt8(Layout.getNumArgsByte()), |
1196 | Builder.CreateConstByteGEP(BufAddr, Offset++, "numArgs")); |
1197 | |
1198 | unsigned I = 1; |
1199 | for (const auto &Item : Layout.Items) { |
1200 | Builder.CreateStore( |
1201 | Builder.getInt8(Item.getDescriptorByte()), |
1202 | Builder.CreateConstByteGEP(BufAddr, Offset++, "argDescriptor")); |
1203 | Builder.CreateStore( |
1204 | Builder.getInt8(Item.getSizeByte()), |
1205 | Builder.CreateConstByteGEP(BufAddr, Offset++, "argSize")); |
1206 | |
1207 | CharUnits Size = Item.size(); |
1208 | if (!Size.getQuantity()) |
1209 | continue; |
1210 | |
1211 | Address Arg = GetAddrOfLocalVar(&Params[I]); |
1212 | Address Addr = Builder.CreateConstByteGEP(BufAddr, Offset, "argData"); |
1213 | Addr = Builder.CreateBitCast(Addr, Arg.getPointer()->getType(), |
1214 | "argDataCast"); |
1215 | Builder.CreateStore(Builder.CreateLoad(Arg), Addr); |
1216 | Offset += Size; |
1217 | ++I; |
1218 | } |
1219 | |
1220 | FinishFunction(); |
1221 | |
1222 | return Fn; |
1223 | } |
1224 | |
1225 | RValue CodeGenFunction::emitBuiltinOSLogFormat(const CallExpr &E) { |
1226 | assert(E.getNumArgs() >= 2 &&((E.getNumArgs() >= 2 && "__builtin_os_log_format takes at least 2 arguments" ) ? static_cast<void> (0) : __assert_fail ("E.getNumArgs() >= 2 && \"__builtin_os_log_format takes at least 2 arguments\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 1227, __PRETTY_FUNCTION__)) |
1227 | "__builtin_os_log_format takes at least 2 arguments")((E.getNumArgs() >= 2 && "__builtin_os_log_format takes at least 2 arguments" ) ? static_cast<void> (0) : __assert_fail ("E.getNumArgs() >= 2 && \"__builtin_os_log_format takes at least 2 arguments\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 1227, __PRETTY_FUNCTION__)); |
1228 | ASTContext &Ctx = getContext(); |
1229 | analyze_os_log::OSLogBufferLayout Layout; |
1230 | analyze_os_log::computeOSLogBufferLayout(Ctx, &E, Layout); |
1231 | Address BufAddr = EmitPointerWithAlignment(E.getArg(0)); |
1232 | llvm::SmallVector<llvm::Value *, 4> RetainableOperands; |
1233 | |
1234 | // Ignore argument 1, the format string. It is not currently used. |
1235 | CallArgList Args; |
1236 | Args.add(RValue::get(BufAddr.getPointer()), Ctx.VoidPtrTy); |
1237 | |
1238 | for (const auto &Item : Layout.Items) { |
1239 | int Size = Item.getSizeByte(); |
1240 | if (!Size) |
1241 | continue; |
1242 | |
1243 | llvm::Value *ArgVal; |
1244 | |
1245 | if (Item.getKind() == analyze_os_log::OSLogBufferItem::MaskKind) { |
1246 | uint64_t Val = 0; |
1247 | for (unsigned I = 0, E = Item.getMaskType().size(); I < E; ++I) |
1248 | Val |= ((uint64_t)Item.getMaskType()[I]) << I * 8; |
1249 | ArgVal = llvm::Constant::getIntegerValue(Int64Ty, llvm::APInt(64, Val)); |
1250 | } else if (const Expr *TheExpr = Item.getExpr()) { |
1251 | ArgVal = EmitScalarExpr(TheExpr, /*Ignore*/ false); |
1252 | |
1253 | // Check if this is a retainable type. |
1254 | if (TheExpr->getType()->isObjCRetainableType()) { |
1255 | assert(getEvaluationKind(TheExpr->getType()) == TEK_Scalar &&((getEvaluationKind(TheExpr->getType()) == TEK_Scalar && "Only scalar can be a ObjC retainable type") ? static_cast< void> (0) : __assert_fail ("getEvaluationKind(TheExpr->getType()) == TEK_Scalar && \"Only scalar can be a ObjC retainable type\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 1256, __PRETTY_FUNCTION__)) |
1256 | "Only scalar can be a ObjC retainable type")((getEvaluationKind(TheExpr->getType()) == TEK_Scalar && "Only scalar can be a ObjC retainable type") ? static_cast< void> (0) : __assert_fail ("getEvaluationKind(TheExpr->getType()) == TEK_Scalar && \"Only scalar can be a ObjC retainable type\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 1256, __PRETTY_FUNCTION__)); |
1257 | // Check if the object is constant, if not, save it in |
1258 | // RetainableOperands. |
1259 | if (!isa<Constant>(ArgVal)) |
1260 | RetainableOperands.push_back(ArgVal); |
1261 | } |
1262 | } else { |
1263 | ArgVal = Builder.getInt32(Item.getConstValue().getQuantity()); |
1264 | } |
1265 | |
1266 | unsigned ArgValSize = |
1267 | CGM.getDataLayout().getTypeSizeInBits(ArgVal->getType()); |
1268 | llvm::IntegerType *IntTy = llvm::Type::getIntNTy(getLLVMContext(), |
1269 | ArgValSize); |
1270 | ArgVal = Builder.CreateBitOrPointerCast(ArgVal, IntTy); |
1271 | CanQualType ArgTy = getOSLogArgType(Ctx, Size); |
1272 | // If ArgVal has type x86_fp80, zero-extend ArgVal. |
1273 | ArgVal = Builder.CreateZExtOrBitCast(ArgVal, ConvertType(ArgTy)); |
1274 | Args.add(RValue::get(ArgVal), ArgTy); |
1275 | } |
1276 | |
1277 | const CGFunctionInfo &FI = |
1278 | CGM.getTypes().arrangeBuiltinFunctionCall(Ctx.VoidTy, Args); |
1279 | llvm::Function *F = CodeGenFunction(CGM).generateBuiltinOSLogHelperFunction( |
1280 | Layout, BufAddr.getAlignment()); |
1281 | EmitCall(FI, CGCallee::forDirect(F), ReturnValueSlot(), Args); |
1282 | |
1283 | // Push a clang.arc.use cleanup for each object in RetainableOperands. The |
1284 | // cleanup will cause the use to appear after the final log call, keeping |
1285 | // the object valid while it’s held in the log buffer. Note that if there’s |
1286 | // a release cleanup on the object, it will already be active; since |
1287 | // cleanups are emitted in reverse order, the use will occur before the |
1288 | // object is released. |
1289 | if (!RetainableOperands.empty() && getLangOpts().ObjCAutoRefCount && |
1290 | CGM.getCodeGenOpts().OptimizationLevel != 0) |
1291 | for (llvm::Value *Object : RetainableOperands) |
1292 | pushFullExprCleanup<CallObjCArcUse>(getARCCleanupKind(), Object); |
1293 | |
1294 | return RValue::get(BufAddr.getPointer()); |
1295 | } |
1296 | |
1297 | /// Determine if a binop is a checked mixed-sign multiply we can specialize. |
1298 | static bool isSpecialMixedSignMultiply(unsigned BuiltinID, |
1299 | WidthAndSignedness Op1Info, |
1300 | WidthAndSignedness Op2Info, |
1301 | WidthAndSignedness ResultInfo) { |
1302 | return BuiltinID == Builtin::BI__builtin_mul_overflow && |
1303 | std::max(Op1Info.Width, Op2Info.Width) >= ResultInfo.Width && |
1304 | Op1Info.Signed != Op2Info.Signed; |
1305 | } |
1306 | |
1307 | /// Emit a checked mixed-sign multiply. This is a cheaper specialization of |
1308 | /// the generic checked-binop irgen. |
1309 | static RValue |
1310 | EmitCheckedMixedSignMultiply(CodeGenFunction &CGF, const clang::Expr *Op1, |
1311 | WidthAndSignedness Op1Info, const clang::Expr *Op2, |
1312 | WidthAndSignedness Op2Info, |
1313 | const clang::Expr *ResultArg, QualType ResultQTy, |
1314 | WidthAndSignedness ResultInfo) { |
1315 | assert(isSpecialMixedSignMultiply(Builtin::BI__builtin_mul_overflow, Op1Info,((isSpecialMixedSignMultiply(Builtin::BI__builtin_mul_overflow , Op1Info, Op2Info, ResultInfo) && "Not a mixed-sign multipliction we can specialize" ) ? static_cast<void> (0) : __assert_fail ("isSpecialMixedSignMultiply(Builtin::BI__builtin_mul_overflow, Op1Info, Op2Info, ResultInfo) && \"Not a mixed-sign multipliction we can specialize\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 1317, __PRETTY_FUNCTION__)) |
1316 | Op2Info, ResultInfo) &&((isSpecialMixedSignMultiply(Builtin::BI__builtin_mul_overflow , Op1Info, Op2Info, ResultInfo) && "Not a mixed-sign multipliction we can specialize" ) ? static_cast<void> (0) : __assert_fail ("isSpecialMixedSignMultiply(Builtin::BI__builtin_mul_overflow, Op1Info, Op2Info, ResultInfo) && \"Not a mixed-sign multipliction we can specialize\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 1317, __PRETTY_FUNCTION__)) |
1317 | "Not a mixed-sign multipliction we can specialize")((isSpecialMixedSignMultiply(Builtin::BI__builtin_mul_overflow , Op1Info, Op2Info, ResultInfo) && "Not a mixed-sign multipliction we can specialize" ) ? static_cast<void> (0) : __assert_fail ("isSpecialMixedSignMultiply(Builtin::BI__builtin_mul_overflow, Op1Info, Op2Info, ResultInfo) && \"Not a mixed-sign multipliction we can specialize\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 1317, __PRETTY_FUNCTION__)); |
1318 | |
1319 | // Emit the signed and unsigned operands. |
1320 | const clang::Expr *SignedOp = Op1Info.Signed ? Op1 : Op2; |
1321 | const clang::Expr *UnsignedOp = Op1Info.Signed ? Op2 : Op1; |
1322 | llvm::Value *Signed = CGF.EmitScalarExpr(SignedOp); |
1323 | llvm::Value *Unsigned = CGF.EmitScalarExpr(UnsignedOp); |
1324 | unsigned SignedOpWidth = Op1Info.Signed ? Op1Info.Width : Op2Info.Width; |
1325 | unsigned UnsignedOpWidth = Op1Info.Signed ? Op2Info.Width : Op1Info.Width; |
1326 | |
1327 | // One of the operands may be smaller than the other. If so, [s|z]ext it. |
1328 | if (SignedOpWidth < UnsignedOpWidth) |
1329 | Signed = CGF.Builder.CreateSExt(Signed, Unsigned->getType(), "op.sext"); |
1330 | if (UnsignedOpWidth < SignedOpWidth) |
1331 | Unsigned = CGF.Builder.CreateZExt(Unsigned, Signed->getType(), "op.zext"); |
1332 | |
1333 | llvm::Type *OpTy = Signed->getType(); |
1334 | llvm::Value *Zero = llvm::Constant::getNullValue(OpTy); |
1335 | Address ResultPtr = CGF.EmitPointerWithAlignment(ResultArg); |
1336 | llvm::Type *ResTy = ResultPtr.getElementType(); |
1337 | unsigned OpWidth = std::max(Op1Info.Width, Op2Info.Width); |
1338 | |
1339 | // Take the absolute value of the signed operand. |
1340 | llvm::Value *IsNegative = CGF.Builder.CreateICmpSLT(Signed, Zero); |
1341 | llvm::Value *AbsOfNegative = CGF.Builder.CreateSub(Zero, Signed); |
1342 | llvm::Value *AbsSigned = |
1343 | CGF.Builder.CreateSelect(IsNegative, AbsOfNegative, Signed); |
1344 | |
1345 | // Perform a checked unsigned multiplication. |
1346 | llvm::Value *UnsignedOverflow; |
1347 | llvm::Value *UnsignedResult = |
1348 | EmitOverflowIntrinsic(CGF, llvm::Intrinsic::umul_with_overflow, AbsSigned, |
1349 | Unsigned, UnsignedOverflow); |
1350 | |
1351 | llvm::Value *Overflow, *Result; |
1352 | if (ResultInfo.Signed) { |
1353 | // Signed overflow occurs if the result is greater than INT_MAX or lesser |
1354 | // than INT_MIN, i.e when |Result| > (INT_MAX + IsNegative). |
1355 | auto IntMax = |
1356 | llvm::APInt::getSignedMaxValue(ResultInfo.Width).zextOrSelf(OpWidth); |
1357 | llvm::Value *MaxResult = |
1358 | CGF.Builder.CreateAdd(llvm::ConstantInt::get(OpTy, IntMax), |
1359 | CGF.Builder.CreateZExt(IsNegative, OpTy)); |
1360 | llvm::Value *SignedOverflow = |
1361 | CGF.Builder.CreateICmpUGT(UnsignedResult, MaxResult); |
1362 | Overflow = CGF.Builder.CreateOr(UnsignedOverflow, SignedOverflow); |
1363 | |
1364 | // Prepare the signed result (possibly by negating it). |
1365 | llvm::Value *NegativeResult = CGF.Builder.CreateNeg(UnsignedResult); |
1366 | llvm::Value *SignedResult = |
1367 | CGF.Builder.CreateSelect(IsNegative, NegativeResult, UnsignedResult); |
1368 | Result = CGF.Builder.CreateTrunc(SignedResult, ResTy); |
1369 | } else { |
1370 | // Unsigned overflow occurs if the result is < 0 or greater than UINT_MAX. |
1371 | llvm::Value *Underflow = CGF.Builder.CreateAnd( |
1372 | IsNegative, CGF.Builder.CreateIsNotNull(UnsignedResult)); |
1373 | Overflow = CGF.Builder.CreateOr(UnsignedOverflow, Underflow); |
1374 | if (ResultInfo.Width < OpWidth) { |
1375 | auto IntMax = |
1376 | llvm::APInt::getMaxValue(ResultInfo.Width).zext(OpWidth); |
1377 | llvm::Value *TruncOverflow = CGF.Builder.CreateICmpUGT( |
1378 | UnsignedResult, llvm::ConstantInt::get(OpTy, IntMax)); |
1379 | Overflow = CGF.Builder.CreateOr(Overflow, TruncOverflow); |
1380 | } |
1381 | |
1382 | // Negate the product if it would be negative in infinite precision. |
1383 | Result = CGF.Builder.CreateSelect( |
1384 | IsNegative, CGF.Builder.CreateNeg(UnsignedResult), UnsignedResult); |
1385 | |
1386 | Result = CGF.Builder.CreateTrunc(Result, ResTy); |
1387 | } |
1388 | assert(Overflow && Result && "Missing overflow or result")((Overflow && Result && "Missing overflow or result" ) ? static_cast<void> (0) : __assert_fail ("Overflow && Result && \"Missing overflow or result\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 1388, __PRETTY_FUNCTION__)); |
1389 | |
1390 | bool isVolatile = |
1391 | ResultArg->getType()->getPointeeType().isVolatileQualified(); |
1392 | CGF.Builder.CreateStore(CGF.EmitToMemory(Result, ResultQTy), ResultPtr, |
1393 | isVolatile); |
1394 | return RValue::get(Overflow); |
1395 | } |
1396 | |
1397 | static llvm::Value *dumpRecord(CodeGenFunction &CGF, QualType RType, |
1398 | Value *&RecordPtr, CharUnits Align, |
1399 | llvm::FunctionCallee Func, int Lvl) { |
1400 | const auto *RT = RType->getAs<RecordType>(); |
1401 | ASTContext &Context = CGF.getContext(); |
1402 | RecordDecl *RD = RT->getDecl()->getDefinition(); |
1403 | ASTContext &Ctx = RD->getASTContext(); |
1404 | const ASTRecordLayout &RL = Ctx.getASTRecordLayout(RD); |
1405 | std::string Pad = std::string(Lvl * 4, ' '); |
1406 | |
1407 | Value *GString = |
1408 | CGF.Builder.CreateGlobalStringPtr(RType.getAsString() + " {\n"); |
1409 | Value *Res = CGF.Builder.CreateCall(Func, {GString}); |
1410 | |
1411 | static llvm::DenseMap<QualType, const char *> Types; |
1412 | if (Types.empty()) { |
1413 | Types[Context.CharTy] = "%c"; |
1414 | Types[Context.BoolTy] = "%d"; |
1415 | Types[Context.SignedCharTy] = "%hhd"; |
1416 | Types[Context.UnsignedCharTy] = "%hhu"; |
1417 | Types[Context.IntTy] = "%d"; |
1418 | Types[Context.UnsignedIntTy] = "%u"; |
1419 | Types[Context.LongTy] = "%ld"; |
1420 | Types[Context.UnsignedLongTy] = "%lu"; |
1421 | Types[Context.LongLongTy] = "%lld"; |
1422 | Types[Context.UnsignedLongLongTy] = "%llu"; |
1423 | Types[Context.ShortTy] = "%hd"; |
1424 | Types[Context.UnsignedShortTy] = "%hu"; |
1425 | Types[Context.VoidPtrTy] = "%p"; |
1426 | Types[Context.FloatTy] = "%f"; |
1427 | Types[Context.DoubleTy] = "%f"; |
1428 | Types[Context.LongDoubleTy] = "%Lf"; |
1429 | Types[Context.getPointerType(Context.CharTy)] = "%s"; |
1430 | Types[Context.getPointerType(Context.getConstType(Context.CharTy))] = "%s"; |
1431 | } |
1432 | |
1433 | for (const auto *FD : RD->fields()) { |
1434 | uint64_t Off = RL.getFieldOffset(FD->getFieldIndex()); |
1435 | Off = Ctx.toCharUnitsFromBits(Off).getQuantity(); |
Value stored to 'Off' is never read | |
1436 | |
1437 | Value *FieldPtr = RecordPtr; |
1438 | if (RD->isUnion()) |
1439 | FieldPtr = CGF.Builder.CreatePointerCast( |
1440 | FieldPtr, CGF.ConvertType(Context.getPointerType(FD->getType()))); |
1441 | else |
1442 | FieldPtr = CGF.Builder.CreateStructGEP(CGF.ConvertType(RType), FieldPtr, |
1443 | FD->getFieldIndex()); |
1444 | |
1445 | GString = CGF.Builder.CreateGlobalStringPtr( |
1446 | llvm::Twine(Pad) |
1447 | .concat(FD->getType().getAsString()) |
1448 | .concat(llvm::Twine(' ')) |
1449 | .concat(FD->getNameAsString()) |
1450 | .concat(" : ") |
1451 | .str()); |
1452 | Value *TmpRes = CGF.Builder.CreateCall(Func, {GString}); |
1453 | Res = CGF.Builder.CreateAdd(Res, TmpRes); |
1454 | |
1455 | QualType CanonicalType = |
1456 | FD->getType().getUnqualifiedType().getCanonicalType(); |
1457 | |
1458 | // We check whether we are in a recursive type |
1459 | if (CanonicalType->isRecordType()) { |
1460 | Value *TmpRes = |
1461 | dumpRecord(CGF, CanonicalType, FieldPtr, Align, Func, Lvl + 1); |
1462 | Res = CGF.Builder.CreateAdd(TmpRes, Res); |
1463 | continue; |
1464 | } |
1465 | |
1466 | // We try to determine the best format to print the current field |
1467 | llvm::Twine Format = Types.find(CanonicalType) == Types.end() |
1468 | ? Types[Context.VoidPtrTy] |
1469 | : Types[CanonicalType]; |
1470 | |
1471 | Address FieldAddress = Address(FieldPtr, Align); |
1472 | FieldPtr = CGF.Builder.CreateLoad(FieldAddress); |
1473 | |
1474 | // FIXME Need to handle bitfield here |
1475 | GString = CGF.Builder.CreateGlobalStringPtr( |
1476 | Format.concat(llvm::Twine('\n')).str()); |
1477 | TmpRes = CGF.Builder.CreateCall(Func, {GString, FieldPtr}); |
1478 | Res = CGF.Builder.CreateAdd(Res, TmpRes); |
1479 | } |
1480 | |
1481 | GString = CGF.Builder.CreateGlobalStringPtr(Pad + "}\n"); |
1482 | Value *TmpRes = CGF.Builder.CreateCall(Func, {GString}); |
1483 | Res = CGF.Builder.CreateAdd(Res, TmpRes); |
1484 | return Res; |
1485 | } |
1486 | |
1487 | static bool |
1488 | TypeRequiresBuiltinLaunderImp(const ASTContext &Ctx, QualType Ty, |
1489 | llvm::SmallPtrSetImpl<const Decl *> &Seen) { |
1490 | if (const auto *Arr = Ctx.getAsArrayType(Ty)) |
1491 | Ty = Ctx.getBaseElementType(Arr); |
1492 | |
1493 | const auto *Record = Ty->getAsCXXRecordDecl(); |
1494 | if (!Record) |
1495 | return false; |
1496 | |
1497 | // We've already checked this type, or are in the process of checking it. |
1498 | if (!Seen.insert(Record).second) |
1499 | return false; |
1500 | |
1501 | assert(Record->hasDefinition() &&((Record->hasDefinition() && "Incomplete types should already be diagnosed" ) ? static_cast<void> (0) : __assert_fail ("Record->hasDefinition() && \"Incomplete types should already be diagnosed\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 1502, __PRETTY_FUNCTION__)) |
1502 | "Incomplete types should already be diagnosed")((Record->hasDefinition() && "Incomplete types should already be diagnosed" ) ? static_cast<void> (0) : __assert_fail ("Record->hasDefinition() && \"Incomplete types should already be diagnosed\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 1502, __PRETTY_FUNCTION__)); |
1503 | |
1504 | if (Record->isDynamicClass()) |
1505 | return true; |
1506 | |
1507 | for (FieldDecl *F : Record->fields()) { |
1508 | if (TypeRequiresBuiltinLaunderImp(Ctx, F->getType(), Seen)) |
1509 | return true; |
1510 | } |
1511 | return false; |
1512 | } |
1513 | |
1514 | /// Determine if the specified type requires laundering by checking if it is a |
1515 | /// dynamic class type or contains a subobject which is a dynamic class type. |
1516 | static bool TypeRequiresBuiltinLaunder(CodeGenModule &CGM, QualType Ty) { |
1517 | if (!CGM.getCodeGenOpts().StrictVTablePointers) |
1518 | return false; |
1519 | llvm::SmallPtrSet<const Decl *, 16> Seen; |
1520 | return TypeRequiresBuiltinLaunderImp(CGM.getContext(), Ty, Seen); |
1521 | } |
1522 | |
1523 | RValue CodeGenFunction::emitRotate(const CallExpr *E, bool IsRotateRight) { |
1524 | llvm::Value *Src = EmitScalarExpr(E->getArg(0)); |
1525 | llvm::Value *ShiftAmt = EmitScalarExpr(E->getArg(1)); |
1526 | |
1527 | // The builtin's shift arg may have a different type than the source arg and |
1528 | // result, but the LLVM intrinsic uses the same type for all values. |
1529 | llvm::Type *Ty = Src->getType(); |
1530 | ShiftAmt = Builder.CreateIntCast(ShiftAmt, Ty, false); |
1531 | |
1532 | // Rotate is a special case of LLVM funnel shift - 1st 2 args are the same. |
1533 | unsigned IID = IsRotateRight ? Intrinsic::fshr : Intrinsic::fshl; |
1534 | Function *F = CGM.getIntrinsic(IID, Ty); |
1535 | return RValue::get(Builder.CreateCall(F, { Src, Src, ShiftAmt })); |
1536 | } |
1537 | |
1538 | RValue CodeGenFunction::EmitBuiltinExpr(const GlobalDecl GD, unsigned BuiltinID, |
1539 | const CallExpr *E, |
1540 | ReturnValueSlot ReturnValue) { |
1541 | const FunctionDecl *FD = GD.getDecl()->getAsFunction(); |
1542 | // See if we can constant fold this builtin. If so, don't emit it at all. |
1543 | Expr::EvalResult Result; |
1544 | if (E->EvaluateAsRValue(Result, CGM.getContext()) && |
1545 | !Result.hasSideEffects()) { |
1546 | if (Result.Val.isInt()) |
1547 | return RValue::get(llvm::ConstantInt::get(getLLVMContext(), |
1548 | Result.Val.getInt())); |
1549 | if (Result.Val.isFloat()) |
1550 | return RValue::get(llvm::ConstantFP::get(getLLVMContext(), |
1551 | Result.Val.getFloat())); |
1552 | } |
1553 | |
1554 | // There are LLVM math intrinsics/instructions corresponding to math library |
1555 | // functions except the LLVM op will never set errno while the math library |
1556 | // might. Also, math builtins have the same semantics as their math library |
1557 | // twins. Thus, we can transform math library and builtin calls to their |
1558 | // LLVM counterparts if the call is marked 'const' (known to never set errno). |
1559 | if (FD->hasAttr<ConstAttr>()) { |
1560 | switch (BuiltinID) { |
1561 | case Builtin::BIceil: |
1562 | case Builtin::BIceilf: |
1563 | case Builtin::BIceill: |
1564 | case Builtin::BI__builtin_ceil: |
1565 | case Builtin::BI__builtin_ceilf: |
1566 | case Builtin::BI__builtin_ceill: |
1567 | return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::ceil)); |
1568 | |
1569 | case Builtin::BIcopysign: |
1570 | case Builtin::BIcopysignf: |
1571 | case Builtin::BIcopysignl: |
1572 | case Builtin::BI__builtin_copysign: |
1573 | case Builtin::BI__builtin_copysignf: |
1574 | case Builtin::BI__builtin_copysignl: |
1575 | case Builtin::BI__builtin_copysignf128: |
1576 | return RValue::get(emitBinaryBuiltin(*this, E, Intrinsic::copysign)); |
1577 | |
1578 | case Builtin::BIcos: |
1579 | case Builtin::BIcosf: |
1580 | case Builtin::BIcosl: |
1581 | case Builtin::BI__builtin_cos: |
1582 | case Builtin::BI__builtin_cosf: |
1583 | case Builtin::BI__builtin_cosl: |
1584 | return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::cos)); |
1585 | |
1586 | case Builtin::BIexp: |
1587 | case Builtin::BIexpf: |
1588 | case Builtin::BIexpl: |
1589 | case Builtin::BI__builtin_exp: |
1590 | case Builtin::BI__builtin_expf: |
1591 | case Builtin::BI__builtin_expl: |
1592 | return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::exp)); |
1593 | |
1594 | case Builtin::BIexp2: |
1595 | case Builtin::BIexp2f: |
1596 | case Builtin::BIexp2l: |
1597 | case Builtin::BI__builtin_exp2: |
1598 | case Builtin::BI__builtin_exp2f: |
1599 | case Builtin::BI__builtin_exp2l: |
1600 | return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::exp2)); |
1601 | |
1602 | case Builtin::BIfabs: |
1603 | case Builtin::BIfabsf: |
1604 | case Builtin::BIfabsl: |
1605 | case Builtin::BI__builtin_fabs: |
1606 | case Builtin::BI__builtin_fabsf: |
1607 | case Builtin::BI__builtin_fabsl: |
1608 | case Builtin::BI__builtin_fabsf128: |
1609 | return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::fabs)); |
1610 | |
1611 | case Builtin::BIfloor: |
1612 | case Builtin::BIfloorf: |
1613 | case Builtin::BIfloorl: |
1614 | case Builtin::BI__builtin_floor: |
1615 | case Builtin::BI__builtin_floorf: |
1616 | case Builtin::BI__builtin_floorl: |
1617 | return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::floor)); |
1618 | |
1619 | case Builtin::BIfma: |
1620 | case Builtin::BIfmaf: |
1621 | case Builtin::BIfmal: |
1622 | case Builtin::BI__builtin_fma: |
1623 | case Builtin::BI__builtin_fmaf: |
1624 | case Builtin::BI__builtin_fmal: |
1625 | return RValue::get(emitTernaryBuiltin(*this, E, Intrinsic::fma)); |
1626 | |
1627 | case Builtin::BIfmax: |
1628 | case Builtin::BIfmaxf: |
1629 | case Builtin::BIfmaxl: |
1630 | case Builtin::BI__builtin_fmax: |
1631 | case Builtin::BI__builtin_fmaxf: |
1632 | case Builtin::BI__builtin_fmaxl: |
1633 | return RValue::get(emitBinaryBuiltin(*this, E, Intrinsic::maxnum)); |
1634 | |
1635 | case Builtin::BIfmin: |
1636 | case Builtin::BIfminf: |
1637 | case Builtin::BIfminl: |
1638 | case Builtin::BI__builtin_fmin: |
1639 | case Builtin::BI__builtin_fminf: |
1640 | case Builtin::BI__builtin_fminl: |
1641 | return RValue::get(emitBinaryBuiltin(*this, E, Intrinsic::minnum)); |
1642 | |
1643 | // fmod() is a special-case. It maps to the frem instruction rather than an |
1644 | // LLVM intrinsic. |
1645 | case Builtin::BIfmod: |
1646 | case Builtin::BIfmodf: |
1647 | case Builtin::BIfmodl: |
1648 | case Builtin::BI__builtin_fmod: |
1649 | case Builtin::BI__builtin_fmodf: |
1650 | case Builtin::BI__builtin_fmodl: { |
1651 | Value *Arg1 = EmitScalarExpr(E->getArg(0)); |
1652 | Value *Arg2 = EmitScalarExpr(E->getArg(1)); |
1653 | return RValue::get(Builder.CreateFRem(Arg1, Arg2, "fmod")); |
1654 | } |
1655 | |
1656 | case Builtin::BIlog: |
1657 | case Builtin::BIlogf: |
1658 | case Builtin::BIlogl: |
1659 | case Builtin::BI__builtin_log: |
1660 | case Builtin::BI__builtin_logf: |
1661 | case Builtin::BI__builtin_logl: |
1662 | return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::log)); |
1663 | |
1664 | case Builtin::BIlog10: |
1665 | case Builtin::BIlog10f: |
1666 | case Builtin::BIlog10l: |
1667 | case Builtin::BI__builtin_log10: |
1668 | case Builtin::BI__builtin_log10f: |
1669 | case Builtin::BI__builtin_log10l: |
1670 | return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::log10)); |
1671 | |
1672 | case Builtin::BIlog2: |
1673 | case Builtin::BIlog2f: |
1674 | case Builtin::BIlog2l: |
1675 | case Builtin::BI__builtin_log2: |
1676 | case Builtin::BI__builtin_log2f: |
1677 | case Builtin::BI__builtin_log2l: |
1678 | return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::log2)); |
1679 | |
1680 | case Builtin::BInearbyint: |
1681 | case Builtin::BInearbyintf: |
1682 | case Builtin::BInearbyintl: |
1683 | case Builtin::BI__builtin_nearbyint: |
1684 | case Builtin::BI__builtin_nearbyintf: |
1685 | case Builtin::BI__builtin_nearbyintl: |
1686 | return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::nearbyint)); |
1687 | |
1688 | case Builtin::BIpow: |
1689 | case Builtin::BIpowf: |
1690 | case Builtin::BIpowl: |
1691 | case Builtin::BI__builtin_pow: |
1692 | case Builtin::BI__builtin_powf: |
1693 | case Builtin::BI__builtin_powl: |
1694 | return RValue::get(emitBinaryBuiltin(*this, E, Intrinsic::pow)); |
1695 | |
1696 | case Builtin::BIrint: |
1697 | case Builtin::BIrintf: |
1698 | case Builtin::BIrintl: |
1699 | case Builtin::BI__builtin_rint: |
1700 | case Builtin::BI__builtin_rintf: |
1701 | case Builtin::BI__builtin_rintl: |
1702 | return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::rint)); |
1703 | |
1704 | case Builtin::BIround: |
1705 | case Builtin::BIroundf: |
1706 | case Builtin::BIroundl: |
1707 | case Builtin::BI__builtin_round: |
1708 | case Builtin::BI__builtin_roundf: |
1709 | case Builtin::BI__builtin_roundl: |
1710 | return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::round)); |
1711 | |
1712 | case Builtin::BIsin: |
1713 | case Builtin::BIsinf: |
1714 | case Builtin::BIsinl: |
1715 | case Builtin::BI__builtin_sin: |
1716 | case Builtin::BI__builtin_sinf: |
1717 | case Builtin::BI__builtin_sinl: |
1718 | return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::sin)); |
1719 | |
1720 | case Builtin::BIsqrt: |
1721 | case Builtin::BIsqrtf: |
1722 | case Builtin::BIsqrtl: |
1723 | case Builtin::BI__builtin_sqrt: |
1724 | case Builtin::BI__builtin_sqrtf: |
1725 | case Builtin::BI__builtin_sqrtl: |
1726 | return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::sqrt)); |
1727 | |
1728 | case Builtin::BItrunc: |
1729 | case Builtin::BItruncf: |
1730 | case Builtin::BItruncl: |
1731 | case Builtin::BI__builtin_trunc: |
1732 | case Builtin::BI__builtin_truncf: |
1733 | case Builtin::BI__builtin_truncl: |
1734 | return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::trunc)); |
1735 | |
1736 | case Builtin::BIlround: |
1737 | case Builtin::BIlroundf: |
1738 | case Builtin::BIlroundl: |
1739 | case Builtin::BI__builtin_lround: |
1740 | case Builtin::BI__builtin_lroundf: |
1741 | case Builtin::BI__builtin_lroundl: |
1742 | return RValue::get(emitFPToIntRoundBuiltin(*this, E, Intrinsic::lround)); |
1743 | |
1744 | case Builtin::BIllround: |
1745 | case Builtin::BIllroundf: |
1746 | case Builtin::BIllroundl: |
1747 | case Builtin::BI__builtin_llround: |
1748 | case Builtin::BI__builtin_llroundf: |
1749 | case Builtin::BI__builtin_llroundl: |
1750 | return RValue::get(emitFPToIntRoundBuiltin(*this, E, Intrinsic::llround)); |
1751 | |
1752 | default: |
1753 | break; |
1754 | } |
1755 | } |
1756 | |
1757 | switch (BuiltinID) { |
1758 | default: break; |
1759 | case Builtin::BI__builtin___CFStringMakeConstantString: |
1760 | case Builtin::BI__builtin___NSStringMakeConstantString: |
1761 | return RValue::get(ConstantEmitter(*this).emitAbstract(E, E->getType())); |
1762 | case Builtin::BI__builtin_stdarg_start: |
1763 | case Builtin::BI__builtin_va_start: |
1764 | case Builtin::BI__va_start: |
1765 | case Builtin::BI__builtin_va_end: |
1766 | return RValue::get( |
1767 | EmitVAStartEnd(BuiltinID == Builtin::BI__va_start |
1768 | ? EmitScalarExpr(E->getArg(0)) |
1769 | : EmitVAListRef(E->getArg(0)).getPointer(), |
1770 | BuiltinID != Builtin::BI__builtin_va_end)); |
1771 | case Builtin::BI__builtin_va_copy: { |
1772 | Value *DstPtr = EmitVAListRef(E->getArg(0)).getPointer(); |
1773 | Value *SrcPtr = EmitVAListRef(E->getArg(1)).getPointer(); |
1774 | |
1775 | llvm::Type *Type = Int8PtrTy; |
1776 | |
1777 | DstPtr = Builder.CreateBitCast(DstPtr, Type); |
1778 | SrcPtr = Builder.CreateBitCast(SrcPtr, Type); |
1779 | return RValue::get(Builder.CreateCall(CGM.getIntrinsic(Intrinsic::vacopy), |
1780 | {DstPtr, SrcPtr})); |
1781 | } |
1782 | case Builtin::BI__builtin_abs: |
1783 | case Builtin::BI__builtin_labs: |
1784 | case Builtin::BI__builtin_llabs: { |
1785 | // X < 0 ? -X : X |
1786 | // The negation has 'nsw' because abs of INT_MIN is undefined. |
1787 | Value *ArgValue = EmitScalarExpr(E->getArg(0)); |
1788 | Value *NegOp = Builder.CreateNSWNeg(ArgValue, "neg"); |
1789 | Constant *Zero = llvm::Constant::getNullValue(ArgValue->getType()); |
1790 | Value *CmpResult = Builder.CreateICmpSLT(ArgValue, Zero, "abscond"); |
1791 | Value *Result = Builder.CreateSelect(CmpResult, NegOp, ArgValue, "abs"); |
1792 | return RValue::get(Result); |
1793 | } |
1794 | case Builtin::BI__builtin_conj: |
1795 | case Builtin::BI__builtin_conjf: |
1796 | case Builtin::BI__builtin_conjl: { |
1797 | ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); |
1798 | Value *Real = ComplexVal.first; |
1799 | Value *Imag = ComplexVal.second; |
1800 | Value *Zero = |
1801 | Imag->getType()->isFPOrFPVectorTy() |
1802 | ? llvm::ConstantFP::getZeroValueForNegation(Imag->getType()) |
1803 | : llvm::Constant::getNullValue(Imag->getType()); |
1804 | |
1805 | Imag = Builder.CreateFSub(Zero, Imag, "sub"); |
1806 | return RValue::getComplex(std::make_pair(Real, Imag)); |
1807 | } |
1808 | case Builtin::BI__builtin_creal: |
1809 | case Builtin::BI__builtin_crealf: |
1810 | case Builtin::BI__builtin_creall: |
1811 | case Builtin::BIcreal: |
1812 | case Builtin::BIcrealf: |
1813 | case Builtin::BIcreall: { |
1814 | ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); |
1815 | return RValue::get(ComplexVal.first); |
1816 | } |
1817 | |
1818 | case Builtin::BI__builtin_dump_struct: { |
1819 | llvm::Type *LLVMIntTy = getTypes().ConvertType(getContext().IntTy); |
1820 | llvm::FunctionType *LLVMFuncType = llvm::FunctionType::get( |
1821 | LLVMIntTy, {llvm::Type::getInt8PtrTy(getLLVMContext())}, true); |
1822 | |
1823 | Value *Func = EmitScalarExpr(E->getArg(1)->IgnoreImpCasts()); |
1824 | CharUnits Arg0Align = EmitPointerWithAlignment(E->getArg(0)).getAlignment(); |
1825 | |
1826 | const Expr *Arg0 = E->getArg(0)->IgnoreImpCasts(); |
1827 | QualType Arg0Type = Arg0->getType()->getPointeeType(); |
1828 | |
1829 | Value *RecordPtr = EmitScalarExpr(Arg0); |
1830 | Value *Res = dumpRecord(*this, Arg0Type, RecordPtr, Arg0Align, |
1831 | {LLVMFuncType, Func}, 0); |
1832 | return RValue::get(Res); |
1833 | } |
1834 | |
1835 | case Builtin::BI__builtin_cimag: |
1836 | case Builtin::BI__builtin_cimagf: |
1837 | case Builtin::BI__builtin_cimagl: |
1838 | case Builtin::BIcimag: |
1839 | case Builtin::BIcimagf: |
1840 | case Builtin::BIcimagl: { |
1841 | ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); |
1842 | return RValue::get(ComplexVal.second); |
1843 | } |
1844 | |
1845 | case Builtin::BI__builtin_clrsb: |
1846 | case Builtin::BI__builtin_clrsbl: |
1847 | case Builtin::BI__builtin_clrsbll: { |
1848 | // clrsb(x) -> clz(x < 0 ? ~x : x) - 1 or |
1849 | Value *ArgValue = EmitScalarExpr(E->getArg(0)); |
1850 | |
1851 | llvm::Type *ArgType = ArgValue->getType(); |
1852 | Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType); |
1853 | |
1854 | llvm::Type *ResultType = ConvertType(E->getType()); |
1855 | Value *Zero = llvm::Constant::getNullValue(ArgType); |
1856 | Value *IsNeg = Builder.CreateICmpSLT(ArgValue, Zero, "isneg"); |
1857 | Value *Inverse = Builder.CreateNot(ArgValue, "not"); |
1858 | Value *Tmp = Builder.CreateSelect(IsNeg, Inverse, ArgValue); |
1859 | Value *Ctlz = Builder.CreateCall(F, {Tmp, Builder.getFalse()}); |
1860 | Value *Result = Builder.CreateSub(Ctlz, llvm::ConstantInt::get(ArgType, 1)); |
1861 | Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, |
1862 | "cast"); |
1863 | return RValue::get(Result); |
1864 | } |
1865 | case Builtin::BI__builtin_ctzs: |
1866 | case Builtin::BI__builtin_ctz: |
1867 | case Builtin::BI__builtin_ctzl: |
1868 | case Builtin::BI__builtin_ctzll: { |
1869 | Value *ArgValue = EmitCheckedArgForBuiltin(E->getArg(0), BCK_CTZPassedZero); |
1870 | |
1871 | llvm::Type *ArgType = ArgValue->getType(); |
1872 | Function *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); |
1873 | |
1874 | llvm::Type *ResultType = ConvertType(E->getType()); |
1875 | Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef()); |
1876 | Value *Result = Builder.CreateCall(F, {ArgValue, ZeroUndef}); |
1877 | if (Result->getType() != ResultType) |
1878 | Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, |
1879 | "cast"); |
1880 | return RValue::get(Result); |
1881 | } |
1882 | case Builtin::BI__builtin_clzs: |
1883 | case Builtin::BI__builtin_clz: |
1884 | case Builtin::BI__builtin_clzl: |
1885 | case Builtin::BI__builtin_clzll: { |
1886 | Value *ArgValue = EmitCheckedArgForBuiltin(E->getArg(0), BCK_CLZPassedZero); |
1887 | |
1888 | llvm::Type *ArgType = ArgValue->getType(); |
1889 | Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType); |
1890 | |
1891 | llvm::Type *ResultType = ConvertType(E->getType()); |
1892 | Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef()); |
1893 | Value *Result = Builder.CreateCall(F, {ArgValue, ZeroUndef}); |
1894 | if (Result->getType() != ResultType) |
1895 | Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, |
1896 | "cast"); |
1897 | return RValue::get(Result); |
1898 | } |
1899 | case Builtin::BI__builtin_ffs: |
1900 | case Builtin::BI__builtin_ffsl: |
1901 | case Builtin::BI__builtin_ffsll: { |
1902 | // ffs(x) -> x ? cttz(x) + 1 : 0 |
1903 | Value *ArgValue = EmitScalarExpr(E->getArg(0)); |
1904 | |
1905 | llvm::Type *ArgType = ArgValue->getType(); |
1906 | Function *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); |
1907 | |
1908 | llvm::Type *ResultType = ConvertType(E->getType()); |
1909 | Value *Tmp = |
1910 | Builder.CreateAdd(Builder.CreateCall(F, {ArgValue, Builder.getTrue()}), |
1911 | llvm::ConstantInt::get(ArgType, 1)); |
1912 | Value *Zero = llvm::Constant::getNullValue(ArgType); |
1913 | Value *IsZero = Builder.CreateICmpEQ(ArgValue, Zero, "iszero"); |
1914 | Value *Result = Builder.CreateSelect(IsZero, Zero, Tmp, "ffs"); |
1915 | if (Result->getType() != ResultType) |
1916 | Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, |
1917 | "cast"); |
1918 | return RValue::get(Result); |
1919 | } |
1920 | case Builtin::BI__builtin_parity: |
1921 | case Builtin::BI__builtin_parityl: |
1922 | case Builtin::BI__builtin_parityll: { |
1923 | // parity(x) -> ctpop(x) & 1 |
1924 | Value *ArgValue = EmitScalarExpr(E->getArg(0)); |
1925 | |
1926 | llvm::Type *ArgType = ArgValue->getType(); |
1927 | Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); |
1928 | |
1929 | llvm::Type *ResultType = ConvertType(E->getType()); |
1930 | Value *Tmp = Builder.CreateCall(F, ArgValue); |
1931 | Value *Result = Builder.CreateAnd(Tmp, llvm::ConstantInt::get(ArgType, 1)); |
1932 | if (Result->getType() != ResultType) |
1933 | Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, |
1934 | "cast"); |
1935 | return RValue::get(Result); |
1936 | } |
1937 | case Builtin::BI__lzcnt16: |
1938 | case Builtin::BI__lzcnt: |
1939 | case Builtin::BI__lzcnt64: { |
1940 | Value *ArgValue = EmitScalarExpr(E->getArg(0)); |
1941 | |
1942 | llvm::Type *ArgType = ArgValue->getType(); |
1943 | Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType); |
1944 | |
1945 | llvm::Type *ResultType = ConvertType(E->getType()); |
1946 | Value *Result = Builder.CreateCall(F, {ArgValue, Builder.getFalse()}); |
1947 | if (Result->getType() != ResultType) |
1948 | Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, |
1949 | "cast"); |
1950 | return RValue::get(Result); |
1951 | } |
1952 | case Builtin::BI__popcnt16: |
1953 | case Builtin::BI__popcnt: |
1954 | case Builtin::BI__popcnt64: |
1955 | case Builtin::BI__builtin_popcount: |
1956 | case Builtin::BI__builtin_popcountl: |
1957 | case Builtin::BI__builtin_popcountll: { |
1958 | Value *ArgValue = EmitScalarExpr(E->getArg(0)); |
1959 | |
1960 | llvm::Type *ArgType = ArgValue->getType(); |
1961 | Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); |
1962 | |
1963 | llvm::Type *ResultType = ConvertType(E->getType()); |
1964 | Value *Result = Builder.CreateCall(F, ArgValue); |
1965 | if (Result->getType() != ResultType) |
1966 | Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, |
1967 | "cast"); |
1968 | return RValue::get(Result); |
1969 | } |
1970 | case Builtin::BI__builtin_unpredictable: { |
1971 | // Always return the argument of __builtin_unpredictable. LLVM does not |
1972 | // handle this builtin. Metadata for this builtin should be added directly |
1973 | // to instructions such as branches or switches that use it. |
1974 | return RValue::get(EmitScalarExpr(E->getArg(0))); |
1975 | } |
1976 | case Builtin::BI__builtin_expect: { |
1977 | Value *ArgValue = EmitScalarExpr(E->getArg(0)); |
1978 | llvm::Type *ArgType = ArgValue->getType(); |
1979 | |
1980 | Value *ExpectedValue = EmitScalarExpr(E->getArg(1)); |
1981 | // Don't generate llvm.expect on -O0 as the backend won't use it for |
1982 | // anything. |
1983 | // Note, we still IRGen ExpectedValue because it could have side-effects. |
1984 | if (CGM.getCodeGenOpts().OptimizationLevel == 0) |
1985 | return RValue::get(ArgValue); |
1986 | |
1987 | Function *FnExpect = CGM.getIntrinsic(Intrinsic::expect, ArgType); |
1988 | Value *Result = |
1989 | Builder.CreateCall(FnExpect, {ArgValue, ExpectedValue}, "expval"); |
1990 | return RValue::get(Result); |
1991 | } |
1992 | case Builtin::BI__builtin_assume_aligned: { |
1993 | const Expr *Ptr = E->getArg(0); |
1994 | Value *PtrValue = EmitScalarExpr(Ptr); |
1995 | Value *OffsetValue = |
1996 | (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : nullptr; |
1997 | |
1998 | Value *AlignmentValue = EmitScalarExpr(E->getArg(1)); |
1999 | ConstantInt *AlignmentCI = cast<ConstantInt>(AlignmentValue); |
2000 | unsigned Alignment = (unsigned)AlignmentCI->getZExtValue(); |
2001 | |
2002 | EmitAlignmentAssumption(PtrValue, Ptr, |
2003 | /*The expr loc is sufficient.*/ SourceLocation(), |
2004 | Alignment, OffsetValue); |
2005 | return RValue::get(PtrValue); |
2006 | } |
2007 | case Builtin::BI__assume: |
2008 | case Builtin::BI__builtin_assume: { |
2009 | if (E->getArg(0)->HasSideEffects(getContext())) |
2010 | return RValue::get(nullptr); |
2011 | |
2012 | Value *ArgValue = EmitScalarExpr(E->getArg(0)); |
2013 | Function *FnAssume = CGM.getIntrinsic(Intrinsic::assume); |
2014 | return RValue::get(Builder.CreateCall(FnAssume, ArgValue)); |
2015 | } |
2016 | case Builtin::BI__builtin_bswap16: |
2017 | case Builtin::BI__builtin_bswap32: |
2018 | case Builtin::BI__builtin_bswap64: { |
2019 | return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::bswap)); |
2020 | } |
2021 | case Builtin::BI__builtin_bitreverse8: |
2022 | case Builtin::BI__builtin_bitreverse16: |
2023 | case Builtin::BI__builtin_bitreverse32: |
2024 | case Builtin::BI__builtin_bitreverse64: { |
2025 | return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::bitreverse)); |
2026 | } |
2027 | case Builtin::BI__builtin_rotateleft8: |
2028 | case Builtin::BI__builtin_rotateleft16: |
2029 | case Builtin::BI__builtin_rotateleft32: |
2030 | case Builtin::BI__builtin_rotateleft64: |
2031 | case Builtin::BI_rotl8: // Microsoft variants of rotate left |
2032 | case Builtin::BI_rotl16: |
2033 | case Builtin::BI_rotl: |
2034 | case Builtin::BI_lrotl: |
2035 | case Builtin::BI_rotl64: |
2036 | return emitRotate(E, false); |
2037 | |
2038 | case Builtin::BI__builtin_rotateright8: |
2039 | case Builtin::BI__builtin_rotateright16: |
2040 | case Builtin::BI__builtin_rotateright32: |
2041 | case Builtin::BI__builtin_rotateright64: |
2042 | case Builtin::BI_rotr8: // Microsoft variants of rotate right |
2043 | case Builtin::BI_rotr16: |
2044 | case Builtin::BI_rotr: |
2045 | case Builtin::BI_lrotr: |
2046 | case Builtin::BI_rotr64: |
2047 | return emitRotate(E, true); |
2048 | |
2049 | case Builtin::BI__builtin_constant_p: { |
2050 | llvm::Type *ResultType = ConvertType(E->getType()); |
2051 | if (CGM.getCodeGenOpts().OptimizationLevel == 0) |
2052 | // At -O0, we don't perform inlining, so we don't need to delay the |
2053 | // processing. |
2054 | return RValue::get(ConstantInt::get(ResultType, 0)); |
2055 | |
2056 | const Expr *Arg = E->getArg(0); |
2057 | QualType ArgType = Arg->getType(); |
2058 | // FIXME: The allowance for Obj-C pointers and block pointers is historical |
2059 | // and likely a mistake. |
2060 | if (!ArgType->isIntegralOrEnumerationType() && !ArgType->isFloatingType() && |
2061 | !ArgType->isObjCObjectPointerType() && !ArgType->isBlockPointerType()) |
2062 | // Per the GCC documentation, only numeric constants are recognized after |
2063 | // inlining. |
2064 | return RValue::get(ConstantInt::get(ResultType, 0)); |
2065 | |
2066 | if (Arg->HasSideEffects(getContext())) |
2067 | // The argument is unevaluated, so be conservative if it might have |
2068 | // side-effects. |
2069 | return RValue::get(ConstantInt::get(ResultType, 0)); |
2070 | |
2071 | Value *ArgValue = EmitScalarExpr(Arg); |
2072 | if (ArgType->isObjCObjectPointerType()) { |
2073 | // Convert Objective-C objects to id because we cannot distinguish between |
2074 | // LLVM types for Obj-C classes as they are opaque. |
2075 | ArgType = CGM.getContext().getObjCIdType(); |
2076 | ArgValue = Builder.CreateBitCast(ArgValue, ConvertType(ArgType)); |
2077 | } |
2078 | Function *F = |
2079 | CGM.getIntrinsic(Intrinsic::is_constant, ConvertType(ArgType)); |
2080 | Value *Result = Builder.CreateCall(F, ArgValue); |
2081 | if (Result->getType() != ResultType) |
2082 | Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/false); |
2083 | return RValue::get(Result); |
2084 | } |
2085 | case Builtin::BI__builtin_dynamic_object_size: |
2086 | case Builtin::BI__builtin_object_size: { |
2087 | unsigned Type = |
2088 | E->getArg(1)->EvaluateKnownConstInt(getContext()).getZExtValue(); |
2089 | auto *ResType = cast<llvm::IntegerType>(ConvertType(E->getType())); |
2090 | |
2091 | // We pass this builtin onto the optimizer so that it can figure out the |
2092 | // object size in more complex cases. |
2093 | bool IsDynamic = BuiltinID == Builtin::BI__builtin_dynamic_object_size; |
2094 | return RValue::get(emitBuiltinObjectSize(E->getArg(0), Type, ResType, |
2095 | /*EmittedE=*/nullptr, IsDynamic)); |
2096 | } |
2097 | case Builtin::BI__builtin_prefetch: { |
2098 | Value *Locality, *RW, *Address = EmitScalarExpr(E->getArg(0)); |
2099 | // FIXME: Technically these constants should of type 'int', yes? |
2100 | RW = (E->getNumArgs() > 1) ? EmitScalarExpr(E->getArg(1)) : |
2101 | llvm::ConstantInt::get(Int32Ty, 0); |
2102 | Locality = (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : |
2103 | llvm::ConstantInt::get(Int32Ty, 3); |
2104 | Value *Data = llvm::ConstantInt::get(Int32Ty, 1); |
2105 | Function *F = CGM.getIntrinsic(Intrinsic::prefetch); |
2106 | return RValue::get(Builder.CreateCall(F, {Address, RW, Locality, Data})); |
2107 | } |
2108 | case Builtin::BI__builtin_readcyclecounter: { |
2109 | Function *F = CGM.getIntrinsic(Intrinsic::readcyclecounter); |
2110 | return RValue::get(Builder.CreateCall(F)); |
2111 | } |
2112 | case Builtin::BI__builtin___clear_cache: { |
2113 | Value *Begin = EmitScalarExpr(E->getArg(0)); |
2114 | Value *End = EmitScalarExpr(E->getArg(1)); |
2115 | Function *F = CGM.getIntrinsic(Intrinsic::clear_cache); |
2116 | return RValue::get(Builder.CreateCall(F, {Begin, End})); |
2117 | } |
2118 | case Builtin::BI__builtin_trap: |
2119 | return RValue::get(EmitTrapCall(Intrinsic::trap)); |
2120 | case Builtin::BI__debugbreak: |
2121 | return RValue::get(EmitTrapCall(Intrinsic::debugtrap)); |
2122 | case Builtin::BI__builtin_unreachable: { |
2123 | EmitUnreachable(E->getExprLoc()); |
2124 | |
2125 | // We do need to preserve an insertion point. |
2126 | EmitBlock(createBasicBlock("unreachable.cont")); |
2127 | |
2128 | return RValue::get(nullptr); |
2129 | } |
2130 | |
2131 | case Builtin::BI__builtin_powi: |
2132 | case Builtin::BI__builtin_powif: |
2133 | case Builtin::BI__builtin_powil: { |
2134 | Value *Base = EmitScalarExpr(E->getArg(0)); |
2135 | Value *Exponent = EmitScalarExpr(E->getArg(1)); |
2136 | llvm::Type *ArgType = Base->getType(); |
2137 | Function *F = CGM.getIntrinsic(Intrinsic::powi, ArgType); |
2138 | return RValue::get(Builder.CreateCall(F, {Base, Exponent})); |
2139 | } |
2140 | |
2141 | case Builtin::BI__builtin_isgreater: |
2142 | case Builtin::BI__builtin_isgreaterequal: |
2143 | case Builtin::BI__builtin_isless: |
2144 | case Builtin::BI__builtin_islessequal: |
2145 | case Builtin::BI__builtin_islessgreater: |
2146 | case Builtin::BI__builtin_isunordered: { |
2147 | // Ordered comparisons: we know the arguments to these are matching scalar |
2148 | // floating point values. |
2149 | Value *LHS = EmitScalarExpr(E->getArg(0)); |
2150 | Value *RHS = EmitScalarExpr(E->getArg(1)); |
2151 | |
2152 | switch (BuiltinID) { |
2153 | default: llvm_unreachable("Unknown ordered comparison")::llvm::llvm_unreachable_internal("Unknown ordered comparison" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 2153); |
2154 | case Builtin::BI__builtin_isgreater: |
2155 | LHS = Builder.CreateFCmpOGT(LHS, RHS, "cmp"); |
2156 | break; |
2157 | case Builtin::BI__builtin_isgreaterequal: |
2158 | LHS = Builder.CreateFCmpOGE(LHS, RHS, "cmp"); |
2159 | break; |
2160 | case Builtin::BI__builtin_isless: |
2161 | LHS = Builder.CreateFCmpOLT(LHS, RHS, "cmp"); |
2162 | break; |
2163 | case Builtin::BI__builtin_islessequal: |
2164 | LHS = Builder.CreateFCmpOLE(LHS, RHS, "cmp"); |
2165 | break; |
2166 | case Builtin::BI__builtin_islessgreater: |
2167 | LHS = Builder.CreateFCmpONE(LHS, RHS, "cmp"); |
2168 | break; |
2169 | case Builtin::BI__builtin_isunordered: |
2170 | LHS = Builder.CreateFCmpUNO(LHS, RHS, "cmp"); |
2171 | break; |
2172 | } |
2173 | // ZExt bool to int type. |
2174 | return RValue::get(Builder.CreateZExt(LHS, ConvertType(E->getType()))); |
2175 | } |
2176 | case Builtin::BI__builtin_isnan: { |
2177 | Value *V = EmitScalarExpr(E->getArg(0)); |
2178 | V = Builder.CreateFCmpUNO(V, V, "cmp"); |
2179 | return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); |
2180 | } |
2181 | |
2182 | case Builtin::BIfinite: |
2183 | case Builtin::BI__finite: |
2184 | case Builtin::BIfinitef: |
2185 | case Builtin::BI__finitef: |
2186 | case Builtin::BIfinitel: |
2187 | case Builtin::BI__finitel: |
2188 | case Builtin::BI__builtin_isinf: |
2189 | case Builtin::BI__builtin_isfinite: { |
2190 | // isinf(x) --> fabs(x) == infinity |
2191 | // isfinite(x) --> fabs(x) != infinity |
2192 | // x != NaN via the ordered compare in either case. |
2193 | Value *V = EmitScalarExpr(E->getArg(0)); |
2194 | Value *Fabs = EmitFAbs(*this, V); |
2195 | Constant *Infinity = ConstantFP::getInfinity(V->getType()); |
2196 | CmpInst::Predicate Pred = (BuiltinID == Builtin::BI__builtin_isinf) |
2197 | ? CmpInst::FCMP_OEQ |
2198 | : CmpInst::FCMP_ONE; |
2199 | Value *FCmp = Builder.CreateFCmp(Pred, Fabs, Infinity, "cmpinf"); |
2200 | return RValue::get(Builder.CreateZExt(FCmp, ConvertType(E->getType()))); |
2201 | } |
2202 | |
2203 | case Builtin::BI__builtin_isinf_sign: { |
2204 | // isinf_sign(x) -> fabs(x) == infinity ? (signbit(x) ? -1 : 1) : 0 |
2205 | Value *Arg = EmitScalarExpr(E->getArg(0)); |
2206 | Value *AbsArg = EmitFAbs(*this, Arg); |
2207 | Value *IsInf = Builder.CreateFCmpOEQ( |
2208 | AbsArg, ConstantFP::getInfinity(Arg->getType()), "isinf"); |
2209 | Value *IsNeg = EmitSignBit(*this, Arg); |
2210 | |
2211 | llvm::Type *IntTy = ConvertType(E->getType()); |
2212 | Value *Zero = Constant::getNullValue(IntTy); |
2213 | Value *One = ConstantInt::get(IntTy, 1); |
2214 | Value *NegativeOne = ConstantInt::get(IntTy, -1); |
2215 | Value *SignResult = Builder.CreateSelect(IsNeg, NegativeOne, One); |
2216 | Value *Result = Builder.CreateSelect(IsInf, SignResult, Zero); |
2217 | return RValue::get(Result); |
2218 | } |
2219 | |
2220 | case Builtin::BI__builtin_isnormal: { |
2221 | // isnormal(x) --> x == x && fabsf(x) < infinity && fabsf(x) >= float_min |
2222 | Value *V = EmitScalarExpr(E->getArg(0)); |
2223 | Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); |
2224 | |
2225 | Value *Abs = EmitFAbs(*this, V); |
2226 | Value *IsLessThanInf = |
2227 | Builder.CreateFCmpULT(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); |
2228 | APFloat Smallest = APFloat::getSmallestNormalized( |
2229 | getContext().getFloatTypeSemantics(E->getArg(0)->getType())); |
2230 | Value *IsNormal = |
2231 | Builder.CreateFCmpUGE(Abs, ConstantFP::get(V->getContext(), Smallest), |
2232 | "isnormal"); |
2233 | V = Builder.CreateAnd(Eq, IsLessThanInf, "and"); |
2234 | V = Builder.CreateAnd(V, IsNormal, "and"); |
2235 | return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); |
2236 | } |
2237 | |
2238 | case Builtin::BI__builtin_flt_rounds: { |
2239 | Function *F = CGM.getIntrinsic(Intrinsic::flt_rounds); |
2240 | |
2241 | llvm::Type *ResultType = ConvertType(E->getType()); |
2242 | Value *Result = Builder.CreateCall(F); |
2243 | if (Result->getType() != ResultType) |
2244 | Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, |
2245 | "cast"); |
2246 | return RValue::get(Result); |
2247 | } |
2248 | |
2249 | case Builtin::BI__builtin_fpclassify: { |
2250 | Value *V = EmitScalarExpr(E->getArg(5)); |
2251 | llvm::Type *Ty = ConvertType(E->getArg(5)->getType()); |
2252 | |
2253 | // Create Result |
2254 | BasicBlock *Begin = Builder.GetInsertBlock(); |
2255 | BasicBlock *End = createBasicBlock("fpclassify_end", this->CurFn); |
2256 | Builder.SetInsertPoint(End); |
2257 | PHINode *Result = |
2258 | Builder.CreatePHI(ConvertType(E->getArg(0)->getType()), 4, |
2259 | "fpclassify_result"); |
2260 | |
2261 | // if (V==0) return FP_ZERO |
2262 | Builder.SetInsertPoint(Begin); |
2263 | Value *IsZero = Builder.CreateFCmpOEQ(V, Constant::getNullValue(Ty), |
2264 | "iszero"); |
2265 | Value *ZeroLiteral = EmitScalarExpr(E->getArg(4)); |
2266 | BasicBlock *NotZero = createBasicBlock("fpclassify_not_zero", this->CurFn); |
2267 | Builder.CreateCondBr(IsZero, End, NotZero); |
2268 | Result->addIncoming(ZeroLiteral, Begin); |
2269 | |
2270 | // if (V != V) return FP_NAN |
2271 | Builder.SetInsertPoint(NotZero); |
2272 | Value *IsNan = Builder.CreateFCmpUNO(V, V, "cmp"); |
2273 | Value *NanLiteral = EmitScalarExpr(E->getArg(0)); |
2274 | BasicBlock *NotNan = createBasicBlock("fpclassify_not_nan", this->CurFn); |
2275 | Builder.CreateCondBr(IsNan, End, NotNan); |
2276 | Result->addIncoming(NanLiteral, NotZero); |
2277 | |
2278 | // if (fabs(V) == infinity) return FP_INFINITY |
2279 | Builder.SetInsertPoint(NotNan); |
2280 | Value *VAbs = EmitFAbs(*this, V); |
2281 | Value *IsInf = |
2282 | Builder.CreateFCmpOEQ(VAbs, ConstantFP::getInfinity(V->getType()), |
2283 | "isinf"); |
2284 | Value *InfLiteral = EmitScalarExpr(E->getArg(1)); |
2285 | BasicBlock *NotInf = createBasicBlock("fpclassify_not_inf", this->CurFn); |
2286 | Builder.CreateCondBr(IsInf, End, NotInf); |
2287 | Result->addIncoming(InfLiteral, NotNan); |
2288 | |
2289 | // if (fabs(V) >= MIN_NORMAL) return FP_NORMAL else FP_SUBNORMAL |
2290 | Builder.SetInsertPoint(NotInf); |
2291 | APFloat Smallest = APFloat::getSmallestNormalized( |
2292 | getContext().getFloatTypeSemantics(E->getArg(5)->getType())); |
2293 | Value *IsNormal = |
2294 | Builder.CreateFCmpUGE(VAbs, ConstantFP::get(V->getContext(), Smallest), |
2295 | "isnormal"); |
2296 | Value *NormalResult = |
2297 | Builder.CreateSelect(IsNormal, EmitScalarExpr(E->getArg(2)), |
2298 | EmitScalarExpr(E->getArg(3))); |
2299 | Builder.CreateBr(End); |
2300 | Result->addIncoming(NormalResult, NotInf); |
2301 | |
2302 | // return Result |
2303 | Builder.SetInsertPoint(End); |
2304 | return RValue::get(Result); |
2305 | } |
2306 | |
2307 | case Builtin::BIalloca: |
2308 | case Builtin::BI_alloca: |
2309 | case Builtin::BI__builtin_alloca: { |
2310 | Value *Size = EmitScalarExpr(E->getArg(0)); |
2311 | const TargetInfo &TI = getContext().getTargetInfo(); |
2312 | // The alignment of the alloca should correspond to __BIGGEST_ALIGNMENT__. |
2313 | unsigned SuitableAlignmentInBytes = |
2314 | CGM.getContext() |
2315 | .toCharUnitsFromBits(TI.getSuitableAlign()) |
2316 | .getQuantity(); |
2317 | AllocaInst *AI = Builder.CreateAlloca(Builder.getInt8Ty(), Size); |
2318 | AI->setAlignment(SuitableAlignmentInBytes); |
2319 | initializeAlloca(*this, AI, Size, SuitableAlignmentInBytes); |
2320 | return RValue::get(AI); |
2321 | } |
2322 | |
2323 | case Builtin::BI__builtin_alloca_with_align: { |
2324 | Value *Size = EmitScalarExpr(E->getArg(0)); |
2325 | Value *AlignmentInBitsValue = EmitScalarExpr(E->getArg(1)); |
2326 | auto *AlignmentInBitsCI = cast<ConstantInt>(AlignmentInBitsValue); |
2327 | unsigned AlignmentInBits = AlignmentInBitsCI->getZExtValue(); |
2328 | unsigned AlignmentInBytes = |
2329 | CGM.getContext().toCharUnitsFromBits(AlignmentInBits).getQuantity(); |
2330 | AllocaInst *AI = Builder.CreateAlloca(Builder.getInt8Ty(), Size); |
2331 | AI->setAlignment(AlignmentInBytes); |
2332 | initializeAlloca(*this, AI, Size, AlignmentInBytes); |
2333 | return RValue::get(AI); |
2334 | } |
2335 | |
2336 | case Builtin::BIbzero: |
2337 | case Builtin::BI__builtin_bzero: { |
2338 | Address Dest = EmitPointerWithAlignment(E->getArg(0)); |
2339 | Value *SizeVal = EmitScalarExpr(E->getArg(1)); |
2340 | EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), |
2341 | E->getArg(0)->getExprLoc(), FD, 0); |
2342 | Builder.CreateMemSet(Dest, Builder.getInt8(0), SizeVal, false); |
2343 | return RValue::get(nullptr); |
2344 | } |
2345 | case Builtin::BImemcpy: |
2346 | case Builtin::BI__builtin_memcpy: { |
2347 | Address Dest = EmitPointerWithAlignment(E->getArg(0)); |
2348 | Address Src = EmitPointerWithAlignment(E->getArg(1)); |
2349 | Value *SizeVal = EmitScalarExpr(E->getArg(2)); |
2350 | EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), |
2351 | E->getArg(0)->getExprLoc(), FD, 0); |
2352 | EmitNonNullArgCheck(RValue::get(Src.getPointer()), E->getArg(1)->getType(), |
2353 | E->getArg(1)->getExprLoc(), FD, 1); |
2354 | Builder.CreateMemCpy(Dest, Src, SizeVal, false); |
2355 | return RValue::get(Dest.getPointer()); |
2356 | } |
2357 | |
2358 | case Builtin::BI__builtin_char_memchr: |
2359 | BuiltinID = Builtin::BI__builtin_memchr; |
2360 | break; |
2361 | |
2362 | case Builtin::BI__builtin___memcpy_chk: { |
2363 | // fold __builtin_memcpy_chk(x, y, cst1, cst2) to memcpy iff cst1<=cst2. |
2364 | Expr::EvalResult SizeResult, DstSizeResult; |
2365 | if (!E->getArg(2)->EvaluateAsInt(SizeResult, CGM.getContext()) || |
2366 | !E->getArg(3)->EvaluateAsInt(DstSizeResult, CGM.getContext())) |
2367 | break; |
2368 | llvm::APSInt Size = SizeResult.Val.getInt(); |
2369 | llvm::APSInt DstSize = DstSizeResult.Val.getInt(); |
2370 | if (Size.ugt(DstSize)) |
2371 | break; |
2372 | Address Dest = EmitPointerWithAlignment(E->getArg(0)); |
2373 | Address Src = EmitPointerWithAlignment(E->getArg(1)); |
2374 | Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); |
2375 | Builder.CreateMemCpy(Dest, Src, SizeVal, false); |
2376 | return RValue::get(Dest.getPointer()); |
2377 | } |
2378 | |
2379 | case Builtin::BI__builtin_objc_memmove_collectable: { |
2380 | Address DestAddr = EmitPointerWithAlignment(E->getArg(0)); |
2381 | Address SrcAddr = EmitPointerWithAlignment(E->getArg(1)); |
2382 | Value *SizeVal = EmitScalarExpr(E->getArg(2)); |
2383 | CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this, |
2384 | DestAddr, SrcAddr, SizeVal); |
2385 | return RValue::get(DestAddr.getPointer()); |
2386 | } |
2387 | |
2388 | case Builtin::BI__builtin___memmove_chk: { |
2389 | // fold __builtin_memmove_chk(x, y, cst1, cst2) to memmove iff cst1<=cst2. |
2390 | Expr::EvalResult SizeResult, DstSizeResult; |
2391 | if (!E->getArg(2)->EvaluateAsInt(SizeResult, CGM.getContext()) || |
2392 | !E->getArg(3)->EvaluateAsInt(DstSizeResult, CGM.getContext())) |
2393 | break; |
2394 | llvm::APSInt Size = SizeResult.Val.getInt(); |
2395 | llvm::APSInt DstSize = DstSizeResult.Val.getInt(); |
2396 | if (Size.ugt(DstSize)) |
2397 | break; |
2398 | Address Dest = EmitPointerWithAlignment(E->getArg(0)); |
2399 | Address Src = EmitPointerWithAlignment(E->getArg(1)); |
2400 | Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); |
2401 | Builder.CreateMemMove(Dest, Src, SizeVal, false); |
2402 | return RValue::get(Dest.getPointer()); |
2403 | } |
2404 | |
2405 | case Builtin::BImemmove: |
2406 | case Builtin::BI__builtin_memmove: { |
2407 | Address Dest = EmitPointerWithAlignment(E->getArg(0)); |
2408 | Address Src = EmitPointerWithAlignment(E->getArg(1)); |
2409 | Value *SizeVal = EmitScalarExpr(E->getArg(2)); |
2410 | EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), |
2411 | E->getArg(0)->getExprLoc(), FD, 0); |
2412 | EmitNonNullArgCheck(RValue::get(Src.getPointer()), E->getArg(1)->getType(), |
2413 | E->getArg(1)->getExprLoc(), FD, 1); |
2414 | Builder.CreateMemMove(Dest, Src, SizeVal, false); |
2415 | return RValue::get(Dest.getPointer()); |
2416 | } |
2417 | case Builtin::BImemset: |
2418 | case Builtin::BI__builtin_memset: { |
2419 | Address Dest = EmitPointerWithAlignment(E->getArg(0)); |
2420 | Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), |
2421 | Builder.getInt8Ty()); |
2422 | Value *SizeVal = EmitScalarExpr(E->getArg(2)); |
2423 | EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), |
2424 | E->getArg(0)->getExprLoc(), FD, 0); |
2425 | Builder.CreateMemSet(Dest, ByteVal, SizeVal, false); |
2426 | return RValue::get(Dest.getPointer()); |
2427 | } |
2428 | case Builtin::BI__builtin___memset_chk: { |
2429 | // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2. |
2430 | Expr::EvalResult SizeResult, DstSizeResult; |
2431 | if (!E->getArg(2)->EvaluateAsInt(SizeResult, CGM.getContext()) || |
2432 | !E->getArg(3)->EvaluateAsInt(DstSizeResult, CGM.getContext())) |
2433 | break; |
2434 | llvm::APSInt Size = SizeResult.Val.getInt(); |
2435 | llvm::APSInt DstSize = DstSizeResult.Val.getInt(); |
2436 | if (Size.ugt(DstSize)) |
2437 | break; |
2438 | Address Dest = EmitPointerWithAlignment(E->getArg(0)); |
2439 | Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), |
2440 | Builder.getInt8Ty()); |
2441 | Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); |
2442 | Builder.CreateMemSet(Dest, ByteVal, SizeVal, false); |
2443 | return RValue::get(Dest.getPointer()); |
2444 | } |
2445 | case Builtin::BI__builtin_wmemcmp: { |
2446 | // The MSVC runtime library does not provide a definition of wmemcmp, so we |
2447 | // need an inline implementation. |
2448 | if (!getTarget().getTriple().isOSMSVCRT()) |
2449 | break; |
2450 | |
2451 | llvm::Type *WCharTy = ConvertType(getContext().WCharTy); |
2452 | |
2453 | Value *Dst = EmitScalarExpr(E->getArg(0)); |
2454 | Value *Src = EmitScalarExpr(E->getArg(1)); |
2455 | Value *Size = EmitScalarExpr(E->getArg(2)); |
2456 | |
2457 | BasicBlock *Entry = Builder.GetInsertBlock(); |
2458 | BasicBlock *CmpGT = createBasicBlock("wmemcmp.gt"); |
2459 | BasicBlock *CmpLT = createBasicBlock("wmemcmp.lt"); |
2460 | BasicBlock *Next = createBasicBlock("wmemcmp.next"); |
2461 | BasicBlock *Exit = createBasicBlock("wmemcmp.exit"); |
2462 | Value *SizeEq0 = Builder.CreateICmpEQ(Size, ConstantInt::get(SizeTy, 0)); |
2463 | Builder.CreateCondBr(SizeEq0, Exit, CmpGT); |
2464 | |
2465 | EmitBlock(CmpGT); |
2466 | PHINode *DstPhi = Builder.CreatePHI(Dst->getType(), 2); |
2467 | DstPhi->addIncoming(Dst, Entry); |
2468 | PHINode *SrcPhi = Builder.CreatePHI(Src->getType(), 2); |
2469 | SrcPhi->addIncoming(Src, Entry); |
2470 | PHINode *SizePhi = Builder.CreatePHI(SizeTy, 2); |
2471 | SizePhi->addIncoming(Size, Entry); |
2472 | CharUnits WCharAlign = |
2473 | getContext().getTypeAlignInChars(getContext().WCharTy); |
2474 | Value *DstCh = Builder.CreateAlignedLoad(WCharTy, DstPhi, WCharAlign); |
2475 | Value *SrcCh = Builder.CreateAlignedLoad(WCharTy, SrcPhi, WCharAlign); |
2476 | Value *DstGtSrc = Builder.CreateICmpUGT(DstCh, SrcCh); |
2477 | Builder.CreateCondBr(DstGtSrc, Exit, CmpLT); |
2478 | |
2479 | EmitBlock(CmpLT); |
2480 | Value *DstLtSrc = Builder.CreateICmpULT(DstCh, SrcCh); |
2481 | Builder.CreateCondBr(DstLtSrc, Exit, Next); |
2482 | |
2483 | EmitBlock(Next); |
2484 | Value *NextDst = Builder.CreateConstInBoundsGEP1_32(WCharTy, DstPhi, 1); |
2485 | Value *NextSrc = Builder.CreateConstInBoundsGEP1_32(WCharTy, SrcPhi, 1); |
2486 | Value *NextSize = Builder.CreateSub(SizePhi, ConstantInt::get(SizeTy, 1)); |
2487 | Value *NextSizeEq0 = |
2488 | Builder.CreateICmpEQ(NextSize, ConstantInt::get(SizeTy, 0)); |
2489 | Builder.CreateCondBr(NextSizeEq0, Exit, CmpGT); |
2490 | DstPhi->addIncoming(NextDst, Next); |
2491 | SrcPhi->addIncoming(NextSrc, Next); |
2492 | SizePhi->addIncoming(NextSize, Next); |
2493 | |
2494 | EmitBlock(Exit); |
2495 | PHINode *Ret = Builder.CreatePHI(IntTy, 4); |
2496 | Ret->addIncoming(ConstantInt::get(IntTy, 0), Entry); |
2497 | Ret->addIncoming(ConstantInt::get(IntTy, 1), CmpGT); |
2498 | Ret->addIncoming(ConstantInt::get(IntTy, -1), CmpLT); |
2499 | Ret->addIncoming(ConstantInt::get(IntTy, 0), Next); |
2500 | return RValue::get(Ret); |
2501 | } |
2502 | case Builtin::BI__builtin_dwarf_cfa: { |
2503 | // The offset in bytes from the first argument to the CFA. |
2504 | // |
2505 | // Why on earth is this in the frontend? Is there any reason at |
2506 | // all that the backend can't reasonably determine this while |
2507 | // lowering llvm.eh.dwarf.cfa()? |
2508 | // |
2509 | // TODO: If there's a satisfactory reason, add a target hook for |
2510 | // this instead of hard-coding 0, which is correct for most targets. |
2511 | int32_t Offset = 0; |
2512 | |
2513 | Function *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa); |
2514 | return RValue::get(Builder.CreateCall(F, |
2515 | llvm::ConstantInt::get(Int32Ty, Offset))); |
2516 | } |
2517 | case Builtin::BI__builtin_return_address: { |
2518 | Value *Depth = ConstantEmitter(*this).emitAbstract(E->getArg(0), |
2519 | getContext().UnsignedIntTy); |
2520 | Function *F = CGM.getIntrinsic(Intrinsic::returnaddress); |
2521 | return RValue::get(Builder.CreateCall(F, Depth)); |
2522 | } |
2523 | case Builtin::BI_ReturnAddress: { |
2524 | Function *F = CGM.getIntrinsic(Intrinsic::returnaddress); |
2525 | return RValue::get(Builder.CreateCall(F, Builder.getInt32(0))); |
2526 | } |
2527 | case Builtin::BI__builtin_frame_address: { |
2528 | Value *Depth = ConstantEmitter(*this).emitAbstract(E->getArg(0), |
2529 | getContext().UnsignedIntTy); |
2530 | Function *F = CGM.getIntrinsic(Intrinsic::frameaddress); |
2531 | return RValue::get(Builder.CreateCall(F, Depth)); |
2532 | } |
2533 | case Builtin::BI__builtin_extract_return_addr: { |
2534 | Value *Address = EmitScalarExpr(E->getArg(0)); |
2535 | Value *Result = getTargetHooks().decodeReturnAddress(*this, Address); |
2536 | return RValue::get(Result); |
2537 | } |
2538 | case Builtin::BI__builtin_frob_return_addr: { |
2539 | Value *Address = EmitScalarExpr(E->getArg(0)); |
2540 | Value *Result = getTargetHooks().encodeReturnAddress(*this, Address); |
2541 | return RValue::get(Result); |
2542 | } |
2543 | case Builtin::BI__builtin_dwarf_sp_column: { |
2544 | llvm::IntegerType *Ty |
2545 | = cast<llvm::IntegerType>(ConvertType(E->getType())); |
2546 | int Column = getTargetHooks().getDwarfEHStackPointer(CGM); |
2547 | if (Column == -1) { |
2548 | CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column"); |
2549 | return RValue::get(llvm::UndefValue::get(Ty)); |
2550 | } |
2551 | return RValue::get(llvm::ConstantInt::get(Ty, Column, true)); |
2552 | } |
2553 | case Builtin::BI__builtin_init_dwarf_reg_size_table: { |
2554 | Value *Address = EmitScalarExpr(E->getArg(0)); |
2555 | if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address)) |
2556 | CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table"); |
2557 | return RValue::get(llvm::UndefValue::get(ConvertType(E->getType()))); |
2558 | } |
2559 | case Builtin::BI__builtin_eh_return: { |
2560 | Value *Int = EmitScalarExpr(E->getArg(0)); |
2561 | Value *Ptr = EmitScalarExpr(E->getArg(1)); |
2562 | |
2563 | llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType()); |
2564 | assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) &&(((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) && "LLVM's __builtin_eh_return only supports 32- and 64-bit variants" ) ? static_cast<void> (0) : __assert_fail ("(IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) && \"LLVM's __builtin_eh_return only supports 32- and 64-bit variants\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 2565, __PRETTY_FUNCTION__)) |
2565 | "LLVM's __builtin_eh_return only supports 32- and 64-bit variants")(((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) && "LLVM's __builtin_eh_return only supports 32- and 64-bit variants" ) ? static_cast<void> (0) : __assert_fail ("(IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) && \"LLVM's __builtin_eh_return only supports 32- and 64-bit variants\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 2565, __PRETTY_FUNCTION__)); |
2566 | Function *F = |
2567 | CGM.getIntrinsic(IntTy->getBitWidth() == 32 ? Intrinsic::eh_return_i32 |
2568 | : Intrinsic::eh_return_i64); |
2569 | Builder.CreateCall(F, {Int, Ptr}); |
2570 | Builder.CreateUnreachable(); |
2571 | |
2572 | // We do need to preserve an insertion point. |
2573 | EmitBlock(createBasicBlock("builtin_eh_return.cont")); |
2574 | |
2575 | return RValue::get(nullptr); |
2576 | } |
2577 | case Builtin::BI__builtin_unwind_init: { |
2578 | Function *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init); |
2579 | return RValue::get(Builder.CreateCall(F)); |
2580 | } |
2581 | case Builtin::BI__builtin_extend_pointer: { |
2582 | // Extends a pointer to the size of an _Unwind_Word, which is |
2583 | // uint64_t on all platforms. Generally this gets poked into a |
2584 | // register and eventually used as an address, so if the |
2585 | // addressing registers are wider than pointers and the platform |
2586 | // doesn't implicitly ignore high-order bits when doing |
2587 | // addressing, we need to make sure we zext / sext based on |
2588 | // the platform's expectations. |
2589 | // |
2590 | // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html |
2591 | |
2592 | // Cast the pointer to intptr_t. |
2593 | Value *Ptr = EmitScalarExpr(E->getArg(0)); |
2594 | Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast"); |
2595 | |
2596 | // If that's 64 bits, we're done. |
2597 | if (IntPtrTy->getBitWidth() == 64) |
2598 | return RValue::get(Result); |
2599 | |
2600 | // Otherwise, ask the codegen data what to do. |
2601 | if (getTargetHooks().extendPointerWithSExt()) |
2602 | return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext")); |
2603 | else |
2604 | return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext")); |
2605 | } |
2606 | case Builtin::BI__builtin_setjmp: { |
2607 | // Buffer is a void**. |
2608 | Address Buf = EmitPointerWithAlignment(E->getArg(0)); |
2609 | |
2610 | // Store the frame pointer to the setjmp buffer. |
2611 | Value *FrameAddr = |
2612 | Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), |
2613 | ConstantInt::get(Int32Ty, 0)); |
2614 | Builder.CreateStore(FrameAddr, Buf); |
2615 | |
2616 | // Store the stack pointer to the setjmp buffer. |
2617 | Value *StackAddr = |
2618 | Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave)); |
2619 | Address StackSaveSlot = Builder.CreateConstInBoundsGEP(Buf, 2); |
2620 | Builder.CreateStore(StackAddr, StackSaveSlot); |
2621 | |
2622 | // Call LLVM's EH setjmp, which is lightweight. |
2623 | Function *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp); |
2624 | Buf = Builder.CreateBitCast(Buf, Int8PtrTy); |
2625 | return RValue::get(Builder.CreateCall(F, Buf.getPointer())); |
2626 | } |
2627 | case Builtin::BI__builtin_longjmp: { |
2628 | Value *Buf = EmitScalarExpr(E->getArg(0)); |
2629 | Buf = Builder.CreateBitCast(Buf, Int8PtrTy); |
2630 | |
2631 | // Call LLVM's EH longjmp, which is lightweight. |
2632 | Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf); |
2633 | |
2634 | // longjmp doesn't return; mark this as unreachable. |
2635 | Builder.CreateUnreachable(); |
2636 | |
2637 | // We do need to preserve an insertion point. |
2638 | EmitBlock(createBasicBlock("longjmp.cont")); |
2639 | |
2640 | return RValue::get(nullptr); |
2641 | } |
2642 | case Builtin::BI__builtin_launder: { |
2643 | const Expr *Arg = E->getArg(0); |
2644 | QualType ArgTy = Arg->getType()->getPointeeType(); |
2645 | Value *Ptr = EmitScalarExpr(Arg); |
2646 | if (TypeRequiresBuiltinLaunder(CGM, ArgTy)) |
2647 | Ptr = Builder.CreateLaunderInvariantGroup(Ptr); |
2648 | |
2649 | return RValue::get(Ptr); |
2650 | } |
2651 | case Builtin::BI__sync_fetch_and_add: |
2652 | case Builtin::BI__sync_fetch_and_sub: |
2653 | case Builtin::BI__sync_fetch_and_or: |
2654 | case Builtin::BI__sync_fetch_and_and: |
2655 | case Builtin::BI__sync_fetch_and_xor: |
2656 | case Builtin::BI__sync_fetch_and_nand: |
2657 | case Builtin::BI__sync_add_and_fetch: |
2658 | case Builtin::BI__sync_sub_and_fetch: |
2659 | case Builtin::BI__sync_and_and_fetch: |
2660 | case Builtin::BI__sync_or_and_fetch: |
2661 | case Builtin::BI__sync_xor_and_fetch: |
2662 | case Builtin::BI__sync_nand_and_fetch: |
2663 | case Builtin::BI__sync_val_compare_and_swap: |
2664 | case Builtin::BI__sync_bool_compare_and_swap: |
2665 | case Builtin::BI__sync_lock_test_and_set: |
2666 | case Builtin::BI__sync_lock_release: |
2667 | case Builtin::BI__sync_swap: |
2668 | llvm_unreachable("Shouldn't make it through sema")::llvm::llvm_unreachable_internal("Shouldn't make it through sema" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 2668); |
2669 | case Builtin::BI__sync_fetch_and_add_1: |
2670 | case Builtin::BI__sync_fetch_and_add_2: |
2671 | case Builtin::BI__sync_fetch_and_add_4: |
2672 | case Builtin::BI__sync_fetch_and_add_8: |
2673 | case Builtin::BI__sync_fetch_and_add_16: |
2674 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E); |
2675 | case Builtin::BI__sync_fetch_and_sub_1: |
2676 | case Builtin::BI__sync_fetch_and_sub_2: |
2677 | case Builtin::BI__sync_fetch_and_sub_4: |
2678 | case Builtin::BI__sync_fetch_and_sub_8: |
2679 | case Builtin::BI__sync_fetch_and_sub_16: |
2680 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E); |
2681 | case Builtin::BI__sync_fetch_and_or_1: |
2682 | case Builtin::BI__sync_fetch_and_or_2: |
2683 | case Builtin::BI__sync_fetch_and_or_4: |
2684 | case Builtin::BI__sync_fetch_and_or_8: |
2685 | case Builtin::BI__sync_fetch_and_or_16: |
2686 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E); |
2687 | case Builtin::BI__sync_fetch_and_and_1: |
2688 | case Builtin::BI__sync_fetch_and_and_2: |
2689 | case Builtin::BI__sync_fetch_and_and_4: |
2690 | case Builtin::BI__sync_fetch_and_and_8: |
2691 | case Builtin::BI__sync_fetch_and_and_16: |
2692 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E); |
2693 | case Builtin::BI__sync_fetch_and_xor_1: |
2694 | case Builtin::BI__sync_fetch_and_xor_2: |
2695 | case Builtin::BI__sync_fetch_and_xor_4: |
2696 | case Builtin::BI__sync_fetch_and_xor_8: |
2697 | case Builtin::BI__sync_fetch_and_xor_16: |
2698 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E); |
2699 | case Builtin::BI__sync_fetch_and_nand_1: |
2700 | case Builtin::BI__sync_fetch_and_nand_2: |
2701 | case Builtin::BI__sync_fetch_and_nand_4: |
2702 | case Builtin::BI__sync_fetch_and_nand_8: |
2703 | case Builtin::BI__sync_fetch_and_nand_16: |
2704 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Nand, E); |
2705 | |
2706 | // Clang extensions: not overloaded yet. |
2707 | case Builtin::BI__sync_fetch_and_min: |
2708 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E); |
2709 | case Builtin::BI__sync_fetch_and_max: |
2710 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E); |
2711 | case Builtin::BI__sync_fetch_and_umin: |
2712 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E); |
2713 | case Builtin::BI__sync_fetch_and_umax: |
2714 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E); |
2715 | |
2716 | case Builtin::BI__sync_add_and_fetch_1: |
2717 | case Builtin::BI__sync_add_and_fetch_2: |
2718 | case Builtin::BI__sync_add_and_fetch_4: |
2719 | case Builtin::BI__sync_add_and_fetch_8: |
2720 | case Builtin::BI__sync_add_and_fetch_16: |
2721 | return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E, |
2722 | llvm::Instruction::Add); |
2723 | case Builtin::BI__sync_sub_and_fetch_1: |
2724 | case Builtin::BI__sync_sub_and_fetch_2: |
2725 | case Builtin::BI__sync_sub_and_fetch_4: |
2726 | case Builtin::BI__sync_sub_and_fetch_8: |
2727 | case Builtin::BI__sync_sub_and_fetch_16: |
2728 | return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E, |
2729 | llvm::Instruction::Sub); |
2730 | case Builtin::BI__sync_and_and_fetch_1: |
2731 | case Builtin::BI__sync_and_and_fetch_2: |
2732 | case Builtin::BI__sync_and_and_fetch_4: |
2733 | case Builtin::BI__sync_and_and_fetch_8: |
2734 | case Builtin::BI__sync_and_and_fetch_16: |
2735 | return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::And, E, |
2736 | llvm::Instruction::And); |
2737 | case Builtin::BI__sync_or_and_fetch_1: |
2738 | case Builtin::BI__sync_or_and_fetch_2: |
2739 | case Builtin::BI__sync_or_and_fetch_4: |
2740 | case Builtin::BI__sync_or_and_fetch_8: |
2741 | case Builtin::BI__sync_or_and_fetch_16: |
2742 | return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E, |
2743 | llvm::Instruction::Or); |
2744 | case Builtin::BI__sync_xor_and_fetch_1: |
2745 | case Builtin::BI__sync_xor_and_fetch_2: |
2746 | case Builtin::BI__sync_xor_and_fetch_4: |
2747 | case Builtin::BI__sync_xor_and_fetch_8: |
2748 | case Builtin::BI__sync_xor_and_fetch_16: |
2749 | return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E, |
2750 | llvm::Instruction::Xor); |
2751 | case Builtin::BI__sync_nand_and_fetch_1: |
2752 | case Builtin::BI__sync_nand_and_fetch_2: |
2753 | case Builtin::BI__sync_nand_and_fetch_4: |
2754 | case Builtin::BI__sync_nand_and_fetch_8: |
2755 | case Builtin::BI__sync_nand_and_fetch_16: |
2756 | return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Nand, E, |
2757 | llvm::Instruction::And, true); |
2758 | |
2759 | case Builtin::BI__sync_val_compare_and_swap_1: |
2760 | case Builtin::BI__sync_val_compare_and_swap_2: |
2761 | case Builtin::BI__sync_val_compare_and_swap_4: |
2762 | case Builtin::BI__sync_val_compare_and_swap_8: |
2763 | case Builtin::BI__sync_val_compare_and_swap_16: |
2764 | return RValue::get(MakeAtomicCmpXchgValue(*this, E, false)); |
2765 | |
2766 | case Builtin::BI__sync_bool_compare_and_swap_1: |
2767 | case Builtin::BI__sync_bool_compare_and_swap_2: |
2768 | case Builtin::BI__sync_bool_compare_and_swap_4: |
2769 | case Builtin::BI__sync_bool_compare_and_swap_8: |
2770 | case Builtin::BI__sync_bool_compare_and_swap_16: |
2771 | return RValue::get(MakeAtomicCmpXchgValue(*this, E, true)); |
2772 | |
2773 | case Builtin::BI__sync_swap_1: |
2774 | case Builtin::BI__sync_swap_2: |
2775 | case Builtin::BI__sync_swap_4: |
2776 | case Builtin::BI__sync_swap_8: |
2777 | case Builtin::BI__sync_swap_16: |
2778 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); |
2779 | |
2780 | case Builtin::BI__sync_lock_test_and_set_1: |
2781 | case Builtin::BI__sync_lock_test_and_set_2: |
2782 | case Builtin::BI__sync_lock_test_and_set_4: |
2783 | case Builtin::BI__sync_lock_test_and_set_8: |
2784 | case Builtin::BI__sync_lock_test_and_set_16: |
2785 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); |
2786 | |
2787 | case Builtin::BI__sync_lock_release_1: |
2788 | case Builtin::BI__sync_lock_release_2: |
2789 | case Builtin::BI__sync_lock_release_4: |
2790 | case Builtin::BI__sync_lock_release_8: |
2791 | case Builtin::BI__sync_lock_release_16: { |
2792 | Value *Ptr = EmitScalarExpr(E->getArg(0)); |
2793 | QualType ElTy = E->getArg(0)->getType()->getPointeeType(); |
2794 | CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy); |
2795 | llvm::Type *ITy = llvm::IntegerType::get(getLLVMContext(), |
2796 | StoreSize.getQuantity() * 8); |
2797 | Ptr = Builder.CreateBitCast(Ptr, ITy->getPointerTo()); |
2798 | llvm::StoreInst *Store = |
2799 | Builder.CreateAlignedStore(llvm::Constant::getNullValue(ITy), Ptr, |
2800 | StoreSize); |
2801 | Store->setAtomic(llvm::AtomicOrdering::Release); |
2802 | return RValue::get(nullptr); |
2803 | } |
2804 | |
2805 | case Builtin::BI__sync_synchronize: { |
2806 | // We assume this is supposed to correspond to a C++0x-style |
2807 | // sequentially-consistent fence (i.e. this is only usable for |
2808 | // synchronization, not device I/O or anything like that). This intrinsic |
2809 | // is really badly designed in the sense that in theory, there isn't |
2810 | // any way to safely use it... but in practice, it mostly works |
2811 | // to use it with non-atomic loads and stores to get acquire/release |
2812 | // semantics. |
2813 | Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent); |
2814 | return RValue::get(nullptr); |
2815 | } |
2816 | |
2817 | case Builtin::BI__builtin_nontemporal_load: |
2818 | return RValue::get(EmitNontemporalLoad(*this, E)); |
2819 | case Builtin::BI__builtin_nontemporal_store: |
2820 | return RValue::get(EmitNontemporalStore(*this, E)); |
2821 | case Builtin::BI__c11_atomic_is_lock_free: |
2822 | case Builtin::BI__atomic_is_lock_free: { |
2823 | // Call "bool __atomic_is_lock_free(size_t size, void *ptr)". For the |
2824 | // __c11 builtin, ptr is 0 (indicating a properly-aligned object), since |
2825 | // _Atomic(T) is always properly-aligned. |
2826 | const char *LibCallName = "__atomic_is_lock_free"; |
2827 | CallArgList Args; |
2828 | Args.add(RValue::get(EmitScalarExpr(E->getArg(0))), |
2829 | getContext().getSizeType()); |
2830 | if (BuiltinID == Builtin::BI__atomic_is_lock_free) |
2831 | Args.add(RValue::get(EmitScalarExpr(E->getArg(1))), |
2832 | getContext().VoidPtrTy); |
2833 | else |
2834 | Args.add(RValue::get(llvm::Constant::getNullValue(VoidPtrTy)), |
2835 | getContext().VoidPtrTy); |
2836 | const CGFunctionInfo &FuncInfo = |
2837 | CGM.getTypes().arrangeBuiltinFunctionCall(E->getType(), Args); |
2838 | llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FuncInfo); |
2839 | llvm::FunctionCallee Func = CGM.CreateRuntimeFunction(FTy, LibCallName); |
2840 | return EmitCall(FuncInfo, CGCallee::forDirect(Func), |
2841 | ReturnValueSlot(), Args); |
2842 | } |
2843 | |
2844 | case Builtin::BI__atomic_test_and_set: { |
2845 | // Look at the argument type to determine whether this is a volatile |
2846 | // operation. The parameter type is always volatile. |
2847 | QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); |
2848 | bool Volatile = |
2849 | PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); |
2850 | |
2851 | Value *Ptr = EmitScalarExpr(E->getArg(0)); |
2852 | unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); |
2853 | Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); |
2854 | Value *NewVal = Builder.getInt8(1); |
2855 | Value *Order = EmitScalarExpr(E->getArg(1)); |
2856 | if (isa<llvm::ConstantInt>(Order)) { |
2857 | int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); |
2858 | AtomicRMWInst *Result = nullptr; |
2859 | switch (ord) { |
2860 | case 0: // memory_order_relaxed |
2861 | default: // invalid order |
2862 | Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal, |
2863 | llvm::AtomicOrdering::Monotonic); |
2864 | break; |
2865 | case 1: // memory_order_consume |
2866 | case 2: // memory_order_acquire |
2867 | Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal, |
2868 | llvm::AtomicOrdering::Acquire); |
2869 | break; |
2870 | case 3: // memory_order_release |
2871 | Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal, |
2872 | llvm::AtomicOrdering::Release); |
2873 | break; |
2874 | case 4: // memory_order_acq_rel |
2875 | |
2876 | Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal, |
2877 | llvm::AtomicOrdering::AcquireRelease); |
2878 | break; |
2879 | case 5: // memory_order_seq_cst |
2880 | Result = Builder.CreateAtomicRMW( |
2881 | llvm::AtomicRMWInst::Xchg, Ptr, NewVal, |
2882 | llvm::AtomicOrdering::SequentiallyConsistent); |
2883 | break; |
2884 | } |
2885 | Result->setVolatile(Volatile); |
2886 | return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); |
2887 | } |
2888 | |
2889 | llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); |
2890 | |
2891 | llvm::BasicBlock *BBs[5] = { |
2892 | createBasicBlock("monotonic", CurFn), |
2893 | createBasicBlock("acquire", CurFn), |
2894 | createBasicBlock("release", CurFn), |
2895 | createBasicBlock("acqrel", CurFn), |
2896 | createBasicBlock("seqcst", CurFn) |
2897 | }; |
2898 | llvm::AtomicOrdering Orders[5] = { |
2899 | llvm::AtomicOrdering::Monotonic, llvm::AtomicOrdering::Acquire, |
2900 | llvm::AtomicOrdering::Release, llvm::AtomicOrdering::AcquireRelease, |
2901 | llvm::AtomicOrdering::SequentiallyConsistent}; |
2902 | |
2903 | Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); |
2904 | llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); |
2905 | |
2906 | Builder.SetInsertPoint(ContBB); |
2907 | PHINode *Result = Builder.CreatePHI(Int8Ty, 5, "was_set"); |
2908 | |
2909 | for (unsigned i = 0; i < 5; ++i) { |
2910 | Builder.SetInsertPoint(BBs[i]); |
2911 | AtomicRMWInst *RMW = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, |
2912 | Ptr, NewVal, Orders[i]); |
2913 | RMW->setVolatile(Volatile); |
2914 | Result->addIncoming(RMW, BBs[i]); |
2915 | Builder.CreateBr(ContBB); |
2916 | } |
2917 | |
2918 | SI->addCase(Builder.getInt32(0), BBs[0]); |
2919 | SI->addCase(Builder.getInt32(1), BBs[1]); |
2920 | SI->addCase(Builder.getInt32(2), BBs[1]); |
2921 | SI->addCase(Builder.getInt32(3), BBs[2]); |
2922 | SI->addCase(Builder.getInt32(4), BBs[3]); |
2923 | SI->addCase(Builder.getInt32(5), BBs[4]); |
2924 | |
2925 | Builder.SetInsertPoint(ContBB); |
2926 | return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); |
2927 | } |
2928 | |
2929 | case Builtin::BI__atomic_clear: { |
2930 | QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); |
2931 | bool Volatile = |
2932 | PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); |
2933 | |
2934 | Address Ptr = EmitPointerWithAlignment(E->getArg(0)); |
2935 | unsigned AddrSpace = Ptr.getPointer()->getType()->getPointerAddressSpace(); |
2936 | Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); |
2937 | Value *NewVal = Builder.getInt8(0); |
2938 | Value *Order = EmitScalarExpr(E->getArg(1)); |
2939 | if (isa<llvm::ConstantInt>(Order)) { |
2940 | int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); |
2941 | StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); |
2942 | switch (ord) { |
2943 | case 0: // memory_order_relaxed |
2944 | default: // invalid order |
2945 | Store->setOrdering(llvm::AtomicOrdering::Monotonic); |
2946 | break; |
2947 | case 3: // memory_order_release |
2948 | Store->setOrdering(llvm::AtomicOrdering::Release); |
2949 | break; |
2950 | case 5: // memory_order_seq_cst |
2951 | Store->setOrdering(llvm::AtomicOrdering::SequentiallyConsistent); |
2952 | break; |
2953 | } |
2954 | return RValue::get(nullptr); |
2955 | } |
2956 | |
2957 | llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); |
2958 | |
2959 | llvm::BasicBlock *BBs[3] = { |
2960 | createBasicBlock("monotonic", CurFn), |
2961 | createBasicBlock("release", CurFn), |
2962 | createBasicBlock("seqcst", CurFn) |
2963 | }; |
2964 | llvm::AtomicOrdering Orders[3] = { |
2965 | llvm::AtomicOrdering::Monotonic, llvm::AtomicOrdering::Release, |
2966 | llvm::AtomicOrdering::SequentiallyConsistent}; |
2967 | |
2968 | Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); |
2969 | llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); |
2970 | |
2971 | for (unsigned i = 0; i < 3; ++i) { |
2972 | Builder.SetInsertPoint(BBs[i]); |
2973 | StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); |
2974 | Store->setOrdering(Orders[i]); |
2975 | Builder.CreateBr(ContBB); |
2976 | } |
2977 | |
2978 | SI->addCase(Builder.getInt32(0), BBs[0]); |
2979 | SI->addCase(Builder.getInt32(3), BBs[1]); |
2980 | SI->addCase(Builder.getInt32(5), BBs[2]); |
2981 | |
2982 | Builder.SetInsertPoint(ContBB); |
2983 | return RValue::get(nullptr); |
2984 | } |
2985 | |
2986 | case Builtin::BI__atomic_thread_fence: |
2987 | case Builtin::BI__atomic_signal_fence: |
2988 | case Builtin::BI__c11_atomic_thread_fence: |
2989 | case Builtin::BI__c11_atomic_signal_fence: { |
2990 | llvm::SyncScope::ID SSID; |
2991 | if (BuiltinID == Builtin::BI__atomic_signal_fence || |
2992 | BuiltinID == Builtin::BI__c11_atomic_signal_fence) |
2993 | SSID = llvm::SyncScope::SingleThread; |
2994 | else |
2995 | SSID = llvm::SyncScope::System; |
2996 | Value *Order = EmitScalarExpr(E->getArg(0)); |
2997 | if (isa<llvm::ConstantInt>(Order)) { |
2998 | int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); |
2999 | switch (ord) { |
3000 | case 0: // memory_order_relaxed |
3001 | default: // invalid order |
3002 | break; |
3003 | case 1: // memory_order_consume |
3004 | case 2: // memory_order_acquire |
3005 | Builder.CreateFence(llvm::AtomicOrdering::Acquire, SSID); |
3006 | break; |
3007 | case 3: // memory_order_release |
3008 | Builder.CreateFence(llvm::AtomicOrdering::Release, SSID); |
3009 | break; |
3010 | case 4: // memory_order_acq_rel |
3011 | Builder.CreateFence(llvm::AtomicOrdering::AcquireRelease, SSID); |
3012 | break; |
3013 | case 5: // memory_order_seq_cst |
3014 | Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent, SSID); |
3015 | break; |
3016 | } |
3017 | return RValue::get(nullptr); |
3018 | } |
3019 | |
3020 | llvm::BasicBlock *AcquireBB, *ReleaseBB, *AcqRelBB, *SeqCstBB; |
3021 | AcquireBB = createBasicBlock("acquire", CurFn); |
3022 | ReleaseBB = createBasicBlock("release", CurFn); |
3023 | AcqRelBB = createBasicBlock("acqrel", CurFn); |
3024 | SeqCstBB = createBasicBlock("seqcst", CurFn); |
3025 | llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); |
3026 | |
3027 | Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); |
3028 | llvm::SwitchInst *SI = Builder.CreateSwitch(Order, ContBB); |
3029 | |
3030 | Builder.SetInsertPoint(AcquireBB); |
3031 | Builder.CreateFence(llvm::AtomicOrdering::Acquire, SSID); |
3032 | Builder.CreateBr(ContBB); |
3033 | SI->addCase(Builder.getInt32(1), AcquireBB); |
3034 | SI->addCase(Builder.getInt32(2), AcquireBB); |
3035 | |
3036 | Builder.SetInsertPoint(ReleaseBB); |
3037 | Builder.CreateFence(llvm::AtomicOrdering::Release, SSID); |
3038 | Builder.CreateBr(ContBB); |
3039 | SI->addCase(Builder.getInt32(3), ReleaseBB); |
3040 | |
3041 | Builder.SetInsertPoint(AcqRelBB); |
3042 | Builder.CreateFence(llvm::AtomicOrdering::AcquireRelease, SSID); |
3043 | Builder.CreateBr(ContBB); |
3044 | SI->addCase(Builder.getInt32(4), AcqRelBB); |
3045 | |
3046 | Builder.SetInsertPoint(SeqCstBB); |
3047 | Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent, SSID); |
3048 | Builder.CreateBr(ContBB); |
3049 | SI->addCase(Builder.getInt32(5), SeqCstBB); |
3050 | |
3051 | Builder.SetInsertPoint(ContBB); |
3052 | return RValue::get(nullptr); |
3053 | } |
3054 | |
3055 | case Builtin::BI__builtin_signbit: |
3056 | case Builtin::BI__builtin_signbitf: |
3057 | case Builtin::BI__builtin_signbitl: { |
3058 | return RValue::get( |
3059 | Builder.CreateZExt(EmitSignBit(*this, EmitScalarExpr(E->getArg(0))), |
3060 | ConvertType(E->getType()))); |
3061 | } |
3062 | case Builtin::BI__annotation: { |
3063 | // Re-encode each wide string to UTF8 and make an MDString. |
3064 | SmallVector<Metadata *, 1> Strings; |
3065 | for (const Expr *Arg : E->arguments()) { |
3066 | const auto *Str = cast<StringLiteral>(Arg->IgnoreParenCasts()); |
3067 | assert(Str->getCharByteWidth() == 2)((Str->getCharByteWidth() == 2) ? static_cast<void> ( 0) : __assert_fail ("Str->getCharByteWidth() == 2", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 3067, __PRETTY_FUNCTION__)); |
3068 | StringRef WideBytes = Str->getBytes(); |
3069 | std::string StrUtf8; |
3070 | if (!convertUTF16ToUTF8String( |
3071 | makeArrayRef(WideBytes.data(), WideBytes.size()), StrUtf8)) { |
3072 | CGM.ErrorUnsupported(E, "non-UTF16 __annotation argument"); |
3073 | continue; |
3074 | } |
3075 | Strings.push_back(llvm::MDString::get(getLLVMContext(), StrUtf8)); |
3076 | } |
3077 | |
3078 | // Build and MDTuple of MDStrings and emit the intrinsic call. |
3079 | llvm::Function *F = |
3080 | CGM.getIntrinsic(llvm::Intrinsic::codeview_annotation, {}); |
3081 | MDTuple *StrTuple = MDTuple::get(getLLVMContext(), Strings); |
3082 | Builder.CreateCall(F, MetadataAsValue::get(getLLVMContext(), StrTuple)); |
3083 | return RValue::getIgnored(); |
3084 | } |
3085 | case Builtin::BI__builtin_annotation: { |
3086 | llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0)); |
3087 | llvm::Function *F = CGM.getIntrinsic(llvm::Intrinsic::annotation, |
3088 | AnnVal->getType()); |
3089 | |
3090 | // Get the annotation string, go through casts. Sema requires this to be a |
3091 | // non-wide string literal, potentially casted, so the cast<> is safe. |
3092 | const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts(); |
3093 | StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString(); |
3094 | return RValue::get(EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc())); |
3095 | } |
3096 | case Builtin::BI__builtin_addcb: |
3097 | case Builtin::BI__builtin_addcs: |
3098 | case Builtin::BI__builtin_addc: |
3099 | case Builtin::BI__builtin_addcl: |
3100 | case Builtin::BI__builtin_addcll: |
3101 | case Builtin::BI__builtin_subcb: |
3102 | case Builtin::BI__builtin_subcs: |
3103 | case Builtin::BI__builtin_subc: |
3104 | case Builtin::BI__builtin_subcl: |
3105 | case Builtin::BI__builtin_subcll: { |
3106 | |
3107 | // We translate all of these builtins from expressions of the form: |
3108 | // int x = ..., y = ..., carryin = ..., carryout, result; |
3109 | // result = __builtin_addc(x, y, carryin, &carryout); |
3110 | // |
3111 | // to LLVM IR of the form: |
3112 | // |
3113 | // %tmp1 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %x, i32 %y) |
3114 | // %tmpsum1 = extractvalue {i32, i1} %tmp1, 0 |
3115 | // %carry1 = extractvalue {i32, i1} %tmp1, 1 |
3116 | // %tmp2 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %tmpsum1, |
3117 | // i32 %carryin) |
3118 | // %result = extractvalue {i32, i1} %tmp2, 0 |
3119 | // %carry2 = extractvalue {i32, i1} %tmp2, 1 |
3120 | // %tmp3 = or i1 %carry1, %carry2 |
3121 | // %tmp4 = zext i1 %tmp3 to i32 |
3122 | // store i32 %tmp4, i32* %carryout |
3123 | |
3124 | // Scalarize our inputs. |
3125 | llvm::Value *X = EmitScalarExpr(E->getArg(0)); |
3126 | llvm::Value *Y = EmitScalarExpr(E->getArg(1)); |
3127 | llvm::Value *Carryin = EmitScalarExpr(E->getArg(2)); |
3128 | Address CarryOutPtr = EmitPointerWithAlignment(E->getArg(3)); |
3129 | |
3130 | // Decide if we are lowering to a uadd.with.overflow or usub.with.overflow. |
3131 | llvm::Intrinsic::ID IntrinsicId; |
3132 | switch (BuiltinID) { |
3133 | default: llvm_unreachable("Unknown multiprecision builtin id.")::llvm::llvm_unreachable_internal("Unknown multiprecision builtin id." , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 3133); |
3134 | case Builtin::BI__builtin_addcb: |
3135 | case Builtin::BI__builtin_addcs: |
3136 | case Builtin::BI__builtin_addc: |
3137 | case Builtin::BI__builtin_addcl: |
3138 | case Builtin::BI__builtin_addcll: |
3139 | IntrinsicId = llvm::Intrinsic::uadd_with_overflow; |
3140 | break; |
3141 | case Builtin::BI__builtin_subcb: |
3142 | case Builtin::BI__builtin_subcs: |
3143 | case Builtin::BI__builtin_subc: |
3144 | case Builtin::BI__builtin_subcl: |
3145 | case Builtin::BI__builtin_subcll: |
3146 | IntrinsicId = llvm::Intrinsic::usub_with_overflow; |
3147 | break; |
3148 | } |
3149 | |
3150 | // Construct our resulting LLVM IR expression. |
3151 | llvm::Value *Carry1; |
3152 | llvm::Value *Sum1 = EmitOverflowIntrinsic(*this, IntrinsicId, |
3153 | X, Y, Carry1); |
3154 | llvm::Value *Carry2; |
3155 | llvm::Value *Sum2 = EmitOverflowIntrinsic(*this, IntrinsicId, |
3156 | Sum1, Carryin, Carry2); |
3157 | llvm::Value *CarryOut = Builder.CreateZExt(Builder.CreateOr(Carry1, Carry2), |
3158 | X->getType()); |
3159 | Builder.CreateStore(CarryOut, CarryOutPtr); |
3160 | return RValue::get(Sum2); |
3161 | } |
3162 | |
3163 | case Builtin::BI__builtin_add_overflow: |
3164 | case Builtin::BI__builtin_sub_overflow: |
3165 | case Builtin::BI__builtin_mul_overflow: { |
3166 | const clang::Expr *LeftArg = E->getArg(0); |
3167 | const clang::Expr *RightArg = E->getArg(1); |
3168 | const clang::Expr *ResultArg = E->getArg(2); |
3169 | |
3170 | clang::QualType ResultQTy = |
3171 | ResultArg->getType()->castAs<PointerType>()->getPointeeType(); |
3172 | |
3173 | WidthAndSignedness LeftInfo = |
3174 | getIntegerWidthAndSignedness(CGM.getContext(), LeftArg->getType()); |
3175 | WidthAndSignedness RightInfo = |
3176 | getIntegerWidthAndSignedness(CGM.getContext(), RightArg->getType()); |
3177 | WidthAndSignedness ResultInfo = |
3178 | getIntegerWidthAndSignedness(CGM.getContext(), ResultQTy); |
3179 | |
3180 | // Handle mixed-sign multiplication as a special case, because adding |
3181 | // runtime or backend support for our generic irgen would be too expensive. |
3182 | if (isSpecialMixedSignMultiply(BuiltinID, LeftInfo, RightInfo, ResultInfo)) |
3183 | return EmitCheckedMixedSignMultiply(*this, LeftArg, LeftInfo, RightArg, |
3184 | RightInfo, ResultArg, ResultQTy, |
3185 | ResultInfo); |
3186 | |
3187 | WidthAndSignedness EncompassingInfo = |
3188 | EncompassingIntegerType({LeftInfo, RightInfo, ResultInfo}); |
3189 | |
3190 | llvm::Type *EncompassingLLVMTy = |
3191 | llvm::IntegerType::get(CGM.getLLVMContext(), EncompassingInfo.Width); |
3192 | |
3193 | llvm::Type *ResultLLVMTy = CGM.getTypes().ConvertType(ResultQTy); |
3194 | |
3195 | llvm::Intrinsic::ID IntrinsicId; |
3196 | switch (BuiltinID) { |
3197 | default: |
3198 | llvm_unreachable("Unknown overflow builtin id.")::llvm::llvm_unreachable_internal("Unknown overflow builtin id." , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 3198); |
3199 | case Builtin::BI__builtin_add_overflow: |
3200 | IntrinsicId = EncompassingInfo.Signed |
3201 | ? llvm::Intrinsic::sadd_with_overflow |
3202 | : llvm::Intrinsic::uadd_with_overflow; |
3203 | break; |
3204 | case Builtin::BI__builtin_sub_overflow: |
3205 | IntrinsicId = EncompassingInfo.Signed |
3206 | ? llvm::Intrinsic::ssub_with_overflow |
3207 | : llvm::Intrinsic::usub_with_overflow; |
3208 | break; |
3209 | case Builtin::BI__builtin_mul_overflow: |
3210 | IntrinsicId = EncompassingInfo.Signed |
3211 | ? llvm::Intrinsic::smul_with_overflow |
3212 | : llvm::Intrinsic::umul_with_overflow; |
3213 | break; |
3214 | } |
3215 | |
3216 | llvm::Value *Left = EmitScalarExpr(LeftArg); |
3217 | llvm::Value *Right = EmitScalarExpr(RightArg); |
3218 | Address ResultPtr = EmitPointerWithAlignment(ResultArg); |
3219 | |
3220 | // Extend each operand to the encompassing type. |
3221 | Left = Builder.CreateIntCast(Left, EncompassingLLVMTy, LeftInfo.Signed); |
3222 | Right = Builder.CreateIntCast(Right, EncompassingLLVMTy, RightInfo.Signed); |
3223 | |
3224 | // Perform the operation on the extended values. |
3225 | llvm::Value *Overflow, *Result; |
3226 | Result = EmitOverflowIntrinsic(*this, IntrinsicId, Left, Right, Overflow); |
3227 | |
3228 | if (EncompassingInfo.Width > ResultInfo.Width) { |
3229 | // The encompassing type is wider than the result type, so we need to |
3230 | // truncate it. |
3231 | llvm::Value *ResultTrunc = Builder.CreateTrunc(Result, ResultLLVMTy); |
3232 | |
3233 | // To see if the truncation caused an overflow, we will extend |
3234 | // the result and then compare it to the original result. |
3235 | llvm::Value *ResultTruncExt = Builder.CreateIntCast( |
3236 | ResultTrunc, EncompassingLLVMTy, ResultInfo.Signed); |
3237 | llvm::Value *TruncationOverflow = |
3238 | Builder.CreateICmpNE(Result, ResultTruncExt); |
3239 | |
3240 | Overflow = Builder.CreateOr(Overflow, TruncationOverflow); |
3241 | Result = ResultTrunc; |
3242 | } |
3243 | |
3244 | // Finally, store the result using the pointer. |
3245 | bool isVolatile = |
3246 | ResultArg->getType()->getPointeeType().isVolatileQualified(); |
3247 | Builder.CreateStore(EmitToMemory(Result, ResultQTy), ResultPtr, isVolatile); |
3248 | |
3249 | return RValue::get(Overflow); |
3250 | } |
3251 | |
3252 | case Builtin::BI__builtin_uadd_overflow: |
3253 | case Builtin::BI__builtin_uaddl_overflow: |
3254 | case Builtin::BI__builtin_uaddll_overflow: |
3255 | case Builtin::BI__builtin_usub_overflow: |
3256 | case Builtin::BI__builtin_usubl_overflow: |
3257 | case Builtin::BI__builtin_usubll_overflow: |
3258 | case Builtin::BI__builtin_umul_overflow: |
3259 | case Builtin::BI__builtin_umull_overflow: |
3260 | case Builtin::BI__builtin_umulll_overflow: |
3261 | case Builtin::BI__builtin_sadd_overflow: |
3262 | case Builtin::BI__builtin_saddl_overflow: |
3263 | case Builtin::BI__builtin_saddll_overflow: |
3264 | case Builtin::BI__builtin_ssub_overflow: |
3265 | case Builtin::BI__builtin_ssubl_overflow: |
3266 | case Builtin::BI__builtin_ssubll_overflow: |
3267 | case Builtin::BI__builtin_smul_overflow: |
3268 | case Builtin::BI__builtin_smull_overflow: |
3269 | case Builtin::BI__builtin_smulll_overflow: { |
3270 | |
3271 | // We translate all of these builtins directly to the relevant llvm IR node. |
3272 | |
3273 | // Scalarize our inputs. |
3274 | llvm::Value *X = EmitScalarExpr(E->getArg(0)); |
3275 | llvm::Value *Y = EmitScalarExpr(E->getArg(1)); |
3276 | Address SumOutPtr = EmitPointerWithAlignment(E->getArg(2)); |
3277 | |
3278 | // Decide which of the overflow intrinsics we are lowering to: |
3279 | llvm::Intrinsic::ID IntrinsicId; |
3280 | switch (BuiltinID) { |
3281 | default: llvm_unreachable("Unknown overflow builtin id.")::llvm::llvm_unreachable_internal("Unknown overflow builtin id." , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 3281); |
3282 | case Builtin::BI__builtin_uadd_overflow: |
3283 | case Builtin::BI__builtin_uaddl_overflow: |
3284 | case Builtin::BI__builtin_uaddll_overflow: |
3285 | IntrinsicId = llvm::Intrinsic::uadd_with_overflow; |
3286 | break; |
3287 | case Builtin::BI__builtin_usub_overflow: |
3288 | case Builtin::BI__builtin_usubl_overflow: |
3289 | case Builtin::BI__builtin_usubll_overflow: |
3290 | IntrinsicId = llvm::Intrinsic::usub_with_overflow; |
3291 | break; |
3292 | case Builtin::BI__builtin_umul_overflow: |
3293 | case Builtin::BI__builtin_umull_overflow: |
3294 | case Builtin::BI__builtin_umulll_overflow: |
3295 | IntrinsicId = llvm::Intrinsic::umul_with_overflow; |
3296 | break; |
3297 | case Builtin::BI__builtin_sadd_overflow: |
3298 | case Builtin::BI__builtin_saddl_overflow: |
3299 | case Builtin::BI__builtin_saddll_overflow: |
3300 | IntrinsicId = llvm::Intrinsic::sadd_with_overflow; |
3301 | break; |
3302 | case Builtin::BI__builtin_ssub_overflow: |
3303 | case Builtin::BI__builtin_ssubl_overflow: |
3304 | case Builtin::BI__builtin_ssubll_overflow: |
3305 | IntrinsicId = llvm::Intrinsic::ssub_with_overflow; |
3306 | break; |
3307 | case Builtin::BI__builtin_smul_overflow: |
3308 | case Builtin::BI__builtin_smull_overflow: |
3309 | case Builtin::BI__builtin_smulll_overflow: |
3310 | IntrinsicId = llvm::Intrinsic::smul_with_overflow; |
3311 | break; |
3312 | } |
3313 | |
3314 | |
3315 | llvm::Value *Carry; |
3316 | llvm::Value *Sum = EmitOverflowIntrinsic(*this, IntrinsicId, X, Y, Carry); |
3317 | Builder.CreateStore(Sum, SumOutPtr); |
3318 | |
3319 | return RValue::get(Carry); |
3320 | } |
3321 | case Builtin::BI__builtin_addressof: |
3322 | return RValue::get(EmitLValue(E->getArg(0)).getPointer()); |
3323 | case Builtin::BI__builtin_operator_new: |
3324 | return EmitBuiltinNewDeleteCall( |
3325 | E->getCallee()->getType()->castAs<FunctionProtoType>(), E, false); |
3326 | case Builtin::BI__builtin_operator_delete: |
3327 | return EmitBuiltinNewDeleteCall( |
3328 | E->getCallee()->getType()->castAs<FunctionProtoType>(), E, true); |
3329 | |
3330 | case Builtin::BI__noop: |
3331 | // __noop always evaluates to an integer literal zero. |
3332 | return RValue::get(ConstantInt::get(IntTy, 0)); |
3333 | case Builtin::BI__builtin_call_with_static_chain: { |
3334 | const CallExpr *Call = cast<CallExpr>(E->getArg(0)); |
3335 | const Expr *Chain = E->getArg(1); |
3336 | return EmitCall(Call->getCallee()->getType(), |
3337 | EmitCallee(Call->getCallee()), Call, ReturnValue, |
3338 | EmitScalarExpr(Chain)); |
3339 | } |
3340 | case Builtin::BI_InterlockedExchange8: |
3341 | case Builtin::BI_InterlockedExchange16: |
3342 | case Builtin::BI_InterlockedExchange: |
3343 | case Builtin::BI_InterlockedExchangePointer: |
3344 | return RValue::get( |
3345 | EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchange, E)); |
3346 | case Builtin::BI_InterlockedCompareExchangePointer: |
3347 | case Builtin::BI_InterlockedCompareExchangePointer_nf: { |
3348 | llvm::Type *RTy; |
3349 | llvm::IntegerType *IntType = |
3350 | IntegerType::get(getLLVMContext(), |
3351 | getContext().getTypeSize(E->getType())); |
3352 | llvm::Type *IntPtrType = IntType->getPointerTo(); |
3353 | |
3354 | llvm::Value *Destination = |
3355 | Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), IntPtrType); |
3356 | |
3357 | llvm::Value *Exchange = EmitScalarExpr(E->getArg(1)); |
3358 | RTy = Exchange->getType(); |
3359 | Exchange = Builder.CreatePtrToInt(Exchange, IntType); |
3360 | |
3361 | llvm::Value *Comparand = |
3362 | Builder.CreatePtrToInt(EmitScalarExpr(E->getArg(2)), IntType); |
3363 | |
3364 | auto Ordering = |
3365 | BuiltinID == Builtin::BI_InterlockedCompareExchangePointer_nf ? |
3366 | AtomicOrdering::Monotonic : AtomicOrdering::SequentiallyConsistent; |
3367 | |
3368 | auto Result = Builder.CreateAtomicCmpXchg(Destination, Comparand, Exchange, |
3369 | Ordering, Ordering); |
3370 | Result->setVolatile(true); |
3371 | |
3372 | return RValue::get(Builder.CreateIntToPtr(Builder.CreateExtractValue(Result, |
3373 | 0), |
3374 | RTy)); |
3375 | } |
3376 | case Builtin::BI_InterlockedCompareExchange8: |
3377 | case Builtin::BI_InterlockedCompareExchange16: |
3378 | case Builtin::BI_InterlockedCompareExchange: |
3379 | case Builtin::BI_InterlockedCompareExchange64: |
3380 | return RValue::get(EmitAtomicCmpXchgForMSIntrin(*this, E)); |
3381 | case Builtin::BI_InterlockedIncrement16: |
3382 | case Builtin::BI_InterlockedIncrement: |
3383 | return RValue::get( |
3384 | EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedIncrement, E)); |
3385 | case Builtin::BI_InterlockedDecrement16: |
3386 | case Builtin::BI_InterlockedDecrement: |
3387 | return RValue::get( |
3388 | EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedDecrement, E)); |
3389 | case Builtin::BI_InterlockedAnd8: |
3390 | case Builtin::BI_InterlockedAnd16: |
3391 | case Builtin::BI_InterlockedAnd: |
3392 | return RValue::get(EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedAnd, E)); |
3393 | case Builtin::BI_InterlockedExchangeAdd8: |
3394 | case Builtin::BI_InterlockedExchangeAdd16: |
3395 | case Builtin::BI_InterlockedExchangeAdd: |
3396 | return RValue::get( |
3397 | EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeAdd, E)); |
3398 | case Builtin::BI_InterlockedExchangeSub8: |
3399 | case Builtin::BI_InterlockedExchangeSub16: |
3400 | case Builtin::BI_InterlockedExchangeSub: |
3401 | return RValue::get( |
3402 | EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeSub, E)); |
3403 | case Builtin::BI_InterlockedOr8: |
3404 | case Builtin::BI_InterlockedOr16: |
3405 | case Builtin::BI_InterlockedOr: |
3406 | return RValue::get(EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedOr, E)); |
3407 | case Builtin::BI_InterlockedXor8: |
3408 | case Builtin::BI_InterlockedXor16: |
3409 | case Builtin::BI_InterlockedXor: |
3410 | return RValue::get(EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedXor, E)); |
3411 | |
3412 | case Builtin::BI_bittest64: |
3413 | case Builtin::BI_bittest: |
3414 | case Builtin::BI_bittestandcomplement64: |
3415 | case Builtin::BI_bittestandcomplement: |
3416 | case Builtin::BI_bittestandreset64: |
3417 | case Builtin::BI_bittestandreset: |
3418 | case Builtin::BI_bittestandset64: |
3419 | case Builtin::BI_bittestandset: |
3420 | case Builtin::BI_interlockedbittestandreset: |
3421 | case Builtin::BI_interlockedbittestandreset64: |
3422 | case Builtin::BI_interlockedbittestandset64: |
3423 | case Builtin::BI_interlockedbittestandset: |
3424 | case Builtin::BI_interlockedbittestandset_acq: |
3425 | case Builtin::BI_interlockedbittestandset_rel: |
3426 | case Builtin::BI_interlockedbittestandset_nf: |
3427 | case Builtin::BI_interlockedbittestandreset_acq: |
3428 | case Builtin::BI_interlockedbittestandreset_rel: |
3429 | case Builtin::BI_interlockedbittestandreset_nf: |
3430 | return RValue::get(EmitBitTestIntrinsic(*this, BuiltinID, E)); |
3431 | |
3432 | // These builtins exist to emit regular volatile loads and stores not |
3433 | // affected by the -fms-volatile setting. |
3434 | case Builtin::BI__iso_volatile_load8: |
3435 | case Builtin::BI__iso_volatile_load16: |
3436 | case Builtin::BI__iso_volatile_load32: |
3437 | case Builtin::BI__iso_volatile_load64: |
3438 | return RValue::get(EmitISOVolatileLoad(*this, E)); |
3439 | case Builtin::BI__iso_volatile_store8: |
3440 | case Builtin::BI__iso_volatile_store16: |
3441 | case Builtin::BI__iso_volatile_store32: |
3442 | case Builtin::BI__iso_volatile_store64: |
3443 | return RValue::get(EmitISOVolatileStore(*this, E)); |
3444 | |
3445 | case Builtin::BI__exception_code: |
3446 | case Builtin::BI_exception_code: |
3447 | return RValue::get(EmitSEHExceptionCode()); |
3448 | case Builtin::BI__exception_info: |
3449 | case Builtin::BI_exception_info: |
3450 | return RValue::get(EmitSEHExceptionInfo()); |
3451 | case Builtin::BI__abnormal_termination: |
3452 | case Builtin::BI_abnormal_termination: |
3453 | return RValue::get(EmitSEHAbnormalTermination()); |
3454 | case Builtin::BI_setjmpex: |
3455 | if (getTarget().getTriple().isOSMSVCRT()) |
3456 | return EmitMSVCRTSetJmp(*this, MSVCSetJmpKind::_setjmpex, E); |
3457 | break; |
3458 | case Builtin::BI_setjmp: |
3459 | if (getTarget().getTriple().isOSMSVCRT()) { |
3460 | if (getTarget().getTriple().getArch() == llvm::Triple::x86) |
3461 | return EmitMSVCRTSetJmp(*this, MSVCSetJmpKind::_setjmp3, E); |
3462 | else if (getTarget().getTriple().getArch() == llvm::Triple::aarch64) |
3463 | return EmitMSVCRTSetJmp(*this, MSVCSetJmpKind::_setjmpex, E); |
3464 | return EmitMSVCRTSetJmp(*this, MSVCSetJmpKind::_setjmp, E); |
3465 | } |
3466 | break; |
3467 | |
3468 | case Builtin::BI__GetExceptionInfo: { |
3469 | if (llvm::GlobalVariable *GV = |
3470 | CGM.getCXXABI().getThrowInfo(FD->getParamDecl(0)->getType())) |
3471 | return RValue::get(llvm::ConstantExpr::getBitCast(GV, CGM.Int8PtrTy)); |
3472 | break; |
3473 | } |
3474 | |
3475 | case Builtin::BI__fastfail: |
3476 | return RValue::get(EmitMSVCBuiltinExpr(MSVCIntrin::__fastfail, E)); |
3477 | |
3478 | case Builtin::BI__builtin_coro_size: { |
3479 | auto & Context = getContext(); |
3480 | auto SizeTy = Context.getSizeType(); |
3481 | auto T = Builder.getIntNTy(Context.getTypeSize(SizeTy)); |
3482 | Function *F = CGM.getIntrinsic(Intrinsic::coro_size, T); |
3483 | return RValue::get(Builder.CreateCall(F)); |
3484 | } |
3485 | |
3486 | case Builtin::BI__builtin_coro_id: |
3487 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_id); |
3488 | case Builtin::BI__builtin_coro_promise: |
3489 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_promise); |
3490 | case Builtin::BI__builtin_coro_resume: |
3491 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_resume); |
3492 | case Builtin::BI__builtin_coro_frame: |
3493 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_frame); |
3494 | case Builtin::BI__builtin_coro_noop: |
3495 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_noop); |
3496 | case Builtin::BI__builtin_coro_free: |
3497 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_free); |
3498 | case Builtin::BI__builtin_coro_destroy: |
3499 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_destroy); |
3500 | case Builtin::BI__builtin_coro_done: |
3501 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_done); |
3502 | case Builtin::BI__builtin_coro_alloc: |
3503 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_alloc); |
3504 | case Builtin::BI__builtin_coro_begin: |
3505 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_begin); |
3506 | case Builtin::BI__builtin_coro_end: |
3507 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_end); |
3508 | case Builtin::BI__builtin_coro_suspend: |
3509 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_suspend); |
3510 | case Builtin::BI__builtin_coro_param: |
3511 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_param); |
3512 | |
3513 | // OpenCL v2.0 s6.13.16.2, Built-in pipe read and write functions |
3514 | case Builtin::BIread_pipe: |
3515 | case Builtin::BIwrite_pipe: { |
3516 | Value *Arg0 = EmitScalarExpr(E->getArg(0)), |
3517 | *Arg1 = EmitScalarExpr(E->getArg(1)); |
3518 | CGOpenCLRuntime OpenCLRT(CGM); |
3519 | Value *PacketSize = OpenCLRT.getPipeElemSize(E->getArg(0)); |
3520 | Value *PacketAlign = OpenCLRT.getPipeElemAlign(E->getArg(0)); |
3521 | |
3522 | // Type of the generic packet parameter. |
3523 | unsigned GenericAS = |
3524 | getContext().getTargetAddressSpace(LangAS::opencl_generic); |
3525 | llvm::Type *I8PTy = llvm::PointerType::get( |
3526 | llvm::Type::getInt8Ty(getLLVMContext()), GenericAS); |
3527 | |
3528 | // Testing which overloaded version we should generate the call for. |
3529 | if (2U == E->getNumArgs()) { |
3530 | const char *Name = (BuiltinID == Builtin::BIread_pipe) ? "__read_pipe_2" |
3531 | : "__write_pipe_2"; |
3532 | // Creating a generic function type to be able to call with any builtin or |
3533 | // user defined type. |
3534 | llvm::Type *ArgTys[] = {Arg0->getType(), I8PTy, Int32Ty, Int32Ty}; |
3535 | llvm::FunctionType *FTy = llvm::FunctionType::get( |
3536 | Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); |
3537 | Value *BCast = Builder.CreatePointerCast(Arg1, I8PTy); |
3538 | return RValue::get( |
3539 | Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), |
3540 | {Arg0, BCast, PacketSize, PacketAlign})); |
3541 | } else { |
3542 | assert(4 == E->getNumArgs() &&((4 == E->getNumArgs() && "Illegal number of parameters to pipe function" ) ? static_cast<void> (0) : __assert_fail ("4 == E->getNumArgs() && \"Illegal number of parameters to pipe function\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 3543, __PRETTY_FUNCTION__)) |
3543 | "Illegal number of parameters to pipe function")((4 == E->getNumArgs() && "Illegal number of parameters to pipe function" ) ? static_cast<void> (0) : __assert_fail ("4 == E->getNumArgs() && \"Illegal number of parameters to pipe function\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 3543, __PRETTY_FUNCTION__)); |
3544 | const char *Name = (BuiltinID == Builtin::BIread_pipe) ? "__read_pipe_4" |
3545 | : "__write_pipe_4"; |
3546 | |
3547 | llvm::Type *ArgTys[] = {Arg0->getType(), Arg1->getType(), Int32Ty, I8PTy, |
3548 | Int32Ty, Int32Ty}; |
3549 | Value *Arg2 = EmitScalarExpr(E->getArg(2)), |
3550 | *Arg3 = EmitScalarExpr(E->getArg(3)); |
3551 | llvm::FunctionType *FTy = llvm::FunctionType::get( |
3552 | Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); |
3553 | Value *BCast = Builder.CreatePointerCast(Arg3, I8PTy); |
3554 | // We know the third argument is an integer type, but we may need to cast |
3555 | // it to i32. |
3556 | if (Arg2->getType() != Int32Ty) |
3557 | Arg2 = Builder.CreateZExtOrTrunc(Arg2, Int32Ty); |
3558 | return RValue::get(Builder.CreateCall( |
3559 | CGM.CreateRuntimeFunction(FTy, Name), |
3560 | {Arg0, Arg1, Arg2, BCast, PacketSize, PacketAlign})); |
3561 | } |
3562 | } |
3563 | // OpenCL v2.0 s6.13.16 ,s9.17.3.5 - Built-in pipe reserve read and write |
3564 | // functions |
3565 | case Builtin::BIreserve_read_pipe: |
3566 | case Builtin::BIreserve_write_pipe: |
3567 | case Builtin::BIwork_group_reserve_read_pipe: |
3568 | case Builtin::BIwork_group_reserve_write_pipe: |
3569 | case Builtin::BIsub_group_reserve_read_pipe: |
3570 | case Builtin::BIsub_group_reserve_write_pipe: { |
3571 | // Composing the mangled name for the function. |
3572 | const char *Name; |
3573 | if (BuiltinID == Builtin::BIreserve_read_pipe) |
3574 | Name = "__reserve_read_pipe"; |
3575 | else if (BuiltinID == Builtin::BIreserve_write_pipe) |
3576 | Name = "__reserve_write_pipe"; |
3577 | else if (BuiltinID == Builtin::BIwork_group_reserve_read_pipe) |
3578 | Name = "__work_group_reserve_read_pipe"; |
3579 | else if (BuiltinID == Builtin::BIwork_group_reserve_write_pipe) |
3580 | Name = "__work_group_reserve_write_pipe"; |
3581 | else if (BuiltinID == Builtin::BIsub_group_reserve_read_pipe) |
3582 | Name = "__sub_group_reserve_read_pipe"; |
3583 | else |
3584 | Name = "__sub_group_reserve_write_pipe"; |
3585 | |
3586 | Value *Arg0 = EmitScalarExpr(E->getArg(0)), |
3587 | *Arg1 = EmitScalarExpr(E->getArg(1)); |
3588 | llvm::Type *ReservedIDTy = ConvertType(getContext().OCLReserveIDTy); |
3589 | CGOpenCLRuntime OpenCLRT(CGM); |
3590 | Value *PacketSize = OpenCLRT.getPipeElemSize(E->getArg(0)); |
3591 | Value *PacketAlign = OpenCLRT.getPipeElemAlign(E->getArg(0)); |
3592 | |
3593 | // Building the generic function prototype. |
3594 | llvm::Type *ArgTys[] = {Arg0->getType(), Int32Ty, Int32Ty, Int32Ty}; |
3595 | llvm::FunctionType *FTy = llvm::FunctionType::get( |
3596 | ReservedIDTy, llvm::ArrayRef<llvm::Type *>(ArgTys), false); |
3597 | // We know the second argument is an integer type, but we may need to cast |
3598 | // it to i32. |
3599 | if (Arg1->getType() != Int32Ty) |
3600 | Arg1 = Builder.CreateZExtOrTrunc(Arg1, Int32Ty); |
3601 | return RValue::get( |
3602 | Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), |
3603 | {Arg0, Arg1, PacketSize, PacketAlign})); |
3604 | } |
3605 | // OpenCL v2.0 s6.13.16, s9.17.3.5 - Built-in pipe commit read and write |
3606 | // functions |
3607 | case Builtin::BIcommit_read_pipe: |
3608 | case Builtin::BIcommit_write_pipe: |
3609 | case Builtin::BIwork_group_commit_read_pipe: |
3610 | case Builtin::BIwork_group_commit_write_pipe: |
3611 | case Builtin::BIsub_group_commit_read_pipe: |
3612 | case Builtin::BIsub_group_commit_write_pipe: { |
3613 | const char *Name; |
3614 | if (BuiltinID == Builtin::BIcommit_read_pipe) |
3615 | Name = "__commit_read_pipe"; |
3616 | else if (BuiltinID == Builtin::BIcommit_write_pipe) |
3617 | Name = "__commit_write_pipe"; |
3618 | else if (BuiltinID == Builtin::BIwork_group_commit_read_pipe) |
3619 | Name = "__work_group_commit_read_pipe"; |
3620 | else if (BuiltinID == Builtin::BIwork_group_commit_write_pipe) |
3621 | Name = "__work_group_commit_write_pipe"; |
3622 | else if (BuiltinID == Builtin::BIsub_group_commit_read_pipe) |
3623 | Name = "__sub_group_commit_read_pipe"; |
3624 | else |
3625 | Name = "__sub_group_commit_write_pipe"; |
3626 | |
3627 | Value *Arg0 = EmitScalarExpr(E->getArg(0)), |
3628 | *Arg1 = EmitScalarExpr(E->getArg(1)); |
3629 | CGOpenCLRuntime OpenCLRT(CGM); |
3630 | Value *PacketSize = OpenCLRT.getPipeElemSize(E->getArg(0)); |
3631 | Value *PacketAlign = OpenCLRT.getPipeElemAlign(E->getArg(0)); |
3632 | |
3633 | // Building the generic function prototype. |
3634 | llvm::Type *ArgTys[] = {Arg0->getType(), Arg1->getType(), Int32Ty, Int32Ty}; |
3635 | llvm::FunctionType *FTy = |
3636 | llvm::FunctionType::get(llvm::Type::getVoidTy(getLLVMContext()), |
3637 | llvm::ArrayRef<llvm::Type *>(ArgTys), false); |
3638 | |
3639 | return RValue::get( |
3640 | Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), |
3641 | {Arg0, Arg1, PacketSize, PacketAlign})); |
3642 | } |
3643 | // OpenCL v2.0 s6.13.16.4 Built-in pipe query functions |
3644 | case Builtin::BIget_pipe_num_packets: |
3645 | case Builtin::BIget_pipe_max_packets: { |
3646 | const char *BaseName; |
3647 | const PipeType *PipeTy = E->getArg(0)->getType()->getAs<PipeType>(); |
3648 | if (BuiltinID == Builtin::BIget_pipe_num_packets) |
3649 | BaseName = "__get_pipe_num_packets"; |
3650 | else |
3651 | BaseName = "__get_pipe_max_packets"; |
3652 | auto Name = std::string(BaseName) + |
3653 | std::string(PipeTy->isReadOnly() ? "_ro" : "_wo"); |
3654 | |
3655 | // Building the generic function prototype. |
3656 | Value *Arg0 = EmitScalarExpr(E->getArg(0)); |
3657 | CGOpenCLRuntime OpenCLRT(CGM); |
3658 | Value *PacketSize = OpenCLRT.getPipeElemSize(E->getArg(0)); |
3659 | Value *PacketAlign = OpenCLRT.getPipeElemAlign(E->getArg(0)); |
3660 | llvm::Type *ArgTys[] = {Arg0->getType(), Int32Ty, Int32Ty}; |
3661 | llvm::FunctionType *FTy = llvm::FunctionType::get( |
3662 | Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); |
3663 | |
3664 | return RValue::get(Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), |
3665 | {Arg0, PacketSize, PacketAlign})); |
3666 | } |
3667 | |
3668 | // OpenCL v2.0 s6.13.9 - Address space qualifier functions. |
3669 | case Builtin::BIto_global: |
3670 | case Builtin::BIto_local: |
3671 | case Builtin::BIto_private: { |
3672 | auto Arg0 = EmitScalarExpr(E->getArg(0)); |
3673 | auto NewArgT = llvm::PointerType::get(Int8Ty, |
3674 | CGM.getContext().getTargetAddressSpace(LangAS::opencl_generic)); |
3675 | auto NewRetT = llvm::PointerType::get(Int8Ty, |
3676 | CGM.getContext().getTargetAddressSpace( |
3677 | E->getType()->getPointeeType().getAddressSpace())); |
3678 | auto FTy = llvm::FunctionType::get(NewRetT, {NewArgT}, false); |
3679 | llvm::Value *NewArg; |
3680 | if (Arg0->getType()->getPointerAddressSpace() != |
3681 | NewArgT->getPointerAddressSpace()) |
3682 | NewArg = Builder.CreateAddrSpaceCast(Arg0, NewArgT); |
3683 | else |
3684 | NewArg = Builder.CreateBitOrPointerCast(Arg0, NewArgT); |
3685 | auto NewName = std::string("__") + E->getDirectCallee()->getName().str(); |
3686 | auto NewCall = |
3687 | Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, NewName), {NewArg}); |
3688 | return RValue::get(Builder.CreateBitOrPointerCast(NewCall, |
3689 | ConvertType(E->getType()))); |
3690 | } |
3691 | |
3692 | // OpenCL v2.0, s6.13.17 - Enqueue kernel function. |
3693 | // It contains four different overload formats specified in Table 6.13.17.1. |
3694 | case Builtin::BIenqueue_kernel: { |
3695 | StringRef Name; // Generated function call name |
3696 | unsigned NumArgs = E->getNumArgs(); |
3697 | |
3698 | llvm::Type *QueueTy = ConvertType(getContext().OCLQueueTy); |
3699 | llvm::Type *GenericVoidPtrTy = Builder.getInt8PtrTy( |
3700 | getContext().getTargetAddressSpace(LangAS::opencl_generic)); |
3701 | |
3702 | llvm::Value *Queue = EmitScalarExpr(E->getArg(0)); |
3703 | llvm::Value *Flags = EmitScalarExpr(E->getArg(1)); |
3704 | LValue NDRangeL = EmitAggExprToLValue(E->getArg(2)); |
3705 | llvm::Value *Range = NDRangeL.getAddress().getPointer(); |
3706 | llvm::Type *RangeTy = NDRangeL.getAddress().getType(); |
3707 | |
3708 | if (NumArgs == 4) { |
3709 | // The most basic form of the call with parameters: |
3710 | // queue_t, kernel_enqueue_flags_t, ndrange_t, block(void) |
3711 | Name = "__enqueue_kernel_basic"; |
3712 | llvm::Type *ArgTys[] = {QueueTy, Int32Ty, RangeTy, GenericVoidPtrTy, |
3713 | GenericVoidPtrTy}; |
3714 | llvm::FunctionType *FTy = llvm::FunctionType::get( |
3715 | Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); |
3716 | |
3717 | auto Info = |
3718 | CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(3)); |
3719 | llvm::Value *Kernel = |
3720 | Builder.CreatePointerCast(Info.Kernel, GenericVoidPtrTy); |
3721 | llvm::Value *Block = |
3722 | Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy); |
3723 | |
3724 | AttrBuilder B; |
3725 | B.addAttribute(Attribute::ByVal); |
3726 | llvm::AttributeList ByValAttrSet = |
3727 | llvm::AttributeList::get(CGM.getModule().getContext(), 3U, B); |
3728 | |
3729 | auto RTCall = |
3730 | Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name, ByValAttrSet), |
3731 | {Queue, Flags, Range, Kernel, Block}); |
3732 | RTCall->setAttributes(ByValAttrSet); |
3733 | return RValue::get(RTCall); |
3734 | } |
3735 | assert(NumArgs >= 5 && "Invalid enqueue_kernel signature")((NumArgs >= 5 && "Invalid enqueue_kernel signature" ) ? static_cast<void> (0) : __assert_fail ("NumArgs >= 5 && \"Invalid enqueue_kernel signature\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 3735, __PRETTY_FUNCTION__)); |
3736 | |
3737 | // Create a temporary array to hold the sizes of local pointer arguments |
3738 | // for the block. \p First is the position of the first size argument. |
3739 | auto CreateArrayForSizeVar = [=](unsigned First) |
3740 | -> std::tuple<llvm::Value *, llvm::Value *, llvm::Value *> { |
3741 | llvm::APInt ArraySize(32, NumArgs - First); |
3742 | QualType SizeArrayTy = getContext().getConstantArrayType( |
3743 | getContext().getSizeType(), ArraySize, ArrayType::Normal, |
3744 | /*IndexTypeQuals=*/0); |
3745 | auto Tmp = CreateMemTemp(SizeArrayTy, "block_sizes"); |
3746 | llvm::Value *TmpPtr = Tmp.getPointer(); |
3747 | llvm::Value *TmpSize = EmitLifetimeStart( |
3748 | CGM.getDataLayout().getTypeAllocSize(Tmp.getElementType()), TmpPtr); |
3749 | llvm::Value *ElemPtr; |
3750 | // Each of the following arguments specifies the size of the corresponding |
3751 | // argument passed to the enqueued block. |
3752 | auto *Zero = llvm::ConstantInt::get(IntTy, 0); |
3753 | for (unsigned I = First; I < NumArgs; ++I) { |
3754 | auto *Index = llvm::ConstantInt::get(IntTy, I - First); |
3755 | auto *GEP = Builder.CreateGEP(TmpPtr, {Zero, Index}); |
3756 | if (I == First) |
3757 | ElemPtr = GEP; |
3758 | auto *V = |
3759 | Builder.CreateZExtOrTrunc(EmitScalarExpr(E->getArg(I)), SizeTy); |
3760 | Builder.CreateAlignedStore( |
3761 | V, GEP, CGM.getDataLayout().getPrefTypeAlignment(SizeTy)); |
3762 | } |
3763 | return std::tie(ElemPtr, TmpSize, TmpPtr); |
3764 | }; |
3765 | |
3766 | // Could have events and/or varargs. |
3767 | if (E->getArg(3)->getType()->isBlockPointerType()) { |
3768 | // No events passed, but has variadic arguments. |
3769 | Name = "__enqueue_kernel_varargs"; |
3770 | auto Info = |
3771 | CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(3)); |
3772 | llvm::Value *Kernel = |
3773 | Builder.CreatePointerCast(Info.Kernel, GenericVoidPtrTy); |
3774 | auto *Block = Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy); |
3775 | llvm::Value *ElemPtr, *TmpSize, *TmpPtr; |
3776 | std::tie(ElemPtr, TmpSize, TmpPtr) = CreateArrayForSizeVar(4); |
3777 | |
3778 | // Create a vector of the arguments, as well as a constant value to |
3779 | // express to the runtime the number of variadic arguments. |
3780 | std::vector<llvm::Value *> Args = { |
3781 | Queue, Flags, Range, |
3782 | Kernel, Block, ConstantInt::get(IntTy, NumArgs - 4), |
3783 | ElemPtr}; |
3784 | std::vector<llvm::Type *> ArgTys = { |
3785 | QueueTy, IntTy, RangeTy, GenericVoidPtrTy, |
3786 | GenericVoidPtrTy, IntTy, ElemPtr->getType()}; |
3787 | |
3788 | llvm::FunctionType *FTy = llvm::FunctionType::get( |
3789 | Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); |
3790 | auto Call = |
3791 | RValue::get(Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), |
3792 | llvm::ArrayRef<llvm::Value *>(Args))); |
3793 | if (TmpSize) |
3794 | EmitLifetimeEnd(TmpSize, TmpPtr); |
3795 | return Call; |
3796 | } |
3797 | // Any calls now have event arguments passed. |
3798 | if (NumArgs >= 7) { |
3799 | llvm::Type *EventTy = ConvertType(getContext().OCLClkEventTy); |
3800 | llvm::PointerType *EventPtrTy = EventTy->getPointerTo( |
3801 | CGM.getContext().getTargetAddressSpace(LangAS::opencl_generic)); |
3802 | |
3803 | llvm::Value *NumEvents = |
3804 | Builder.CreateZExtOrTrunc(EmitScalarExpr(E->getArg(3)), Int32Ty); |
3805 | |
3806 | // Since SemaOpenCLBuiltinEnqueueKernel allows fifth and sixth arguments |
3807 | // to be a null pointer constant (including `0` literal), we can take it |
3808 | // into account and emit null pointer directly. |
3809 | llvm::Value *EventWaitList = nullptr; |
3810 | if (E->getArg(4)->isNullPointerConstant( |
3811 | getContext(), Expr::NPC_ValueDependentIsNotNull)) { |
3812 | EventWaitList = llvm::ConstantPointerNull::get(EventPtrTy); |
3813 | } else { |
3814 | EventWaitList = E->getArg(4)->getType()->isArrayType() |
3815 | ? EmitArrayToPointerDecay(E->getArg(4)).getPointer() |
3816 | : EmitScalarExpr(E->getArg(4)); |
3817 | // Convert to generic address space. |
3818 | EventWaitList = Builder.CreatePointerCast(EventWaitList, EventPtrTy); |
3819 | } |
3820 | llvm::Value *EventRet = nullptr; |
3821 | if (E->getArg(5)->isNullPointerConstant( |
3822 | getContext(), Expr::NPC_ValueDependentIsNotNull)) { |
3823 | EventRet = llvm::ConstantPointerNull::get(EventPtrTy); |
3824 | } else { |
3825 | EventRet = |
3826 | Builder.CreatePointerCast(EmitScalarExpr(E->getArg(5)), EventPtrTy); |
3827 | } |
3828 | |
3829 | auto Info = |
3830 | CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(6)); |
3831 | llvm::Value *Kernel = |
3832 | Builder.CreatePointerCast(Info.Kernel, GenericVoidPtrTy); |
3833 | llvm::Value *Block = |
3834 | Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy); |
3835 | |
3836 | std::vector<llvm::Type *> ArgTys = { |
3837 | QueueTy, Int32Ty, RangeTy, Int32Ty, |
3838 | EventPtrTy, EventPtrTy, GenericVoidPtrTy, GenericVoidPtrTy}; |
3839 | |
3840 | std::vector<llvm::Value *> Args = {Queue, Flags, Range, |
3841 | NumEvents, EventWaitList, EventRet, |
3842 | Kernel, Block}; |
3843 | |
3844 | if (NumArgs == 7) { |
3845 | // Has events but no variadics. |
3846 | Name = "__enqueue_kernel_basic_events"; |
3847 | llvm::FunctionType *FTy = llvm::FunctionType::get( |
3848 | Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); |
3849 | return RValue::get( |
3850 | Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), |
3851 | llvm::ArrayRef<llvm::Value *>(Args))); |
3852 | } |
3853 | // Has event info and variadics |
3854 | // Pass the number of variadics to the runtime function too. |
3855 | Args.push_back(ConstantInt::get(Int32Ty, NumArgs - 7)); |
3856 | ArgTys.push_back(Int32Ty); |
3857 | Name = "__enqueue_kernel_events_varargs"; |
3858 | |
3859 | llvm::Value *ElemPtr, *TmpSize, *TmpPtr; |
3860 | std::tie(ElemPtr, TmpSize, TmpPtr) = CreateArrayForSizeVar(7); |
3861 | Args.push_back(ElemPtr); |
3862 | ArgTys.push_back(ElemPtr->getType()); |
3863 | |
3864 | llvm::FunctionType *FTy = llvm::FunctionType::get( |
3865 | Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); |
3866 | auto Call = |
3867 | RValue::get(Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), |
3868 | llvm::ArrayRef<llvm::Value *>(Args))); |
3869 | if (TmpSize) |
3870 | EmitLifetimeEnd(TmpSize, TmpPtr); |
3871 | return Call; |
3872 | } |
3873 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
3874 | } |
3875 | // OpenCL v2.0 s6.13.17.6 - Kernel query functions need bitcast of block |
3876 | // parameter. |
3877 | case Builtin::BIget_kernel_work_group_size: { |
3878 | llvm::Type *GenericVoidPtrTy = Builder.getInt8PtrTy( |
3879 | getContext().getTargetAddressSpace(LangAS::opencl_generic)); |
3880 | auto Info = |
3881 | CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(0)); |
3882 | Value *Kernel = Builder.CreatePointerCast(Info.Kernel, GenericVoidPtrTy); |
3883 | Value *Arg = Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy); |
3884 | return RValue::get(Builder.CreateCall( |
3885 | CGM.CreateRuntimeFunction( |
3886 | llvm::FunctionType::get(IntTy, {GenericVoidPtrTy, GenericVoidPtrTy}, |
3887 | false), |
3888 | "__get_kernel_work_group_size_impl"), |
3889 | {Kernel, Arg})); |
3890 | } |
3891 | case Builtin::BIget_kernel_preferred_work_group_size_multiple: { |
3892 | llvm::Type *GenericVoidPtrTy = Builder.getInt8PtrTy( |
3893 | getContext().getTargetAddressSpace(LangAS::opencl_generic)); |
3894 | auto Info = |
3895 | CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(0)); |
3896 | Value *Kernel = Builder.CreatePointerCast(Info.Kernel, GenericVoidPtrTy); |
3897 | Value *Arg = Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy); |
3898 | return RValue::get(Builder.CreateCall( |
3899 | CGM.CreateRuntimeFunction( |
3900 | llvm::FunctionType::get(IntTy, {GenericVoidPtrTy, GenericVoidPtrTy}, |
3901 | false), |
3902 | "__get_kernel_preferred_work_group_size_multiple_impl"), |
3903 | {Kernel, Arg})); |
3904 | } |
3905 | case Builtin::BIget_kernel_max_sub_group_size_for_ndrange: |
3906 | case Builtin::BIget_kernel_sub_group_count_for_ndrange: { |
3907 | llvm::Type *GenericVoidPtrTy = Builder.getInt8PtrTy( |
3908 | getContext().getTargetAddressSpace(LangAS::opencl_generic)); |
3909 | LValue NDRangeL = EmitAggExprToLValue(E->getArg(0)); |
3910 | llvm::Value *NDRange = NDRangeL.getAddress().getPointer(); |
3911 | auto Info = |
3912 | CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(1)); |
3913 | Value *Kernel = Builder.CreatePointerCast(Info.Kernel, GenericVoidPtrTy); |
3914 | Value *Block = Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy); |
3915 | const char *Name = |
3916 | BuiltinID == Builtin::BIget_kernel_max_sub_group_size_for_ndrange |
3917 | ? "__get_kernel_max_sub_group_size_for_ndrange_impl" |
3918 | : "__get_kernel_sub_group_count_for_ndrange_impl"; |
3919 | return RValue::get(Builder.CreateCall( |
3920 | CGM.CreateRuntimeFunction( |
3921 | llvm::FunctionType::get( |
3922 | IntTy, {NDRange->getType(), GenericVoidPtrTy, GenericVoidPtrTy}, |
3923 | false), |
3924 | Name), |
3925 | {NDRange, Kernel, Block})); |
3926 | } |
3927 | |
3928 | case Builtin::BI__builtin_store_half: |
3929 | case Builtin::BI__builtin_store_halff: { |
3930 | Value *Val = EmitScalarExpr(E->getArg(0)); |
3931 | Address Address = EmitPointerWithAlignment(E->getArg(1)); |
3932 | Value *HalfVal = Builder.CreateFPTrunc(Val, Builder.getHalfTy()); |
3933 | return RValue::get(Builder.CreateStore(HalfVal, Address)); |
3934 | } |
3935 | case Builtin::BI__builtin_load_half: { |
3936 | Address Address = EmitPointerWithAlignment(E->getArg(0)); |
3937 | Value *HalfVal = Builder.CreateLoad(Address); |
3938 | return RValue::get(Builder.CreateFPExt(HalfVal, Builder.getDoubleTy())); |
3939 | } |
3940 | case Builtin::BI__builtin_load_halff: { |
3941 | Address Address = EmitPointerWithAlignment(E->getArg(0)); |
3942 | Value *HalfVal = Builder.CreateLoad(Address); |
3943 | return RValue::get(Builder.CreateFPExt(HalfVal, Builder.getFloatTy())); |
3944 | } |
3945 | case Builtin::BIprintf: |
3946 | if (getTarget().getTriple().isNVPTX()) |
3947 | return EmitNVPTXDevicePrintfCallExpr(E, ReturnValue); |
3948 | break; |
3949 | case Builtin::BI__builtin_canonicalize: |
3950 | case Builtin::BI__builtin_canonicalizef: |
3951 | case Builtin::BI__builtin_canonicalizel: |
3952 | return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::canonicalize)); |
3953 | |
3954 | case Builtin::BI__builtin_thread_pointer: { |
3955 | if (!getContext().getTargetInfo().isTLSSupported()) |
3956 | CGM.ErrorUnsupported(E, "__builtin_thread_pointer"); |
3957 | // Fall through - it's already mapped to the intrinsic by GCCBuiltin. |
3958 | break; |
3959 | } |
3960 | case Builtin::BI__builtin_os_log_format: |
3961 | return emitBuiltinOSLogFormat(*E); |
3962 | |
3963 | case Builtin::BI__xray_customevent: { |
3964 | if (!ShouldXRayInstrumentFunction()) |
3965 | return RValue::getIgnored(); |
3966 | |
3967 | if (!CGM.getCodeGenOpts().XRayInstrumentationBundle.has( |
3968 | XRayInstrKind::Custom)) |
3969 | return RValue::getIgnored(); |
3970 | |
3971 | if (const auto *XRayAttr = CurFuncDecl->getAttr<XRayInstrumentAttr>()) |
3972 | if (XRayAttr->neverXRayInstrument() && !AlwaysEmitXRayCustomEvents()) |
3973 | return RValue::getIgnored(); |
3974 | |
3975 | Function *F = CGM.getIntrinsic(Intrinsic::xray_customevent); |
3976 | auto FTy = F->getFunctionType(); |
3977 | auto Arg0 = E->getArg(0); |
3978 | auto Arg0Val = EmitScalarExpr(Arg0); |
3979 | auto Arg0Ty = Arg0->getType(); |
3980 | auto PTy0 = FTy->getParamType(0); |
3981 | if (PTy0 != Arg0Val->getType()) { |
3982 | if (Arg0Ty->isArrayType()) |
3983 | Arg0Val = EmitArrayToPointerDecay(Arg0).getPointer(); |
3984 | else |
3985 | Arg0Val = Builder.CreatePointerCast(Arg0Val, PTy0); |
3986 | } |
3987 | auto Arg1 = EmitScalarExpr(E->getArg(1)); |
3988 | auto PTy1 = FTy->getParamType(1); |
3989 | if (PTy1 != Arg1->getType()) |
3990 | Arg1 = Builder.CreateTruncOrBitCast(Arg1, PTy1); |
3991 | return RValue::get(Builder.CreateCall(F, {Arg0Val, Arg1})); |
3992 | } |
3993 | |
3994 | case Builtin::BI__xray_typedevent: { |
3995 | // TODO: There should be a way to always emit events even if the current |
3996 | // function is not instrumented. Losing events in a stream can cripple |
3997 | // a trace. |
3998 | if (!ShouldXRayInstrumentFunction()) |
3999 | return RValue::getIgnored(); |
4000 | |
4001 | if (!CGM.getCodeGenOpts().XRayInstrumentationBundle.has( |
4002 | XRayInstrKind::Typed)) |
4003 | return RValue::getIgnored(); |
4004 | |
4005 | if (const auto *XRayAttr = CurFuncDecl->getAttr<XRayInstrumentAttr>()) |
4006 | if (XRayAttr->neverXRayInstrument() && !AlwaysEmitXRayTypedEvents()) |
4007 | return RValue::getIgnored(); |
4008 | |
4009 | Function *F = CGM.getIntrinsic(Intrinsic::xray_typedevent); |
4010 | auto FTy = F->getFunctionType(); |
4011 | auto Arg0 = EmitScalarExpr(E->getArg(0)); |
4012 | auto PTy0 = FTy->getParamType(0); |
4013 | if (PTy0 != Arg0->getType()) |
4014 | Arg0 = Builder.CreateTruncOrBitCast(Arg0, PTy0); |
4015 | auto Arg1 = E->getArg(1); |
4016 | auto Arg1Val = EmitScalarExpr(Arg1); |
4017 | auto Arg1Ty = Arg1->getType(); |
4018 | auto PTy1 = FTy->getParamType(1); |
4019 | if (PTy1 != Arg1Val->getType()) { |
4020 | if (Arg1Ty->isArrayType()) |
4021 | Arg1Val = EmitArrayToPointerDecay(Arg1).getPointer(); |
4022 | else |
4023 | Arg1Val = Builder.CreatePointerCast(Arg1Val, PTy1); |
4024 | } |
4025 | auto Arg2 = EmitScalarExpr(E->getArg(2)); |
4026 | auto PTy2 = FTy->getParamType(2); |
4027 | if (PTy2 != Arg2->getType()) |
4028 | Arg2 = Builder.CreateTruncOrBitCast(Arg2, PTy2); |
4029 | return RValue::get(Builder.CreateCall(F, {Arg0, Arg1Val, Arg2})); |
4030 | } |
4031 | |
4032 | case Builtin::BI__builtin_ms_va_start: |
4033 | case Builtin::BI__builtin_ms_va_end: |
4034 | return RValue::get( |
4035 | EmitVAStartEnd(EmitMSVAListRef(E->getArg(0)).getPointer(), |
4036 | BuiltinID == Builtin::BI__builtin_ms_va_start)); |
4037 | |
4038 | case Builtin::BI__builtin_ms_va_copy: { |
4039 | // Lower this manually. We can't reliably determine whether or not any |
4040 | // given va_copy() is for a Win64 va_list from the calling convention |
4041 | // alone, because it's legal to do this from a System V ABI function. |
4042 | // With opaque pointer types, we won't have enough information in LLVM |
4043 | // IR to determine this from the argument types, either. Best to do it |
4044 | // now, while we have enough information. |
4045 | Address DestAddr = EmitMSVAListRef(E->getArg(0)); |
4046 | Address SrcAddr = EmitMSVAListRef(E->getArg(1)); |
4047 | |
4048 | llvm::Type *BPP = Int8PtrPtrTy; |
4049 | |
4050 | DestAddr = Address(Builder.CreateBitCast(DestAddr.getPointer(), BPP, "cp"), |
4051 | DestAddr.getAlignment()); |
4052 | SrcAddr = Address(Builder.CreateBitCast(SrcAddr.getPointer(), BPP, "ap"), |
4053 | SrcAddr.getAlignment()); |
4054 | |
4055 | Value *ArgPtr = Builder.CreateLoad(SrcAddr, "ap.val"); |
4056 | return RValue::get(Builder.CreateStore(ArgPtr, DestAddr)); |
4057 | } |
4058 | } |
4059 | |
4060 | // If this is an alias for a lib function (e.g. __builtin_sin), emit |
4061 | // the call using the normal call path, but using the unmangled |
4062 | // version of the function name. |
4063 | if (getContext().BuiltinInfo.isLibFunction(BuiltinID)) |
4064 | return emitLibraryCall(*this, FD, E, |
4065 | CGM.getBuiltinLibFunction(FD, BuiltinID)); |
4066 | |
4067 | // If this is a predefined lib function (e.g. malloc), emit the call |
4068 | // using exactly the normal call path. |
4069 | if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID)) |
4070 | return emitLibraryCall(*this, FD, E, |
4071 | cast<llvm::Constant>(EmitScalarExpr(E->getCallee()))); |
4072 | |
4073 | // Check that a call to a target specific builtin has the correct target |
4074 | // features. |
4075 | // This is down here to avoid non-target specific builtins, however, if |
4076 | // generic builtins start to require generic target features then we |
4077 | // can move this up to the beginning of the function. |
4078 | checkTargetFeatures(E, FD); |
4079 | |
4080 | if (unsigned VectorWidth = getContext().BuiltinInfo.getRequiredVectorWidth(BuiltinID)) |
4081 | LargestVectorWidth = std::max(LargestVectorWidth, VectorWidth); |
4082 | |
4083 | // See if we have a target specific intrinsic. |
4084 | const char *Name = getContext().BuiltinInfo.getName(BuiltinID); |
4085 | Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic; |
4086 | StringRef Prefix = |
4087 | llvm::Triple::getArchTypePrefix(getTarget().getTriple().getArch()); |
4088 | if (!Prefix.empty()) { |
4089 | IntrinsicID = Intrinsic::getIntrinsicForGCCBuiltin(Prefix.data(), Name); |
4090 | // NOTE we don't need to perform a compatibility flag check here since the |
4091 | // intrinsics are declared in Builtins*.def via LANGBUILTIN which filter the |
4092 | // MS builtins via ALL_MS_LANGUAGES and are filtered earlier. |
4093 | if (IntrinsicID == Intrinsic::not_intrinsic) |
4094 | IntrinsicID = Intrinsic::getIntrinsicForMSBuiltin(Prefix.data(), Name); |
4095 | } |
4096 | |
4097 | if (IntrinsicID != Intrinsic::not_intrinsic) { |
4098 | SmallVector<Value*, 16> Args; |
4099 | |
4100 | // Find out if any arguments are required to be integer constant |
4101 | // expressions. |
4102 | unsigned ICEArguments = 0; |
4103 | ASTContext::GetBuiltinTypeError Error; |
4104 | getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); |
4105 | assert(Error == ASTContext::GE_None && "Should not codegen an error")((Error == ASTContext::GE_None && "Should not codegen an error" ) ? static_cast<void> (0) : __assert_fail ("Error == ASTContext::GE_None && \"Should not codegen an error\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 4105, __PRETTY_FUNCTION__)); |
4106 | |
4107 | Function *F = CGM.getIntrinsic(IntrinsicID); |
4108 | llvm::FunctionType *FTy = F->getFunctionType(); |
4109 | |
4110 | for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) { |
4111 | Value *ArgValue; |
4112 | // If this is a normal argument, just emit it as a scalar. |
4113 | if ((ICEArguments & (1 << i)) == 0) { |
4114 | ArgValue = EmitScalarExpr(E->getArg(i)); |
4115 | } else { |
4116 | // If this is required to be a constant, constant fold it so that we |
4117 | // know that the generated intrinsic gets a ConstantInt. |
4118 | llvm::APSInt Result; |
4119 | bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result,getContext()); |
4120 | assert(IsConst && "Constant arg isn't actually constant?")((IsConst && "Constant arg isn't actually constant?") ? static_cast<void> (0) : __assert_fail ("IsConst && \"Constant arg isn't actually constant?\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 4120, __PRETTY_FUNCTION__)); |
4121 | (void)IsConst; |
4122 | ArgValue = llvm::ConstantInt::get(getLLVMContext(), Result); |
4123 | } |
4124 | |
4125 | // If the intrinsic arg type is different from the builtin arg type |
4126 | // we need to do a bit cast. |
4127 | llvm::Type *PTy = FTy->getParamType(i); |
4128 | if (PTy != ArgValue->getType()) { |
4129 | // XXX - vector of pointers? |
4130 | if (auto *PtrTy = dyn_cast<llvm::PointerType>(PTy)) { |
4131 | if (PtrTy->getAddressSpace() != |
4132 | ArgValue->getType()->getPointerAddressSpace()) { |
4133 | ArgValue = Builder.CreateAddrSpaceCast( |
4134 | ArgValue, |
4135 | ArgValue->getType()->getPointerTo(PtrTy->getAddressSpace())); |
4136 | } |
4137 | } |
4138 | |
4139 | assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) &&((PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) && "Must be able to losslessly bit cast to param") ? static_cast <void> (0) : __assert_fail ("PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) && \"Must be able to losslessly bit cast to param\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 4140, __PRETTY_FUNCTION__)) |
4140 | "Must be able to losslessly bit cast to param")((PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) && "Must be able to losslessly bit cast to param") ? static_cast <void> (0) : __assert_fail ("PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) && \"Must be able to losslessly bit cast to param\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 4140, __PRETTY_FUNCTION__)); |
4141 | ArgValue = Builder.CreateBitCast(ArgValue, PTy); |
4142 | } |
4143 | |
4144 | Args.push_back(ArgValue); |
4145 | } |
4146 | |
4147 | Value *V = Builder.CreateCall(F, Args); |
4148 | QualType BuiltinRetType = E->getType(); |
4149 | |
4150 | llvm::Type *RetTy = VoidTy; |
4151 | if (!BuiltinRetType->isVoidType()) |
4152 | RetTy = ConvertType(BuiltinRetType); |
4153 | |
4154 | if (RetTy != V->getType()) { |
4155 | // XXX - vector of pointers? |
4156 | if (auto *PtrTy = dyn_cast<llvm::PointerType>(RetTy)) { |
4157 | if (PtrTy->getAddressSpace() != V->getType()->getPointerAddressSpace()) { |
4158 | V = Builder.CreateAddrSpaceCast( |
4159 | V, V->getType()->getPointerTo(PtrTy->getAddressSpace())); |
4160 | } |
4161 | } |
4162 | |
4163 | assert(V->getType()->canLosslesslyBitCastTo(RetTy) &&((V->getType()->canLosslesslyBitCastTo(RetTy) && "Must be able to losslessly bit cast result type") ? static_cast <void> (0) : __assert_fail ("V->getType()->canLosslesslyBitCastTo(RetTy) && \"Must be able to losslessly bit cast result type\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 4164, __PRETTY_FUNCTION__)) |
4164 | "Must be able to losslessly bit cast result type")((V->getType()->canLosslesslyBitCastTo(RetTy) && "Must be able to losslessly bit cast result type") ? static_cast <void> (0) : __assert_fail ("V->getType()->canLosslesslyBitCastTo(RetTy) && \"Must be able to losslessly bit cast result type\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 4164, __PRETTY_FUNCTION__)); |
4165 | V = Builder.CreateBitCast(V, RetTy); |
4166 | } |
4167 | |
4168 | return RValue::get(V); |
4169 | } |
4170 | |
4171 | // See if we have a target specific builtin that needs to be lowered. |
4172 | if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E)) |
4173 | return RValue::get(V); |
4174 | |
4175 | ErrorUnsupported(E, "builtin function"); |
4176 | |
4177 | // Unknown builtin, for now just dump it out and return undef. |
4178 | return GetUndefRValue(E->getType()); |
4179 | } |
4180 | |
4181 | static Value *EmitTargetArchBuiltinExpr(CodeGenFunction *CGF, |
4182 | unsigned BuiltinID, const CallExpr *E, |
4183 | llvm::Triple::ArchType Arch) { |
4184 | switch (Arch) { |
4185 | case llvm::Triple::arm: |
4186 | case llvm::Triple::armeb: |
4187 | case llvm::Triple::thumb: |
4188 | case llvm::Triple::thumbeb: |
4189 | return CGF->EmitARMBuiltinExpr(BuiltinID, E, Arch); |
4190 | case llvm::Triple::aarch64: |
4191 | case llvm::Triple::aarch64_be: |
4192 | return CGF->EmitAArch64BuiltinExpr(BuiltinID, E, Arch); |
4193 | case llvm::Triple::x86: |
4194 | case llvm::Triple::x86_64: |
4195 | return CGF->EmitX86BuiltinExpr(BuiltinID, E); |
4196 | case llvm::Triple::ppc: |
4197 | case llvm::Triple::ppc64: |
4198 | case llvm::Triple::ppc64le: |
4199 | return CGF->EmitPPCBuiltinExpr(BuiltinID, E); |
4200 | case llvm::Triple::r600: |
4201 | case llvm::Triple::amdgcn: |
4202 | return CGF->EmitAMDGPUBuiltinExpr(BuiltinID, E); |
4203 | case llvm::Triple::systemz: |
4204 | return CGF->EmitSystemZBuiltinExpr(BuiltinID, E); |
4205 | case llvm::Triple::nvptx: |
4206 | case llvm::Triple::nvptx64: |
4207 | return CGF->EmitNVPTXBuiltinExpr(BuiltinID, E); |
4208 | case llvm::Triple::wasm32: |
4209 | case llvm::Triple::wasm64: |
4210 | return CGF->EmitWebAssemblyBuiltinExpr(BuiltinID, E); |
4211 | case llvm::Triple::hexagon: |
4212 | return CGF->EmitHexagonBuiltinExpr(BuiltinID, E); |
4213 | default: |
4214 | return nullptr; |
4215 | } |
4216 | } |
4217 | |
4218 | Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID, |
4219 | const CallExpr *E) { |
4220 | if (getContext().BuiltinInfo.isAuxBuiltinID(BuiltinID)) { |
4221 | assert(getContext().getAuxTargetInfo() && "Missing aux target info")((getContext().getAuxTargetInfo() && "Missing aux target info" ) ? static_cast<void> (0) : __assert_fail ("getContext().getAuxTargetInfo() && \"Missing aux target info\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 4221, __PRETTY_FUNCTION__)); |
4222 | return EmitTargetArchBuiltinExpr( |
4223 | this, getContext().BuiltinInfo.getAuxBuiltinID(BuiltinID), E, |
4224 | getContext().getAuxTargetInfo()->getTriple().getArch()); |
4225 | } |
4226 | |
4227 | return EmitTargetArchBuiltinExpr(this, BuiltinID, E, |
4228 | getTarget().getTriple().getArch()); |
4229 | } |
4230 | |
4231 | static llvm::VectorType *GetNeonType(CodeGenFunction *CGF, |
4232 | NeonTypeFlags TypeFlags, |
4233 | bool HasLegalHalfType=true, |
4234 | bool V1Ty=false) { |
4235 | int IsQuad = TypeFlags.isQuad(); |
4236 | switch (TypeFlags.getEltType()) { |
4237 | case NeonTypeFlags::Int8: |
4238 | case NeonTypeFlags::Poly8: |
4239 | return llvm::VectorType::get(CGF->Int8Ty, V1Ty ? 1 : (8 << IsQuad)); |
4240 | case NeonTypeFlags::Int16: |
4241 | case NeonTypeFlags::Poly16: |
4242 | return llvm::VectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad)); |
4243 | case NeonTypeFlags::Float16: |
4244 | if (HasLegalHalfType) |
4245 | return llvm::VectorType::get(CGF->HalfTy, V1Ty ? 1 : (4 << IsQuad)); |
4246 | else |
4247 | return llvm::VectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad)); |
4248 | case NeonTypeFlags::Int32: |
4249 | return llvm::VectorType::get(CGF->Int32Ty, V1Ty ? 1 : (2 << IsQuad)); |
4250 | case NeonTypeFlags::Int64: |
4251 | case NeonTypeFlags::Poly64: |
4252 | return llvm::VectorType::get(CGF->Int64Ty, V1Ty ? 1 : (1 << IsQuad)); |
4253 | case NeonTypeFlags::Poly128: |
4254 | // FIXME: i128 and f128 doesn't get fully support in Clang and llvm. |
4255 | // There is a lot of i128 and f128 API missing. |
4256 | // so we use v16i8 to represent poly128 and get pattern matched. |
4257 | return llvm::VectorType::get(CGF->Int8Ty, 16); |
4258 | case NeonTypeFlags::Float32: |
4259 | return llvm::VectorType::get(CGF->FloatTy, V1Ty ? 1 : (2 << IsQuad)); |
4260 | case NeonTypeFlags::Float64: |
4261 | return llvm::VectorType::get(CGF->DoubleTy, V1Ty ? 1 : (1 << IsQuad)); |
4262 | } |
4263 | llvm_unreachable("Unknown vector element type!")::llvm::llvm_unreachable_internal("Unknown vector element type!" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 4263); |
4264 | } |
4265 | |
4266 | static llvm::VectorType *GetFloatNeonType(CodeGenFunction *CGF, |
4267 | NeonTypeFlags IntTypeFlags) { |
4268 | int IsQuad = IntTypeFlags.isQuad(); |
4269 | switch (IntTypeFlags.getEltType()) { |
4270 | case NeonTypeFlags::Int16: |
4271 | return llvm::VectorType::get(CGF->HalfTy, (4 << IsQuad)); |
4272 | case NeonTypeFlags::Int32: |
4273 | return llvm::VectorType::get(CGF->FloatTy, (2 << IsQuad)); |
4274 | case NeonTypeFlags::Int64: |
4275 | return llvm::VectorType::get(CGF->DoubleTy, (1 << IsQuad)); |
4276 | default: |
4277 | llvm_unreachable("Type can't be converted to floating-point!")::llvm::llvm_unreachable_internal("Type can't be converted to floating-point!" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 4277); |
4278 | } |
4279 | } |
4280 | |
4281 | Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) { |
4282 | unsigned nElts = V->getType()->getVectorNumElements(); |
4283 | Value* SV = llvm::ConstantVector::getSplat(nElts, C); |
4284 | return Builder.CreateShuffleVector(V, V, SV, "lane"); |
4285 | } |
4286 | |
4287 | Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops, |
4288 | const char *name, |
4289 | unsigned shift, bool rightshift) { |
4290 | unsigned j = 0; |
4291 | for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); |
4292 | ai != ae; ++ai, ++j) |
4293 | if (shift > 0 && shift == j) |
4294 | Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift); |
4295 | else |
4296 | Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name); |
4297 | |
4298 | return Builder.CreateCall(F, Ops, name); |
4299 | } |
4300 | |
4301 | Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty, |
4302 | bool neg) { |
4303 | int SV = cast<ConstantInt>(V)->getSExtValue(); |
4304 | return ConstantInt::get(Ty, neg ? -SV : SV); |
4305 | } |
4306 | |
4307 | // Right-shift a vector by a constant. |
4308 | Value *CodeGenFunction::EmitNeonRShiftImm(Value *Vec, Value *Shift, |
4309 | llvm::Type *Ty, bool usgn, |
4310 | const char *name) { |
4311 | llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); |
4312 | |
4313 | int ShiftAmt = cast<ConstantInt>(Shift)->getSExtValue(); |
4314 | int EltSize = VTy->getScalarSizeInBits(); |
4315 | |
4316 | Vec = Builder.CreateBitCast(Vec, Ty); |
4317 | |
4318 | // lshr/ashr are undefined when the shift amount is equal to the vector |
4319 | // element size. |
4320 | if (ShiftAmt == EltSize) { |
4321 | if (usgn) { |
4322 | // Right-shifting an unsigned value by its size yields 0. |
4323 | return llvm::ConstantAggregateZero::get(VTy); |
4324 | } else { |
4325 | // Right-shifting a signed value by its size is equivalent |
4326 | // to a shift of size-1. |
4327 | --ShiftAmt; |
4328 | Shift = ConstantInt::get(VTy->getElementType(), ShiftAmt); |
4329 | } |
4330 | } |
4331 | |
4332 | Shift = EmitNeonShiftVector(Shift, Ty, false); |
4333 | if (usgn) |
4334 | return Builder.CreateLShr(Vec, Shift, name); |
4335 | else |
4336 | return Builder.CreateAShr(Vec, Shift, name); |
4337 | } |
4338 | |
4339 | enum { |
4340 | AddRetType = (1 << 0), |
4341 | Add1ArgType = (1 << 1), |
4342 | Add2ArgTypes = (1 << 2), |
4343 | |
4344 | VectorizeRetType = (1 << 3), |
4345 | VectorizeArgTypes = (1 << 4), |
4346 | |
4347 | InventFloatType = (1 << 5), |
4348 | UnsignedAlts = (1 << 6), |
4349 | |
4350 | Use64BitVectors = (1 << 7), |
4351 | Use128BitVectors = (1 << 8), |
4352 | |
4353 | Vectorize1ArgType = Add1ArgType | VectorizeArgTypes, |
4354 | VectorRet = AddRetType | VectorizeRetType, |
4355 | VectorRetGetArgs01 = |
4356 | AddRetType | Add2ArgTypes | VectorizeRetType | VectorizeArgTypes, |
4357 | FpCmpzModifiers = |
4358 | AddRetType | VectorizeRetType | Add1ArgType | InventFloatType |
4359 | }; |
4360 | |
4361 | namespace { |
4362 | struct NeonIntrinsicInfo { |
4363 | const char *NameHint; |
4364 | unsigned BuiltinID; |
4365 | unsigned LLVMIntrinsic; |
4366 | unsigned AltLLVMIntrinsic; |
4367 | unsigned TypeModifier; |
4368 | |
4369 | bool operator<(unsigned RHSBuiltinID) const { |
4370 | return BuiltinID < RHSBuiltinID; |
4371 | } |
4372 | bool operator<(const NeonIntrinsicInfo &TE) const { |
4373 | return BuiltinID < TE.BuiltinID; |
4374 | } |
4375 | }; |
4376 | } // end anonymous namespace |
4377 | |
4378 | #define NEONMAP0(NameBase) \ |
4379 | { #NameBase, NEON::BI__builtin_neon_ ## NameBase, 0, 0, 0 } |
4380 | |
4381 | #define NEONMAP1(NameBase, LLVMIntrinsic, TypeModifier) \ |
4382 | { #NameBase, NEON:: BI__builtin_neon_ ## NameBase, \ |
4383 | Intrinsic::LLVMIntrinsic, 0, TypeModifier } |
4384 | |
4385 | #define NEONMAP2(NameBase, LLVMIntrinsic, AltLLVMIntrinsic, TypeModifier) \ |
4386 | { #NameBase, NEON:: BI__builtin_neon_ ## NameBase, \ |
4387 | Intrinsic::LLVMIntrinsic, Intrinsic::AltLLVMIntrinsic, \ |
4388 | TypeModifier } |
4389 | |
4390 | static const NeonIntrinsicInfo ARMSIMDIntrinsicMap [] = { |
4391 | NEONMAP2(vabd_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), |
4392 | NEONMAP2(vabdq_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), |
4393 | NEONMAP1(vabs_v, arm_neon_vabs, 0), |
4394 | NEONMAP1(vabsq_v, arm_neon_vabs, 0), |
4395 | NEONMAP0(vaddhn_v), |
4396 | NEONMAP1(vaesdq_v, arm_neon_aesd, 0), |
4397 | NEONMAP1(vaeseq_v, arm_neon_aese, 0), |
4398 | NEONMAP1(vaesimcq_v, arm_neon_aesimc, 0), |
4399 | NEONMAP1(vaesmcq_v, arm_neon_aesmc, 0), |
4400 | NEONMAP1(vbsl_v, arm_neon_vbsl, AddRetType), |
4401 | NEONMAP1(vbslq_v, arm_neon_vbsl, AddRetType), |
4402 | NEONMAP1(vcage_v, arm_neon_vacge, 0), |
4403 | NEONMAP1(vcageq_v, arm_neon_vacge, 0), |
4404 | NEONMAP1(vcagt_v, arm_neon_vacgt, 0), |
4405 | NEONMAP1(vcagtq_v, arm_neon_vacgt, 0), |
4406 | NEONMAP1(vcale_v, arm_neon_vacge, 0), |
4407 | NEONMAP1(vcaleq_v, arm_neon_vacge, 0), |
4408 | NEONMAP1(vcalt_v, arm_neon_vacgt, 0), |
4409 | NEONMAP1(vcaltq_v, arm_neon_vacgt, 0), |
4410 | NEONMAP0(vceqz_v), |
4411 | NEONMAP0(vceqzq_v), |
4412 | NEONMAP0(vcgez_v), |
4413 | NEONMAP0(vcgezq_v), |
4414 | NEONMAP0(vcgtz_v), |
4415 | NEONMAP0(vcgtzq_v), |
4416 | NEONMAP0(vclez_v), |
4417 | NEONMAP0(vclezq_v), |
4418 | NEONMAP1(vcls_v, arm_neon_vcls, Add1ArgType), |
4419 | NEONMAP1(vclsq_v, arm_neon_vcls, Add1ArgType), |
4420 | NEONMAP0(vcltz_v), |
4421 | NEONMAP0(vcltzq_v), |
4422 | NEONMAP1(vclz_v, ctlz, Add1ArgType), |
4423 | NEONMAP1(vclzq_v, ctlz, Add1ArgType), |
4424 | NEONMAP1(vcnt_v, ctpop, Add1ArgType), |
4425 | NEONMAP1(vcntq_v, ctpop, Add1ArgType), |
4426 | NEONMAP1(vcvt_f16_f32, arm_neon_vcvtfp2hf, 0), |
4427 | NEONMAP0(vcvt_f16_v), |
4428 | NEONMAP1(vcvt_f32_f16, arm_neon_vcvthf2fp, 0), |
4429 | NEONMAP0(vcvt_f32_v), |
4430 | NEONMAP2(vcvt_n_f16_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), |
4431 | NEONMAP2(vcvt_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), |
4432 | NEONMAP1(vcvt_n_s16_v, arm_neon_vcvtfp2fxs, 0), |
4433 | NEONMAP1(vcvt_n_s32_v, arm_neon_vcvtfp2fxs, 0), |
4434 | NEONMAP1(vcvt_n_s64_v, arm_neon_vcvtfp2fxs, 0), |
4435 | NEONMAP1(vcvt_n_u16_v, arm_neon_vcvtfp2fxu, 0), |
4436 | NEONMAP1(vcvt_n_u32_v, arm_neon_vcvtfp2fxu, 0), |
4437 | NEONMAP1(vcvt_n_u64_v, arm_neon_vcvtfp2fxu, 0), |
4438 | NEONMAP0(vcvt_s16_v), |
4439 | NEONMAP0(vcvt_s32_v), |
4440 | NEONMAP0(vcvt_s64_v), |
4441 | NEONMAP0(vcvt_u16_v), |
4442 | NEONMAP0(vcvt_u32_v), |
4443 | NEONMAP0(vcvt_u64_v), |
4444 | NEONMAP1(vcvta_s16_v, arm_neon_vcvtas, 0), |
4445 | NEONMAP1(vcvta_s32_v, arm_neon_vcvtas, 0), |
4446 | NEONMAP1(vcvta_s64_v, arm_neon_vcvtas, 0), |
4447 | NEONMAP1(vcvta_u16_v, arm_neon_vcvtau, 0), |
4448 | NEONMAP1(vcvta_u32_v, arm_neon_vcvtau, 0), |
4449 | NEONMAP1(vcvta_u64_v, arm_neon_vcvtau, 0), |
4450 | NEONMAP1(vcvtaq_s16_v, arm_neon_vcvtas, 0), |
4451 | NEONMAP1(vcvtaq_s32_v, arm_neon_vcvtas, 0), |
4452 | NEONMAP1(vcvtaq_s64_v, arm_neon_vcvtas, 0), |
4453 | NEONMAP1(vcvtaq_u16_v, arm_neon_vcvtau, 0), |
4454 | NEONMAP1(vcvtaq_u32_v, arm_neon_vcvtau, 0), |
4455 | NEONMAP1(vcvtaq_u64_v, arm_neon_vcvtau, 0), |
4456 | NEONMAP1(vcvtm_s16_v, arm_neon_vcvtms, 0), |
4457 | NEONMAP1(vcvtm_s32_v, arm_neon_vcvtms, 0), |
4458 | NEONMAP1(vcvtm_s64_v, arm_neon_vcvtms, 0), |
4459 | NEONMAP1(vcvtm_u16_v, arm_neon_vcvtmu, 0), |
4460 | NEONMAP1(vcvtm_u32_v, arm_neon_vcvtmu, 0), |
4461 | NEONMAP1(vcvtm_u64_v, arm_neon_vcvtmu, 0), |
4462 | NEONMAP1(vcvtmq_s16_v, arm_neon_vcvtms, 0), |
4463 | NEONMAP1(vcvtmq_s32_v, arm_neon_vcvtms, 0), |
4464 | NEONMAP1(vcvtmq_s64_v, arm_neon_vcvtms, 0), |
4465 | NEONMAP1(vcvtmq_u16_v, arm_neon_vcvtmu, 0), |
4466 | NEONMAP1(vcvtmq_u32_v, arm_neon_vcvtmu, 0), |
4467 | NEONMAP1(vcvtmq_u64_v, arm_neon_vcvtmu, 0), |
4468 | NEONMAP1(vcvtn_s16_v, arm_neon_vcvtns, 0), |
4469 | NEONMAP1(vcvtn_s32_v, arm_neon_vcvtns, 0), |
4470 | NEONMAP1(vcvtn_s64_v, arm_neon_vcvtns, 0), |
4471 | NEONMAP1(vcvtn_u16_v, arm_neon_vcvtnu, 0), |
4472 | NEONMAP1(vcvtn_u32_v, arm_neon_vcvtnu, 0), |
4473 | NEONMAP1(vcvtn_u64_v, arm_neon_vcvtnu, 0), |
4474 | NEONMAP1(vcvtnq_s16_v, arm_neon_vcvtns, 0), |
4475 | NEONMAP1(vcvtnq_s32_v, arm_neon_vcvtns, 0), |
4476 | NEONMAP1(vcvtnq_s64_v, arm_neon_vcvtns, 0), |
4477 | NEONMAP1(vcvtnq_u16_v, arm_neon_vcvtnu, 0), |
4478 | NEONMAP1(vcvtnq_u32_v, arm_neon_vcvtnu, 0), |
4479 | NEONMAP1(vcvtnq_u64_v, arm_neon_vcvtnu, 0), |
4480 | NEONMAP1(vcvtp_s16_v, arm_neon_vcvtps, 0), |
4481 | NEONMAP1(vcvtp_s32_v, arm_neon_vcvtps, 0), |
4482 | NEONMAP1(vcvtp_s64_v, arm_neon_vcvtps, 0), |
4483 | NEONMAP1(vcvtp_u16_v, arm_neon_vcvtpu, 0), |
4484 | NEONMAP1(vcvtp_u32_v, arm_neon_vcvtpu, 0), |
4485 | NEONMAP1(vcvtp_u64_v, arm_neon_vcvtpu, 0), |
4486 | NEONMAP1(vcvtpq_s16_v, arm_neon_vcvtps, 0), |
4487 | NEONMAP1(vcvtpq_s32_v, arm_neon_vcvtps, 0), |
4488 | NEONMAP1(vcvtpq_s64_v, arm_neon_vcvtps, 0), |
4489 | NEONMAP1(vcvtpq_u16_v, arm_neon_vcvtpu, 0), |
4490 | NEONMAP1(vcvtpq_u32_v, arm_neon_vcvtpu, 0), |
4491 | NEONMAP1(vcvtpq_u64_v, arm_neon_vcvtpu, 0), |
4492 | NEONMAP0(vcvtq_f16_v), |
4493 | NEONMAP0(vcvtq_f32_v), |
4494 | NEONMAP2(vcvtq_n_f16_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), |
4495 | NEONMAP2(vcvtq_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), |
4496 | NEONMAP1(vcvtq_n_s16_v, arm_neon_vcvtfp2fxs, 0), |
4497 | NEONMAP1(vcvtq_n_s32_v, arm_neon_vcvtfp2fxs, 0), |
4498 | NEONMAP1(vcvtq_n_s64_v, arm_neon_vcvtfp2fxs, 0), |
4499 | NEONMAP1(vcvtq_n_u16_v, arm_neon_vcvtfp2fxu, 0), |
4500 | NEONMAP1(vcvtq_n_u32_v, arm_neon_vcvtfp2fxu, 0), |
4501 | NEONMAP1(vcvtq_n_u64_v, arm_neon_vcvtfp2fxu, 0), |
4502 | NEONMAP0(vcvtq_s16_v), |
4503 | NEONMAP0(vcvtq_s32_v), |
4504 | NEONMAP0(vcvtq_s64_v), |
4505 | NEONMAP0(vcvtq_u16_v), |
4506 | NEONMAP0(vcvtq_u32_v), |
4507 | NEONMAP0(vcvtq_u64_v), |
4508 | NEONMAP2(vdot_v, arm_neon_udot, arm_neon_sdot, 0), |
4509 | NEONMAP2(vdotq_v, arm_neon_udot, arm_neon_sdot, 0), |
4510 | NEONMAP0(vext_v), |
4511 | NEONMAP0(vextq_v), |
4512 | NEONMAP0(vfma_v), |
4513 | NEONMAP0(vfmaq_v), |
4514 | NEONMAP2(vhadd_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), |
4515 | NEONMAP2(vhaddq_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), |
4516 | NEONMAP2(vhsub_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), |
4517 | NEONMAP2(vhsubq_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), |
4518 | NEONMAP0(vld1_dup_v), |
4519 | NEONMAP1(vld1_v, arm_neon_vld1, 0), |
4520 | NEONMAP1(vld1_x2_v, arm_neon_vld1x2, 0), |
4521 | NEONMAP1(vld1_x3_v, arm_neon_vld1x3, 0), |
4522 | NEONMAP1(vld1_x4_v, arm_neon_vld1x4, 0), |
4523 | NEONMAP0(vld1q_dup_v), |
4524 | NEONMAP1(vld1q_v, arm_neon_vld1, 0), |
4525 | NEONMAP1(vld1q_x2_v, arm_neon_vld1x2, 0), |
4526 | NEONMAP1(vld1q_x3_v, arm_neon_vld1x3, 0), |
4527 | NEONMAP1(vld1q_x4_v, arm_neon_vld1x4, 0), |
4528 | NEONMAP1(vld2_dup_v, arm_neon_vld2dup, 0), |
4529 | NEONMAP1(vld2_lane_v, arm_neon_vld2lane, 0), |
4530 | NEONMAP1(vld2_v, arm_neon_vld2, 0), |
4531 | NEONMAP1(vld2q_dup_v, arm_neon_vld2dup, 0), |
4532 | NEONMAP1(vld2q_lane_v, arm_neon_vld2lane, 0), |
4533 | NEONMAP1(vld2q_v, arm_neon_vld2, 0), |
4534 | NEONMAP1(vld3_dup_v, arm_neon_vld3dup, 0), |
4535 | NEONMAP1(vld3_lane_v, arm_neon_vld3lane, 0), |
4536 | NEONMAP1(vld3_v, arm_neon_vld3, 0), |
4537 | NEONMAP1(vld3q_dup_v, arm_neon_vld3dup, 0), |
4538 | NEONMAP1(vld3q_lane_v, arm_neon_vld3lane, 0), |
4539 | NEONMAP1(vld3q_v, arm_neon_vld3, 0), |
4540 | NEONMAP1(vld4_dup_v, arm_neon_vld4dup, 0), |
4541 | NEONMAP1(vld4_lane_v, arm_neon_vld4lane, 0), |
4542 | NEONMAP1(vld4_v, arm_neon_vld4, 0), |
4543 | NEONMAP1(vld4q_dup_v, arm_neon_vld4dup, 0), |
4544 | NEONMAP1(vld4q_lane_v, arm_neon_vld4lane, 0), |
4545 | NEONMAP1(vld4q_v, arm_neon_vld4, 0), |
4546 | NEONMAP2(vmax_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), |
4547 | NEONMAP1(vmaxnm_v, arm_neon_vmaxnm, Add1ArgType), |
4548 | NEONMAP1(vmaxnmq_v, arm_neon_vmaxnm, Add1ArgType), |
4549 | NEONMAP2(vmaxq_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), |
4550 | NEONMAP2(vmin_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), |
4551 | NEONMAP1(vminnm_v, arm_neon_vminnm, Add1ArgType), |
4552 | NEONMAP1(vminnmq_v, arm_neon_vminnm, Add1ArgType), |
4553 | NEONMAP2(vminq_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), |
4554 | NEONMAP0(vmovl_v), |
4555 | NEONMAP0(vmovn_v), |
4556 | NEONMAP1(vmul_v, arm_neon_vmulp, Add1ArgType), |
4557 | NEONMAP0(vmull_v), |
4558 | NEONMAP1(vmulq_v, arm_neon_vmulp, Add1ArgType), |
4559 | NEONMAP2(vpadal_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), |
4560 | NEONMAP2(vpadalq_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), |
4561 | NEONMAP1(vpadd_v, arm_neon_vpadd, Add1ArgType), |
4562 | NEONMAP2(vpaddl_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), |
4563 | NEONMAP2(vpaddlq_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), |
4564 | NEONMAP1(vpaddq_v, arm_neon_vpadd, Add1ArgType), |
4565 | NEONMAP2(vpmax_v, arm_neon_vpmaxu, arm_neon_vpmaxs, Add1ArgType | UnsignedAlts), |
4566 | NEONMAP2(vpmin_v, arm_neon_vpminu, arm_neon_vpmins, Add1ArgType | UnsignedAlts), |
4567 | NEONMAP1(vqabs_v, arm_neon_vqabs, Add1ArgType), |
4568 | NEONMAP1(vqabsq_v, arm_neon_vqabs, Add1ArgType), |
4569 | NEONMAP2(vqadd_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), |
4570 | NEONMAP2(vqaddq_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), |
4571 | NEONMAP2(vqdmlal_v, arm_neon_vqdmull, arm_neon_vqadds, 0), |
4572 | NEONMAP2(vqdmlsl_v, arm_neon_vqdmull, arm_neon_vqsubs, 0), |
4573 | NEONMAP1(vqdmulh_v, arm_neon_vqdmulh, Add1ArgType), |
4574 | NEONMAP1(vqdmulhq_v, arm_neon_vqdmulh, Add1ArgType), |
4575 | NEONMAP1(vqdmull_v, arm_neon_vqdmull, Add1ArgType), |
4576 | NEONMAP2(vqmovn_v, arm_neon_vqmovnu, arm_neon_vqmovns, Add1ArgType | UnsignedAlts), |
4577 | NEONMAP1(vqmovun_v, arm_neon_vqmovnsu, Add1ArgType), |
4578 | NEONMAP1(vqneg_v, arm_neon_vqneg, Add1ArgType), |
4579 | NEONMAP1(vqnegq_v, arm_neon_vqneg, Add1ArgType), |
4580 | NEONMAP1(vqrdmulh_v, arm_neon_vqrdmulh, Add1ArgType), |
4581 | NEONMAP1(vqrdmulhq_v, arm_neon_vqrdmulh, Add1ArgType), |
4582 | NEONMAP2(vqrshl_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), |
4583 | NEONMAP2(vqrshlq_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), |
4584 | NEONMAP2(vqshl_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), |
4585 | NEONMAP2(vqshl_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), |
4586 | NEONMAP2(vqshlq_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), |
4587 | NEONMAP2(vqshlq_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), |
4588 | NEONMAP1(vqshlu_n_v, arm_neon_vqshiftsu, 0), |
4589 | NEONMAP1(vqshluq_n_v, arm_neon_vqshiftsu, 0), |
4590 | NEONMAP2(vqsub_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), |
4591 | NEONMAP2(vqsubq_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), |
4592 | NEONMAP1(vraddhn_v, arm_neon_vraddhn, Add1ArgType), |
4593 | NEONMAP2(vrecpe_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), |
4594 | NEONMAP2(vrecpeq_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), |
4595 | NEONMAP1(vrecps_v, arm_neon_vrecps, Add1ArgType), |
4596 | NEONMAP1(vrecpsq_v, arm_neon_vrecps, Add1ArgType), |
4597 | NEONMAP2(vrhadd_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), |
4598 | NEONMAP2(vrhaddq_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), |
4599 | NEONMAP1(vrnd_v, arm_neon_vrintz, Add1ArgType), |
4600 | NEONMAP1(vrnda_v, arm_neon_vrinta, Add1ArgType), |
4601 | NEONMAP1(vrndaq_v, arm_neon_vrinta, Add1ArgType), |
4602 | NEONMAP0(vrndi_v), |
4603 | NEONMAP0(vrndiq_v), |
4604 | NEONMAP1(vrndm_v, arm_neon_vrintm, Add1ArgType), |
4605 | NEONMAP1(vrndmq_v, arm_neon_vrintm, Add1ArgType), |
4606 | NEONMAP1(vrndn_v, arm_neon_vrintn, Add1ArgType), |
4607 | NEONMAP1(vrndnq_v, arm_neon_vrintn, Add1ArgType), |
4608 | NEONMAP1(vrndp_v, arm_neon_vrintp, Add1ArgType), |
4609 | NEONMAP1(vrndpq_v, arm_neon_vrintp, Add1ArgType), |
4610 | NEONMAP1(vrndq_v, arm_neon_vrintz, Add1ArgType), |
4611 | NEONMAP1(vrndx_v, arm_neon_vrintx, Add1ArgType), |
4612 | NEONMAP1(vrndxq_v, arm_neon_vrintx, Add1ArgType), |
4613 | NEONMAP2(vrshl_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), |
4614 | NEONMAP2(vrshlq_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), |
4615 | NEONMAP2(vrshr_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), |
4616 | NEONMAP2(vrshrq_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), |
4617 | NEONMAP2(vrsqrte_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), |
4618 | NEONMAP2(vrsqrteq_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), |
4619 | NEONMAP1(vrsqrts_v, arm_neon_vrsqrts, Add1ArgType), |
4620 | NEONMAP1(vrsqrtsq_v, arm_neon_vrsqrts, Add1ArgType), |
4621 | NEONMAP1(vrsubhn_v, arm_neon_vrsubhn, Add1ArgType), |
4622 | NEONMAP1(vsha1su0q_v, arm_neon_sha1su0, 0), |
4623 | NEONMAP1(vsha1su1q_v, arm_neon_sha1su1, 0), |
4624 | NEONMAP1(vsha256h2q_v, arm_neon_sha256h2, 0), |
4625 | NEONMAP1(vsha256hq_v, arm_neon_sha256h, 0), |
4626 | NEONMAP1(vsha256su0q_v, arm_neon_sha256su0, 0), |
4627 | NEONMAP1(vsha256su1q_v, arm_neon_sha256su1, 0), |
4628 | NEONMAP0(vshl_n_v), |
4629 | NEONMAP2(vshl_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), |
4630 | NEONMAP0(vshll_n_v), |
4631 | NEONMAP0(vshlq_n_v), |
4632 | NEONMAP2(vshlq_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), |
4633 | NEONMAP0(vshr_n_v), |
4634 | NEONMAP0(vshrn_n_v), |
4635 | NEONMAP0(vshrq_n_v), |
4636 | NEONMAP1(vst1_v, arm_neon_vst1, 0), |
4637 | NEONMAP1(vst1_x2_v, arm_neon_vst1x2, 0), |
4638 | NEONMAP1(vst1_x3_v, arm_neon_vst1x3, 0), |
4639 | NEONMAP1(vst1_x4_v, arm_neon_vst1x4, 0), |
4640 | NEONMAP1(vst1q_v, arm_neon_vst1, 0), |
4641 | NEONMAP1(vst1q_x2_v, arm_neon_vst1x2, 0), |
4642 | NEONMAP1(vst1q_x3_v, arm_neon_vst1x3, 0), |
4643 | NEONMAP1(vst1q_x4_v, arm_neon_vst1x4, 0), |
4644 | NEONMAP1(vst2_lane_v, arm_neon_vst2lane, 0), |
4645 | NEONMAP1(vst2_v, arm_neon_vst2, 0), |
4646 | NEONMAP1(vst2q_lane_v, arm_neon_vst2lane, 0), |
4647 | NEONMAP1(vst2q_v, arm_neon_vst2, 0), |
4648 | NEONMAP1(vst3_lane_v, arm_neon_vst3lane, 0), |
4649 | NEONMAP1(vst3_v, arm_neon_vst3, 0), |
4650 | NEONMAP1(vst3q_lane_v, arm_neon_vst3lane, 0), |
4651 | NEONMAP1(vst3q_v, arm_neon_vst3, 0), |
4652 | NEONMAP1(vst4_lane_v, arm_neon_vst4lane, 0), |
4653 | NEONMAP1(vst4_v, arm_neon_vst4, 0), |
4654 | NEONMAP1(vst4q_lane_v, arm_neon_vst4lane, 0), |
4655 | NEONMAP1(vst4q_v, arm_neon_vst4, 0), |
4656 | NEONMAP0(vsubhn_v), |
4657 | NEONMAP0(vtrn_v), |
4658 | NEONMAP0(vtrnq_v), |
4659 | NEONMAP0(vtst_v), |
4660 | NEONMAP0(vtstq_v), |
4661 | NEONMAP0(vuzp_v), |
4662 | NEONMAP0(vuzpq_v), |
4663 | NEONMAP0(vzip_v), |
4664 | NEONMAP0(vzipq_v) |
4665 | }; |
4666 | |
4667 | static const NeonIntrinsicInfo AArch64SIMDIntrinsicMap[] = { |
4668 | NEONMAP1(vabs_v, aarch64_neon_abs, 0), |
4669 | NEONMAP1(vabsq_v, aarch64_neon_abs, 0), |
4670 | NEONMAP0(vaddhn_v), |
4671 | NEONMAP1(vaesdq_v, aarch64_crypto_aesd, 0), |
4672 | NEONMAP1(vaeseq_v, aarch64_crypto_aese, 0), |
4673 | NEONMAP1(vaesimcq_v, aarch64_crypto_aesimc, 0), |
4674 | NEONMAP1(vaesmcq_v, aarch64_crypto_aesmc, 0), |
4675 | NEONMAP1(vcage_v, aarch64_neon_facge, 0), |
4676 | NEONMAP1(vcageq_v, aarch64_neon_facge, 0), |
4677 | NEONMAP1(vcagt_v, aarch64_neon_facgt, 0), |
4678 | NEONMAP1(vcagtq_v, aarch64_neon_facgt, 0), |
4679 | NEONMAP1(vcale_v, aarch64_neon_facge, 0), |
4680 | NEONMAP1(vcaleq_v, aarch64_neon_facge, 0), |
4681 | NEONMAP1(vcalt_v, aarch64_neon_facgt, 0), |
4682 | NEONMAP1(vcaltq_v, aarch64_neon_facgt, 0), |
4683 | NEONMAP0(vceqz_v), |
4684 | NEONMAP0(vceqzq_v), |
4685 | NEONMAP0(vcgez_v), |
4686 | NEONMAP0(vcgezq_v), |
4687 | NEONMAP0(vcgtz_v), |
4688 | NEONMAP0(vcgtzq_v), |
4689 | NEONMAP0(vclez_v), |
4690 | NEONMAP0(vclezq_v), |
4691 | NEONMAP1(vcls_v, aarch64_neon_cls, Add1ArgType), |
4692 | NEONMAP1(vclsq_v, aarch64_neon_cls, Add1ArgType), |
4693 | NEONMAP0(vcltz_v), |
4694 | NEONMAP0(vcltzq_v), |
4695 | NEONMAP1(vclz_v, ctlz, Add1ArgType), |
4696 | NEONMAP1(vclzq_v, ctlz, Add1ArgType), |
4697 | NEONMAP1(vcnt_v, ctpop, Add1ArgType), |
4698 | NEONMAP1(vcntq_v, ctpop, Add1ArgType), |
4699 | NEONMAP1(vcvt_f16_f32, aarch64_neon_vcvtfp2hf, 0), |
4700 | NEONMAP0(vcvt_f16_v), |
4701 | NEONMAP1(vcvt_f32_f16, aarch64_neon_vcvthf2fp, 0), |
4702 | NEONMAP0(vcvt_f32_v), |
4703 | NEONMAP2(vcvt_n_f16_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), |
4704 | NEONMAP2(vcvt_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), |
4705 | NEONMAP2(vcvt_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), |
4706 | NEONMAP1(vcvt_n_s16_v, aarch64_neon_vcvtfp2fxs, 0), |
4707 | NEONMAP1(vcvt_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), |
4708 | NEONMAP1(vcvt_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), |
4709 | NEONMAP1(vcvt_n_u16_v, aarch64_neon_vcvtfp2fxu, 0), |
4710 | NEONMAP1(vcvt_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), |
4711 | NEONMAP1(vcvt_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), |
4712 | NEONMAP0(vcvtq_f16_v), |
4713 | NEONMAP0(vcvtq_f32_v), |
4714 | NEONMAP2(vcvtq_n_f16_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), |
4715 | NEONMAP2(vcvtq_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), |
4716 | NEONMAP2(vcvtq_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), |
4717 | NEONMAP1(vcvtq_n_s16_v, aarch64_neon_vcvtfp2fxs, 0), |
4718 | NEONMAP1(vcvtq_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), |
4719 | NEONMAP1(vcvtq_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), |
4720 | NEONMAP1(vcvtq_n_u16_v, aarch64_neon_vcvtfp2fxu, 0), |
4721 | NEONMAP1(vcvtq_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), |
4722 | NEONMAP1(vcvtq_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), |
4723 | NEONMAP1(vcvtx_f32_v, aarch64_neon_fcvtxn, AddRetType | Add1ArgType), |
4724 | NEONMAP2(vdot_v, aarch64_neon_udot, aarch64_neon_sdot, 0), |
4725 | NEONMAP2(vdotq_v, aarch64_neon_udot, aarch64_neon_sdot, 0), |
4726 | NEONMAP0(vext_v), |
4727 | NEONMAP0(vextq_v), |
4728 | NEONMAP0(vfma_v), |
4729 | NEONMAP0(vfmaq_v), |
4730 | NEONMAP1(vfmlal_high_v, aarch64_neon_fmlal2, 0), |
4731 | NEONMAP1(vfmlal_low_v, aarch64_neon_fmlal, 0), |
4732 | NEONMAP1(vfmlalq_high_v, aarch64_neon_fmlal2, 0), |
4733 | NEONMAP1(vfmlalq_low_v, aarch64_neon_fmlal, 0), |
4734 | NEONMAP1(vfmlsl_high_v, aarch64_neon_fmlsl2, 0), |
4735 | NEONMAP1(vfmlsl_low_v, aarch64_neon_fmlsl, 0), |
4736 | NEONMAP1(vfmlslq_high_v, aarch64_neon_fmlsl2, 0), |
4737 | NEONMAP1(vfmlslq_low_v, aarch64_neon_fmlsl, 0), |
4738 | NEONMAP2(vhadd_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), |
4739 | NEONMAP2(vhaddq_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), |
4740 | NEONMAP2(vhsub_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), |
4741 | NEONMAP2(vhsubq_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), |
4742 | NEONMAP1(vld1_x2_v, aarch64_neon_ld1x2, 0), |
4743 | NEONMAP1(vld1_x3_v, aarch64_neon_ld1x3, 0), |
4744 | NEONMAP1(vld1_x4_v, aarch64_neon_ld1x4, 0), |
4745 | NEONMAP1(vld1q_x2_v, aarch64_neon_ld1x2, 0), |
4746 | NEONMAP1(vld1q_x3_v, aarch64_neon_ld1x3, 0), |
4747 | NEONMAP1(vld1q_x4_v, aarch64_neon_ld1x4, 0), |
4748 | NEONMAP0(vmovl_v), |
4749 | NEONMAP0(vmovn_v), |
4750 | NEONMAP1(vmul_v, aarch64_neon_pmul, Add1ArgType), |
4751 | NEONMAP1(vmulq_v, aarch64_neon_pmul, Add1ArgType), |
4752 | NEONMAP1(vpadd_v, aarch64_neon_addp, Add1ArgType), |
4753 | NEONMAP2(vpaddl_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), |
4754 | NEONMAP2(vpaddlq_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), |
4755 | NEONMAP1(vpaddq_v, aarch64_neon_addp, Add1ArgType), |
4756 | NEONMAP1(vqabs_v, aarch64_neon_sqabs, Add1ArgType), |
4757 | NEONMAP1(vqabsq_v, aarch64_neon_sqabs, Add1ArgType), |
4758 | NEONMAP2(vqadd_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), |
4759 | NEONMAP2(vqaddq_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), |
4760 | NEONMAP2(vqdmlal_v, aarch64_neon_sqdmull, aarch64_neon_sqadd, 0), |
4761 | NEONMAP2(vqdmlsl_v, aarch64_neon_sqdmull, aarch64_neon_sqsub, 0), |
4762 | NEONMAP1(vqdmulh_v, aarch64_neon_sqdmulh, Add1ArgType), |
4763 | NEONMAP1(vqdmulhq_v, aarch64_neon_sqdmulh, Add1ArgType), |
4764 | NEONMAP1(vqdmull_v, aarch64_neon_sqdmull, Add1ArgType), |
4765 | NEONMAP2(vqmovn_v, aarch64_neon_uqxtn, aarch64_neon_sqxtn, Add1ArgType | UnsignedAlts), |
4766 | NEONMAP1(vqmovun_v, aarch64_neon_sqxtun, Add1ArgType), |
4767 | NEONMAP1(vqneg_v, aarch64_neon_sqneg, Add1ArgType), |
4768 | NEONMAP1(vqnegq_v, aarch64_neon_sqneg, Add1ArgType), |
4769 | NEONMAP1(vqrdmulh_v, aarch64_neon_sqrdmulh, Add1ArgType), |
4770 | NEONMAP1(vqrdmulhq_v, aarch64_neon_sqrdmulh, Add1ArgType), |
4771 | NEONMAP2(vqrshl_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), |
4772 | NEONMAP2(vqrshlq_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), |
4773 | NEONMAP2(vqshl_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl, UnsignedAlts), |
4774 | NEONMAP2(vqshl_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), |
4775 | NEONMAP2(vqshlq_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl,UnsignedAlts), |
4776 | NEONMAP2(vqshlq_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), |
4777 | NEONMAP1(vqshlu_n_v, aarch64_neon_sqshlu, 0), |
4778 | NEONMAP1(vqshluq_n_v, aarch64_neon_sqshlu, 0), |
4779 | NEONMAP2(vqsub_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), |
4780 | NEONMAP2(vqsubq_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), |
4781 | NEONMAP1(vraddhn_v, aarch64_neon_raddhn, Add1ArgType), |
4782 | NEONMAP2(vrecpe_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), |
4783 | NEONMAP2(vrecpeq_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), |
4784 | NEONMAP1(vrecps_v, aarch64_neon_frecps, Add1ArgType), |
4785 | NEONMAP1(vrecpsq_v, aarch64_neon_frecps, Add1ArgType), |
4786 | NEONMAP2(vrhadd_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), |
4787 | NEONMAP2(vrhaddq_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), |
4788 | NEONMAP0(vrndi_v), |
4789 | NEONMAP0(vrndiq_v), |
4790 | NEONMAP2(vrshl_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), |
4791 | NEONMAP2(vrshlq_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), |
4792 | NEONMAP2(vrshr_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), |
4793 | NEONMAP2(vrshrq_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), |
4794 | NEONMAP2(vrsqrte_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), |
4795 | NEONMAP2(vrsqrteq_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), |
4796 | NEONMAP1(vrsqrts_v, aarch64_neon_frsqrts, Add1ArgType), |
4797 | NEONMAP1(vrsqrtsq_v, aarch64_neon_frsqrts, Add1ArgType), |
4798 | NEONMAP1(vrsubhn_v, aarch64_neon_rsubhn, Add1ArgType), |
4799 | NEONMAP1(vsha1su0q_v, aarch64_crypto_sha1su0, 0), |
4800 | NEONMAP1(vsha1su1q_v, aarch64_crypto_sha1su1, 0), |
4801 | NEONMAP1(vsha256h2q_v, aarch64_crypto_sha256h2, 0), |
4802 | NEONMAP1(vsha256hq_v, aarch64_crypto_sha256h, 0), |
4803 | NEONMAP1(vsha256su0q_v, aarch64_crypto_sha256su0, 0), |
4804 | NEONMAP1(vsha256su1q_v, aarch64_crypto_sha256su1, 0), |
4805 | NEONMAP0(vshl_n_v), |
4806 | NEONMAP2(vshl_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), |
4807 | NEONMAP0(vshll_n_v), |
4808 | NEONMAP0(vshlq_n_v), |
4809 | NEONMAP2(vshlq_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), |
4810 | NEONMAP0(vshr_n_v), |
4811 | NEONMAP0(vshrn_n_v), |
4812 | NEONMAP0(vshrq_n_v), |
4813 | NEONMAP1(vst1_x2_v, aarch64_neon_st1x2, 0), |
4814 | NEONMAP1(vst1_x3_v, aarch64_neon_st1x3, 0), |
4815 | NEONMAP1(vst1_x4_v, aarch64_neon_st1x4, 0), |
4816 | NEONMAP1(vst1q_x2_v, aarch64_neon_st1x2, 0), |
4817 | NEONMAP1(vst1q_x3_v, aarch64_neon_st1x3, 0), |
4818 | NEONMAP1(vst1q_x4_v, aarch64_neon_st1x4, 0), |
4819 | NEONMAP0(vsubhn_v), |
4820 | NEONMAP0(vtst_v), |
4821 | NEONMAP0(vtstq_v), |
4822 | }; |
4823 | |
4824 | static const NeonIntrinsicInfo AArch64SISDIntrinsicMap[] = { |
4825 | NEONMAP1(vabdd_f64, aarch64_sisd_fabd, Add1ArgType), |
4826 | NEONMAP1(vabds_f32, aarch64_sisd_fabd, Add1ArgType), |
4827 | NEONMAP1(vabsd_s64, aarch64_neon_abs, Add1ArgType), |
4828 | NEONMAP1(vaddlv_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), |
4829 | NEONMAP1(vaddlv_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), |
4830 | NEONMAP1(vaddlvq_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), |
4831 | NEONMAP1(vaddlvq_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), |
4832 | NEONMAP1(vaddv_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), |
4833 | NEONMAP1(vaddv_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), |
4834 | NEONMAP1(vaddv_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), |
4835 | NEONMAP1(vaddvq_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), |
4836 | NEONMAP1(vaddvq_f64, aarch64_neon_faddv, AddRetType | Add1ArgType), |
4837 | NEONMAP1(vaddvq_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), |
4838 | NEONMAP1(vaddvq_s64, aarch64_neon_saddv, AddRetType | Add1ArgType), |
4839 | NEONMAP1(vaddvq_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), |
4840 | NEONMAP1(vaddvq_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), |
4841 | NEONMAP1(vcaged_f64, aarch64_neon_facge, AddRetType | Add1ArgType), |
4842 | NEONMAP1(vcages_f32, aarch64_neon_facge, AddRetType | Add1ArgType), |
4843 | NEONMAP1(vcagtd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), |
4844 | NEONMAP1(vcagts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), |
4845 | NEONMAP1(vcaled_f64, aarch64_neon_facge, AddRetType | Add1ArgType), |
4846 | NEONMAP1(vcales_f32, aarch64_neon_facge, AddRetType | Add1ArgType), |
4847 | NEONMAP1(vcaltd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), |
4848 | NEONMAP1(vcalts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), |
4849 | NEONMAP1(vcvtad_s64_f64, aarch64_neon_fcvtas, AddRetType | Add1ArgType), |
4850 | NEONMAP1(vcvtad_u64_f64, aarch64_neon_fcvtau, AddRetType | Add1ArgType), |
4851 | NEONMAP1(vcvtas_s32_f32, aarch64_neon_fcvtas, AddRetType | Add1ArgType), |
4852 | NEONMAP1(vcvtas_u32_f32, aarch64_neon_fcvtau, AddRetType | Add1ArgType), |
4853 | NEONMAP1(vcvtd_n_f64_s64, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), |
4854 | NEONMAP1(vcvtd_n_f64_u64, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), |
4855 | NEONMAP1(vcvtd_n_s64_f64, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), |
4856 | NEONMAP1(vcvtd_n_u64_f64, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), |
4857 | NEONMAP1(vcvtmd_s64_f64, aarch64_neon_fcvtms, AddRetType | Add1ArgType), |
4858 | NEONMAP1(vcvtmd_u64_f64, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), |
4859 | NEONMAP1(vcvtms_s32_f32, aarch64_neon_fcvtms, AddRetType | Add1ArgType), |
4860 | NEONMAP1(vcvtms_u32_f32, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), |
4861 | NEONMAP1(vcvtnd_s64_f64, aarch64_neon_fcvtns, AddRetType | Add1ArgType), |
4862 | NEONMAP1(vcvtnd_u64_f64, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), |
4863 | NEONMAP1(vcvtns_s32_f32, aarch64_neon_fcvtns, AddRetType | Add1ArgType), |
4864 | NEONMAP1(vcvtns_u32_f32, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), |
4865 | NEONMAP1(vcvtpd_s64_f64, aarch64_neon_fcvtps, AddRetType | Add1ArgType), |
4866 | NEONMAP1(vcvtpd_u64_f64, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), |
4867 | NEONMAP1(vcvtps_s32_f32, aarch64_neon_fcvtps, AddRetType | Add1ArgType), |
4868 | NEONMAP1(vcvtps_u32_f32, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), |
4869 | NEONMAP1(vcvts_n_f32_s32, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), |
4870 | NEONMAP1(vcvts_n_f32_u32, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), |
4871 | NEONMAP1(vcvts_n_s32_f32, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), |
4872 | NEONMAP1(vcvts_n_u32_f32, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), |
4873 | NEONMAP1(vcvtxd_f32_f64, aarch64_sisd_fcvtxn, 0), |
4874 | NEONMAP1(vmaxnmv_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), |
4875 | NEONMAP1(vmaxnmvq_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), |
4876 | NEONMAP1(vmaxnmvq_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), |
4877 | NEONMAP1(vmaxv_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), |
4878 | NEONMAP1(vmaxv_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), |
4879 | NEONMAP1(vmaxv_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), |
4880 | NEONMAP1(vmaxvq_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), |
4881 | NEONMAP1(vmaxvq_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), |
4882 | NEONMAP1(vmaxvq_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), |
4883 | NEONMAP1(vmaxvq_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), |
4884 | NEONMAP1(vminnmv_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), |
4885 | NEONMAP1(vminnmvq_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), |
4886 | NEONMAP1(vminnmvq_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), |
4887 | NEONMAP1(vminv_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), |
4888 | NEONMAP1(vminv_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), |
4889 | NEONMAP1(vminv_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), |
4890 | NEONMAP1(vminvq_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), |
4891 | NEONMAP1(vminvq_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), |
4892 | NEONMAP1(vminvq_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), |
4893 | NEONMAP1(vminvq_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), |
4894 | NEONMAP1(vmull_p64, aarch64_neon_pmull64, 0), |
4895 | NEONMAP1(vmulxd_f64, aarch64_neon_fmulx, Add1ArgType), |
4896 | NEONMAP1(vmulxs_f32, aarch64_neon_fmulx, Add1ArgType), |
4897 | NEONMAP1(vpaddd_s64, aarch64_neon_uaddv, AddRetType | Add1ArgType), |
4898 | NEONMAP1(vpaddd_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), |
4899 | NEONMAP1(vpmaxnmqd_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), |
4900 | NEONMAP1(vpmaxnms_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), |
4901 | NEONMAP1(vpmaxqd_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), |
4902 | NEONMAP1(vpmaxs_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), |
4903 | NEONMAP1(vpminnmqd_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), |
4904 | NEONMAP1(vpminnms_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), |
4905 | NEONMAP1(vpminqd_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), |
4906 | NEONMAP1(vpmins_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), |
4907 | NEONMAP1(vqabsb_s8, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), |
4908 | NEONMAP1(vqabsd_s64, aarch64_neon_sqabs, Add1ArgType), |
4909 | NEONMAP1(vqabsh_s16, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), |
4910 | NEONMAP1(vqabss_s32, aarch64_neon_sqabs, Add1ArgType), |
4911 | NEONMAP1(vqaddb_s8, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), |
4912 | NEONMAP1(vqaddb_u8, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), |
4913 | NEONMAP1(vqaddd_s64, aarch64_neon_sqadd, Add1ArgType), |
4914 | NEONMAP1(vqaddd_u64, aarch64_neon_uqadd, Add1ArgType), |
4915 | NEONMAP1(vqaddh_s16, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), |
4916 | NEONMAP1(vqaddh_u16, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), |
4917 | NEONMAP1(vqadds_s32, aarch64_neon_sqadd, Add1ArgType), |
4918 | NEONMAP1(vqadds_u32, aarch64_neon_uqadd, Add1ArgType), |
4919 | NEONMAP1(vqdmulhh_s16, aarch64_neon_sqdmulh, Vectorize1ArgType | Use64BitVectors), |
4920 | NEONMAP1(vqdmulhs_s32, aarch64_neon_sqdmulh, Add1ArgType), |
4921 | NEONMAP1(vqdmullh_s16, aarch64_neon_sqdmull, VectorRet | Use128BitVectors), |
4922 | NEONMAP1(vqdmulls_s32, aarch64_neon_sqdmulls_scalar, 0), |
4923 | NEONMAP1(vqmovnd_s64, aarch64_neon_scalar_sqxtn, AddRetType | Add1ArgType), |
4924 | NEONMAP1(vqmovnd_u64, aarch64_neon_scalar_uqxtn, AddRetType | Add1ArgType), |
4925 | NEONMAP1(vqmovnh_s16, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), |
4926 | NEONMAP1(vqmovnh_u16, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), |
4927 | NEONMAP1(vqmovns_s32, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), |
4928 | NEONMAP1(vqmovns_u32, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), |
4929 | NEONMAP1(vqmovund_s64, aarch64_neon_scalar_sqxtun, AddRetType | Add1ArgType), |
4930 | NEONMAP1(vqmovunh_s16, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), |
4931 | NEONMAP1(vqmovuns_s32, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), |
4932 | NEONMAP1(vqnegb_s8, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), |
4933 | NEONMAP1(vqnegd_s64, aarch64_neon_sqneg, Add1ArgType), |
4934 | NEONMAP1(vqnegh_s16, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), |
4935 | NEONMAP1(vqnegs_s32, aarch64_neon_sqneg, Add1ArgType), |
4936 | NEONMAP1(vqrdmulhh_s16, aarch64_neon_sqrdmulh, Vectorize1ArgType | Use64BitVectors), |
4937 | NEONMAP1(vqrdmulhs_s32, aarch64_neon_sqrdmulh, Add1ArgType), |
4938 | NEONMAP1(vqrshlb_s8, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), |
4939 | NEONMAP1(vqrshlb_u8, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), |
4940 | NEONMAP1(vqrshld_s64, aarch64_neon_sqrshl, Add1ArgType), |
4941 | NEONMAP1(vqrshld_u64, aarch64_neon_uqrshl, Add1ArgType), |
4942 | NEONMAP1(vqrshlh_s16, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), |
4943 | NEONMAP1(vqrshlh_u16, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), |
4944 | NEONMAP1(vqrshls_s32, aarch64_neon_sqrshl, Add1ArgType), |
4945 | NEONMAP1(vqrshls_u32, aarch64_neon_uqrshl, Add1ArgType), |
4946 | NEONMAP1(vqrshrnd_n_s64, aarch64_neon_sqrshrn, AddRetType), |
4947 | NEONMAP1(vqrshrnd_n_u64, aarch64_neon_uqrshrn, AddRetType), |
4948 | NEONMAP1(vqrshrnh_n_s16, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), |
4949 | NEONMAP1(vqrshrnh_n_u16, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), |
4950 | NEONMAP1(vqrshrns_n_s32, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), |
4951 | NEONMAP1(vqrshrns_n_u32, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), |
4952 | NEONMAP1(vqrshrund_n_s64, aarch64_neon_sqrshrun, AddRetType), |
4953 | NEONMAP1(vqrshrunh_n_s16, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), |
4954 | NEONMAP1(vqrshruns_n_s32, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), |
4955 | NEONMAP1(vqshlb_n_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), |
4956 | NEONMAP1(vqshlb_n_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), |
4957 | NEONMAP1(vqshlb_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), |
4958 | NEONMAP1(vqshlb_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), |
4959 | NEONMAP1(vqshld_s64, aarch64_neon_sqshl, Add1ArgType), |
4960 | NEONMAP1(vqshld_u64, aarch64_neon_uqshl, Add1ArgType), |
4961 | NEONMAP1(vqshlh_n_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), |
4962 | NEONMAP1(vqshlh_n_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), |
4963 | NEONMAP1(vqshlh_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), |
4964 | NEONMAP1(vqshlh_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), |
4965 | NEONMAP1(vqshls_n_s32, aarch64_neon_sqshl, Add1ArgType), |
4966 | NEONMAP1(vqshls_n_u32, aarch64_neon_uqshl, Add1ArgType), |
4967 | NEONMAP1(vqshls_s32, aarch64_neon_sqshl, Add1ArgType), |
4968 | NEONMAP1(vqshls_u32, aarch64_neon_uqshl, Add1ArgType), |
4969 | NEONMAP1(vqshlub_n_s8, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), |
4970 | NEONMAP1(vqshluh_n_s16, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), |
4971 | NEONMAP1(vqshlus_n_s32, aarch64_neon_sqshlu, Add1ArgType), |
4972 | NEONMAP1(vqshrnd_n_s64, aarch64_neon_sqshrn, AddRetType), |
4973 | NEONMAP1(vqshrnd_n_u64, aarch64_neon_uqshrn, AddRetType), |
4974 | NEONMAP1(vqshrnh_n_s16, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), |
4975 | NEONMAP1(vqshrnh_n_u16, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), |
4976 | NEONMAP1(vqshrns_n_s32, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), |
4977 | NEONMAP1(vqshrns_n_u32, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), |
4978 | NEONMAP1(vqshrund_n_s64, aarch64_neon_sqshrun, AddRetType), |
4979 | NEONMAP1(vqshrunh_n_s16, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), |
4980 | NEONMAP1(vqshruns_n_s32, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), |
4981 | NEONMAP1(vqsubb_s8, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), |
4982 | NEONMAP1(vqsubb_u8, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), |
4983 | NEONMAP1(vqsubd_s64, aarch64_neon_sqsub, Add1ArgType), |
4984 | NEONMAP1(vqsubd_u64, aarch64_neon_uqsub, Add1ArgType), |
4985 | NEONMAP1(vqsubh_s16, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), |
4986 | NEONMAP1(vqsubh_u16, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), |
4987 | NEONMAP1(vqsubs_s32, aarch64_neon_sqsub, Add1ArgType), |
4988 | NEONMAP1(vqsubs_u32, aarch64_neon_uqsub, Add1ArgType), |
4989 | NEONMAP1(vrecped_f64, aarch64_neon_frecpe, Add1ArgType), |
4990 | NEONMAP1(vrecpes_f32, aarch64_neon_frecpe, Add1ArgType), |
4991 | NEONMAP1(vrecpxd_f64, aarch64_neon_frecpx, Add1ArgType), |
4992 | NEONMAP1(vrecpxs_f32, aarch64_neon_frecpx, Add1ArgType), |
4993 | NEONMAP1(vrshld_s64, aarch64_neon_srshl, Add1ArgType), |
4994 | NEONMAP1(vrshld_u64, aarch64_neon_urshl, Add1ArgType), |
4995 | NEONMAP1(vrsqrted_f64, aarch64_neon_frsqrte, Add1ArgType), |
4996 | NEONMAP1(vrsqrtes_f32, aarch64_neon_frsqrte, Add1ArgType), |
4997 | NEONMAP1(vrsqrtsd_f64, aarch64_neon_frsqrts, Add1ArgType), |
4998 | NEONMAP1(vrsqrtss_f32, aarch64_neon_frsqrts, Add1ArgType), |
4999 | NEONMAP1(vsha1cq_u32, aarch64_crypto_sha1c, 0), |
5000 | NEONMAP1(vsha1h_u32, aarch64_crypto_sha1h, 0), |
5001 | NEONMAP1(vsha1mq_u32, aarch64_crypto_sha1m, 0), |
5002 | NEONMAP1(vsha1pq_u32, aarch64_crypto_sha1p, 0), |
5003 | NEONMAP1(vshld_s64, aarch64_neon_sshl, Add1ArgType), |
5004 | NEONMAP1(vshld_u64, aarch64_neon_ushl, Add1ArgType), |
5005 | NEONMAP1(vslid_n_s64, aarch64_neon_vsli, Vectorize1ArgType), |
5006 | NEONMAP1(vslid_n_u64, aarch64_neon_vsli, Vectorize1ArgType), |
5007 | NEONMAP1(vsqaddb_u8, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), |
5008 | NEONMAP1(vsqaddd_u64, aarch64_neon_usqadd, Add1ArgType), |
5009 | NEONMAP1(vsqaddh_u16, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), |
5010 | NEONMAP1(vsqadds_u32, aarch64_neon_usqadd, Add1ArgType), |
5011 | NEONMAP1(vsrid_n_s64, aarch64_neon_vsri, Vectorize1ArgType), |
5012 | NEONMAP1(vsrid_n_u64, aarch64_neon_vsri, Vectorize1ArgType), |
5013 | NEONMAP1(vuqaddb_s8, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), |
5014 | NEONMAP1(vuqaddd_s64, aarch64_neon_suqadd, Add1ArgType), |
5015 | NEONMAP1(vuqaddh_s16, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), |
5016 | NEONMAP1(vuqadds_s32, aarch64_neon_suqadd, Add1ArgType), |
5017 | // FP16 scalar intrinisics go here. |
5018 | NEONMAP1(vabdh_f16, aarch64_sisd_fabd, Add1ArgType), |
5019 | NEONMAP1(vcvtah_s32_f16, aarch64_neon_fcvtas, AddRetType | Add1ArgType), |
5020 | NEONMAP1(vcvtah_s64_f16, aarch64_neon_fcvtas, AddRetType | Add1ArgType), |
5021 | NEONMAP1(vcvtah_u32_f16, aarch64_neon_fcvtau, AddRetType | Add1ArgType), |
5022 | NEONMAP1(vcvtah_u64_f16, aarch64_neon_fcvtau, AddRetType | Add1ArgType), |
5023 | NEONMAP1(vcvth_n_f16_s32, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), |
5024 | NEONMAP1(vcvth_n_f16_s64, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), |
5025 | NEONMAP1(vcvth_n_f16_u32, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), |
5026 | NEONMAP1(vcvth_n_f16_u64, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), |
5027 | NEONMAP1(vcvth_n_s32_f16, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), |
5028 | NEONMAP1(vcvth_n_s64_f16, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), |
5029 | NEONMAP1(vcvth_n_u32_f16, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), |
5030 | NEONMAP1(vcvth_n_u64_f16, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), |
5031 | NEONMAP1(vcvtmh_s32_f16, aarch64_neon_fcvtms, AddRetType | Add1ArgType), |
5032 | NEONMAP1(vcvtmh_s64_f16, aarch64_neon_fcvtms, AddRetType | Add1ArgType), |
5033 | NEONMAP1(vcvtmh_u32_f16, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), |
5034 | NEONMAP1(vcvtmh_u64_f16, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), |
5035 | NEONMAP1(vcvtnh_s32_f16, aarch64_neon_fcvtns, AddRetType | Add1ArgType), |
5036 | NEONMAP1(vcvtnh_s64_f16, aarch64_neon_fcvtns, AddRetType | Add1ArgType), |
5037 | NEONMAP1(vcvtnh_u32_f16, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), |
5038 | NEONMAP1(vcvtnh_u64_f16, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), |
5039 | NEONMAP1(vcvtph_s32_f16, aarch64_neon_fcvtps, AddRetType | Add1ArgType), |
5040 | NEONMAP1(vcvtph_s64_f16, aarch64_neon_fcvtps, AddRetType | Add1ArgType), |
5041 | NEONMAP1(vcvtph_u32_f16, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), |
5042 | NEONMAP1(vcvtph_u64_f16, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), |
5043 | NEONMAP1(vmulxh_f16, aarch64_neon_fmulx, Add1ArgType), |
5044 | NEONMAP1(vrecpeh_f16, aarch64_neon_frecpe, Add1ArgType), |
5045 | NEONMAP1(vrecpxh_f16, aarch64_neon_frecpx, Add1ArgType), |
5046 | NEONMAP1(vrsqrteh_f16, aarch64_neon_frsqrte, Add1ArgType), |
5047 | NEONMAP1(vrsqrtsh_f16, aarch64_neon_frsqrts, Add1ArgType), |
5048 | }; |
5049 | |
5050 | #undef NEONMAP0 |
5051 | #undef NEONMAP1 |
5052 | #undef NEONMAP2 |
5053 | |
5054 | static bool NEONSIMDIntrinsicsProvenSorted = false; |
5055 | |
5056 | static bool AArch64SIMDIntrinsicsProvenSorted = false; |
5057 | static bool AArch64SISDIntrinsicsProvenSorted = false; |
5058 | |
5059 | |
5060 | static const NeonIntrinsicInfo * |
5061 | findNeonIntrinsicInMap(ArrayRef<NeonIntrinsicInfo> IntrinsicMap, |
5062 | unsigned BuiltinID, bool &MapProvenSorted) { |
5063 | |
5064 | #ifndef NDEBUG |
5065 | if (!MapProvenSorted) { |
5066 | assert(std::is_sorted(std::begin(IntrinsicMap), std::end(IntrinsicMap)))((std::is_sorted(std::begin(IntrinsicMap), std::end(IntrinsicMap ))) ? static_cast<void> (0) : __assert_fail ("std::is_sorted(std::begin(IntrinsicMap), std::end(IntrinsicMap))" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 5066, __PRETTY_FUNCTION__)); |
5067 | MapProvenSorted = true; |
5068 | } |
5069 | #endif |
5070 | |
5071 | const NeonIntrinsicInfo *Builtin = |
5072 | std::lower_bound(IntrinsicMap.begin(), IntrinsicMap.end(), BuiltinID); |
5073 | |
5074 | if (Builtin != IntrinsicMap.end() && Builtin->BuiltinID == BuiltinID) |
5075 | return Builtin; |
5076 | |
5077 | return nullptr; |
5078 | } |
5079 | |
5080 | Function *CodeGenFunction::LookupNeonLLVMIntrinsic(unsigned IntrinsicID, |
5081 | unsigned Modifier, |
5082 | llvm::Type *ArgType, |
5083 | const CallExpr *E) { |
5084 | int VectorSize = 0; |
5085 | if (Modifier & Use64BitVectors) |
5086 | VectorSize = 64; |
5087 | else if (Modifier & Use128BitVectors) |
5088 | VectorSize = 128; |
5089 | |
5090 | // Return type. |
5091 | SmallVector<llvm::Type *, 3> Tys; |
5092 | if (Modifier & AddRetType) { |
5093 | llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext())); |
5094 | if (Modifier & VectorizeRetType) |
5095 | Ty = llvm::VectorType::get( |
5096 | Ty, VectorSize ? VectorSize / Ty->getPrimitiveSizeInBits() : 1); |
5097 | |
5098 | Tys.push_back(Ty); |
5099 | } |
5100 | |
5101 | // Arguments. |
5102 | if (Modifier & VectorizeArgTypes) { |
5103 | int Elts = VectorSize ? VectorSize / ArgType->getPrimitiveSizeInBits() : 1; |
5104 | ArgType = llvm::VectorType::get(ArgType, Elts); |
5105 | } |
5106 | |
5107 | if (Modifier & (Add1ArgType | Add2ArgTypes)) |
5108 | Tys.push_back(ArgType); |
5109 | |
5110 | if (Modifier & Add2ArgTypes) |
5111 | Tys.push_back(ArgType); |
5112 | |
5113 | if (Modifier & InventFloatType) |
5114 | Tys.push_back(FloatTy); |
5115 | |
5116 | return CGM.getIntrinsic(IntrinsicID, Tys); |
5117 | } |
5118 | |
5119 | static Value *EmitCommonNeonSISDBuiltinExpr(CodeGenFunction &CGF, |
5120 | const NeonIntrinsicInfo &SISDInfo, |
5121 | SmallVectorImpl<Value *> &Ops, |
5122 | const CallExpr *E) { |
5123 | unsigned BuiltinID = SISDInfo.BuiltinID; |
5124 | unsigned int Int = SISDInfo.LLVMIntrinsic; |
5125 | unsigned Modifier = SISDInfo.TypeModifier; |
5126 | const char *s = SISDInfo.NameHint; |
5127 | |
5128 | switch (BuiltinID) { |
5129 | case NEON::BI__builtin_neon_vcled_s64: |
5130 | case NEON::BI__builtin_neon_vcled_u64: |
5131 | case NEON::BI__builtin_neon_vcles_f32: |
5132 | case NEON::BI__builtin_neon_vcled_f64: |
5133 | case NEON::BI__builtin_neon_vcltd_s64: |
5134 | case NEON::BI__builtin_neon_vcltd_u64: |
5135 | case NEON::BI__builtin_neon_vclts_f32: |
5136 | case NEON::BI__builtin_neon_vcltd_f64: |
5137 | case NEON::BI__builtin_neon_vcales_f32: |
5138 | case NEON::BI__builtin_neon_vcaled_f64: |
5139 | case NEON::BI__builtin_neon_vcalts_f32: |
5140 | case NEON::BI__builtin_neon_vcaltd_f64: |
5141 | // Only one direction of comparisons actually exist, cmle is actually a cmge |
5142 | // with swapped operands. The table gives us the right intrinsic but we |
5143 | // still need to do the swap. |
5144 | std::swap(Ops[0], Ops[1]); |
5145 | break; |
5146 | } |
5147 | |
5148 | assert(Int && "Generic code assumes a valid intrinsic")((Int && "Generic code assumes a valid intrinsic") ? static_cast <void> (0) : __assert_fail ("Int && \"Generic code assumes a valid intrinsic\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 5148, __PRETTY_FUNCTION__)); |
5149 | |
5150 | // Determine the type(s) of this overloaded AArch64 intrinsic. |
5151 | const Expr *Arg = E->getArg(0); |
5152 | llvm::Type *ArgTy = CGF.ConvertType(Arg->getType()); |
5153 | Function *F = CGF.LookupNeonLLVMIntrinsic(Int, Modifier, ArgTy, E); |
5154 | |
5155 | int j = 0; |
5156 | ConstantInt *C0 = ConstantInt::get(CGF.SizeTy, 0); |
5157 | for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); |
5158 | ai != ae; ++ai, ++j) { |
5159 | llvm::Type *ArgTy = ai->getType(); |
5160 | if (Ops[j]->getType()->getPrimitiveSizeInBits() == |
5161 | ArgTy->getPrimitiveSizeInBits()) |
5162 | continue; |
5163 | |
5164 | assert(ArgTy->isVectorTy() && !Ops[j]->getType()->isVectorTy())((ArgTy->isVectorTy() && !Ops[j]->getType()-> isVectorTy()) ? static_cast<void> (0) : __assert_fail ( "ArgTy->isVectorTy() && !Ops[j]->getType()->isVectorTy()" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 5164, __PRETTY_FUNCTION__)); |
5165 | // The constant argument to an _n_ intrinsic always has Int32Ty, so truncate |
5166 | // it before inserting. |
5167 | Ops[j] = |
5168 | CGF.Builder.CreateTruncOrBitCast(Ops[j], ArgTy->getVectorElementType()); |
5169 | Ops[j] = |
5170 | CGF.Builder.CreateInsertElement(UndefValue::get(ArgTy), Ops[j], C0); |
5171 | } |
5172 | |
5173 | Value *Result = CGF.EmitNeonCall(F, Ops, s); |
5174 | llvm::Type *ResultType = CGF.ConvertType(E->getType()); |
5175 | if (ResultType->getPrimitiveSizeInBits() < |
5176 | Result->getType()->getPrimitiveSizeInBits()) |
5177 | return CGF.Builder.CreateExtractElement(Result, C0); |
5178 | |
5179 | return CGF.Builder.CreateBitCast(Result, ResultType, s); |
5180 | } |
5181 | |
5182 | Value *CodeGenFunction::EmitCommonNeonBuiltinExpr( |
5183 | unsigned BuiltinID, unsigned LLVMIntrinsic, unsigned AltLLVMIntrinsic, |
5184 | const char *NameHint, unsigned Modifier, const CallExpr *E, |
5185 | SmallVectorImpl<llvm::Value *> &Ops, Address PtrOp0, Address PtrOp1, |
5186 | llvm::Triple::ArchType Arch) { |
5187 | // Get the last argument, which specifies the vector type. |
5188 | llvm::APSInt NeonTypeConst; |
5189 | const Expr *Arg = E->getArg(E->getNumArgs() - 1); |
5190 | if (!Arg->isIntegerConstantExpr(NeonTypeConst, getContext())) |
5191 | return nullptr; |
5192 | |
5193 | // Determine the type of this overloaded NEON intrinsic. |
5194 | NeonTypeFlags Type(NeonTypeConst.getZExtValue()); |
5195 | bool Usgn = Type.isUnsigned(); |
5196 | bool Quad = Type.isQuad(); |
5197 | const bool HasLegalHalfType = getTarget().hasLegalHalfType(); |
5198 | |
5199 | llvm::VectorType *VTy = GetNeonType(this, Type, HasLegalHalfType); |
5200 | llvm::Type *Ty = VTy; |
5201 | if (!Ty) |
5202 | return nullptr; |
5203 | |
5204 | auto getAlignmentValue32 = [&](Address addr) -> Value* { |
5205 | return Builder.getInt32(addr.getAlignment().getQuantity()); |
5206 | }; |
5207 | |
5208 | unsigned Int = LLVMIntrinsic; |
5209 | if ((Modifier & UnsignedAlts) && !Usgn) |
5210 | Int = AltLLVMIntrinsic; |
5211 | |
5212 | switch (BuiltinID) { |
5213 | default: break; |
5214 | case NEON::BI__builtin_neon_vpadd_v: |
5215 | case NEON::BI__builtin_neon_vpaddq_v: |
5216 | // We don't allow fp/int overloading of intrinsics. |
5217 | if (VTy->getElementType()->isFloatingPointTy() && |
5218 | Int == Intrinsic::aarch64_neon_addp) |
5219 | Int = Intrinsic::aarch64_neon_faddp; |
5220 | break; |
5221 | case NEON::BI__builtin_neon_vabs_v: |
5222 | case NEON::BI__builtin_neon_vabsq_v: |
5223 | if (VTy->getElementType()->isFloatingPointTy()) |
5224 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::fabs, Ty), Ops, "vabs"); |
5225 | return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vabs"); |
5226 | case NEON::BI__builtin_neon_vaddhn_v: { |
5227 | llvm::VectorType *SrcTy = |
5228 | llvm::VectorType::getExtendedElementVectorType(VTy); |
5229 | |
5230 | // %sum = add <4 x i32> %lhs, %rhs |
5231 | Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); |
5232 | Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); |
5233 | Ops[0] = Builder.CreateAdd(Ops[0], Ops[1], "vaddhn"); |
5234 | |
5235 | // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> |
5236 | Constant *ShiftAmt = |
5237 | ConstantInt::get(SrcTy, SrcTy->getScalarSizeInBits() / 2); |
5238 | Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vaddhn"); |
5239 | |
5240 | // %res = trunc <4 x i32> %high to <4 x i16> |
5241 | return Builder.CreateTrunc(Ops[0], VTy, "vaddhn"); |
5242 | } |
5243 | case NEON::BI__builtin_neon_vcale_v: |
5244 | case NEON::BI__builtin_neon_vcaleq_v: |
5245 | case NEON::BI__builtin_neon_vcalt_v: |
5246 | case NEON::BI__builtin_neon_vcaltq_v: |
5247 | std::swap(Ops[0], Ops[1]); |
5248 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
5249 | case NEON::BI__builtin_neon_vcage_v: |
5250 | case NEON::BI__builtin_neon_vcageq_v: |
5251 | case NEON::BI__builtin_neon_vcagt_v: |
5252 | case NEON::BI__builtin_neon_vcagtq_v: { |
5253 | llvm::Type *Ty; |
5254 | switch (VTy->getScalarSizeInBits()) { |
5255 | default: llvm_unreachable("unexpected type")::llvm::llvm_unreachable_internal("unexpected type", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 5255); |
5256 | case 32: |
5257 | Ty = FloatTy; |
5258 | break; |
5259 | case 64: |
5260 | Ty = DoubleTy; |
5261 | break; |
5262 | case 16: |
5263 | Ty = HalfTy; |
5264 | break; |
5265 | } |
5266 | llvm::Type *VecFlt = llvm::VectorType::get(Ty, VTy->getNumElements()); |
5267 | llvm::Type *Tys[] = { VTy, VecFlt }; |
5268 | Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); |
5269 | return EmitNeonCall(F, Ops, NameHint); |
5270 | } |
5271 | case NEON::BI__builtin_neon_vceqz_v: |
5272 | case NEON::BI__builtin_neon_vceqzq_v: |
5273 | return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OEQ, |
5274 | ICmpInst::ICMP_EQ, "vceqz"); |
5275 | case NEON::BI__builtin_neon_vcgez_v: |
5276 | case NEON::BI__builtin_neon_vcgezq_v: |
5277 | return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGE, |
5278 | ICmpInst::ICMP_SGE, "vcgez"); |
5279 | case NEON::BI__builtin_neon_vclez_v: |
5280 | case NEON::BI__builtin_neon_vclezq_v: |
5281 | return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLE, |
5282 | ICmpInst::ICMP_SLE, "vclez"); |
5283 | case NEON::BI__builtin_neon_vcgtz_v: |
5284 | case NEON::BI__builtin_neon_vcgtzq_v: |
5285 | return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGT, |
5286 | ICmpInst::ICMP_SGT, "vcgtz"); |
5287 | case NEON::BI__builtin_neon_vcltz_v: |
5288 | case NEON::BI__builtin_neon_vcltzq_v: |
5289 | return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLT, |
5290 | ICmpInst::ICMP_SLT, "vcltz"); |
5291 | case NEON::BI__builtin_neon_vclz_v: |
5292 | case NEON::BI__builtin_neon_vclzq_v: |
5293 | // We generate target-independent intrinsic, which needs a second argument |
5294 | // for whether or not clz of zero is undefined; on ARM it isn't. |
5295 | Ops.push_back(Builder.getInt1(getTarget().isCLZForZeroUndef())); |
5296 | break; |
5297 | case NEON::BI__builtin_neon_vcvt_f32_v: |
5298 | case NEON::BI__builtin_neon_vcvtq_f32_v: |
5299 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
5300 | Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, Quad), |
5301 | HasLegalHalfType); |
5302 | return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") |
5303 | : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); |
5304 | case NEON::BI__builtin_neon_vcvt_f16_v: |
5305 | case NEON::BI__builtin_neon_vcvtq_f16_v: |
5306 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
5307 | Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float16, false, Quad), |
5308 | HasLegalHalfType); |
5309 | return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") |
5310 | : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); |
5311 | case NEON::BI__builtin_neon_vcvt_n_f16_v: |
5312 | case NEON::BI__builtin_neon_vcvt_n_f32_v: |
5313 | case NEON::BI__builtin_neon_vcvt_n_f64_v: |
5314 | case NEON::BI__builtin_neon_vcvtq_n_f16_v: |
5315 | case NEON::BI__builtin_neon_vcvtq_n_f32_v: |
5316 | case NEON::BI__builtin_neon_vcvtq_n_f64_v: { |
5317 | llvm::Type *Tys[2] = { GetFloatNeonType(this, Type), Ty }; |
5318 | Int = Usgn ? LLVMIntrinsic : AltLLVMIntrinsic; |
5319 | Function *F = CGM.getIntrinsic(Int, Tys); |
5320 | return EmitNeonCall(F, Ops, "vcvt_n"); |
5321 | } |
5322 | case NEON::BI__builtin_neon_vcvt_n_s16_v: |
5323 | case NEON::BI__builtin_neon_vcvt_n_s32_v: |
5324 | case NEON::BI__builtin_neon_vcvt_n_u16_v: |
5325 | case NEON::BI__builtin_neon_vcvt_n_u32_v: |
5326 | case NEON::BI__builtin_neon_vcvt_n_s64_v: |
5327 | case NEON::BI__builtin_neon_vcvt_n_u64_v: |
5328 | case NEON::BI__builtin_neon_vcvtq_n_s16_v: |
5329 | case NEON::BI__builtin_neon_vcvtq_n_s32_v: |
5330 | case NEON::BI__builtin_neon_vcvtq_n_u16_v: |
5331 | case NEON::BI__builtin_neon_vcvtq_n_u32_v: |
5332 | case NEON::BI__builtin_neon_vcvtq_n_s64_v: |
5333 | case NEON::BI__builtin_neon_vcvtq_n_u64_v: { |
5334 | llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; |
5335 | Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); |
5336 | return EmitNeonCall(F, Ops, "vcvt_n"); |
5337 | } |
5338 | case NEON::BI__builtin_neon_vcvt_s32_v: |
5339 | case NEON::BI__builtin_neon_vcvt_u32_v: |
5340 | case NEON::BI__builtin_neon_vcvt_s64_v: |
5341 | case NEON::BI__builtin_neon_vcvt_u64_v: |
5342 | case NEON::BI__builtin_neon_vcvt_s16_v: |
5343 | case NEON::BI__builtin_neon_vcvt_u16_v: |
5344 | case NEON::BI__builtin_neon_vcvtq_s32_v: |
5345 | case NEON::BI__builtin_neon_vcvtq_u32_v: |
5346 | case NEON::BI__builtin_neon_vcvtq_s64_v: |
5347 | case NEON::BI__builtin_neon_vcvtq_u64_v: |
5348 | case NEON::BI__builtin_neon_vcvtq_s16_v: |
5349 | case NEON::BI__builtin_neon_vcvtq_u16_v: { |
5350 | Ops[0] = Builder.CreateBitCast(Ops[0], GetFloatNeonType(this, Type)); |
5351 | return Usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt") |
5352 | : Builder.CreateFPToSI(Ops[0], Ty, "vcvt"); |
5353 | } |
5354 | case NEON::BI__builtin_neon_vcvta_s16_v: |
5355 | case NEON::BI__builtin_neon_vcvta_s32_v: |
5356 | case NEON::BI__builtin_neon_vcvta_s64_v: |
5357 | case NEON::BI__builtin_neon_vcvta_u16_v: |
5358 | case NEON::BI__builtin_neon_vcvta_u32_v: |
5359 | case NEON::BI__builtin_neon_vcvta_u64_v: |
5360 | case NEON::BI__builtin_neon_vcvtaq_s16_v: |
5361 | case NEON::BI__builtin_neon_vcvtaq_s32_v: |
5362 | case NEON::BI__builtin_neon_vcvtaq_s64_v: |
5363 | case NEON::BI__builtin_neon_vcvtaq_u16_v: |
5364 | case NEON::BI__builtin_neon_vcvtaq_u32_v: |
5365 | case NEON::BI__builtin_neon_vcvtaq_u64_v: |
5366 | case NEON::BI__builtin_neon_vcvtn_s16_v: |
5367 | case NEON::BI__builtin_neon_vcvtn_s32_v: |
5368 | case NEON::BI__builtin_neon_vcvtn_s64_v: |
5369 | case NEON::BI__builtin_neon_vcvtn_u16_v: |
5370 | case NEON::BI__builtin_neon_vcvtn_u32_v: |
5371 | case NEON::BI__builtin_neon_vcvtn_u64_v: |
5372 | case NEON::BI__builtin_neon_vcvtnq_s16_v: |
5373 | case NEON::BI__builtin_neon_vcvtnq_s32_v: |
5374 | case NEON::BI__builtin_neon_vcvtnq_s64_v: |
5375 | case NEON::BI__builtin_neon_vcvtnq_u16_v: |
5376 | case NEON::BI__builtin_neon_vcvtnq_u32_v: |
5377 | case NEON::BI__builtin_neon_vcvtnq_u64_v: |
5378 | case NEON::BI__builtin_neon_vcvtp_s16_v: |
5379 | case NEON::BI__builtin_neon_vcvtp_s32_v: |
5380 | case NEON::BI__builtin_neon_vcvtp_s64_v: |
5381 | case NEON::BI__builtin_neon_vcvtp_u16_v: |
5382 | case NEON::BI__builtin_neon_vcvtp_u32_v: |
5383 | case NEON::BI__builtin_neon_vcvtp_u64_v: |
5384 | case NEON::BI__builtin_neon_vcvtpq_s16_v: |
5385 | case NEON::BI__builtin_neon_vcvtpq_s32_v: |
5386 | case NEON::BI__builtin_neon_vcvtpq_s64_v: |
5387 | case NEON::BI__builtin_neon_vcvtpq_u16_v: |
5388 | case NEON::BI__builtin_neon_vcvtpq_u32_v: |
5389 | case NEON::BI__builtin_neon_vcvtpq_u64_v: |
5390 | case NEON::BI__builtin_neon_vcvtm_s16_v: |
5391 | case NEON::BI__builtin_neon_vcvtm_s32_v: |
5392 | case NEON::BI__builtin_neon_vcvtm_s64_v: |
5393 | case NEON::BI__builtin_neon_vcvtm_u16_v: |
5394 | case NEON::BI__builtin_neon_vcvtm_u32_v: |
5395 | case NEON::BI__builtin_neon_vcvtm_u64_v: |
5396 | case NEON::BI__builtin_neon_vcvtmq_s16_v: |
5397 | case NEON::BI__builtin_neon_vcvtmq_s32_v: |
5398 | case NEON::BI__builtin_neon_vcvtmq_s64_v: |
5399 | case NEON::BI__builtin_neon_vcvtmq_u16_v: |
5400 | case NEON::BI__builtin_neon_vcvtmq_u32_v: |
5401 | case NEON::BI__builtin_neon_vcvtmq_u64_v: { |
5402 | llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; |
5403 | return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, NameHint); |
5404 | } |
5405 | case NEON::BI__builtin_neon_vext_v: |
5406 | case NEON::BI__builtin_neon_vextq_v: { |
5407 | int CV = cast<ConstantInt>(Ops[2])->getSExtValue(); |
5408 | SmallVector<uint32_t, 16> Indices; |
5409 | for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) |
5410 | Indices.push_back(i+CV); |
5411 | |
5412 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
5413 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
5414 | return Builder.CreateShuffleVector(Ops[0], Ops[1], Indices, "vext"); |
5415 | } |
5416 | case NEON::BI__builtin_neon_vfma_v: |
5417 | case NEON::BI__builtin_neon_vfmaq_v: { |
5418 | Function *F = CGM.getIntrinsic(Intrinsic::fma, Ty); |
5419 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
5420 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
5421 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
5422 | |
5423 | // NEON intrinsic puts accumulator first, unlike the LLVM fma. |
5424 | return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); |
5425 | } |
5426 | case NEON::BI__builtin_neon_vld1_v: |
5427 | case NEON::BI__builtin_neon_vld1q_v: { |
5428 | llvm::Type *Tys[] = {Ty, Int8PtrTy}; |
5429 | Ops.push_back(getAlignmentValue32(PtrOp0)); |
5430 | return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, "vld1"); |
5431 | } |
5432 | case NEON::BI__builtin_neon_vld1_x2_v: |
5433 | case NEON::BI__builtin_neon_vld1q_x2_v: |
5434 | case NEON::BI__builtin_neon_vld1_x3_v: |
5435 | case NEON::BI__builtin_neon_vld1q_x3_v: |
5436 | case NEON::BI__builtin_neon_vld1_x4_v: |
5437 | case NEON::BI__builtin_neon_vld1q_x4_v: { |
5438 | llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); |
5439 | Ops[1] = Builder.CreateBitCast(Ops[1], PTy); |
5440 | llvm::Type *Tys[2] = { VTy, PTy }; |
5441 | Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); |
5442 | Ops[1] = Builder.CreateCall(F, Ops[1], "vld1xN"); |
5443 | Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); |
5444 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
5445 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
5446 | } |
5447 | case NEON::BI__builtin_neon_vld2_v: |
5448 | case NEON::BI__builtin_neon_vld2q_v: |
5449 | case NEON::BI__builtin_neon_vld3_v: |
5450 | case NEON::BI__builtin_neon_vld3q_v: |
5451 | case NEON::BI__builtin_neon_vld4_v: |
5452 | case NEON::BI__builtin_neon_vld4q_v: |
5453 | case NEON::BI__builtin_neon_vld2_dup_v: |
5454 | case NEON::BI__builtin_neon_vld2q_dup_v: |
5455 | case NEON::BI__builtin_neon_vld3_dup_v: |
5456 | case NEON::BI__builtin_neon_vld3q_dup_v: |
5457 | case NEON::BI__builtin_neon_vld4_dup_v: |
5458 | case NEON::BI__builtin_neon_vld4q_dup_v: { |
5459 | llvm::Type *Tys[] = {Ty, Int8PtrTy}; |
5460 | Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); |
5461 | Value *Align = getAlignmentValue32(PtrOp1); |
5462 | Ops[1] = Builder.CreateCall(F, {Ops[1], Align}, NameHint); |
5463 | Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); |
5464 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
5465 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
5466 | } |
5467 | case NEON::BI__builtin_neon_vld1_dup_v: |
5468 | case NEON::BI__builtin_neon_vld1q_dup_v: { |
5469 | Value *V = UndefValue::get(Ty); |
5470 | Ty = llvm::PointerType::getUnqual(VTy->getElementType()); |
5471 | PtrOp0 = Builder.CreateBitCast(PtrOp0, Ty); |
5472 | LoadInst *Ld = Builder.CreateLoad(PtrOp0); |
5473 | llvm::Constant *CI = ConstantInt::get(SizeTy, 0); |
5474 | Ops[0] = Builder.CreateInsertElement(V, Ld, CI); |
5475 | return EmitNeonSplat(Ops[0], CI); |
5476 | } |
5477 | case NEON::BI__builtin_neon_vld2_lane_v: |
5478 | case NEON::BI__builtin_neon_vld2q_lane_v: |
5479 | case NEON::BI__builtin_neon_vld3_lane_v: |
5480 | case NEON::BI__builtin_neon_vld3q_lane_v: |
5481 | case NEON::BI__builtin_neon_vld4_lane_v: |
5482 | case NEON::BI__builtin_neon_vld4q_lane_v: { |
5483 | llvm::Type *Tys[] = {Ty, Int8PtrTy}; |
5484 | Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); |
5485 | for (unsigned I = 2; I < Ops.size() - 1; ++I) |
5486 | Ops[I] = Builder.CreateBitCast(Ops[I], Ty); |
5487 | Ops.push_back(getAlignmentValue32(PtrOp1)); |
5488 | Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), NameHint); |
5489 | Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); |
5490 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
5491 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
5492 | } |
5493 | case NEON::BI__builtin_neon_vmovl_v: { |
5494 | llvm::Type *DTy =llvm::VectorType::getTruncatedElementVectorType(VTy); |
5495 | Ops[0] = Builder.CreateBitCast(Ops[0], DTy); |
5496 | if (Usgn) |
5497 | return Builder.CreateZExt(Ops[0], Ty, "vmovl"); |
5498 | return Builder.CreateSExt(Ops[0], Ty, "vmovl"); |
5499 | } |
5500 | case NEON::BI__builtin_neon_vmovn_v: { |
5501 | llvm::Type *QTy = llvm::VectorType::getExtendedElementVectorType(VTy); |
5502 | Ops[0] = Builder.CreateBitCast(Ops[0], QTy); |
5503 | return Builder.CreateTrunc(Ops[0], Ty, "vmovn"); |
5504 | } |
5505 | case NEON::BI__builtin_neon_vmull_v: |
5506 | // FIXME: the integer vmull operations could be emitted in terms of pure |
5507 | // LLVM IR (2 exts followed by a mul). Unfortunately LLVM has a habit of |
5508 | // hoisting the exts outside loops. Until global ISel comes along that can |
5509 | // see through such movement this leads to bad CodeGen. So we need an |
5510 | // intrinsic for now. |
5511 | Int = Usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls; |
5512 | Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp : Int; |
5513 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); |
5514 | case NEON::BI__builtin_neon_vpadal_v: |
5515 | case NEON::BI__builtin_neon_vpadalq_v: { |
5516 | // The source operand type has twice as many elements of half the size. |
5517 | unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); |
5518 | llvm::Type *EltTy = |
5519 | llvm::IntegerType::get(getLLVMContext(), EltBits / 2); |
5520 | llvm::Type *NarrowTy = |
5521 | llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); |
5522 | llvm::Type *Tys[2] = { Ty, NarrowTy }; |
5523 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, NameHint); |
5524 | } |
5525 | case NEON::BI__builtin_neon_vpaddl_v: |
5526 | case NEON::BI__builtin_neon_vpaddlq_v: { |
5527 | // The source operand type has twice as many elements of half the size. |
5528 | unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); |
5529 | llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2); |
5530 | llvm::Type *NarrowTy = |
5531 | llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); |
5532 | llvm::Type *Tys[2] = { Ty, NarrowTy }; |
5533 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl"); |
5534 | } |
5535 | case NEON::BI__builtin_neon_vqdmlal_v: |
5536 | case NEON::BI__builtin_neon_vqdmlsl_v: { |
5537 | SmallVector<Value *, 2> MulOps(Ops.begin() + 1, Ops.end()); |
5538 | Ops[1] = |
5539 | EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), MulOps, "vqdmlal"); |
5540 | Ops.resize(2); |
5541 | return EmitNeonCall(CGM.getIntrinsic(AltLLVMIntrinsic, Ty), Ops, NameHint); |
5542 | } |
5543 | case NEON::BI__builtin_neon_vqshl_n_v: |
5544 | case NEON::BI__builtin_neon_vqshlq_n_v: |
5545 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n", |
5546 | 1, false); |
5547 | case NEON::BI__builtin_neon_vqshlu_n_v: |
5548 | case NEON::BI__builtin_neon_vqshluq_n_v: |
5549 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshlu_n", |
5550 | 1, false); |
5551 | case NEON::BI__builtin_neon_vrecpe_v: |
5552 | case NEON::BI__builtin_neon_vrecpeq_v: |
5553 | case NEON::BI__builtin_neon_vrsqrte_v: |
5554 | case NEON::BI__builtin_neon_vrsqrteq_v: |
5555 | Int = Ty->isFPOrFPVectorTy() ? LLVMIntrinsic : AltLLVMIntrinsic; |
5556 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, NameHint); |
5557 | case NEON::BI__builtin_neon_vrndi_v: |
5558 | case NEON::BI__builtin_neon_vrndiq_v: |
5559 | Int = Intrinsic::nearbyint; |
5560 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, NameHint); |
5561 | case NEON::BI__builtin_neon_vrshr_n_v: |
5562 | case NEON::BI__builtin_neon_vrshrq_n_v: |
5563 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", |
5564 | 1, true); |
5565 | case NEON::BI__builtin_neon_vshl_n_v: |
5566 | case NEON::BI__builtin_neon_vshlq_n_v: |
5567 | Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); |
5568 | return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1], |
5569 | "vshl_n"); |
5570 | case NEON::BI__builtin_neon_vshll_n_v: { |
5571 | llvm::Type *SrcTy = llvm::VectorType::getTruncatedElementVectorType(VTy); |
5572 | Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); |
5573 | if (Usgn) |
5574 | Ops[0] = Builder.CreateZExt(Ops[0], VTy); |
5575 | else |
5576 | Ops[0] = Builder.CreateSExt(Ops[0], VTy); |
5577 | Ops[1] = EmitNeonShiftVector(Ops[1], VTy, false); |
5578 | return Builder.CreateShl(Ops[0], Ops[1], "vshll_n"); |
5579 | } |
5580 | case NEON::BI__builtin_neon_vshrn_n_v: { |
5581 | llvm::Type *SrcTy = llvm::VectorType::getExtendedElementVectorType(VTy); |
5582 | Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); |
5583 | Ops[1] = EmitNeonShiftVector(Ops[1], SrcTy, false); |
5584 | if (Usgn) |
5585 | Ops[0] = Builder.CreateLShr(Ops[0], Ops[1]); |
5586 | else |
5587 | Ops[0] = Builder.CreateAShr(Ops[0], Ops[1]); |
5588 | return Builder.CreateTrunc(Ops[0], Ty, "vshrn_n"); |
5589 | } |
5590 | case NEON::BI__builtin_neon_vshr_n_v: |
5591 | case NEON::BI__builtin_neon_vshrq_n_v: |
5592 | return EmitNeonRShiftImm(Ops[0], Ops[1], Ty, Usgn, "vshr_n"); |
5593 | case NEON::BI__builtin_neon_vst1_v: |
5594 | case NEON::BI__builtin_neon_vst1q_v: |
5595 | case NEON::BI__builtin_neon_vst2_v: |
5596 | case NEON::BI__builtin_neon_vst2q_v: |
5597 | case NEON::BI__builtin_neon_vst3_v: |
5598 | case NEON::BI__builtin_neon_vst3q_v: |
5599 | case NEON::BI__builtin_neon_vst4_v: |
5600 | case NEON::BI__builtin_neon_vst4q_v: |
5601 | case NEON::BI__builtin_neon_vst2_lane_v: |
5602 | case NEON::BI__builtin_neon_vst2q_lane_v: |
5603 | case NEON::BI__builtin_neon_vst3_lane_v: |
5604 | case NEON::BI__builtin_neon_vst3q_lane_v: |
5605 | case NEON::BI__builtin_neon_vst4_lane_v: |
5606 | case NEON::BI__builtin_neon_vst4q_lane_v: { |
5607 | llvm::Type *Tys[] = {Int8PtrTy, Ty}; |
5608 | Ops.push_back(getAlignmentValue32(PtrOp0)); |
5609 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, ""); |
5610 | } |
5611 | case NEON::BI__builtin_neon_vst1_x2_v: |
5612 | case NEON::BI__builtin_neon_vst1q_x2_v: |
5613 | case NEON::BI__builtin_neon_vst1_x3_v: |
5614 | case NEON::BI__builtin_neon_vst1q_x3_v: |
5615 | case NEON::BI__builtin_neon_vst1_x4_v: |
5616 | case NEON::BI__builtin_neon_vst1q_x4_v: { |
5617 | llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); |
5618 | // TODO: Currently in AArch32 mode the pointer operand comes first, whereas |
5619 | // in AArch64 it comes last. We may want to stick to one or another. |
5620 | if (Arch == llvm::Triple::aarch64 || Arch == llvm::Triple::aarch64_be) { |
5621 | llvm::Type *Tys[2] = { VTy, PTy }; |
5622 | std::rotate(Ops.begin(), Ops.begin() + 1, Ops.end()); |
5623 | return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, ""); |
5624 | } |
5625 | llvm::Type *Tys[2] = { PTy, VTy }; |
5626 | return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, ""); |
5627 | } |
5628 | case NEON::BI__builtin_neon_vsubhn_v: { |
5629 | llvm::VectorType *SrcTy = |
5630 | llvm::VectorType::getExtendedElementVectorType(VTy); |
5631 | |
5632 | // %sum = add <4 x i32> %lhs, %rhs |
5633 | Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); |
5634 | Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); |
5635 | Ops[0] = Builder.CreateSub(Ops[0], Ops[1], "vsubhn"); |
5636 | |
5637 | // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> |
5638 | Constant *ShiftAmt = |
5639 | ConstantInt::get(SrcTy, SrcTy->getScalarSizeInBits() / 2); |
5640 | Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vsubhn"); |
5641 | |
5642 | // %res = trunc <4 x i32> %high to <4 x i16> |
5643 | return Builder.CreateTrunc(Ops[0], VTy, "vsubhn"); |
5644 | } |
5645 | case NEON::BI__builtin_neon_vtrn_v: |
5646 | case NEON::BI__builtin_neon_vtrnq_v: { |
5647 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); |
5648 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
5649 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
5650 | Value *SV = nullptr; |
5651 | |
5652 | for (unsigned vi = 0; vi != 2; ++vi) { |
5653 | SmallVector<uint32_t, 16> Indices; |
5654 | for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { |
5655 | Indices.push_back(i+vi); |
5656 | Indices.push_back(i+e+vi); |
5657 | } |
5658 | Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); |
5659 | SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vtrn"); |
5660 | SV = Builder.CreateDefaultAlignedStore(SV, Addr); |
5661 | } |
5662 | return SV; |
5663 | } |
5664 | case NEON::BI__builtin_neon_vtst_v: |
5665 | case NEON::BI__builtin_neon_vtstq_v: { |
5666 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
5667 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
5668 | Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); |
5669 | Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], |
5670 | ConstantAggregateZero::get(Ty)); |
5671 | return Builder.CreateSExt(Ops[0], Ty, "vtst"); |
5672 | } |
5673 | case NEON::BI__builtin_neon_vuzp_v: |
5674 | case NEON::BI__builtin_neon_vuzpq_v: { |
5675 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); |
5676 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
5677 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
5678 | Value *SV = nullptr; |
5679 | |
5680 | for (unsigned vi = 0; vi != 2; ++vi) { |
5681 | SmallVector<uint32_t, 16> Indices; |
5682 | for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) |
5683 | Indices.push_back(2*i+vi); |
5684 | |
5685 | Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); |
5686 | SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vuzp"); |
5687 | SV = Builder.CreateDefaultAlignedStore(SV, Addr); |
5688 | } |
5689 | return SV; |
5690 | } |
5691 | case NEON::BI__builtin_neon_vzip_v: |
5692 | case NEON::BI__builtin_neon_vzipq_v: { |
5693 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); |
5694 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
5695 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
5696 | Value *SV = nullptr; |
5697 | |
5698 | for (unsigned vi = 0; vi != 2; ++vi) { |
5699 | SmallVector<uint32_t, 16> Indices; |
5700 | for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { |
5701 | Indices.push_back((i + vi*e) >> 1); |
5702 | Indices.push_back(((i + vi*e) >> 1)+e); |
5703 | } |
5704 | Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); |
5705 | SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vzip"); |
5706 | SV = Builder.CreateDefaultAlignedStore(SV, Addr); |
5707 | } |
5708 | return SV; |
5709 | } |
5710 | case NEON::BI__builtin_neon_vdot_v: |
5711 | case NEON::BI__builtin_neon_vdotq_v: { |
5712 | llvm::Type *InputTy = |
5713 | llvm::VectorType::get(Int8Ty, Ty->getPrimitiveSizeInBits() / 8); |
5714 | llvm::Type *Tys[2] = { Ty, InputTy }; |
5715 | Int = Usgn ? LLVMIntrinsic : AltLLVMIntrinsic; |
5716 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vdot"); |
5717 | } |
5718 | case NEON::BI__builtin_neon_vfmlal_low_v: |
5719 | case NEON::BI__builtin_neon_vfmlalq_low_v: { |
5720 | llvm::Type *InputTy = |
5721 | llvm::VectorType::get(HalfTy, Ty->getPrimitiveSizeInBits() / 16); |
5722 | llvm::Type *Tys[2] = { Ty, InputTy }; |
5723 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vfmlal_low"); |
5724 | } |
5725 | case NEON::BI__builtin_neon_vfmlsl_low_v: |
5726 | case NEON::BI__builtin_neon_vfmlslq_low_v: { |
5727 | llvm::Type *InputTy = |
5728 | llvm::VectorType::get(HalfTy, Ty->getPrimitiveSizeInBits() / 16); |
5729 | llvm::Type *Tys[2] = { Ty, InputTy }; |
5730 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vfmlsl_low"); |
5731 | } |
5732 | case NEON::BI__builtin_neon_vfmlal_high_v: |
5733 | case NEON::BI__builtin_neon_vfmlalq_high_v: { |
5734 | llvm::Type *InputTy = |
5735 | llvm::VectorType::get(HalfTy, Ty->getPrimitiveSizeInBits() / 16); |
5736 | llvm::Type *Tys[2] = { Ty, InputTy }; |
5737 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vfmlal_high"); |
5738 | } |
5739 | case NEON::BI__builtin_neon_vfmlsl_high_v: |
5740 | case NEON::BI__builtin_neon_vfmlslq_high_v: { |
5741 | llvm::Type *InputTy = |
5742 | llvm::VectorType::get(HalfTy, Ty->getPrimitiveSizeInBits() / 16); |
5743 | llvm::Type *Tys[2] = { Ty, InputTy }; |
5744 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vfmlsl_high"); |
5745 | } |
5746 | } |
5747 | |
5748 | assert(Int && "Expected valid intrinsic number")((Int && "Expected valid intrinsic number") ? static_cast <void> (0) : __assert_fail ("Int && \"Expected valid intrinsic number\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 5748, __PRETTY_FUNCTION__)); |
5749 | |
5750 | // Determine the type(s) of this overloaded AArch64 intrinsic. |
5751 | Function *F = LookupNeonLLVMIntrinsic(Int, Modifier, Ty, E); |
5752 | |
5753 | Value *Result = EmitNeonCall(F, Ops, NameHint); |
5754 | llvm::Type *ResultType = ConvertType(E->getType()); |
5755 | // AArch64 intrinsic one-element vector type cast to |
5756 | // scalar type expected by the builtin |
5757 | return Builder.CreateBitCast(Result, ResultType, NameHint); |
5758 | } |
5759 | |
5760 | Value *CodeGenFunction::EmitAArch64CompareBuiltinExpr( |
5761 | Value *Op, llvm::Type *Ty, const CmpInst::Predicate Fp, |
5762 | const CmpInst::Predicate Ip, const Twine &Name) { |
5763 | llvm::Type *OTy = Op->getType(); |
5764 | |
5765 | // FIXME: this is utterly horrific. We should not be looking at previous |
5766 | // codegen context to find out what needs doing. Unfortunately TableGen |
5767 | // currently gives us exactly the same calls for vceqz_f32 and vceqz_s32 |
5768 | // (etc). |
5769 | if (BitCastInst *BI = dyn_cast<BitCastInst>(Op)) |
5770 | OTy = BI->getOperand(0)->getType(); |
5771 | |
5772 | Op = Builder.CreateBitCast(Op, OTy); |
5773 | if (OTy->getScalarType()->isFloatingPointTy()) { |
5774 | Op = Builder.CreateFCmp(Fp, Op, Constant::getNullValue(OTy)); |
5775 | } else { |
5776 | Op = Builder.CreateICmp(Ip, Op, Constant::getNullValue(OTy)); |
5777 | } |
5778 | return Builder.CreateSExt(Op, Ty, Name); |
5779 | } |
5780 | |
5781 | static Value *packTBLDVectorList(CodeGenFunction &CGF, ArrayRef<Value *> Ops, |
5782 | Value *ExtOp, Value *IndexOp, |
5783 | llvm::Type *ResTy, unsigned IntID, |
5784 | const char *Name) { |
5785 | SmallVector<Value *, 2> TblOps; |
5786 | if (ExtOp) |
5787 | TblOps.push_back(ExtOp); |
5788 | |
5789 | // Build a vector containing sequential number like (0, 1, 2, ..., 15) |
5790 | SmallVector<uint32_t, 16> Indices; |
5791 | llvm::VectorType *TblTy = cast<llvm::VectorType>(Ops[0]->getType()); |
5792 | for (unsigned i = 0, e = TblTy->getNumElements(); i != e; ++i) { |
5793 | Indices.push_back(2*i); |
5794 | Indices.push_back(2*i+1); |
5795 | } |
5796 | |
5797 | int PairPos = 0, End = Ops.size() - 1; |
5798 | while (PairPos < End) { |
5799 | TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], |
5800 | Ops[PairPos+1], Indices, |
5801 | Name)); |
5802 | PairPos += 2; |
5803 | } |
5804 | |
5805 | // If there's an odd number of 64-bit lookup table, fill the high 64-bit |
5806 | // of the 128-bit lookup table with zero. |
5807 | if (PairPos == End) { |
5808 | Value *ZeroTbl = ConstantAggregateZero::get(TblTy); |
5809 | TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], |
5810 | ZeroTbl, Indices, Name)); |
5811 | } |
5812 | |
5813 | Function *TblF; |
5814 | TblOps.push_back(IndexOp); |
5815 | TblF = CGF.CGM.getIntrinsic(IntID, ResTy); |
5816 | |
5817 | return CGF.EmitNeonCall(TblF, TblOps, Name); |
5818 | } |
5819 | |
5820 | Value *CodeGenFunction::GetValueForARMHint(unsigned BuiltinID) { |
5821 | unsigned Value; |
5822 | switch (BuiltinID) { |
5823 | default: |
5824 | return nullptr; |
5825 | case ARM::BI__builtin_arm_nop: |
5826 | Value = 0; |
5827 | break; |
5828 | case ARM::BI__builtin_arm_yield: |
5829 | case ARM::BI__yield: |
5830 | Value = 1; |
5831 | break; |
5832 | case ARM::BI__builtin_arm_wfe: |
5833 | case ARM::BI__wfe: |
5834 | Value = 2; |
5835 | break; |
5836 | case ARM::BI__builtin_arm_wfi: |
5837 | case ARM::BI__wfi: |
5838 | Value = 3; |
5839 | break; |
5840 | case ARM::BI__builtin_arm_sev: |
5841 | case ARM::BI__sev: |
5842 | Value = 4; |
5843 | break; |
5844 | case ARM::BI__builtin_arm_sevl: |
5845 | case ARM::BI__sevl: |
5846 | Value = 5; |
5847 | break; |
5848 | } |
5849 | |
5850 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), |
5851 | llvm::ConstantInt::get(Int32Ty, Value)); |
5852 | } |
5853 | |
5854 | // Generates the IR for the read/write special register builtin, |
5855 | // ValueType is the type of the value that is to be written or read, |
5856 | // RegisterType is the type of the register being written to or read from. |
5857 | static Value *EmitSpecialRegisterBuiltin(CodeGenFunction &CGF, |
5858 | const CallExpr *E, |
5859 | llvm::Type *RegisterType, |
5860 | llvm::Type *ValueType, |
5861 | bool IsRead, |
5862 | StringRef SysReg = "") { |
5863 | // write and register intrinsics only support 32 and 64 bit operations. |
5864 | assert((RegisterType->isIntegerTy(32) || RegisterType->isIntegerTy(64))(((RegisterType->isIntegerTy(32) || RegisterType->isIntegerTy (64)) && "Unsupported size for register.") ? static_cast <void> (0) : __assert_fail ("(RegisterType->isIntegerTy(32) || RegisterType->isIntegerTy(64)) && \"Unsupported size for register.\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 5865, __PRETTY_FUNCTION__)) |
5865 | && "Unsupported size for register.")(((RegisterType->isIntegerTy(32) || RegisterType->isIntegerTy (64)) && "Unsupported size for register.") ? static_cast <void> (0) : __assert_fail ("(RegisterType->isIntegerTy(32) || RegisterType->isIntegerTy(64)) && \"Unsupported size for register.\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 5865, __PRETTY_FUNCTION__)); |
5866 | |
5867 | CodeGen::CGBuilderTy &Builder = CGF.Builder; |
5868 | CodeGen::CodeGenModule &CGM = CGF.CGM; |
5869 | LLVMContext &Context = CGM.getLLVMContext(); |
5870 | |
5871 | if (SysReg.empty()) { |
5872 | const Expr *SysRegStrExpr = E->getArg(0)->IgnoreParenCasts(); |
5873 | SysReg = cast<clang::StringLiteral>(SysRegStrExpr)->getString(); |
5874 | } |
5875 | |
5876 | llvm::Metadata *Ops[] = { llvm::MDString::get(Context, SysReg) }; |
5877 | llvm::MDNode *RegName = llvm::MDNode::get(Context, Ops); |
5878 | llvm::Value *Metadata = llvm::MetadataAsValue::get(Context, RegName); |
5879 | |
5880 | llvm::Type *Types[] = { RegisterType }; |
5881 | |
5882 | bool MixedTypes = RegisterType->isIntegerTy(64) && ValueType->isIntegerTy(32); |
5883 | assert(!(RegisterType->isIntegerTy(32) && ValueType->isIntegerTy(64))((!(RegisterType->isIntegerTy(32) && ValueType-> isIntegerTy(64)) && "Can't fit 64-bit value in 32-bit register" ) ? static_cast<void> (0) : __assert_fail ("!(RegisterType->isIntegerTy(32) && ValueType->isIntegerTy(64)) && \"Can't fit 64-bit value in 32-bit register\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 5884, __PRETTY_FUNCTION__)) |
5884 | && "Can't fit 64-bit value in 32-bit register")((!(RegisterType->isIntegerTy(32) && ValueType-> isIntegerTy(64)) && "Can't fit 64-bit value in 32-bit register" ) ? static_cast<void> (0) : __assert_fail ("!(RegisterType->isIntegerTy(32) && ValueType->isIntegerTy(64)) && \"Can't fit 64-bit value in 32-bit register\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 5884, __PRETTY_FUNCTION__)); |
5885 | |
5886 | if (IsRead) { |
5887 | llvm::Function *F = CGM.getIntrinsic(llvm::Intrinsic::read_register, Types); |
5888 | llvm::Value *Call = Builder.CreateCall(F, Metadata); |
5889 | |
5890 | if (MixedTypes) |
5891 | // Read into 64 bit register and then truncate result to 32 bit. |
5892 | return Builder.CreateTrunc(Call, ValueType); |
5893 | |
5894 | if (ValueType->isPointerTy()) |
5895 | // Have i32/i64 result (Call) but want to return a VoidPtrTy (i8*). |
5896 | return Builder.CreateIntToPtr(Call, ValueType); |
5897 | |
5898 | return Call; |
5899 | } |
5900 | |
5901 | llvm::Function *F = CGM.getIntrinsic(llvm::Intrinsic::write_register, Types); |
5902 | llvm::Value *ArgValue = CGF.EmitScalarExpr(E->getArg(1)); |
5903 | if (MixedTypes) { |
5904 | // Extend 32 bit write value to 64 bit to pass to write. |
5905 | ArgValue = Builder.CreateZExt(ArgValue, RegisterType); |
5906 | return Builder.CreateCall(F, { Metadata, ArgValue }); |
5907 | } |
5908 | |
5909 | if (ValueType->isPointerTy()) { |
5910 | // Have VoidPtrTy ArgValue but want to return an i32/i64. |
5911 | ArgValue = Builder.CreatePtrToInt(ArgValue, RegisterType); |
5912 | return Builder.CreateCall(F, { Metadata, ArgValue }); |
5913 | } |
5914 | |
5915 | return Builder.CreateCall(F, { Metadata, ArgValue }); |
5916 | } |
5917 | |
5918 | /// Return true if BuiltinID is an overloaded Neon intrinsic with an extra |
5919 | /// argument that specifies the vector type. |
5920 | static bool HasExtraNeonArgument(unsigned BuiltinID) { |
5921 | switch (BuiltinID) { |
5922 | default: break; |
5923 | case NEON::BI__builtin_neon_vget_lane_i8: |
5924 | case NEON::BI__builtin_neon_vget_lane_i16: |
5925 | case NEON::BI__builtin_neon_vget_lane_i32: |
5926 | case NEON::BI__builtin_neon_vget_lane_i64: |
5927 | case NEON::BI__builtin_neon_vget_lane_f32: |
5928 | case NEON::BI__builtin_neon_vgetq_lane_i8: |
5929 | case NEON::BI__builtin_neon_vgetq_lane_i16: |
5930 | case NEON::BI__builtin_neon_vgetq_lane_i32: |
5931 | case NEON::BI__builtin_neon_vgetq_lane_i64: |
5932 | case NEON::BI__builtin_neon_vgetq_lane_f32: |
5933 | case NEON::BI__builtin_neon_vset_lane_i8: |
5934 | case NEON::BI__builtin_neon_vset_lane_i16: |
5935 | case NEON::BI__builtin_neon_vset_lane_i32: |
5936 | case NEON::BI__builtin_neon_vset_lane_i64: |
5937 | case NEON::BI__builtin_neon_vset_lane_f32: |
5938 | case NEON::BI__builtin_neon_vsetq_lane_i8: |
5939 | case NEON::BI__builtin_neon_vsetq_lane_i16: |
5940 | case NEON::BI__builtin_neon_vsetq_lane_i32: |
5941 | case NEON::BI__builtin_neon_vsetq_lane_i64: |
5942 | case NEON::BI__builtin_neon_vsetq_lane_f32: |
5943 | case NEON::BI__builtin_neon_vsha1h_u32: |
5944 | case NEON::BI__builtin_neon_vsha1cq_u32: |
5945 | case NEON::BI__builtin_neon_vsha1pq_u32: |
5946 | case NEON::BI__builtin_neon_vsha1mq_u32: |
5947 | case clang::ARM::BI_MoveToCoprocessor: |
5948 | case clang::ARM::BI_MoveToCoprocessor2: |
5949 | return false; |
5950 | } |
5951 | return true; |
5952 | } |
5953 | |
5954 | Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID, |
5955 | const CallExpr *E, |
5956 | llvm::Triple::ArchType Arch) { |
5957 | if (auto Hint = GetValueForARMHint(BuiltinID)) |
5958 | return Hint; |
5959 | |
5960 | if (BuiltinID == ARM::BI__emit) { |
5961 | bool IsThumb = getTarget().getTriple().getArch() == llvm::Triple::thumb; |
5962 | llvm::FunctionType *FTy = |
5963 | llvm::FunctionType::get(VoidTy, /*Variadic=*/false); |
5964 | |
5965 | Expr::EvalResult Result; |
5966 | if (!E->getArg(0)->EvaluateAsInt(Result, CGM.getContext())) |
5967 | llvm_unreachable("Sema will ensure that the parameter is constant")::llvm::llvm_unreachable_internal("Sema will ensure that the parameter is constant" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 5967); |
5968 | |
5969 | llvm::APSInt Value = Result.Val.getInt(); |
5970 | uint64_t ZExtValue = Value.zextOrTrunc(IsThumb ? 16 : 32).getZExtValue(); |
5971 | |
5972 | llvm::InlineAsm *Emit = |
5973 | IsThumb ? InlineAsm::get(FTy, ".inst.n 0x" + utohexstr(ZExtValue), "", |
5974 | /*SideEffects=*/true) |
5975 | : InlineAsm::get(FTy, ".inst 0x" + utohexstr(ZExtValue), "", |
5976 | /*SideEffects=*/true); |
5977 | |
5978 | return Builder.CreateCall(Emit); |
5979 | } |
5980 | |
5981 | if (BuiltinID == ARM::BI__builtin_arm_dbg) { |
5982 | Value *Option = EmitScalarExpr(E->getArg(0)); |
5983 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_dbg), Option); |
5984 | } |
5985 | |
5986 | if (BuiltinID == ARM::BI__builtin_arm_prefetch) { |
5987 | Value *Address = EmitScalarExpr(E->getArg(0)); |
5988 | Value *RW = EmitScalarExpr(E->getArg(1)); |
5989 | Value *IsData = EmitScalarExpr(E->getArg(2)); |
5990 | |
5991 | // Locality is not supported on ARM target |
5992 | Value *Locality = llvm::ConstantInt::get(Int32Ty, 3); |
5993 | |
5994 | Function *F = CGM.getIntrinsic(Intrinsic::prefetch); |
5995 | return Builder.CreateCall(F, {Address, RW, Locality, IsData}); |
5996 | } |
5997 | |
5998 | if (BuiltinID == ARM::BI__builtin_arm_rbit) { |
5999 | llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); |
6000 | return Builder.CreateCall( |
6001 | CGM.getIntrinsic(Intrinsic::bitreverse, Arg->getType()), Arg, "rbit"); |
6002 | } |
6003 | |
6004 | if (BuiltinID == ARM::BI__clear_cache) { |
6005 | assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments")((E->getNumArgs() == 2 && "__clear_cache takes 2 arguments" ) ? static_cast<void> (0) : __assert_fail ("E->getNumArgs() == 2 && \"__clear_cache takes 2 arguments\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 6005, __PRETTY_FUNCTION__)); |
6006 | const FunctionDecl *FD = E->getDirectCallee(); |
6007 | Value *Ops[2]; |
6008 | for (unsigned i = 0; i < 2; i++) |
6009 | Ops[i] = EmitScalarExpr(E->getArg(i)); |
6010 | llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); |
6011 | llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); |
6012 | StringRef Name = FD->getName(); |
6013 | return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); |
6014 | } |
6015 | |
6016 | if (BuiltinID == ARM::BI__builtin_arm_mcrr || |
6017 | BuiltinID == ARM::BI__builtin_arm_mcrr2) { |
6018 | Function *F; |
6019 | |
6020 | switch (BuiltinID) { |
6021 | default: llvm_unreachable("unexpected builtin")::llvm::llvm_unreachable_internal("unexpected builtin", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 6021); |
6022 | case ARM::BI__builtin_arm_mcrr: |
6023 | F = CGM.getIntrinsic(Intrinsic::arm_mcrr); |
6024 | break; |
6025 | case ARM::BI__builtin_arm_mcrr2: |
6026 | F = CGM.getIntrinsic(Intrinsic::arm_mcrr2); |
6027 | break; |
6028 | } |
6029 | |
6030 | // MCRR{2} instruction has 5 operands but |
6031 | // the intrinsic has 4 because Rt and Rt2 |
6032 | // are represented as a single unsigned 64 |
6033 | // bit integer in the intrinsic definition |
6034 | // but internally it's represented as 2 32 |
6035 | // bit integers. |
6036 | |
6037 | Value *Coproc = EmitScalarExpr(E->getArg(0)); |
6038 | Value *Opc1 = EmitScalarExpr(E->getArg(1)); |
6039 | Value *RtAndRt2 = EmitScalarExpr(E->getArg(2)); |
6040 | Value *CRm = EmitScalarExpr(E->getArg(3)); |
6041 | |
6042 | Value *C1 = llvm::ConstantInt::get(Int64Ty, 32); |
6043 | Value *Rt = Builder.CreateTruncOrBitCast(RtAndRt2, Int32Ty); |
6044 | Value *Rt2 = Builder.CreateLShr(RtAndRt2, C1); |
6045 | Rt2 = Builder.CreateTruncOrBitCast(Rt2, Int32Ty); |
6046 | |
6047 | return Builder.CreateCall(F, {Coproc, Opc1, Rt, Rt2, CRm}); |
6048 | } |
6049 | |
6050 | if (BuiltinID == ARM::BI__builtin_arm_mrrc || |
6051 | BuiltinID == ARM::BI__builtin_arm_mrrc2) { |
6052 | Function *F; |
6053 | |
6054 | switch (BuiltinID) { |
6055 | default: llvm_unreachable("unexpected builtin")::llvm::llvm_unreachable_internal("unexpected builtin", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 6055); |
6056 | case ARM::BI__builtin_arm_mrrc: |
6057 | F = CGM.getIntrinsic(Intrinsic::arm_mrrc); |
6058 | break; |
6059 | case ARM::BI__builtin_arm_mrrc2: |
6060 | F = CGM.getIntrinsic(Intrinsic::arm_mrrc2); |
6061 | break; |
6062 | } |
6063 | |
6064 | Value *Coproc = EmitScalarExpr(E->getArg(0)); |
6065 | Value *Opc1 = EmitScalarExpr(E->getArg(1)); |
6066 | Value *CRm = EmitScalarExpr(E->getArg(2)); |
6067 | Value *RtAndRt2 = Builder.CreateCall(F, {Coproc, Opc1, CRm}); |
6068 | |
6069 | // Returns an unsigned 64 bit integer, represented |
6070 | // as two 32 bit integers. |
6071 | |
6072 | Value *Rt = Builder.CreateExtractValue(RtAndRt2, 1); |
6073 | Value *Rt1 = Builder.CreateExtractValue(RtAndRt2, 0); |
6074 | Rt = Builder.CreateZExt(Rt, Int64Ty); |
6075 | Rt1 = Builder.CreateZExt(Rt1, Int64Ty); |
6076 | |
6077 | Value *ShiftCast = llvm::ConstantInt::get(Int64Ty, 32); |
6078 | RtAndRt2 = Builder.CreateShl(Rt, ShiftCast, "shl", true); |
6079 | RtAndRt2 = Builder.CreateOr(RtAndRt2, Rt1); |
6080 | |
6081 | return Builder.CreateBitCast(RtAndRt2, ConvertType(E->getType())); |
6082 | } |
6083 | |
6084 | if (BuiltinID == ARM::BI__builtin_arm_ldrexd || |
6085 | ((BuiltinID == ARM::BI__builtin_arm_ldrex || |
6086 | BuiltinID == ARM::BI__builtin_arm_ldaex) && |
6087 | getContext().getTypeSize(E->getType()) == 64) || |
6088 | BuiltinID == ARM::BI__ldrexd) { |
6089 | Function *F; |
6090 | |
6091 | switch (BuiltinID) { |
6092 | default: llvm_unreachable("unexpected builtin")::llvm::llvm_unreachable_internal("unexpected builtin", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 6092); |
6093 | case ARM::BI__builtin_arm_ldaex: |
6094 | F = CGM.getIntrinsic(Intrinsic::arm_ldaexd); |
6095 | break; |
6096 | case ARM::BI__builtin_arm_ldrexd: |
6097 | case ARM::BI__builtin_arm_ldrex: |
6098 | case ARM::BI__ldrexd: |
6099 | F = CGM.getIntrinsic(Intrinsic::arm_ldrexd); |
6100 | break; |
6101 | } |
6102 | |
6103 | Value *LdPtr = EmitScalarExpr(E->getArg(0)); |
6104 | Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), |
6105 | "ldrexd"); |
6106 | |
6107 | Value *Val0 = Builder.CreateExtractValue(Val, 1); |
6108 | Value *Val1 = Builder.CreateExtractValue(Val, 0); |
6109 | Val0 = Builder.CreateZExt(Val0, Int64Ty); |
6110 | Val1 = Builder.CreateZExt(Val1, Int64Ty); |
6111 | |
6112 | Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32); |
6113 | Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); |
6114 | Val = Builder.CreateOr(Val, Val1); |
6115 | return Builder.CreateBitCast(Val, ConvertType(E->getType())); |
6116 | } |
6117 | |
6118 | if (BuiltinID == ARM::BI__builtin_arm_ldrex || |
6119 | BuiltinID == ARM::BI__builtin_arm_ldaex) { |
6120 | Value *LoadAddr = EmitScalarExpr(E->getArg(0)); |
6121 | |
6122 | QualType Ty = E->getType(); |
6123 | llvm::Type *RealResTy = ConvertType(Ty); |
6124 | llvm::Type *PtrTy = llvm::IntegerType::get( |
6125 | getLLVMContext(), getContext().getTypeSize(Ty))->getPointerTo(); |
6126 | LoadAddr = Builder.CreateBitCast(LoadAddr, PtrTy); |
6127 | |
6128 | Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_ldaex |
6129 | ? Intrinsic::arm_ldaex |
6130 | : Intrinsic::arm_ldrex, |
6131 | PtrTy); |
6132 | Value *Val = Builder.CreateCall(F, LoadAddr, "ldrex"); |
6133 | |
6134 | if (RealResTy->isPointerTy()) |
6135 | return Builder.CreateIntToPtr(Val, RealResTy); |
6136 | else { |
6137 | llvm::Type *IntResTy = llvm::IntegerType::get( |
6138 | getLLVMContext(), CGM.getDataLayout().getTypeSizeInBits(RealResTy)); |
6139 | Val = Builder.CreateTruncOrBitCast(Val, IntResTy); |
6140 | return Builder.CreateBitCast(Val, RealResTy); |
6141 | } |
6142 | } |
6143 | |
6144 | if (BuiltinID == ARM::BI__builtin_arm_strexd || |
6145 | ((BuiltinID == ARM::BI__builtin_arm_stlex || |
6146 | BuiltinID == ARM::BI__builtin_arm_strex) && |
6147 | getContext().getTypeSize(E->getArg(0)->getType()) == 64)) { |
6148 | Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex |
6149 | ? Intrinsic::arm_stlexd |
6150 | : Intrinsic::arm_strexd); |
6151 | llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty); |
6152 | |
6153 | Address Tmp = CreateMemTemp(E->getArg(0)->getType()); |
6154 | Value *Val = EmitScalarExpr(E->getArg(0)); |
6155 | Builder.CreateStore(Val, Tmp); |
6156 | |
6157 | Address LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); |
6158 | Val = Builder.CreateLoad(LdPtr); |
6159 | |
6160 | Value *Arg0 = Builder.CreateExtractValue(Val, 0); |
6161 | Value *Arg1 = Builder.CreateExtractValue(Val, 1); |
6162 | Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), Int8PtrTy); |
6163 | return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "strexd"); |
6164 | } |
6165 | |
6166 | if (BuiltinID == ARM::BI__builtin_arm_strex || |
6167 | BuiltinID == ARM::BI__builtin_arm_stlex) { |
6168 | Value *StoreVal = EmitScalarExpr(E->getArg(0)); |
6169 | Value *StoreAddr = EmitScalarExpr(E->getArg(1)); |
6170 | |
6171 | QualType Ty = E->getArg(0)->getType(); |
6172 | llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), |
6173 | getContext().getTypeSize(Ty)); |
6174 | StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); |
6175 | |
6176 | if (StoreVal->getType()->isPointerTy()) |
6177 | StoreVal = Builder.CreatePtrToInt(StoreVal, Int32Ty); |
6178 | else { |
6179 | llvm::Type *IntTy = llvm::IntegerType::get( |
6180 | getLLVMContext(), |
6181 | CGM.getDataLayout().getTypeSizeInBits(StoreVal->getType())); |
6182 | StoreVal = Builder.CreateBitCast(StoreVal, IntTy); |
6183 | StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int32Ty); |
6184 | } |
6185 | |
6186 | Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex |
6187 | ? Intrinsic::arm_stlex |
6188 | : Intrinsic::arm_strex, |
6189 | StoreAddr->getType()); |
6190 | return Builder.CreateCall(F, {StoreVal, StoreAddr}, "strex"); |
6191 | } |
6192 | |
6193 | if (BuiltinID == ARM::BI__builtin_arm_clrex) { |
6194 | Function *F = CGM.getIntrinsic(Intrinsic::arm_clrex); |
6195 | return Builder.CreateCall(F); |
6196 | } |
6197 | |
6198 | // CRC32 |
6199 | Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; |
6200 | switch (BuiltinID) { |
6201 | case ARM::BI__builtin_arm_crc32b: |
6202 | CRCIntrinsicID = Intrinsic::arm_crc32b; break; |
6203 | case ARM::BI__builtin_arm_crc32cb: |
6204 | CRCIntrinsicID = Intrinsic::arm_crc32cb; break; |
6205 | case ARM::BI__builtin_arm_crc32h: |
6206 | CRCIntrinsicID = Intrinsic::arm_crc32h; break; |
6207 | case ARM::BI__builtin_arm_crc32ch: |
6208 | CRCIntrinsicID = Intrinsic::arm_crc32ch; break; |
6209 | case ARM::BI__builtin_arm_crc32w: |
6210 | case ARM::BI__builtin_arm_crc32d: |
6211 | CRCIntrinsicID = Intrinsic::arm_crc32w; break; |
6212 | case ARM::BI__builtin_arm_crc32cw: |
6213 | case ARM::BI__builtin_arm_crc32cd: |
6214 | CRCIntrinsicID = Intrinsic::arm_crc32cw; break; |
6215 | } |
6216 | |
6217 | if (CRCIntrinsicID != Intrinsic::not_intrinsic) { |
6218 | Value *Arg0 = EmitScalarExpr(E->getArg(0)); |
6219 | Value *Arg1 = EmitScalarExpr(E->getArg(1)); |
6220 | |
6221 | // crc32{c,}d intrinsics are implemnted as two calls to crc32{c,}w |
6222 | // intrinsics, hence we need different codegen for these cases. |
6223 | if (BuiltinID == ARM::BI__builtin_arm_crc32d || |
6224 | BuiltinID == ARM::BI__builtin_arm_crc32cd) { |
6225 | Value *C1 = llvm::ConstantInt::get(Int64Ty, 32); |
6226 | Value *Arg1a = Builder.CreateTruncOrBitCast(Arg1, Int32Ty); |
6227 | Value *Arg1b = Builder.CreateLShr(Arg1, C1); |
6228 | Arg1b = Builder.CreateTruncOrBitCast(Arg1b, Int32Ty); |
6229 | |
6230 | Function *F = CGM.getIntrinsic(CRCIntrinsicID); |
6231 | Value *Res = Builder.CreateCall(F, {Arg0, Arg1a}); |
6232 | return Builder.CreateCall(F, {Res, Arg1b}); |
6233 | } else { |
6234 | Arg1 = Builder.CreateZExtOrBitCast(Arg1, Int32Ty); |
6235 | |
6236 | Function *F = CGM.getIntrinsic(CRCIntrinsicID); |
6237 | return Builder.CreateCall(F, {Arg0, Arg1}); |
6238 | } |
6239 | } |
6240 | |
6241 | if (BuiltinID == ARM::BI__builtin_arm_rsr || |
6242 | BuiltinID == ARM::BI__builtin_arm_rsr64 || |
6243 | BuiltinID == ARM::BI__builtin_arm_rsrp || |
6244 | BuiltinID == ARM::BI__builtin_arm_wsr || |
6245 | BuiltinID == ARM::BI__builtin_arm_wsr64 || |
6246 | BuiltinID == ARM::BI__builtin_arm_wsrp) { |
6247 | |
6248 | bool IsRead = BuiltinID == ARM::BI__builtin_arm_rsr || |
6249 | BuiltinID == ARM::BI__builtin_arm_rsr64 || |
6250 | BuiltinID == ARM::BI__builtin_arm_rsrp; |
6251 | |
6252 | bool IsPointerBuiltin = BuiltinID == ARM::BI__builtin_arm_rsrp || |
6253 | BuiltinID == ARM::BI__builtin_arm_wsrp; |
6254 | |
6255 | bool Is64Bit = BuiltinID == ARM::BI__builtin_arm_rsr64 || |
6256 | BuiltinID == ARM::BI__builtin_arm_wsr64; |
6257 | |
6258 | llvm::Type *ValueType; |
6259 | llvm::Type *RegisterType; |
6260 | if (IsPointerBuiltin) { |
6261 | ValueType = VoidPtrTy; |
6262 | RegisterType = Int32Ty; |
6263 | } else if (Is64Bit) { |
6264 | ValueType = RegisterType = Int64Ty; |
6265 | } else { |
6266 | ValueType = RegisterType = Int32Ty; |
6267 | } |
6268 | |
6269 | return EmitSpecialRegisterBuiltin(*this, E, RegisterType, ValueType, IsRead); |
6270 | } |
6271 | |
6272 | // Find out if any arguments are required to be integer constant |
6273 | // expressions. |
6274 | unsigned ICEArguments = 0; |
6275 | ASTContext::GetBuiltinTypeError Error; |
6276 | getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); |
6277 | assert(Error == ASTContext::GE_None && "Should not codegen an error")((Error == ASTContext::GE_None && "Should not codegen an error" ) ? static_cast<void> (0) : __assert_fail ("Error == ASTContext::GE_None && \"Should not codegen an error\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 6277, __PRETTY_FUNCTION__)); |
6278 | |
6279 | auto getAlignmentValue32 = [&](Address addr) -> Value* { |
6280 | return Builder.getInt32(addr.getAlignment().getQuantity()); |
6281 | }; |
6282 | |
6283 | Address PtrOp0 = Address::invalid(); |
6284 | Address PtrOp1 = Address::invalid(); |
6285 | SmallVector<Value*, 4> Ops; |
6286 | bool HasExtraArg = HasExtraNeonArgument(BuiltinID); |
6287 | unsigned NumArgs = E->getNumArgs() - (HasExtraArg ? 1 : 0); |
6288 | for (unsigned i = 0, e = NumArgs; i != e; i++) { |
6289 | if (i == 0) { |
6290 | switch (BuiltinID) { |
6291 | case NEON::BI__builtin_neon_vld1_v: |
6292 | case NEON::BI__builtin_neon_vld1q_v: |
6293 | case NEON::BI__builtin_neon_vld1q_lane_v: |
6294 | case NEON::BI__builtin_neon_vld1_lane_v: |
6295 | case NEON::BI__builtin_neon_vld1_dup_v: |
6296 | case NEON::BI__builtin_neon_vld1q_dup_v: |
6297 | case NEON::BI__builtin_neon_vst1_v: |
6298 | case NEON::BI__builtin_neon_vst1q_v: |
6299 | case NEON::BI__builtin_neon_vst1q_lane_v: |
6300 | case NEON::BI__builtin_neon_vst1_lane_v: |
6301 | case NEON::BI__builtin_neon_vst2_v: |
6302 | case NEON::BI__builtin_neon_vst2q_v: |
6303 | case NEON::BI__builtin_neon_vst2_lane_v: |
6304 | case NEON::BI__builtin_neon_vst2q_lane_v: |
6305 | case NEON::BI__builtin_neon_vst3_v: |
6306 | case NEON::BI__builtin_neon_vst3q_v: |
6307 | case NEON::BI__builtin_neon_vst3_lane_v: |
6308 | case NEON::BI__builtin_neon_vst3q_lane_v: |
6309 | case NEON::BI__builtin_neon_vst4_v: |
6310 | case NEON::BI__builtin_neon_vst4q_v: |
6311 | case NEON::BI__builtin_neon_vst4_lane_v: |
6312 | case NEON::BI__builtin_neon_vst4q_lane_v: |
6313 | // Get the alignment for the argument in addition to the value; |
6314 | // we'll use it later. |
6315 | PtrOp0 = EmitPointerWithAlignment(E->getArg(0)); |
6316 | Ops.push_back(PtrOp0.getPointer()); |
6317 | continue; |
6318 | } |
6319 | } |
6320 | if (i == 1) { |
6321 | switch (BuiltinID) { |
6322 | case NEON::BI__builtin_neon_vld2_v: |
6323 | case NEON::BI__builtin_neon_vld2q_v: |
6324 | case NEON::BI__builtin_neon_vld3_v: |
6325 | case NEON::BI__builtin_neon_vld3q_v: |
6326 | case NEON::BI__builtin_neon_vld4_v: |
6327 | case NEON::BI__builtin_neon_vld4q_v: |
6328 | case NEON::BI__builtin_neon_vld2_lane_v: |
6329 | case NEON::BI__builtin_neon_vld2q_lane_v: |
6330 | case NEON::BI__builtin_neon_vld3_lane_v: |
6331 | case NEON::BI__builtin_neon_vld3q_lane_v: |
6332 | case NEON::BI__builtin_neon_vld4_lane_v: |
6333 | case NEON::BI__builtin_neon_vld4q_lane_v: |
6334 | case NEON::BI__builtin_neon_vld2_dup_v: |
6335 | case NEON::BI__builtin_neon_vld2q_dup_v: |
6336 | case NEON::BI__builtin_neon_vld3_dup_v: |
6337 | case NEON::BI__builtin_neon_vld3q_dup_v: |
6338 | case NEON::BI__builtin_neon_vld4_dup_v: |
6339 | case NEON::BI__builtin_neon_vld4q_dup_v: |
6340 | // Get the alignment for the argument in addition to the value; |
6341 | // we'll use it later. |
6342 | PtrOp1 = EmitPointerWithAlignment(E->getArg(1)); |
6343 | Ops.push_back(PtrOp1.getPointer()); |
6344 | continue; |
6345 | } |
6346 | } |
6347 | |
6348 | if ((ICEArguments & (1 << i)) == 0) { |
6349 | Ops.push_back(EmitScalarExpr(E->getArg(i))); |
6350 | } else { |
6351 | // If this is required to be a constant, constant fold it so that we know |
6352 | // that the generated intrinsic gets a ConstantInt. |
6353 | llvm::APSInt Result; |
6354 | bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); |
6355 | assert(IsConst && "Constant arg isn't actually constant?")((IsConst && "Constant arg isn't actually constant?") ? static_cast<void> (0) : __assert_fail ("IsConst && \"Constant arg isn't actually constant?\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 6355, __PRETTY_FUNCTION__)); (void)IsConst; |
6356 | Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); |
6357 | } |
6358 | } |
6359 | |
6360 | switch (BuiltinID) { |
6361 | default: break; |
6362 | |
6363 | case NEON::BI__builtin_neon_vget_lane_i8: |
6364 | case NEON::BI__builtin_neon_vget_lane_i16: |
6365 | case NEON::BI__builtin_neon_vget_lane_i32: |
6366 | case NEON::BI__builtin_neon_vget_lane_i64: |
6367 | case NEON::BI__builtin_neon_vget_lane_f32: |
6368 | case NEON::BI__builtin_neon_vgetq_lane_i8: |
6369 | case NEON::BI__builtin_neon_vgetq_lane_i16: |
6370 | case NEON::BI__builtin_neon_vgetq_lane_i32: |
6371 | case NEON::BI__builtin_neon_vgetq_lane_i64: |
6372 | case NEON::BI__builtin_neon_vgetq_lane_f32: |
6373 | return Builder.CreateExtractElement(Ops[0], Ops[1], "vget_lane"); |
6374 | |
6375 | case NEON::BI__builtin_neon_vrndns_f32: { |
6376 | Value *Arg = EmitScalarExpr(E->getArg(0)); |
6377 | llvm::Type *Tys[] = {Arg->getType()}; |
6378 | Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vrintn, Tys); |
6379 | return Builder.CreateCall(F, {Arg}, "vrndn"); } |
6380 | |
6381 | case NEON::BI__builtin_neon_vset_lane_i8: |
6382 | case NEON::BI__builtin_neon_vset_lane_i16: |
6383 | case NEON::BI__builtin_neon_vset_lane_i32: |
6384 | case NEON::BI__builtin_neon_vset_lane_i64: |
6385 | case NEON::BI__builtin_neon_vset_lane_f32: |
6386 | case NEON::BI__builtin_neon_vsetq_lane_i8: |
6387 | case NEON::BI__builtin_neon_vsetq_lane_i16: |
6388 | case NEON::BI__builtin_neon_vsetq_lane_i32: |
6389 | case NEON::BI__builtin_neon_vsetq_lane_i64: |
6390 | case NEON::BI__builtin_neon_vsetq_lane_f32: |
6391 | return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); |
6392 | |
6393 | case NEON::BI__builtin_neon_vsha1h_u32: |
6394 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1h), Ops, |
6395 | "vsha1h"); |
6396 | case NEON::BI__builtin_neon_vsha1cq_u32: |
6397 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1c), Ops, |
6398 | "vsha1h"); |
6399 | case NEON::BI__builtin_neon_vsha1pq_u32: |
6400 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1p), Ops, |
6401 | "vsha1h"); |
6402 | case NEON::BI__builtin_neon_vsha1mq_u32: |
6403 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1m), Ops, |
6404 | "vsha1h"); |
6405 | |
6406 | // The ARM _MoveToCoprocessor builtins put the input register value as |
6407 | // the first argument, but the LLVM intrinsic expects it as the third one. |
6408 | case ARM::BI_MoveToCoprocessor: |
6409 | case ARM::BI_MoveToCoprocessor2: { |
6410 | Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI_MoveToCoprocessor ? |
6411 | Intrinsic::arm_mcr : Intrinsic::arm_mcr2); |
6412 | return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0], |
6413 | Ops[3], Ops[4], Ops[5]}); |
6414 | } |
6415 | case ARM::BI_BitScanForward: |
6416 | case ARM::BI_BitScanForward64: |
6417 | return EmitMSVCBuiltinExpr(MSVCIntrin::_BitScanForward, E); |
6418 | case ARM::BI_BitScanReverse: |
6419 | case ARM::BI_BitScanReverse64: |
6420 | return EmitMSVCBuiltinExpr(MSVCIntrin::_BitScanReverse, E); |
6421 | |
6422 | case ARM::BI_InterlockedAnd64: |
6423 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedAnd, E); |
6424 | case ARM::BI_InterlockedExchange64: |
6425 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchange, E); |
6426 | case ARM::BI_InterlockedExchangeAdd64: |
6427 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeAdd, E); |
6428 | case ARM::BI_InterlockedExchangeSub64: |
6429 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeSub, E); |
6430 | case ARM::BI_InterlockedOr64: |
6431 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedOr, E); |
6432 | case ARM::BI_InterlockedXor64: |
6433 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedXor, E); |
6434 | case ARM::BI_InterlockedDecrement64: |
6435 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedDecrement, E); |
6436 | case ARM::BI_InterlockedIncrement64: |
6437 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedIncrement, E); |
6438 | case ARM::BI_InterlockedExchangeAdd8_acq: |
6439 | case ARM::BI_InterlockedExchangeAdd16_acq: |
6440 | case ARM::BI_InterlockedExchangeAdd_acq: |
6441 | case ARM::BI_InterlockedExchangeAdd64_acq: |
6442 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeAdd_acq, E); |
6443 | case ARM::BI_InterlockedExchangeAdd8_rel: |
6444 | case ARM::BI_InterlockedExchangeAdd16_rel: |
6445 | case ARM::BI_InterlockedExchangeAdd_rel: |
6446 | case ARM::BI_InterlockedExchangeAdd64_rel: |
6447 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeAdd_rel, E); |
6448 | case ARM::BI_InterlockedExchangeAdd8_nf: |
6449 | case ARM::BI_InterlockedExchangeAdd16_nf: |
6450 | case ARM::BI_InterlockedExchangeAdd_nf: |
6451 | case ARM::BI_InterlockedExchangeAdd64_nf: |
6452 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeAdd_nf, E); |
6453 | case ARM::BI_InterlockedExchange8_acq: |
6454 | case ARM::BI_InterlockedExchange16_acq: |
6455 | case ARM::BI_InterlockedExchange_acq: |
6456 | case ARM::BI_InterlockedExchange64_acq: |
6457 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchange_acq, E); |
6458 | case ARM::BI_InterlockedExchange8_rel: |
6459 | case ARM::BI_InterlockedExchange16_rel: |
6460 | case ARM::BI_InterlockedExchange_rel: |
6461 | case ARM::BI_InterlockedExchange64_rel: |
6462 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchange_rel, E); |
6463 | case ARM::BI_InterlockedExchange8_nf: |
6464 | case ARM::BI_InterlockedExchange16_nf: |
6465 | case ARM::BI_InterlockedExchange_nf: |
6466 | case ARM::BI_InterlockedExchange64_nf: |
6467 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchange_nf, E); |
6468 | case ARM::BI_InterlockedCompareExchange8_acq: |
6469 | case ARM::BI_InterlockedCompareExchange16_acq: |
6470 | case ARM::BI_InterlockedCompareExchange_acq: |
6471 | case ARM::BI_InterlockedCompareExchange64_acq: |
6472 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedCompareExchange_acq, E); |
6473 | case ARM::BI_InterlockedCompareExchange8_rel: |
6474 | case ARM::BI_InterlockedCompareExchange16_rel: |
6475 | case ARM::BI_InterlockedCompareExchange_rel: |
6476 | case ARM::BI_InterlockedCompareExchange64_rel: |
6477 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedCompareExchange_rel, E); |
6478 | case ARM::BI_InterlockedCompareExchange8_nf: |
6479 | case ARM::BI_InterlockedCompareExchange16_nf: |
6480 | case ARM::BI_InterlockedCompareExchange_nf: |
6481 | case ARM::BI_InterlockedCompareExchange64_nf: |
6482 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedCompareExchange_nf, E); |
6483 | case ARM::BI_InterlockedOr8_acq: |
6484 | case ARM::BI_InterlockedOr16_acq: |
6485 | case ARM::BI_InterlockedOr_acq: |
6486 | case ARM::BI_InterlockedOr64_acq: |
6487 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedOr_acq, E); |
6488 | case ARM::BI_InterlockedOr8_rel: |
6489 | case ARM::BI_InterlockedOr16_rel: |
6490 | case ARM::BI_InterlockedOr_rel: |
6491 | case ARM::BI_InterlockedOr64_rel: |
6492 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedOr_rel, E); |
6493 | case ARM::BI_InterlockedOr8_nf: |
6494 | case ARM::BI_InterlockedOr16_nf: |
6495 | case ARM::BI_InterlockedOr_nf: |
6496 | case ARM::BI_InterlockedOr64_nf: |
6497 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedOr_nf, E); |
6498 | case ARM::BI_InterlockedXor8_acq: |
6499 | case ARM::BI_InterlockedXor16_acq: |
6500 | case ARM::BI_InterlockedXor_acq: |
6501 | case ARM::BI_InterlockedXor64_acq: |
6502 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedXor_acq, E); |
6503 | case ARM::BI_InterlockedXor8_rel: |
6504 | case ARM::BI_InterlockedXor16_rel: |
6505 | case ARM::BI_InterlockedXor_rel: |
6506 | case ARM::BI_InterlockedXor64_rel: |
6507 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedXor_rel, E); |
6508 | case ARM::BI_InterlockedXor8_nf: |
6509 | case ARM::BI_InterlockedXor16_nf: |
6510 | case ARM::BI_InterlockedXor_nf: |
6511 | case ARM::BI_InterlockedXor64_nf: |
6512 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedXor_nf, E); |
6513 | case ARM::BI_InterlockedAnd8_acq: |
6514 | case ARM::BI_InterlockedAnd16_acq: |
6515 | case ARM::BI_InterlockedAnd_acq: |
6516 | case ARM::BI_InterlockedAnd64_acq: |
6517 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedAnd_acq, E); |
6518 | case ARM::BI_InterlockedAnd8_rel: |
6519 | case ARM::BI_InterlockedAnd16_rel: |
6520 | case ARM::BI_InterlockedAnd_rel: |
6521 | case ARM::BI_InterlockedAnd64_rel: |
6522 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedAnd_rel, E); |
6523 | case ARM::BI_InterlockedAnd8_nf: |
6524 | case ARM::BI_InterlockedAnd16_nf: |
6525 | case ARM::BI_InterlockedAnd_nf: |
6526 | case ARM::BI_InterlockedAnd64_nf: |
6527 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedAnd_nf, E); |
6528 | case ARM::BI_InterlockedIncrement16_acq: |
6529 | case ARM::BI_InterlockedIncrement_acq: |
6530 | case ARM::BI_InterlockedIncrement64_acq: |
6531 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedIncrement_acq, E); |
6532 | case ARM::BI_InterlockedIncrement16_rel: |
6533 | case ARM::BI_InterlockedIncrement_rel: |
6534 | case ARM::BI_InterlockedIncrement64_rel: |
6535 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedIncrement_rel, E); |
6536 | case ARM::BI_InterlockedIncrement16_nf: |
6537 | case ARM::BI_InterlockedIncrement_nf: |
6538 | case ARM::BI_InterlockedIncrement64_nf: |
6539 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedIncrement_nf, E); |
6540 | case ARM::BI_InterlockedDecrement16_acq: |
6541 | case ARM::BI_InterlockedDecrement_acq: |
6542 | case ARM::BI_InterlockedDecrement64_acq: |
6543 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedDecrement_acq, E); |
6544 | case ARM::BI_InterlockedDecrement16_rel: |
6545 | case ARM::BI_InterlockedDecrement_rel: |
6546 | case ARM::BI_InterlockedDecrement64_rel: |
6547 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedDecrement_rel, E); |
6548 | case ARM::BI_InterlockedDecrement16_nf: |
6549 | case ARM::BI_InterlockedDecrement_nf: |
6550 | case ARM::BI_InterlockedDecrement64_nf: |
6551 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedDecrement_nf, E); |
6552 | } |
6553 | |
6554 | // Get the last argument, which specifies the vector type. |
6555 | assert(HasExtraArg)((HasExtraArg) ? static_cast<void> (0) : __assert_fail ( "HasExtraArg", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 6555, __PRETTY_FUNCTION__)); |
6556 | llvm::APSInt Result; |
6557 | const Expr *Arg = E->getArg(E->getNumArgs()-1); |
6558 | if (!Arg->isIntegerConstantExpr(Result, getContext())) |
6559 | return nullptr; |
6560 | |
6561 | if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f || |
6562 | BuiltinID == ARM::BI__builtin_arm_vcvtr_d) { |
6563 | // Determine the overloaded type of this builtin. |
6564 | llvm::Type *Ty; |
6565 | if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f) |
6566 | Ty = FloatTy; |
6567 | else |
6568 | Ty = DoubleTy; |
6569 | |
6570 | // Determine whether this is an unsigned conversion or not. |
6571 | bool usgn = Result.getZExtValue() == 1; |
6572 | unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr; |
6573 | |
6574 | // Call the appropriate intrinsic. |
6575 | Function *F = CGM.getIntrinsic(Int, Ty); |
6576 | return Builder.CreateCall(F, Ops, "vcvtr"); |
6577 | } |
6578 | |
6579 | // Determine the type of this overloaded NEON intrinsic. |
6580 | NeonTypeFlags Type(Result.getZExtValue()); |
6581 | bool usgn = Type.isUnsigned(); |
6582 | bool rightShift = false; |
6583 | |
6584 | llvm::VectorType *VTy = GetNeonType(this, Type, |
6585 | getTarget().hasLegalHalfType()); |
6586 | llvm::Type *Ty = VTy; |
6587 | if (!Ty) |
6588 | return nullptr; |
6589 | |
6590 | // Many NEON builtins have identical semantics and uses in ARM and |
6591 | // AArch64. Emit these in a single function. |
6592 | auto IntrinsicMap = makeArrayRef(ARMSIMDIntrinsicMap); |
6593 | const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( |
6594 | IntrinsicMap, BuiltinID, NEONSIMDIntrinsicsProvenSorted); |
6595 | if (Builtin) |
6596 | return EmitCommonNeonBuiltinExpr( |
6597 | Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, |
6598 | Builtin->NameHint, Builtin->TypeModifier, E, Ops, PtrOp0, PtrOp1, Arch); |
6599 | |
6600 | unsigned Int; |
6601 | switch (BuiltinID) { |
6602 | default: return nullptr; |
6603 | case NEON::BI__builtin_neon_vld1q_lane_v: |
6604 | // Handle 64-bit integer elements as a special case. Use shuffles of |
6605 | // one-element vectors to avoid poor code for i64 in the backend. |
6606 | if (VTy->getElementType()->isIntegerTy(64)) { |
6607 | // Extract the other lane. |
6608 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
6609 | uint32_t Lane = cast<ConstantInt>(Ops[2])->getZExtValue(); |
6610 | Value *SV = llvm::ConstantVector::get(ConstantInt::get(Int32Ty, 1-Lane)); |
6611 | Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); |
6612 | // Load the value as a one-element vector. |
6613 | Ty = llvm::VectorType::get(VTy->getElementType(), 1); |
6614 | llvm::Type *Tys[] = {Ty, Int8PtrTy}; |
6615 | Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Tys); |
6616 | Value *Align = getAlignmentValue32(PtrOp0); |
6617 | Value *Ld = Builder.CreateCall(F, {Ops[0], Align}); |
6618 | // Combine them. |
6619 | uint32_t Indices[] = {1 - Lane, Lane}; |
6620 | SV = llvm::ConstantDataVector::get(getLLVMContext(), Indices); |
6621 | return Builder.CreateShuffleVector(Ops[1], Ld, SV, "vld1q_lane"); |
6622 | } |
6623 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
6624 | case NEON::BI__builtin_neon_vld1_lane_v: { |
6625 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
6626 | PtrOp0 = Builder.CreateElementBitCast(PtrOp0, VTy->getElementType()); |
6627 | Value *Ld = Builder.CreateLoad(PtrOp0); |
6628 | return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane"); |
6629 | } |
6630 | case NEON::BI__builtin_neon_vqrshrn_n_v: |
6631 | Int = |
6632 | usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns; |
6633 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n", |
6634 | 1, true); |
6635 | case NEON::BI__builtin_neon_vqrshrun_n_v: |
6636 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty), |
6637 | Ops, "vqrshrun_n", 1, true); |
6638 | case NEON::BI__builtin_neon_vqshrn_n_v: |
6639 | Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns; |
6640 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n", |
6641 | 1, true); |
6642 | case NEON::BI__builtin_neon_vqshrun_n_v: |
6643 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty), |
6644 | Ops, "vqshrun_n", 1, true); |
6645 | case NEON::BI__builtin_neon_vrecpe_v: |
6646 | case NEON::BI__builtin_neon_vrecpeq_v: |
6647 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty), |
6648 | Ops, "vrecpe"); |
6649 | case NEON::BI__builtin_neon_vrshrn_n_v: |
6650 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty), |
6651 | Ops, "vrshrn_n", 1, true); |
6652 | case NEON::BI__builtin_neon_vrsra_n_v: |
6653 | case NEON::BI__builtin_neon_vrsraq_n_v: |
6654 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
6655 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
6656 | Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true); |
6657 | Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; |
6658 | Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Ty), {Ops[1], Ops[2]}); |
6659 | return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n"); |
6660 | case NEON::BI__builtin_neon_vsri_n_v: |
6661 | case NEON::BI__builtin_neon_vsriq_n_v: |
6662 | rightShift = true; |
6663 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
6664 | case NEON::BI__builtin_neon_vsli_n_v: |
6665 | case NEON::BI__builtin_neon_vsliq_n_v: |
6666 | Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift); |
6667 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty), |
6668 | Ops, "vsli_n"); |
6669 | case NEON::BI__builtin_neon_vsra_n_v: |
6670 | case NEON::BI__builtin_neon_vsraq_n_v: |
6671 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
6672 | Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); |
6673 | return Builder.CreateAdd(Ops[0], Ops[1]); |
6674 | case NEON::BI__builtin_neon_vst1q_lane_v: |
6675 | // Handle 64-bit integer elements as a special case. Use a shuffle to get |
6676 | // a one-element vector and avoid poor code for i64 in the backend. |
6677 | if (VTy->getElementType()->isIntegerTy(64)) { |
6678 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
6679 | Value *SV = llvm::ConstantVector::get(cast<llvm::Constant>(Ops[2])); |
6680 | Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); |
6681 | Ops[2] = getAlignmentValue32(PtrOp0); |
6682 | llvm::Type *Tys[] = {Int8PtrTy, Ops[1]->getType()}; |
6683 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, |
6684 | Tys), Ops); |
6685 | } |
6686 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
6687 | case NEON::BI__builtin_neon_vst1_lane_v: { |
6688 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
6689 | Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); |
6690 | Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); |
6691 | auto St = Builder.CreateStore(Ops[1], Builder.CreateBitCast(PtrOp0, Ty)); |
6692 | return St; |
6693 | } |
6694 | case NEON::BI__builtin_neon_vtbl1_v: |
6695 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1), |
6696 | Ops, "vtbl1"); |
6697 | case NEON::BI__builtin_neon_vtbl2_v: |
6698 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2), |
6699 | Ops, "vtbl2"); |
6700 | case NEON::BI__builtin_neon_vtbl3_v: |
6701 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3), |
6702 | Ops, "vtbl3"); |
6703 | case NEON::BI__builtin_neon_vtbl4_v: |
6704 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4), |
6705 | Ops, "vtbl4"); |
6706 | case NEON::BI__builtin_neon_vtbx1_v: |
6707 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1), |
6708 | Ops, "vtbx1"); |
6709 | case NEON::BI__builtin_neon_vtbx2_v: |
6710 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2), |
6711 | Ops, "vtbx2"); |
6712 | case NEON::BI__builtin_neon_vtbx3_v: |
6713 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3), |
6714 | Ops, "vtbx3"); |
6715 | case NEON::BI__builtin_neon_vtbx4_v: |
6716 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4), |
6717 | Ops, "vtbx4"); |
6718 | } |
6719 | } |
6720 | |
6721 | static Value *EmitAArch64TblBuiltinExpr(CodeGenFunction &CGF, unsigned BuiltinID, |
6722 | const CallExpr *E, |
6723 | SmallVectorImpl<Value *> &Ops, |
6724 | llvm::Triple::ArchType Arch) { |
6725 | unsigned int Int = 0; |
6726 | const char *s = nullptr; |
6727 | |
6728 | switch (BuiltinID) { |
6729 | default: |
6730 | return nullptr; |
6731 | case NEON::BI__builtin_neon_vtbl1_v: |
6732 | case NEON::BI__builtin_neon_vqtbl1_v: |
6733 | case NEON::BI__builtin_neon_vqtbl1q_v: |
6734 | case NEON::BI__builtin_neon_vtbl2_v: |
6735 | case NEON::BI__builtin_neon_vqtbl2_v: |
6736 | case NEON::BI__builtin_neon_vqtbl2q_v: |
6737 | case NEON::BI__builtin_neon_vtbl3_v: |
6738 | case NEON::BI__builtin_neon_vqtbl3_v: |
6739 | case NEON::BI__builtin_neon_vqtbl3q_v: |
6740 | case NEON::BI__builtin_neon_vtbl4_v: |
6741 | case NEON::BI__builtin_neon_vqtbl4_v: |
6742 | case NEON::BI__builtin_neon_vqtbl4q_v: |
6743 | break; |
6744 | case NEON::BI__builtin_neon_vtbx1_v: |
6745 | case NEON::BI__builtin_neon_vqtbx1_v: |
6746 | case NEON::BI__builtin_neon_vqtbx1q_v: |
6747 | case NEON::BI__builtin_neon_vtbx2_v: |
6748 | case NEON::BI__builtin_neon_vqtbx2_v: |
6749 | case NEON::BI__builtin_neon_vqtbx2q_v: |
6750 | case NEON::BI__builtin_neon_vtbx3_v: |
6751 | case NEON::BI__builtin_neon_vqtbx3_v: |
6752 | case NEON::BI__builtin_neon_vqtbx3q_v: |
6753 | case NEON::BI__builtin_neon_vtbx4_v: |
6754 | case NEON::BI__builtin_neon_vqtbx4_v: |
6755 | case NEON::BI__builtin_neon_vqtbx4q_v: |
6756 | break; |
6757 | } |
6758 | |
6759 | assert(E->getNumArgs() >= 3)((E->getNumArgs() >= 3) ? static_cast<void> (0) : __assert_fail ("E->getNumArgs() >= 3", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 6759, __PRETTY_FUNCTION__)); |
6760 | |
6761 | // Get the last argument, which specifies the vector type. |
6762 | llvm::APSInt Result; |
6763 | const Expr *Arg = E->getArg(E->getNumArgs() - 1); |
6764 | if (!Arg->isIntegerConstantExpr(Result, CGF.getContext())) |
6765 | return nullptr; |
6766 | |
6767 | // Determine the type of this overloaded NEON intrinsic. |
6768 | NeonTypeFlags Type(Result.getZExtValue()); |
6769 | llvm::VectorType *Ty = GetNeonType(&CGF, Type); |
6770 | if (!Ty) |
6771 | return nullptr; |
6772 | |
6773 | CodeGen::CGBuilderTy &Builder = CGF.Builder; |
6774 | |
6775 | // AArch64 scalar builtins are not overloaded, they do not have an extra |
6776 | // argument that specifies the vector type, need to handle each case. |
6777 | switch (BuiltinID) { |
6778 | case NEON::BI__builtin_neon_vtbl1_v: { |
6779 | return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 1), nullptr, |
6780 | Ops[1], Ty, Intrinsic::aarch64_neon_tbl1, |
6781 | "vtbl1"); |
6782 | } |
6783 | case NEON::BI__builtin_neon_vtbl2_v: { |
6784 | return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 2), nullptr, |
6785 | Ops[2], Ty, Intrinsic::aarch64_neon_tbl1, |
6786 | "vtbl1"); |
6787 | } |
6788 | case NEON::BI__builtin_neon_vtbl3_v: { |
6789 | return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 3), nullptr, |
6790 | Ops[3], Ty, Intrinsic::aarch64_neon_tbl2, |
6791 | "vtbl2"); |
6792 | } |
6793 | case NEON::BI__builtin_neon_vtbl4_v: { |
6794 | return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 4), nullptr, |
6795 | Ops[4], Ty, Intrinsic::aarch64_neon_tbl2, |
6796 | "vtbl2"); |
6797 | } |
6798 | case NEON::BI__builtin_neon_vtbx1_v: { |
6799 | Value *TblRes = |
6800 | packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 1), nullptr, Ops[2], |
6801 | Ty, Intrinsic::aarch64_neon_tbl1, "vtbl1"); |
6802 | |
6803 | llvm::Constant *EightV = ConstantInt::get(Ty, 8); |
6804 | Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[2], EightV); |
6805 | CmpRes = Builder.CreateSExt(CmpRes, Ty); |
6806 | |
6807 | Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); |
6808 | Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); |
6809 | return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); |
6810 | } |
6811 | case NEON::BI__builtin_neon_vtbx2_v: { |
6812 | return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 2), Ops[0], |
6813 | Ops[3], Ty, Intrinsic::aarch64_neon_tbx1, |
6814 | "vtbx1"); |
6815 | } |
6816 | case NEON::BI__builtin_neon_vtbx3_v: { |
6817 | Value *TblRes = |
6818 | packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 3), nullptr, Ops[4], |
6819 | Ty, Intrinsic::aarch64_neon_tbl2, "vtbl2"); |
6820 | |
6821 | llvm::Constant *TwentyFourV = ConstantInt::get(Ty, 24); |
6822 | Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[4], |
6823 | TwentyFourV); |
6824 | CmpRes = Builder.CreateSExt(CmpRes, Ty); |
6825 | |
6826 | Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); |
6827 | Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); |
6828 | return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); |
6829 | } |
6830 | case NEON::BI__builtin_neon_vtbx4_v: { |
6831 | return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 4), Ops[0], |
6832 | Ops[5], Ty, Intrinsic::aarch64_neon_tbx2, |
6833 | "vtbx2"); |
6834 | } |
6835 | case NEON::BI__builtin_neon_vqtbl1_v: |
6836 | case NEON::BI__builtin_neon_vqtbl1q_v: |
6837 | Int = Intrinsic::aarch64_neon_tbl1; s = "vtbl1"; break; |
6838 | case NEON::BI__builtin_neon_vqtbl2_v: |
6839 | case NEON::BI__builtin_neon_vqtbl2q_v: { |
6840 | Int = Intrinsic::aarch64_neon_tbl2; s = "vtbl2"; break; |
6841 | case NEON::BI__builtin_neon_vqtbl3_v: |
6842 | case NEON::BI__builtin_neon_vqtbl3q_v: |
6843 | Int = Intrinsic::aarch64_neon_tbl3; s = "vtbl3"; break; |
6844 | case NEON::BI__builtin_neon_vqtbl4_v: |
6845 | case NEON::BI__builtin_neon_vqtbl4q_v: |
6846 | Int = Intrinsic::aarch64_neon_tbl4; s = "vtbl4"; break; |
6847 | case NEON::BI__builtin_neon_vqtbx1_v: |
6848 | case NEON::BI__builtin_neon_vqtbx1q_v: |
6849 | Int = Intrinsic::aarch64_neon_tbx1; s = "vtbx1"; break; |
6850 | case NEON::BI__builtin_neon_vqtbx2_v: |
6851 | case NEON::BI__builtin_neon_vqtbx2q_v: |
6852 | Int = Intrinsic::aarch64_neon_tbx2; s = "vtbx2"; break; |
6853 | case NEON::BI__builtin_neon_vqtbx3_v: |
6854 | case NEON::BI__builtin_neon_vqtbx3q_v: |
6855 | Int = Intrinsic::aarch64_neon_tbx3; s = "vtbx3"; break; |
6856 | case NEON::BI__builtin_neon_vqtbx4_v: |
6857 | case NEON::BI__builtin_neon_vqtbx4q_v: |
6858 | Int = Intrinsic::aarch64_neon_tbx4; s = "vtbx4"; break; |
6859 | } |
6860 | } |
6861 | |
6862 | if (!Int) |
6863 | return nullptr; |
6864 | |
6865 | Function *F = CGF.CGM.getIntrinsic(Int, Ty); |
6866 | return CGF.EmitNeonCall(F, Ops, s); |
6867 | } |
6868 | |
6869 | Value *CodeGenFunction::vectorWrapScalar16(Value *Op) { |
6870 | llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4); |
6871 | Op = Builder.CreateBitCast(Op, Int16Ty); |
6872 | Value *V = UndefValue::get(VTy); |
6873 | llvm::Constant *CI = ConstantInt::get(SizeTy, 0); |
6874 | Op = Builder.CreateInsertElement(V, Op, CI); |
6875 | return Op; |
6876 | } |
6877 | |
6878 | Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID, |
6879 | const CallExpr *E, |
6880 | llvm::Triple::ArchType Arch) { |
6881 | unsigned HintID = static_cast<unsigned>(-1); |
6882 | switch (BuiltinID) { |
6883 | default: break; |
6884 | case AArch64::BI__builtin_arm_nop: |
6885 | HintID = 0; |
6886 | break; |
6887 | case AArch64::BI__builtin_arm_yield: |
6888 | case AArch64::BI__yield: |
6889 | HintID = 1; |
6890 | break; |
6891 | case AArch64::BI__builtin_arm_wfe: |
6892 | case AArch64::BI__wfe: |
6893 | HintID = 2; |
6894 | break; |
6895 | case AArch64::BI__builtin_arm_wfi: |
6896 | case AArch64::BI__wfi: |
6897 | HintID = 3; |
6898 | break; |
6899 | case AArch64::BI__builtin_arm_sev: |
6900 | case AArch64::BI__sev: |
6901 | HintID = 4; |
6902 | break; |
6903 | case AArch64::BI__builtin_arm_sevl: |
6904 | case AArch64::BI__sevl: |
6905 | HintID = 5; |
6906 | break; |
6907 | } |
6908 | |
6909 | if (HintID != static_cast<unsigned>(-1)) { |
6910 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_hint); |
6911 | return Builder.CreateCall(F, llvm::ConstantInt::get(Int32Ty, HintID)); |
6912 | } |
6913 | |
6914 | if (BuiltinID == AArch64::BI__builtin_arm_prefetch) { |
6915 | Value *Address = EmitScalarExpr(E->getArg(0)); |
6916 | Value *RW = EmitScalarExpr(E->getArg(1)); |
6917 | Value *CacheLevel = EmitScalarExpr(E->getArg(2)); |
6918 | Value *RetentionPolicy = EmitScalarExpr(E->getArg(3)); |
6919 | Value *IsData = EmitScalarExpr(E->getArg(4)); |
6920 | |
6921 | Value *Locality = nullptr; |
6922 | if (cast<llvm::ConstantInt>(RetentionPolicy)->isZero()) { |
6923 | // Temporal fetch, needs to convert cache level to locality. |
6924 | Locality = llvm::ConstantInt::get(Int32Ty, |
6925 | -cast<llvm::ConstantInt>(CacheLevel)->getValue() + 3); |
6926 | } else { |
6927 | // Streaming fetch. |
6928 | Locality = llvm::ConstantInt::get(Int32Ty, 0); |
6929 | } |
6930 | |
6931 | // FIXME: We need AArch64 specific LLVM intrinsic if we want to specify |
6932 | // PLDL3STRM or PLDL2STRM. |
6933 | Function *F = CGM.getIntrinsic(Intrinsic::prefetch); |
6934 | return Builder.CreateCall(F, {Address, RW, Locality, IsData}); |
6935 | } |
6936 | |
6937 | if (BuiltinID == AArch64::BI__builtin_arm_rbit) { |
6938 | assert((getContext().getTypeSize(E->getType()) == 32) &&(((getContext().getTypeSize(E->getType()) == 32) && "rbit of unusual size!") ? static_cast<void> (0) : __assert_fail ("(getContext().getTypeSize(E->getType()) == 32) && \"rbit of unusual size!\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 6939, __PRETTY_FUNCTION__)) |
6939 | "rbit of unusual size!")(((getContext().getTypeSize(E->getType()) == 32) && "rbit of unusual size!") ? static_cast<void> (0) : __assert_fail ("(getContext().getTypeSize(E->getType()) == 32) && \"rbit of unusual size!\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 6939, __PRETTY_FUNCTION__)); |
6940 | llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); |
6941 | return Builder.CreateCall( |
6942 | CGM.getIntrinsic(Intrinsic::bitreverse, Arg->getType()), Arg, "rbit"); |
6943 | } |
6944 | if (BuiltinID == AArch64::BI__builtin_arm_rbit64) { |
6945 | assert((getContext().getTypeSize(E->getType()) == 64) &&(((getContext().getTypeSize(E->getType()) == 64) && "rbit of unusual size!") ? static_cast<void> (0) : __assert_fail ("(getContext().getTypeSize(E->getType()) == 64) && \"rbit of unusual size!\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 6946, __PRETTY_FUNCTION__)) |
6946 | "rbit of unusual size!")(((getContext().getTypeSize(E->getType()) == 64) && "rbit of unusual size!") ? static_cast<void> (0) : __assert_fail ("(getContext().getTypeSize(E->getType()) == 64) && \"rbit of unusual size!\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 6946, __PRETTY_FUNCTION__)); |
6947 | llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); |
6948 | return Builder.CreateCall( |
6949 | CGM.getIntrinsic(Intrinsic::bitreverse, Arg->getType()), Arg, "rbit"); |
6950 | } |
6951 | |
6952 | if (BuiltinID == AArch64::BI__clear_cache) { |
6953 | assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments")((E->getNumArgs() == 2 && "__clear_cache takes 2 arguments" ) ? static_cast<void> (0) : __assert_fail ("E->getNumArgs() == 2 && \"__clear_cache takes 2 arguments\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 6953, __PRETTY_FUNCTION__)); |
6954 | const FunctionDecl *FD = E->getDirectCallee(); |
6955 | Value *Ops[2]; |
6956 | for (unsigned i = 0; i < 2; i++) |
6957 | Ops[i] = EmitScalarExpr(E->getArg(i)); |
6958 | llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); |
6959 | llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); |
6960 | StringRef Name = FD->getName(); |
6961 | return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); |
6962 | } |
6963 | |
6964 | if ((BuiltinID == AArch64::BI__builtin_arm_ldrex || |
6965 | BuiltinID == AArch64::BI__builtin_arm_ldaex) && |
6966 | getContext().getTypeSize(E->getType()) == 128) { |
6967 | Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex |
6968 | ? Intrinsic::aarch64_ldaxp |
6969 | : Intrinsic::aarch64_ldxp); |
6970 | |
6971 | Value *LdPtr = EmitScalarExpr(E->getArg(0)); |
6972 | Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), |
6973 | "ldxp"); |
6974 | |
6975 | Value *Val0 = Builder.CreateExtractValue(Val, 1); |
6976 | Value *Val1 = Builder.CreateExtractValue(Val, 0); |
6977 | llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128); |
6978 | Val0 = Builder.CreateZExt(Val0, Int128Ty); |
6979 | Val1 = Builder.CreateZExt(Val1, Int128Ty); |
6980 | |
6981 | Value *ShiftCst = llvm::ConstantInt::get(Int128Ty, 64); |
6982 | Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); |
6983 | Val = Builder.CreateOr(Val, Val1); |
6984 | return Builder.CreateBitCast(Val, ConvertType(E->getType())); |
6985 | } else if (BuiltinID == AArch64::BI__builtin_arm_ldrex || |
6986 | BuiltinID == AArch64::BI__builtin_arm_ldaex) { |
6987 | Value *LoadAddr = EmitScalarExpr(E->getArg(0)); |
6988 | |
6989 | QualType Ty = E->getType(); |
6990 | llvm::Type *RealResTy = ConvertType(Ty); |
6991 | llvm::Type *PtrTy = llvm::IntegerType::get( |
6992 | getLLVMContext(), getContext().getTypeSize(Ty))->getPointerTo(); |
6993 | LoadAddr = Builder.CreateBitCast(LoadAddr, PtrTy); |
6994 | |
6995 | Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex |
6996 | ? Intrinsic::aarch64_ldaxr |
6997 | : Intrinsic::aarch64_ldxr, |
6998 | PtrTy); |
6999 | Value *Val = Builder.CreateCall(F, LoadAddr, "ldxr"); |
7000 | |
7001 | if (RealResTy->isPointerTy()) |
7002 | return Builder.CreateIntToPtr(Val, RealResTy); |
7003 | |
7004 | llvm::Type *IntResTy = llvm::IntegerType::get( |
7005 | getLLVMContext(), CGM.getDataLayout().getTypeSizeInBits(RealResTy)); |
7006 | Val = Builder.CreateTruncOrBitCast(Val, IntResTy); |
7007 | return Builder.CreateBitCast(Val, RealResTy); |
7008 | } |
7009 | |
7010 | if ((BuiltinID == AArch64::BI__builtin_arm_strex || |
7011 | BuiltinID == AArch64::BI__builtin_arm_stlex) && |
7012 | getContext().getTypeSize(E->getArg(0)->getType()) == 128) { |
7013 | Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex |
7014 | ? Intrinsic::aarch64_stlxp |
7015 | : Intrinsic::aarch64_stxp); |
7016 | llvm::Type *STy = llvm::StructType::get(Int64Ty, Int64Ty); |
7017 | |
7018 | Address Tmp = CreateMemTemp(E->getArg(0)->getType()); |
7019 | EmitAnyExprToMem(E->getArg(0), Tmp, Qualifiers(), /*init*/ true); |
7020 | |
7021 | Tmp = Builder.CreateBitCast(Tmp, llvm::PointerType::getUnqual(STy)); |
7022 | llvm::Value *Val = Builder.CreateLoad(Tmp); |
7023 | |
7024 | Value *Arg0 = Builder.CreateExtractValue(Val, 0); |
7025 | Value *Arg1 = Builder.CreateExtractValue(Val, 1); |
7026 | Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), |
7027 | Int8PtrTy); |
7028 | return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "stxp"); |
7029 | } |
7030 | |
7031 | if (BuiltinID == AArch64::BI__builtin_arm_strex || |
7032 | BuiltinID == AArch64::BI__builtin_arm_stlex) { |
7033 | Value *StoreVal = EmitScalarExpr(E->getArg(0)); |
7034 | Value *StoreAddr = EmitScalarExpr(E->getArg(1)); |
7035 | |
7036 | QualType Ty = E->getArg(0)->getType(); |
7037 | llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), |
7038 | getContext().getTypeSize(Ty)); |
7039 | StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); |
7040 | |
7041 | if (StoreVal->getType()->isPointerTy()) |
7042 | StoreVal = Builder.CreatePtrToInt(StoreVal, Int64Ty); |
7043 | else { |
7044 | llvm::Type *IntTy = llvm::IntegerType::get( |
7045 | getLLVMContext(), |
7046 | CGM.getDataLayout().getTypeSizeInBits(StoreVal->getType())); |
7047 | StoreVal = Builder.CreateBitCast(StoreVal, IntTy); |
7048 | StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int64Ty); |
7049 | } |
7050 | |
7051 | Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex |
7052 | ? Intrinsic::aarch64_stlxr |
7053 | : Intrinsic::aarch64_stxr, |
7054 | StoreAddr->getType()); |
7055 | return Builder.CreateCall(F, {StoreVal, StoreAddr}, "stxr"); |
7056 | } |
7057 | |
7058 | if (BuiltinID == AArch64::BI__getReg) { |
7059 | Expr::EvalResult Result; |
7060 | if (!E->getArg(0)->EvaluateAsInt(Result, CGM.getContext())) |
7061 | llvm_unreachable("Sema will ensure that the parameter is constant")::llvm::llvm_unreachable_internal("Sema will ensure that the parameter is constant" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 7061); |
7062 | |
7063 | llvm::APSInt Value = Result.Val.getInt(); |
7064 | LLVMContext &Context = CGM.getLLVMContext(); |
7065 | std::string Reg = Value == 31 ? "sp" : "x" + Value.toString(10); |
7066 | |
7067 | llvm::Metadata *Ops[] = {llvm::MDString::get(Context, Reg)}; |
7068 | llvm::MDNode *RegName = llvm::MDNode::get(Context, Ops); |
7069 | llvm::Value *Metadata = llvm::MetadataAsValue::get(Context, RegName); |
7070 | |
7071 | llvm::Function *F = |
7072 | CGM.getIntrinsic(llvm::Intrinsic::read_register, {Int64Ty}); |
7073 | return Builder.CreateCall(F, Metadata); |
7074 | } |
7075 | |
7076 | if (BuiltinID == AArch64::BI__builtin_arm_clrex) { |
7077 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_clrex); |
7078 | return Builder.CreateCall(F); |
7079 | } |
7080 | |
7081 | if (BuiltinID == AArch64::BI_ReadWriteBarrier) |
7082 | return Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent, |
7083 | llvm::SyncScope::SingleThread); |
7084 | |
7085 | // CRC32 |
7086 | Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; |
7087 | switch (BuiltinID) { |
7088 | case AArch64::BI__builtin_arm_crc32b: |
7089 | CRCIntrinsicID = Intrinsic::aarch64_crc32b; break; |
7090 | case AArch64::BI__builtin_arm_crc32cb: |
7091 | CRCIntrinsicID = Intrinsic::aarch64_crc32cb; break; |
7092 | case AArch64::BI__builtin_arm_crc32h: |
7093 | CRCIntrinsicID = Intrinsic::aarch64_crc32h; break; |
7094 | case AArch64::BI__builtin_arm_crc32ch: |
7095 | CRCIntrinsicID = Intrinsic::aarch64_crc32ch; break; |
7096 | case AArch64::BI__builtin_arm_crc32w: |
7097 | CRCIntrinsicID = Intrinsic::aarch64_crc32w; break; |
7098 | case AArch64::BI__builtin_arm_crc32cw: |
7099 | CRCIntrinsicID = Intrinsic::aarch64_crc32cw; break; |
7100 | case AArch64::BI__builtin_arm_crc32d: |
7101 | CRCIntrinsicID = Intrinsic::aarch64_crc32x; break; |
7102 | case AArch64::BI__builtin_arm_crc32cd: |
7103 | CRCIntrinsicID = Intrinsic::aarch64_crc32cx; break; |
7104 | } |
7105 | |
7106 | if (CRCIntrinsicID != Intrinsic::not_intrinsic) { |
7107 | Value *Arg0 = EmitScalarExpr(E->getArg(0)); |
7108 | Value *Arg1 = EmitScalarExpr(E->getArg(1)); |
7109 | Function *F = CGM.getIntrinsic(CRCIntrinsicID); |
7110 | |
7111 | llvm::Type *DataTy = F->getFunctionType()->getParamType(1); |
7112 | Arg1 = Builder.CreateZExtOrBitCast(Arg1, DataTy); |
7113 | |
7114 | return Builder.CreateCall(F, {Arg0, Arg1}); |
7115 | } |
7116 | |
7117 | // Memory Tagging Extensions (MTE) Intrinsics |
7118 | Intrinsic::ID MTEIntrinsicID = Intrinsic::not_intrinsic; |
7119 | switch (BuiltinID) { |
7120 | case AArch64::BI__builtin_arm_irg: |
7121 | MTEIntrinsicID = Intrinsic::aarch64_irg; break; |
7122 | case AArch64::BI__builtin_arm_addg: |
7123 | MTEIntrinsicID = Intrinsic::aarch64_addg; break; |
7124 | case AArch64::BI__builtin_arm_gmi: |
7125 | MTEIntrinsicID = Intrinsic::aarch64_gmi; break; |
7126 | case AArch64::BI__builtin_arm_ldg: |
7127 | MTEIntrinsicID = Intrinsic::aarch64_ldg; break; |
7128 | case AArch64::BI__builtin_arm_stg: |
7129 | MTEIntrinsicID = Intrinsic::aarch64_stg; break; |
7130 | case AArch64::BI__builtin_arm_subp: |
7131 | MTEIntrinsicID = Intrinsic::aarch64_subp; break; |
7132 | } |
7133 | |
7134 | if (MTEIntrinsicID != Intrinsic::not_intrinsic) { |
7135 | llvm::Type *T = ConvertType(E->getType()); |
7136 | |
7137 | if (MTEIntrinsicID == Intrinsic::aarch64_irg) { |
7138 | Value *Pointer = EmitScalarExpr(E->getArg(0)); |
7139 | Value *Mask = EmitScalarExpr(E->getArg(1)); |
7140 | |
7141 | Pointer = Builder.CreatePointerCast(Pointer, Int8PtrTy); |
7142 | Mask = Builder.CreateZExt(Mask, Int64Ty); |
7143 | Value *RV = Builder.CreateCall( |
7144 | CGM.getIntrinsic(MTEIntrinsicID), {Pointer, Mask}); |
7145 | return Builder.CreatePointerCast(RV, T); |
7146 | } |
7147 | if (MTEIntrinsicID == Intrinsic::aarch64_addg) { |
7148 | Value *Pointer = EmitScalarExpr(E->getArg(0)); |
7149 | Value *TagOffset = EmitScalarExpr(E->getArg(1)); |
7150 | |
7151 | Pointer = Builder.CreatePointerCast(Pointer, Int8PtrTy); |
7152 | TagOffset = Builder.CreateZExt(TagOffset, Int64Ty); |
7153 | Value *RV = Builder.CreateCall( |
7154 | CGM.getIntrinsic(MTEIntrinsicID), {Pointer, TagOffset}); |
7155 | return Builder.CreatePointerCast(RV, T); |
7156 | } |
7157 | if (MTEIntrinsicID == Intrinsic::aarch64_gmi) { |
7158 | Value *Pointer = EmitScalarExpr(E->getArg(0)); |
7159 | Value *ExcludedMask = EmitScalarExpr(E->getArg(1)); |
7160 | |
7161 | ExcludedMask = Builder.CreateZExt(ExcludedMask, Int64Ty); |
7162 | Pointer = Builder.CreatePointerCast(Pointer, Int8PtrTy); |
7163 | return Builder.CreateCall( |
7164 | CGM.getIntrinsic(MTEIntrinsicID), {Pointer, ExcludedMask}); |
7165 | } |
7166 | // Although it is possible to supply a different return |
7167 | // address (first arg) to this intrinsic, for now we set |
7168 | // return address same as input address. |
7169 | if (MTEIntrinsicID == Intrinsic::aarch64_ldg) { |
7170 | Value *TagAddress = EmitScalarExpr(E->getArg(0)); |
7171 | TagAddress = Builder.CreatePointerCast(TagAddress, Int8PtrTy); |
7172 | Value *RV = Builder.CreateCall( |
7173 | CGM.getIntrinsic(MTEIntrinsicID), {TagAddress, TagAddress}); |
7174 | return Builder.CreatePointerCast(RV, T); |
7175 | } |
7176 | // Although it is possible to supply a different tag (to set) |
7177 | // to this intrinsic (as first arg), for now we supply |
7178 | // the tag that is in input address arg (common use case). |
7179 | if (MTEIntrinsicID == Intrinsic::aarch64_stg) { |
7180 | Value *TagAddress = EmitScalarExpr(E->getArg(0)); |
7181 | TagAddress = Builder.CreatePointerCast(TagAddress, Int8PtrTy); |
7182 | return Builder.CreateCall( |
7183 | CGM.getIntrinsic(MTEIntrinsicID), {TagAddress, TagAddress}); |
7184 | } |
7185 | if (MTEIntrinsicID == Intrinsic::aarch64_subp) { |
7186 | Value *PointerA = EmitScalarExpr(E->getArg(0)); |
7187 | Value *PointerB = EmitScalarExpr(E->getArg(1)); |
7188 | PointerA = Builder.CreatePointerCast(PointerA, Int8PtrTy); |
7189 | PointerB = Builder.CreatePointerCast(PointerB, Int8PtrTy); |
7190 | return Builder.CreateCall( |
7191 | CGM.getIntrinsic(MTEIntrinsicID), {PointerA, PointerB}); |
7192 | } |
7193 | } |
7194 | |
7195 | if (BuiltinID == AArch64::BI__builtin_arm_rsr || |
7196 | BuiltinID == AArch64::BI__builtin_arm_rsr64 || |
7197 | BuiltinID == AArch64::BI__builtin_arm_rsrp || |
7198 | BuiltinID == AArch64::BI__builtin_arm_wsr || |
7199 | BuiltinID == AArch64::BI__builtin_arm_wsr64 || |
7200 | BuiltinID == AArch64::BI__builtin_arm_wsrp) { |
7201 | |
7202 | bool IsRead = BuiltinID == AArch64::BI__builtin_arm_rsr || |
7203 | BuiltinID == AArch64::BI__builtin_arm_rsr64 || |
7204 | BuiltinID == AArch64::BI__builtin_arm_rsrp; |
7205 | |
7206 | bool IsPointerBuiltin = BuiltinID == AArch64::BI__builtin_arm_rsrp || |
7207 | BuiltinID == AArch64::BI__builtin_arm_wsrp; |
7208 | |
7209 | bool Is64Bit = BuiltinID != AArch64::BI__builtin_arm_rsr && |
7210 | BuiltinID != AArch64::BI__builtin_arm_wsr; |
7211 | |
7212 | llvm::Type *ValueType; |
7213 | llvm::Type *RegisterType = Int64Ty; |
7214 | if (IsPointerBuiltin) { |
7215 | ValueType = VoidPtrTy; |
7216 | } else if (Is64Bit) { |
7217 | ValueType = Int64Ty; |
7218 | } else { |
7219 | ValueType = Int32Ty; |
7220 | } |
7221 | |
7222 | return EmitSpecialRegisterBuiltin(*this, E, RegisterType, ValueType, IsRead); |
7223 | } |
7224 | |
7225 | if (BuiltinID == AArch64::BI_ReadStatusReg || |
7226 | BuiltinID == AArch64::BI_WriteStatusReg) { |
7227 | LLVMContext &Context = CGM.getLLVMContext(); |
7228 | |
7229 | unsigned SysReg = |
7230 | E->getArg(0)->EvaluateKnownConstInt(getContext()).getZExtValue(); |
7231 | |
7232 | std::string SysRegStr; |
7233 | llvm::raw_string_ostream(SysRegStr) << |
7234 | ((1 << 1) | ((SysReg >> 14) & 1)) << ":" << |
7235 | ((SysReg >> 11) & 7) << ":" << |
7236 | ((SysReg >> 7) & 15) << ":" << |
7237 | ((SysReg >> 3) & 15) << ":" << |
7238 | ( SysReg & 7); |
7239 | |
7240 | llvm::Metadata *Ops[] = { llvm::MDString::get(Context, SysRegStr) }; |
7241 | llvm::MDNode *RegName = llvm::MDNode::get(Context, Ops); |
7242 | llvm::Value *Metadata = llvm::MetadataAsValue::get(Context, RegName); |
7243 | |
7244 | llvm::Type *RegisterType = Int64Ty; |
7245 | llvm::Type *Types[] = { RegisterType }; |
7246 | |
7247 | if (BuiltinID == AArch64::BI_ReadStatusReg) { |
7248 | llvm::Function *F = CGM.getIntrinsic(llvm::Intrinsic::read_register, Types); |
7249 | |
7250 | return Builder.CreateCall(F, Metadata); |
7251 | } |
7252 | |
7253 | llvm::Function *F = CGM.getIntrinsic(llvm::Intrinsic::write_register, Types); |
7254 | llvm::Value *ArgValue = EmitScalarExpr(E->getArg(1)); |
7255 | |
7256 | return Builder.CreateCall(F, { Metadata, ArgValue }); |
7257 | } |
7258 | |
7259 | if (BuiltinID == AArch64::BI_AddressOfReturnAddress) { |
7260 | llvm::Function *F = CGM.getIntrinsic(Intrinsic::addressofreturnaddress); |
7261 | return Builder.CreateCall(F); |
7262 | } |
7263 | |
7264 | if (BuiltinID == AArch64::BI__builtin_sponentry) { |
7265 | llvm::Function *F = CGM.getIntrinsic(Intrinsic::sponentry); |
7266 | return Builder.CreateCall(F); |
7267 | } |
7268 | |
7269 | // Find out if any arguments are required to be integer constant |
7270 | // expressions. |
7271 | unsigned ICEArguments = 0; |
7272 | ASTContext::GetBuiltinTypeError Error; |
7273 | getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); |
7274 | assert(Error == ASTContext::GE_None && "Should not codegen an error")((Error == ASTContext::GE_None && "Should not codegen an error" ) ? static_cast<void> (0) : __assert_fail ("Error == ASTContext::GE_None && \"Should not codegen an error\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 7274, __PRETTY_FUNCTION__)); |
7275 | |
7276 | llvm::SmallVector<Value*, 4> Ops; |
7277 | for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) { |
7278 | if ((ICEArguments & (1 << i)) == 0) { |
7279 | Ops.push_back(EmitScalarExpr(E->getArg(i))); |
7280 | } else { |
7281 | // If this is required to be a constant, constant fold it so that we know |
7282 | // that the generated intrinsic gets a ConstantInt. |
7283 | llvm::APSInt Result; |
7284 | bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); |
7285 | assert(IsConst && "Constant arg isn't actually constant?")((IsConst && "Constant arg isn't actually constant?") ? static_cast<void> (0) : __assert_fail ("IsConst && \"Constant arg isn't actually constant?\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 7285, __PRETTY_FUNCTION__)); |
7286 | (void)IsConst; |
7287 | Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); |
7288 | } |
7289 | } |
7290 | |
7291 | auto SISDMap = makeArrayRef(AArch64SISDIntrinsicMap); |
7292 | const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( |
7293 | SISDMap, BuiltinID, AArch64SISDIntrinsicsProvenSorted); |
7294 | |
7295 | if (Builtin) { |
7296 | Ops.push_back(EmitScalarExpr(E->getArg(E->getNumArgs() - 1))); |
7297 | Value *Result = EmitCommonNeonSISDBuiltinExpr(*this, *Builtin, Ops, E); |
7298 | assert(Result && "SISD intrinsic should have been handled")((Result && "SISD intrinsic should have been handled" ) ? static_cast<void> (0) : __assert_fail ("Result && \"SISD intrinsic should have been handled\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 7298, __PRETTY_FUNCTION__)); |
7299 | return Result; |
7300 | } |
7301 | |
7302 | llvm::APSInt Result; |
7303 | const Expr *Arg = E->getArg(E->getNumArgs()-1); |
7304 | NeonTypeFlags Type(0); |
7305 | if (Arg->isIntegerConstantExpr(Result, getContext())) |
7306 | // Determine the type of this overloaded NEON intrinsic. |
7307 | Type = NeonTypeFlags(Result.getZExtValue()); |
7308 | |
7309 | bool usgn = Type.isUnsigned(); |
7310 | bool quad = Type.isQuad(); |
7311 | |
7312 | // Handle non-overloaded intrinsics first. |
7313 | switch (BuiltinID) { |
7314 | default: break; |
7315 | case NEON::BI__builtin_neon_vabsh_f16: |
7316 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
7317 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::fabs, HalfTy), Ops, "vabs"); |
7318 | case NEON::BI__builtin_neon_vldrq_p128: { |
7319 | llvm::Type *Int128Ty = llvm::Type::getIntNTy(getLLVMContext(), 128); |
7320 | llvm::Type *Int128PTy = llvm::PointerType::get(Int128Ty, 0); |
7321 | Value *Ptr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int128PTy); |
7322 | return Builder.CreateAlignedLoad(Int128Ty, Ptr, |
7323 | CharUnits::fromQuantity(16)); |
7324 | } |
7325 | case NEON::BI__builtin_neon_vstrq_p128: { |
7326 | llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); |
7327 | Value *Ptr = Builder.CreateBitCast(Ops[0], Int128PTy); |
7328 | return Builder.CreateDefaultAlignedStore(EmitScalarExpr(E->getArg(1)), Ptr); |
7329 | } |
7330 | case NEON::BI__builtin_neon_vcvts_u32_f32: |
7331 | case NEON::BI__builtin_neon_vcvtd_u64_f64: |
7332 | usgn = true; |
7333 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
7334 | case NEON::BI__builtin_neon_vcvts_s32_f32: |
7335 | case NEON::BI__builtin_neon_vcvtd_s64_f64: { |
7336 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
7337 | bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; |
7338 | llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; |
7339 | llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; |
7340 | Ops[0] = Builder.CreateBitCast(Ops[0], FTy); |
7341 | if (usgn) |
7342 | return Builder.CreateFPToUI(Ops[0], InTy); |
7343 | return Builder.CreateFPToSI(Ops[0], InTy); |
7344 | } |
7345 | case NEON::BI__builtin_neon_vcvts_f32_u32: |
7346 | case NEON::BI__builtin_neon_vcvtd_f64_u64: |
7347 | usgn = true; |
7348 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
7349 | case NEON::BI__builtin_neon_vcvts_f32_s32: |
7350 | case NEON::BI__builtin_neon_vcvtd_f64_s64: { |
7351 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
7352 | bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; |
7353 | llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; |
7354 | llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; |
7355 | Ops[0] = Builder.CreateBitCast(Ops[0], InTy); |
7356 | if (usgn) |
7357 | return Builder.CreateUIToFP(Ops[0], FTy); |
7358 | return Builder.CreateSIToFP(Ops[0], FTy); |
7359 | } |
7360 | case NEON::BI__builtin_neon_vcvth_f16_u16: |
7361 | case NEON::BI__builtin_neon_vcvth_f16_u32: |
7362 | case NEON::BI__builtin_neon_vcvth_f16_u64: |
7363 | usgn = true; |
7364 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
7365 | case NEON::BI__builtin_neon_vcvth_f16_s16: |
7366 | case NEON::BI__builtin_neon_vcvth_f16_s32: |
7367 | case NEON::BI__builtin_neon_vcvth_f16_s64: { |
7368 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
7369 | llvm::Type *FTy = HalfTy; |
7370 | llvm::Type *InTy; |
7371 | if (Ops[0]->getType()->getPrimitiveSizeInBits() == 64) |
7372 | InTy = Int64Ty; |
7373 | else if (Ops[0]->getType()->getPrimitiveSizeInBits() == 32) |
7374 | InTy = Int32Ty; |
7375 | else |
7376 | InTy = Int16Ty; |
7377 | Ops[0] = Builder.CreateBitCast(Ops[0], InTy); |
7378 | if (usgn) |
7379 | return Builder.CreateUIToFP(Ops[0], FTy); |
7380 | return Builder.CreateSIToFP(Ops[0], FTy); |
7381 | } |
7382 | case NEON::BI__builtin_neon_vcvth_u16_f16: |
7383 | usgn = true; |
7384 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
7385 | case NEON::BI__builtin_neon_vcvth_s16_f16: { |
7386 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
7387 | Ops[0] = Builder.CreateBitCast(Ops[0], HalfTy); |
7388 | if (usgn) |
7389 | return Builder.CreateFPToUI(Ops[0], Int16Ty); |
7390 | return Builder.CreateFPToSI(Ops[0], Int16Ty); |
7391 | } |
7392 | case NEON::BI__builtin_neon_vcvth_u32_f16: |
7393 | usgn = true; |
7394 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
7395 | case NEON::BI__builtin_neon_vcvth_s32_f16: { |
7396 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
7397 | Ops[0] = Builder.CreateBitCast(Ops[0], HalfTy); |
7398 | if (usgn) |
7399 | return Builder.CreateFPToUI(Ops[0], Int32Ty); |
7400 | return Builder.CreateFPToSI(Ops[0], Int32Ty); |
7401 | } |
7402 | case NEON::BI__builtin_neon_vcvth_u64_f16: |
7403 | usgn = true; |
7404 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
7405 | case NEON::BI__builtin_neon_vcvth_s64_f16: { |
7406 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
7407 | Ops[0] = Builder.CreateBitCast(Ops[0], HalfTy); |
7408 | if (usgn) |
7409 | return Builder.CreateFPToUI(Ops[0], Int64Ty); |
7410 | return Builder.CreateFPToSI(Ops[0], Int64Ty); |
7411 | } |
7412 | case NEON::BI__builtin_neon_vcvtah_u16_f16: |
7413 | case NEON::BI__builtin_neon_vcvtmh_u16_f16: |
7414 | case NEON::BI__builtin_neon_vcvtnh_u16_f16: |
7415 | case NEON::BI__builtin_neon_vcvtph_u16_f16: |
7416 | case NEON::BI__builtin_neon_vcvtah_s16_f16: |
7417 | case NEON::BI__builtin_neon_vcvtmh_s16_f16: |
7418 | case NEON::BI__builtin_neon_vcvtnh_s16_f16: |
7419 | case NEON::BI__builtin_neon_vcvtph_s16_f16: { |
7420 | unsigned Int; |
7421 | llvm::Type* InTy = Int32Ty; |
7422 | llvm::Type* FTy = HalfTy; |
7423 | llvm::Type *Tys[2] = {InTy, FTy}; |
7424 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
7425 | switch (BuiltinID) { |
7426 | default: llvm_unreachable("missing builtin ID in switch!")::llvm::llvm_unreachable_internal("missing builtin ID in switch!" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 7426); |
7427 | case NEON::BI__builtin_neon_vcvtah_u16_f16: |
7428 | Int = Intrinsic::aarch64_neon_fcvtau; break; |
7429 | case NEON::BI__builtin_neon_vcvtmh_u16_f16: |
7430 | Int = Intrinsic::aarch64_neon_fcvtmu; break; |
7431 | case NEON::BI__builtin_neon_vcvtnh_u16_f16: |
7432 | Int = Intrinsic::aarch64_neon_fcvtnu; break; |
7433 | case NEON::BI__builtin_neon_vcvtph_u16_f16: |
7434 | Int = Intrinsic::aarch64_neon_fcvtpu; break; |
7435 | case NEON::BI__builtin_neon_vcvtah_s16_f16: |
7436 | Int = Intrinsic::aarch64_neon_fcvtas; break; |
7437 | case NEON::BI__builtin_neon_vcvtmh_s16_f16: |
7438 | Int = Intrinsic::aarch64_neon_fcvtms; break; |
7439 | case NEON::BI__builtin_neon_vcvtnh_s16_f16: |
7440 | Int = Intrinsic::aarch64_neon_fcvtns; break; |
7441 | case NEON::BI__builtin_neon_vcvtph_s16_f16: |
7442 | Int = Intrinsic::aarch64_neon_fcvtps; break; |
7443 | } |
7444 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "fcvt"); |
7445 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
7446 | } |
7447 | case NEON::BI__builtin_neon_vcaleh_f16: |
7448 | case NEON::BI__builtin_neon_vcalth_f16: |
7449 | case NEON::BI__builtin_neon_vcageh_f16: |
7450 | case NEON::BI__builtin_neon_vcagth_f16: { |
7451 | unsigned Int; |
7452 | llvm::Type* InTy = Int32Ty; |
7453 | llvm::Type* FTy = HalfTy; |
7454 | llvm::Type *Tys[2] = {InTy, FTy}; |
7455 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
7456 | switch (BuiltinID) { |
7457 | default: llvm_unreachable("missing builtin ID in switch!")::llvm::llvm_unreachable_internal("missing builtin ID in switch!" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 7457); |
7458 | case NEON::BI__builtin_neon_vcageh_f16: |
7459 | Int = Intrinsic::aarch64_neon_facge; break; |
7460 | case NEON::BI__builtin_neon_vcagth_f16: |
7461 | Int = Intrinsic::aarch64_neon_facgt; break; |
7462 | case NEON::BI__builtin_neon_vcaleh_f16: |
7463 | Int = Intrinsic::aarch64_neon_facge; std::swap(Ops[0], Ops[1]); break; |
7464 | case NEON::BI__builtin_neon_vcalth_f16: |
7465 | Int = Intrinsic::aarch64_neon_facgt; std::swap(Ops[0], Ops[1]); break; |
7466 | } |
7467 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "facg"); |
7468 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
7469 | } |
7470 | case NEON::BI__builtin_neon_vcvth_n_s16_f16: |
7471 | case NEON::BI__builtin_neon_vcvth_n_u16_f16: { |
7472 | unsigned Int; |
7473 | llvm::Type* InTy = Int32Ty; |
7474 | llvm::Type* FTy = HalfTy; |
7475 | llvm::Type *Tys[2] = {InTy, FTy}; |
7476 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
7477 | switch (BuiltinID) { |
7478 | default: llvm_unreachable("missing builtin ID in switch!")::llvm::llvm_unreachable_internal("missing builtin ID in switch!" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 7478); |
7479 | case NEON::BI__builtin_neon_vcvth_n_s16_f16: |
7480 | Int = Intrinsic::aarch64_neon_vcvtfp2fxs; break; |
7481 | case NEON::BI__builtin_neon_vcvth_n_u16_f16: |
7482 | Int = Intrinsic::aarch64_neon_vcvtfp2fxu; break; |
7483 | } |
7484 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "fcvth_n"); |
7485 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
7486 | } |
7487 | case NEON::BI__builtin_neon_vcvth_n_f16_s16: |
7488 | case NEON::BI__builtin_neon_vcvth_n_f16_u16: { |
7489 | unsigned Int; |
7490 | llvm::Type* FTy = HalfTy; |
7491 | llvm::Type* InTy = Int32Ty; |
7492 | llvm::Type *Tys[2] = {FTy, InTy}; |
7493 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
7494 | switch (BuiltinID) { |
7495 | default: llvm_unreachable("missing builtin ID in switch!")::llvm::llvm_unreachable_internal("missing builtin ID in switch!" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 7495); |
7496 | case NEON::BI__builtin_neon_vcvth_n_f16_s16: |
7497 | Int = Intrinsic::aarch64_neon_vcvtfxs2fp; |
7498 | Ops[0] = Builder.CreateSExt(Ops[0], InTy, "sext"); |
7499 | break; |
7500 | case NEON::BI__builtin_neon_vcvth_n_f16_u16: |
7501 | Int = Intrinsic::aarch64_neon_vcvtfxu2fp; |
7502 | Ops[0] = Builder.CreateZExt(Ops[0], InTy); |
7503 | break; |
7504 | } |
7505 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "fcvth_n"); |
7506 | } |
7507 | case NEON::BI__builtin_neon_vpaddd_s64: { |
7508 | llvm::Type *Ty = llvm::VectorType::get(Int64Ty, 2); |
7509 | Value *Vec = EmitScalarExpr(E->getArg(0)); |
7510 | // The vector is v2f64, so make sure it's bitcast to that. |
7511 | Vec = Builder.CreateBitCast(Vec, Ty, "v2i64"); |
7512 | llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); |
7513 | llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); |
7514 | Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); |
7515 | Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); |
7516 | // Pairwise addition of a v2f64 into a scalar f64. |
7517 | return Builder.CreateAdd(Op0, Op1, "vpaddd"); |
7518 | } |
7519 | case NEON::BI__builtin_neon_vpaddd_f64: { |
7520 | llvm::Type *Ty = |
7521 | llvm::VectorType::get(DoubleTy, 2); |
7522 | Value *Vec = EmitScalarExpr(E->getArg(0)); |
7523 | // The vector is v2f64, so make sure it's bitcast to that. |
7524 | Vec = Builder.CreateBitCast(Vec, Ty, "v2f64"); |
7525 | llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); |
7526 | llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); |
7527 | Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); |
7528 | Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); |
7529 | // Pairwise addition of a v2f64 into a scalar f64. |
7530 | return Builder.CreateFAdd(Op0, Op1, "vpaddd"); |
7531 | } |
7532 | case NEON::BI__builtin_neon_vpadds_f32: { |
7533 | llvm::Type *Ty = |
7534 | llvm::VectorType::get(FloatTy, 2); |
7535 | Value *Vec = EmitScalarExpr(E->getArg(0)); |
7536 | // The vector is v2f32, so make sure it's bitcast to that. |
7537 | Vec = Builder.CreateBitCast(Vec, Ty, "v2f32"); |
7538 | llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); |
7539 | llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); |
7540 | Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); |
7541 | Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); |
7542 | // Pairwise addition of a v2f32 into a scalar f32. |
7543 | return Builder.CreateFAdd(Op0, Op1, "vpaddd"); |
7544 | } |
7545 | case NEON::BI__builtin_neon_vceqzd_s64: |
7546 | case NEON::BI__builtin_neon_vceqzd_f64: |
7547 | case NEON::BI__builtin_neon_vceqzs_f32: |
7548 | case NEON::BI__builtin_neon_vceqzh_f16: |
7549 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
7550 | return EmitAArch64CompareBuiltinExpr( |
7551 | Ops[0], ConvertType(E->getCallReturnType(getContext())), |
7552 | ICmpInst::FCMP_OEQ, ICmpInst::ICMP_EQ, "vceqz"); |
7553 | case NEON::BI__builtin_neon_vcgezd_s64: |
7554 | case NEON::BI__builtin_neon_vcgezd_f64: |
7555 | case NEON::BI__builtin_neon_vcgezs_f32: |
7556 | case NEON::BI__builtin_neon_vcgezh_f16: |
7557 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
7558 | return EmitAArch64CompareBuiltinExpr( |
7559 | Ops[0], ConvertType(E->getCallReturnType(getContext())), |
7560 | ICmpInst::FCMP_OGE, ICmpInst::ICMP_SGE, "vcgez"); |
7561 | case NEON::BI__builtin_neon_vclezd_s64: |
7562 | case NEON::BI__builtin_neon_vclezd_f64: |
7563 | case NEON::BI__builtin_neon_vclezs_f32: |
7564 | case NEON::BI__builtin_neon_vclezh_f16: |
7565 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
7566 | return EmitAArch64CompareBuiltinExpr( |
7567 | Ops[0], ConvertType(E->getCallReturnType(getContext())), |
7568 | ICmpInst::FCMP_OLE, ICmpInst::ICMP_SLE, "vclez"); |
7569 | case NEON::BI__builtin_neon_vcgtzd_s64: |
7570 | case NEON::BI__builtin_neon_vcgtzd_f64: |
7571 | case NEON::BI__builtin_neon_vcgtzs_f32: |
7572 | case NEON::BI__builtin_neon_vcgtzh_f16: |
7573 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
7574 | return EmitAArch64CompareBuiltinExpr( |
7575 | Ops[0], ConvertType(E->getCallReturnType(getContext())), |
7576 | ICmpInst::FCMP_OGT, ICmpInst::ICMP_SGT, "vcgtz"); |
7577 | case NEON::BI__builtin_neon_vcltzd_s64: |
7578 | case NEON::BI__builtin_neon_vcltzd_f64: |
7579 | case NEON::BI__builtin_neon_vcltzs_f32: |
7580 | case NEON::BI__builtin_neon_vcltzh_f16: |
7581 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
7582 | return EmitAArch64CompareBuiltinExpr( |
7583 | Ops[0], ConvertType(E->getCallReturnType(getContext())), |
7584 | ICmpInst::FCMP_OLT, ICmpInst::ICMP_SLT, "vcltz"); |
7585 | |
7586 | case NEON::BI__builtin_neon_vceqzd_u64: { |
7587 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
7588 | Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); |
7589 | Ops[0] = |
7590 | Builder.CreateICmpEQ(Ops[0], llvm::Constant::getNullValue(Int64Ty)); |
7591 | return Builder.CreateSExt(Ops[0], Int64Ty, "vceqzd"); |
7592 | } |
7593 | case NEON::BI__builtin_neon_vceqd_f64: |
7594 | case NEON::BI__builtin_neon_vcled_f64: |
7595 | case NEON::BI__builtin_neon_vcltd_f64: |
7596 | case NEON::BI__builtin_neon_vcged_f64: |
7597 | case NEON::BI__builtin_neon_vcgtd_f64: { |
7598 | llvm::CmpInst::Predicate P; |
7599 | switch (BuiltinID) { |
7600 | default: llvm_unreachable("missing builtin ID in switch!")::llvm::llvm_unreachable_internal("missing builtin ID in switch!" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 7600); |
7601 | case NEON::BI__builtin_neon_vceqd_f64: P = llvm::FCmpInst::FCMP_OEQ; break; |
7602 | case NEON::BI__builtin_neon_vcled_f64: P = llvm::FCmpInst::FCMP_OLE; break; |
7603 | case NEON::BI__builtin_neon_vcltd_f64: P = llvm::FCmpInst::FCMP_OLT; break; |
7604 | case NEON::BI__builtin_neon_vcged_f64: P = llvm::FCmpInst::FCMP_OGE; break; |
7605 | case NEON::BI__builtin_neon_vcgtd_f64: P = llvm::FCmpInst::FCMP_OGT; break; |
7606 | } |
7607 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
7608 | Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); |
7609 | Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); |
7610 | Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); |
7611 | return Builder.CreateSExt(Ops[0], Int64Ty, "vcmpd"); |
7612 | } |
7613 | case NEON::BI__builtin_neon_vceqs_f32: |
7614 | case NEON::BI__builtin_neon_vcles_f32: |
7615 | case NEON::BI__builtin_neon_vclts_f32: |
7616 | case NEON::BI__builtin_neon_vcges_f32: |
7617 | case NEON::BI__builtin_neon_vcgts_f32: { |
7618 | llvm::CmpInst::Predicate P; |
7619 | switch (BuiltinID) { |
7620 | default: llvm_unreachable("missing builtin ID in switch!")::llvm::llvm_unreachable_internal("missing builtin ID in switch!" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 7620); |
7621 | case NEON::BI__builtin_neon_vceqs_f32: P = llvm::FCmpInst::FCMP_OEQ; break; |
7622 | case NEON::BI__builtin_neon_vcles_f32: P = llvm::FCmpInst::FCMP_OLE; break; |
7623 | case NEON::BI__builtin_neon_vclts_f32: P = llvm::FCmpInst::FCMP_OLT; break; |
7624 | case NEON::BI__builtin_neon_vcges_f32: P = llvm::FCmpInst::FCMP_OGE; break; |
7625 | case NEON::BI__builtin_neon_vcgts_f32: P = llvm::FCmpInst::FCMP_OGT; break; |
7626 | } |
7627 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
7628 | Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); |
7629 | Ops[1] = Builder.CreateBitCast(Ops[1], FloatTy); |
7630 | Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); |
7631 | return Builder.CreateSExt(Ops[0], Int32Ty, "vcmpd"); |
7632 | } |
7633 | case NEON::BI__builtin_neon_vceqh_f16: |
7634 | case NEON::BI__builtin_neon_vcleh_f16: |
7635 | case NEON::BI__builtin_neon_vclth_f16: |
7636 | case NEON::BI__builtin_neon_vcgeh_f16: |
7637 | case NEON::BI__builtin_neon_vcgth_f16: { |
7638 | llvm::CmpInst::Predicate P; |
7639 | switch (BuiltinID) { |
7640 | default: llvm_unreachable("missing builtin ID in switch!")::llvm::llvm_unreachable_internal("missing builtin ID in switch!" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 7640); |
7641 | case NEON::BI__builtin_neon_vceqh_f16: P = llvm::FCmpInst::FCMP_OEQ; break; |
7642 | case NEON::BI__builtin_neon_vcleh_f16: P = llvm::FCmpInst::FCMP_OLE; break; |
7643 | case NEON::BI__builtin_neon_vclth_f16: P = llvm::FCmpInst::FCMP_OLT; break; |
7644 | case NEON::BI__builtin_neon_vcgeh_f16: P = llvm::FCmpInst::FCMP_OGE; break; |
7645 | case NEON::BI__builtin_neon_vcgth_f16: P = llvm::FCmpInst::FCMP_OGT; break; |
7646 | } |
7647 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
7648 | Ops[0] = Builder.CreateBitCast(Ops[0], HalfTy); |
7649 | Ops[1] = Builder.CreateBitCast(Ops[1], HalfTy); |
7650 | Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); |
7651 | return Builder.CreateSExt(Ops[0], Int16Ty, "vcmpd"); |
7652 | } |
7653 | case NEON::BI__builtin_neon_vceqd_s64: |
7654 | case NEON::BI__builtin_neon_vceqd_u64: |
7655 | case NEON::BI__builtin_neon_vcgtd_s64: |
7656 | case NEON::BI__builtin_neon_vcgtd_u64: |
7657 | case NEON::BI__builtin_neon_vcltd_s64: |
7658 | case NEON::BI__builtin_neon_vcltd_u64: |
7659 | case NEON::BI__builtin_neon_vcged_u64: |
7660 | case NEON::BI__builtin_neon_vcged_s64: |
7661 | case NEON::BI__builtin_neon_vcled_u64: |
7662 | case NEON::BI__builtin_neon_vcled_s64: { |
7663 | llvm::CmpInst::Predicate P; |
7664 | switch (BuiltinID) { |
7665 | default: llvm_unreachable("missing builtin ID in switch!")::llvm::llvm_unreachable_internal("missing builtin ID in switch!" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 7665); |
7666 | case NEON::BI__builtin_neon_vceqd_s64: |
7667 | case NEON::BI__builtin_neon_vceqd_u64:P = llvm::ICmpInst::ICMP_EQ;break; |
7668 | case NEON::BI__builtin_neon_vcgtd_s64:P = llvm::ICmpInst::ICMP_SGT;break; |
7669 | case NEON::BI__builtin_neon_vcgtd_u64:P = llvm::ICmpInst::ICMP_UGT;break; |
7670 | case NEON::BI__builtin_neon_vcltd_s64:P = llvm::ICmpInst::ICMP_SLT;break; |
7671 | case NEON::BI__builtin_neon_vcltd_u64:P = llvm::ICmpInst::ICMP_ULT;break; |
7672 | case NEON::BI__builtin_neon_vcged_u64:P = llvm::ICmpInst::ICMP_UGE;break; |
7673 | case NEON::BI__builtin_neon_vcged_s64:P = llvm::ICmpInst::ICMP_SGE;break; |
7674 | case NEON::BI__builtin_neon_vcled_u64:P = llvm::ICmpInst::ICMP_ULE;break; |
7675 | case NEON::BI__builtin_neon_vcled_s64:P = llvm::ICmpInst::ICMP_SLE;break; |
7676 | } |
7677 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
7678 | Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); |
7679 | Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); |
7680 | Ops[0] = Builder.CreateICmp(P, Ops[0], Ops[1]); |
7681 | return Builder.CreateSExt(Ops[0], Int64Ty, "vceqd"); |
7682 | } |
7683 | case NEON::BI__builtin_neon_vtstd_s64: |
7684 | case NEON::BI__builtin_neon_vtstd_u64: { |
7685 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
7686 | Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); |
7687 | Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); |
7688 | Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); |
7689 | Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], |
7690 | llvm::Constant::getNullValue(Int64Ty)); |
7691 | return Builder.CreateSExt(Ops[0], Int64Ty, "vtstd"); |
7692 | } |
7693 | case NEON::BI__builtin_neon_vset_lane_i8: |
7694 | case NEON::BI__builtin_neon_vset_lane_i16: |
7695 | case NEON::BI__builtin_neon_vset_lane_i32: |
7696 | case NEON::BI__builtin_neon_vset_lane_i64: |
7697 | case NEON::BI__builtin_neon_vset_lane_f32: |
7698 | case NEON::BI__builtin_neon_vsetq_lane_i8: |
7699 | case NEON::BI__builtin_neon_vsetq_lane_i16: |
7700 | case NEON::BI__builtin_neon_vsetq_lane_i32: |
7701 | case NEON::BI__builtin_neon_vsetq_lane_i64: |
7702 | case NEON::BI__builtin_neon_vsetq_lane_f32: |
7703 | Ops.push_back(EmitScalarExpr(E->getArg(2))); |
7704 | return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); |
7705 | case NEON::BI__builtin_neon_vset_lane_f64: |
7706 | // The vector type needs a cast for the v1f64 variant. |
7707 | Ops[1] = Builder.CreateBitCast(Ops[1], |
7708 | llvm::VectorType::get(DoubleTy, 1)); |
7709 | Ops.push_back(EmitScalarExpr(E->getArg(2))); |
7710 | return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); |
7711 | case NEON::BI__builtin_neon_vsetq_lane_f64: |
7712 | // The vector type needs a cast for the v2f64 variant. |
7713 | Ops[1] = Builder.CreateBitCast(Ops[1], |
7714 | llvm::VectorType::get(DoubleTy, 2)); |
7715 | Ops.push_back(EmitScalarExpr(E->getArg(2))); |
7716 | return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); |
7717 | |
7718 | case NEON::BI__builtin_neon_vget_lane_i8: |
7719 | case NEON::BI__builtin_neon_vdupb_lane_i8: |
7720 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int8Ty, 8)); |
7721 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
7722 | "vget_lane"); |
7723 | case NEON::BI__builtin_neon_vgetq_lane_i8: |
7724 | case NEON::BI__builtin_neon_vdupb_laneq_i8: |
7725 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int8Ty, 16)); |
7726 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
7727 | "vgetq_lane"); |
7728 | case NEON::BI__builtin_neon_vget_lane_i16: |
7729 | case NEON::BI__builtin_neon_vduph_lane_i16: |
7730 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int16Ty, 4)); |
7731 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
7732 | "vget_lane"); |
7733 | case NEON::BI__builtin_neon_vgetq_lane_i16: |
7734 | case NEON::BI__builtin_neon_vduph_laneq_i16: |
7735 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int16Ty, 8)); |
7736 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
7737 | "vgetq_lane"); |
7738 | case NEON::BI__builtin_neon_vget_lane_i32: |
7739 | case NEON::BI__builtin_neon_vdups_lane_i32: |
7740 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int32Ty, 2)); |
7741 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
7742 | "vget_lane"); |
7743 | case NEON::BI__builtin_neon_vdups_lane_f32: |
7744 | Ops[0] = Builder.CreateBitCast(Ops[0], |
7745 | llvm::VectorType::get(FloatTy, 2)); |
7746 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
7747 | "vdups_lane"); |
7748 | case NEON::BI__builtin_neon_vgetq_lane_i32: |
7749 | case NEON::BI__builtin_neon_vdups_laneq_i32: |
7750 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int32Ty, 4)); |
7751 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
7752 | "vgetq_lane"); |
7753 | case NEON::BI__builtin_neon_vget_lane_i64: |
7754 | case NEON::BI__builtin_neon_vdupd_lane_i64: |
7755 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int64Ty, 1)); |
7756 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
7757 | "vget_lane"); |
7758 | case NEON::BI__builtin_neon_vdupd_lane_f64: |
7759 | Ops[0] = Builder.CreateBitCast(Ops[0], |
7760 | llvm::VectorType::get(DoubleTy, 1)); |
7761 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
7762 | "vdupd_lane"); |
7763 | case NEON::BI__builtin_neon_vgetq_lane_i64: |
7764 | case NEON::BI__builtin_neon_vdupd_laneq_i64: |
7765 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int64Ty, 2)); |
7766 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
7767 | "vgetq_lane"); |
7768 | case NEON::BI__builtin_neon_vget_lane_f32: |
7769 | Ops[0] = Builder.CreateBitCast(Ops[0], |
7770 | llvm::VectorType::get(FloatTy, 2)); |
7771 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
7772 | "vget_lane"); |
7773 | case NEON::BI__builtin_neon_vget_lane_f64: |
7774 | Ops[0] = Builder.CreateBitCast(Ops[0], |
7775 | llvm::VectorType::get(DoubleTy, 1)); |
7776 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
7777 | "vget_lane"); |
7778 | case NEON::BI__builtin_neon_vgetq_lane_f32: |
7779 | case NEON::BI__builtin_neon_vdups_laneq_f32: |
7780 | Ops[0] = Builder.CreateBitCast(Ops[0], |
7781 | llvm::VectorType::get(FloatTy, 4)); |
7782 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
7783 | "vgetq_lane"); |
7784 | case NEON::BI__builtin_neon_vgetq_lane_f64: |
7785 | case NEON::BI__builtin_neon_vdupd_laneq_f64: |
7786 | Ops[0] = Builder.CreateBitCast(Ops[0], |
7787 | llvm::VectorType::get(DoubleTy, 2)); |
7788 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
7789 | "vgetq_lane"); |
7790 | case NEON::BI__builtin_neon_vaddh_f16: |
7791 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
7792 | return Builder.CreateFAdd(Ops[0], Ops[1], "vaddh"); |
7793 | case NEON::BI__builtin_neon_vsubh_f16: |
7794 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
7795 | return Builder.CreateFSub(Ops[0], Ops[1], "vsubh"); |
7796 | case NEON::BI__builtin_neon_vmulh_f16: |
7797 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
7798 | return Builder.CreateFMul(Ops[0], Ops[1], "vmulh"); |
7799 | case NEON::BI__builtin_neon_vdivh_f16: |
7800 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
7801 | return Builder.CreateFDiv(Ops[0], Ops[1], "vdivh"); |
7802 | case NEON::BI__builtin_neon_vfmah_f16: { |
7803 | Function *F = CGM.getIntrinsic(Intrinsic::fma, HalfTy); |
7804 | // NEON intrinsic puts accumulator first, unlike the LLVM fma. |
7805 | return Builder.CreateCall(F, |
7806 | {EmitScalarExpr(E->getArg(1)), EmitScalarExpr(E->getArg(2)), Ops[0]}); |
7807 | } |
7808 | case NEON::BI__builtin_neon_vfmsh_f16: { |
7809 | Function *F = CGM.getIntrinsic(Intrinsic::fma, HalfTy); |
7810 | Value *Zero = llvm::ConstantFP::getZeroValueForNegation(HalfTy); |
7811 | Value* Sub = Builder.CreateFSub(Zero, EmitScalarExpr(E->getArg(1)), "vsubh"); |
7812 | // NEON intrinsic puts accumulator first, unlike the LLVM fma. |
7813 | return Builder.CreateCall(F, {Sub, EmitScalarExpr(E->getArg(2)), Ops[0]}); |
7814 | } |
7815 | case NEON::BI__builtin_neon_vaddd_s64: |
7816 | case NEON::BI__builtin_neon_vaddd_u64: |
7817 | return Builder.CreateAdd(Ops[0], EmitScalarExpr(E->getArg(1)), "vaddd"); |
7818 | case NEON::BI__builtin_neon_vsubd_s64: |
7819 | case NEON::BI__builtin_neon_vsubd_u64: |
7820 | return Builder.CreateSub(Ops[0], EmitScalarExpr(E->getArg(1)), "vsubd"); |
7821 | case NEON::BI__builtin_neon_vqdmlalh_s16: |
7822 | case NEON::BI__builtin_neon_vqdmlslh_s16: { |
7823 | SmallVector<Value *, 2> ProductOps; |
7824 | ProductOps.push_back(vectorWrapScalar16(Ops[1])); |
7825 | ProductOps.push_back(vectorWrapScalar16(EmitScalarExpr(E->getArg(2)))); |
7826 | llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); |
7827 | Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), |
7828 | ProductOps, "vqdmlXl"); |
7829 | Constant *CI = ConstantInt::get(SizeTy, 0); |
7830 | Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); |
7831 | |
7832 | unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlalh_s16 |
7833 | ? Intrinsic::aarch64_neon_sqadd |
7834 | : Intrinsic::aarch64_neon_sqsub; |
7835 | return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int32Ty), Ops, "vqdmlXl"); |
7836 | } |
7837 | case NEON::BI__builtin_neon_vqshlud_n_s64: { |
7838 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
7839 | Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); |
7840 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqshlu, Int64Ty), |
7841 | Ops, "vqshlu_n"); |
7842 | } |
7843 | case NEON::BI__builtin_neon_vqshld_n_u64: |
7844 | case NEON::BI__builtin_neon_vqshld_n_s64: { |
7845 | unsigned Int = BuiltinID == NEON::BI__builtin_neon_vqshld_n_u64 |
7846 | ? Intrinsic::aarch64_neon_uqshl |
7847 | : Intrinsic::aarch64_neon_sqshl; |
7848 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
7849 | Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); |
7850 | return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vqshl_n"); |
7851 | } |
7852 | case NEON::BI__builtin_neon_vrshrd_n_u64: |
7853 | case NEON::BI__builtin_neon_vrshrd_n_s64: { |
7854 | unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrshrd_n_u64 |
7855 | ? Intrinsic::aarch64_neon_urshl |
7856 | : Intrinsic::aarch64_neon_srshl; |
7857 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
7858 | int SV = cast<ConstantInt>(Ops[1])->getSExtValue(); |
7859 | Ops[1] = ConstantInt::get(Int64Ty, -SV); |
7860 | return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vrshr_n"); |
7861 | } |
7862 | case NEON::BI__builtin_neon_vrsrad_n_u64: |
7863 | case NEON::BI__builtin_neon_vrsrad_n_s64: { |
7864 | unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrsrad_n_u64 |
7865 | ? Intrinsic::aarch64_neon_urshl |
7866 | : Intrinsic::aarch64_neon_srshl; |
7867 | Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); |
7868 | Ops.push_back(Builder.CreateNeg(EmitScalarExpr(E->getArg(2)))); |
7869 | Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Int64Ty), |
7870 | {Ops[1], Builder.CreateSExt(Ops[2], Int64Ty)}); |
7871 | return Builder.CreateAdd(Ops[0], Builder.CreateBitCast(Ops[1], Int64Ty)); |
7872 | } |
7873 | case NEON::BI__builtin_neon_vshld_n_s64: |
7874 | case NEON::BI__builtin_neon_vshld_n_u64: { |
7875 | llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); |
7876 | return Builder.CreateShl( |
7877 | Ops[0], ConstantInt::get(Int64Ty, Amt->getZExtValue()), "shld_n"); |
7878 | } |
7879 | case NEON::BI__builtin_neon_vshrd_n_s64: { |
7880 | llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); |
7881 | return Builder.CreateAShr( |
7882 | Ops[0], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), |
7883 | Amt->getZExtValue())), |
7884 | "shrd_n"); |
7885 | } |
7886 | case NEON::BI__builtin_neon_vshrd_n_u64: { |
7887 | llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); |
7888 | uint64_t ShiftAmt = Amt->getZExtValue(); |
7889 | // Right-shifting an unsigned value by its size yields 0. |
7890 | if (ShiftAmt == 64) |
7891 | return ConstantInt::get(Int64Ty, 0); |
7892 | return Builder.CreateLShr(Ops[0], ConstantInt::get(Int64Ty, ShiftAmt), |
7893 | "shrd_n"); |
7894 | } |
7895 | case NEON::BI__builtin_neon_vsrad_n_s64: { |
7896 | llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); |
7897 | Ops[1] = Builder.CreateAShr( |
7898 | Ops[1], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), |
7899 | Amt->getZExtValue())), |
7900 | "shrd_n"); |
7901 | return Builder.CreateAdd(Ops[0], Ops[1]); |
7902 | } |
7903 | case NEON::BI__builtin_neon_vsrad_n_u64: { |
7904 | llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); |
7905 | uint64_t ShiftAmt = Amt->getZExtValue(); |
7906 | // Right-shifting an unsigned value by its size yields 0. |
7907 | // As Op + 0 = Op, return Ops[0] directly. |
7908 | if (ShiftAmt == 64) |
7909 | return Ops[0]; |
7910 | Ops[1] = Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, ShiftAmt), |
7911 | "shrd_n"); |
7912 | return Builder.CreateAdd(Ops[0], Ops[1]); |
7913 | } |
7914 | case NEON::BI__builtin_neon_vqdmlalh_lane_s16: |
7915 | case NEON::BI__builtin_neon_vqdmlalh_laneq_s16: |
7916 | case NEON::BI__builtin_neon_vqdmlslh_lane_s16: |
7917 | case NEON::BI__builtin_neon_vqdmlslh_laneq_s16: { |
7918 | Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), |
7919 | "lane"); |
7920 | SmallVector<Value *, 2> ProductOps; |
7921 | ProductOps.push_back(vectorWrapScalar16(Ops[1])); |
7922 | ProductOps.push_back(vectorWrapScalar16(Ops[2])); |
7923 | llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); |
7924 | Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), |
7925 | ProductOps, "vqdmlXl"); |
7926 | Constant *CI = ConstantInt::get(SizeTy, 0); |
7927 | Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); |
7928 | Ops.pop_back(); |
7929 | |
7930 | unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlalh_lane_s16 || |
7931 | BuiltinID == NEON::BI__builtin_neon_vqdmlalh_laneq_s16) |
7932 | ? Intrinsic::aarch64_neon_sqadd |
7933 | : Intrinsic::aarch64_neon_sqsub; |
7934 | return EmitNeonCall(CGM.getIntrinsic(AccInt, Int32Ty), Ops, "vqdmlXl"); |
7935 | } |
7936 | case NEON::BI__builtin_neon_vqdmlals_s32: |
7937 | case NEON::BI__builtin_neon_vqdmlsls_s32: { |
7938 | SmallVector<Value *, 2> ProductOps; |
7939 | ProductOps.push_back(Ops[1]); |
7940 | ProductOps.push_back(EmitScalarExpr(E->getArg(2))); |
7941 | Ops[1] = |
7942 | EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), |
7943 | ProductOps, "vqdmlXl"); |
7944 | |
7945 | unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlals_s32 |
7946 | ? Intrinsic::aarch64_neon_sqadd |
7947 | : Intrinsic::aarch64_neon_sqsub; |
7948 | return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int64Ty), Ops, "vqdmlXl"); |
7949 | } |
7950 | case NEON::BI__builtin_neon_vqdmlals_lane_s32: |
7951 | case NEON::BI__builtin_neon_vqdmlals_laneq_s32: |
7952 | case NEON::BI__builtin_neon_vqdmlsls_lane_s32: |
7953 | case NEON::BI__builtin_neon_vqdmlsls_laneq_s32: { |
7954 | Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), |
7955 | "lane"); |
7956 | SmallVector<Value *, 2> ProductOps; |
7957 | ProductOps.push_back(Ops[1]); |
7958 | ProductOps.push_back(Ops[2]); |
7959 | Ops[1] = |
7960 | EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), |
7961 | ProductOps, "vqdmlXl"); |
7962 | Ops.pop_back(); |
7963 | |
7964 | unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlals_lane_s32 || |
7965 | BuiltinID == NEON::BI__builtin_neon_vqdmlals_laneq_s32) |
7966 | ? Intrinsic::aarch64_neon_sqadd |
7967 | : Intrinsic::aarch64_neon_sqsub; |
7968 | return EmitNeonCall(CGM.getIntrinsic(AccInt, Int64Ty), Ops, "vqdmlXl"); |
7969 | } |
7970 | case NEON::BI__builtin_neon_vduph_lane_f16: { |
7971 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
7972 | "vget_lane"); |
7973 | } |
7974 | case NEON::BI__builtin_neon_vduph_laneq_f16: { |
7975 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
7976 | "vgetq_lane"); |
7977 | } |
7978 | } |
7979 | |
7980 | llvm::VectorType *VTy = GetNeonType(this, Type); |
7981 | llvm::Type *Ty = VTy; |
7982 | if (!Ty) |
7983 | return nullptr; |
7984 | |
7985 | // Not all intrinsics handled by the common case work for AArch64 yet, so only |
7986 | // defer to common code if it's been added to our special map. |
7987 | Builtin = findNeonIntrinsicInMap(AArch64SIMDIntrinsicMap, BuiltinID, |
7988 | AArch64SIMDIntrinsicsProvenSorted); |
7989 | |
7990 | if (Builtin) |
7991 | return EmitCommonNeonBuiltinExpr( |
7992 | Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, |
7993 | Builtin->NameHint, Builtin->TypeModifier, E, Ops, |
7994 | /*never use addresses*/ Address::invalid(), Address::invalid(), Arch); |
7995 | |
7996 | if (Value *V = EmitAArch64TblBuiltinExpr(*this, BuiltinID, E, Ops, Arch)) |
7997 | return V; |
7998 | |
7999 | unsigned Int; |
8000 | switch (BuiltinID) { |
8001 | default: return nullptr; |
8002 | case NEON::BI__builtin_neon_vbsl_v: |
8003 | case NEON::BI__builtin_neon_vbslq_v: { |
8004 | llvm::Type *BitTy = llvm::VectorType::getInteger(VTy); |
8005 | Ops[0] = Builder.CreateBitCast(Ops[0], BitTy, "vbsl"); |
8006 | Ops[1] = Builder.CreateBitCast(Ops[1], BitTy, "vbsl"); |
8007 | Ops[2] = Builder.CreateBitCast(Ops[2], BitTy, "vbsl"); |
8008 | |
8009 | Ops[1] = Builder.CreateAnd(Ops[0], Ops[1], "vbsl"); |
8010 | Ops[2] = Builder.CreateAnd(Builder.CreateNot(Ops[0]), Ops[2], "vbsl"); |
8011 | Ops[0] = Builder.CreateOr(Ops[1], Ops[2], "vbsl"); |
8012 | return Builder.CreateBitCast(Ops[0], Ty); |
8013 | } |
8014 | case NEON::BI__builtin_neon_vfma_lane_v: |
8015 | case NEON::BI__builtin_neon_vfmaq_lane_v: { // Only used for FP types |
8016 | // The ARM builtins (and instructions) have the addend as the first |
8017 | // operand, but the 'fma' intrinsics have it last. Swap it around here. |
8018 | Value *Addend = Ops[0]; |
8019 | Value *Multiplicand = Ops[1]; |
8020 | Value *LaneSource = Ops[2]; |
8021 | Ops[0] = Multiplicand; |
8022 | Ops[1] = LaneSource; |
8023 | Ops[2] = Addend; |
8024 | |
8025 | // Now adjust things to handle the lane access. |
8026 | llvm::Type *SourceTy = BuiltinID == NEON::BI__builtin_neon_vfmaq_lane_v ? |
8027 | llvm::VectorType::get(VTy->getElementType(), VTy->getNumElements() / 2) : |
8028 | VTy; |
8029 | llvm::Constant *cst = cast<Constant>(Ops[3]); |
8030 | Value *SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), cst); |
8031 | Ops[1] = Builder.CreateBitCast(Ops[1], SourceTy); |
8032 | Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV, "lane"); |
8033 | |
8034 | Ops.pop_back(); |
8035 | Int = Intrinsic::fma; |
8036 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmla"); |
8037 | } |
8038 | case NEON::BI__builtin_neon_vfma_laneq_v: { |
8039 | llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); |
8040 | // v1f64 fma should be mapped to Neon scalar f64 fma |
8041 | if (VTy && VTy->getElementType() == DoubleTy) { |
8042 | Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); |
8043 | Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); |
8044 | llvm::Type *VTy = GetNeonType(this, |
8045 | NeonTypeFlags(NeonTypeFlags::Float64, false, true)); |
8046 | Ops[2] = Builder.CreateBitCast(Ops[2], VTy); |
8047 | Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); |
8048 | Function *F = CGM.getIntrinsic(Intrinsic::fma, DoubleTy); |
8049 | Value *Result = Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); |
8050 | return Builder.CreateBitCast(Result, Ty); |
8051 | } |
8052 | Function *F = CGM.getIntrinsic(Intrinsic::fma, Ty); |
8053 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
8054 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
8055 | |
8056 | llvm::Type *STy = llvm::VectorType::get(VTy->getElementType(), |
8057 | VTy->getNumElements() * 2); |
8058 | Ops[2] = Builder.CreateBitCast(Ops[2], STy); |
8059 | Value* SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), |
8060 | cast<ConstantInt>(Ops[3])); |
8061 | Ops[2] = Builder.CreateShuffleVector(Ops[2], Ops[2], SV, "lane"); |
8062 | |
8063 | return Builder.CreateCall(F, {Ops[2], Ops[1], Ops[0]}); |
8064 | } |
8065 | case NEON::BI__builtin_neon_vfmaq_laneq_v: { |
8066 | Function *F = CGM.getIntrinsic(Intrinsic::fma, Ty); |
8067 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
8068 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
8069 | |
8070 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
8071 | Ops[2] = EmitNeonSplat(Ops[2], cast<ConstantInt>(Ops[3])); |
8072 | return Builder.CreateCall(F, {Ops[2], Ops[1], Ops[0]}); |
8073 | } |
8074 | case NEON::BI__builtin_neon_vfmah_lane_f16: |
8075 | case NEON::BI__builtin_neon_vfmas_lane_f32: |
8076 | case NEON::BI__builtin_neon_vfmah_laneq_f16: |
8077 | case NEON::BI__builtin_neon_vfmas_laneq_f32: |
8078 | case NEON::BI__builtin_neon_vfmad_lane_f64: |
8079 | case NEON::BI__builtin_neon_vfmad_laneq_f64: { |
8080 | Ops.push_back(EmitScalarExpr(E->getArg(3))); |
8081 | llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext())); |
8082 | Function *F = CGM.getIntrinsic(Intrinsic::fma, Ty); |
8083 | Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); |
8084 | return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); |
8085 | } |
8086 | case NEON::BI__builtin_neon_vmull_v: |
8087 | // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. |
8088 | Int = usgn ? Intrinsic::aarch64_neon_umull : Intrinsic::aarch64_neon_smull; |
8089 | if (Type.isPoly()) Int = Intrinsic::aarch64_neon_pmull; |
8090 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); |
8091 | case NEON::BI__builtin_neon_vmax_v: |
8092 | case NEON::BI__builtin_neon_vmaxq_v: |
8093 | // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. |
8094 | Int = usgn ? Intrinsic::aarch64_neon_umax : Intrinsic::aarch64_neon_smax; |
8095 | if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmax; |
8096 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax"); |
8097 | case NEON::BI__builtin_neon_vmaxh_f16: { |
8098 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
8099 | Int = Intrinsic::aarch64_neon_fmax; |
8100 | return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vmax"); |
8101 | } |
8102 | case NEON::BI__builtin_neon_vmin_v: |
8103 | case NEON::BI__builtin_neon_vminq_v: |
8104 | // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. |
8105 | Int = usgn ? Intrinsic::aarch64_neon_umin : Intrinsic::aarch64_neon_smin; |
8106 | if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmin; |
8107 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin"); |
8108 | case NEON::BI__builtin_neon_vminh_f16: { |
8109 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
8110 | Int = Intrinsic::aarch64_neon_fmin; |
8111 | return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vmin"); |
8112 | } |
8113 | case NEON::BI__builtin_neon_vabd_v: |
8114 | case NEON::BI__builtin_neon_vabdq_v: |
8115 | // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. |
8116 | Int = usgn ? Intrinsic::aarch64_neon_uabd : Intrinsic::aarch64_neon_sabd; |
8117 | if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fabd; |
8118 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd"); |
8119 | case NEON::BI__builtin_neon_vpadal_v: |
8120 | case NEON::BI__builtin_neon_vpadalq_v: { |
8121 | unsigned ArgElts = VTy->getNumElements(); |
8122 | llvm::IntegerType *EltTy = cast<IntegerType>(VTy->getElementType()); |
8123 | unsigned BitWidth = EltTy->getBitWidth(); |
8124 | llvm::Type *ArgTy = llvm::VectorType::get( |
8125 | llvm::IntegerType::get(getLLVMContext(), BitWidth/2), 2*ArgElts); |
8126 | llvm::Type* Tys[2] = { VTy, ArgTy }; |
8127 | Int = usgn ? Intrinsic::aarch64_neon_uaddlp : Intrinsic::aarch64_neon_saddlp; |
8128 | SmallVector<llvm::Value*, 1> TmpOps; |
8129 | TmpOps.push_back(Ops[1]); |
8130 | Function *F = CGM.getIntrinsic(Int, Tys); |
8131 | llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vpadal"); |
8132 | llvm::Value *addend = Builder.CreateBitCast(Ops[0], tmp->getType()); |
8133 | return Builder.CreateAdd(tmp, addend); |
8134 | } |
8135 | case NEON::BI__builtin_neon_vpmin_v: |
8136 | case NEON::BI__builtin_neon_vpminq_v: |
8137 | // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. |
8138 | Int = usgn ? Intrinsic::aarch64_neon_uminp : Intrinsic::aarch64_neon_sminp; |
8139 | if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fminp; |
8140 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin"); |
8141 | case NEON::BI__builtin_neon_vpmax_v: |
8142 | case NEON::BI__builtin_neon_vpmaxq_v: |
8143 | // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. |
8144 | Int = usgn ? Intrinsic::aarch64_neon_umaxp : Intrinsic::aarch64_neon_smaxp; |
8145 | if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmaxp; |
8146 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax"); |
8147 | case NEON::BI__builtin_neon_vminnm_v: |
8148 | case NEON::BI__builtin_neon_vminnmq_v: |
8149 | Int = Intrinsic::aarch64_neon_fminnm; |
8150 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vminnm"); |
8151 | case NEON::BI__builtin_neon_vminnmh_f16: |
8152 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
8153 | Int = Intrinsic::aarch64_neon_fminnm; |
8154 | return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vminnm"); |
8155 | case NEON::BI__builtin_neon_vmaxnm_v: |
8156 | case NEON::BI__builtin_neon_vmaxnmq_v: |
8157 | Int = Intrinsic::aarch64_neon_fmaxnm; |
8158 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmaxnm"); |
8159 | case NEON::BI__builtin_neon_vmaxnmh_f16: |
8160 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
8161 | Int = Intrinsic::aarch64_neon_fmaxnm; |
8162 | return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vmaxnm"); |
8163 | case NEON::BI__builtin_neon_vrecpss_f32: { |
8164 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
8165 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, FloatTy), |
8166 | Ops, "vrecps"); |
8167 | } |
8168 | case NEON::BI__builtin_neon_vrecpsd_f64: |
8169 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
8170 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, DoubleTy), |
8171 | Ops, "vrecps"); |
8172 | case NEON::BI__builtin_neon_vrecpsh_f16: |
8173 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
8174 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, HalfTy), |
8175 | Ops, "vrecps"); |
8176 | case NEON::BI__builtin_neon_vqshrun_n_v: |
8177 | Int = Intrinsic::aarch64_neon_sqshrun; |
8178 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrun_n"); |
8179 | case NEON::BI__builtin_neon_vqrshrun_n_v: |
8180 | Int = Intrinsic::aarch64_neon_sqrshrun; |
8181 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrun_n"); |
8182 | case NEON::BI__builtin_neon_vqshrn_n_v: |
8183 | Int = usgn ? Intrinsic::aarch64_neon_uqshrn : Intrinsic::aarch64_neon_sqshrn; |
8184 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n"); |
8185 | case NEON::BI__builtin_neon_vrshrn_n_v: |
8186 | Int = Intrinsic::aarch64_neon_rshrn; |
8187 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshrn_n"); |
8188 | case NEON::BI__builtin_neon_vqrshrn_n_v: |
8189 | Int = usgn ? Intrinsic::aarch64_neon_uqrshrn : Intrinsic::aarch64_neon_sqrshrn; |
8190 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n"); |
8191 | case NEON::BI__builtin_neon_vrndah_f16: { |
8192 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8193 | Int = Intrinsic::round; |
8194 | return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrnda"); |
8195 | } |
8196 | case NEON::BI__builtin_neon_vrnda_v: |
8197 | case NEON::BI__builtin_neon_vrndaq_v: { |
8198 | Int = Intrinsic::round; |
8199 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnda"); |
8200 | } |
8201 | case NEON::BI__builtin_neon_vrndih_f16: { |
8202 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8203 | Int = Intrinsic::nearbyint; |
8204 | return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrndi"); |
8205 | } |
8206 | case NEON::BI__builtin_neon_vrndmh_f16: { |
8207 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8208 | Int = Intrinsic::floor; |
8209 | return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrndm"); |
8210 | } |
8211 | case NEON::BI__builtin_neon_vrndm_v: |
8212 | case NEON::BI__builtin_neon_vrndmq_v: { |
8213 | Int = Intrinsic::floor; |
8214 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndm"); |
8215 | } |
8216 | case NEON::BI__builtin_neon_vrndnh_f16: { |
8217 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8218 | Int = Intrinsic::aarch64_neon_frintn; |
8219 | return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrndn"); |
8220 | } |
8221 | case NEON::BI__builtin_neon_vrndn_v: |
8222 | case NEON::BI__builtin_neon_vrndnq_v: { |
8223 | Int = Intrinsic::aarch64_neon_frintn; |
8224 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndn"); |
8225 | } |
8226 | case NEON::BI__builtin_neon_vrndns_f32: { |
8227 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8228 | Int = Intrinsic::aarch64_neon_frintn; |
8229 | return EmitNeonCall(CGM.getIntrinsic(Int, FloatTy), Ops, "vrndn"); |
8230 | } |
8231 | case NEON::BI__builtin_neon_vrndph_f16: { |
8232 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8233 | Int = Intrinsic::ceil; |
8234 | return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrndp"); |
8235 | } |
8236 | case NEON::BI__builtin_neon_vrndp_v: |
8237 | case NEON::BI__builtin_neon_vrndpq_v: { |
8238 | Int = Intrinsic::ceil; |
8239 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndp"); |
8240 | } |
8241 | case NEON::BI__builtin_neon_vrndxh_f16: { |
8242 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8243 | Int = Intrinsic::rint; |
8244 | return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrndx"); |
8245 | } |
8246 | case NEON::BI__builtin_neon_vrndx_v: |
8247 | case NEON::BI__builtin_neon_vrndxq_v: { |
8248 | Int = Intrinsic::rint; |
8249 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndx"); |
8250 | } |
8251 | case NEON::BI__builtin_neon_vrndh_f16: { |
8252 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8253 | Int = Intrinsic::trunc; |
8254 | return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrndz"); |
8255 | } |
8256 | case NEON::BI__builtin_neon_vrnd_v: |
8257 | case NEON::BI__builtin_neon_vrndq_v: { |
8258 | Int = Intrinsic::trunc; |
8259 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndz"); |
8260 | } |
8261 | case NEON::BI__builtin_neon_vcvt_f64_v: |
8262 | case NEON::BI__builtin_neon_vcvtq_f64_v: |
8263 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
8264 | Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, quad)); |
8265 | return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") |
8266 | : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); |
8267 | case NEON::BI__builtin_neon_vcvt_f64_f32: { |
8268 | assert(Type.getEltType() == NeonTypeFlags::Float64 && quad &&((Type.getEltType() == NeonTypeFlags::Float64 && quad && "unexpected vcvt_f64_f32 builtin") ? static_cast< void> (0) : __assert_fail ("Type.getEltType() == NeonTypeFlags::Float64 && quad && \"unexpected vcvt_f64_f32 builtin\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 8269, __PRETTY_FUNCTION__)) |
8269 | "unexpected vcvt_f64_f32 builtin")((Type.getEltType() == NeonTypeFlags::Float64 && quad && "unexpected vcvt_f64_f32 builtin") ? static_cast< void> (0) : __assert_fail ("Type.getEltType() == NeonTypeFlags::Float64 && quad && \"unexpected vcvt_f64_f32 builtin\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 8269, __PRETTY_FUNCTION__)); |
8270 | NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float32, false, false); |
8271 | Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); |
8272 | |
8273 | return Builder.CreateFPExt(Ops[0], Ty, "vcvt"); |
8274 | } |
8275 | case NEON::BI__builtin_neon_vcvt_f32_f64: { |
8276 | assert(Type.getEltType() == NeonTypeFlags::Float32 &&((Type.getEltType() == NeonTypeFlags::Float32 && "unexpected vcvt_f32_f64 builtin" ) ? static_cast<void> (0) : __assert_fail ("Type.getEltType() == NeonTypeFlags::Float32 && \"unexpected vcvt_f32_f64 builtin\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 8277, __PRETTY_FUNCTION__)) |
8277 | "unexpected vcvt_f32_f64 builtin")((Type.getEltType() == NeonTypeFlags::Float32 && "unexpected vcvt_f32_f64 builtin" ) ? static_cast<void> (0) : __assert_fail ("Type.getEltType() == NeonTypeFlags::Float32 && \"unexpected vcvt_f32_f64 builtin\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 8277, __PRETTY_FUNCTION__)); |
8278 | NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float64, false, true); |
8279 | Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); |
8280 | |
8281 | return Builder.CreateFPTrunc(Ops[0], Ty, "vcvt"); |
8282 | } |
8283 | case NEON::BI__builtin_neon_vcvt_s32_v: |
8284 | case NEON::BI__builtin_neon_vcvt_u32_v: |
8285 | case NEON::BI__builtin_neon_vcvt_s64_v: |
8286 | case NEON::BI__builtin_neon_vcvt_u64_v: |
8287 | case NEON::BI__builtin_neon_vcvt_s16_v: |
8288 | case NEON::BI__builtin_neon_vcvt_u16_v: |
8289 | case NEON::BI__builtin_neon_vcvtq_s32_v: |
8290 | case NEON::BI__builtin_neon_vcvtq_u32_v: |
8291 | case NEON::BI__builtin_neon_vcvtq_s64_v: |
8292 | case NEON::BI__builtin_neon_vcvtq_u64_v: |
8293 | case NEON::BI__builtin_neon_vcvtq_s16_v: |
8294 | case NEON::BI__builtin_neon_vcvtq_u16_v: { |
8295 | Ops[0] = Builder.CreateBitCast(Ops[0], GetFloatNeonType(this, Type)); |
8296 | if (usgn) |
8297 | return Builder.CreateFPToUI(Ops[0], Ty); |
8298 | return Builder.CreateFPToSI(Ops[0], Ty); |
8299 | } |
8300 | case NEON::BI__builtin_neon_vcvta_s16_v: |
8301 | case NEON::BI__builtin_neon_vcvta_u16_v: |
8302 | case NEON::BI__builtin_neon_vcvta_s32_v: |
8303 | case NEON::BI__builtin_neon_vcvtaq_s16_v: |
8304 | case NEON::BI__builtin_neon_vcvtaq_s32_v: |
8305 | case NEON::BI__builtin_neon_vcvta_u32_v: |
8306 | case NEON::BI__builtin_neon_vcvtaq_u16_v: |
8307 | case NEON::BI__builtin_neon_vcvtaq_u32_v: |
8308 | case NEON::BI__builtin_neon_vcvta_s64_v: |
8309 | case NEON::BI__builtin_neon_vcvtaq_s64_v: |
8310 | case NEON::BI__builtin_neon_vcvta_u64_v: |
8311 | case NEON::BI__builtin_neon_vcvtaq_u64_v: { |
8312 | Int = usgn ? Intrinsic::aarch64_neon_fcvtau : Intrinsic::aarch64_neon_fcvtas; |
8313 | llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; |
8314 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvta"); |
8315 | } |
8316 | case NEON::BI__builtin_neon_vcvtm_s16_v: |
8317 | case NEON::BI__builtin_neon_vcvtm_s32_v: |
8318 | case NEON::BI__builtin_neon_vcvtmq_s16_v: |
8319 | case NEON::BI__builtin_neon_vcvtmq_s32_v: |
8320 | case NEON::BI__builtin_neon_vcvtm_u16_v: |
8321 | case NEON::BI__builtin_neon_vcvtm_u32_v: |
8322 | case NEON::BI__builtin_neon_vcvtmq_u16_v: |
8323 | case NEON::BI__builtin_neon_vcvtmq_u32_v: |
8324 | case NEON::BI__builtin_neon_vcvtm_s64_v: |
8325 | case NEON::BI__builtin_neon_vcvtmq_s64_v: |
8326 | case NEON::BI__builtin_neon_vcvtm_u64_v: |
8327 | case NEON::BI__builtin_neon_vcvtmq_u64_v: { |
8328 | Int = usgn ? Intrinsic::aarch64_neon_fcvtmu : Intrinsic::aarch64_neon_fcvtms; |
8329 | llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; |
8330 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtm"); |
8331 | } |
8332 | case NEON::BI__builtin_neon_vcvtn_s16_v: |
8333 | case NEON::BI__builtin_neon_vcvtn_s32_v: |
8334 | case NEON::BI__builtin_neon_vcvtnq_s16_v: |
8335 | case NEON::BI__builtin_neon_vcvtnq_s32_v: |
8336 | case NEON::BI__builtin_neon_vcvtn_u16_v: |
8337 | case NEON::BI__builtin_neon_vcvtn_u32_v: |
8338 | case NEON::BI__builtin_neon_vcvtnq_u16_v: |
8339 | case NEON::BI__builtin_neon_vcvtnq_u32_v: |
8340 | case NEON::BI__builtin_neon_vcvtn_s64_v: |
8341 | case NEON::BI__builtin_neon_vcvtnq_s64_v: |
8342 | case NEON::BI__builtin_neon_vcvtn_u64_v: |
8343 | case NEON::BI__builtin_neon_vcvtnq_u64_v: { |
8344 | Int = usgn ? Intrinsic::aarch64_neon_fcvtnu : Intrinsic::aarch64_neon_fcvtns; |
8345 | llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; |
8346 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtn"); |
8347 | } |
8348 | case NEON::BI__builtin_neon_vcvtp_s16_v: |
8349 | case NEON::BI__builtin_neon_vcvtp_s32_v: |
8350 | case NEON::BI__builtin_neon_vcvtpq_s16_v: |
8351 | case NEON::BI__builtin_neon_vcvtpq_s32_v: |
8352 | case NEON::BI__builtin_neon_vcvtp_u16_v: |
8353 | case NEON::BI__builtin_neon_vcvtp_u32_v: |
8354 | case NEON::BI__builtin_neon_vcvtpq_u16_v: |
8355 | case NEON::BI__builtin_neon_vcvtpq_u32_v: |
8356 | case NEON::BI__builtin_neon_vcvtp_s64_v: |
8357 | case NEON::BI__builtin_neon_vcvtpq_s64_v: |
8358 | case NEON::BI__builtin_neon_vcvtp_u64_v: |
8359 | case NEON::BI__builtin_neon_vcvtpq_u64_v: { |
8360 | Int = usgn ? Intrinsic::aarch64_neon_fcvtpu : Intrinsic::aarch64_neon_fcvtps; |
8361 | llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; |
8362 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtp"); |
8363 | } |
8364 | case NEON::BI__builtin_neon_vmulx_v: |
8365 | case NEON::BI__builtin_neon_vmulxq_v: { |
8366 | Int = Intrinsic::aarch64_neon_fmulx; |
8367 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmulx"); |
8368 | } |
8369 | case NEON::BI__builtin_neon_vmulxh_lane_f16: |
8370 | case NEON::BI__builtin_neon_vmulxh_laneq_f16: { |
8371 | // vmulx_lane should be mapped to Neon scalar mulx after |
8372 | // extracting the scalar element |
8373 | Ops.push_back(EmitScalarExpr(E->getArg(2))); |
8374 | Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract"); |
8375 | Ops.pop_back(); |
8376 | Int = Intrinsic::aarch64_neon_fmulx; |
8377 | return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vmulx"); |
8378 | } |
8379 | case NEON::BI__builtin_neon_vmul_lane_v: |
8380 | case NEON::BI__builtin_neon_vmul_laneq_v: { |
8381 | // v1f64 vmul_lane should be mapped to Neon scalar mul lane |
8382 | bool Quad = false; |
8383 | if (BuiltinID == NEON::BI__builtin_neon_vmul_laneq_v) |
8384 | Quad = true; |
8385 | Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); |
8386 | llvm::Type *VTy = GetNeonType(this, |
8387 | NeonTypeFlags(NeonTypeFlags::Float64, false, Quad)); |
8388 | Ops[1] = Builder.CreateBitCast(Ops[1], VTy); |
8389 | Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract"); |
8390 | Value *Result = Builder.CreateFMul(Ops[0], Ops[1]); |
8391 | return Builder.CreateBitCast(Result, Ty); |
8392 | } |
8393 | case NEON::BI__builtin_neon_vnegd_s64: |
8394 | return Builder.CreateNeg(EmitScalarExpr(E->getArg(0)), "vnegd"); |
8395 | case NEON::BI__builtin_neon_vnegh_f16: |
8396 | return Builder.CreateFNeg(EmitScalarExpr(E->getArg(0)), "vnegh"); |
8397 | case NEON::BI__builtin_neon_vpmaxnm_v: |
8398 | case NEON::BI__builtin_neon_vpmaxnmq_v: { |
8399 | Int = Intrinsic::aarch64_neon_fmaxnmp; |
8400 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmaxnm"); |
8401 | } |
8402 | case NEON::BI__builtin_neon_vpminnm_v: |
8403 | case NEON::BI__builtin_neon_vpminnmq_v: { |
8404 | Int = Intrinsic::aarch64_neon_fminnmp; |
8405 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpminnm"); |
8406 | } |
8407 | case NEON::BI__builtin_neon_vsqrth_f16: { |
8408 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8409 | Int = Intrinsic::sqrt; |
8410 | return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vsqrt"); |
8411 | } |
8412 | case NEON::BI__builtin_neon_vsqrt_v: |
8413 | case NEON::BI__builtin_neon_vsqrtq_v: { |
8414 | Int = Intrinsic::sqrt; |
8415 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
8416 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqrt"); |
8417 | } |
8418 | case NEON::BI__builtin_neon_vrbit_v: |
8419 | case NEON::BI__builtin_neon_vrbitq_v: { |
8420 | Int = Intrinsic::aarch64_neon_rbit; |
8421 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrbit"); |
8422 | } |
8423 | case NEON::BI__builtin_neon_vaddv_u8: |
8424 | // FIXME: These are handled by the AArch64 scalar code. |
8425 | usgn = true; |
8426 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
8427 | case NEON::BI__builtin_neon_vaddv_s8: { |
8428 | Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; |
8429 | Ty = Int32Ty; |
8430 | VTy = llvm::VectorType::get(Int8Ty, 8); |
8431 | llvm::Type *Tys[2] = { Ty, VTy }; |
8432 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8433 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); |
8434 | return Builder.CreateTrunc(Ops[0], Int8Ty); |
8435 | } |
8436 | case NEON::BI__builtin_neon_vaddv_u16: |
8437 | usgn = true; |
8438 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
8439 | case NEON::BI__builtin_neon_vaddv_s16: { |
8440 | Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; |
8441 | Ty = Int32Ty; |
8442 | VTy = llvm::VectorType::get(Int16Ty, 4); |
8443 | llvm::Type *Tys[2] = { Ty, VTy }; |
8444 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8445 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); |
8446 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
8447 | } |
8448 | case NEON::BI__builtin_neon_vaddvq_u8: |
8449 | usgn = true; |
8450 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
8451 | case NEON::BI__builtin_neon_vaddvq_s8: { |
8452 | Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; |
8453 | Ty = Int32Ty; |
8454 | VTy = llvm::VectorType::get(Int8Ty, 16); |
8455 | llvm::Type *Tys[2] = { Ty, VTy }; |
8456 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8457 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); |
8458 | return Builder.CreateTrunc(Ops[0], Int8Ty); |
8459 | } |
8460 | case NEON::BI__builtin_neon_vaddvq_u16: |
8461 | usgn = true; |
8462 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
8463 | case NEON::BI__builtin_neon_vaddvq_s16: { |
8464 | Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; |
8465 | Ty = Int32Ty; |
8466 | VTy = llvm::VectorType::get(Int16Ty, 8); |
8467 | llvm::Type *Tys[2] = { Ty, VTy }; |
8468 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8469 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); |
8470 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
8471 | } |
8472 | case NEON::BI__builtin_neon_vmaxv_u8: { |
8473 | Int = Intrinsic::aarch64_neon_umaxv; |
8474 | Ty = Int32Ty; |
8475 | VTy = llvm::VectorType::get(Int8Ty, 8); |
8476 | llvm::Type *Tys[2] = { Ty, VTy }; |
8477 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8478 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); |
8479 | return Builder.CreateTrunc(Ops[0], Int8Ty); |
8480 | } |
8481 | case NEON::BI__builtin_neon_vmaxv_u16: { |
8482 | Int = Intrinsic::aarch64_neon_umaxv; |
8483 | Ty = Int32Ty; |
8484 | VTy = llvm::VectorType::get(Int16Ty, 4); |
8485 | llvm::Type *Tys[2] = { Ty, VTy }; |
8486 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8487 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); |
8488 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
8489 | } |
8490 | case NEON::BI__builtin_neon_vmaxvq_u8: { |
8491 | Int = Intrinsic::aarch64_neon_umaxv; |
8492 | Ty = Int32Ty; |
8493 | VTy = llvm::VectorType::get(Int8Ty, 16); |
8494 | llvm::Type *Tys[2] = { Ty, VTy }; |
8495 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8496 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); |
8497 | return Builder.CreateTrunc(Ops[0], Int8Ty); |
8498 | } |
8499 | case NEON::BI__builtin_neon_vmaxvq_u16: { |
8500 | Int = Intrinsic::aarch64_neon_umaxv; |
8501 | Ty = Int32Ty; |
8502 | VTy = llvm::VectorType::get(Int16Ty, 8); |
8503 | llvm::Type *Tys[2] = { Ty, VTy }; |
8504 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8505 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); |
8506 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
8507 | } |
8508 | case NEON::BI__builtin_neon_vmaxv_s8: { |
8509 | Int = Intrinsic::aarch64_neon_smaxv; |
8510 | Ty = Int32Ty; |
8511 | VTy = llvm::VectorType::get(Int8Ty, 8); |
8512 | llvm::Type *Tys[2] = { Ty, VTy }; |
8513 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8514 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); |
8515 | return Builder.CreateTrunc(Ops[0], Int8Ty); |
8516 | } |
8517 | case NEON::BI__builtin_neon_vmaxv_s16: { |
8518 | Int = Intrinsic::aarch64_neon_smaxv; |
8519 | Ty = Int32Ty; |
8520 | VTy = llvm::VectorType::get(Int16Ty, 4); |
8521 | llvm::Type *Tys[2] = { Ty, VTy }; |
8522 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8523 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); |
8524 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
8525 | } |
8526 | case NEON::BI__builtin_neon_vmaxvq_s8: { |
8527 | Int = Intrinsic::aarch64_neon_smaxv; |
8528 | Ty = Int32Ty; |
8529 | VTy = llvm::VectorType::get(Int8Ty, 16); |
8530 | llvm::Type *Tys[2] = { Ty, VTy }; |
8531 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8532 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); |
8533 | return Builder.CreateTrunc(Ops[0], Int8Ty); |
8534 | } |
8535 | case NEON::BI__builtin_neon_vmaxvq_s16: { |
8536 | Int = Intrinsic::aarch64_neon_smaxv; |
8537 | Ty = Int32Ty; |
8538 | VTy = llvm::VectorType::get(Int16Ty, 8); |
8539 | llvm::Type *Tys[2] = { Ty, VTy }; |
8540 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8541 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); |
8542 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
8543 | } |
8544 | case NEON::BI__builtin_neon_vmaxv_f16: { |
8545 | Int = Intrinsic::aarch64_neon_fmaxv; |
8546 | Ty = HalfTy; |
8547 | VTy = llvm::VectorType::get(HalfTy, 4); |
8548 | llvm::Type *Tys[2] = { Ty, VTy }; |
8549 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8550 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); |
8551 | return Builder.CreateTrunc(Ops[0], HalfTy); |
8552 | } |
8553 | case NEON::BI__builtin_neon_vmaxvq_f16: { |
8554 | Int = Intrinsic::aarch64_neon_fmaxv; |
8555 | Ty = HalfTy; |
8556 | VTy = llvm::VectorType::get(HalfTy, 8); |
8557 | llvm::Type *Tys[2] = { Ty, VTy }; |
8558 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8559 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); |
8560 | return Builder.CreateTrunc(Ops[0], HalfTy); |
8561 | } |
8562 | case NEON::BI__builtin_neon_vminv_u8: { |
8563 | Int = Intrinsic::aarch64_neon_uminv; |
8564 | Ty = Int32Ty; |
8565 | VTy = llvm::VectorType::get(Int8Ty, 8); |
8566 | llvm::Type *Tys[2] = { Ty, VTy }; |
8567 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8568 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); |
8569 | return Builder.CreateTrunc(Ops[0], Int8Ty); |
8570 | } |
8571 | case NEON::BI__builtin_neon_vminv_u16: { |
8572 | Int = Intrinsic::aarch64_neon_uminv; |
8573 | Ty = Int32Ty; |
8574 | VTy = llvm::VectorType::get(Int16Ty, 4); |
8575 | llvm::Type *Tys[2] = { Ty, VTy }; |
8576 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8577 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); |
8578 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
8579 | } |
8580 | case NEON::BI__builtin_neon_vminvq_u8: { |
8581 | Int = Intrinsic::aarch64_neon_uminv; |
8582 | Ty = Int32Ty; |
8583 | VTy = llvm::VectorType::get(Int8Ty, 16); |
8584 | llvm::Type *Tys[2] = { Ty, VTy }; |
8585 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8586 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); |
8587 | return Builder.CreateTrunc(Ops[0], Int8Ty); |
8588 | } |
8589 | case NEON::BI__builtin_neon_vminvq_u16: { |
8590 | Int = Intrinsic::aarch64_neon_uminv; |
8591 | Ty = Int32Ty; |
8592 | VTy = llvm::VectorType::get(Int16Ty, 8); |
8593 | llvm::Type *Tys[2] = { Ty, VTy }; |
8594 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8595 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); |
8596 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
8597 | } |
8598 | case NEON::BI__builtin_neon_vminv_s8: { |
8599 | Int = Intrinsic::aarch64_neon_sminv; |
8600 | Ty = Int32Ty; |
8601 | VTy = llvm::VectorType::get(Int8Ty, 8); |
8602 | llvm::Type *Tys[2] = { Ty, VTy }; |
8603 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8604 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); |
8605 | return Builder.CreateTrunc(Ops[0], Int8Ty); |
8606 | } |
8607 | case NEON::BI__builtin_neon_vminv_s16: { |
8608 | Int = Intrinsic::aarch64_neon_sminv; |
8609 | Ty = Int32Ty; |
8610 | VTy = llvm::VectorType::get(Int16Ty, 4); |
8611 | llvm::Type *Tys[2] = { Ty, VTy }; |
8612 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8613 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); |
8614 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
8615 | } |
8616 | case NEON::BI__builtin_neon_vminvq_s8: { |
8617 | Int = Intrinsic::aarch64_neon_sminv; |
8618 | Ty = Int32Ty; |
8619 | VTy = llvm::VectorType::get(Int8Ty, 16); |
8620 | llvm::Type *Tys[2] = { Ty, VTy }; |
8621 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8622 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); |
8623 | return Builder.CreateTrunc(Ops[0], Int8Ty); |
8624 | } |
8625 | case NEON::BI__builtin_neon_vminvq_s16: { |
8626 | Int = Intrinsic::aarch64_neon_sminv; |
8627 | Ty = Int32Ty; |
8628 | VTy = llvm::VectorType::get(Int16Ty, 8); |
8629 | llvm::Type *Tys[2] = { Ty, VTy }; |
8630 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8631 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); |
8632 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
8633 | } |
8634 | case NEON::BI__builtin_neon_vminv_f16: { |
8635 | Int = Intrinsic::aarch64_neon_fminv; |
8636 | Ty = HalfTy; |
8637 | VTy = llvm::VectorType::get(HalfTy, 4); |
8638 | llvm::Type *Tys[2] = { Ty, VTy }; |
8639 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8640 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); |
8641 | return Builder.CreateTrunc(Ops[0], HalfTy); |
8642 | } |
8643 | case NEON::BI__builtin_neon_vminvq_f16: { |
8644 | Int = Intrinsic::aarch64_neon_fminv; |
8645 | Ty = HalfTy; |
8646 | VTy = llvm::VectorType::get(HalfTy, 8); |
8647 | llvm::Type *Tys[2] = { Ty, VTy }; |
8648 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8649 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); |
8650 | return Builder.CreateTrunc(Ops[0], HalfTy); |
8651 | } |
8652 | case NEON::BI__builtin_neon_vmaxnmv_f16: { |
8653 | Int = Intrinsic::aarch64_neon_fmaxnmv; |
8654 | Ty = HalfTy; |
8655 | VTy = llvm::VectorType::get(HalfTy, 4); |
8656 | llvm::Type *Tys[2] = { Ty, VTy }; |
8657 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8658 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxnmv"); |
8659 | return Builder.CreateTrunc(Ops[0], HalfTy); |
8660 | } |
8661 | case NEON::BI__builtin_neon_vmaxnmvq_f16: { |
8662 | Int = Intrinsic::aarch64_neon_fmaxnmv; |
8663 | Ty = HalfTy; |
8664 | VTy = llvm::VectorType::get(HalfTy, 8); |
8665 | llvm::Type *Tys[2] = { Ty, VTy }; |
8666 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8667 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxnmv"); |
8668 | return Builder.CreateTrunc(Ops[0], HalfTy); |
8669 | } |
8670 | case NEON::BI__builtin_neon_vminnmv_f16: { |
8671 | Int = Intrinsic::aarch64_neon_fminnmv; |
8672 | Ty = HalfTy; |
8673 | VTy = llvm::VectorType::get(HalfTy, 4); |
8674 | llvm::Type *Tys[2] = { Ty, VTy }; |
8675 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8676 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminnmv"); |
8677 | return Builder.CreateTrunc(Ops[0], HalfTy); |
8678 | } |
8679 | case NEON::BI__builtin_neon_vminnmvq_f16: { |
8680 | Int = Intrinsic::aarch64_neon_fminnmv; |
8681 | Ty = HalfTy; |
8682 | VTy = llvm::VectorType::get(HalfTy, 8); |
8683 | llvm::Type *Tys[2] = { Ty, VTy }; |
8684 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8685 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminnmv"); |
8686 | return Builder.CreateTrunc(Ops[0], HalfTy); |
8687 | } |
8688 | case NEON::BI__builtin_neon_vmul_n_f64: { |
8689 | Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); |
8690 | Value *RHS = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), DoubleTy); |
8691 | return Builder.CreateFMul(Ops[0], RHS); |
8692 | } |
8693 | case NEON::BI__builtin_neon_vaddlv_u8: { |
8694 | Int = Intrinsic::aarch64_neon_uaddlv; |
8695 | Ty = Int32Ty; |
8696 | VTy = llvm::VectorType::get(Int8Ty, 8); |
8697 | llvm::Type *Tys[2] = { Ty, VTy }; |
8698 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8699 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); |
8700 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
8701 | } |
8702 | case NEON::BI__builtin_neon_vaddlv_u16: { |
8703 | Int = Intrinsic::aarch64_neon_uaddlv; |
8704 | Ty = Int32Ty; |
8705 | VTy = llvm::VectorType::get(Int16Ty, 4); |
8706 | llvm::Type *Tys[2] = { Ty, VTy }; |
8707 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8708 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); |
8709 | } |
8710 | case NEON::BI__builtin_neon_vaddlvq_u8: { |
8711 | Int = Intrinsic::aarch64_neon_uaddlv; |
8712 | Ty = Int32Ty; |
8713 | VTy = llvm::VectorType::get(Int8Ty, 16); |
8714 | llvm::Type *Tys[2] = { Ty, VTy }; |
8715 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8716 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); |
8717 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
8718 | } |
8719 | case NEON::BI__builtin_neon_vaddlvq_u16: { |
8720 | Int = Intrinsic::aarch64_neon_uaddlv; |
8721 | Ty = Int32Ty; |
8722 | VTy = llvm::VectorType::get(Int16Ty, 8); |
8723 | llvm::Type *Tys[2] = { Ty, VTy }; |
8724 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8725 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); |
8726 | } |
8727 | case NEON::BI__builtin_neon_vaddlv_s8: { |
8728 | Int = Intrinsic::aarch64_neon_saddlv; |
8729 | Ty = Int32Ty; |
8730 | VTy = llvm::VectorType::get(Int8Ty, 8); |
8731 | llvm::Type *Tys[2] = { Ty, VTy }; |
8732 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8733 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); |
8734 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
8735 | } |
8736 | case NEON::BI__builtin_neon_vaddlv_s16: { |
8737 | Int = Intrinsic::aarch64_neon_saddlv; |
8738 | Ty = Int32Ty; |
8739 | VTy = llvm::VectorType::get(Int16Ty, 4); |
8740 | llvm::Type *Tys[2] = { Ty, VTy }; |
8741 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8742 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); |
8743 | } |
8744 | case NEON::BI__builtin_neon_vaddlvq_s8: { |
8745 | Int = Intrinsic::aarch64_neon_saddlv; |
8746 | Ty = Int32Ty; |
8747 | VTy = llvm::VectorType::get(Int8Ty, 16); |
8748 | llvm::Type *Tys[2] = { Ty, VTy }; |
8749 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8750 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); |
8751 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
8752 | } |
8753 | case NEON::BI__builtin_neon_vaddlvq_s16: { |
8754 | Int = Intrinsic::aarch64_neon_saddlv; |
8755 | Ty = Int32Ty; |
8756 | VTy = llvm::VectorType::get(Int16Ty, 8); |
8757 | llvm::Type *Tys[2] = { Ty, VTy }; |
8758 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
8759 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); |
8760 | } |
8761 | case NEON::BI__builtin_neon_vsri_n_v: |
8762 | case NEON::BI__builtin_neon_vsriq_n_v: { |
8763 | Int = Intrinsic::aarch64_neon_vsri; |
8764 | llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); |
8765 | return EmitNeonCall(Intrin, Ops, "vsri_n"); |
8766 | } |
8767 | case NEON::BI__builtin_neon_vsli_n_v: |
8768 | case NEON::BI__builtin_neon_vsliq_n_v: { |
8769 | Int = Intrinsic::aarch64_neon_vsli; |
8770 | llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); |
8771 | return EmitNeonCall(Intrin, Ops, "vsli_n"); |
8772 | } |
8773 | case NEON::BI__builtin_neon_vsra_n_v: |
8774 | case NEON::BI__builtin_neon_vsraq_n_v: |
8775 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
8776 | Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); |
8777 | return Builder.CreateAdd(Ops[0], Ops[1]); |
8778 | case NEON::BI__builtin_neon_vrsra_n_v: |
8779 | case NEON::BI__builtin_neon_vrsraq_n_v: { |
8780 | Int = usgn ? Intrinsic::aarch64_neon_urshl : Intrinsic::aarch64_neon_srshl; |
8781 | SmallVector<llvm::Value*,2> TmpOps; |
8782 | TmpOps.push_back(Ops[1]); |
8783 | TmpOps.push_back(Ops[2]); |
8784 | Function* F = CGM.getIntrinsic(Int, Ty); |
8785 | llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vrshr_n", 1, true); |
8786 | Ops[0] = Builder.CreateBitCast(Ops[0], VTy); |
8787 | return Builder.CreateAdd(Ops[0], tmp); |
8788 | } |
8789 | case NEON::BI__builtin_neon_vld1_v: |
8790 | case NEON::BI__builtin_neon_vld1q_v: { |
8791 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); |
8792 | auto Alignment = CharUnits::fromQuantity( |
8793 | BuiltinID == NEON::BI__builtin_neon_vld1_v ? 8 : 16); |
8794 | return Builder.CreateAlignedLoad(VTy, Ops[0], Alignment); |
8795 | } |
8796 | case NEON::BI__builtin_neon_vst1_v: |
8797 | case NEON::BI__builtin_neon_vst1q_v: |
8798 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); |
8799 | Ops[1] = Builder.CreateBitCast(Ops[1], VTy); |
8800 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
8801 | case NEON::BI__builtin_neon_vld1_lane_v: |
8802 | case NEON::BI__builtin_neon_vld1q_lane_v: { |
8803 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
8804 | Ty = llvm::PointerType::getUnqual(VTy->getElementType()); |
8805 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
8806 | auto Alignment = CharUnits::fromQuantity( |
8807 | BuiltinID == NEON::BI__builtin_neon_vld1_lane_v ? 8 : 16); |
8808 | Ops[0] = |
8809 | Builder.CreateAlignedLoad(VTy->getElementType(), Ops[0], Alignment); |
8810 | return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vld1_lane"); |
8811 | } |
8812 | case NEON::BI__builtin_neon_vld1_dup_v: |
8813 | case NEON::BI__builtin_neon_vld1q_dup_v: { |
8814 | Value *V = UndefValue::get(Ty); |
8815 | Ty = llvm::PointerType::getUnqual(VTy->getElementType()); |
8816 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
8817 | auto Alignment = CharUnits::fromQuantity( |
8818 | BuiltinID == NEON::BI__builtin_neon_vld1_dup_v ? 8 : 16); |
8819 | Ops[0] = |
8820 | Builder.CreateAlignedLoad(VTy->getElementType(), Ops[0], Alignment); |
8821 | llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); |
8822 | Ops[0] = Builder.CreateInsertElement(V, Ops[0], CI); |
8823 | return EmitNeonSplat(Ops[0], CI); |
8824 | } |
8825 | case NEON::BI__builtin_neon_vst1_lane_v: |
8826 | case NEON::BI__builtin_neon_vst1q_lane_v: |
8827 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
8828 | Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); |
8829 | Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); |
8830 | return Builder.CreateDefaultAlignedStore(Ops[1], |
8831 | Builder.CreateBitCast(Ops[0], Ty)); |
8832 | case NEON::BI__builtin_neon_vld2_v: |
8833 | case NEON::BI__builtin_neon_vld2q_v: { |
8834 | llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); |
8835 | Ops[1] = Builder.CreateBitCast(Ops[1], PTy); |
8836 | llvm::Type *Tys[2] = { VTy, PTy }; |
8837 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2, Tys); |
8838 | Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); |
8839 | Ops[0] = Builder.CreateBitCast(Ops[0], |
8840 | llvm::PointerType::getUnqual(Ops[1]->getType())); |
8841 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
8842 | } |
8843 | case NEON::BI__builtin_neon_vld3_v: |
8844 | case NEON::BI__builtin_neon_vld3q_v: { |
8845 | llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); |
8846 | Ops[1] = Builder.CreateBitCast(Ops[1], PTy); |
8847 | llvm::Type *Tys[2] = { VTy, PTy }; |
8848 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3, Tys); |
8849 | Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); |
8850 | Ops[0] = Builder.CreateBitCast(Ops[0], |
8851 | llvm::PointerType::getUnqual(Ops[1]->getType())); |
8852 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
8853 | } |
8854 | case NEON::BI__builtin_neon_vld4_v: |
8855 | case NEON::BI__builtin_neon_vld4q_v: { |
8856 | llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); |
8857 | Ops[1] = Builder.CreateBitCast(Ops[1], PTy); |
8858 | llvm::Type *Tys[2] = { VTy, PTy }; |
8859 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4, Tys); |
8860 | Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); |
8861 | Ops[0] = Builder.CreateBitCast(Ops[0], |
8862 | llvm::PointerType::getUnqual(Ops[1]->getType())); |
8863 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
8864 | } |
8865 | case NEON::BI__builtin_neon_vld2_dup_v: |
8866 | case NEON::BI__builtin_neon_vld2q_dup_v: { |
8867 | llvm::Type *PTy = |
8868 | llvm::PointerType::getUnqual(VTy->getElementType()); |
8869 | Ops[1] = Builder.CreateBitCast(Ops[1], PTy); |
8870 | llvm::Type *Tys[2] = { VTy, PTy }; |
8871 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2r, Tys); |
8872 | Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); |
8873 | Ops[0] = Builder.CreateBitCast(Ops[0], |
8874 | llvm::PointerType::getUnqual(Ops[1]->getType())); |
8875 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
8876 | } |
8877 | case NEON::BI__builtin_neon_vld3_dup_v: |
8878 | case NEON::BI__builtin_neon_vld3q_dup_v: { |
8879 | llvm::Type *PTy = |
8880 | llvm::PointerType::getUnqual(VTy->getElementType()); |
8881 | Ops[1] = Builder.CreateBitCast(Ops[1], PTy); |
8882 | llvm::Type *Tys[2] = { VTy, PTy }; |
8883 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3r, Tys); |
8884 | Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); |
8885 | Ops[0] = Builder.CreateBitCast(Ops[0], |
8886 | llvm::PointerType::getUnqual(Ops[1]->getType())); |
8887 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
8888 | } |
8889 | case NEON::BI__builtin_neon_vld4_dup_v: |
8890 | case NEON::BI__builtin_neon_vld4q_dup_v: { |
8891 | llvm::Type *PTy = |
8892 | llvm::PointerType::getUnqual(VTy->getElementType()); |
8893 | Ops[1] = Builder.CreateBitCast(Ops[1], PTy); |
8894 | llvm::Type *Tys[2] = { VTy, PTy }; |
8895 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4r, Tys); |
8896 | Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); |
8897 | Ops[0] = Builder.CreateBitCast(Ops[0], |
8898 | llvm::PointerType::getUnqual(Ops[1]->getType())); |
8899 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
8900 | } |
8901 | case NEON::BI__builtin_neon_vld2_lane_v: |
8902 | case NEON::BI__builtin_neon_vld2q_lane_v: { |
8903 | llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; |
8904 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2lane, Tys); |
8905 | Ops.push_back(Ops[1]); |
8906 | Ops.erase(Ops.begin()+1); |
8907 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
8908 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
8909 | Ops[3] = Builder.CreateZExt(Ops[3], Int64Ty); |
8910 | Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld2_lane"); |
8911 | Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); |
8912 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
8913 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
8914 | } |
8915 | case NEON::BI__builtin_neon_vld3_lane_v: |
8916 | case NEON::BI__builtin_neon_vld3q_lane_v: { |
8917 | llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; |
8918 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3lane, Tys); |
8919 | Ops.push_back(Ops[1]); |
8920 | Ops.erase(Ops.begin()+1); |
8921 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
8922 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
8923 | Ops[3] = Builder.CreateBitCast(Ops[3], Ty); |
8924 | Ops[4] = Builder.CreateZExt(Ops[4], Int64Ty); |
8925 | Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane"); |
8926 | Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); |
8927 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
8928 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
8929 | } |
8930 | case NEON::BI__builtin_neon_vld4_lane_v: |
8931 | case NEON::BI__builtin_neon_vld4q_lane_v: { |
8932 | llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; |
8933 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4lane, Tys); |
8934 | Ops.push_back(Ops[1]); |
8935 | Ops.erase(Ops.begin()+1); |
8936 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
8937 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
8938 | Ops[3] = Builder.CreateBitCast(Ops[3], Ty); |
8939 | Ops[4] = Builder.CreateBitCast(Ops[4], Ty); |
8940 | Ops[5] = Builder.CreateZExt(Ops[5], Int64Ty); |
8941 | Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld4_lane"); |
8942 | Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); |
8943 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
8944 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
8945 | } |
8946 | case NEON::BI__builtin_neon_vst2_v: |
8947 | case NEON::BI__builtin_neon_vst2q_v: { |
8948 | Ops.push_back(Ops[0]); |
8949 | Ops.erase(Ops.begin()); |
8950 | llvm::Type *Tys[2] = { VTy, Ops[2]->getType() }; |
8951 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2, Tys), |
8952 | Ops, ""); |
8953 | } |
8954 | case NEON::BI__builtin_neon_vst2_lane_v: |
8955 | case NEON::BI__builtin_neon_vst2q_lane_v: { |
8956 | Ops.push_back(Ops[0]); |
8957 | Ops.erase(Ops.begin()); |
8958 | Ops[2] = Builder.CreateZExt(Ops[2], Int64Ty); |
8959 | llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; |
8960 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2lane, Tys), |
8961 | Ops, ""); |
8962 | } |
8963 | case NEON::BI__builtin_neon_vst3_v: |
8964 | case NEON::BI__builtin_neon_vst3q_v: { |
8965 | Ops.push_back(Ops[0]); |
8966 | Ops.erase(Ops.begin()); |
8967 | llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; |
8968 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3, Tys), |
8969 | Ops, ""); |
8970 | } |
8971 | case NEON::BI__builtin_neon_vst3_lane_v: |
8972 | case NEON::BI__builtin_neon_vst3q_lane_v: { |
8973 | Ops.push_back(Ops[0]); |
8974 | Ops.erase(Ops.begin()); |
8975 | Ops[3] = Builder.CreateZExt(Ops[3], Int64Ty); |
8976 | llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; |
8977 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3lane, Tys), |
8978 | Ops, ""); |
8979 | } |
8980 | case NEON::BI__builtin_neon_vst4_v: |
8981 | case NEON::BI__builtin_neon_vst4q_v: { |
8982 | Ops.push_back(Ops[0]); |
8983 | Ops.erase(Ops.begin()); |
8984 | llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; |
8985 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4, Tys), |
8986 | Ops, ""); |
8987 | } |
8988 | case NEON::BI__builtin_neon_vst4_lane_v: |
8989 | case NEON::BI__builtin_neon_vst4q_lane_v: { |
8990 | Ops.push_back(Ops[0]); |
8991 | Ops.erase(Ops.begin()); |
8992 | Ops[4] = Builder.CreateZExt(Ops[4], Int64Ty); |
8993 | llvm::Type *Tys[2] = { VTy, Ops[5]->getType() }; |
8994 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4lane, Tys), |
8995 | Ops, ""); |
8996 | } |
8997 | case NEON::BI__builtin_neon_vtrn_v: |
8998 | case NEON::BI__builtin_neon_vtrnq_v: { |
8999 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); |
9000 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
9001 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
9002 | Value *SV = nullptr; |
9003 | |
9004 | for (unsigned vi = 0; vi != 2; ++vi) { |
9005 | SmallVector<uint32_t, 16> Indices; |
9006 | for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { |
9007 | Indices.push_back(i+vi); |
9008 | Indices.push_back(i+e+vi); |
9009 | } |
9010 | Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); |
9011 | SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vtrn"); |
9012 | SV = Builder.CreateDefaultAlignedStore(SV, Addr); |
9013 | } |
9014 | return SV; |
9015 | } |
9016 | case NEON::BI__builtin_neon_vuzp_v: |
9017 | case NEON::BI__builtin_neon_vuzpq_v: { |
9018 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); |
9019 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
9020 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
9021 | Value *SV = nullptr; |
9022 | |
9023 | for (unsigned vi = 0; vi != 2; ++vi) { |
9024 | SmallVector<uint32_t, 16> Indices; |
9025 | for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) |
9026 | Indices.push_back(2*i+vi); |
9027 | |
9028 | Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); |
9029 | SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vuzp"); |
9030 | SV = Builder.CreateDefaultAlignedStore(SV, Addr); |
9031 | } |
9032 | return SV; |
9033 | } |
9034 | case NEON::BI__builtin_neon_vzip_v: |
9035 | case NEON::BI__builtin_neon_vzipq_v: { |
9036 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); |
9037 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
9038 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
9039 | Value *SV = nullptr; |
9040 | |
9041 | for (unsigned vi = 0; vi != 2; ++vi) { |
9042 | SmallVector<uint32_t, 16> Indices; |
9043 | for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { |
9044 | Indices.push_back((i + vi*e) >> 1); |
9045 | Indices.push_back(((i + vi*e) >> 1)+e); |
9046 | } |
9047 | Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); |
9048 | SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vzip"); |
9049 | SV = Builder.CreateDefaultAlignedStore(SV, Addr); |
9050 | } |
9051 | return SV; |
9052 | } |
9053 | case NEON::BI__builtin_neon_vqtbl1q_v: { |
9054 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl1, Ty), |
9055 | Ops, "vtbl1"); |
9056 | } |
9057 | case NEON::BI__builtin_neon_vqtbl2q_v: { |
9058 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl2, Ty), |
9059 | Ops, "vtbl2"); |
9060 | } |
9061 | case NEON::BI__builtin_neon_vqtbl3q_v: { |
9062 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl3, Ty), |
9063 | Ops, "vtbl3"); |
9064 | } |
9065 | case NEON::BI__builtin_neon_vqtbl4q_v: { |
9066 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl4, Ty), |
9067 | Ops, "vtbl4"); |
9068 | } |
9069 | case NEON::BI__builtin_neon_vqtbx1q_v: { |
9070 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx1, Ty), |
9071 | Ops, "vtbx1"); |
9072 | } |
9073 | case NEON::BI__builtin_neon_vqtbx2q_v: { |
9074 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx2, Ty), |
9075 | Ops, "vtbx2"); |
9076 | } |
9077 | case NEON::BI__builtin_neon_vqtbx3q_v: { |
9078 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx3, Ty), |
9079 | Ops, "vtbx3"); |
9080 | } |
9081 | case NEON::BI__builtin_neon_vqtbx4q_v: { |
9082 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx4, Ty), |
9083 | Ops, "vtbx4"); |
9084 | } |
9085 | case NEON::BI__builtin_neon_vsqadd_v: |
9086 | case NEON::BI__builtin_neon_vsqaddq_v: { |
9087 | Int = Intrinsic::aarch64_neon_usqadd; |
9088 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqadd"); |
9089 | } |
9090 | case NEON::BI__builtin_neon_vuqadd_v: |
9091 | case NEON::BI__builtin_neon_vuqaddq_v: { |
9092 | Int = Intrinsic::aarch64_neon_suqadd; |
9093 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vuqadd"); |
9094 | } |
9095 | case AArch64::BI_BitScanForward: |
9096 | case AArch64::BI_BitScanForward64: |
9097 | return EmitMSVCBuiltinExpr(MSVCIntrin::_BitScanForward, E); |
9098 | case AArch64::BI_BitScanReverse: |
9099 | case AArch64::BI_BitScanReverse64: |
9100 | return EmitMSVCBuiltinExpr(MSVCIntrin::_BitScanReverse, E); |
9101 | case AArch64::BI_InterlockedAnd64: |
9102 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedAnd, E); |
9103 | case AArch64::BI_InterlockedExchange64: |
9104 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchange, E); |
9105 | case AArch64::BI_InterlockedExchangeAdd64: |
9106 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeAdd, E); |
9107 | case AArch64::BI_InterlockedExchangeSub64: |
9108 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeSub, E); |
9109 | case AArch64::BI_InterlockedOr64: |
9110 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedOr, E); |
9111 | case AArch64::BI_InterlockedXor64: |
9112 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedXor, E); |
9113 | case AArch64::BI_InterlockedDecrement64: |
9114 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedDecrement, E); |
9115 | case AArch64::BI_InterlockedIncrement64: |
9116 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedIncrement, E); |
9117 | case AArch64::BI_InterlockedExchangeAdd8_acq: |
9118 | case AArch64::BI_InterlockedExchangeAdd16_acq: |
9119 | case AArch64::BI_InterlockedExchangeAdd_acq: |
9120 | case AArch64::BI_InterlockedExchangeAdd64_acq: |
9121 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeAdd_acq, E); |
9122 | case AArch64::BI_InterlockedExchangeAdd8_rel: |
9123 | case AArch64::BI_InterlockedExchangeAdd16_rel: |
9124 | case AArch64::BI_InterlockedExchangeAdd_rel: |
9125 | case AArch64::BI_InterlockedExchangeAdd64_rel: |
9126 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeAdd_rel, E); |
9127 | case AArch64::BI_InterlockedExchangeAdd8_nf: |
9128 | case AArch64::BI_InterlockedExchangeAdd16_nf: |
9129 | case AArch64::BI_InterlockedExchangeAdd_nf: |
9130 | case AArch64::BI_InterlockedExchangeAdd64_nf: |
9131 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeAdd_nf, E); |
9132 | case AArch64::BI_InterlockedExchange8_acq: |
9133 | case AArch64::BI_InterlockedExchange16_acq: |
9134 | case AArch64::BI_InterlockedExchange_acq: |
9135 | case AArch64::BI_InterlockedExchange64_acq: |
9136 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchange_acq, E); |
9137 | case AArch64::BI_InterlockedExchange8_rel: |
9138 | case AArch64::BI_InterlockedExchange16_rel: |
9139 | case AArch64::BI_InterlockedExchange_rel: |
9140 | case AArch64::BI_InterlockedExchange64_rel: |
9141 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchange_rel, E); |
9142 | case AArch64::BI_InterlockedExchange8_nf: |
9143 | case AArch64::BI_InterlockedExchange16_nf: |
9144 | case AArch64::BI_InterlockedExchange_nf: |
9145 | case AArch64::BI_InterlockedExchange64_nf: |
9146 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchange_nf, E); |
9147 | case AArch64::BI_InterlockedCompareExchange8_acq: |
9148 | case AArch64::BI_InterlockedCompareExchange16_acq: |
9149 | case AArch64::BI_InterlockedCompareExchange_acq: |
9150 | case AArch64::BI_InterlockedCompareExchange64_acq: |
9151 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedCompareExchange_acq, E); |
9152 | case AArch64::BI_InterlockedCompareExchange8_rel: |
9153 | case AArch64::BI_InterlockedCompareExchange16_rel: |
9154 | case AArch64::BI_InterlockedCompareExchange_rel: |
9155 | case AArch64::BI_InterlockedCompareExchange64_rel: |
9156 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedCompareExchange_rel, E); |
9157 | case AArch64::BI_InterlockedCompareExchange8_nf: |
9158 | case AArch64::BI_InterlockedCompareExchange16_nf: |
9159 | case AArch64::BI_InterlockedCompareExchange_nf: |
9160 | case AArch64::BI_InterlockedCompareExchange64_nf: |
9161 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedCompareExchange_nf, E); |
9162 | case AArch64::BI_InterlockedOr8_acq: |
9163 | case AArch64::BI_InterlockedOr16_acq: |
9164 | case AArch64::BI_InterlockedOr_acq: |
9165 | case AArch64::BI_InterlockedOr64_acq: |
9166 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedOr_acq, E); |
9167 | case AArch64::BI_InterlockedOr8_rel: |
9168 | case AArch64::BI_InterlockedOr16_rel: |
9169 | case AArch64::BI_InterlockedOr_rel: |
9170 | case AArch64::BI_InterlockedOr64_rel: |
9171 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedOr_rel, E); |
9172 | case AArch64::BI_InterlockedOr8_nf: |
9173 | case AArch64::BI_InterlockedOr16_nf: |
9174 | case AArch64::BI_InterlockedOr_nf: |
9175 | case AArch64::BI_InterlockedOr64_nf: |
9176 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedOr_nf, E); |
9177 | case AArch64::BI_InterlockedXor8_acq: |
9178 | case AArch64::BI_InterlockedXor16_acq: |
9179 | case AArch64::BI_InterlockedXor_acq: |
9180 | case AArch64::BI_InterlockedXor64_acq: |
9181 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedXor_acq, E); |
9182 | case AArch64::BI_InterlockedXor8_rel: |
9183 | case AArch64::BI_InterlockedXor16_rel: |
9184 | case AArch64::BI_InterlockedXor_rel: |
9185 | case AArch64::BI_InterlockedXor64_rel: |
9186 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedXor_rel, E); |
9187 | case AArch64::BI_InterlockedXor8_nf: |
9188 | case AArch64::BI_InterlockedXor16_nf: |
9189 | case AArch64::BI_InterlockedXor_nf: |
9190 | case AArch64::BI_InterlockedXor64_nf: |
9191 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedXor_nf, E); |
9192 | case AArch64::BI_InterlockedAnd8_acq: |
9193 | case AArch64::BI_InterlockedAnd16_acq: |
9194 | case AArch64::BI_InterlockedAnd_acq: |
9195 | case AArch64::BI_InterlockedAnd64_acq: |
9196 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedAnd_acq, E); |
9197 | case AArch64::BI_InterlockedAnd8_rel: |
9198 | case AArch64::BI_InterlockedAnd16_rel: |
9199 | case AArch64::BI_InterlockedAnd_rel: |
9200 | case AArch64::BI_InterlockedAnd64_rel: |
9201 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedAnd_rel, E); |
9202 | case AArch64::BI_InterlockedAnd8_nf: |
9203 | case AArch64::BI_InterlockedAnd16_nf: |
9204 | case AArch64::BI_InterlockedAnd_nf: |
9205 | case AArch64::BI_InterlockedAnd64_nf: |
9206 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedAnd_nf, E); |
9207 | case AArch64::BI_InterlockedIncrement16_acq: |
9208 | case AArch64::BI_InterlockedIncrement_acq: |
9209 | case AArch64::BI_InterlockedIncrement64_acq: |
9210 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedIncrement_acq, E); |
9211 | case AArch64::BI_InterlockedIncrement16_rel: |
9212 | case AArch64::BI_InterlockedIncrement_rel: |
9213 | case AArch64::BI_InterlockedIncrement64_rel: |
9214 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedIncrement_rel, E); |
9215 | case AArch64::BI_InterlockedIncrement16_nf: |
9216 | case AArch64::BI_InterlockedIncrement_nf: |
9217 | case AArch64::BI_InterlockedIncrement64_nf: |
9218 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedIncrement_nf, E); |
9219 | case AArch64::BI_InterlockedDecrement16_acq: |
9220 | case AArch64::BI_InterlockedDecrement_acq: |
9221 | case AArch64::BI_InterlockedDecrement64_acq: |
9222 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedDecrement_acq, E); |
9223 | case AArch64::BI_InterlockedDecrement16_rel: |
9224 | case AArch64::BI_InterlockedDecrement_rel: |
9225 | case AArch64::BI_InterlockedDecrement64_rel: |
9226 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedDecrement_rel, E); |
9227 | case AArch64::BI_InterlockedDecrement16_nf: |
9228 | case AArch64::BI_InterlockedDecrement_nf: |
9229 | case AArch64::BI_InterlockedDecrement64_nf: |
9230 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedDecrement_nf, E); |
9231 | |
9232 | case AArch64::BI_InterlockedAdd: { |
9233 | Value *Arg0 = EmitScalarExpr(E->getArg(0)); |
9234 | Value *Arg1 = EmitScalarExpr(E->getArg(1)); |
9235 | AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( |
9236 | AtomicRMWInst::Add, Arg0, Arg1, |
9237 | llvm::AtomicOrdering::SequentiallyConsistent); |
9238 | return Builder.CreateAdd(RMWI, Arg1); |
9239 | } |
9240 | } |
9241 | } |
9242 | |
9243 | llvm::Value *CodeGenFunction:: |
9244 | BuildVector(ArrayRef<llvm::Value*> Ops) { |
9245 | assert((Ops.size() & (Ops.size() - 1)) == 0 &&(((Ops.size() & (Ops.size() - 1)) == 0 && "Not a power-of-two sized vector!" ) ? static_cast<void> (0) : __assert_fail ("(Ops.size() & (Ops.size() - 1)) == 0 && \"Not a power-of-two sized vector!\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 9246, __PRETTY_FUNCTION__)) |
9246 | "Not a power-of-two sized vector!")(((Ops.size() & (Ops.size() - 1)) == 0 && "Not a power-of-two sized vector!" ) ? static_cast<void> (0) : __assert_fail ("(Ops.size() & (Ops.size() - 1)) == 0 && \"Not a power-of-two sized vector!\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 9246, __PRETTY_FUNCTION__)); |
9247 | bool AllConstants = true; |
9248 | for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i) |
9249 | AllConstants &= isa<Constant>(Ops[i]); |
9250 | |
9251 | // If this is a constant vector, create a ConstantVector. |
9252 | if (AllConstants) { |
9253 | SmallVector<llvm::Constant*, 16> CstOps; |
9254 | for (unsigned i = 0, e = Ops.size(); i != e; ++i) |
9255 | CstOps.push_back(cast<Constant>(Ops[i])); |
9256 | return llvm::ConstantVector::get(CstOps); |
9257 | } |
9258 | |
9259 | // Otherwise, insertelement the values to build the vector. |
9260 | Value *Result = |
9261 | llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size())); |
9262 | |
9263 | for (unsigned i = 0, e = Ops.size(); i != e; ++i) |
9264 | Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt32(i)); |
9265 | |
9266 | return Result; |
9267 | } |
9268 | |
9269 | // Convert the mask from an integer type to a vector of i1. |
9270 | static Value *getMaskVecValue(CodeGenFunction &CGF, Value *Mask, |
9271 | unsigned NumElts) { |
9272 | |
9273 | llvm::VectorType *MaskTy = llvm::VectorType::get(CGF.Builder.getInt1Ty(), |
9274 | cast<IntegerType>(Mask->getType())->getBitWidth()); |
9275 | Value *MaskVec = CGF.Builder.CreateBitCast(Mask, MaskTy); |
9276 | |
9277 | // If we have less than 8 elements, then the starting mask was an i8 and |
9278 | // we need to extract down to the right number of elements. |
9279 | if (NumElts < 8) { |
9280 | uint32_t Indices[4]; |
9281 | for (unsigned i = 0; i != NumElts; ++i) |
9282 | Indices[i] = i; |
9283 | MaskVec = CGF.Builder.CreateShuffleVector(MaskVec, MaskVec, |
9284 | makeArrayRef(Indices, NumElts), |
9285 | "extract"); |
9286 | } |
9287 | return MaskVec; |
9288 | } |
9289 | |
9290 | static Value *EmitX86MaskedStore(CodeGenFunction &CGF, |
9291 | ArrayRef<Value *> Ops, |
9292 | unsigned Align) { |
9293 | // Cast the pointer to right type. |
9294 | Value *Ptr = CGF.Builder.CreateBitCast(Ops[0], |
9295 | llvm::PointerType::getUnqual(Ops[1]->getType())); |
9296 | |
9297 | Value *MaskVec = getMaskVecValue(CGF, Ops[2], |
9298 | Ops[1]->getType()->getVectorNumElements()); |
9299 | |
9300 | return CGF.Builder.CreateMaskedStore(Ops[1], Ptr, Align, MaskVec); |
9301 | } |
9302 | |
9303 | static Value *EmitX86MaskedLoad(CodeGenFunction &CGF, |
9304 | ArrayRef<Value *> Ops, unsigned Align) { |
9305 | // Cast the pointer to right type. |
9306 | Value *Ptr = CGF.Builder.CreateBitCast(Ops[0], |
9307 | llvm::PointerType::getUnqual(Ops[1]->getType())); |
9308 | |
9309 | Value *MaskVec = getMaskVecValue(CGF, Ops[2], |
9310 | Ops[1]->getType()->getVectorNumElements()); |
9311 | |
9312 | return CGF.Builder.CreateMaskedLoad(Ptr, Align, MaskVec, Ops[1]); |
9313 | } |
9314 | |
9315 | static Value *EmitX86ExpandLoad(CodeGenFunction &CGF, |
9316 | ArrayRef<Value *> Ops) { |
9317 | llvm::Type *ResultTy = Ops[1]->getType(); |
9318 | llvm::Type *PtrTy = ResultTy->getVectorElementType(); |
9319 | |
9320 | // Cast the pointer to element type. |
9321 | Value *Ptr = CGF.Builder.CreateBitCast(Ops[0], |
9322 | llvm::PointerType::getUnqual(PtrTy)); |
9323 | |
9324 | Value *MaskVec = getMaskVecValue(CGF, Ops[2], |
9325 | ResultTy->getVectorNumElements()); |
9326 | |
9327 | llvm::Function *F = CGF.CGM.getIntrinsic(Intrinsic::masked_expandload, |
9328 | ResultTy); |
9329 | return CGF.Builder.CreateCall(F, { Ptr, MaskVec, Ops[1] }); |
9330 | } |
9331 | |
9332 | static Value *EmitX86CompressExpand(CodeGenFunction &CGF, |
9333 | ArrayRef<Value *> Ops, |
9334 | bool IsCompress) { |
9335 | llvm::Type *ResultTy = Ops[1]->getType(); |
9336 | |
9337 | Value *MaskVec = getMaskVecValue(CGF, Ops[2], |
9338 | ResultTy->getVectorNumElements()); |
9339 | |
9340 | Intrinsic::ID IID = IsCompress ? Intrinsic::x86_avx512_mask_compress |
9341 | : Intrinsic::x86_avx512_mask_expand; |
9342 | llvm::Function *F = CGF.CGM.getIntrinsic(IID, ResultTy); |
9343 | return CGF.Builder.CreateCall(F, { Ops[0], Ops[1], MaskVec }); |
9344 | } |
9345 | |
9346 | static Value *EmitX86CompressStore(CodeGenFunction &CGF, |
9347 | ArrayRef<Value *> Ops) { |
9348 | llvm::Type *ResultTy = Ops[1]->getType(); |
9349 | llvm::Type *PtrTy = ResultTy->getVectorElementType(); |
9350 | |
9351 | // Cast the pointer to element type. |
9352 | Value *Ptr = CGF.Builder.CreateBitCast(Ops[0], |
9353 | llvm::PointerType::getUnqual(PtrTy)); |
9354 | |
9355 | Value *MaskVec = getMaskVecValue(CGF, Ops[2], |
9356 | ResultTy->getVectorNumElements()); |
9357 | |
9358 | llvm::Function *F = CGF.CGM.getIntrinsic(Intrinsic::masked_compressstore, |
9359 | ResultTy); |
9360 | return CGF.Builder.CreateCall(F, { Ops[1], Ptr, MaskVec }); |
9361 | } |
9362 | |
9363 | static Value *EmitX86MaskLogic(CodeGenFunction &CGF, Instruction::BinaryOps Opc, |
9364 | ArrayRef<Value *> Ops, |
9365 | bool InvertLHS = false) { |
9366 | unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth(); |
9367 | Value *LHS = getMaskVecValue(CGF, Ops[0], NumElts); |
9368 | Value *RHS = getMaskVecValue(CGF, Ops[1], NumElts); |
9369 | |
9370 | if (InvertLHS) |
9371 | LHS = CGF.Builder.CreateNot(LHS); |
9372 | |
9373 | return CGF.Builder.CreateBitCast(CGF.Builder.CreateBinOp(Opc, LHS, RHS), |
9374 | Ops[0]->getType()); |
9375 | } |
9376 | |
9377 | static Value *EmitX86FunnelShift(CodeGenFunction &CGF, Value *Op0, Value *Op1, |
9378 | Value *Amt, bool IsRight) { |
9379 | llvm::Type *Ty = Op0->getType(); |
9380 | |
9381 | // Amount may be scalar immediate, in which case create a splat vector. |
9382 | // Funnel shifts amounts are treated as modulo and types are all power-of-2 so |
9383 | // we only care about the lowest log2 bits anyway. |
9384 | if (Amt->getType() != Ty) { |
9385 | unsigned NumElts = Ty->getVectorNumElements(); |
9386 | Amt = CGF.Builder.CreateIntCast(Amt, Ty->getScalarType(), false); |
9387 | Amt = CGF.Builder.CreateVectorSplat(NumElts, Amt); |
9388 | } |
9389 | |
9390 | unsigned IID = IsRight ? Intrinsic::fshr : Intrinsic::fshl; |
9391 | Function *F = CGF.CGM.getIntrinsic(IID, Ty); |
9392 | return CGF.Builder.CreateCall(F, {Op0, Op1, Amt}); |
9393 | } |
9394 | |
9395 | static Value *EmitX86vpcom(CodeGenFunction &CGF, ArrayRef<Value *> Ops, |
9396 | bool IsSigned) { |
9397 | Value *Op0 = Ops[0]; |
9398 | Value *Op1 = Ops[1]; |
9399 | llvm::Type *Ty = Op0->getType(); |
9400 | uint64_t Imm = cast<llvm::ConstantInt>(Ops[2])->getZExtValue() & 0x7; |
9401 | |
9402 | CmpInst::Predicate Pred; |
9403 | switch (Imm) { |
9404 | case 0x0: |
9405 | Pred = IsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; |
9406 | break; |
9407 | case 0x1: |
9408 | Pred = IsSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; |
9409 | break; |
9410 | case 0x2: |
9411 | Pred = IsSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; |
9412 | break; |
9413 | case 0x3: |
9414 | Pred = IsSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; |
9415 | break; |
9416 | case 0x4: |
9417 | Pred = ICmpInst::ICMP_EQ; |
9418 | break; |
9419 | case 0x5: |
9420 | Pred = ICmpInst::ICMP_NE; |
9421 | break; |
9422 | case 0x6: |
9423 | return llvm::Constant::getNullValue(Ty); // FALSE |
9424 | case 0x7: |
9425 | return llvm::Constant::getAllOnesValue(Ty); // TRUE |
9426 | default: |
9427 | llvm_unreachable("Unexpected XOP vpcom/vpcomu predicate")::llvm::llvm_unreachable_internal("Unexpected XOP vpcom/vpcomu predicate" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 9427); |
9428 | } |
9429 | |
9430 | Value *Cmp = CGF.Builder.CreateICmp(Pred, Op0, Op1); |
9431 | Value *Res = CGF.Builder.CreateSExt(Cmp, Ty); |
9432 | return Res; |
9433 | } |
9434 | |
9435 | static Value *EmitX86Select(CodeGenFunction &CGF, |
9436 | Value *Mask, Value *Op0, Value *Op1) { |
9437 | |
9438 | // If the mask is all ones just return first argument. |
9439 | if (const auto *C = dyn_cast<Constant>(Mask)) |
9440 | if (C->isAllOnesValue()) |
9441 | return Op0; |
9442 | |
9443 | Mask = getMaskVecValue(CGF, Mask, Op0->getType()->getVectorNumElements()); |
9444 | |
9445 | return CGF.Builder.CreateSelect(Mask, Op0, Op1); |
9446 | } |
9447 | |
9448 | static Value *EmitX86ScalarSelect(CodeGenFunction &CGF, |
9449 | Value *Mask, Value *Op0, Value *Op1) { |
9450 | // If the mask is all ones just return first argument. |
9451 | if (const auto *C = dyn_cast<Constant>(Mask)) |
9452 | if (C->isAllOnesValue()) |
9453 | return Op0; |
9454 | |
9455 | llvm::VectorType *MaskTy = |
9456 | llvm::VectorType::get(CGF.Builder.getInt1Ty(), |
9457 | Mask->getType()->getIntegerBitWidth()); |
9458 | Mask = CGF.Builder.CreateBitCast(Mask, MaskTy); |
9459 | Mask = CGF.Builder.CreateExtractElement(Mask, (uint64_t)0); |
9460 | return CGF.Builder.CreateSelect(Mask, Op0, Op1); |
9461 | } |
9462 | |
9463 | static Value *EmitX86MaskedCompareResult(CodeGenFunction &CGF, Value *Cmp, |
9464 | unsigned NumElts, Value *MaskIn) { |
9465 | if (MaskIn) { |
9466 | const auto *C = dyn_cast<Constant>(MaskIn); |
9467 | if (!C || !C->isAllOnesValue()) |
9468 | Cmp = CGF.Builder.CreateAnd(Cmp, getMaskVecValue(CGF, MaskIn, NumElts)); |
9469 | } |
9470 | |
9471 | if (NumElts < 8) { |
9472 | uint32_t Indices[8]; |
9473 | for (unsigned i = 0; i != NumElts; ++i) |
9474 | Indices[i] = i; |
9475 | for (unsigned i = NumElts; i != 8; ++i) |
9476 | Indices[i] = i % NumElts + NumElts; |
9477 | Cmp = CGF.Builder.CreateShuffleVector( |
9478 | Cmp, llvm::Constant::getNullValue(Cmp->getType()), Indices); |
9479 | } |
9480 | |
9481 | return CGF.Builder.CreateBitCast(Cmp, |
9482 | IntegerType::get(CGF.getLLVMContext(), |
9483 | std::max(NumElts, 8U))); |
9484 | } |
9485 | |
9486 | static Value *EmitX86MaskedCompare(CodeGenFunction &CGF, unsigned CC, |
9487 | bool Signed, ArrayRef<Value *> Ops) { |
9488 | assert((Ops.size() == 2 || Ops.size() == 4) &&(((Ops.size() == 2 || Ops.size() == 4) && "Unexpected number of arguments" ) ? static_cast<void> (0) : __assert_fail ("(Ops.size() == 2 || Ops.size() == 4) && \"Unexpected number of arguments\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 9489, __PRETTY_FUNCTION__)) |
9489 | "Unexpected number of arguments")(((Ops.size() == 2 || Ops.size() == 4) && "Unexpected number of arguments" ) ? static_cast<void> (0) : __assert_fail ("(Ops.size() == 2 || Ops.size() == 4) && \"Unexpected number of arguments\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 9489, __PRETTY_FUNCTION__)); |
9490 | unsigned NumElts = Ops[0]->getType()->getVectorNumElements(); |
9491 | Value *Cmp; |
9492 | |
9493 | if (CC == 3) { |
9494 | Cmp = Constant::getNullValue( |
9495 | llvm::VectorType::get(CGF.Builder.getInt1Ty(), NumElts)); |
9496 | } else if (CC == 7) { |
9497 | Cmp = Constant::getAllOnesValue( |
9498 | llvm::VectorType::get(CGF.Builder.getInt1Ty(), NumElts)); |
9499 | } else { |
9500 | ICmpInst::Predicate Pred; |
9501 | switch (CC) { |
9502 | default: llvm_unreachable("Unknown condition code")::llvm::llvm_unreachable_internal("Unknown condition code", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 9502); |
9503 | case 0: Pred = ICmpInst::ICMP_EQ; break; |
9504 | case 1: Pred = Signed ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; break; |
9505 | case 2: Pred = Signed ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; break; |
9506 | case 4: Pred = ICmpInst::ICMP_NE; break; |
9507 | case 5: Pred = Signed ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; break; |
9508 | case 6: Pred = Signed ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; break; |
9509 | } |
9510 | Cmp = CGF.Builder.CreateICmp(Pred, Ops[0], Ops[1]); |
9511 | } |
9512 | |
9513 | Value *MaskIn = nullptr; |
9514 | if (Ops.size() == 4) |
9515 | MaskIn = Ops[3]; |
9516 | |
9517 | return EmitX86MaskedCompareResult(CGF, Cmp, NumElts, MaskIn); |
9518 | } |
9519 | |
9520 | static Value *EmitX86ConvertToMask(CodeGenFunction &CGF, Value *In) { |
9521 | Value *Zero = Constant::getNullValue(In->getType()); |
9522 | return EmitX86MaskedCompare(CGF, 1, true, { In, Zero }); |
9523 | } |
9524 | |
9525 | static Value *EmitX86ConvertIntToFp(CodeGenFunction &CGF, |
9526 | ArrayRef<Value *> Ops, bool IsSigned) { |
9527 | unsigned Rnd = cast<llvm::ConstantInt>(Ops[3])->getZExtValue(); |
9528 | llvm::Type *Ty = Ops[1]->getType(); |
9529 | |
9530 | Value *Res; |
9531 | if (Rnd != 4) { |
9532 | Intrinsic::ID IID = IsSigned ? Intrinsic::x86_avx512_sitofp_round |
9533 | : Intrinsic::x86_avx512_uitofp_round; |
9534 | Function *F = CGF.CGM.getIntrinsic(IID, { Ty, Ops[0]->getType() }); |
9535 | Res = CGF.Builder.CreateCall(F, { Ops[0], Ops[3] }); |
9536 | } else { |
9537 | Res = IsSigned ? CGF.Builder.CreateSIToFP(Ops[0], Ty) |
9538 | : CGF.Builder.CreateUIToFP(Ops[0], Ty); |
9539 | } |
9540 | |
9541 | return EmitX86Select(CGF, Ops[2], Res, Ops[1]); |
9542 | } |
9543 | |
9544 | static Value *EmitX86Abs(CodeGenFunction &CGF, ArrayRef<Value *> Ops) { |
9545 | |
9546 | llvm::Type *Ty = Ops[0]->getType(); |
9547 | Value *Zero = llvm::Constant::getNullValue(Ty); |
9548 | Value *Sub = CGF.Builder.CreateSub(Zero, Ops[0]); |
9549 | Value *Cmp = CGF.Builder.CreateICmp(ICmpInst::ICMP_SGT, Ops[0], Zero); |
9550 | Value *Res = CGF.Builder.CreateSelect(Cmp, Ops[0], Sub); |
9551 | return Res; |
9552 | } |
9553 | |
9554 | static Value *EmitX86MinMax(CodeGenFunction &CGF, ICmpInst::Predicate Pred, |
9555 | ArrayRef<Value *> Ops) { |
9556 | Value *Cmp = CGF.Builder.CreateICmp(Pred, Ops[0], Ops[1]); |
9557 | Value *Res = CGF.Builder.CreateSelect(Cmp, Ops[0], Ops[1]); |
9558 | |
9559 | assert(Ops.size() == 2)((Ops.size() == 2) ? static_cast<void> (0) : __assert_fail ("Ops.size() == 2", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 9559, __PRETTY_FUNCTION__)); |
9560 | return Res; |
9561 | } |
9562 | |
9563 | // Lowers X86 FMA intrinsics to IR. |
9564 | static Value *EmitX86FMAExpr(CodeGenFunction &CGF, ArrayRef<Value *> Ops, |
9565 | unsigned BuiltinID, bool IsAddSub) { |
9566 | |
9567 | bool Subtract = false; |
9568 | Intrinsic::ID IID = Intrinsic::not_intrinsic; |
9569 | switch (BuiltinID) { |
9570 | default: break; |
9571 | case clang::X86::BI__builtin_ia32_vfmsubps512_mask3: |
9572 | Subtract = true; |
9573 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
9574 | case clang::X86::BI__builtin_ia32_vfmaddps512_mask: |
9575 | case clang::X86::BI__builtin_ia32_vfmaddps512_maskz: |
9576 | case clang::X86::BI__builtin_ia32_vfmaddps512_mask3: |
9577 | IID = llvm::Intrinsic::x86_avx512_vfmadd_ps_512; break; |
9578 | case clang::X86::BI__builtin_ia32_vfmsubpd512_mask3: |
9579 | Subtract = true; |
9580 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
9581 | case clang::X86::BI__builtin_ia32_vfmaddpd512_mask: |
9582 | case clang::X86::BI__builtin_ia32_vfmaddpd512_maskz: |
9583 | case clang::X86::BI__builtin_ia32_vfmaddpd512_mask3: |
9584 | IID = llvm::Intrinsic::x86_avx512_vfmadd_pd_512; break; |
9585 | case clang::X86::BI__builtin_ia32_vfmsubaddps512_mask3: |
9586 | Subtract = true; |
9587 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
9588 | case clang::X86::BI__builtin_ia32_vfmaddsubps512_mask: |
9589 | case clang::X86::BI__builtin_ia32_vfmaddsubps512_maskz: |
9590 | case clang::X86::BI__builtin_ia32_vfmaddsubps512_mask3: |
9591 | IID = llvm::Intrinsic::x86_avx512_vfmaddsub_ps_512; |
9592 | break; |
9593 | case clang::X86::BI__builtin_ia32_vfmsubaddpd512_mask3: |
9594 | Subtract = true; |
9595 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
9596 | case clang::X86::BI__builtin_ia32_vfmaddsubpd512_mask: |
9597 | case clang::X86::BI__builtin_ia32_vfmaddsubpd512_maskz: |
9598 | case clang::X86::BI__builtin_ia32_vfmaddsubpd512_mask3: |
9599 | IID = llvm::Intrinsic::x86_avx512_vfmaddsub_pd_512; |
9600 | break; |
9601 | } |
9602 | |
9603 | Value *A = Ops[0]; |
9604 | Value *B = Ops[1]; |
9605 | Value *C = Ops[2]; |
9606 | |
9607 | if (Subtract) |
9608 | C = CGF.Builder.CreateFNeg(C); |
9609 | |
9610 | Value *Res; |
9611 | |
9612 | // Only handle in case of _MM_FROUND_CUR_DIRECTION/4 (no rounding). |
9613 | if (IID != Intrinsic::not_intrinsic && |
9614 | cast<llvm::ConstantInt>(Ops.back())->getZExtValue() != (uint64_t)4) { |
9615 | Function *Intr = CGF.CGM.getIntrinsic(IID); |
9616 | Res = CGF.Builder.CreateCall(Intr, {A, B, C, Ops.back() }); |
9617 | } else { |
9618 | llvm::Type *Ty = A->getType(); |
9619 | Function *FMA = CGF.CGM.getIntrinsic(Intrinsic::fma, Ty); |
9620 | Res = CGF.Builder.CreateCall(FMA, {A, B, C} ); |
9621 | |
9622 | if (IsAddSub) { |
9623 | // Negate even elts in C using a mask. |
9624 | unsigned NumElts = Ty->getVectorNumElements(); |
9625 | SmallVector<uint32_t, 16> Indices(NumElts); |
9626 | for (unsigned i = 0; i != NumElts; ++i) |
9627 | Indices[i] = i + (i % 2) * NumElts; |
9628 | |
9629 | Value *NegC = CGF.Builder.CreateFNeg(C); |
9630 | Value *FMSub = CGF.Builder.CreateCall(FMA, {A, B, NegC} ); |
9631 | Res = CGF.Builder.CreateShuffleVector(FMSub, Res, Indices); |
9632 | } |
9633 | } |
9634 | |
9635 | // Handle any required masking. |
9636 | Value *MaskFalseVal = nullptr; |
9637 | switch (BuiltinID) { |
9638 | case clang::X86::BI__builtin_ia32_vfmaddps512_mask: |
9639 | case clang::X86::BI__builtin_ia32_vfmaddpd512_mask: |
9640 | case clang::X86::BI__builtin_ia32_vfmaddsubps512_mask: |
9641 | case clang::X86::BI__builtin_ia32_vfmaddsubpd512_mask: |
9642 | MaskFalseVal = Ops[0]; |
9643 | break; |
9644 | case clang::X86::BI__builtin_ia32_vfmaddps512_maskz: |
9645 | case clang::X86::BI__builtin_ia32_vfmaddpd512_maskz: |
9646 | case clang::X86::BI__builtin_ia32_vfmaddsubps512_maskz: |
9647 | case clang::X86::BI__builtin_ia32_vfmaddsubpd512_maskz: |
9648 | MaskFalseVal = Constant::getNullValue(Ops[0]->getType()); |
9649 | break; |
9650 | case clang::X86::BI__builtin_ia32_vfmsubps512_mask3: |
9651 | case clang::X86::BI__builtin_ia32_vfmaddps512_mask3: |
9652 | case clang::X86::BI__builtin_ia32_vfmsubpd512_mask3: |
9653 | case clang::X86::BI__builtin_ia32_vfmaddpd512_mask3: |
9654 | case clang::X86::BI__builtin_ia32_vfmsubaddps512_mask3: |
9655 | case clang::X86::BI__builtin_ia32_vfmaddsubps512_mask3: |
9656 | case clang::X86::BI__builtin_ia32_vfmsubaddpd512_mask3: |
9657 | case clang::X86::BI__builtin_ia32_vfmaddsubpd512_mask3: |
9658 | MaskFalseVal = Ops[2]; |
9659 | break; |
9660 | } |
9661 | |
9662 | if (MaskFalseVal) |
9663 | return EmitX86Select(CGF, Ops[3], Res, MaskFalseVal); |
9664 | |
9665 | return Res; |
9666 | } |
9667 | |
9668 | static Value * |
9669 | EmitScalarFMAExpr(CodeGenFunction &CGF, MutableArrayRef<Value *> Ops, |
9670 | Value *Upper, bool ZeroMask = false, unsigned PTIdx = 0, |
9671 | bool NegAcc = false) { |
9672 | unsigned Rnd = 4; |
9673 | if (Ops.size() > 4) |
9674 | Rnd = cast<llvm::ConstantInt>(Ops[4])->getZExtValue(); |
9675 | |
9676 | if (NegAcc) |
9677 | Ops[2] = CGF.Builder.CreateFNeg(Ops[2]); |
9678 | |
9679 | Ops[0] = CGF.Builder.CreateExtractElement(Ops[0], (uint64_t)0); |
9680 | Ops[1] = CGF.Builder.CreateExtractElement(Ops[1], (uint64_t)0); |
9681 | Ops[2] = CGF.Builder.CreateExtractElement(Ops[2], (uint64_t)0); |
9682 | Value *Res; |
9683 | if (Rnd != 4) { |
9684 | Intrinsic::ID IID = Ops[0]->getType()->getPrimitiveSizeInBits() == 32 ? |
9685 | Intrinsic::x86_avx512_vfmadd_f32 : |
9686 | Intrinsic::x86_avx512_vfmadd_f64; |
9687 | Res = CGF.Builder.CreateCall(CGF.CGM.getIntrinsic(IID), |
9688 | {Ops[0], Ops[1], Ops[2], Ops[4]}); |
9689 | } else { |
9690 | Function *FMA = CGF.CGM.getIntrinsic(Intrinsic::fma, Ops[0]->getType()); |
9691 | Res = CGF.Builder.CreateCall(FMA, Ops.slice(0, 3)); |
9692 | } |
9693 | // If we have more than 3 arguments, we need to do masking. |
9694 | if (Ops.size() > 3) { |
9695 | Value *PassThru = ZeroMask ? Constant::getNullValue(Res->getType()) |
9696 | : Ops[PTIdx]; |
9697 | |
9698 | // If we negated the accumulator and the its the PassThru value we need to |
9699 | // bypass the negate. Conveniently Upper should be the same thing in this |
9700 | // case. |
9701 | if (NegAcc && PTIdx == 2) |
9702 | PassThru = CGF.Builder.CreateExtractElement(Upper, (uint64_t)0); |
9703 | |
9704 | Res = EmitX86ScalarSelect(CGF, Ops[3], Res, PassThru); |
9705 | } |
9706 | return CGF.Builder.CreateInsertElement(Upper, Res, (uint64_t)0); |
9707 | } |
9708 | |
9709 | static Value *EmitX86Muldq(CodeGenFunction &CGF, bool IsSigned, |
9710 | ArrayRef<Value *> Ops) { |
9711 | llvm::Type *Ty = Ops[0]->getType(); |
9712 | // Arguments have a vXi32 type so cast to vXi64. |
9713 | Ty = llvm::VectorType::get(CGF.Int64Ty, |
9714 | Ty->getPrimitiveSizeInBits() / 64); |
9715 | Value *LHS = CGF.Builder.CreateBitCast(Ops[0], Ty); |
9716 | Value *RHS = CGF.Builder.CreateBitCast(Ops[1], Ty); |
9717 | |
9718 | if (IsSigned) { |
9719 | // Shift left then arithmetic shift right. |
9720 | Constant *ShiftAmt = ConstantInt::get(Ty, 32); |
9721 | LHS = CGF.Builder.CreateShl(LHS, ShiftAmt); |
9722 | LHS = CGF.Builder.CreateAShr(LHS, ShiftAmt); |
9723 | RHS = CGF.Builder.CreateShl(RHS, ShiftAmt); |
9724 | RHS = CGF.Builder.CreateAShr(RHS, ShiftAmt); |
9725 | } else { |
9726 | // Clear the upper bits. |
9727 | Constant *Mask = ConstantInt::get(Ty, 0xffffffff); |
9728 | LHS = CGF.Builder.CreateAnd(LHS, Mask); |
9729 | RHS = CGF.Builder.CreateAnd(RHS, Mask); |
9730 | } |
9731 | |
9732 | return CGF.Builder.CreateMul(LHS, RHS); |
9733 | } |
9734 | |
9735 | // Emit a masked pternlog intrinsic. This only exists because the header has to |
9736 | // use a macro and we aren't able to pass the input argument to a pternlog |
9737 | // builtin and a select builtin without evaluating it twice. |
9738 | static Value *EmitX86Ternlog(CodeGenFunction &CGF, bool ZeroMask, |
9739 | ArrayRef<Value *> Ops) { |
9740 | llvm::Type *Ty = Ops[0]->getType(); |
9741 | |
9742 | unsigned VecWidth = Ty->getPrimitiveSizeInBits(); |
9743 | unsigned EltWidth = Ty->getScalarSizeInBits(); |
9744 | Intrinsic::ID IID; |
9745 | if (VecWidth == 128 && EltWidth == 32) |
9746 | IID = Intrinsic::x86_avx512_pternlog_d_128; |
9747 | else if (VecWidth == 256 && EltWidth == 32) |
9748 | IID = Intrinsic::x86_avx512_pternlog_d_256; |
9749 | else if (VecWidth == 512 && EltWidth == 32) |
9750 | IID = Intrinsic::x86_avx512_pternlog_d_512; |
9751 | else if (VecWidth == 128 && EltWidth == 64) |
9752 | IID = Intrinsic::x86_avx512_pternlog_q_128; |
9753 | else if (VecWidth == 256 && EltWidth == 64) |
9754 | IID = Intrinsic::x86_avx512_pternlog_q_256; |
9755 | else if (VecWidth == 512 && EltWidth == 64) |
9756 | IID = Intrinsic::x86_avx512_pternlog_q_512; |
9757 | else |
9758 | llvm_unreachable("Unexpected intrinsic")::llvm::llvm_unreachable_internal("Unexpected intrinsic", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 9758); |
9759 | |
9760 | Value *Ternlog = CGF.Builder.CreateCall(CGF.CGM.getIntrinsic(IID), |
9761 | Ops.drop_back()); |
9762 | Value *PassThru = ZeroMask ? ConstantAggregateZero::get(Ty) : Ops[0]; |
9763 | return EmitX86Select(CGF, Ops[4], Ternlog, PassThru); |
9764 | } |
9765 | |
9766 | static Value *EmitX86SExtMask(CodeGenFunction &CGF, Value *Op, |
9767 | llvm::Type *DstTy) { |
9768 | unsigned NumberOfElements = DstTy->getVectorNumElements(); |
9769 | Value *Mask = getMaskVecValue(CGF, Op, NumberOfElements); |
9770 | return CGF.Builder.CreateSExt(Mask, DstTy, "vpmovm2"); |
9771 | } |
9772 | |
9773 | // Emit addition or subtraction with signed/unsigned saturation. |
9774 | static Value *EmitX86AddSubSatExpr(CodeGenFunction &CGF, |
9775 | ArrayRef<Value *> Ops, bool IsSigned, |
9776 | bool IsAddition) { |
9777 | Intrinsic::ID IID = |
9778 | IsSigned ? (IsAddition ? Intrinsic::sadd_sat : Intrinsic::ssub_sat) |
9779 | : (IsAddition ? Intrinsic::uadd_sat : Intrinsic::usub_sat); |
9780 | llvm::Function *F = CGF.CGM.getIntrinsic(IID, Ops[0]->getType()); |
9781 | return CGF.Builder.CreateCall(F, {Ops[0], Ops[1]}); |
9782 | } |
9783 | |
9784 | Value *CodeGenFunction::EmitX86CpuIs(const CallExpr *E) { |
9785 | const Expr *CPUExpr = E->getArg(0)->IgnoreParenCasts(); |
9786 | StringRef CPUStr = cast<clang::StringLiteral>(CPUExpr)->getString(); |
9787 | return EmitX86CpuIs(CPUStr); |
9788 | } |
9789 | |
9790 | Value *CodeGenFunction::EmitX86CpuIs(StringRef CPUStr) { |
9791 | |
9792 | llvm::Type *Int32Ty = Builder.getInt32Ty(); |
9793 | |
9794 | // Matching the struct layout from the compiler-rt/libgcc structure that is |
9795 | // filled in: |
9796 | // unsigned int __cpu_vendor; |
9797 | // unsigned int __cpu_type; |
9798 | // unsigned int __cpu_subtype; |
9799 | // unsigned int __cpu_features[1]; |
9800 | llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, Int32Ty, |
9801 | llvm::ArrayType::get(Int32Ty, 1)); |
9802 | |
9803 | // Grab the global __cpu_model. |
9804 | llvm::Constant *CpuModel = CGM.CreateRuntimeVariable(STy, "__cpu_model"); |
9805 | cast<llvm::GlobalValue>(CpuModel)->setDSOLocal(true); |
9806 | |
9807 | // Calculate the index needed to access the correct field based on the |
9808 | // range. Also adjust the expected value. |
9809 | unsigned Index; |
9810 | unsigned Value; |
9811 | std::tie(Index, Value) = StringSwitch<std::pair<unsigned, unsigned>>(CPUStr) |
9812 | #define X86_VENDOR(ENUM, STRING) \ |
9813 | .Case(STRING, {0u, static_cast<unsigned>(llvm::X86::ENUM)}) |
9814 | #define X86_CPU_TYPE_COMPAT_WITH_ALIAS(ARCHNAME, ENUM, STR, ALIAS) \ |
9815 | .Cases(STR, ALIAS, {1u, static_cast<unsigned>(llvm::X86::ENUM)}) |
9816 | #define X86_CPU_TYPE_COMPAT(ARCHNAME, ENUM, STR) \ |
9817 | .Case(STR, {1u, static_cast<unsigned>(llvm::X86::ENUM)}) |
9818 | #define X86_CPU_SUBTYPE_COMPAT(ARCHNAME, ENUM, STR) \ |
9819 | .Case(STR, {2u, static_cast<unsigned>(llvm::X86::ENUM)}) |
9820 | #include "llvm/Support/X86TargetParser.def" |
9821 | .Default({0, 0}); |
9822 | assert(Value != 0 && "Invalid CPUStr passed to CpuIs")((Value != 0 && "Invalid CPUStr passed to CpuIs") ? static_cast <void> (0) : __assert_fail ("Value != 0 && \"Invalid CPUStr passed to CpuIs\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 9822, __PRETTY_FUNCTION__)); |
9823 | |
9824 | // Grab the appropriate field from __cpu_model. |
9825 | llvm::Value *Idxs[] = {ConstantInt::get(Int32Ty, 0), |
9826 | ConstantInt::get(Int32Ty, Index)}; |
9827 | llvm::Value *CpuValue = Builder.CreateGEP(STy, CpuModel, Idxs); |
9828 | CpuValue = Builder.CreateAlignedLoad(CpuValue, CharUnits::fromQuantity(4)); |
9829 | |
9830 | // Check the value of the field against the requested value. |
9831 | return Builder.CreateICmpEQ(CpuValue, |
9832 | llvm::ConstantInt::get(Int32Ty, Value)); |
9833 | } |
9834 | |
9835 | Value *CodeGenFunction::EmitX86CpuSupports(const CallExpr *E) { |
9836 | const Expr *FeatureExpr = E->getArg(0)->IgnoreParenCasts(); |
9837 | StringRef FeatureStr = cast<StringLiteral>(FeatureExpr)->getString(); |
9838 | return EmitX86CpuSupports(FeatureStr); |
9839 | } |
9840 | |
9841 | uint64_t |
9842 | CodeGenFunction::GetX86CpuSupportsMask(ArrayRef<StringRef> FeatureStrs) { |
9843 | // Processor features and mapping to processor feature value. |
9844 | uint64_t FeaturesMask = 0; |
9845 | for (const StringRef &FeatureStr : FeatureStrs) { |
9846 | unsigned Feature = |
9847 | StringSwitch<unsigned>(FeatureStr) |
9848 | #define X86_FEATURE_COMPAT(VAL, ENUM, STR) .Case(STR, VAL) |
9849 | #include "llvm/Support/X86TargetParser.def" |
9850 | ; |
9851 | FeaturesMask |= (1ULL << Feature); |
9852 | } |
9853 | return FeaturesMask; |
9854 | } |
9855 | |
9856 | Value *CodeGenFunction::EmitX86CpuSupports(ArrayRef<StringRef> FeatureStrs) { |
9857 | return EmitX86CpuSupports(GetX86CpuSupportsMask(FeatureStrs)); |
9858 | } |
9859 | |
9860 | llvm::Value *CodeGenFunction::EmitX86CpuSupports(uint64_t FeaturesMask) { |
9861 | uint32_t Features1 = Lo_32(FeaturesMask); |
9862 | uint32_t Features2 = Hi_32(FeaturesMask); |
9863 | |
9864 | Value *Result = Builder.getTrue(); |
9865 | |
9866 | if (Features1 != 0) { |
9867 | // Matching the struct layout from the compiler-rt/libgcc structure that is |
9868 | // filled in: |
9869 | // unsigned int __cpu_vendor; |
9870 | // unsigned int __cpu_type; |
9871 | // unsigned int __cpu_subtype; |
9872 | // unsigned int __cpu_features[1]; |
9873 | llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, Int32Ty, |
9874 | llvm::ArrayType::get(Int32Ty, 1)); |
9875 | |
9876 | // Grab the global __cpu_model. |
9877 | llvm::Constant *CpuModel = CGM.CreateRuntimeVariable(STy, "__cpu_model"); |
9878 | cast<llvm::GlobalValue>(CpuModel)->setDSOLocal(true); |
9879 | |
9880 | // Grab the first (0th) element from the field __cpu_features off of the |
9881 | // global in the struct STy. |
9882 | Value *Idxs[] = {Builder.getInt32(0), Builder.getInt32(3), |
9883 | Builder.getInt32(0)}; |
9884 | Value *CpuFeatures = Builder.CreateGEP(STy, CpuModel, Idxs); |
9885 | Value *Features = |
9886 | Builder.CreateAlignedLoad(CpuFeatures, CharUnits::fromQuantity(4)); |
9887 | |
9888 | // Check the value of the bit corresponding to the feature requested. |
9889 | Value *Mask = Builder.getInt32(Features1); |
9890 | Value *Bitset = Builder.CreateAnd(Features, Mask); |
9891 | Value *Cmp = Builder.CreateICmpEQ(Bitset, Mask); |
9892 | Result = Builder.CreateAnd(Result, Cmp); |
9893 | } |
9894 | |
9895 | if (Features2 != 0) { |
9896 | llvm::Constant *CpuFeatures2 = CGM.CreateRuntimeVariable(Int32Ty, |
9897 | "__cpu_features2"); |
9898 | cast<llvm::GlobalValue>(CpuFeatures2)->setDSOLocal(true); |
9899 | |
9900 | Value *Features = |
9901 | Builder.CreateAlignedLoad(CpuFeatures2, CharUnits::fromQuantity(4)); |
9902 | |
9903 | // Check the value of the bit corresponding to the feature requested. |
9904 | Value *Mask = Builder.getInt32(Features2); |
9905 | Value *Bitset = Builder.CreateAnd(Features, Mask); |
9906 | Value *Cmp = Builder.CreateICmpEQ(Bitset, Mask); |
9907 | Result = Builder.CreateAnd(Result, Cmp); |
9908 | } |
9909 | |
9910 | return Result; |
9911 | } |
9912 | |
9913 | Value *CodeGenFunction::EmitX86CpuInit() { |
9914 | llvm::FunctionType *FTy = llvm::FunctionType::get(VoidTy, |
9915 | /*Variadic*/ false); |
9916 | llvm::FunctionCallee Func = |
9917 | CGM.CreateRuntimeFunction(FTy, "__cpu_indicator_init"); |
9918 | cast<llvm::GlobalValue>(Func.getCallee())->setDSOLocal(true); |
9919 | cast<llvm::GlobalValue>(Func.getCallee()) |
9920 | ->setDLLStorageClass(llvm::GlobalValue::DefaultStorageClass); |
9921 | return Builder.CreateCall(Func); |
9922 | } |
9923 | |
9924 | Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID, |
9925 | const CallExpr *E) { |
9926 | if (BuiltinID == X86::BI__builtin_cpu_is) |
9927 | return EmitX86CpuIs(E); |
9928 | if (BuiltinID == X86::BI__builtin_cpu_supports) |
9929 | return EmitX86CpuSupports(E); |
9930 | if (BuiltinID == X86::BI__builtin_cpu_init) |
9931 | return EmitX86CpuInit(); |
9932 | |
9933 | SmallVector<Value*, 4> Ops; |
9934 | |
9935 | // Find out if any arguments are required to be integer constant expressions. |
9936 | unsigned ICEArguments = 0; |
9937 | ASTContext::GetBuiltinTypeError Error; |
9938 | getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); |
9939 | assert(Error == ASTContext::GE_None && "Should not codegen an error")((Error == ASTContext::GE_None && "Should not codegen an error" ) ? static_cast<void> (0) : __assert_fail ("Error == ASTContext::GE_None && \"Should not codegen an error\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 9939, __PRETTY_FUNCTION__)); |
9940 | |
9941 | for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) { |
9942 | // If this is a normal argument, just emit it as a scalar. |
9943 | if ((ICEArguments & (1 << i)) == 0) { |
9944 | Ops.push_back(EmitScalarExpr(E->getArg(i))); |
9945 | continue; |
9946 | } |
9947 | |
9948 | // If this is required to be a constant, constant fold it so that we know |
9949 | // that the generated intrinsic gets a ConstantInt. |
9950 | llvm::APSInt Result; |
9951 | bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); |
9952 | assert(IsConst && "Constant arg isn't actually constant?")((IsConst && "Constant arg isn't actually constant?") ? static_cast<void> (0) : __assert_fail ("IsConst && \"Constant arg isn't actually constant?\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 9952, __PRETTY_FUNCTION__)); (void)IsConst; |
9953 | Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); |
9954 | } |
9955 | |
9956 | // These exist so that the builtin that takes an immediate can be bounds |
9957 | // checked by clang to avoid passing bad immediates to the backend. Since |
9958 | // AVX has a larger immediate than SSE we would need separate builtins to |
9959 | // do the different bounds checking. Rather than create a clang specific |
9960 | // SSE only builtin, this implements eight separate builtins to match gcc |
9961 | // implementation. |
9962 | auto getCmpIntrinsicCall = [this, &Ops](Intrinsic::ID ID, unsigned Imm) { |
9963 | Ops.push_back(llvm::ConstantInt::get(Int8Ty, Imm)); |
9964 | llvm::Function *F = CGM.getIntrinsic(ID); |
9965 | return Builder.CreateCall(F, Ops); |
9966 | }; |
9967 | |
9968 | // For the vector forms of FP comparisons, translate the builtins directly to |
9969 | // IR. |
9970 | // TODO: The builtins could be removed if the SSE header files used vector |
9971 | // extension comparisons directly (vector ordered/unordered may need |
9972 | // additional support via __builtin_isnan()). |
9973 | auto getVectorFCmpIR = [this, &Ops](CmpInst::Predicate Pred) { |
9974 | Value *Cmp = Builder.CreateFCmp(Pred, Ops[0], Ops[1]); |
9975 | llvm::VectorType *FPVecTy = cast<llvm::VectorType>(Ops[0]->getType()); |
9976 | llvm::VectorType *IntVecTy = llvm::VectorType::getInteger(FPVecTy); |
9977 | Value *Sext = Builder.CreateSExt(Cmp, IntVecTy); |
9978 | return Builder.CreateBitCast(Sext, FPVecTy); |
9979 | }; |
9980 | |
9981 | switch (BuiltinID) { |
9982 | default: return nullptr; |
9983 | case X86::BI_mm_prefetch: { |
9984 | Value *Address = Ops[0]; |
9985 | ConstantInt *C = cast<ConstantInt>(Ops[1]); |
9986 | Value *RW = ConstantInt::get(Int32Ty, (C->getZExtValue() >> 2) & 0x1); |
9987 | Value *Locality = ConstantInt::get(Int32Ty, C->getZExtValue() & 0x3); |
9988 | Value *Data = ConstantInt::get(Int32Ty, 1); |
9989 | Function *F = CGM.getIntrinsic(Intrinsic::prefetch); |
9990 | return Builder.CreateCall(F, {Address, RW, Locality, Data}); |
9991 | } |
9992 | case X86::BI_mm_clflush: { |
9993 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse2_clflush), |
9994 | Ops[0]); |
9995 | } |
9996 | case X86::BI_mm_lfence: { |
9997 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse2_lfence)); |
9998 | } |
9999 | case X86::BI_mm_mfence: { |
10000 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse2_mfence)); |
10001 | } |
10002 | case X86::BI_mm_sfence: { |
10003 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_sfence)); |
10004 | } |
10005 | case X86::BI_mm_pause: { |
10006 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse2_pause)); |
10007 | } |
10008 | case X86::BI__rdtsc: { |
10009 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_rdtsc)); |
10010 | } |
10011 | case X86::BI__builtin_ia32_rdtscp: { |
10012 | Value *Call = Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_rdtscp)); |
10013 | Builder.CreateDefaultAlignedStore(Builder.CreateExtractValue(Call, 1), |
10014 | Ops[0]); |
10015 | return Builder.CreateExtractValue(Call, 0); |
10016 | } |
10017 | case X86::BI__builtin_ia32_lzcnt_u16: |
10018 | case X86::BI__builtin_ia32_lzcnt_u32: |
10019 | case X86::BI__builtin_ia32_lzcnt_u64: { |
10020 | Function *F = CGM.getIntrinsic(Intrinsic::ctlz, Ops[0]->getType()); |
10021 | return Builder.CreateCall(F, {Ops[0], Builder.getInt1(false)}); |
10022 | } |
10023 | case X86::BI__builtin_ia32_tzcnt_u16: |
10024 | case X86::BI__builtin_ia32_tzcnt_u32: |
10025 | case X86::BI__builtin_ia32_tzcnt_u64: { |
10026 | Function *F = CGM.getIntrinsic(Intrinsic::cttz, Ops[0]->getType()); |
10027 | return Builder.CreateCall(F, {Ops[0], Builder.getInt1(false)}); |
10028 | } |
10029 | case X86::BI__builtin_ia32_undef128: |
10030 | case X86::BI__builtin_ia32_undef256: |
10031 | case X86::BI__builtin_ia32_undef512: |
10032 | // The x86 definition of "undef" is not the same as the LLVM definition |
10033 | // (PR32176). We leave optimizing away an unnecessary zero constant to the |
10034 | // IR optimizer and backend. |
10035 | // TODO: If we had a "freeze" IR instruction to generate a fixed undef |
10036 | // value, we should use that here instead of a zero. |
10037 | return llvm::Constant::getNullValue(ConvertType(E->getType())); |
10038 | case X86::BI__builtin_ia32_vec_init_v8qi: |
10039 | case X86::BI__builtin_ia32_vec_init_v4hi: |
10040 | case X86::BI__builtin_ia32_vec_init_v2si: |
10041 | return Builder.CreateBitCast(BuildVector(Ops), |
10042 | llvm::Type::getX86_MMXTy(getLLVMContext())); |
10043 | case X86::BI__builtin_ia32_vec_ext_v2si: |
10044 | case X86::BI__builtin_ia32_vec_ext_v16qi: |
10045 | case X86::BI__builtin_ia32_vec_ext_v8hi: |
10046 | case X86::BI__builtin_ia32_vec_ext_v4si: |
10047 | case X86::BI__builtin_ia32_vec_ext_v4sf: |
10048 | case X86::BI__builtin_ia32_vec_ext_v2di: |
10049 | case X86::BI__builtin_ia32_vec_ext_v32qi: |
10050 | case X86::BI__builtin_ia32_vec_ext_v16hi: |
10051 | case X86::BI__builtin_ia32_vec_ext_v8si: |
10052 | case X86::BI__builtin_ia32_vec_ext_v4di: { |
10053 | unsigned NumElts = Ops[0]->getType()->getVectorNumElements(); |
10054 | uint64_t Index = cast<ConstantInt>(Ops[1])->getZExtValue(); |
10055 | Index &= NumElts - 1; |
10056 | // These builtins exist so we can ensure the index is an ICE and in range. |
10057 | // Otherwise we could just do this in the header file. |
10058 | return Builder.CreateExtractElement(Ops[0], Index); |
10059 | } |
10060 | case X86::BI__builtin_ia32_vec_set_v16qi: |
10061 | case X86::BI__builtin_ia32_vec_set_v8hi: |
10062 | case X86::BI__builtin_ia32_vec_set_v4si: |
10063 | case X86::BI__builtin_ia32_vec_set_v2di: |
10064 | case X86::BI__builtin_ia32_vec_set_v32qi: |
10065 | case X86::BI__builtin_ia32_vec_set_v16hi: |
10066 | case X86::BI__builtin_ia32_vec_set_v8si: |
10067 | case X86::BI__builtin_ia32_vec_set_v4di: { |
10068 | unsigned NumElts = Ops[0]->getType()->getVectorNumElements(); |
10069 | unsigned Index = cast<ConstantInt>(Ops[2])->getZExtValue(); |
10070 | Index &= NumElts - 1; |
10071 | // These builtins exist so we can ensure the index is an ICE and in range. |
10072 | // Otherwise we could just do this in the header file. |
10073 | return Builder.CreateInsertElement(Ops[0], Ops[1], Index); |
10074 | } |
10075 | case X86::BI_mm_setcsr: |
10076 | case X86::BI__builtin_ia32_ldmxcsr: { |
10077 | Address Tmp = CreateMemTemp(E->getArg(0)->getType()); |
10078 | Builder.CreateStore(Ops[0], Tmp); |
10079 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr), |
10080 | Builder.CreateBitCast(Tmp.getPointer(), Int8PtrTy)); |
10081 | } |
10082 | case X86::BI_mm_getcsr: |
10083 | case X86::BI__builtin_ia32_stmxcsr: { |
10084 | Address Tmp = CreateMemTemp(E->getType()); |
10085 | Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr), |
10086 | Builder.CreateBitCast(Tmp.getPointer(), Int8PtrTy)); |
10087 | return Builder.CreateLoad(Tmp, "stmxcsr"); |
10088 | } |
10089 | case X86::BI__builtin_ia32_xsave: |
10090 | case X86::BI__builtin_ia32_xsave64: |
10091 | case X86::BI__builtin_ia32_xrstor: |
10092 | case X86::BI__builtin_ia32_xrstor64: |
10093 | case X86::BI__builtin_ia32_xsaveopt: |
10094 | case X86::BI__builtin_ia32_xsaveopt64: |
10095 | case X86::BI__builtin_ia32_xrstors: |
10096 | case X86::BI__builtin_ia32_xrstors64: |
10097 | case X86::BI__builtin_ia32_xsavec: |
10098 | case X86::BI__builtin_ia32_xsavec64: |
10099 | case X86::BI__builtin_ia32_xsaves: |
10100 | case X86::BI__builtin_ia32_xsaves64: |
10101 | case X86::BI__builtin_ia32_xsetbv: |
10102 | case X86::BI_xsetbv: { |
10103 | Intrinsic::ID ID; |
10104 | #define INTRINSIC_X86_XSAVE_ID(NAME) \ |
10105 | case X86::BI__builtin_ia32_##NAME: \ |
10106 | ID = Intrinsic::x86_##NAME; \ |
10107 | break |
10108 | switch (BuiltinID) { |
10109 | default: llvm_unreachable("Unsupported intrinsic!")::llvm::llvm_unreachable_internal("Unsupported intrinsic!", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 10109); |
10110 | INTRINSIC_X86_XSAVE_ID(xsave); |
10111 | INTRINSIC_X86_XSAVE_ID(xsave64); |
10112 | INTRINSIC_X86_XSAVE_ID(xrstor); |
10113 | INTRINSIC_X86_XSAVE_ID(xrstor64); |
10114 | INTRINSIC_X86_XSAVE_ID(xsaveopt); |
10115 | INTRINSIC_X86_XSAVE_ID(xsaveopt64); |
10116 | INTRINSIC_X86_XSAVE_ID(xrstors); |
10117 | INTRINSIC_X86_XSAVE_ID(xrstors64); |
10118 | INTRINSIC_X86_XSAVE_ID(xsavec); |
10119 | INTRINSIC_X86_XSAVE_ID(xsavec64); |
10120 | INTRINSIC_X86_XSAVE_ID(xsaves); |
10121 | INTRINSIC_X86_XSAVE_ID(xsaves64); |
10122 | INTRINSIC_X86_XSAVE_ID(xsetbv); |
10123 | case X86::BI_xsetbv: |
10124 | ID = Intrinsic::x86_xsetbv; |
10125 | break; |
10126 | } |
10127 | #undef INTRINSIC_X86_XSAVE_ID |
10128 | Value *Mhi = Builder.CreateTrunc( |
10129 | Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, 32)), Int32Ty); |
10130 | Value *Mlo = Builder.CreateTrunc(Ops[1], Int32Ty); |
10131 | Ops[1] = Mhi; |
10132 | Ops.push_back(Mlo); |
10133 | return Builder.CreateCall(CGM.getIntrinsic(ID), Ops); |
10134 | } |
10135 | case X86::BI__builtin_ia32_xgetbv: |
10136 | case X86::BI_xgetbv: |
10137 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_xgetbv), Ops); |
10138 | case X86::BI__builtin_ia32_storedqudi128_mask: |
10139 | case X86::BI__builtin_ia32_storedqusi128_mask: |
10140 | case X86::BI__builtin_ia32_storedquhi128_mask: |
10141 | case X86::BI__builtin_ia32_storedquqi128_mask: |
10142 | case X86::BI__builtin_ia32_storeupd128_mask: |
10143 | case X86::BI__builtin_ia32_storeups128_mask: |
10144 | case X86::BI__builtin_ia32_storedqudi256_mask: |
10145 | case X86::BI__builtin_ia32_storedqusi256_mask: |
10146 | case X86::BI__builtin_ia32_storedquhi256_mask: |
10147 | case X86::BI__builtin_ia32_storedquqi256_mask: |
10148 | case X86::BI__builtin_ia32_storeupd256_mask: |
10149 | case X86::BI__builtin_ia32_storeups256_mask: |
10150 | case X86::BI__builtin_ia32_storedqudi512_mask: |
10151 | case X86::BI__builtin_ia32_storedqusi512_mask: |
10152 | case X86::BI__builtin_ia32_storedquhi512_mask: |
10153 | case X86::BI__builtin_ia32_storedquqi512_mask: |
10154 | case X86::BI__builtin_ia32_storeupd512_mask: |
10155 | case X86::BI__builtin_ia32_storeups512_mask: |
10156 | return EmitX86MaskedStore(*this, Ops, 1); |
10157 | |
10158 | case X86::BI__builtin_ia32_storess128_mask: |
10159 | case X86::BI__builtin_ia32_storesd128_mask: { |
10160 | return EmitX86MaskedStore(*this, Ops, 1); |
10161 | } |
10162 | case X86::BI__builtin_ia32_vpopcntb_128: |
10163 | case X86::BI__builtin_ia32_vpopcntd_128: |
10164 | case X86::BI__builtin_ia32_vpopcntq_128: |
10165 | case X86::BI__builtin_ia32_vpopcntw_128: |
10166 | case X86::BI__builtin_ia32_vpopcntb_256: |
10167 | case X86::BI__builtin_ia32_vpopcntd_256: |
10168 | case X86::BI__builtin_ia32_vpopcntq_256: |
10169 | case X86::BI__builtin_ia32_vpopcntw_256: |
10170 | case X86::BI__builtin_ia32_vpopcntb_512: |
10171 | case X86::BI__builtin_ia32_vpopcntd_512: |
10172 | case X86::BI__builtin_ia32_vpopcntq_512: |
10173 | case X86::BI__builtin_ia32_vpopcntw_512: { |
10174 | llvm::Type *ResultType = ConvertType(E->getType()); |
10175 | llvm::Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ResultType); |
10176 | return Builder.CreateCall(F, Ops); |
10177 | } |
10178 | case X86::BI__builtin_ia32_cvtmask2b128: |
10179 | case X86::BI__builtin_ia32_cvtmask2b256: |
10180 | case X86::BI__builtin_ia32_cvtmask2b512: |
10181 | case X86::BI__builtin_ia32_cvtmask2w128: |
10182 | case X86::BI__builtin_ia32_cvtmask2w256: |
10183 | case X86::BI__builtin_ia32_cvtmask2w512: |
10184 | case X86::BI__builtin_ia32_cvtmask2d128: |
10185 | case X86::BI__builtin_ia32_cvtmask2d256: |
10186 | case X86::BI__builtin_ia32_cvtmask2d512: |
10187 | case X86::BI__builtin_ia32_cvtmask2q128: |
10188 | case X86::BI__builtin_ia32_cvtmask2q256: |
10189 | case X86::BI__builtin_ia32_cvtmask2q512: |
10190 | return EmitX86SExtMask(*this, Ops[0], ConvertType(E->getType())); |
10191 | |
10192 | case X86::BI__builtin_ia32_cvtb2mask128: |
10193 | case X86::BI__builtin_ia32_cvtb2mask256: |
10194 | case X86::BI__builtin_ia32_cvtb2mask512: |
10195 | case X86::BI__builtin_ia32_cvtw2mask128: |
10196 | case X86::BI__builtin_ia32_cvtw2mask256: |
10197 | case X86::BI__builtin_ia32_cvtw2mask512: |
10198 | case X86::BI__builtin_ia32_cvtd2mask128: |
10199 | case X86::BI__builtin_ia32_cvtd2mask256: |
10200 | case X86::BI__builtin_ia32_cvtd2mask512: |
10201 | case X86::BI__builtin_ia32_cvtq2mask128: |
10202 | case X86::BI__builtin_ia32_cvtq2mask256: |
10203 | case X86::BI__builtin_ia32_cvtq2mask512: |
10204 | return EmitX86ConvertToMask(*this, Ops[0]); |
10205 | |
10206 | case X86::BI__builtin_ia32_cvtdq2ps512_mask: |
10207 | case X86::BI__builtin_ia32_cvtqq2ps512_mask: |
10208 | case X86::BI__builtin_ia32_cvtqq2pd512_mask: |
10209 | return EmitX86ConvertIntToFp(*this, Ops, /*IsSigned*/true); |
10210 | case X86::BI__builtin_ia32_cvtudq2ps512_mask: |
10211 | case X86::BI__builtin_ia32_cvtuqq2ps512_mask: |
10212 | case X86::BI__builtin_ia32_cvtuqq2pd512_mask: |
10213 | return EmitX86ConvertIntToFp(*this, Ops, /*IsSigned*/false); |
10214 | |
10215 | case X86::BI__builtin_ia32_vfmaddss3: |
10216 | case X86::BI__builtin_ia32_vfmaddsd3: |
10217 | case X86::BI__builtin_ia32_vfmaddss3_mask: |
10218 | case X86::BI__builtin_ia32_vfmaddsd3_mask: |
10219 | return EmitScalarFMAExpr(*this, Ops, Ops[0]); |
10220 | case X86::BI__builtin_ia32_vfmaddss: |
10221 | case X86::BI__builtin_ia32_vfmaddsd: |
10222 | return EmitScalarFMAExpr(*this, Ops, |
10223 | Constant::getNullValue(Ops[0]->getType())); |
10224 | case X86::BI__builtin_ia32_vfmaddss3_maskz: |
10225 | case X86::BI__builtin_ia32_vfmaddsd3_maskz: |
10226 | return EmitScalarFMAExpr(*this, Ops, Ops[0], /*ZeroMask*/true); |
10227 | case X86::BI__builtin_ia32_vfmaddss3_mask3: |
10228 | case X86::BI__builtin_ia32_vfmaddsd3_mask3: |
10229 | return EmitScalarFMAExpr(*this, Ops, Ops[2], /*ZeroMask*/false, 2); |
10230 | case X86::BI__builtin_ia32_vfmsubss3_mask3: |
10231 | case X86::BI__builtin_ia32_vfmsubsd3_mask3: |
10232 | return EmitScalarFMAExpr(*this, Ops, Ops[2], /*ZeroMask*/false, 2, |
10233 | /*NegAcc*/true); |
10234 | case X86::BI__builtin_ia32_vfmaddps: |
10235 | case X86::BI__builtin_ia32_vfmaddpd: |
10236 | case X86::BI__builtin_ia32_vfmaddps256: |
10237 | case X86::BI__builtin_ia32_vfmaddpd256: |
10238 | case X86::BI__builtin_ia32_vfmaddps512_mask: |
10239 | case X86::BI__builtin_ia32_vfmaddps512_maskz: |
10240 | case X86::BI__builtin_ia32_vfmaddps512_mask3: |
10241 | case X86::BI__builtin_ia32_vfmsubps512_mask3: |
10242 | case X86::BI__builtin_ia32_vfmaddpd512_mask: |
10243 | case X86::BI__builtin_ia32_vfmaddpd512_maskz: |
10244 | case X86::BI__builtin_ia32_vfmaddpd512_mask3: |
10245 | case X86::BI__builtin_ia32_vfmsubpd512_mask3: |
10246 | return EmitX86FMAExpr(*this, Ops, BuiltinID, /*IsAddSub*/false); |
10247 | case X86::BI__builtin_ia32_vfmaddsubps: |
10248 | case X86::BI__builtin_ia32_vfmaddsubpd: |
10249 | case X86::BI__builtin_ia32_vfmaddsubps256: |
10250 | case X86::BI__builtin_ia32_vfmaddsubpd256: |
10251 | case X86::BI__builtin_ia32_vfmaddsubps512_mask: |
10252 | case X86::BI__builtin_ia32_vfmaddsubps512_maskz: |
10253 | case X86::BI__builtin_ia32_vfmaddsubps512_mask3: |
10254 | case X86::BI__builtin_ia32_vfmsubaddps512_mask3: |
10255 | case X86::BI__builtin_ia32_vfmaddsubpd512_mask: |
10256 | case X86::BI__builtin_ia32_vfmaddsubpd512_maskz: |
10257 | case X86::BI__builtin_ia32_vfmaddsubpd512_mask3: |
10258 | case X86::BI__builtin_ia32_vfmsubaddpd512_mask3: |
10259 | return EmitX86FMAExpr(*this, Ops, BuiltinID, /*IsAddSub*/true); |
10260 | |
10261 | case X86::BI__builtin_ia32_movdqa32store128_mask: |
10262 | case X86::BI__builtin_ia32_movdqa64store128_mask: |
10263 | case X86::BI__builtin_ia32_storeaps128_mask: |
10264 | case X86::BI__builtin_ia32_storeapd128_mask: |
10265 | case X86::BI__builtin_ia32_movdqa32store256_mask: |
10266 | case X86::BI__builtin_ia32_movdqa64store256_mask: |
10267 | case X86::BI__builtin_ia32_storeaps256_mask: |
10268 | case X86::BI__builtin_ia32_storeapd256_mask: |
10269 | case X86::BI__builtin_ia32_movdqa32store512_mask: |
10270 | case X86::BI__builtin_ia32_movdqa64store512_mask: |
10271 | case X86::BI__builtin_ia32_storeaps512_mask: |
10272 | case X86::BI__builtin_ia32_storeapd512_mask: { |
10273 | unsigned Align = |
10274 | getContext().getTypeAlignInChars(E->getArg(1)->getType()).getQuantity(); |
10275 | return EmitX86MaskedStore(*this, Ops, Align); |
10276 | } |
10277 | case X86::BI__builtin_ia32_loadups128_mask: |
10278 | case X86::BI__builtin_ia32_loadups256_mask: |
10279 | case X86::BI__builtin_ia32_loadups512_mask: |
10280 | case X86::BI__builtin_ia32_loadupd128_mask: |
10281 | case X86::BI__builtin_ia32_loadupd256_mask: |
10282 | case X86::BI__builtin_ia32_loadupd512_mask: |
10283 | case X86::BI__builtin_ia32_loaddquqi128_mask: |
10284 | case X86::BI__builtin_ia32_loaddquqi256_mask: |
10285 | case X86::BI__builtin_ia32_loaddquqi512_mask: |
10286 | case X86::BI__builtin_ia32_loaddquhi128_mask: |
10287 | case X86::BI__builtin_ia32_loaddquhi256_mask: |
10288 | case X86::BI__builtin_ia32_loaddquhi512_mask: |
10289 | case X86::BI__builtin_ia32_loaddqusi128_mask: |
10290 | case X86::BI__builtin_ia32_loaddqusi256_mask: |
10291 | case X86::BI__builtin_ia32_loaddqusi512_mask: |
10292 | case X86::BI__builtin_ia32_loaddqudi128_mask: |
10293 | case X86::BI__builtin_ia32_loaddqudi256_mask: |
10294 | case X86::BI__builtin_ia32_loaddqudi512_mask: |
10295 | return EmitX86MaskedLoad(*this, Ops, 1); |
10296 | |
10297 | case X86::BI__builtin_ia32_loadss128_mask: |
10298 | case X86::BI__builtin_ia32_loadsd128_mask: |
10299 | return EmitX86MaskedLoad(*this, Ops, 1); |
10300 | |
10301 | case X86::BI__builtin_ia32_loadaps128_mask: |
10302 | case X86::BI__builtin_ia32_loadaps256_mask: |
10303 | case X86::BI__builtin_ia32_loadaps512_mask: |
10304 | case X86::BI__builtin_ia32_loadapd128_mask: |
10305 | case X86::BI__builtin_ia32_loadapd256_mask: |
10306 | case X86::BI__builtin_ia32_loadapd512_mask: |
10307 | case X86::BI__builtin_ia32_movdqa32load128_mask: |
10308 | case X86::BI__builtin_ia32_movdqa32load256_mask: |
10309 | case X86::BI__builtin_ia32_movdqa32load512_mask: |
10310 | case X86::BI__builtin_ia32_movdqa64load128_mask: |
10311 | case X86::BI__builtin_ia32_movdqa64load256_mask: |
10312 | case X86::BI__builtin_ia32_movdqa64load512_mask: { |
10313 | unsigned Align = |
10314 | getContext().getTypeAlignInChars(E->getArg(1)->getType()).getQuantity(); |
10315 | return EmitX86MaskedLoad(*this, Ops, Align); |
10316 | } |
10317 | |
10318 | case X86::BI__builtin_ia32_expandloaddf128_mask: |
10319 | case X86::BI__builtin_ia32_expandloaddf256_mask: |
10320 | case X86::BI__builtin_ia32_expandloaddf512_mask: |
10321 | case X86::BI__builtin_ia32_expandloadsf128_mask: |
10322 | case X86::BI__builtin_ia32_expandloadsf256_mask: |
10323 | case X86::BI__builtin_ia32_expandloadsf512_mask: |
10324 | case X86::BI__builtin_ia32_expandloaddi128_mask: |
10325 | case X86::BI__builtin_ia32_expandloaddi256_mask: |
10326 | case X86::BI__builtin_ia32_expandloaddi512_mask: |
10327 | case X86::BI__builtin_ia32_expandloadsi128_mask: |
10328 | case X86::BI__builtin_ia32_expandloadsi256_mask: |
10329 | case X86::BI__builtin_ia32_expandloadsi512_mask: |
10330 | case X86::BI__builtin_ia32_expandloadhi128_mask: |
10331 | case X86::BI__builtin_ia32_expandloadhi256_mask: |
10332 | case X86::BI__builtin_ia32_expandloadhi512_mask: |
10333 | case X86::BI__builtin_ia32_expandloadqi128_mask: |
10334 | case X86::BI__builtin_ia32_expandloadqi256_mask: |
10335 | case X86::BI__builtin_ia32_expandloadqi512_mask: |
10336 | return EmitX86ExpandLoad(*this, Ops); |
10337 | |
10338 | case X86::BI__builtin_ia32_compressstoredf128_mask: |
10339 | case X86::BI__builtin_ia32_compressstoredf256_mask: |
10340 | case X86::BI__builtin_ia32_compressstoredf512_mask: |
10341 | case X86::BI__builtin_ia32_compressstoresf128_mask: |
10342 | case X86::BI__builtin_ia32_compressstoresf256_mask: |
10343 | case X86::BI__builtin_ia32_compressstoresf512_mask: |
10344 | case X86::BI__builtin_ia32_compressstoredi128_mask: |
10345 | case X86::BI__builtin_ia32_compressstoredi256_mask: |
10346 | case X86::BI__builtin_ia32_compressstoredi512_mask: |
10347 | case X86::BI__builtin_ia32_compressstoresi128_mask: |
10348 | case X86::BI__builtin_ia32_compressstoresi256_mask: |
10349 | case X86::BI__builtin_ia32_compressstoresi512_mask: |
10350 | case X86::BI__builtin_ia32_compressstorehi128_mask: |
10351 | case X86::BI__builtin_ia32_compressstorehi256_mask: |
10352 | case X86::BI__builtin_ia32_compressstorehi512_mask: |
10353 | case X86::BI__builtin_ia32_compressstoreqi128_mask: |
10354 | case X86::BI__builtin_ia32_compressstoreqi256_mask: |
10355 | case X86::BI__builtin_ia32_compressstoreqi512_mask: |
10356 | return EmitX86CompressStore(*this, Ops); |
10357 | |
10358 | case X86::BI__builtin_ia32_expanddf128_mask: |
10359 | case X86::BI__builtin_ia32_expanddf256_mask: |
10360 | case X86::BI__builtin_ia32_expanddf512_mask: |
10361 | case X86::BI__builtin_ia32_expandsf128_mask: |
10362 | case X86::BI__builtin_ia32_expandsf256_mask: |
10363 | case X86::BI__builtin_ia32_expandsf512_mask: |
10364 | case X86::BI__builtin_ia32_expanddi128_mask: |
10365 | case X86::BI__builtin_ia32_expanddi256_mask: |
10366 | case X86::BI__builtin_ia32_expanddi512_mask: |
10367 | case X86::BI__builtin_ia32_expandsi128_mask: |
10368 | case X86::BI__builtin_ia32_expandsi256_mask: |
10369 | case X86::BI__builtin_ia32_expandsi512_mask: |
10370 | case X86::BI__builtin_ia32_expandhi128_mask: |
10371 | case X86::BI__builtin_ia32_expandhi256_mask: |
10372 | case X86::BI__builtin_ia32_expandhi512_mask: |
10373 | case X86::BI__builtin_ia32_expandqi128_mask: |
10374 | case X86::BI__builtin_ia32_expandqi256_mask: |
10375 | case X86::BI__builtin_ia32_expandqi512_mask: |
10376 | return EmitX86CompressExpand(*this, Ops, /*IsCompress*/false); |
10377 | |
10378 | case X86::BI__builtin_ia32_compressdf128_mask: |
10379 | case X86::BI__builtin_ia32_compressdf256_mask: |
10380 | case X86::BI__builtin_ia32_compressdf512_mask: |
10381 | case X86::BI__builtin_ia32_compresssf128_mask: |
10382 | case X86::BI__builtin_ia32_compresssf256_mask: |
10383 | case X86::BI__builtin_ia32_compresssf512_mask: |
10384 | case X86::BI__builtin_ia32_compressdi128_mask: |
10385 | case X86::BI__builtin_ia32_compressdi256_mask: |
10386 | case X86::BI__builtin_ia32_compressdi512_mask: |
10387 | case X86::BI__builtin_ia32_compresssi128_mask: |
10388 | case X86::BI__builtin_ia32_compresssi256_mask: |
10389 | case X86::BI__builtin_ia32_compresssi512_mask: |
10390 | case X86::BI__builtin_ia32_compresshi128_mask: |
10391 | case X86::BI__builtin_ia32_compresshi256_mask: |
10392 | case X86::BI__builtin_ia32_compresshi512_mask: |
10393 | case X86::BI__builtin_ia32_compressqi128_mask: |
10394 | case X86::BI__builtin_ia32_compressqi256_mask: |
10395 | case X86::BI__builtin_ia32_compressqi512_mask: |
10396 | return EmitX86CompressExpand(*this, Ops, /*IsCompress*/true); |
10397 | |
10398 | case X86::BI__builtin_ia32_gather3div2df: |
10399 | case X86::BI__builtin_ia32_gather3div2di: |
10400 | case X86::BI__builtin_ia32_gather3div4df: |
10401 | case X86::BI__builtin_ia32_gather3div4di: |
10402 | case X86::BI__builtin_ia32_gather3div4sf: |
10403 | case X86::BI__builtin_ia32_gather3div4si: |
10404 | case X86::BI__builtin_ia32_gather3div8sf: |
10405 | case X86::BI__builtin_ia32_gather3div8si: |
10406 | case X86::BI__builtin_ia32_gather3siv2df: |
10407 | case X86::BI__builtin_ia32_gather3siv2di: |
10408 | case X86::BI__builtin_ia32_gather3siv4df: |
10409 | case X86::BI__builtin_ia32_gather3siv4di: |
10410 | case X86::BI__builtin_ia32_gather3siv4sf: |
10411 | case X86::BI__builtin_ia32_gather3siv4si: |
10412 | case X86::BI__builtin_ia32_gather3siv8sf: |
10413 | case X86::BI__builtin_ia32_gather3siv8si: |
10414 | case X86::BI__builtin_ia32_gathersiv8df: |
10415 | case X86::BI__builtin_ia32_gathersiv16sf: |
10416 | case X86::BI__builtin_ia32_gatherdiv8df: |
10417 | case X86::BI__builtin_ia32_gatherdiv16sf: |
10418 | case X86::BI__builtin_ia32_gathersiv8di: |
10419 | case X86::BI__builtin_ia32_gathersiv16si: |
10420 | case X86::BI__builtin_ia32_gatherdiv8di: |
10421 | case X86::BI__builtin_ia32_gatherdiv16si: { |
10422 | Intrinsic::ID IID; |
10423 | switch (BuiltinID) { |
10424 | default: llvm_unreachable("Unexpected builtin")::llvm::llvm_unreachable_internal("Unexpected builtin", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 10424); |
10425 | case X86::BI__builtin_ia32_gather3div2df: |
10426 | IID = Intrinsic::x86_avx512_mask_gather3div2_df; |
10427 | break; |
10428 | case X86::BI__builtin_ia32_gather3div2di: |
10429 | IID = Intrinsic::x86_avx512_mask_gather3div2_di; |
10430 | break; |
10431 | case X86::BI__builtin_ia32_gather3div4df: |
10432 | IID = Intrinsic::x86_avx512_mask_gather3div4_df; |
10433 | break; |
10434 | case X86::BI__builtin_ia32_gather3div4di: |
10435 | IID = Intrinsic::x86_avx512_mask_gather3div4_di; |
10436 | break; |
10437 | case X86::BI__builtin_ia32_gather3div4sf: |
10438 | IID = Intrinsic::x86_avx512_mask_gather3div4_sf; |
10439 | break; |
10440 | case X86::BI__builtin_ia32_gather3div4si: |
10441 | IID = Intrinsic::x86_avx512_mask_gather3div4_si; |
10442 | break; |
10443 | case X86::BI__builtin_ia32_gather3div8sf: |
10444 | IID = Intrinsic::x86_avx512_mask_gather3div8_sf; |
10445 | break; |
10446 | case X86::BI__builtin_ia32_gather3div8si: |
10447 | IID = Intrinsic::x86_avx512_mask_gather3div8_si; |
10448 | break; |
10449 | case X86::BI__builtin_ia32_gather3siv2df: |
10450 | IID = Intrinsic::x86_avx512_mask_gather3siv2_df; |
10451 | break; |
10452 | case X86::BI__builtin_ia32_gather3siv2di: |
10453 | IID = Intrinsic::x86_avx512_mask_gather3siv2_di; |
10454 | break; |
10455 | case X86::BI__builtin_ia32_gather3siv4df: |
10456 | IID = Intrinsic::x86_avx512_mask_gather3siv4_df; |
10457 | break; |
10458 | case X86::BI__builtin_ia32_gather3siv4di: |
10459 | IID = Intrinsic::x86_avx512_mask_gather3siv4_di; |
10460 | break; |
10461 | case X86::BI__builtin_ia32_gather3siv4sf: |
10462 | IID = Intrinsic::x86_avx512_mask_gather3siv4_sf; |
10463 | break; |
10464 | case X86::BI__builtin_ia32_gather3siv4si: |
10465 | IID = Intrinsic::x86_avx512_mask_gather3siv4_si; |
10466 | break; |
10467 | case X86::BI__builtin_ia32_gather3siv8sf: |
10468 | IID = Intrinsic::x86_avx512_mask_gather3siv8_sf; |
10469 | break; |
10470 | case X86::BI__builtin_ia32_gather3siv8si: |
10471 | IID = Intrinsic::x86_avx512_mask_gather3siv8_si; |
10472 | break; |
10473 | case X86::BI__builtin_ia32_gathersiv8df: |
10474 | IID = Intrinsic::x86_avx512_mask_gather_dpd_512; |
10475 | break; |
10476 | case X86::BI__builtin_ia32_gathersiv16sf: |
10477 | IID = Intrinsic::x86_avx512_mask_gather_dps_512; |
10478 | break; |
10479 | case X86::BI__builtin_ia32_gatherdiv8df: |
10480 | IID = Intrinsic::x86_avx512_mask_gather_qpd_512; |
10481 | break; |
10482 | case X86::BI__builtin_ia32_gatherdiv16sf: |
10483 | IID = Intrinsic::x86_avx512_mask_gather_qps_512; |
10484 | break; |
10485 | case X86::BI__builtin_ia32_gathersiv8di: |
10486 | IID = Intrinsic::x86_avx512_mask_gather_dpq_512; |
10487 | break; |
10488 | case X86::BI__builtin_ia32_gathersiv16si: |
10489 | IID = Intrinsic::x86_avx512_mask_gather_dpi_512; |
10490 | break; |
10491 | case X86::BI__builtin_ia32_gatherdiv8di: |
10492 | IID = Intrinsic::x86_avx512_mask_gather_qpq_512; |
10493 | break; |
10494 | case X86::BI__builtin_ia32_gatherdiv16si: |
10495 | IID = Intrinsic::x86_avx512_mask_gather_qpi_512; |
10496 | break; |
10497 | } |
10498 | |
10499 | unsigned MinElts = std::min(Ops[0]->getType()->getVectorNumElements(), |
10500 | Ops[2]->getType()->getVectorNumElements()); |
10501 | Ops[3] = getMaskVecValue(*this, Ops[3], MinElts); |
10502 | Function *Intr = CGM.getIntrinsic(IID); |
10503 | return Builder.CreateCall(Intr, Ops); |
10504 | } |
10505 | |
10506 | case X86::BI__builtin_ia32_scattersiv8df: |
10507 | case X86::BI__builtin_ia32_scattersiv16sf: |
10508 | case X86::BI__builtin_ia32_scatterdiv8df: |
10509 | case X86::BI__builtin_ia32_scatterdiv16sf: |
10510 | case X86::BI__builtin_ia32_scattersiv8di: |
10511 | case X86::BI__builtin_ia32_scattersiv16si: |
10512 | case X86::BI__builtin_ia32_scatterdiv8di: |
10513 | case X86::BI__builtin_ia32_scatterdiv16si: |
10514 | case X86::BI__builtin_ia32_scatterdiv2df: |
10515 | case X86::BI__builtin_ia32_scatterdiv2di: |
10516 | case X86::BI__builtin_ia32_scatterdiv4df: |
10517 | case X86::BI__builtin_ia32_scatterdiv4di: |
10518 | case X86::BI__builtin_ia32_scatterdiv4sf: |
10519 | case X86::BI__builtin_ia32_scatterdiv4si: |
10520 | case X86::BI__builtin_ia32_scatterdiv8sf: |
10521 | case X86::BI__builtin_ia32_scatterdiv8si: |
10522 | case X86::BI__builtin_ia32_scattersiv2df: |
10523 | case X86::BI__builtin_ia32_scattersiv2di: |
10524 | case X86::BI__builtin_ia32_scattersiv4df: |
10525 | case X86::BI__builtin_ia32_scattersiv4di: |
10526 | case X86::BI__builtin_ia32_scattersiv4sf: |
10527 | case X86::BI__builtin_ia32_scattersiv4si: |
10528 | case X86::BI__builtin_ia32_scattersiv8sf: |
10529 | case X86::BI__builtin_ia32_scattersiv8si: { |
10530 | Intrinsic::ID IID; |
10531 | switch (BuiltinID) { |
10532 | default: llvm_unreachable("Unexpected builtin")::llvm::llvm_unreachable_internal("Unexpected builtin", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 10532); |
10533 | case X86::BI__builtin_ia32_scattersiv8df: |
10534 | IID = Intrinsic::x86_avx512_mask_scatter_dpd_512; |
10535 | break; |
10536 | case X86::BI__builtin_ia32_scattersiv16sf: |
10537 | IID = Intrinsic::x86_avx512_mask_scatter_dps_512; |
10538 | break; |
10539 | case X86::BI__builtin_ia32_scatterdiv8df: |
10540 | IID = Intrinsic::x86_avx512_mask_scatter_qpd_512; |
10541 | break; |
10542 | case X86::BI__builtin_ia32_scatterdiv16sf: |
10543 | IID = Intrinsic::x86_avx512_mask_scatter_qps_512; |
10544 | break; |
10545 | case X86::BI__builtin_ia32_scattersiv8di: |
10546 | IID = Intrinsic::x86_avx512_mask_scatter_dpq_512; |
10547 | break; |
10548 | case X86::BI__builtin_ia32_scattersiv16si: |
10549 | IID = Intrinsic::x86_avx512_mask_scatter_dpi_512; |
10550 | break; |
10551 | case X86::BI__builtin_ia32_scatterdiv8di: |
10552 | IID = Intrinsic::x86_avx512_mask_scatter_qpq_512; |
10553 | break; |
10554 | case X86::BI__builtin_ia32_scatterdiv16si: |
10555 | IID = Intrinsic::x86_avx512_mask_scatter_qpi_512; |
10556 | break; |
10557 | case X86::BI__builtin_ia32_scatterdiv2df: |
10558 | IID = Intrinsic::x86_avx512_mask_scatterdiv2_df; |
10559 | break; |
10560 | case X86::BI__builtin_ia32_scatterdiv2di: |
10561 | IID = Intrinsic::x86_avx512_mask_scatterdiv2_di; |
10562 | break; |
10563 | case X86::BI__builtin_ia32_scatterdiv4df: |
10564 | IID = Intrinsic::x86_avx512_mask_scatterdiv4_df; |
10565 | break; |
10566 | case X86::BI__builtin_ia32_scatterdiv4di: |
10567 | IID = Intrinsic::x86_avx512_mask_scatterdiv4_di; |
10568 | break; |
10569 | case X86::BI__builtin_ia32_scatterdiv4sf: |
10570 | IID = Intrinsic::x86_avx512_mask_scatterdiv4_sf; |
10571 | break; |
10572 | case X86::BI__builtin_ia32_scatterdiv4si: |
10573 | IID = Intrinsic::x86_avx512_mask_scatterdiv4_si; |
10574 | break; |
10575 | case X86::BI__builtin_ia32_scatterdiv8sf: |
10576 | IID = Intrinsic::x86_avx512_mask_scatterdiv8_sf; |
10577 | break; |
10578 | case X86::BI__builtin_ia32_scatterdiv8si: |
10579 | IID = Intrinsic::x86_avx512_mask_scatterdiv8_si; |
10580 | break; |
10581 | case X86::BI__builtin_ia32_scattersiv2df: |
10582 | IID = Intrinsic::x86_avx512_mask_scattersiv2_df; |
10583 | break; |
10584 | case X86::BI__builtin_ia32_scattersiv2di: |
10585 | IID = Intrinsic::x86_avx512_mask_scattersiv2_di; |
10586 | break; |
10587 | case X86::BI__builtin_ia32_scattersiv4df: |
10588 | IID = Intrinsic::x86_avx512_mask_scattersiv4_df; |
10589 | break; |
10590 | case X86::BI__builtin_ia32_scattersiv4di: |
10591 | IID = Intrinsic::x86_avx512_mask_scattersiv4_di; |
10592 | break; |
10593 | case X86::BI__builtin_ia32_scattersiv4sf: |
10594 | IID = Intrinsic::x86_avx512_mask_scattersiv4_sf; |
10595 | break; |
10596 | case X86::BI__builtin_ia32_scattersiv4si: |
10597 | IID = Intrinsic::x86_avx512_mask_scattersiv4_si; |
10598 | break; |
10599 | case X86::BI__builtin_ia32_scattersiv8sf: |
10600 | IID = Intrinsic::x86_avx512_mask_scattersiv8_sf; |
10601 | break; |
10602 | case X86::BI__builtin_ia32_scattersiv8si: |
10603 | IID = Intrinsic::x86_avx512_mask_scattersiv8_si; |
10604 | break; |
10605 | } |
10606 | |
10607 | unsigned MinElts = std::min(Ops[2]->getType()->getVectorNumElements(), |
10608 | Ops[3]->getType()->getVectorNumElements()); |
10609 | Ops[1] = getMaskVecValue(*this, Ops[1], MinElts); |
10610 | Function *Intr = CGM.getIntrinsic(IID); |
10611 | return Builder.CreateCall(Intr, Ops); |
10612 | } |
10613 | |
10614 | case X86::BI__builtin_ia32_storehps: |
10615 | case X86::BI__builtin_ia32_storelps: { |
10616 | llvm::Type *PtrTy = llvm::PointerType::getUnqual(Int64Ty); |
10617 | llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); |
10618 | |
10619 | // cast val v2i64 |
10620 | Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast"); |
10621 | |
10622 | // extract (0, 1) |
10623 | unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1; |
10624 | Ops[1] = Builder.CreateExtractElement(Ops[1], Index, "extract"); |
10625 | |
10626 | // cast pointer to i64 & store |
10627 | Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy); |
10628 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
10629 | } |
10630 | case X86::BI__builtin_ia32_vextractf128_pd256: |
10631 | case X86::BI__builtin_ia32_vextractf128_ps256: |
10632 | case X86::BI__builtin_ia32_vextractf128_si256: |
10633 | case X86::BI__builtin_ia32_extract128i256: |
10634 | case X86::BI__builtin_ia32_extractf64x4_mask: |
10635 | case X86::BI__builtin_ia32_extractf32x4_mask: |
10636 | case X86::BI__builtin_ia32_extracti64x4_mask: |
10637 | case X86::BI__builtin_ia32_extracti32x4_mask: |
10638 | case X86::BI__builtin_ia32_extractf32x8_mask: |
10639 | case X86::BI__builtin_ia32_extracti32x8_mask: |
10640 | case X86::BI__builtin_ia32_extractf32x4_256_mask: |
10641 | case X86::BI__builtin_ia32_extracti32x4_256_mask: |
10642 | case X86::BI__builtin_ia32_extractf64x2_256_mask: |
10643 | case X86::BI__builtin_ia32_extracti64x2_256_mask: |
10644 | case X86::BI__builtin_ia32_extractf64x2_512_mask: |
10645 | case X86::BI__builtin_ia32_extracti64x2_512_mask: { |
10646 | llvm::Type *DstTy = ConvertType(E->getType()); |
10647 | unsigned NumElts = DstTy->getVectorNumElements(); |
10648 | unsigned SrcNumElts = Ops[0]->getType()->getVectorNumElements(); |
10649 | unsigned SubVectors = SrcNumElts / NumElts; |
10650 | unsigned Index = cast<ConstantInt>(Ops[1])->getZExtValue(); |
10651 | assert(llvm::isPowerOf2_32(SubVectors) && "Expected power of 2 subvectors")((llvm::isPowerOf2_32(SubVectors) && "Expected power of 2 subvectors" ) ? static_cast<void> (0) : __assert_fail ("llvm::isPowerOf2_32(SubVectors) && \"Expected power of 2 subvectors\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 10651, __PRETTY_FUNCTION__)); |
10652 | Index &= SubVectors - 1; // Remove any extra bits. |
10653 | Index *= NumElts; |
10654 | |
10655 | uint32_t Indices[16]; |
10656 | for (unsigned i = 0; i != NumElts; ++i) |
10657 | Indices[i] = i + Index; |
10658 | |
10659 | Value *Res = Builder.CreateShuffleVector(Ops[0], |
10660 | UndefValue::get(Ops[0]->getType()), |
10661 | makeArrayRef(Indices, NumElts), |
10662 | "extract"); |
10663 | |
10664 | if (Ops.size() == 4) |
10665 | Res = EmitX86Select(*this, Ops[3], Res, Ops[2]); |
10666 | |
10667 | return Res; |
10668 | } |
10669 | case X86::BI__builtin_ia32_vinsertf128_pd256: |
10670 | case X86::BI__builtin_ia32_vinsertf128_ps256: |
10671 | case X86::BI__builtin_ia32_vinsertf128_si256: |
10672 | case X86::BI__builtin_ia32_insert128i256: |
10673 | case X86::BI__builtin_ia32_insertf64x4: |
10674 | case X86::BI__builtin_ia32_insertf32x4: |
10675 | case X86::BI__builtin_ia32_inserti64x4: |
10676 | case X86::BI__builtin_ia32_inserti32x4: |
10677 | case X86::BI__builtin_ia32_insertf32x8: |
10678 | case X86::BI__builtin_ia32_inserti32x8: |
10679 | case X86::BI__builtin_ia32_insertf32x4_256: |
10680 | case X86::BI__builtin_ia32_inserti32x4_256: |
10681 | case X86::BI__builtin_ia32_insertf64x2_256: |
10682 | case X86::BI__builtin_ia32_inserti64x2_256: |
10683 | case X86::BI__builtin_ia32_insertf64x2_512: |
10684 | case X86::BI__builtin_ia32_inserti64x2_512: { |
10685 | unsigned DstNumElts = Ops[0]->getType()->getVectorNumElements(); |
10686 | unsigned SrcNumElts = Ops[1]->getType()->getVectorNumElements(); |
10687 | unsigned SubVectors = DstNumElts / SrcNumElts; |
10688 | unsigned Index = cast<ConstantInt>(Ops[2])->getZExtValue(); |
10689 | assert(llvm::isPowerOf2_32(SubVectors) && "Expected power of 2 subvectors")((llvm::isPowerOf2_32(SubVectors) && "Expected power of 2 subvectors" ) ? static_cast<void> (0) : __assert_fail ("llvm::isPowerOf2_32(SubVectors) && \"Expected power of 2 subvectors\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 10689, __PRETTY_FUNCTION__)); |
10690 | Index &= SubVectors - 1; // Remove any extra bits. |
10691 | Index *= SrcNumElts; |
10692 | |
10693 | uint32_t Indices[16]; |
10694 | for (unsigned i = 0; i != DstNumElts; ++i) |
10695 | Indices[i] = (i >= SrcNumElts) ? SrcNumElts + (i % SrcNumElts) : i; |
10696 | |
10697 | Value *Op1 = Builder.CreateShuffleVector(Ops[1], |
10698 | UndefValue::get(Ops[1]->getType()), |
10699 | makeArrayRef(Indices, DstNumElts), |
10700 | "widen"); |
10701 | |
10702 | for (unsigned i = 0; i != DstNumElts; ++i) { |
10703 | if (i >= Index && i < (Index + SrcNumElts)) |
10704 | Indices[i] = (i - Index) + DstNumElts; |
10705 | else |
10706 | Indices[i] = i; |
10707 | } |
10708 | |
10709 | return Builder.CreateShuffleVector(Ops[0], Op1, |
10710 | makeArrayRef(Indices, DstNumElts), |
10711 | "insert"); |
10712 | } |
10713 | case X86::BI__builtin_ia32_pmovqd512_mask: |
10714 | case X86::BI__builtin_ia32_pmovwb512_mask: { |
10715 | Value *Res = Builder.CreateTrunc(Ops[0], Ops[1]->getType()); |
10716 | return EmitX86Select(*this, Ops[2], Res, Ops[1]); |
10717 | } |
10718 | case X86::BI__builtin_ia32_pmovdb512_mask: |
10719 | case X86::BI__builtin_ia32_pmovdw512_mask: |
10720 | case X86::BI__builtin_ia32_pmovqw512_mask: { |
10721 | if (const auto *C = dyn_cast<Constant>(Ops[2])) |
10722 | if (C->isAllOnesValue()) |
10723 | return Builder.CreateTrunc(Ops[0], Ops[1]->getType()); |
10724 | |
10725 | Intrinsic::ID IID; |
10726 | switch (BuiltinID) { |
10727 | default: llvm_unreachable("Unsupported intrinsic!")::llvm::llvm_unreachable_internal("Unsupported intrinsic!", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 10727); |
10728 | case X86::BI__builtin_ia32_pmovdb512_mask: |
10729 | IID = Intrinsic::x86_avx512_mask_pmov_db_512; |
10730 | break; |
10731 | case X86::BI__builtin_ia32_pmovdw512_mask: |
10732 | IID = Intrinsic::x86_avx512_mask_pmov_dw_512; |
10733 | break; |
10734 | case X86::BI__builtin_ia32_pmovqw512_mask: |
10735 | IID = Intrinsic::x86_avx512_mask_pmov_qw_512; |
10736 | break; |
10737 | } |
10738 | |
10739 | Function *Intr = CGM.getIntrinsic(IID); |
10740 | return Builder.CreateCall(Intr, Ops); |
10741 | } |
10742 | case X86::BI__builtin_ia32_pblendw128: |
10743 | case X86::BI__builtin_ia32_blendpd: |
10744 | case X86::BI__builtin_ia32_blendps: |
10745 | case X86::BI__builtin_ia32_blendpd256: |
10746 | case X86::BI__builtin_ia32_blendps256: |
10747 | case X86::BI__builtin_ia32_pblendw256: |
10748 | case X86::BI__builtin_ia32_pblendd128: |
10749 | case X86::BI__builtin_ia32_pblendd256: { |
10750 | unsigned NumElts = Ops[0]->getType()->getVectorNumElements(); |
10751 | unsigned Imm = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); |
10752 | |
10753 | uint32_t Indices[16]; |
10754 | // If there are more than 8 elements, the immediate is used twice so make |
10755 | // sure we handle that. |
10756 | for (unsigned i = 0; i != NumElts; ++i) |
10757 | Indices[i] = ((Imm >> (i % 8)) & 0x1) ? NumElts + i : i; |
10758 | |
10759 | return Builder.CreateShuffleVector(Ops[0], Ops[1], |
10760 | makeArrayRef(Indices, NumElts), |
10761 | "blend"); |
10762 | } |
10763 | case X86::BI__builtin_ia32_pshuflw: |
10764 | case X86::BI__builtin_ia32_pshuflw256: |
10765 | case X86::BI__builtin_ia32_pshuflw512: { |
10766 | uint32_t Imm = cast<llvm::ConstantInt>(Ops[1])->getZExtValue(); |
10767 | llvm::Type *Ty = Ops[0]->getType(); |
10768 | unsigned NumElts = Ty->getVectorNumElements(); |
10769 | |
10770 | // Splat the 8-bits of immediate 4 times to help the loop wrap around. |
10771 | Imm = (Imm & 0xff) * 0x01010101; |
10772 | |
10773 | uint32_t Indices[32]; |
10774 | for (unsigned l = 0; l != NumElts; l += 8) { |
10775 | for (unsigned i = 0; i != 4; ++i) { |
10776 | Indices[l + i] = l + (Imm & 3); |
10777 | Imm >>= 2; |
10778 | } |
10779 | for (unsigned i = 4; i != 8; ++i) |
10780 | Indices[l + i] = l + i; |
10781 | } |
10782 | |
10783 | return Builder.CreateShuffleVector(Ops[0], UndefValue::get(Ty), |
10784 | makeArrayRef(Indices, NumElts), |
10785 | "pshuflw"); |
10786 | } |
10787 | case X86::BI__builtin_ia32_pshufhw: |
10788 | case X86::BI__builtin_ia32_pshufhw256: |
10789 | case X86::BI__builtin_ia32_pshufhw512: { |
10790 | uint32_t Imm = cast<llvm::ConstantInt>(Ops[1])->getZExtValue(); |
10791 | llvm::Type *Ty = Ops[0]->getType(); |
10792 | unsigned NumElts = Ty->getVectorNumElements(); |
10793 | |
10794 | // Splat the 8-bits of immediate 4 times to help the loop wrap around. |
10795 | Imm = (Imm & 0xff) * 0x01010101; |
10796 | |
10797 | uint32_t Indices[32]; |
10798 | for (unsigned l = 0; l != NumElts; l += 8) { |
10799 | for (unsigned i = 0; i != 4; ++i) |
10800 | Indices[l + i] = l + i; |
10801 | for (unsigned i = 4; i != 8; ++i) { |
10802 | Indices[l + i] = l + 4 + (Imm & 3); |
10803 | Imm >>= 2; |
10804 | } |
10805 | } |
10806 | |
10807 | return Builder.CreateShuffleVector(Ops[0], UndefValue::get(Ty), |
10808 | makeArrayRef(Indices, NumElts), |
10809 | "pshufhw"); |
10810 | } |
10811 | case X86::BI__builtin_ia32_pshufd: |
10812 | case X86::BI__builtin_ia32_pshufd256: |
10813 | case X86::BI__builtin_ia32_pshufd512: |
10814 | case X86::BI__builtin_ia32_vpermilpd: |
10815 | case X86::BI__builtin_ia32_vpermilps: |
10816 | case X86::BI__builtin_ia32_vpermilpd256: |
10817 | case X86::BI__builtin_ia32_vpermilps256: |
10818 | case X86::BI__builtin_ia32_vpermilpd512: |
10819 | case X86::BI__builtin_ia32_vpermilps512: { |
10820 | uint32_t Imm = cast<llvm::ConstantInt>(Ops[1])->getZExtValue(); |
10821 | llvm::Type *Ty = Ops[0]->getType(); |
10822 | unsigned NumElts = Ty->getVectorNumElements(); |
10823 | unsigned NumLanes = Ty->getPrimitiveSizeInBits() / 128; |
10824 | unsigned NumLaneElts = NumElts / NumLanes; |
10825 | |
10826 | // Splat the 8-bits of immediate 4 times to help the loop wrap around. |
10827 | Imm = (Imm & 0xff) * 0x01010101; |
10828 | |
10829 | uint32_t Indices[16]; |
10830 | for (unsigned l = 0; l != NumElts; l += NumLaneElts) { |
10831 | for (unsigned i = 0; i != NumLaneElts; ++i) { |
10832 | Indices[i + l] = (Imm % NumLaneElts) + l; |
10833 | Imm /= NumLaneElts; |
10834 | } |
10835 | } |
10836 | |
10837 | return Builder.CreateShuffleVector(Ops[0], UndefValue::get(Ty), |
10838 | makeArrayRef(Indices, NumElts), |
10839 | "permil"); |
10840 | } |
10841 | case X86::BI__builtin_ia32_shufpd: |
10842 | case X86::BI__builtin_ia32_shufpd256: |
10843 | case X86::BI__builtin_ia32_shufpd512: |
10844 | case X86::BI__builtin_ia32_shufps: |
10845 | case X86::BI__builtin_ia32_shufps256: |
10846 | case X86::BI__builtin_ia32_shufps512: { |
10847 | uint32_t Imm = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); |
10848 | llvm::Type *Ty = Ops[0]->getType(); |
10849 | unsigned NumElts = Ty->getVectorNumElements(); |
10850 | unsigned NumLanes = Ty->getPrimitiveSizeInBits() / 128; |
10851 | unsigned NumLaneElts = NumElts / NumLanes; |
10852 | |
10853 | // Splat the 8-bits of immediate 4 times to help the loop wrap around. |
10854 | Imm = (Imm & 0xff) * 0x01010101; |
10855 | |
10856 | uint32_t Indices[16]; |
10857 | for (unsigned l = 0; l != NumElts; l += NumLaneElts) { |
10858 | for (unsigned i = 0; i != NumLaneElts; ++i) { |
10859 | unsigned Index = Imm % NumLaneElts; |
10860 | Imm /= NumLaneElts; |
10861 | if (i >= (NumLaneElts / 2)) |
10862 | Index += NumElts; |
10863 | Indices[l + i] = l + Index; |
10864 | } |
10865 | } |
10866 | |
10867 | return Builder.CreateShuffleVector(Ops[0], Ops[1], |
10868 | makeArrayRef(Indices, NumElts), |
10869 | "shufp"); |
10870 | } |
10871 | case X86::BI__builtin_ia32_permdi256: |
10872 | case X86::BI__builtin_ia32_permdf256: |
10873 | case X86::BI__builtin_ia32_permdi512: |
10874 | case X86::BI__builtin_ia32_permdf512: { |
10875 | unsigned Imm = cast<llvm::ConstantInt>(Ops[1])->getZExtValue(); |
10876 | llvm::Type *Ty = Ops[0]->getType(); |
10877 | unsigned NumElts = Ty->getVectorNumElements(); |
10878 | |
10879 | // These intrinsics operate on 256-bit lanes of four 64-bit elements. |
10880 | uint32_t Indices[8]; |
10881 | for (unsigned l = 0; l != NumElts; l += 4) |
10882 | for (unsigned i = 0; i != 4; ++i) |
10883 | Indices[l + i] = l + ((Imm >> (2 * i)) & 0x3); |
10884 | |
10885 | return Builder.CreateShuffleVector(Ops[0], UndefValue::get(Ty), |
10886 | makeArrayRef(Indices, NumElts), |
10887 | "perm"); |
10888 | } |
10889 | case X86::BI__builtin_ia32_palignr128: |
10890 | case X86::BI__builtin_ia32_palignr256: |
10891 | case X86::BI__builtin_ia32_palignr512: { |
10892 | unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue() & 0xff; |
10893 | |
10894 | unsigned NumElts = Ops[0]->getType()->getVectorNumElements(); |
10895 | assert(NumElts % 16 == 0)((NumElts % 16 == 0) ? static_cast<void> (0) : __assert_fail ("NumElts % 16 == 0", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 10895, __PRETTY_FUNCTION__)); |
10896 | |
10897 | // If palignr is shifting the pair of vectors more than the size of two |
10898 | // lanes, emit zero. |
10899 | if (ShiftVal >= 32) |
10900 | return llvm::Constant::getNullValue(ConvertType(E->getType())); |
10901 | |
10902 | // If palignr is shifting the pair of input vectors more than one lane, |
10903 | // but less than two lanes, convert to shifting in zeroes. |
10904 | if (ShiftVal > 16) { |
10905 | ShiftVal -= 16; |
10906 | Ops[1] = Ops[0]; |
10907 | Ops[0] = llvm::Constant::getNullValue(Ops[0]->getType()); |
10908 | } |
10909 | |
10910 | uint32_t Indices[64]; |
10911 | // 256-bit palignr operates on 128-bit lanes so we need to handle that |
10912 | for (unsigned l = 0; l != NumElts; l += 16) { |
10913 | for (unsigned i = 0; i != 16; ++i) { |
10914 | unsigned Idx = ShiftVal + i; |
10915 | if (Idx >= 16) |
10916 | Idx += NumElts - 16; // End of lane, switch operand. |
10917 | Indices[l + i] = Idx + l; |
10918 | } |
10919 | } |
10920 | |
10921 | return Builder.CreateShuffleVector(Ops[1], Ops[0], |
10922 | makeArrayRef(Indices, NumElts), |
10923 | "palignr"); |
10924 | } |
10925 | case X86::BI__builtin_ia32_alignd128: |
10926 | case X86::BI__builtin_ia32_alignd256: |
10927 | case X86::BI__builtin_ia32_alignd512: |
10928 | case X86::BI__builtin_ia32_alignq128: |
10929 | case X86::BI__builtin_ia32_alignq256: |
10930 | case X86::BI__builtin_ia32_alignq512: { |
10931 | unsigned NumElts = Ops[0]->getType()->getVectorNumElements(); |
10932 | unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue() & 0xff; |
10933 | |
10934 | // Mask the shift amount to width of two vectors. |
10935 | ShiftVal &= (2 * NumElts) - 1; |
10936 | |
10937 | uint32_t Indices[16]; |
10938 | for (unsigned i = 0; i != NumElts; ++i) |
10939 | Indices[i] = i + ShiftVal; |
10940 | |
10941 | return Builder.CreateShuffleVector(Ops[1], Ops[0], |
10942 | makeArrayRef(Indices, NumElts), |
10943 | "valign"); |
10944 | } |
10945 | case X86::BI__builtin_ia32_shuf_f32x4_256: |
10946 | case X86::BI__builtin_ia32_shuf_f64x2_256: |
10947 | case X86::BI__builtin_ia32_shuf_i32x4_256: |
10948 | case X86::BI__builtin_ia32_shuf_i64x2_256: |
10949 | case X86::BI__builtin_ia32_shuf_f32x4: |
10950 | case X86::BI__builtin_ia32_shuf_f64x2: |
10951 | case X86::BI__builtin_ia32_shuf_i32x4: |
10952 | case X86::BI__builtin_ia32_shuf_i64x2: { |
10953 | unsigned Imm = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); |
10954 | llvm::Type *Ty = Ops[0]->getType(); |
10955 | unsigned NumElts = Ty->getVectorNumElements(); |
10956 | unsigned NumLanes = Ty->getPrimitiveSizeInBits() == 512 ? 4 : 2; |
10957 | unsigned NumLaneElts = NumElts / NumLanes; |
10958 | |
10959 | uint32_t Indices[16]; |
10960 | for (unsigned l = 0; l != NumElts; l += NumLaneElts) { |
10961 | unsigned Index = (Imm % NumLanes) * NumLaneElts; |
10962 | Imm /= NumLanes; // Discard the bits we just used. |
10963 | if (l >= (NumElts / 2)) |
10964 | Index += NumElts; // Switch to other source. |
10965 | for (unsigned i = 0; i != NumLaneElts; ++i) { |
10966 | Indices[l + i] = Index + i; |
10967 | } |
10968 | } |
10969 | |
10970 | return Builder.CreateShuffleVector(Ops[0], Ops[1], |
10971 | makeArrayRef(Indices, NumElts), |
10972 | "shuf"); |
10973 | } |
10974 | |
10975 | case X86::BI__builtin_ia32_vperm2f128_pd256: |
10976 | case X86::BI__builtin_ia32_vperm2f128_ps256: |
10977 | case X86::BI__builtin_ia32_vperm2f128_si256: |
10978 | case X86::BI__builtin_ia32_permti256: { |
10979 | unsigned Imm = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); |
10980 | unsigned NumElts = Ops[0]->getType()->getVectorNumElements(); |
10981 | |
10982 | // This takes a very simple approach since there are two lanes and a |
10983 | // shuffle can have 2 inputs. So we reserve the first input for the first |
10984 | // lane and the second input for the second lane. This may result in |
10985 | // duplicate sources, but this can be dealt with in the backend. |
10986 | |
10987 | Value *OutOps[2]; |
10988 | uint32_t Indices[8]; |
10989 | for (unsigned l = 0; l != 2; ++l) { |
10990 | // Determine the source for this lane. |
10991 | if (Imm & (1 << ((l * 4) + 3))) |
10992 | OutOps[l] = llvm::ConstantAggregateZero::get(Ops[0]->getType()); |
10993 | else if (Imm & (1 << ((l * 4) + 1))) |
10994 | OutOps[l] = Ops[1]; |
10995 | else |
10996 | OutOps[l] = Ops[0]; |
10997 | |
10998 | for (unsigned i = 0; i != NumElts/2; ++i) { |
10999 | // Start with ith element of the source for this lane. |
11000 | unsigned Idx = (l * NumElts) + i; |
11001 | // If bit 0 of the immediate half is set, switch to the high half of |
11002 | // the source. |
11003 | if (Imm & (1 << (l * 4))) |
11004 | Idx += NumElts/2; |
11005 | Indices[(l * (NumElts/2)) + i] = Idx; |
11006 | } |
11007 | } |
11008 | |
11009 | return Builder.CreateShuffleVector(OutOps[0], OutOps[1], |
11010 | makeArrayRef(Indices, NumElts), |
11011 | "vperm"); |
11012 | } |
11013 | |
11014 | case X86::BI__builtin_ia32_pslldqi128_byteshift: |
11015 | case X86::BI__builtin_ia32_pslldqi256_byteshift: |
11016 | case X86::BI__builtin_ia32_pslldqi512_byteshift: { |
11017 | unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() & 0xff; |
11018 | llvm::Type *ResultType = Ops[0]->getType(); |
11019 | // Builtin type is vXi64 so multiply by 8 to get bytes. |
11020 | unsigned NumElts = ResultType->getVectorNumElements() * 8; |
11021 | |
11022 | // If pslldq is shifting the vector more than 15 bytes, emit zero. |
11023 | if (ShiftVal >= 16) |
11024 | return llvm::Constant::getNullValue(ResultType); |
11025 | |
11026 | uint32_t Indices[64]; |
11027 | // 256/512-bit pslldq operates on 128-bit lanes so we need to handle that |
11028 | for (unsigned l = 0; l != NumElts; l += 16) { |
11029 | for (unsigned i = 0; i != 16; ++i) { |
11030 | unsigned Idx = NumElts + i - ShiftVal; |
11031 | if (Idx < NumElts) Idx -= NumElts - 16; // end of lane, switch operand. |
11032 | Indices[l + i] = Idx + l; |
11033 | } |
11034 | } |
11035 | |
11036 | llvm::Type *VecTy = llvm::VectorType::get(Int8Ty, NumElts); |
11037 | Value *Cast = Builder.CreateBitCast(Ops[0], VecTy, "cast"); |
11038 | Value *Zero = llvm::Constant::getNullValue(VecTy); |
11039 | Value *SV = Builder.CreateShuffleVector(Zero, Cast, |
11040 | makeArrayRef(Indices, NumElts), |
11041 | "pslldq"); |
11042 | return Builder.CreateBitCast(SV, Ops[0]->getType(), "cast"); |
11043 | } |
11044 | case X86::BI__builtin_ia32_psrldqi128_byteshift: |
11045 | case X86::BI__builtin_ia32_psrldqi256_byteshift: |
11046 | case X86::BI__builtin_ia32_psrldqi512_byteshift: { |
11047 | unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() & 0xff; |
11048 | llvm::Type *ResultType = Ops[0]->getType(); |
11049 | // Builtin type is vXi64 so multiply by 8 to get bytes. |
11050 | unsigned NumElts = ResultType->getVectorNumElements() * 8; |
11051 | |
11052 | // If psrldq is shifting the vector more than 15 bytes, emit zero. |
11053 | if (ShiftVal >= 16) |
11054 | return llvm::Constant::getNullValue(ResultType); |
11055 | |
11056 | uint32_t Indices[64]; |
11057 | // 256/512-bit psrldq operates on 128-bit lanes so we need to handle that |
11058 | for (unsigned l = 0; l != NumElts; l += 16) { |
11059 | for (unsigned i = 0; i != 16; ++i) { |
11060 | unsigned Idx = i + ShiftVal; |
11061 | if (Idx >= 16) Idx += NumElts - 16; // end of lane, switch operand. |
11062 | Indices[l + i] = Idx + l; |
11063 | } |
11064 | } |
11065 | |
11066 | llvm::Type *VecTy = llvm::VectorType::get(Int8Ty, NumElts); |
11067 | Value *Cast = Builder.CreateBitCast(Ops[0], VecTy, "cast"); |
11068 | Value *Zero = llvm::Constant::getNullValue(VecTy); |
11069 | Value *SV = Builder.CreateShuffleVector(Cast, Zero, |
11070 | makeArrayRef(Indices, NumElts), |
11071 | "psrldq"); |
11072 | return Builder.CreateBitCast(SV, ResultType, "cast"); |
11073 | } |
11074 | case X86::BI__builtin_ia32_kshiftliqi: |
11075 | case X86::BI__builtin_ia32_kshiftlihi: |
11076 | case X86::BI__builtin_ia32_kshiftlisi: |
11077 | case X86::BI__builtin_ia32_kshiftlidi: { |
11078 | unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() & 0xff; |
11079 | unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth(); |
11080 | |
11081 | if (ShiftVal >= NumElts) |
11082 | return llvm::Constant::getNullValue(Ops[0]->getType()); |
11083 | |
11084 | Value *In = getMaskVecValue(*this, Ops[0], NumElts); |
11085 | |
11086 | uint32_t Indices[64]; |
11087 | for (unsigned i = 0; i != NumElts; ++i) |
11088 | Indices[i] = NumElts + i - ShiftVal; |
11089 | |
11090 | Value *Zero = llvm::Constant::getNullValue(In->getType()); |
11091 | Value *SV = Builder.CreateShuffleVector(Zero, In, |
11092 | makeArrayRef(Indices, NumElts), |
11093 | "kshiftl"); |
11094 | return Builder.CreateBitCast(SV, Ops[0]->getType()); |
11095 | } |
11096 | case X86::BI__builtin_ia32_kshiftriqi: |
11097 | case X86::BI__builtin_ia32_kshiftrihi: |
11098 | case X86::BI__builtin_ia32_kshiftrisi: |
11099 | case X86::BI__builtin_ia32_kshiftridi: { |
11100 | unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() & 0xff; |
11101 | unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth(); |
11102 | |
11103 | if (ShiftVal >= NumElts) |
11104 | return llvm::Constant::getNullValue(Ops[0]->getType()); |
11105 | |
11106 | Value *In = getMaskVecValue(*this, Ops[0], NumElts); |
11107 | |
11108 | uint32_t Indices[64]; |
11109 | for (unsigned i = 0; i != NumElts; ++i) |
11110 | Indices[i] = i + ShiftVal; |
11111 | |
11112 | Value *Zero = llvm::Constant::getNullValue(In->getType()); |
11113 | Value *SV = Builder.CreateShuffleVector(In, Zero, |
11114 | makeArrayRef(Indices, NumElts), |
11115 | "kshiftr"); |
11116 | return Builder.CreateBitCast(SV, Ops[0]->getType()); |
11117 | } |
11118 | case X86::BI__builtin_ia32_movnti: |
11119 | case X86::BI__builtin_ia32_movnti64: |
11120 | case X86::BI__builtin_ia32_movntsd: |
11121 | case X86::BI__builtin_ia32_movntss: { |
11122 | llvm::MDNode *Node = llvm::MDNode::get( |
11123 | getLLVMContext(), llvm::ConstantAsMetadata::get(Builder.getInt32(1))); |
11124 | |
11125 | Value *Ptr = Ops[0]; |
11126 | Value *Src = Ops[1]; |
11127 | |
11128 | // Extract the 0'th element of the source vector. |
11129 | if (BuiltinID == X86::BI__builtin_ia32_movntsd || |
11130 | BuiltinID == X86::BI__builtin_ia32_movntss) |
11131 | Src = Builder.CreateExtractElement(Src, (uint64_t)0, "extract"); |
11132 | |
11133 | // Convert the type of the pointer to a pointer to the stored type. |
11134 | Value *BC = Builder.CreateBitCast( |
11135 | Ptr, llvm::PointerType::getUnqual(Src->getType()), "cast"); |
11136 | |
11137 | // Unaligned nontemporal store of the scalar value. |
11138 | StoreInst *SI = Builder.CreateDefaultAlignedStore(Src, BC); |
11139 | SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node); |
11140 | SI->setAlignment(1); |
11141 | return SI; |
11142 | } |
11143 | // Rotate is a special case of funnel shift - 1st 2 args are the same. |
11144 | case X86::BI__builtin_ia32_vprotb: |
11145 | case X86::BI__builtin_ia32_vprotw: |
11146 | case X86::BI__builtin_ia32_vprotd: |
11147 | case X86::BI__builtin_ia32_vprotq: |
11148 | case X86::BI__builtin_ia32_vprotbi: |
11149 | case X86::BI__builtin_ia32_vprotwi: |
11150 | case X86::BI__builtin_ia32_vprotdi: |
11151 | case X86::BI__builtin_ia32_vprotqi: |
11152 | case X86::BI__builtin_ia32_prold128: |
11153 | case X86::BI__builtin_ia32_prold256: |
11154 | case X86::BI__builtin_ia32_prold512: |
11155 | case X86::BI__builtin_ia32_prolq128: |
11156 | case X86::BI__builtin_ia32_prolq256: |
11157 | case X86::BI__builtin_ia32_prolq512: |
11158 | case X86::BI__builtin_ia32_prolvd128: |
11159 | case X86::BI__builtin_ia32_prolvd256: |
11160 | case X86::BI__builtin_ia32_prolvd512: |
11161 | case X86::BI__builtin_ia32_prolvq128: |
11162 | case X86::BI__builtin_ia32_prolvq256: |
11163 | case X86::BI__builtin_ia32_prolvq512: |
11164 | return EmitX86FunnelShift(*this, Ops[0], Ops[0], Ops[1], false); |
11165 | case X86::BI__builtin_ia32_prord128: |
11166 | case X86::BI__builtin_ia32_prord256: |
11167 | case X86::BI__builtin_ia32_prord512: |
11168 | case X86::BI__builtin_ia32_prorq128: |
11169 | case X86::BI__builtin_ia32_prorq256: |
11170 | case X86::BI__builtin_ia32_prorq512: |
11171 | case X86::BI__builtin_ia32_prorvd128: |
11172 | case X86::BI__builtin_ia32_prorvd256: |
11173 | case X86::BI__builtin_ia32_prorvd512: |
11174 | case X86::BI__builtin_ia32_prorvq128: |
11175 | case X86::BI__builtin_ia32_prorvq256: |
11176 | case X86::BI__builtin_ia32_prorvq512: |
11177 | return EmitX86FunnelShift(*this, Ops[0], Ops[0], Ops[1], true); |
11178 | case X86::BI__builtin_ia32_selectb_128: |
11179 | case X86::BI__builtin_ia32_selectb_256: |
11180 | case X86::BI__builtin_ia32_selectb_512: |
11181 | case X86::BI__builtin_ia32_selectw_128: |
11182 | case X86::BI__builtin_ia32_selectw_256: |
11183 | case X86::BI__builtin_ia32_selectw_512: |
11184 | case X86::BI__builtin_ia32_selectd_128: |
11185 | case X86::BI__builtin_ia32_selectd_256: |
11186 | case X86::BI__builtin_ia32_selectd_512: |
11187 | case X86::BI__builtin_ia32_selectq_128: |
11188 | case X86::BI__builtin_ia32_selectq_256: |
11189 | case X86::BI__builtin_ia32_selectq_512: |
11190 | case X86::BI__builtin_ia32_selectps_128: |
11191 | case X86::BI__builtin_ia32_selectps_256: |
11192 | case X86::BI__builtin_ia32_selectps_512: |
11193 | case X86::BI__builtin_ia32_selectpd_128: |
11194 | case X86::BI__builtin_ia32_selectpd_256: |
11195 | case X86::BI__builtin_ia32_selectpd_512: |
11196 | return EmitX86Select(*this, Ops[0], Ops[1], Ops[2]); |
11197 | case X86::BI__builtin_ia32_selectss_128: |
11198 | case X86::BI__builtin_ia32_selectsd_128: { |
11199 | Value *A = Builder.CreateExtractElement(Ops[1], (uint64_t)0); |
11200 | Value *B = Builder.CreateExtractElement(Ops[2], (uint64_t)0); |
11201 | A = EmitX86ScalarSelect(*this, Ops[0], A, B); |
11202 | return Builder.CreateInsertElement(Ops[1], A, (uint64_t)0); |
11203 | } |
11204 | case X86::BI__builtin_ia32_cmpb128_mask: |
11205 | case X86::BI__builtin_ia32_cmpb256_mask: |
11206 | case X86::BI__builtin_ia32_cmpb512_mask: |
11207 | case X86::BI__builtin_ia32_cmpw128_mask: |
11208 | case X86::BI__builtin_ia32_cmpw256_mask: |
11209 | case X86::BI__builtin_ia32_cmpw512_mask: |
11210 | case X86::BI__builtin_ia32_cmpd128_mask: |
11211 | case X86::BI__builtin_ia32_cmpd256_mask: |
11212 | case X86::BI__builtin_ia32_cmpd512_mask: |
11213 | case X86::BI__builtin_ia32_cmpq128_mask: |
11214 | case X86::BI__builtin_ia32_cmpq256_mask: |
11215 | case X86::BI__builtin_ia32_cmpq512_mask: { |
11216 | unsigned CC = cast<llvm::ConstantInt>(Ops[2])->getZExtValue() & 0x7; |
11217 | return EmitX86MaskedCompare(*this, CC, true, Ops); |
11218 | } |
11219 | case X86::BI__builtin_ia32_ucmpb128_mask: |
11220 | case X86::BI__builtin_ia32_ucmpb256_mask: |
11221 | case X86::BI__builtin_ia32_ucmpb512_mask: |
11222 | case X86::BI__builtin_ia32_ucmpw128_mask: |
11223 | case X86::BI__builtin_ia32_ucmpw256_mask: |
11224 | case X86::BI__builtin_ia32_ucmpw512_mask: |
11225 | case X86::BI__builtin_ia32_ucmpd128_mask: |
11226 | case X86::BI__builtin_ia32_ucmpd256_mask: |
11227 | case X86::BI__builtin_ia32_ucmpd512_mask: |
11228 | case X86::BI__builtin_ia32_ucmpq128_mask: |
11229 | case X86::BI__builtin_ia32_ucmpq256_mask: |
11230 | case X86::BI__builtin_ia32_ucmpq512_mask: { |
11231 | unsigned CC = cast<llvm::ConstantInt>(Ops[2])->getZExtValue() & 0x7; |
11232 | return EmitX86MaskedCompare(*this, CC, false, Ops); |
11233 | } |
11234 | case X86::BI__builtin_ia32_vpcomb: |
11235 | case X86::BI__builtin_ia32_vpcomw: |
11236 | case X86::BI__builtin_ia32_vpcomd: |
11237 | case X86::BI__builtin_ia32_vpcomq: |
11238 | return EmitX86vpcom(*this, Ops, true); |
11239 | case X86::BI__builtin_ia32_vpcomub: |
11240 | case X86::BI__builtin_ia32_vpcomuw: |
11241 | case X86::BI__builtin_ia32_vpcomud: |
11242 | case X86::BI__builtin_ia32_vpcomuq: |
11243 | return EmitX86vpcom(*this, Ops, false); |
11244 | |
11245 | case X86::BI__builtin_ia32_kortestcqi: |
11246 | case X86::BI__builtin_ia32_kortestchi: |
11247 | case X86::BI__builtin_ia32_kortestcsi: |
11248 | case X86::BI__builtin_ia32_kortestcdi: { |
11249 | Value *Or = EmitX86MaskLogic(*this, Instruction::Or, Ops); |
11250 | Value *C = llvm::Constant::getAllOnesValue(Ops[0]->getType()); |
11251 | Value *Cmp = Builder.CreateICmpEQ(Or, C); |
11252 | return Builder.CreateZExt(Cmp, ConvertType(E->getType())); |
11253 | } |
11254 | case X86::BI__builtin_ia32_kortestzqi: |
11255 | case X86::BI__builtin_ia32_kortestzhi: |
11256 | case X86::BI__builtin_ia32_kortestzsi: |
11257 | case X86::BI__builtin_ia32_kortestzdi: { |
11258 | Value *Or = EmitX86MaskLogic(*this, Instruction::Or, Ops); |
11259 | Value *C = llvm::Constant::getNullValue(Ops[0]->getType()); |
11260 | Value *Cmp = Builder.CreateICmpEQ(Or, C); |
11261 | return Builder.CreateZExt(Cmp, ConvertType(E->getType())); |
11262 | } |
11263 | |
11264 | case X86::BI__builtin_ia32_ktestcqi: |
11265 | case X86::BI__builtin_ia32_ktestzqi: |
11266 | case X86::BI__builtin_ia32_ktestchi: |
11267 | case X86::BI__builtin_ia32_ktestzhi: |
11268 | case X86::BI__builtin_ia32_ktestcsi: |
11269 | case X86::BI__builtin_ia32_ktestzsi: |
11270 | case X86::BI__builtin_ia32_ktestcdi: |
11271 | case X86::BI__builtin_ia32_ktestzdi: { |
11272 | Intrinsic::ID IID; |
11273 | switch (BuiltinID) { |
11274 | default: llvm_unreachable("Unsupported intrinsic!")::llvm::llvm_unreachable_internal("Unsupported intrinsic!", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 11274); |
11275 | case X86::BI__builtin_ia32_ktestcqi: |
11276 | IID = Intrinsic::x86_avx512_ktestc_b; |
11277 | break; |
11278 | case X86::BI__builtin_ia32_ktestzqi: |
11279 | IID = Intrinsic::x86_avx512_ktestz_b; |
11280 | break; |
11281 | case X86::BI__builtin_ia32_ktestchi: |
11282 | IID = Intrinsic::x86_avx512_ktestc_w; |
11283 | break; |
11284 | case X86::BI__builtin_ia32_ktestzhi: |
11285 | IID = Intrinsic::x86_avx512_ktestz_w; |
11286 | break; |
11287 | case X86::BI__builtin_ia32_ktestcsi: |
11288 | IID = Intrinsic::x86_avx512_ktestc_d; |
11289 | break; |
11290 | case X86::BI__builtin_ia32_ktestzsi: |
11291 | IID = Intrinsic::x86_avx512_ktestz_d; |
11292 | break; |
11293 | case X86::BI__builtin_ia32_ktestcdi: |
11294 | IID = Intrinsic::x86_avx512_ktestc_q; |
11295 | break; |
11296 | case X86::BI__builtin_ia32_ktestzdi: |
11297 | IID = Intrinsic::x86_avx512_ktestz_q; |
11298 | break; |
11299 | } |
11300 | |
11301 | unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth(); |
11302 | Value *LHS = getMaskVecValue(*this, Ops[0], NumElts); |
11303 | Value *RHS = getMaskVecValue(*this, Ops[1], NumElts); |
11304 | Function *Intr = CGM.getIntrinsic(IID); |
11305 | return Builder.CreateCall(Intr, {LHS, RHS}); |
11306 | } |
11307 | |
11308 | case X86::BI__builtin_ia32_kaddqi: |
11309 | case X86::BI__builtin_ia32_kaddhi: |
11310 | case X86::BI__builtin_ia32_kaddsi: |
11311 | case X86::BI__builtin_ia32_kadddi: { |
11312 | Intrinsic::ID IID; |
11313 | switch (BuiltinID) { |
11314 | default: llvm_unreachable("Unsupported intrinsic!")::llvm::llvm_unreachable_internal("Unsupported intrinsic!", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 11314); |
11315 | case X86::BI__builtin_ia32_kaddqi: |
11316 | IID = Intrinsic::x86_avx512_kadd_b; |
11317 | break; |
11318 | case X86::BI__builtin_ia32_kaddhi: |
11319 | IID = Intrinsic::x86_avx512_kadd_w; |
11320 | break; |
11321 | case X86::BI__builtin_ia32_kaddsi: |
11322 | IID = Intrinsic::x86_avx512_kadd_d; |
11323 | break; |
11324 | case X86::BI__builtin_ia32_kadddi: |
11325 | IID = Intrinsic::x86_avx512_kadd_q; |
11326 | break; |
11327 | } |
11328 | |
11329 | unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth(); |
11330 | Value *LHS = getMaskVecValue(*this, Ops[0], NumElts); |
11331 | Value *RHS = getMaskVecValue(*this, Ops[1], NumElts); |
11332 | Function *Intr = CGM.getIntrinsic(IID); |
11333 | Value *Res = Builder.CreateCall(Intr, {LHS, RHS}); |
11334 | return Builder.CreateBitCast(Res, Ops[0]->getType()); |
11335 | } |
11336 | case X86::BI__builtin_ia32_kandqi: |
11337 | case X86::BI__builtin_ia32_kandhi: |
11338 | case X86::BI__builtin_ia32_kandsi: |
11339 | case X86::BI__builtin_ia32_kanddi: |
11340 | return EmitX86MaskLogic(*this, Instruction::And, Ops); |
11341 | case X86::BI__builtin_ia32_kandnqi: |
11342 | case X86::BI__builtin_ia32_kandnhi: |
11343 | case X86::BI__builtin_ia32_kandnsi: |
11344 | case X86::BI__builtin_ia32_kandndi: |
11345 | return EmitX86MaskLogic(*this, Instruction::And, Ops, true); |
11346 | case X86::BI__builtin_ia32_korqi: |
11347 | case X86::BI__builtin_ia32_korhi: |
11348 | case X86::BI__builtin_ia32_korsi: |
11349 | case X86::BI__builtin_ia32_kordi: |
11350 | return EmitX86MaskLogic(*this, Instruction::Or, Ops); |
11351 | case X86::BI__builtin_ia32_kxnorqi: |
11352 | case X86::BI__builtin_ia32_kxnorhi: |
11353 | case X86::BI__builtin_ia32_kxnorsi: |
11354 | case X86::BI__builtin_ia32_kxnordi: |
11355 | return EmitX86MaskLogic(*this, Instruction::Xor, Ops, true); |
11356 | case X86::BI__builtin_ia32_kxorqi: |
11357 | case X86::BI__builtin_ia32_kxorhi: |
11358 | case X86::BI__builtin_ia32_kxorsi: |
11359 | case X86::BI__builtin_ia32_kxordi: |
11360 | return EmitX86MaskLogic(*this, Instruction::Xor, Ops); |
11361 | case X86::BI__builtin_ia32_knotqi: |
11362 | case X86::BI__builtin_ia32_knothi: |
11363 | case X86::BI__builtin_ia32_knotsi: |
11364 | case X86::BI__builtin_ia32_knotdi: { |
11365 | unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth(); |
11366 | Value *Res = getMaskVecValue(*this, Ops[0], NumElts); |
11367 | return Builder.CreateBitCast(Builder.CreateNot(Res), |
11368 | Ops[0]->getType()); |
11369 | } |
11370 | case X86::BI__builtin_ia32_kmovb: |
11371 | case X86::BI__builtin_ia32_kmovw: |
11372 | case X86::BI__builtin_ia32_kmovd: |
11373 | case X86::BI__builtin_ia32_kmovq: { |
11374 | // Bitcast to vXi1 type and then back to integer. This gets the mask |
11375 | // register type into the IR, but might be optimized out depending on |
11376 | // what's around it. |
11377 | unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth(); |
11378 | Value *Res = getMaskVecValue(*this, Ops[0], NumElts); |
11379 | return Builder.CreateBitCast(Res, Ops[0]->getType()); |
11380 | } |
11381 | |
11382 | case X86::BI__builtin_ia32_kunpckdi: |
11383 | case X86::BI__builtin_ia32_kunpcksi: |
11384 | case X86::BI__builtin_ia32_kunpckhi: { |
11385 | unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth(); |
11386 | Value *LHS = getMaskVecValue(*this, Ops[0], NumElts); |
11387 | Value *RHS = getMaskVecValue(*this, Ops[1], NumElts); |
11388 | uint32_t Indices[64]; |
11389 | for (unsigned i = 0; i != NumElts; ++i) |
11390 | Indices[i] = i; |
11391 | |
11392 | // First extract half of each vector. This gives better codegen than |
11393 | // doing it in a single shuffle. |
11394 | LHS = Builder.CreateShuffleVector(LHS, LHS, |
11395 | makeArrayRef(Indices, NumElts / 2)); |
11396 | RHS = Builder.CreateShuffleVector(RHS, RHS, |
11397 | makeArrayRef(Indices, NumElts / 2)); |
11398 | // Concat the vectors. |
11399 | // NOTE: Operands are swapped to match the intrinsic definition. |
11400 | Value *Res = Builder.CreateShuffleVector(RHS, LHS, |
11401 | makeArrayRef(Indices, NumElts)); |
11402 | return Builder.CreateBitCast(Res, Ops[0]->getType()); |
11403 | } |
11404 | |
11405 | case X86::BI__builtin_ia32_vplzcntd_128: |
11406 | case X86::BI__builtin_ia32_vplzcntd_256: |
11407 | case X86::BI__builtin_ia32_vplzcntd_512: |
11408 | case X86::BI__builtin_ia32_vplzcntq_128: |
11409 | case X86::BI__builtin_ia32_vplzcntq_256: |
11410 | case X86::BI__builtin_ia32_vplzcntq_512: { |
11411 | Function *F = CGM.getIntrinsic(Intrinsic::ctlz, Ops[0]->getType()); |
11412 | return Builder.CreateCall(F, {Ops[0],Builder.getInt1(false)}); |
11413 | } |
11414 | case X86::BI__builtin_ia32_sqrtss: |
11415 | case X86::BI__builtin_ia32_sqrtsd: { |
11416 | Value *A = Builder.CreateExtractElement(Ops[0], (uint64_t)0); |
11417 | Function *F = CGM.getIntrinsic(Intrinsic::sqrt, A->getType()); |
11418 | A = Builder.CreateCall(F, {A}); |
11419 | return Builder.CreateInsertElement(Ops[0], A, (uint64_t)0); |
11420 | } |
11421 | case X86::BI__builtin_ia32_sqrtsd_round_mask: |
11422 | case X86::BI__builtin_ia32_sqrtss_round_mask: { |
11423 | unsigned CC = cast<llvm::ConstantInt>(Ops[4])->getZExtValue(); |
11424 | // Support only if the rounding mode is 4 (AKA CUR_DIRECTION), |
11425 | // otherwise keep the intrinsic. |
11426 | if (CC != 4) { |
11427 | Intrinsic::ID IID = BuiltinID == X86::BI__builtin_ia32_sqrtsd_round_mask ? |
11428 | Intrinsic::x86_avx512_mask_sqrt_sd : |
11429 | Intrinsic::x86_avx512_mask_sqrt_ss; |
11430 | return Builder.CreateCall(CGM.getIntrinsic(IID), Ops); |
11431 | } |
11432 | Value *A = Builder.CreateExtractElement(Ops[1], (uint64_t)0); |
11433 | Function *F = CGM.getIntrinsic(Intrinsic::sqrt, A->getType()); |
11434 | A = Builder.CreateCall(F, A); |
11435 | Value *Src = Builder.CreateExtractElement(Ops[2], (uint64_t)0); |
11436 | A = EmitX86ScalarSelect(*this, Ops[3], A, Src); |
11437 | return Builder.CreateInsertElement(Ops[0], A, (uint64_t)0); |
11438 | } |
11439 | case X86::BI__builtin_ia32_sqrtpd256: |
11440 | case X86::BI__builtin_ia32_sqrtpd: |
11441 | case X86::BI__builtin_ia32_sqrtps256: |
11442 | case X86::BI__builtin_ia32_sqrtps: |
11443 | case X86::BI__builtin_ia32_sqrtps512: |
11444 | case X86::BI__builtin_ia32_sqrtpd512: { |
11445 | if (Ops.size() == 2) { |
11446 | unsigned CC = cast<llvm::ConstantInt>(Ops[1])->getZExtValue(); |
11447 | // Support only if the rounding mode is 4 (AKA CUR_DIRECTION), |
11448 | // otherwise keep the intrinsic. |
11449 | if (CC != 4) { |
11450 | Intrinsic::ID IID = BuiltinID == X86::BI__builtin_ia32_sqrtps512 ? |
11451 | Intrinsic::x86_avx512_sqrt_ps_512 : |
11452 | Intrinsic::x86_avx512_sqrt_pd_512; |
11453 | return Builder.CreateCall(CGM.getIntrinsic(IID), Ops); |
11454 | } |
11455 | } |
11456 | Function *F = CGM.getIntrinsic(Intrinsic::sqrt, Ops[0]->getType()); |
11457 | return Builder.CreateCall(F, Ops[0]); |
11458 | } |
11459 | case X86::BI__builtin_ia32_pabsb128: |
11460 | case X86::BI__builtin_ia32_pabsw128: |
11461 | case X86::BI__builtin_ia32_pabsd128: |
11462 | case X86::BI__builtin_ia32_pabsb256: |
11463 | case X86::BI__builtin_ia32_pabsw256: |
11464 | case X86::BI__builtin_ia32_pabsd256: |
11465 | case X86::BI__builtin_ia32_pabsq128: |
11466 | case X86::BI__builtin_ia32_pabsq256: |
11467 | case X86::BI__builtin_ia32_pabsb512: |
11468 | case X86::BI__builtin_ia32_pabsw512: |
11469 | case X86::BI__builtin_ia32_pabsd512: |
11470 | case X86::BI__builtin_ia32_pabsq512: |
11471 | return EmitX86Abs(*this, Ops); |
11472 | |
11473 | case X86::BI__builtin_ia32_pmaxsb128: |
11474 | case X86::BI__builtin_ia32_pmaxsw128: |
11475 | case X86::BI__builtin_ia32_pmaxsd128: |
11476 | case X86::BI__builtin_ia32_pmaxsq128: |
11477 | case X86::BI__builtin_ia32_pmaxsb256: |
11478 | case X86::BI__builtin_ia32_pmaxsw256: |
11479 | case X86::BI__builtin_ia32_pmaxsd256: |
11480 | case X86::BI__builtin_ia32_pmaxsq256: |
11481 | case X86::BI__builtin_ia32_pmaxsb512: |
11482 | case X86::BI__builtin_ia32_pmaxsw512: |
11483 | case X86::BI__builtin_ia32_pmaxsd512: |
11484 | case X86::BI__builtin_ia32_pmaxsq512: |
11485 | return EmitX86MinMax(*this, ICmpInst::ICMP_SGT, Ops); |
11486 | case X86::BI__builtin_ia32_pmaxub128: |
11487 | case X86::BI__builtin_ia32_pmaxuw128: |
11488 | case X86::BI__builtin_ia32_pmaxud128: |
11489 | case X86::BI__builtin_ia32_pmaxuq128: |
11490 | case X86::BI__builtin_ia32_pmaxub256: |
11491 | case X86::BI__builtin_ia32_pmaxuw256: |
11492 | case X86::BI__builtin_ia32_pmaxud256: |
11493 | case X86::BI__builtin_ia32_pmaxuq256: |
11494 | case X86::BI__builtin_ia32_pmaxub512: |
11495 | case X86::BI__builtin_ia32_pmaxuw512: |
11496 | case X86::BI__builtin_ia32_pmaxud512: |
11497 | case X86::BI__builtin_ia32_pmaxuq512: |
11498 | return EmitX86MinMax(*this, ICmpInst::ICMP_UGT, Ops); |
11499 | case X86::BI__builtin_ia32_pminsb128: |
11500 | case X86::BI__builtin_ia32_pminsw128: |
11501 | case X86::BI__builtin_ia32_pminsd128: |
11502 | case X86::BI__builtin_ia32_pminsq128: |
11503 | case X86::BI__builtin_ia32_pminsb256: |
11504 | case X86::BI__builtin_ia32_pminsw256: |
11505 | case X86::BI__builtin_ia32_pminsd256: |
11506 | case X86::BI__builtin_ia32_pminsq256: |
11507 | case X86::BI__builtin_ia32_pminsb512: |
11508 | case X86::BI__builtin_ia32_pminsw512: |
11509 | case X86::BI__builtin_ia32_pminsd512: |
11510 | case X86::BI__builtin_ia32_pminsq512: |
11511 | return EmitX86MinMax(*this, ICmpInst::ICMP_SLT, Ops); |
11512 | case X86::BI__builtin_ia32_pminub128: |
11513 | case X86::BI__builtin_ia32_pminuw128: |
11514 | case X86::BI__builtin_ia32_pminud128: |
11515 | case X86::BI__builtin_ia32_pminuq128: |
11516 | case X86::BI__builtin_ia32_pminub256: |
11517 | case X86::BI__builtin_ia32_pminuw256: |
11518 | case X86::BI__builtin_ia32_pminud256: |
11519 | case X86::BI__builtin_ia32_pminuq256: |
11520 | case X86::BI__builtin_ia32_pminub512: |
11521 | case X86::BI__builtin_ia32_pminuw512: |
11522 | case X86::BI__builtin_ia32_pminud512: |
11523 | case X86::BI__builtin_ia32_pminuq512: |
11524 | return EmitX86MinMax(*this, ICmpInst::ICMP_ULT, Ops); |
11525 | |
11526 | case X86::BI__builtin_ia32_pmuludq128: |
11527 | case X86::BI__builtin_ia32_pmuludq256: |
11528 | case X86::BI__builtin_ia32_pmuludq512: |
11529 | return EmitX86Muldq(*this, /*IsSigned*/false, Ops); |
11530 | |
11531 | case X86::BI__builtin_ia32_pmuldq128: |
11532 | case X86::BI__builtin_ia32_pmuldq256: |
11533 | case X86::BI__builtin_ia32_pmuldq512: |
11534 | return EmitX86Muldq(*this, /*IsSigned*/true, Ops); |
11535 | |
11536 | case X86::BI__builtin_ia32_pternlogd512_mask: |
11537 | case X86::BI__builtin_ia32_pternlogq512_mask: |
11538 | case X86::BI__builtin_ia32_pternlogd128_mask: |
11539 | case X86::BI__builtin_ia32_pternlogd256_mask: |
11540 | case X86::BI__builtin_ia32_pternlogq128_mask: |
11541 | case X86::BI__builtin_ia32_pternlogq256_mask: |
11542 | return EmitX86Ternlog(*this, /*ZeroMask*/false, Ops); |
11543 | |
11544 | case X86::BI__builtin_ia32_pternlogd512_maskz: |
11545 | case X86::BI__builtin_ia32_pternlogq512_maskz: |
11546 | case X86::BI__builtin_ia32_pternlogd128_maskz: |
11547 | case X86::BI__builtin_ia32_pternlogd256_maskz: |
11548 | case X86::BI__builtin_ia32_pternlogq128_maskz: |
11549 | case X86::BI__builtin_ia32_pternlogq256_maskz: |
11550 | return EmitX86Ternlog(*this, /*ZeroMask*/true, Ops); |
11551 | |
11552 | case X86::BI__builtin_ia32_vpshldd128: |
11553 | case X86::BI__builtin_ia32_vpshldd256: |
11554 | case X86::BI__builtin_ia32_vpshldd512: |
11555 | case X86::BI__builtin_ia32_vpshldq128: |
11556 | case X86::BI__builtin_ia32_vpshldq256: |
11557 | case X86::BI__builtin_ia32_vpshldq512: |
11558 | case X86::BI__builtin_ia32_vpshldw128: |
11559 | case X86::BI__builtin_ia32_vpshldw256: |
11560 | case X86::BI__builtin_ia32_vpshldw512: |
11561 | return EmitX86FunnelShift(*this, Ops[0], Ops[1], Ops[2], false); |
11562 | |
11563 | case X86::BI__builtin_ia32_vpshrdd128: |
11564 | case X86::BI__builtin_ia32_vpshrdd256: |
11565 | case X86::BI__builtin_ia32_vpshrdd512: |
11566 | case X86::BI__builtin_ia32_vpshrdq128: |
11567 | case X86::BI__builtin_ia32_vpshrdq256: |
11568 | case X86::BI__builtin_ia32_vpshrdq512: |
11569 | case X86::BI__builtin_ia32_vpshrdw128: |
11570 | case X86::BI__builtin_ia32_vpshrdw256: |
11571 | case X86::BI__builtin_ia32_vpshrdw512: |
11572 | // Ops 0 and 1 are swapped. |
11573 | return EmitX86FunnelShift(*this, Ops[1], Ops[0], Ops[2], true); |
11574 | |
11575 | case X86::BI__builtin_ia32_vpshldvd128: |
11576 | case X86::BI__builtin_ia32_vpshldvd256: |
11577 | case X86::BI__builtin_ia32_vpshldvd512: |
11578 | case X86::BI__builtin_ia32_vpshldvq128: |
11579 | case X86::BI__builtin_ia32_vpshldvq256: |
11580 | case X86::BI__builtin_ia32_vpshldvq512: |
11581 | case X86::BI__builtin_ia32_vpshldvw128: |
11582 | case X86::BI__builtin_ia32_vpshldvw256: |
11583 | case X86::BI__builtin_ia32_vpshldvw512: |
11584 | return EmitX86FunnelShift(*this, Ops[0], Ops[1], Ops[2], false); |
11585 | |
11586 | case X86::BI__builtin_ia32_vpshrdvd128: |
11587 | case X86::BI__builtin_ia32_vpshrdvd256: |
11588 | case X86::BI__builtin_ia32_vpshrdvd512: |
11589 | case X86::BI__builtin_ia32_vpshrdvq128: |
11590 | case X86::BI__builtin_ia32_vpshrdvq256: |
11591 | case X86::BI__builtin_ia32_vpshrdvq512: |
11592 | case X86::BI__builtin_ia32_vpshrdvw128: |
11593 | case X86::BI__builtin_ia32_vpshrdvw256: |
11594 | case X86::BI__builtin_ia32_vpshrdvw512: |
11595 | // Ops 0 and 1 are swapped. |
11596 | return EmitX86FunnelShift(*this, Ops[1], Ops[0], Ops[2], true); |
11597 | |
11598 | // 3DNow! |
11599 | case X86::BI__builtin_ia32_pswapdsf: |
11600 | case X86::BI__builtin_ia32_pswapdsi: { |
11601 | llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext()); |
11602 | Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast"); |
11603 | llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_3dnowa_pswapd); |
11604 | return Builder.CreateCall(F, Ops, "pswapd"); |
11605 | } |
11606 | case X86::BI__builtin_ia32_rdrand16_step: |
11607 | case X86::BI__builtin_ia32_rdrand32_step: |
11608 | case X86::BI__builtin_ia32_rdrand64_step: |
11609 | case X86::BI__builtin_ia32_rdseed16_step: |
11610 | case X86::BI__builtin_ia32_rdseed32_step: |
11611 | case X86::BI__builtin_ia32_rdseed64_step: { |
11612 | Intrinsic::ID ID; |
11613 | switch (BuiltinID) { |
11614 | default: llvm_unreachable("Unsupported intrinsic!")::llvm::llvm_unreachable_internal("Unsupported intrinsic!", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 11614); |
11615 | case X86::BI__builtin_ia32_rdrand16_step: |
11616 | ID = Intrinsic::x86_rdrand_16; |
11617 | break; |
11618 | case X86::BI__builtin_ia32_rdrand32_step: |
11619 | ID = Intrinsic::x86_rdrand_32; |
11620 | break; |
11621 | case X86::BI__builtin_ia32_rdrand64_step: |
11622 | ID = Intrinsic::x86_rdrand_64; |
11623 | break; |
11624 | case X86::BI__builtin_ia32_rdseed16_step: |
11625 | ID = Intrinsic::x86_rdseed_16; |
11626 | break; |
11627 | case X86::BI__builtin_ia32_rdseed32_step: |
11628 | ID = Intrinsic::x86_rdseed_32; |
11629 | break; |
11630 | case X86::BI__builtin_ia32_rdseed64_step: |
11631 | ID = Intrinsic::x86_rdseed_64; |
11632 | break; |
11633 | } |
11634 | |
11635 | Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID)); |
11636 | Builder.CreateDefaultAlignedStore(Builder.CreateExtractValue(Call, 0), |
11637 | Ops[0]); |
11638 | return Builder.CreateExtractValue(Call, 1); |
11639 | } |
11640 | case X86::BI__builtin_ia32_addcarryx_u32: |
11641 | case X86::BI__builtin_ia32_addcarryx_u64: |
11642 | case X86::BI__builtin_ia32_subborrow_u32: |
11643 | case X86::BI__builtin_ia32_subborrow_u64: { |
11644 | Intrinsic::ID IID; |
11645 | switch (BuiltinID) { |
11646 | default: llvm_unreachable("Unsupported intrinsic!")::llvm::llvm_unreachable_internal("Unsupported intrinsic!", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 11646); |
11647 | case X86::BI__builtin_ia32_addcarryx_u32: |
11648 | IID = Intrinsic::x86_addcarry_32; |
11649 | break; |
11650 | case X86::BI__builtin_ia32_addcarryx_u64: |
11651 | IID = Intrinsic::x86_addcarry_64; |
11652 | break; |
11653 | case X86::BI__builtin_ia32_subborrow_u32: |
11654 | IID = Intrinsic::x86_subborrow_32; |
11655 | break; |
11656 | case X86::BI__builtin_ia32_subborrow_u64: |
11657 | IID = Intrinsic::x86_subborrow_64; |
11658 | break; |
11659 | } |
11660 | |
11661 | Value *Call = Builder.CreateCall(CGM.getIntrinsic(IID), |
11662 | { Ops[0], Ops[1], Ops[2] }); |
11663 | Builder.CreateDefaultAlignedStore(Builder.CreateExtractValue(Call, 1), |
11664 | Ops[3]); |
11665 | return Builder.CreateExtractValue(Call, 0); |
11666 | } |
11667 | |
11668 | case X86::BI__builtin_ia32_fpclassps128_mask: |
11669 | case X86::BI__builtin_ia32_fpclassps256_mask: |
11670 | case X86::BI__builtin_ia32_fpclassps512_mask: |
11671 | case X86::BI__builtin_ia32_fpclasspd128_mask: |
11672 | case X86::BI__builtin_ia32_fpclasspd256_mask: |
11673 | case X86::BI__builtin_ia32_fpclasspd512_mask: { |
11674 | unsigned NumElts = Ops[0]->getType()->getVectorNumElements(); |
11675 | Value *MaskIn = Ops[2]; |
11676 | Ops.erase(&Ops[2]); |
11677 | |
11678 | Intrinsic::ID ID; |
11679 | switch (BuiltinID) { |
11680 | default: llvm_unreachable("Unsupported intrinsic!")::llvm::llvm_unreachable_internal("Unsupported intrinsic!", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 11680); |
11681 | case X86::BI__builtin_ia32_fpclassps128_mask: |
11682 | ID = Intrinsic::x86_avx512_fpclass_ps_128; |
11683 | break; |
11684 | case X86::BI__builtin_ia32_fpclassps256_mask: |
11685 | ID = Intrinsic::x86_avx512_fpclass_ps_256; |
11686 | break; |
11687 | case X86::BI__builtin_ia32_fpclassps512_mask: |
11688 | ID = Intrinsic::x86_avx512_fpclass_ps_512; |
11689 | break; |
11690 | case X86::BI__builtin_ia32_fpclasspd128_mask: |
11691 | ID = Intrinsic::x86_avx512_fpclass_pd_128; |
11692 | break; |
11693 | case X86::BI__builtin_ia32_fpclasspd256_mask: |
11694 | ID = Intrinsic::x86_avx512_fpclass_pd_256; |
11695 | break; |
11696 | case X86::BI__builtin_ia32_fpclasspd512_mask: |
11697 | ID = Intrinsic::x86_avx512_fpclass_pd_512; |
11698 | break; |
11699 | } |
11700 | |
11701 | Value *Fpclass = Builder.CreateCall(CGM.getIntrinsic(ID), Ops); |
11702 | return EmitX86MaskedCompareResult(*this, Fpclass, NumElts, MaskIn); |
11703 | } |
11704 | |
11705 | case X86::BI__builtin_ia32_vpmultishiftqb128: |
11706 | case X86::BI__builtin_ia32_vpmultishiftqb256: |
11707 | case X86::BI__builtin_ia32_vpmultishiftqb512: { |
11708 | Intrinsic::ID ID; |
11709 | switch (BuiltinID) { |
11710 | default: llvm_unreachable("Unsupported intrinsic!")::llvm::llvm_unreachable_internal("Unsupported intrinsic!", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 11710); |
11711 | case X86::BI__builtin_ia32_vpmultishiftqb128: |
11712 | ID = Intrinsic::x86_avx512_pmultishift_qb_128; |
11713 | break; |
11714 | case X86::BI__builtin_ia32_vpmultishiftqb256: |
11715 | ID = Intrinsic::x86_avx512_pmultishift_qb_256; |
11716 | break; |
11717 | case X86::BI__builtin_ia32_vpmultishiftqb512: |
11718 | ID = Intrinsic::x86_avx512_pmultishift_qb_512; |
11719 | break; |
11720 | } |
11721 | |
11722 | return Builder.CreateCall(CGM.getIntrinsic(ID), Ops); |
11723 | } |
11724 | |
11725 | case X86::BI__builtin_ia32_vpshufbitqmb128_mask: |
11726 | case X86::BI__builtin_ia32_vpshufbitqmb256_mask: |
11727 | case X86::BI__builtin_ia32_vpshufbitqmb512_mask: { |
11728 | unsigned NumElts = Ops[0]->getType()->getVectorNumElements(); |
11729 | Value *MaskIn = Ops[2]; |
11730 | Ops.erase(&Ops[2]); |
11731 | |
11732 | Intrinsic::ID ID; |
11733 | switch (BuiltinID) { |
11734 | default: llvm_unreachable("Unsupported intrinsic!")::llvm::llvm_unreachable_internal("Unsupported intrinsic!", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 11734); |
11735 | case X86::BI__builtin_ia32_vpshufbitqmb128_mask: |
11736 | ID = Intrinsic::x86_avx512_vpshufbitqmb_128; |
11737 | break; |
11738 | case X86::BI__builtin_ia32_vpshufbitqmb256_mask: |
11739 | ID = Intrinsic::x86_avx512_vpshufbitqmb_256; |
11740 | break; |
11741 | case X86::BI__builtin_ia32_vpshufbitqmb512_mask: |
11742 | ID = Intrinsic::x86_avx512_vpshufbitqmb_512; |
11743 | break; |
11744 | } |
11745 | |
11746 | Value *Shufbit = Builder.CreateCall(CGM.getIntrinsic(ID), Ops); |
11747 | return EmitX86MaskedCompareResult(*this, Shufbit, NumElts, MaskIn); |
11748 | } |
11749 | |
11750 | // packed comparison intrinsics |
11751 | case X86::BI__builtin_ia32_cmpeqps: |
11752 | case X86::BI__builtin_ia32_cmpeqpd: |
11753 | return getVectorFCmpIR(CmpInst::FCMP_OEQ); |
11754 | case X86::BI__builtin_ia32_cmpltps: |
11755 | case X86::BI__builtin_ia32_cmpltpd: |
11756 | return getVectorFCmpIR(CmpInst::FCMP_OLT); |
11757 | case X86::BI__builtin_ia32_cmpleps: |
11758 | case X86::BI__builtin_ia32_cmplepd: |
11759 | return getVectorFCmpIR(CmpInst::FCMP_OLE); |
11760 | case X86::BI__builtin_ia32_cmpunordps: |
11761 | case X86::BI__builtin_ia32_cmpunordpd: |
11762 | return getVectorFCmpIR(CmpInst::FCMP_UNO); |
11763 | case X86::BI__builtin_ia32_cmpneqps: |
11764 | case X86::BI__builtin_ia32_cmpneqpd: |
11765 | return getVectorFCmpIR(CmpInst::FCMP_UNE); |
11766 | case X86::BI__builtin_ia32_cmpnltps: |
11767 | case X86::BI__builtin_ia32_cmpnltpd: |
11768 | return getVectorFCmpIR(CmpInst::FCMP_UGE); |
11769 | case X86::BI__builtin_ia32_cmpnleps: |
11770 | case X86::BI__builtin_ia32_cmpnlepd: |
11771 | return getVectorFCmpIR(CmpInst::FCMP_UGT); |
11772 | case X86::BI__builtin_ia32_cmpordps: |
11773 | case X86::BI__builtin_ia32_cmpordpd: |
11774 | return getVectorFCmpIR(CmpInst::FCMP_ORD); |
11775 | case X86::BI__builtin_ia32_cmpps: |
11776 | case X86::BI__builtin_ia32_cmpps256: |
11777 | case X86::BI__builtin_ia32_cmppd: |
11778 | case X86::BI__builtin_ia32_cmppd256: |
11779 | case X86::BI__builtin_ia32_cmpps128_mask: |
11780 | case X86::BI__builtin_ia32_cmpps256_mask: |
11781 | case X86::BI__builtin_ia32_cmpps512_mask: |
11782 | case X86::BI__builtin_ia32_cmppd128_mask: |
11783 | case X86::BI__builtin_ia32_cmppd256_mask: |
11784 | case X86::BI__builtin_ia32_cmppd512_mask: { |
11785 | // Lowering vector comparisons to fcmp instructions, while |
11786 | // ignoring signalling behaviour requested |
11787 | // ignoring rounding mode requested |
11788 | // This is is only possible as long as FENV_ACCESS is not implemented. |
11789 | // See also: https://reviews.llvm.org/D45616 |
11790 | |
11791 | // The third argument is the comparison condition, and integer in the |
11792 | // range [0, 31] |
11793 | unsigned CC = cast<llvm::ConstantInt>(Ops[2])->getZExtValue() & 0x1f; |
11794 | |
11795 | // Lowering to IR fcmp instruction. |
11796 | // Ignoring requested signaling behaviour, |
11797 | // e.g. both _CMP_GT_OS & _CMP_GT_OQ are translated to FCMP_OGT. |
11798 | FCmpInst::Predicate Pred; |
11799 | switch (CC) { |
11800 | case 0x00: Pred = FCmpInst::FCMP_OEQ; break; |
11801 | case 0x01: Pred = FCmpInst::FCMP_OLT; break; |
11802 | case 0x02: Pred = FCmpInst::FCMP_OLE; break; |
11803 | case 0x03: Pred = FCmpInst::FCMP_UNO; break; |
11804 | case 0x04: Pred = FCmpInst::FCMP_UNE; break; |
11805 | case 0x05: Pred = FCmpInst::FCMP_UGE; break; |
11806 | case 0x06: Pred = FCmpInst::FCMP_UGT; break; |
11807 | case 0x07: Pred = FCmpInst::FCMP_ORD; break; |
11808 | case 0x08: Pred = FCmpInst::FCMP_UEQ; break; |
11809 | case 0x09: Pred = FCmpInst::FCMP_ULT; break; |
11810 | case 0x0a: Pred = FCmpInst::FCMP_ULE; break; |
11811 | case 0x0b: Pred = FCmpInst::FCMP_FALSE; break; |
11812 | case 0x0c: Pred = FCmpInst::FCMP_ONE; break; |
11813 | case 0x0d: Pred = FCmpInst::FCMP_OGE; break; |
11814 | case 0x0e: Pred = FCmpInst::FCMP_OGT; break; |
11815 | case 0x0f: Pred = FCmpInst::FCMP_TRUE; break; |
11816 | case 0x10: Pred = FCmpInst::FCMP_OEQ; break; |
11817 | case 0x11: Pred = FCmpInst::FCMP_OLT; break; |
11818 | case 0x12: Pred = FCmpInst::FCMP_OLE; break; |
11819 | case 0x13: Pred = FCmpInst::FCMP_UNO; break; |
11820 | case 0x14: Pred = FCmpInst::FCMP_UNE; break; |
11821 | case 0x15: Pred = FCmpInst::FCMP_UGE; break; |
11822 | case 0x16: Pred = FCmpInst::FCMP_UGT; break; |
11823 | case 0x17: Pred = FCmpInst::FCMP_ORD; break; |
11824 | case 0x18: Pred = FCmpInst::FCMP_UEQ; break; |
11825 | case 0x19: Pred = FCmpInst::FCMP_ULT; break; |
11826 | case 0x1a: Pred = FCmpInst::FCMP_ULE; break; |
11827 | case 0x1b: Pred = FCmpInst::FCMP_FALSE; break; |
11828 | case 0x1c: Pred = FCmpInst::FCMP_ONE; break; |
11829 | case 0x1d: Pred = FCmpInst::FCMP_OGE; break; |
11830 | case 0x1e: Pred = FCmpInst::FCMP_OGT; break; |
11831 | case 0x1f: Pred = FCmpInst::FCMP_TRUE; break; |
11832 | default: llvm_unreachable("Unhandled CC")::llvm::llvm_unreachable_internal("Unhandled CC", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 11832); |
11833 | } |
11834 | |
11835 | // Builtins without the _mask suffix return a vector of integers |
11836 | // of the same width as the input vectors |
11837 | switch (BuiltinID) { |
11838 | case X86::BI__builtin_ia32_cmpps512_mask: |
11839 | case X86::BI__builtin_ia32_cmppd512_mask: |
11840 | case X86::BI__builtin_ia32_cmpps128_mask: |
11841 | case X86::BI__builtin_ia32_cmpps256_mask: |
11842 | case X86::BI__builtin_ia32_cmppd128_mask: |
11843 | case X86::BI__builtin_ia32_cmppd256_mask: { |
11844 | unsigned NumElts = Ops[0]->getType()->getVectorNumElements(); |
11845 | Value *Cmp = Builder.CreateFCmp(Pred, Ops[0], Ops[1]); |
11846 | return EmitX86MaskedCompareResult(*this, Cmp, NumElts, Ops[3]); |
11847 | } |
11848 | default: |
11849 | return getVectorFCmpIR(Pred); |
11850 | } |
11851 | } |
11852 | |
11853 | // SSE scalar comparison intrinsics |
11854 | case X86::BI__builtin_ia32_cmpeqss: |
11855 | return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 0); |
11856 | case X86::BI__builtin_ia32_cmpltss: |
11857 | return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 1); |
11858 | case X86::BI__builtin_ia32_cmpless: |
11859 | return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 2); |
11860 | case X86::BI__builtin_ia32_cmpunordss: |
11861 | return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 3); |
11862 | case X86::BI__builtin_ia32_cmpneqss: |
11863 | return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 4); |
11864 | case X86::BI__builtin_ia32_cmpnltss: |
11865 | return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 5); |
11866 | case X86::BI__builtin_ia32_cmpnless: |
11867 | return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 6); |
11868 | case X86::BI__builtin_ia32_cmpordss: |
11869 | return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 7); |
11870 | case X86::BI__builtin_ia32_cmpeqsd: |
11871 | return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 0); |
11872 | case X86::BI__builtin_ia32_cmpltsd: |
11873 | return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 1); |
11874 | case X86::BI__builtin_ia32_cmplesd: |
11875 | return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 2); |
11876 | case X86::BI__builtin_ia32_cmpunordsd: |
11877 | return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 3); |
11878 | case X86::BI__builtin_ia32_cmpneqsd: |
11879 | return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 4); |
11880 | case X86::BI__builtin_ia32_cmpnltsd: |
11881 | return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 5); |
11882 | case X86::BI__builtin_ia32_cmpnlesd: |
11883 | return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 6); |
11884 | case X86::BI__builtin_ia32_cmpordsd: |
11885 | return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 7); |
11886 | |
11887 | // AVX512 bf16 intrinsics |
11888 | case X86::BI__builtin_ia32_cvtneps2bf16_128_mask: { |
11889 | Ops[2] = getMaskVecValue(*this, Ops[2], |
11890 | Ops[0]->getType()->getVectorNumElements()); |
11891 | Intrinsic::ID IID = Intrinsic::x86_avx512bf16_mask_cvtneps2bf16_128; |
11892 | return Builder.CreateCall(CGM.getIntrinsic(IID), Ops); |
11893 | } |
11894 | |
11895 | case X86::BI__builtin_ia32_cvtneps2bf16_256_mask: |
11896 | case X86::BI__builtin_ia32_cvtneps2bf16_512_mask: { |
11897 | Intrinsic::ID IID; |
11898 | switch (BuiltinID) { |
11899 | default: llvm_unreachable("Unsupported intrinsic!")::llvm::llvm_unreachable_internal("Unsupported intrinsic!", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 11899); |
11900 | case X86::BI__builtin_ia32_cvtneps2bf16_256_mask: |
11901 | IID = Intrinsic::x86_avx512bf16_cvtneps2bf16_256; |
11902 | break; |
11903 | case X86::BI__builtin_ia32_cvtneps2bf16_512_mask: |
11904 | IID = Intrinsic::x86_avx512bf16_cvtneps2bf16_512; |
11905 | break; |
11906 | } |
11907 | Value *Res = Builder.CreateCall(CGM.getIntrinsic(IID), Ops[0]); |
11908 | return EmitX86Select(*this, Ops[2], Res, Ops[1]); |
11909 | } |
11910 | |
11911 | case X86::BI__emul: |
11912 | case X86::BI__emulu: { |
11913 | llvm::Type *Int64Ty = llvm::IntegerType::get(getLLVMContext(), 64); |
11914 | bool isSigned = (BuiltinID == X86::BI__emul); |
11915 | Value *LHS = Builder.CreateIntCast(Ops[0], Int64Ty, isSigned); |
11916 | Value *RHS = Builder.CreateIntCast(Ops[1], Int64Ty, isSigned); |
11917 | return Builder.CreateMul(LHS, RHS, "", !isSigned, isSigned); |
11918 | } |
11919 | case X86::BI__mulh: |
11920 | case X86::BI__umulh: |
11921 | case X86::BI_mul128: |
11922 | case X86::BI_umul128: { |
11923 | llvm::Type *ResType = ConvertType(E->getType()); |
11924 | llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128); |
11925 | |
11926 | bool IsSigned = (BuiltinID == X86::BI__mulh || BuiltinID == X86::BI_mul128); |
11927 | Value *LHS = Builder.CreateIntCast(Ops[0], Int128Ty, IsSigned); |
11928 | Value *RHS = Builder.CreateIntCast(Ops[1], Int128Ty, IsSigned); |
11929 | |
11930 | Value *MulResult, *HigherBits; |
11931 | if (IsSigned) { |
11932 | MulResult = Builder.CreateNSWMul(LHS, RHS); |
11933 | HigherBits = Builder.CreateAShr(MulResult, 64); |
11934 | } else { |
11935 | MulResult = Builder.CreateNUWMul(LHS, RHS); |
11936 | HigherBits = Builder.CreateLShr(MulResult, 64); |
11937 | } |
11938 | HigherBits = Builder.CreateIntCast(HigherBits, ResType, IsSigned); |
11939 | |
11940 | if (BuiltinID == X86::BI__mulh || BuiltinID == X86::BI__umulh) |
11941 | return HigherBits; |
11942 | |
11943 | Address HighBitsAddress = EmitPointerWithAlignment(E->getArg(2)); |
11944 | Builder.CreateStore(HigherBits, HighBitsAddress); |
11945 | return Builder.CreateIntCast(MulResult, ResType, IsSigned); |
11946 | } |
11947 | |
11948 | case X86::BI__faststorefence: { |
11949 | return Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent, |
11950 | llvm::SyncScope::System); |
11951 | } |
11952 | case X86::BI__shiftleft128: |
11953 | case X86::BI__shiftright128: { |
11954 | // FIXME: Once fshl/fshr no longer add an unneeded and and cmov, do this: |
11955 | // llvm::Function *F = CGM.getIntrinsic( |
11956 | // BuiltinID == X86::BI__shiftleft128 ? Intrinsic::fshl : Intrinsic::fshr, |
11957 | // Int64Ty); |
11958 | // Ops[2] = Builder.CreateZExt(Ops[2], Int64Ty); |
11959 | // return Builder.CreateCall(F, Ops); |
11960 | llvm::Type *Int128Ty = Builder.getInt128Ty(); |
11961 | Value *HighPart128 = |
11962 | Builder.CreateShl(Builder.CreateZExt(Ops[1], Int128Ty), 64); |
11963 | Value *LowPart128 = Builder.CreateZExt(Ops[0], Int128Ty); |
11964 | Value *Val = Builder.CreateOr(HighPart128, LowPart128); |
11965 | Value *Amt = Builder.CreateAnd(Builder.CreateZExt(Ops[2], Int128Ty), |
11966 | llvm::ConstantInt::get(Int128Ty, 0x3f)); |
11967 | Value *Res; |
11968 | if (BuiltinID == X86::BI__shiftleft128) |
11969 | Res = Builder.CreateLShr(Builder.CreateShl(Val, Amt), 64); |
11970 | else |
11971 | Res = Builder.CreateLShr(Val, Amt); |
11972 | return Builder.CreateTrunc(Res, Int64Ty); |
11973 | } |
11974 | case X86::BI_ReadWriteBarrier: |
11975 | case X86::BI_ReadBarrier: |
11976 | case X86::BI_WriteBarrier: { |
11977 | return Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent, |
11978 | llvm::SyncScope::SingleThread); |
11979 | } |
11980 | case X86::BI_BitScanForward: |
11981 | case X86::BI_BitScanForward64: |
11982 | return EmitMSVCBuiltinExpr(MSVCIntrin::_BitScanForward, E); |
11983 | case X86::BI_BitScanReverse: |
11984 | case X86::BI_BitScanReverse64: |
11985 | return EmitMSVCBuiltinExpr(MSVCIntrin::_BitScanReverse, E); |
11986 | |
11987 | case X86::BI_InterlockedAnd64: |
11988 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedAnd, E); |
11989 | case X86::BI_InterlockedExchange64: |
11990 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchange, E); |
11991 | case X86::BI_InterlockedExchangeAdd64: |
11992 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeAdd, E); |
11993 | case X86::BI_InterlockedExchangeSub64: |
11994 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeSub, E); |
11995 | case X86::BI_InterlockedOr64: |
11996 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedOr, E); |
11997 | case X86::BI_InterlockedXor64: |
11998 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedXor, E); |
11999 | case X86::BI_InterlockedDecrement64: |
12000 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedDecrement, E); |
12001 | case X86::BI_InterlockedIncrement64: |
12002 | return EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedIncrement, E); |
12003 | case X86::BI_InterlockedCompareExchange128: { |
12004 | // InterlockedCompareExchange128 doesn't directly refer to 128bit ints, |
12005 | // instead it takes pointers to 64bit ints for Destination and |
12006 | // ComparandResult, and exchange is taken as two 64bit ints (high & low). |
12007 | // The previous value is written to ComparandResult, and success is |
12008 | // returned. |
12009 | |
12010 | llvm::Type *Int128Ty = Builder.getInt128Ty(); |
12011 | llvm::Type *Int128PtrTy = Int128Ty->getPointerTo(); |
12012 | |
12013 | Value *Destination = |
12014 | Builder.CreateBitCast(Ops[0], Int128PtrTy); |
12015 | Value *ExchangeHigh128 = Builder.CreateZExt(Ops[1], Int128Ty); |
12016 | Value *ExchangeLow128 = Builder.CreateZExt(Ops[2], Int128Ty); |
12017 | Address ComparandResult(Builder.CreateBitCast(Ops[3], Int128PtrTy), |
12018 | getContext().toCharUnitsFromBits(128)); |
12019 | |
12020 | Value *Exchange = Builder.CreateOr( |
12021 | Builder.CreateShl(ExchangeHigh128, 64, "", false, false), |
12022 | ExchangeLow128); |
12023 | |
12024 | Value *Comparand = Builder.CreateLoad(ComparandResult); |
12025 | |
12026 | AtomicCmpXchgInst *CXI = |
12027 | Builder.CreateAtomicCmpXchg(Destination, Comparand, Exchange, |
12028 | AtomicOrdering::SequentiallyConsistent, |
12029 | AtomicOrdering::SequentiallyConsistent); |
12030 | CXI->setVolatile(true); |
12031 | |
12032 | // Write the result back to the inout pointer. |
12033 | Builder.CreateStore(Builder.CreateExtractValue(CXI, 0), ComparandResult); |
12034 | |
12035 | // Get the success boolean and zero extend it to i8. |
12036 | Value *Success = Builder.CreateExtractValue(CXI, 1); |
12037 | return Builder.CreateZExt(Success, ConvertType(E->getType())); |
12038 | } |
12039 | |
12040 | case X86::BI_AddressOfReturnAddress: { |
12041 | Function *F = CGM.getIntrinsic(Intrinsic::addressofreturnaddress); |
12042 | return Builder.CreateCall(F); |
12043 | } |
12044 | case X86::BI__stosb: { |
12045 | // We treat __stosb as a volatile memset - it may not generate "rep stosb" |
12046 | // instruction, but it will create a memset that won't be optimized away. |
12047 | return Builder.CreateMemSet(Ops[0], Ops[1], Ops[2], 1, true); |
12048 | } |
12049 | case X86::BI__ud2: |
12050 | // llvm.trap makes a ud2a instruction on x86. |
12051 | return EmitTrapCall(Intrinsic::trap); |
12052 | case X86::BI__int2c: { |
12053 | // This syscall signals a driver assertion failure in x86 NT kernels. |
12054 | llvm::FunctionType *FTy = llvm::FunctionType::get(VoidTy, false); |
12055 | llvm::InlineAsm *IA = |
12056 | llvm::InlineAsm::get(FTy, "int $$0x2c", "", /*SideEffects=*/true); |
12057 | llvm::AttributeList NoReturnAttr = llvm::AttributeList::get( |
12058 | getLLVMContext(), llvm::AttributeList::FunctionIndex, |
12059 | llvm::Attribute::NoReturn); |
12060 | llvm::CallInst *CI = Builder.CreateCall(IA); |
12061 | CI->setAttributes(NoReturnAttr); |
12062 | return CI; |
12063 | } |
12064 | case X86::BI__readfsbyte: |
12065 | case X86::BI__readfsword: |
12066 | case X86::BI__readfsdword: |
12067 | case X86::BI__readfsqword: { |
12068 | llvm::Type *IntTy = ConvertType(E->getType()); |
12069 | Value *Ptr = |
12070 | Builder.CreateIntToPtr(Ops[0], llvm::PointerType::get(IntTy, 257)); |
12071 | LoadInst *Load = Builder.CreateAlignedLoad( |
12072 | IntTy, Ptr, getContext().getTypeAlignInChars(E->getType())); |
12073 | Load->setVolatile(true); |
12074 | return Load; |
12075 | } |
12076 | case X86::BI__readgsbyte: |
12077 | case X86::BI__readgsword: |
12078 | case X86::BI__readgsdword: |
12079 | case X86::BI__readgsqword: { |
12080 | llvm::Type *IntTy = ConvertType(E->getType()); |
12081 | Value *Ptr = |
12082 | Builder.CreateIntToPtr(Ops[0], llvm::PointerType::get(IntTy, 256)); |
12083 | LoadInst *Load = Builder.CreateAlignedLoad( |
12084 | IntTy, Ptr, getContext().getTypeAlignInChars(E->getType())); |
12085 | Load->setVolatile(true); |
12086 | return Load; |
12087 | } |
12088 | case X86::BI__builtin_ia32_paddsb512: |
12089 | case X86::BI__builtin_ia32_paddsw512: |
12090 | case X86::BI__builtin_ia32_paddsb256: |
12091 | case X86::BI__builtin_ia32_paddsw256: |
12092 | case X86::BI__builtin_ia32_paddsb128: |
12093 | case X86::BI__builtin_ia32_paddsw128: |
12094 | return EmitX86AddSubSatExpr(*this, Ops, true, true); |
12095 | case X86::BI__builtin_ia32_paddusb512: |
12096 | case X86::BI__builtin_ia32_paddusw512: |
12097 | case X86::BI__builtin_ia32_paddusb256: |
12098 | case X86::BI__builtin_ia32_paddusw256: |
12099 | case X86::BI__builtin_ia32_paddusb128: |
12100 | case X86::BI__builtin_ia32_paddusw128: |
12101 | return EmitX86AddSubSatExpr(*this, Ops, false, true); |
12102 | case X86::BI__builtin_ia32_psubsb512: |
12103 | case X86::BI__builtin_ia32_psubsw512: |
12104 | case X86::BI__builtin_ia32_psubsb256: |
12105 | case X86::BI__builtin_ia32_psubsw256: |
12106 | case X86::BI__builtin_ia32_psubsb128: |
12107 | case X86::BI__builtin_ia32_psubsw128: |
12108 | return EmitX86AddSubSatExpr(*this, Ops, true, false); |
12109 | case X86::BI__builtin_ia32_psubusb512: |
12110 | case X86::BI__builtin_ia32_psubusw512: |
12111 | case X86::BI__builtin_ia32_psubusb256: |
12112 | case X86::BI__builtin_ia32_psubusw256: |
12113 | case X86::BI__builtin_ia32_psubusb128: |
12114 | case X86::BI__builtin_ia32_psubusw128: |
12115 | return EmitX86AddSubSatExpr(*this, Ops, false, false); |
12116 | } |
12117 | } |
12118 | |
12119 | Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID, |
12120 | const CallExpr *E) { |
12121 | SmallVector<Value*, 4> Ops; |
12122 | |
12123 | for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) |
12124 | Ops.push_back(EmitScalarExpr(E->getArg(i))); |
12125 | |
12126 | Intrinsic::ID ID = Intrinsic::not_intrinsic; |
12127 | |
12128 | switch (BuiltinID) { |
12129 | default: return nullptr; |
12130 | |
12131 | // __builtin_ppc_get_timebase is GCC 4.8+'s PowerPC-specific name for what we |
12132 | // call __builtin_readcyclecounter. |
12133 | case PPC::BI__builtin_ppc_get_timebase: |
12134 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::readcyclecounter)); |
12135 | |
12136 | // vec_ld, vec_xl_be, vec_lvsl, vec_lvsr |
12137 | case PPC::BI__builtin_altivec_lvx: |
12138 | case PPC::BI__builtin_altivec_lvxl: |
12139 | case PPC::BI__builtin_altivec_lvebx: |
12140 | case PPC::BI__builtin_altivec_lvehx: |
12141 | case PPC::BI__builtin_altivec_lvewx: |
12142 | case PPC::BI__builtin_altivec_lvsl: |
12143 | case PPC::BI__builtin_altivec_lvsr: |
12144 | case PPC::BI__builtin_vsx_lxvd2x: |
12145 | case PPC::BI__builtin_vsx_lxvw4x: |
12146 | case PPC::BI__builtin_vsx_lxvd2x_be: |
12147 | case PPC::BI__builtin_vsx_lxvw4x_be: |
12148 | case PPC::BI__builtin_vsx_lxvl: |
12149 | case PPC::BI__builtin_vsx_lxvll: |
12150 | { |
12151 | if(BuiltinID == PPC::BI__builtin_vsx_lxvl || |
12152 | BuiltinID == PPC::BI__builtin_vsx_lxvll){ |
12153 | Ops[0] = Builder.CreateBitCast(Ops[0], Int8PtrTy); |
12154 | }else { |
12155 | Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy); |
12156 | Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]); |
12157 | Ops.pop_back(); |
12158 | } |
12159 | |
12160 | switch (BuiltinID) { |
12161 | default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!")::llvm::llvm_unreachable_internal("Unsupported ld/lvsl/lvsr intrinsic!" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 12161); |
12162 | case PPC::BI__builtin_altivec_lvx: |
12163 | ID = Intrinsic::ppc_altivec_lvx; |
12164 | break; |
12165 | case PPC::BI__builtin_altivec_lvxl: |
12166 | ID = Intrinsic::ppc_altivec_lvxl; |
12167 | break; |
12168 | case PPC::BI__builtin_altivec_lvebx: |
12169 | ID = Intrinsic::ppc_altivec_lvebx; |
12170 | break; |
12171 | case PPC::BI__builtin_altivec_lvehx: |
12172 | ID = Intrinsic::ppc_altivec_lvehx; |
12173 | break; |
12174 | case PPC::BI__builtin_altivec_lvewx: |
12175 | ID = Intrinsic::ppc_altivec_lvewx; |
12176 | break; |
12177 | case PPC::BI__builtin_altivec_lvsl: |
12178 | ID = Intrinsic::ppc_altivec_lvsl; |
12179 | break; |
12180 | case PPC::BI__builtin_altivec_lvsr: |
12181 | ID = Intrinsic::ppc_altivec_lvsr; |
12182 | break; |
12183 | case PPC::BI__builtin_vsx_lxvd2x: |
12184 | ID = Intrinsic::ppc_vsx_lxvd2x; |
12185 | break; |
12186 | case PPC::BI__builtin_vsx_lxvw4x: |
12187 | ID = Intrinsic::ppc_vsx_lxvw4x; |
12188 | break; |
12189 | case PPC::BI__builtin_vsx_lxvd2x_be: |
12190 | ID = Intrinsic::ppc_vsx_lxvd2x_be; |
12191 | break; |
12192 | case PPC::BI__builtin_vsx_lxvw4x_be: |
12193 | ID = Intrinsic::ppc_vsx_lxvw4x_be; |
12194 | break; |
12195 | case PPC::BI__builtin_vsx_lxvl: |
12196 | ID = Intrinsic::ppc_vsx_lxvl; |
12197 | break; |
12198 | case PPC::BI__builtin_vsx_lxvll: |
12199 | ID = Intrinsic::ppc_vsx_lxvll; |
12200 | break; |
12201 | } |
12202 | llvm::Function *F = CGM.getIntrinsic(ID); |
12203 | return Builder.CreateCall(F, Ops, ""); |
12204 | } |
12205 | |
12206 | // vec_st, vec_xst_be |
12207 | case PPC::BI__builtin_altivec_stvx: |
12208 | case PPC::BI__builtin_altivec_stvxl: |
12209 | case PPC::BI__builtin_altivec_stvebx: |
12210 | case PPC::BI__builtin_altivec_stvehx: |
12211 | case PPC::BI__builtin_altivec_stvewx: |
12212 | case PPC::BI__builtin_vsx_stxvd2x: |
12213 | case PPC::BI__builtin_vsx_stxvw4x: |
12214 | case PPC::BI__builtin_vsx_stxvd2x_be: |
12215 | case PPC::BI__builtin_vsx_stxvw4x_be: |
12216 | case PPC::BI__builtin_vsx_stxvl: |
12217 | case PPC::BI__builtin_vsx_stxvll: |
12218 | { |
12219 | if(BuiltinID == PPC::BI__builtin_vsx_stxvl || |
12220 | BuiltinID == PPC::BI__builtin_vsx_stxvll ){ |
12221 | Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy); |
12222 | }else { |
12223 | Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy); |
12224 | Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]); |
12225 | Ops.pop_back(); |
12226 | } |
12227 | |
12228 | switch (BuiltinID) { |
12229 | default: llvm_unreachable("Unsupported st intrinsic!")::llvm::llvm_unreachable_internal("Unsupported st intrinsic!" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 12229); |
12230 | case PPC::BI__builtin_altivec_stvx: |
12231 | ID = Intrinsic::ppc_altivec_stvx; |
12232 | break; |
12233 | case PPC::BI__builtin_altivec_stvxl: |
12234 | ID = Intrinsic::ppc_altivec_stvxl; |
12235 | break; |
12236 | case PPC::BI__builtin_altivec_stvebx: |
12237 | ID = Intrinsic::ppc_altivec_stvebx; |
12238 | break; |
12239 | case PPC::BI__builtin_altivec_stvehx: |
12240 | ID = Intrinsic::ppc_altivec_stvehx; |
12241 | break; |
12242 | case PPC::BI__builtin_altivec_stvewx: |
12243 | ID = Intrinsic::ppc_altivec_stvewx; |
12244 | break; |
12245 | case PPC::BI__builtin_vsx_stxvd2x: |
12246 | ID = Intrinsic::ppc_vsx_stxvd2x; |
12247 | break; |
12248 | case PPC::BI__builtin_vsx_stxvw4x: |
12249 | ID = Intrinsic::ppc_vsx_stxvw4x; |
12250 | break; |
12251 | case PPC::BI__builtin_vsx_stxvd2x_be: |
12252 | ID = Intrinsic::ppc_vsx_stxvd2x_be; |
12253 | break; |
12254 | case PPC::BI__builtin_vsx_stxvw4x_be: |
12255 | ID = Intrinsic::ppc_vsx_stxvw4x_be; |
12256 | break; |
12257 | case PPC::BI__builtin_vsx_stxvl: |
12258 | ID = Intrinsic::ppc_vsx_stxvl; |
12259 | break; |
12260 | case PPC::BI__builtin_vsx_stxvll: |
12261 | ID = Intrinsic::ppc_vsx_stxvll; |
12262 | break; |
12263 | } |
12264 | llvm::Function *F = CGM.getIntrinsic(ID); |
12265 | return Builder.CreateCall(F, Ops, ""); |
12266 | } |
12267 | // Square root |
12268 | case PPC::BI__builtin_vsx_xvsqrtsp: |
12269 | case PPC::BI__builtin_vsx_xvsqrtdp: { |
12270 | llvm::Type *ResultType = ConvertType(E->getType()); |
12271 | Value *X = EmitScalarExpr(E->getArg(0)); |
12272 | ID = Intrinsic::sqrt; |
12273 | llvm::Function *F = CGM.getIntrinsic(ID, ResultType); |
12274 | return Builder.CreateCall(F, X); |
12275 | } |
12276 | // Count leading zeros |
12277 | case PPC::BI__builtin_altivec_vclzb: |
12278 | case PPC::BI__builtin_altivec_vclzh: |
12279 | case PPC::BI__builtin_altivec_vclzw: |
12280 | case PPC::BI__builtin_altivec_vclzd: { |
12281 | llvm::Type *ResultType = ConvertType(E->getType()); |
12282 | Value *X = EmitScalarExpr(E->getArg(0)); |
12283 | Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); |
12284 | Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ResultType); |
12285 | return Builder.CreateCall(F, {X, Undef}); |
12286 | } |
12287 | case PPC::BI__builtin_altivec_vctzb: |
12288 | case PPC::BI__builtin_altivec_vctzh: |
12289 | case PPC::BI__builtin_altivec_vctzw: |
12290 | case PPC::BI__builtin_altivec_vctzd: { |
12291 | llvm::Type *ResultType = ConvertType(E->getType()); |
12292 | Value *X = EmitScalarExpr(E->getArg(0)); |
12293 | Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); |
12294 | Function *F = CGM.getIntrinsic(Intrinsic::cttz, ResultType); |
12295 | return Builder.CreateCall(F, {X, Undef}); |
12296 | } |
12297 | case PPC::BI__builtin_altivec_vpopcntb: |
12298 | case PPC::BI__builtin_altivec_vpopcnth: |
12299 | case PPC::BI__builtin_altivec_vpopcntw: |
12300 | case PPC::BI__builtin_altivec_vpopcntd: { |
12301 | llvm::Type *ResultType = ConvertType(E->getType()); |
12302 | Value *X = EmitScalarExpr(E->getArg(0)); |
12303 | llvm::Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ResultType); |
12304 | return Builder.CreateCall(F, X); |
12305 | } |
12306 | // Copy sign |
12307 | case PPC::BI__builtin_vsx_xvcpsgnsp: |
12308 | case PPC::BI__builtin_vsx_xvcpsgndp: { |
12309 | llvm::Type *ResultType = ConvertType(E->getType()); |
12310 | Value *X = EmitScalarExpr(E->getArg(0)); |
12311 | Value *Y = EmitScalarExpr(E->getArg(1)); |
12312 | ID = Intrinsic::copysign; |
12313 | llvm::Function *F = CGM.getIntrinsic(ID, ResultType); |
12314 | return Builder.CreateCall(F, {X, Y}); |
12315 | } |
12316 | // Rounding/truncation |
12317 | case PPC::BI__builtin_vsx_xvrspip: |
12318 | case PPC::BI__builtin_vsx_xvrdpip: |
12319 | case PPC::BI__builtin_vsx_xvrdpim: |
12320 | case PPC::BI__builtin_vsx_xvrspim: |
12321 | case PPC::BI__builtin_vsx_xvrdpi: |
12322 | case PPC::BI__builtin_vsx_xvrspi: |
12323 | case PPC::BI__builtin_vsx_xvrdpic: |
12324 | case PPC::BI__builtin_vsx_xvrspic: |
12325 | case PPC::BI__builtin_vsx_xvrdpiz: |
12326 | case PPC::BI__builtin_vsx_xvrspiz: { |
12327 | llvm::Type *ResultType = ConvertType(E->getType()); |
12328 | Value *X = EmitScalarExpr(E->getArg(0)); |
12329 | if (BuiltinID == PPC::BI__builtin_vsx_xvrdpim || |
12330 | BuiltinID == PPC::BI__builtin_vsx_xvrspim) |
12331 | ID = Intrinsic::floor; |
12332 | else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpi || |
12333 | BuiltinID == PPC::BI__builtin_vsx_xvrspi) |
12334 | ID = Intrinsic::round; |
12335 | else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpic || |
12336 | BuiltinID == PPC::BI__builtin_vsx_xvrspic) |
12337 | ID = Intrinsic::nearbyint; |
12338 | else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpip || |
12339 | BuiltinID == PPC::BI__builtin_vsx_xvrspip) |
12340 | ID = Intrinsic::ceil; |
12341 | else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpiz || |
12342 | BuiltinID == PPC::BI__builtin_vsx_xvrspiz) |
12343 | ID = Intrinsic::trunc; |
12344 | llvm::Function *F = CGM.getIntrinsic(ID, ResultType); |
12345 | return Builder.CreateCall(F, X); |
12346 | } |
12347 | |
12348 | // Absolute value |
12349 | case PPC::BI__builtin_vsx_xvabsdp: |
12350 | case PPC::BI__builtin_vsx_xvabssp: { |
12351 | llvm::Type *ResultType = ConvertType(E->getType()); |
12352 | Value *X = EmitScalarExpr(E->getArg(0)); |
12353 | llvm::Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); |
12354 | return Builder.CreateCall(F, X); |
12355 | } |
12356 | |
12357 | // FMA variations |
12358 | case PPC::BI__builtin_vsx_xvmaddadp: |
12359 | case PPC::BI__builtin_vsx_xvmaddasp: |
12360 | case PPC::BI__builtin_vsx_xvnmaddadp: |
12361 | case PPC::BI__builtin_vsx_xvnmaddasp: |
12362 | case PPC::BI__builtin_vsx_xvmsubadp: |
12363 | case PPC::BI__builtin_vsx_xvmsubasp: |
12364 | case PPC::BI__builtin_vsx_xvnmsubadp: |
12365 | case PPC::BI__builtin_vsx_xvnmsubasp: { |
12366 | llvm::Type *ResultType = ConvertType(E->getType()); |
12367 | Value *X = EmitScalarExpr(E->getArg(0)); |
12368 | Value *Y = EmitScalarExpr(E->getArg(1)); |
12369 | Value *Z = EmitScalarExpr(E->getArg(2)); |
12370 | Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); |
12371 | llvm::Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); |
12372 | switch (BuiltinID) { |
12373 | case PPC::BI__builtin_vsx_xvmaddadp: |
12374 | case PPC::BI__builtin_vsx_xvmaddasp: |
12375 | return Builder.CreateCall(F, {X, Y, Z}); |
12376 | case PPC::BI__builtin_vsx_xvnmaddadp: |
12377 | case PPC::BI__builtin_vsx_xvnmaddasp: |
12378 | return Builder.CreateFSub(Zero, |
12379 | Builder.CreateCall(F, {X, Y, Z}), "sub"); |
12380 | case PPC::BI__builtin_vsx_xvmsubadp: |
12381 | case PPC::BI__builtin_vsx_xvmsubasp: |
12382 | return Builder.CreateCall(F, |
12383 | {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); |
12384 | case PPC::BI__builtin_vsx_xvnmsubadp: |
12385 | case PPC::BI__builtin_vsx_xvnmsubasp: |
12386 | Value *FsubRes = |
12387 | Builder.CreateCall(F, {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); |
12388 | return Builder.CreateFSub(Zero, FsubRes, "sub"); |
12389 | } |
12390 | llvm_unreachable("Unknown FMA operation")::llvm::llvm_unreachable_internal("Unknown FMA operation", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 12390); |
12391 | return nullptr; // Suppress no-return warning |
12392 | } |
12393 | |
12394 | case PPC::BI__builtin_vsx_insertword: { |
12395 | llvm::Function *F = CGM.getIntrinsic(Intrinsic::ppc_vsx_xxinsertw); |
12396 | |
12397 | // Third argument is a compile time constant int. It must be clamped to |
12398 | // to the range [0, 12]. |
12399 | ConstantInt *ArgCI = dyn_cast<ConstantInt>(Ops[2]); |
12400 | assert(ArgCI &&((ArgCI && "Third arg to xxinsertw intrinsic must be constant integer" ) ? static_cast<void> (0) : __assert_fail ("ArgCI && \"Third arg to xxinsertw intrinsic must be constant integer\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 12401, __PRETTY_FUNCTION__)) |
12401 | "Third arg to xxinsertw intrinsic must be constant integer")((ArgCI && "Third arg to xxinsertw intrinsic must be constant integer" ) ? static_cast<void> (0) : __assert_fail ("ArgCI && \"Third arg to xxinsertw intrinsic must be constant integer\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 12401, __PRETTY_FUNCTION__)); |
12402 | const int64_t MaxIndex = 12; |
12403 | int64_t Index = clamp(ArgCI->getSExtValue(), 0, MaxIndex); |
12404 | |
12405 | // The builtin semantics don't exactly match the xxinsertw instructions |
12406 | // semantics (which ppc_vsx_xxinsertw follows). The builtin extracts the |
12407 | // word from the first argument, and inserts it in the second argument. The |
12408 | // instruction extracts the word from its second input register and inserts |
12409 | // it into its first input register, so swap the first and second arguments. |
12410 | std::swap(Ops[0], Ops[1]); |
12411 | |
12412 | // Need to cast the second argument from a vector of unsigned int to a |
12413 | // vector of long long. |
12414 | Ops[1] = Builder.CreateBitCast(Ops[1], llvm::VectorType::get(Int64Ty, 2)); |
12415 | |
12416 | if (getTarget().isLittleEndian()) { |
12417 | // Create a shuffle mask of (1, 0) |
12418 | Constant *ShuffleElts[2] = { ConstantInt::get(Int32Ty, 1), |
12419 | ConstantInt::get(Int32Ty, 0) |
12420 | }; |
12421 | Constant *ShuffleMask = llvm::ConstantVector::get(ShuffleElts); |
12422 | |
12423 | // Reverse the double words in the vector we will extract from. |
12424 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int64Ty, 2)); |
12425 | Ops[0] = Builder.CreateShuffleVector(Ops[0], Ops[0], ShuffleMask); |
12426 | |
12427 | // Reverse the index. |
12428 | Index = MaxIndex - Index; |
12429 | } |
12430 | |
12431 | // Intrinsic expects the first arg to be a vector of int. |
12432 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int32Ty, 4)); |
12433 | Ops[2] = ConstantInt::getSigned(Int32Ty, Index); |
12434 | return Builder.CreateCall(F, Ops); |
12435 | } |
12436 | |
12437 | case PPC::BI__builtin_vsx_extractuword: { |
12438 | llvm::Function *F = CGM.getIntrinsic(Intrinsic::ppc_vsx_xxextractuw); |
12439 | |
12440 | // Intrinsic expects the first argument to be a vector of doublewords. |
12441 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int64Ty, 2)); |
12442 | |
12443 | // The second argument is a compile time constant int that needs to |
12444 | // be clamped to the range [0, 12]. |
12445 | ConstantInt *ArgCI = dyn_cast<ConstantInt>(Ops[1]); |
12446 | assert(ArgCI &&((ArgCI && "Second Arg to xxextractuw intrinsic must be a constant integer!" ) ? static_cast<void> (0) : __assert_fail ("ArgCI && \"Second Arg to xxextractuw intrinsic must be a constant integer!\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 12447, __PRETTY_FUNCTION__)) |
12447 | "Second Arg to xxextractuw intrinsic must be a constant integer!")((ArgCI && "Second Arg to xxextractuw intrinsic must be a constant integer!" ) ? static_cast<void> (0) : __assert_fail ("ArgCI && \"Second Arg to xxextractuw intrinsic must be a constant integer!\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 12447, __PRETTY_FUNCTION__)); |
12448 | const int64_t MaxIndex = 12; |
12449 | int64_t Index = clamp(ArgCI->getSExtValue(), 0, MaxIndex); |
12450 | |
12451 | if (getTarget().isLittleEndian()) { |
12452 | // Reverse the index. |
12453 | Index = MaxIndex - Index; |
12454 | Ops[1] = ConstantInt::getSigned(Int32Ty, Index); |
12455 | |
12456 | // Emit the call, then reverse the double words of the results vector. |
12457 | Value *Call = Builder.CreateCall(F, Ops); |
12458 | |
12459 | // Create a shuffle mask of (1, 0) |
12460 | Constant *ShuffleElts[2] = { ConstantInt::get(Int32Ty, 1), |
12461 | ConstantInt::get(Int32Ty, 0) |
12462 | }; |
12463 | Constant *ShuffleMask = llvm::ConstantVector::get(ShuffleElts); |
12464 | |
12465 | Value *ShuffleCall = Builder.CreateShuffleVector(Call, Call, ShuffleMask); |
12466 | return ShuffleCall; |
12467 | } else { |
12468 | Ops[1] = ConstantInt::getSigned(Int32Ty, Index); |
12469 | return Builder.CreateCall(F, Ops); |
12470 | } |
12471 | } |
12472 | |
12473 | case PPC::BI__builtin_vsx_xxpermdi: { |
12474 | ConstantInt *ArgCI = dyn_cast<ConstantInt>(Ops[2]); |
12475 | assert(ArgCI && "Third arg must be constant integer!")((ArgCI && "Third arg must be constant integer!") ? static_cast <void> (0) : __assert_fail ("ArgCI && \"Third arg must be constant integer!\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 12475, __PRETTY_FUNCTION__)); |
12476 | |
12477 | unsigned Index = ArgCI->getZExtValue(); |
12478 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int64Ty, 2)); |
12479 | Ops[1] = Builder.CreateBitCast(Ops[1], llvm::VectorType::get(Int64Ty, 2)); |
12480 | |
12481 | // Account for endianness by treating this as just a shuffle. So we use the |
12482 | // same indices for both LE and BE in order to produce expected results in |
12483 | // both cases. |
12484 | unsigned ElemIdx0 = (Index & 2) >> 1; |
12485 | unsigned ElemIdx1 = 2 + (Index & 1); |
12486 | |
12487 | Constant *ShuffleElts[2] = {ConstantInt::get(Int32Ty, ElemIdx0), |
12488 | ConstantInt::get(Int32Ty, ElemIdx1)}; |
12489 | Constant *ShuffleMask = llvm::ConstantVector::get(ShuffleElts); |
12490 | |
12491 | Value *ShuffleCall = |
12492 | Builder.CreateShuffleVector(Ops[0], Ops[1], ShuffleMask); |
12493 | QualType BIRetType = E->getType(); |
12494 | auto RetTy = ConvertType(BIRetType); |
12495 | return Builder.CreateBitCast(ShuffleCall, RetTy); |
12496 | } |
12497 | |
12498 | case PPC::BI__builtin_vsx_xxsldwi: { |
12499 | ConstantInt *ArgCI = dyn_cast<ConstantInt>(Ops[2]); |
12500 | assert(ArgCI && "Third argument must be a compile time constant")((ArgCI && "Third argument must be a compile time constant" ) ? static_cast<void> (0) : __assert_fail ("ArgCI && \"Third argument must be a compile time constant\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 12500, __PRETTY_FUNCTION__)); |
12501 | unsigned Index = ArgCI->getZExtValue() & 0x3; |
12502 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int32Ty, 4)); |
12503 | Ops[1] = Builder.CreateBitCast(Ops[1], llvm::VectorType::get(Int32Ty, 4)); |
12504 | |
12505 | // Create a shuffle mask |
12506 | unsigned ElemIdx0; |
12507 | unsigned ElemIdx1; |
12508 | unsigned ElemIdx2; |
12509 | unsigned ElemIdx3; |
12510 | if (getTarget().isLittleEndian()) { |
12511 | // Little endian element N comes from element 8+N-Index of the |
12512 | // concatenated wide vector (of course, using modulo arithmetic on |
12513 | // the total number of elements). |
12514 | ElemIdx0 = (8 - Index) % 8; |
12515 | ElemIdx1 = (9 - Index) % 8; |
12516 | ElemIdx2 = (10 - Index) % 8; |
12517 | ElemIdx3 = (11 - Index) % 8; |
12518 | } else { |
12519 | // Big endian ElemIdx<N> = Index + N |
12520 | ElemIdx0 = Index; |
12521 | ElemIdx1 = Index + 1; |
12522 | ElemIdx2 = Index + 2; |
12523 | ElemIdx3 = Index + 3; |
12524 | } |
12525 | |
12526 | Constant *ShuffleElts[4] = {ConstantInt::get(Int32Ty, ElemIdx0), |
12527 | ConstantInt::get(Int32Ty, ElemIdx1), |
12528 | ConstantInt::get(Int32Ty, ElemIdx2), |
12529 | ConstantInt::get(Int32Ty, ElemIdx3)}; |
12530 | |
12531 | Constant *ShuffleMask = llvm::ConstantVector::get(ShuffleElts); |
12532 | Value *ShuffleCall = |
12533 | Builder.CreateShuffleVector(Ops[0], Ops[1], ShuffleMask); |
12534 | QualType BIRetType = E->getType(); |
12535 | auto RetTy = ConvertType(BIRetType); |
12536 | return Builder.CreateBitCast(ShuffleCall, RetTy); |
12537 | } |
12538 | |
12539 | case PPC::BI__builtin_pack_vector_int128: { |
12540 | bool isLittleEndian = getTarget().isLittleEndian(); |
12541 | Value *UndefValue = |
12542 | llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), 2)); |
12543 | Value *Res = Builder.CreateInsertElement( |
12544 | UndefValue, Ops[0], (uint64_t)(isLittleEndian ? 1 : 0)); |
12545 | Res = Builder.CreateInsertElement(Res, Ops[1], |
12546 | (uint64_t)(isLittleEndian ? 0 : 1)); |
12547 | return Builder.CreateBitCast(Res, ConvertType(E->getType())); |
12548 | } |
12549 | |
12550 | case PPC::BI__builtin_unpack_vector_int128: { |
12551 | ConstantInt *Index = cast<ConstantInt>(Ops[1]); |
12552 | Value *Unpacked = Builder.CreateBitCast( |
12553 | Ops[0], llvm::VectorType::get(ConvertType(E->getType()), 2)); |
12554 | |
12555 | if (getTarget().isLittleEndian()) |
12556 | Index = ConstantInt::get(Index->getType(), 1 - Index->getZExtValue()); |
12557 | |
12558 | return Builder.CreateExtractElement(Unpacked, Index); |
12559 | } |
12560 | } |
12561 | } |
12562 | |
12563 | Value *CodeGenFunction::EmitAMDGPUBuiltinExpr(unsigned BuiltinID, |
12564 | const CallExpr *E) { |
12565 | switch (BuiltinID) { |
12566 | case AMDGPU::BI__builtin_amdgcn_div_scale: |
12567 | case AMDGPU::BI__builtin_amdgcn_div_scalef: { |
12568 | // Translate from the intrinsics's struct return to the builtin's out |
12569 | // argument. |
12570 | |
12571 | Address FlagOutPtr = EmitPointerWithAlignment(E->getArg(3)); |
12572 | |
12573 | llvm::Value *X = EmitScalarExpr(E->getArg(0)); |
12574 | llvm::Value *Y = EmitScalarExpr(E->getArg(1)); |
12575 | llvm::Value *Z = EmitScalarExpr(E->getArg(2)); |
12576 | |
12577 | llvm::Function *Callee = CGM.getIntrinsic(Intrinsic::amdgcn_div_scale, |
12578 | X->getType()); |
12579 | |
12580 | llvm::Value *Tmp = Builder.CreateCall(Callee, {X, Y, Z}); |
12581 | |
12582 | llvm::Value *Result = Builder.CreateExtractValue(Tmp, 0); |
12583 | llvm::Value *Flag = Builder.CreateExtractValue(Tmp, 1); |
12584 | |
12585 | llvm::Type *RealFlagType |
12586 | = FlagOutPtr.getPointer()->getType()->getPointerElementType(); |
12587 | |
12588 | llvm::Value *FlagExt = Builder.CreateZExt(Flag, RealFlagType); |
12589 | Builder.CreateStore(FlagExt, FlagOutPtr); |
12590 | return Result; |
12591 | } |
12592 | case AMDGPU::BI__builtin_amdgcn_div_fmas: |
12593 | case AMDGPU::BI__builtin_amdgcn_div_fmasf: { |
12594 | llvm::Value *Src0 = EmitScalarExpr(E->getArg(0)); |
12595 | llvm::Value *Src1 = EmitScalarExpr(E->getArg(1)); |
12596 | llvm::Value *Src2 = EmitScalarExpr(E->getArg(2)); |
12597 | llvm::Value *Src3 = EmitScalarExpr(E->getArg(3)); |
12598 | |
12599 | llvm::Function *F = CGM.getIntrinsic(Intrinsic::amdgcn_div_fmas, |
12600 | Src0->getType()); |
12601 | llvm::Value *Src3ToBool = Builder.CreateIsNotNull(Src3); |
12602 | return Builder.CreateCall(F, {Src0, Src1, Src2, Src3ToBool}); |
12603 | } |
12604 | |
12605 | case AMDGPU::BI__builtin_amdgcn_ds_swizzle: |
12606 | return emitBinaryBuiltin(*this, E, Intrinsic::amdgcn_ds_swizzle); |
12607 | case AMDGPU::BI__builtin_amdgcn_mov_dpp: |
12608 | case AMDGPU::BI__builtin_amdgcn_update_dpp: { |
12609 | llvm::SmallVector<llvm::Value *, 6> Args; |
12610 | for (unsigned I = 0; I != E->getNumArgs(); ++I) |
12611 | Args.push_back(EmitScalarExpr(E->getArg(I))); |
12612 | assert(Args.size() == 5 || Args.size() == 6)((Args.size() == 5 || Args.size() == 6) ? static_cast<void > (0) : __assert_fail ("Args.size() == 5 || Args.size() == 6" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 12612, __PRETTY_FUNCTION__)); |
12613 | if (Args.size() == 5) |
12614 | Args.insert(Args.begin(), llvm::UndefValue::get(Args[0]->getType())); |
12615 | Function *F = |
12616 | CGM.getIntrinsic(Intrinsic::amdgcn_update_dpp, Args[0]->getType()); |
12617 | return Builder.CreateCall(F, Args); |
12618 | } |
12619 | case AMDGPU::BI__builtin_amdgcn_div_fixup: |
12620 | case AMDGPU::BI__builtin_amdgcn_div_fixupf: |
12621 | case AMDGPU::BI__builtin_amdgcn_div_fixuph: |
12622 | return emitTernaryBuiltin(*this, E, Intrinsic::amdgcn_div_fixup); |
12623 | case AMDGPU::BI__builtin_amdgcn_trig_preop: |
12624 | case AMDGPU::BI__builtin_amdgcn_trig_preopf: |
12625 | return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_trig_preop); |
12626 | case AMDGPU::BI__builtin_amdgcn_rcp: |
12627 | case AMDGPU::BI__builtin_amdgcn_rcpf: |
12628 | case AMDGPU::BI__builtin_amdgcn_rcph: |
12629 | return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rcp); |
12630 | case AMDGPU::BI__builtin_amdgcn_rsq: |
12631 | case AMDGPU::BI__builtin_amdgcn_rsqf: |
12632 | case AMDGPU::BI__builtin_amdgcn_rsqh: |
12633 | return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rsq); |
12634 | case AMDGPU::BI__builtin_amdgcn_rsq_clamp: |
12635 | case AMDGPU::BI__builtin_amdgcn_rsq_clampf: |
12636 | return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rsq_clamp); |
12637 | case AMDGPU::BI__builtin_amdgcn_sinf: |
12638 | case AMDGPU::BI__builtin_amdgcn_sinh: |
12639 | return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_sin); |
12640 | case AMDGPU::BI__builtin_amdgcn_cosf: |
12641 | case AMDGPU::BI__builtin_amdgcn_cosh: |
12642 | return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_cos); |
12643 | case AMDGPU::BI__builtin_amdgcn_log_clampf: |
12644 | return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_log_clamp); |
12645 | case AMDGPU::BI__builtin_amdgcn_ldexp: |
12646 | case AMDGPU::BI__builtin_amdgcn_ldexpf: |
12647 | case AMDGPU::BI__builtin_amdgcn_ldexph: |
12648 | return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_ldexp); |
12649 | case AMDGPU::BI__builtin_amdgcn_frexp_mant: |
12650 | case AMDGPU::BI__builtin_amdgcn_frexp_mantf: |
12651 | case AMDGPU::BI__builtin_amdgcn_frexp_manth: |
12652 | return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_frexp_mant); |
12653 | case AMDGPU::BI__builtin_amdgcn_frexp_exp: |
12654 | case AMDGPU::BI__builtin_amdgcn_frexp_expf: { |
12655 | Value *Src0 = EmitScalarExpr(E->getArg(0)); |
12656 | Function *F = CGM.getIntrinsic(Intrinsic::amdgcn_frexp_exp, |
12657 | { Builder.getInt32Ty(), Src0->getType() }); |
12658 | return Builder.CreateCall(F, Src0); |
12659 | } |
12660 | case AMDGPU::BI__builtin_amdgcn_frexp_exph: { |
12661 | Value *Src0 = EmitScalarExpr(E->getArg(0)); |
12662 | Function *F = CGM.getIntrinsic(Intrinsic::amdgcn_frexp_exp, |
12663 | { Builder.getInt16Ty(), Src0->getType() }); |
12664 | return Builder.CreateCall(F, Src0); |
12665 | } |
12666 | case AMDGPU::BI__builtin_amdgcn_fract: |
12667 | case AMDGPU::BI__builtin_amdgcn_fractf: |
12668 | case AMDGPU::BI__builtin_amdgcn_fracth: |
12669 | return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_fract); |
12670 | case AMDGPU::BI__builtin_amdgcn_lerp: |
12671 | return emitTernaryBuiltin(*this, E, Intrinsic::amdgcn_lerp); |
12672 | case AMDGPU::BI__builtin_amdgcn_uicmp: |
12673 | case AMDGPU::BI__builtin_amdgcn_uicmpl: |
12674 | case AMDGPU::BI__builtin_amdgcn_sicmp: |
12675 | case AMDGPU::BI__builtin_amdgcn_sicmpl: |
12676 | return emitTernaryBuiltin(*this, E, Intrinsic::amdgcn_icmp); |
12677 | case AMDGPU::BI__builtin_amdgcn_fcmp: |
12678 | case AMDGPU::BI__builtin_amdgcn_fcmpf: |
12679 | return emitTernaryBuiltin(*this, E, Intrinsic::amdgcn_fcmp); |
12680 | case AMDGPU::BI__builtin_amdgcn_class: |
12681 | case AMDGPU::BI__builtin_amdgcn_classf: |
12682 | case AMDGPU::BI__builtin_amdgcn_classh: |
12683 | return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_class); |
12684 | case AMDGPU::BI__builtin_amdgcn_fmed3f: |
12685 | case AMDGPU::BI__builtin_amdgcn_fmed3h: |
12686 | return emitTernaryBuiltin(*this, E, Intrinsic::amdgcn_fmed3); |
12687 | case AMDGPU::BI__builtin_amdgcn_ds_append: |
12688 | case AMDGPU::BI__builtin_amdgcn_ds_consume: { |
12689 | Intrinsic::ID Intrin = BuiltinID == AMDGPU::BI__builtin_amdgcn_ds_append ? |
12690 | Intrinsic::amdgcn_ds_append : Intrinsic::amdgcn_ds_consume; |
12691 | Value *Src0 = EmitScalarExpr(E->getArg(0)); |
12692 | Function *F = CGM.getIntrinsic(Intrin, { Src0->getType() }); |
12693 | return Builder.CreateCall(F, { Src0, Builder.getFalse() }); |
12694 | } |
12695 | case AMDGPU::BI__builtin_amdgcn_read_exec: { |
12696 | CallInst *CI = cast<CallInst>( |
12697 | EmitSpecialRegisterBuiltin(*this, E, Int64Ty, Int64Ty, true, "exec")); |
12698 | CI->setConvergent(); |
12699 | return CI; |
12700 | } |
12701 | case AMDGPU::BI__builtin_amdgcn_read_exec_lo: |
12702 | case AMDGPU::BI__builtin_amdgcn_read_exec_hi: { |
12703 | StringRef RegName = BuiltinID == AMDGPU::BI__builtin_amdgcn_read_exec_lo ? |
12704 | "exec_lo" : "exec_hi"; |
12705 | CallInst *CI = cast<CallInst>( |
12706 | EmitSpecialRegisterBuiltin(*this, E, Int32Ty, Int32Ty, true, RegName)); |
12707 | CI->setConvergent(); |
12708 | return CI; |
12709 | } |
12710 | // amdgcn workitem |
12711 | case AMDGPU::BI__builtin_amdgcn_workitem_id_x: |
12712 | return emitRangedBuiltin(*this, Intrinsic::amdgcn_workitem_id_x, 0, 1024); |
12713 | case AMDGPU::BI__builtin_amdgcn_workitem_id_y: |
12714 | return emitRangedBuiltin(*this, Intrinsic::amdgcn_workitem_id_y, 0, 1024); |
12715 | case AMDGPU::BI__builtin_amdgcn_workitem_id_z: |
12716 | return emitRangedBuiltin(*this, Intrinsic::amdgcn_workitem_id_z, 0, 1024); |
12717 | |
12718 | // r600 intrinsics |
12719 | case AMDGPU::BI__builtin_r600_recipsqrt_ieee: |
12720 | case AMDGPU::BI__builtin_r600_recipsqrt_ieeef: |
12721 | return emitUnaryBuiltin(*this, E, Intrinsic::r600_recipsqrt_ieee); |
12722 | case AMDGPU::BI__builtin_r600_read_tidig_x: |
12723 | return emitRangedBuiltin(*this, Intrinsic::r600_read_tidig_x, 0, 1024); |
12724 | case AMDGPU::BI__builtin_r600_read_tidig_y: |
12725 | return emitRangedBuiltin(*this, Intrinsic::r600_read_tidig_y, 0, 1024); |
12726 | case AMDGPU::BI__builtin_r600_read_tidig_z: |
12727 | return emitRangedBuiltin(*this, Intrinsic::r600_read_tidig_z, 0, 1024); |
12728 | default: |
12729 | return nullptr; |
12730 | } |
12731 | } |
12732 | |
12733 | /// Handle a SystemZ function in which the final argument is a pointer |
12734 | /// to an int that receives the post-instruction CC value. At the LLVM level |
12735 | /// this is represented as a function that returns a {result, cc} pair. |
12736 | static Value *EmitSystemZIntrinsicWithCC(CodeGenFunction &CGF, |
12737 | unsigned IntrinsicID, |
12738 | const CallExpr *E) { |
12739 | unsigned NumArgs = E->getNumArgs() - 1; |
12740 | SmallVector<Value *, 8> Args(NumArgs); |
12741 | for (unsigned I = 0; I < NumArgs; ++I) |
12742 | Args[I] = CGF.EmitScalarExpr(E->getArg(I)); |
12743 | Address CCPtr = CGF.EmitPointerWithAlignment(E->getArg(NumArgs)); |
12744 | Function *F = CGF.CGM.getIntrinsic(IntrinsicID); |
12745 | Value *Call = CGF.Builder.CreateCall(F, Args); |
12746 | Value *CC = CGF.Builder.CreateExtractValue(Call, 1); |
12747 | CGF.Builder.CreateStore(CC, CCPtr); |
12748 | return CGF.Builder.CreateExtractValue(Call, 0); |
12749 | } |
12750 | |
12751 | Value *CodeGenFunction::EmitSystemZBuiltinExpr(unsigned BuiltinID, |
12752 | const CallExpr *E) { |
12753 | switch (BuiltinID) { |
12754 | case SystemZ::BI__builtin_tbegin: { |
12755 | Value *TDB = EmitScalarExpr(E->getArg(0)); |
12756 | Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c); |
12757 | Function *F = CGM.getIntrinsic(Intrinsic::s390_tbegin); |
12758 | return Builder.CreateCall(F, {TDB, Control}); |
12759 | } |
12760 | case SystemZ::BI__builtin_tbegin_nofloat: { |
12761 | Value *TDB = EmitScalarExpr(E->getArg(0)); |
12762 | Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c); |
12763 | Function *F = CGM.getIntrinsic(Intrinsic::s390_tbegin_nofloat); |
12764 | return Builder.CreateCall(F, {TDB, Control}); |
12765 | } |
12766 | case SystemZ::BI__builtin_tbeginc: { |
12767 | Value *TDB = llvm::ConstantPointerNull::get(Int8PtrTy); |
12768 | Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff08); |
12769 | Function *F = CGM.getIntrinsic(Intrinsic::s390_tbeginc); |
12770 | return Builder.CreateCall(F, {TDB, Control}); |
12771 | } |
12772 | case SystemZ::BI__builtin_tabort: { |
12773 | Value *Data = EmitScalarExpr(E->getArg(0)); |
12774 | Function *F = CGM.getIntrinsic(Intrinsic::s390_tabort); |
12775 | return Builder.CreateCall(F, Builder.CreateSExt(Data, Int64Ty, "tabort")); |
12776 | } |
12777 | case SystemZ::BI__builtin_non_tx_store: { |
12778 | Value *Address = EmitScalarExpr(E->getArg(0)); |
12779 | Value *Data = EmitScalarExpr(E->getArg(1)); |
12780 | Function *F = CGM.getIntrinsic(Intrinsic::s390_ntstg); |
12781 | return Builder.CreateCall(F, {Data, Address}); |
12782 | } |
12783 | |
12784 | // Vector builtins. Note that most vector builtins are mapped automatically |
12785 | // to target-specific LLVM intrinsics. The ones handled specially here can |
12786 | // be represented via standard LLVM IR, which is preferable to enable common |
12787 | // LLVM optimizations. |
12788 | |
12789 | case SystemZ::BI__builtin_s390_vpopctb: |
12790 | case SystemZ::BI__builtin_s390_vpopcth: |
12791 | case SystemZ::BI__builtin_s390_vpopctf: |
12792 | case SystemZ::BI__builtin_s390_vpopctg: { |
12793 | llvm::Type *ResultType = ConvertType(E->getType()); |
12794 | Value *X = EmitScalarExpr(E->getArg(0)); |
12795 | Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ResultType); |
12796 | return Builder.CreateCall(F, X); |
12797 | } |
12798 | |
12799 | case SystemZ::BI__builtin_s390_vclzb: |
12800 | case SystemZ::BI__builtin_s390_vclzh: |
12801 | case SystemZ::BI__builtin_s390_vclzf: |
12802 | case SystemZ::BI__builtin_s390_vclzg: { |
12803 | llvm::Type *ResultType = ConvertType(E->getType()); |
12804 | Value *X = EmitScalarExpr(E->getArg(0)); |
12805 | Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); |
12806 | Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ResultType); |
12807 | return Builder.CreateCall(F, {X, Undef}); |
12808 | } |
12809 | |
12810 | case SystemZ::BI__builtin_s390_vctzb: |
12811 | case SystemZ::BI__builtin_s390_vctzh: |
12812 | case SystemZ::BI__builtin_s390_vctzf: |
12813 | case SystemZ::BI__builtin_s390_vctzg: { |
12814 | llvm::Type *ResultType = ConvertType(E->getType()); |
12815 | Value *X = EmitScalarExpr(E->getArg(0)); |
12816 | Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); |
12817 | Function *F = CGM.getIntrinsic(Intrinsic::cttz, ResultType); |
12818 | return Builder.CreateCall(F, {X, Undef}); |
12819 | } |
12820 | |
12821 | case SystemZ::BI__builtin_s390_vfsqsb: |
12822 | case SystemZ::BI__builtin_s390_vfsqdb: { |
12823 | llvm::Type *ResultType = ConvertType(E->getType()); |
12824 | Value *X = EmitScalarExpr(E->getArg(0)); |
12825 | Function *F = CGM.getIntrinsic(Intrinsic::sqrt, ResultType); |
12826 | return Builder.CreateCall(F, X); |
12827 | } |
12828 | case SystemZ::BI__builtin_s390_vfmasb: |
12829 | case SystemZ::BI__builtin_s390_vfmadb: { |
12830 | llvm::Type *ResultType = ConvertType(E->getType()); |
12831 | Value *X = EmitScalarExpr(E->getArg(0)); |
12832 | Value *Y = EmitScalarExpr(E->getArg(1)); |
12833 | Value *Z = EmitScalarExpr(E->getArg(2)); |
12834 | Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); |
12835 | return Builder.CreateCall(F, {X, Y, Z}); |
12836 | } |
12837 | case SystemZ::BI__builtin_s390_vfmssb: |
12838 | case SystemZ::BI__builtin_s390_vfmsdb: { |
12839 | llvm::Type *ResultType = ConvertType(E->getType()); |
12840 | Value *X = EmitScalarExpr(E->getArg(0)); |
12841 | Value *Y = EmitScalarExpr(E->getArg(1)); |
12842 | Value *Z = EmitScalarExpr(E->getArg(2)); |
12843 | Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); |
12844 | Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); |
12845 | return Builder.CreateCall(F, {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); |
12846 | } |
12847 | case SystemZ::BI__builtin_s390_vfnmasb: |
12848 | case SystemZ::BI__builtin_s390_vfnmadb: { |
12849 | llvm::Type *ResultType = ConvertType(E->getType()); |
12850 | Value *X = EmitScalarExpr(E->getArg(0)); |
12851 | Value *Y = EmitScalarExpr(E->getArg(1)); |
12852 | Value *Z = EmitScalarExpr(E->getArg(2)); |
12853 | Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); |
12854 | Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); |
12855 | return Builder.CreateFSub(Zero, Builder.CreateCall(F, {X, Y, Z}), "sub"); |
12856 | } |
12857 | case SystemZ::BI__builtin_s390_vfnmssb: |
12858 | case SystemZ::BI__builtin_s390_vfnmsdb: { |
12859 | llvm::Type *ResultType = ConvertType(E->getType()); |
12860 | Value *X = EmitScalarExpr(E->getArg(0)); |
12861 | Value *Y = EmitScalarExpr(E->getArg(1)); |
12862 | Value *Z = EmitScalarExpr(E->getArg(2)); |
12863 | Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); |
12864 | Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); |
12865 | Value *NegZ = Builder.CreateFSub(Zero, Z, "sub"); |
12866 | return Builder.CreateFSub(Zero, Builder.CreateCall(F, {X, Y, NegZ})); |
12867 | } |
12868 | case SystemZ::BI__builtin_s390_vflpsb: |
12869 | case SystemZ::BI__builtin_s390_vflpdb: { |
12870 | llvm::Type *ResultType = ConvertType(E->getType()); |
12871 | Value *X = EmitScalarExpr(E->getArg(0)); |
12872 | Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); |
12873 | return Builder.CreateCall(F, X); |
12874 | } |
12875 | case SystemZ::BI__builtin_s390_vflnsb: |
12876 | case SystemZ::BI__builtin_s390_vflndb: { |
12877 | llvm::Type *ResultType = ConvertType(E->getType()); |
12878 | Value *X = EmitScalarExpr(E->getArg(0)); |
12879 | Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); |
12880 | Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); |
12881 | return Builder.CreateFSub(Zero, Builder.CreateCall(F, X), "sub"); |
12882 | } |
12883 | case SystemZ::BI__builtin_s390_vfisb: |
12884 | case SystemZ::BI__builtin_s390_vfidb: { |
12885 | llvm::Type *ResultType = ConvertType(E->getType()); |
12886 | Value *X = EmitScalarExpr(E->getArg(0)); |
12887 | // Constant-fold the M4 and M5 mask arguments. |
12888 | llvm::APSInt M4, M5; |
12889 | bool IsConstM4 = E->getArg(1)->isIntegerConstantExpr(M4, getContext()); |
12890 | bool IsConstM5 = E->getArg(2)->isIntegerConstantExpr(M5, getContext()); |
12891 | assert(IsConstM4 && IsConstM5 && "Constant arg isn't actually constant?")((IsConstM4 && IsConstM5 && "Constant arg isn't actually constant?" ) ? static_cast<void> (0) : __assert_fail ("IsConstM4 && IsConstM5 && \"Constant arg isn't actually constant?\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 12891, __PRETTY_FUNCTION__)); |
12892 | (void)IsConstM4; (void)IsConstM5; |
12893 | // Check whether this instance can be represented via a LLVM standard |
12894 | // intrinsic. We only support some combinations of M4 and M5. |
12895 | Intrinsic::ID ID = Intrinsic::not_intrinsic; |
12896 | switch (M4.getZExtValue()) { |
12897 | default: break; |
12898 | case 0: // IEEE-inexact exception allowed |
12899 | switch (M5.getZExtValue()) { |
12900 | default: break; |
12901 | case 0: ID = Intrinsic::rint; break; |
12902 | } |
12903 | break; |
12904 | case 4: // IEEE-inexact exception suppressed |
12905 | switch (M5.getZExtValue()) { |
12906 | default: break; |
12907 | case 0: ID = Intrinsic::nearbyint; break; |
12908 | case 1: ID = Intrinsic::round; break; |
12909 | case 5: ID = Intrinsic::trunc; break; |
12910 | case 6: ID = Intrinsic::ceil; break; |
12911 | case 7: ID = Intrinsic::floor; break; |
12912 | } |
12913 | break; |
12914 | } |
12915 | if (ID != Intrinsic::not_intrinsic) { |
12916 | Function *F = CGM.getIntrinsic(ID, ResultType); |
12917 | return Builder.CreateCall(F, X); |
12918 | } |
12919 | switch (BuiltinID) { |
12920 | case SystemZ::BI__builtin_s390_vfisb: ID = Intrinsic::s390_vfisb; break; |
12921 | case SystemZ::BI__builtin_s390_vfidb: ID = Intrinsic::s390_vfidb; break; |
12922 | default: llvm_unreachable("Unknown BuiltinID")::llvm::llvm_unreachable_internal("Unknown BuiltinID", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 12922); |
12923 | } |
12924 | Function *F = CGM.getIntrinsic(ID); |
12925 | Value *M4Value = llvm::ConstantInt::get(getLLVMContext(), M4); |
12926 | Value *M5Value = llvm::ConstantInt::get(getLLVMContext(), M5); |
12927 | return Builder.CreateCall(F, {X, M4Value, M5Value}); |
12928 | } |
12929 | case SystemZ::BI__builtin_s390_vfmaxsb: |
12930 | case SystemZ::BI__builtin_s390_vfmaxdb: { |
12931 | llvm::Type *ResultType = ConvertType(E->getType()); |
12932 | Value *X = EmitScalarExpr(E->getArg(0)); |
12933 | Value *Y = EmitScalarExpr(E->getArg(1)); |
12934 | // Constant-fold the M4 mask argument. |
12935 | llvm::APSInt M4; |
12936 | bool IsConstM4 = E->getArg(2)->isIntegerConstantExpr(M4, getContext()); |
12937 | assert(IsConstM4 && "Constant arg isn't actually constant?")((IsConstM4 && "Constant arg isn't actually constant?" ) ? static_cast<void> (0) : __assert_fail ("IsConstM4 && \"Constant arg isn't actually constant?\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 12937, __PRETTY_FUNCTION__)); |
12938 | (void)IsConstM4; |
12939 | // Check whether this instance can be represented via a LLVM standard |
12940 | // intrinsic. We only support some values of M4. |
12941 | Intrinsic::ID ID = Intrinsic::not_intrinsic; |
12942 | switch (M4.getZExtValue()) { |
12943 | default: break; |
12944 | case 4: ID = Intrinsic::maxnum; break; |
12945 | } |
12946 | if (ID != Intrinsic::not_intrinsic) { |
12947 | Function *F = CGM.getIntrinsic(ID, ResultType); |
12948 | return Builder.CreateCall(F, {X, Y}); |
12949 | } |
12950 | switch (BuiltinID) { |
12951 | case SystemZ::BI__builtin_s390_vfmaxsb: ID = Intrinsic::s390_vfmaxsb; break; |
12952 | case SystemZ::BI__builtin_s390_vfmaxdb: ID = Intrinsic::s390_vfmaxdb; break; |
12953 | default: llvm_unreachable("Unknown BuiltinID")::llvm::llvm_unreachable_internal("Unknown BuiltinID", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 12953); |
12954 | } |
12955 | Function *F = CGM.getIntrinsic(ID); |
12956 | Value *M4Value = llvm::ConstantInt::get(getLLVMContext(), M4); |
12957 | return Builder.CreateCall(F, {X, Y, M4Value}); |
12958 | } |
12959 | case SystemZ::BI__builtin_s390_vfminsb: |
12960 | case SystemZ::BI__builtin_s390_vfmindb: { |
12961 | llvm::Type *ResultType = ConvertType(E->getType()); |
12962 | Value *X = EmitScalarExpr(E->getArg(0)); |
12963 | Value *Y = EmitScalarExpr(E->getArg(1)); |
12964 | // Constant-fold the M4 mask argument. |
12965 | llvm::APSInt M4; |
12966 | bool IsConstM4 = E->getArg(2)->isIntegerConstantExpr(M4, getContext()); |
12967 | assert(IsConstM4 && "Constant arg isn't actually constant?")((IsConstM4 && "Constant arg isn't actually constant?" ) ? static_cast<void> (0) : __assert_fail ("IsConstM4 && \"Constant arg isn't actually constant?\"" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 12967, __PRETTY_FUNCTION__)); |
12968 | (void)IsConstM4; |
12969 | // Check whether this instance can be represented via a LLVM standard |
12970 | // intrinsic. We only support some values of M4. |
12971 | Intrinsic::ID ID = Intrinsic::not_intrinsic; |
12972 | switch (M4.getZExtValue()) { |
12973 | default: break; |
12974 | case 4: ID = Intrinsic::minnum; break; |
12975 | } |
12976 | if (ID != Intrinsic::not_intrinsic) { |
12977 | Function *F = CGM.getIntrinsic(ID, ResultType); |
12978 | return Builder.CreateCall(F, {X, Y}); |
12979 | } |
12980 | switch (BuiltinID) { |
12981 | case SystemZ::BI__builtin_s390_vfminsb: ID = Intrinsic::s390_vfminsb; break; |
12982 | case SystemZ::BI__builtin_s390_vfmindb: ID = Intrinsic::s390_vfmindb; break; |
12983 | default: llvm_unreachable("Unknown BuiltinID")::llvm::llvm_unreachable_internal("Unknown BuiltinID", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 12983); |
12984 | } |
12985 | Function *F = CGM.getIntrinsic(ID); |
12986 | Value *M4Value = llvm::ConstantInt::get(getLLVMContext(), M4); |
12987 | return Builder.CreateCall(F, {X, Y, M4Value}); |
12988 | } |
12989 | |
12990 | // Vector intrinsics that output the post-instruction CC value. |
12991 | |
12992 | #define INTRINSIC_WITH_CC(NAME) \ |
12993 | case SystemZ::BI__builtin_##NAME: \ |
12994 | return EmitSystemZIntrinsicWithCC(*this, Intrinsic::NAME, E) |
12995 | |
12996 | INTRINSIC_WITH_CC(s390_vpkshs); |
12997 | INTRINSIC_WITH_CC(s390_vpksfs); |
12998 | INTRINSIC_WITH_CC(s390_vpksgs); |
12999 | |
13000 | INTRINSIC_WITH_CC(s390_vpklshs); |
13001 | INTRINSIC_WITH_CC(s390_vpklsfs); |
13002 | INTRINSIC_WITH_CC(s390_vpklsgs); |
13003 | |
13004 | INTRINSIC_WITH_CC(s390_vceqbs); |
13005 | INTRINSIC_WITH_CC(s390_vceqhs); |
13006 | INTRINSIC_WITH_CC(s390_vceqfs); |
13007 | INTRINSIC_WITH_CC(s390_vceqgs); |
13008 | |
13009 | INTRINSIC_WITH_CC(s390_vchbs); |
13010 | INTRINSIC_WITH_CC(s390_vchhs); |
13011 | INTRINSIC_WITH_CC(s390_vchfs); |
13012 | INTRINSIC_WITH_CC(s390_vchgs); |
13013 | |
13014 | INTRINSIC_WITH_CC(s390_vchlbs); |
13015 | INTRINSIC_WITH_CC(s390_vchlhs); |
13016 | INTRINSIC_WITH_CC(s390_vchlfs); |
13017 | INTRINSIC_WITH_CC(s390_vchlgs); |
13018 | |
13019 | INTRINSIC_WITH_CC(s390_vfaebs); |
13020 | INTRINSIC_WITH_CC(s390_vfaehs); |
13021 | INTRINSIC_WITH_CC(s390_vfaefs); |
13022 | |
13023 | INTRINSIC_WITH_CC(s390_vfaezbs); |
13024 | INTRINSIC_WITH_CC(s390_vfaezhs); |
13025 | INTRINSIC_WITH_CC(s390_vfaezfs); |
13026 | |
13027 | INTRINSIC_WITH_CC(s390_vfeebs); |
13028 | INTRINSIC_WITH_CC(s390_vfeehs); |
13029 | INTRINSIC_WITH_CC(s390_vfeefs); |
13030 | |
13031 | INTRINSIC_WITH_CC(s390_vfeezbs); |
13032 | INTRINSIC_WITH_CC(s390_vfeezhs); |
13033 | INTRINSIC_WITH_CC(s390_vfeezfs); |
13034 | |
13035 | INTRINSIC_WITH_CC(s390_vfenebs); |
13036 | INTRINSIC_WITH_CC(s390_vfenehs); |
13037 | INTRINSIC_WITH_CC(s390_vfenefs); |
13038 | |
13039 | INTRINSIC_WITH_CC(s390_vfenezbs); |
13040 | INTRINSIC_WITH_CC(s390_vfenezhs); |
13041 | INTRINSIC_WITH_CC(s390_vfenezfs); |
13042 | |
13043 | INTRINSIC_WITH_CC(s390_vistrbs); |
13044 | INTRINSIC_WITH_CC(s390_vistrhs); |
13045 | INTRINSIC_WITH_CC(s390_vistrfs); |
13046 | |
13047 | INTRINSIC_WITH_CC(s390_vstrcbs); |
13048 | INTRINSIC_WITH_CC(s390_vstrchs); |
13049 | INTRINSIC_WITH_CC(s390_vstrcfs); |
13050 | |
13051 | INTRINSIC_WITH_CC(s390_vstrczbs); |
13052 | INTRINSIC_WITH_CC(s390_vstrczhs); |
13053 | INTRINSIC_WITH_CC(s390_vstrczfs); |
13054 | |
13055 | INTRINSIC_WITH_CC(s390_vfcesbs); |
13056 | INTRINSIC_WITH_CC(s390_vfcedbs); |
13057 | INTRINSIC_WITH_CC(s390_vfchsbs); |
13058 | INTRINSIC_WITH_CC(s390_vfchdbs); |
13059 | INTRINSIC_WITH_CC(s390_vfchesbs); |
13060 | INTRINSIC_WITH_CC(s390_vfchedbs); |
13061 | |
13062 | INTRINSIC_WITH_CC(s390_vftcisb); |
13063 | INTRINSIC_WITH_CC(s390_vftcidb); |
13064 | |
13065 | #undef INTRINSIC_WITH_CC |
13066 | |
13067 | default: |
13068 | return nullptr; |
13069 | } |
13070 | } |
13071 | |
13072 | namespace { |
13073 | // Helper classes for mapping MMA builtins to particular LLVM intrinsic variant. |
13074 | struct NVPTXMmaLdstInfo { |
13075 | unsigned NumResults; // Number of elements to load/store |
13076 | // Intrinsic IDs for row/col variants. 0 if particular layout is unsupported. |
13077 | unsigned IID_col; |
13078 | unsigned IID_row; |
13079 | }; |
13080 | |
13081 | #define MMA_INTR(geom_op_type, layout) \ |
13082 | Intrinsic::nvvm_wmma_##geom_op_type##_##layout##_stride |
13083 | #define MMA_LDST(n, geom_op_type) \ |
13084 | { n, MMA_INTR(geom_op_type, col), MMA_INTR(geom_op_type, row) } |
13085 | |
13086 | static NVPTXMmaLdstInfo getNVPTXMmaLdstInfo(unsigned BuiltinID) { |
13087 | switch (BuiltinID) { |
13088 | // FP MMA loads |
13089 | case NVPTX::BI__hmma_m16n16k16_ld_a: |
13090 | return MMA_LDST(8, m16n16k16_load_a_f16); |
13091 | case NVPTX::BI__hmma_m16n16k16_ld_b: |
13092 | return MMA_LDST(8, m16n16k16_load_b_f16); |
13093 | case NVPTX::BI__hmma_m16n16k16_ld_c_f16: |
13094 | return MMA_LDST(4, m16n16k16_load_c_f16); |
13095 | case NVPTX::BI__hmma_m16n16k16_ld_c_f32: |
13096 | return MMA_LDST(8, m16n16k16_load_c_f32); |
13097 | case NVPTX::BI__hmma_m32n8k16_ld_a: |
13098 | return MMA_LDST(8, m32n8k16_load_a_f16); |
13099 | case NVPTX::BI__hmma_m32n8k16_ld_b: |
13100 | return MMA_LDST(8, m32n8k16_load_b_f16); |
13101 | case NVPTX::BI__hmma_m32n8k16_ld_c_f16: |
13102 | return MMA_LDST(4, m32n8k16_load_c_f16); |
13103 | case NVPTX::BI__hmma_m32n8k16_ld_c_f32: |
13104 | return MMA_LDST(8, m32n8k16_load_c_f32); |
13105 | case NVPTX::BI__hmma_m8n32k16_ld_a: |
13106 | return MMA_LDST(8, m8n32k16_load_a_f16); |
13107 | case NVPTX::BI__hmma_m8n32k16_ld_b: |
13108 | return MMA_LDST(8, m8n32k16_load_b_f16); |
13109 | case NVPTX::BI__hmma_m8n32k16_ld_c_f16: |
13110 | return MMA_LDST(4, m8n32k16_load_c_f16); |
13111 | case NVPTX::BI__hmma_m8n32k16_ld_c_f32: |
13112 | return MMA_LDST(8, m8n32k16_load_c_f32); |
13113 | |
13114 | // Integer MMA loads |
13115 | case NVPTX::BI__imma_m16n16k16_ld_a_s8: |
13116 | return MMA_LDST(2, m16n16k16_load_a_s8); |
13117 | case NVPTX::BI__imma_m16n16k16_ld_a_u8: |
13118 | return MMA_LDST(2, m16n16k16_load_a_u8); |
13119 | case NVPTX::BI__imma_m16n16k16_ld_b_s8: |
13120 | return MMA_LDST(2, m16n16k16_load_b_s8); |
13121 | case NVPTX::BI__imma_m16n16k16_ld_b_u8: |
13122 | return MMA_LDST(2, m16n16k16_load_b_u8); |
13123 | case NVPTX::BI__imma_m16n16k16_ld_c: |
13124 | return MMA_LDST(8, m16n16k16_load_c_s32); |
13125 | case NVPTX::BI__imma_m32n8k16_ld_a_s8: |
13126 | return MMA_LDST(4, m32n8k16_load_a_s8); |
13127 | case NVPTX::BI__imma_m32n8k16_ld_a_u8: |
13128 | return MMA_LDST(4, m32n8k16_load_a_u8); |
13129 | case NVPTX::BI__imma_m32n8k16_ld_b_s8: |
13130 | return MMA_LDST(1, m32n8k16_load_b_s8); |
13131 | case NVPTX::BI__imma_m32n8k16_ld_b_u8: |
13132 | return MMA_LDST(1, m32n8k16_load_b_u8); |
13133 | case NVPTX::BI__imma_m32n8k16_ld_c: |
13134 | return MMA_LDST(8, m32n8k16_load_c_s32); |
13135 | case NVPTX::BI__imma_m8n32k16_ld_a_s8: |
13136 | return MMA_LDST(1, m8n32k16_load_a_s8); |
13137 | case NVPTX::BI__imma_m8n32k16_ld_a_u8: |
13138 | return MMA_LDST(1, m8n32k16_load_a_u8); |
13139 | case NVPTX::BI__imma_m8n32k16_ld_b_s8: |
13140 | return MMA_LDST(4, m8n32k16_load_b_s8); |
13141 | case NVPTX::BI__imma_m8n32k16_ld_b_u8: |
13142 | return MMA_LDST(4, m8n32k16_load_b_u8); |
13143 | case NVPTX::BI__imma_m8n32k16_ld_c: |
13144 | return MMA_LDST(8, m8n32k16_load_c_s32); |
13145 | |
13146 | // Sub-integer MMA loads. |
13147 | // Only row/col layout is supported by A/B fragments. |
13148 | case NVPTX::BI__imma_m8n8k32_ld_a_s4: |
13149 | return {1, 0, MMA_INTR(m8n8k32_load_a_s4, row)}; |
13150 | case NVPTX::BI__imma_m8n8k32_ld_a_u4: |
13151 | return {1, 0, MMA_INTR(m8n8k32_load_a_u4, row)}; |
13152 | case NVPTX::BI__imma_m8n8k32_ld_b_s4: |
13153 | return {1, MMA_INTR(m8n8k32_load_b_s4, col), 0}; |
13154 | case NVPTX::BI__imma_m8n8k32_ld_b_u4: |
13155 | return {1, MMA_INTR(m8n8k32_load_b_u4, col), 0}; |
13156 | case NVPTX::BI__imma_m8n8k32_ld_c: |
13157 | return MMA_LDST(2, m8n8k32_load_c_s32); |
13158 | case NVPTX::BI__bmma_m8n8k128_ld_a_b1: |
13159 | return {1, 0, MMA_INTR(m8n8k128_load_a_b1, row)}; |
13160 | case NVPTX::BI__bmma_m8n8k128_ld_b_b1: |
13161 | return {1, MMA_INTR(m8n8k128_load_b_b1, col), 0}; |
13162 | case NVPTX::BI__bmma_m8n8k128_ld_c: |
13163 | return MMA_LDST(2, m8n8k128_load_c_s32); |
13164 | |
13165 | // NOTE: We need to follow inconsitent naming scheme used by NVCC. Unlike |
13166 | // PTX and LLVM IR where stores always use fragment D, NVCC builtins always |
13167 | // use fragment C for both loads and stores. |
13168 | // FP MMA stores. |
13169 | case NVPTX::BI__hmma_m16n16k16_st_c_f16: |
13170 | return MMA_LDST(4, m16n16k16_store_d_f16); |
13171 | case NVPTX::BI__hmma_m16n16k16_st_c_f32: |
13172 | return MMA_LDST(8, m16n16k16_store_d_f32); |
13173 | case NVPTX::BI__hmma_m32n8k16_st_c_f16: |
13174 | return MMA_LDST(4, m32n8k16_store_d_f16); |
13175 | case NVPTX::BI__hmma_m32n8k16_st_c_f32: |
13176 | return MMA_LDST(8, m32n8k16_store_d_f32); |
13177 | case NVPTX::BI__hmma_m8n32k16_st_c_f16: |
13178 | return MMA_LDST(4, m8n32k16_store_d_f16); |
13179 | case NVPTX::BI__hmma_m8n32k16_st_c_f32: |
13180 | return MMA_LDST(8, m8n32k16_store_d_f32); |
13181 | |
13182 | // Integer and sub-integer MMA stores. |
13183 | // Another naming quirk. Unlike other MMA builtins that use PTX types in the |
13184 | // name, integer loads/stores use LLVM's i32. |
13185 | case NVPTX::BI__imma_m16n16k16_st_c_i32: |
13186 | return MMA_LDST(8, m16n16k16_store_d_s32); |
13187 | case NVPTX::BI__imma_m32n8k16_st_c_i32: |
13188 | return MMA_LDST(8, m32n8k16_store_d_s32); |
13189 | case NVPTX::BI__imma_m8n32k16_st_c_i32: |
13190 | return MMA_LDST(8, m8n32k16_store_d_s32); |
13191 | case NVPTX::BI__imma_m8n8k32_st_c_i32: |
13192 | return MMA_LDST(2, m8n8k32_store_d_s32); |
13193 | case NVPTX::BI__bmma_m8n8k128_st_c_i32: |
13194 | return MMA_LDST(2, m8n8k128_store_d_s32); |
13195 | |
13196 | default: |
13197 | llvm_unreachable("Unknown MMA builtin")::llvm::llvm_unreachable_internal("Unknown MMA builtin", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 13197); |
13198 | } |
13199 | } |
13200 | #undef MMA_LDST |
13201 | #undef MMA_INTR |
13202 | |
13203 | |
13204 | struct NVPTXMmaInfo { |
13205 | unsigned NumEltsA; |
13206 | unsigned NumEltsB; |
13207 | unsigned NumEltsC; |
13208 | unsigned NumEltsD; |
13209 | std::array<unsigned, 8> Variants; |
13210 | |
13211 | unsigned getMMAIntrinsic(int Layout, bool Satf) { |
13212 | unsigned Index = Layout * 2 + Satf; |
13213 | if (Index >= Variants.size()) |
13214 | return 0; |
13215 | return Variants[Index]; |
13216 | } |
13217 | }; |
13218 | |
13219 | // Returns an intrinsic that matches Layout and Satf for valid combinations of |
13220 | // Layout and Satf, 0 otherwise. |
13221 | static NVPTXMmaInfo getNVPTXMmaInfo(unsigned BuiltinID) { |
13222 | // clang-format off |
13223 | #define MMA_VARIANTS(geom, type) {{ \ |
13224 | Intrinsic::nvvm_wmma_##geom##_mma_row_row_##type, \ |
13225 | Intrinsic::nvvm_wmma_##geom##_mma_row_row_##type##_satfinite, \ |
13226 | Intrinsic::nvvm_wmma_##geom##_mma_row_col_##type, \ |
13227 | Intrinsic::nvvm_wmma_##geom##_mma_row_col_##type##_satfinite, \ |
13228 | Intrinsic::nvvm_wmma_##geom##_mma_col_row_##type, \ |
13229 | Intrinsic::nvvm_wmma_##geom##_mma_col_row_##type##_satfinite, \ |
13230 | Intrinsic::nvvm_wmma_##geom##_mma_col_col_##type, \ |
13231 | Intrinsic::nvvm_wmma_##geom##_mma_col_col_##type##_satfinite \ |
13232 | }} |
13233 | // Sub-integer MMA only supports row.col layout. |
13234 | #define MMA_VARIANTS_I4(geom, type) {{ \ |
13235 | 0, \ |
13236 | 0, \ |
13237 | Intrinsic::nvvm_wmma_##geom##_mma_row_col_##type, \ |
13238 | Intrinsic::nvvm_wmma_##geom##_mma_row_col_##type##_satfinite, \ |
13239 | 0, \ |
13240 | 0, \ |
13241 | 0, \ |
13242 | 0 \ |
13243 | }} |
13244 | // b1 MMA does not support .satfinite. |
13245 | #define MMA_VARIANTS_B1(geom, type) {{ \ |
13246 | 0, \ |
13247 | 0, \ |
13248 | Intrinsic::nvvm_wmma_##geom##_mma_row_col_##type, \ |
13249 | 0, \ |
13250 | 0, \ |
13251 | 0, \ |
13252 | 0, \ |
13253 | 0 \ |
13254 | }} |
13255 | // clang-format on |
13256 | switch (BuiltinID) { |
13257 | // FP MMA |
13258 | // Note that 'type' argument of MMA_VARIANT uses D_C notation, while |
13259 | // NumEltsN of return value are ordered as A,B,C,D. |
13260 | case NVPTX::BI__hmma_m16n16k16_mma_f16f16: |
13261 | return {8, 8, 4, 4, MMA_VARIANTS(m16n16k16, f16_f16)}; |
13262 | case NVPTX::BI__hmma_m16n16k16_mma_f32f16: |
13263 | return {8, 8, 4, 8, MMA_VARIANTS(m16n16k16, f32_f16)}; |
13264 | case NVPTX::BI__hmma_m16n16k16_mma_f16f32: |
13265 | return {8, 8, 8, 4, MMA_VARIANTS(m16n16k16, f16_f32)}; |
13266 | case NVPTX::BI__hmma_m16n16k16_mma_f32f32: |
13267 | return {8, 8, 8, 8, MMA_VARIANTS(m16n16k16, f32_f32)}; |
13268 | case NVPTX::BI__hmma_m32n8k16_mma_f16f16: |
13269 | return {8, 8, 4, 4, MMA_VARIANTS(m32n8k16, f16_f16)}; |
13270 | case NVPTX::BI__hmma_m32n8k16_mma_f32f16: |
13271 | return {8, 8, 4, 8, MMA_VARIANTS(m32n8k16, f32_f16)}; |
13272 | case NVPTX::BI__hmma_m32n8k16_mma_f16f32: |
13273 | return {8, 8, 8, 4, MMA_VARIANTS(m32n8k16, f16_f32)}; |
13274 | case NVPTX::BI__hmma_m32n8k16_mma_f32f32: |
13275 | return {8, 8, 8, 8, MMA_VARIANTS(m32n8k16, f32_f32)}; |
13276 | case NVPTX::BI__hmma_m8n32k16_mma_f16f16: |
13277 | return {8, 8, 4, 4, MMA_VARIANTS(m8n32k16, f16_f16)}; |
13278 | case NVPTX::BI__hmma_m8n32k16_mma_f32f16: |
13279 | return {8, 8, 4, 8, MMA_VARIANTS(m8n32k16, f32_f16)}; |
13280 | case NVPTX::BI__hmma_m8n32k16_mma_f16f32: |
13281 | return {8, 8, 8, 4, MMA_VARIANTS(m8n32k16, f16_f32)}; |
13282 | case NVPTX::BI__hmma_m8n32k16_mma_f32f32: |
13283 | return {8, 8, 8, 8, MMA_VARIANTS(m8n32k16, f32_f32)}; |
13284 | |
13285 | // Integer MMA |
13286 | case NVPTX::BI__imma_m16n16k16_mma_s8: |
13287 | return {2, 2, 8, 8, MMA_VARIANTS(m16n16k16, s8)}; |
13288 | case NVPTX::BI__imma_m16n16k16_mma_u8: |
13289 | return {2, 2, 8, 8, MMA_VARIANTS(m16n16k16, u8)}; |
13290 | case NVPTX::BI__imma_m32n8k16_mma_s8: |
13291 | return {4, 1, 8, 8, MMA_VARIANTS(m32n8k16, s8)}; |
13292 | case NVPTX::BI__imma_m32n8k16_mma_u8: |
13293 | return {4, 1, 8, 8, MMA_VARIANTS(m32n8k16, u8)}; |
13294 | case NVPTX::BI__imma_m8n32k16_mma_s8: |
13295 | return {1, 4, 8, 8, MMA_VARIANTS(m8n32k16, s8)}; |
13296 | case NVPTX::BI__imma_m8n32k16_mma_u8: |
13297 | return {1, 4, 8, 8, MMA_VARIANTS(m8n32k16, u8)}; |
13298 | |
13299 | // Sub-integer MMA |
13300 | case NVPTX::BI__imma_m8n8k32_mma_s4: |
13301 | return {1, 1, 2, 2, MMA_VARIANTS_I4(m8n8k32, s4)}; |
13302 | case NVPTX::BI__imma_m8n8k32_mma_u4: |
13303 | return {1, 1, 2, 2, MMA_VARIANTS_I4(m8n8k32, u4)}; |
13304 | case NVPTX::BI__bmma_m8n8k128_mma_xor_popc_b1: |
13305 | return {1, 1, 2, 2, MMA_VARIANTS_B1(m8n8k128, b1)}; |
13306 | default: |
13307 | llvm_unreachable("Unexpected builtin ID.")::llvm::llvm_unreachable_internal("Unexpected builtin ID.", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 13307); |
13308 | } |
13309 | #undef MMA_VARIANTS |
13310 | #undef MMA_VARIANTS_I4 |
13311 | #undef MMA_VARIANTS_B1 |
13312 | } |
13313 | |
13314 | } // namespace |
13315 | |
13316 | Value * |
13317 | CodeGenFunction::EmitNVPTXBuiltinExpr(unsigned BuiltinID, const CallExpr *E) { |
13318 | auto MakeLdg = [&](unsigned IntrinsicID) { |
13319 | Value *Ptr = EmitScalarExpr(E->getArg(0)); |
13320 | clang::CharUnits Align = |
13321 | getNaturalPointeeTypeAlignment(E->getArg(0)->getType()); |
13322 | return Builder.CreateCall( |
13323 | CGM.getIntrinsic(IntrinsicID, {Ptr->getType()->getPointerElementType(), |
13324 | Ptr->getType()}), |
13325 | {Ptr, ConstantInt::get(Builder.getInt32Ty(), Align.getQuantity())}); |
13326 | }; |
13327 | auto MakeScopedAtomic = [&](unsigned IntrinsicID) { |
13328 | Value *Ptr = EmitScalarExpr(E->getArg(0)); |
13329 | return Builder.CreateCall( |
13330 | CGM.getIntrinsic(IntrinsicID, {Ptr->getType()->getPointerElementType(), |
13331 | Ptr->getType()}), |
13332 | {Ptr, EmitScalarExpr(E->getArg(1))}); |
13333 | }; |
13334 | switch (BuiltinID) { |
13335 | case NVPTX::BI__nvvm_atom_add_gen_i: |
13336 | case NVPTX::BI__nvvm_atom_add_gen_l: |
13337 | case NVPTX::BI__nvvm_atom_add_gen_ll: |
13338 | return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Add, E); |
13339 | |
13340 | case NVPTX::BI__nvvm_atom_sub_gen_i: |
13341 | case NVPTX::BI__nvvm_atom_sub_gen_l: |
13342 | case NVPTX::BI__nvvm_atom_sub_gen_ll: |
13343 | return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Sub, E); |
13344 | |
13345 | case NVPTX::BI__nvvm_atom_and_gen_i: |
13346 | case NVPTX::BI__nvvm_atom_and_gen_l: |
13347 | case NVPTX::BI__nvvm_atom_and_gen_ll: |
13348 | return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::And, E); |
13349 | |
13350 | case NVPTX::BI__nvvm_atom_or_gen_i: |
13351 | case NVPTX::BI__nvvm_atom_or_gen_l: |
13352 | case NVPTX::BI__nvvm_atom_or_gen_ll: |
13353 | return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Or, E); |
13354 | |
13355 | case NVPTX::BI__nvvm_atom_xor_gen_i: |
13356 | case NVPTX::BI__nvvm_atom_xor_gen_l: |
13357 | case NVPTX::BI__nvvm_atom_xor_gen_ll: |
13358 | return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Xor, E); |
13359 | |
13360 | case NVPTX::BI__nvvm_atom_xchg_gen_i: |
13361 | case NVPTX::BI__nvvm_atom_xchg_gen_l: |
13362 | case NVPTX::BI__nvvm_atom_xchg_gen_ll: |
13363 | return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Xchg, E); |
13364 | |
13365 | case NVPTX::BI__nvvm_atom_max_gen_i: |
13366 | case NVPTX::BI__nvvm_atom_max_gen_l: |
13367 | case NVPTX::BI__nvvm_atom_max_gen_ll: |
13368 | return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Max, E); |
13369 | |
13370 | case NVPTX::BI__nvvm_atom_max_gen_ui: |
13371 | case NVPTX::BI__nvvm_atom_max_gen_ul: |
13372 | case NVPTX::BI__nvvm_atom_max_gen_ull: |
13373 | return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::UMax, E); |
13374 | |
13375 | case NVPTX::BI__nvvm_atom_min_gen_i: |
13376 | case NVPTX::BI__nvvm_atom_min_gen_l: |
13377 | case NVPTX::BI__nvvm_atom_min_gen_ll: |
13378 | return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Min, E); |
13379 | |
13380 | case NVPTX::BI__nvvm_atom_min_gen_ui: |
13381 | case NVPTX::BI__nvvm_atom_min_gen_ul: |
13382 | case NVPTX::BI__nvvm_atom_min_gen_ull: |
13383 | return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::UMin, E); |
13384 | |
13385 | case NVPTX::BI__nvvm_atom_cas_gen_i: |
13386 | case NVPTX::BI__nvvm_atom_cas_gen_l: |
13387 | case NVPTX::BI__nvvm_atom_cas_gen_ll: |
13388 | // __nvvm_atom_cas_gen_* should return the old value rather than the |
13389 | // success flag. |
13390 | return MakeAtomicCmpXchgValue(*this, E, /*ReturnBool=*/false); |
13391 | |
13392 | case NVPTX::BI__nvvm_atom_add_gen_f: { |
13393 | Value *Ptr = EmitScalarExpr(E->getArg(0)); |
13394 | Value *Val = EmitScalarExpr(E->getArg(1)); |
13395 | // atomicrmw only deals with integer arguments so we need to use |
13396 | // LLVM's nvvm_atomic_load_add_f32 intrinsic for that. |
13397 | Function *FnALAF32 = |
13398 | CGM.getIntrinsic(Intrinsic::nvvm_atomic_load_add_f32, Ptr->getType()); |
13399 | return Builder.CreateCall(FnALAF32, {Ptr, Val}); |
13400 | } |
13401 | |
13402 | case NVPTX::BI__nvvm_atom_add_gen_d: { |
13403 | Value *Ptr = EmitScalarExpr(E->getArg(0)); |
13404 | Value *Val = EmitScalarExpr(E->getArg(1)); |
13405 | // atomicrmw only deals with integer arguments, so we need to use |
13406 | // LLVM's nvvm_atomic_load_add_f64 intrinsic. |
13407 | Function *FnALAF64 = |
13408 | CGM.getIntrinsic(Intrinsic::nvvm_atomic_load_add_f64, Ptr->getType()); |
13409 | return Builder.CreateCall(FnALAF64, {Ptr, Val}); |
13410 | } |
13411 | |
13412 | case NVPTX::BI__nvvm_atom_inc_gen_ui: { |
13413 | Value *Ptr = EmitScalarExpr(E->getArg(0)); |
13414 | Value *Val = EmitScalarExpr(E->getArg(1)); |
13415 | Function *FnALI32 = |
13416 | CGM.getIntrinsic(Intrinsic::nvvm_atomic_load_inc_32, Ptr->getType()); |
13417 | return Builder.CreateCall(FnALI32, {Ptr, Val}); |
13418 | } |
13419 | |
13420 | case NVPTX::BI__nvvm_atom_dec_gen_ui: { |
13421 | Value *Ptr = EmitScalarExpr(E->getArg(0)); |
13422 | Value *Val = EmitScalarExpr(E->getArg(1)); |
13423 | Function *FnALD32 = |
13424 | CGM.getIntrinsic(Intrinsic::nvvm_atomic_load_dec_32, Ptr->getType()); |
13425 | return Builder.CreateCall(FnALD32, {Ptr, Val}); |
13426 | } |
13427 | |
13428 | case NVPTX::BI__nvvm_ldg_c: |
13429 | case NVPTX::BI__nvvm_ldg_c2: |
13430 | case NVPTX::BI__nvvm_ldg_c4: |
13431 | case NVPTX::BI__nvvm_ldg_s: |
13432 | case NVPTX::BI__nvvm_ldg_s2: |
13433 | case NVPTX::BI__nvvm_ldg_s4: |
13434 | case NVPTX::BI__nvvm_ldg_i: |
13435 | case NVPTX::BI__nvvm_ldg_i2: |
13436 | case NVPTX::BI__nvvm_ldg_i4: |
13437 | case NVPTX::BI__nvvm_ldg_l: |
13438 | case NVPTX::BI__nvvm_ldg_ll: |
13439 | case NVPTX::BI__nvvm_ldg_ll2: |
13440 | case NVPTX::BI__nvvm_ldg_uc: |
13441 | case NVPTX::BI__nvvm_ldg_uc2: |
13442 | case NVPTX::BI__nvvm_ldg_uc4: |
13443 | case NVPTX::BI__nvvm_ldg_us: |
13444 | case NVPTX::BI__nvvm_ldg_us2: |
13445 | case NVPTX::BI__nvvm_ldg_us4: |
13446 | case NVPTX::BI__nvvm_ldg_ui: |
13447 | case NVPTX::BI__nvvm_ldg_ui2: |
13448 | case NVPTX::BI__nvvm_ldg_ui4: |
13449 | case NVPTX::BI__nvvm_ldg_ul: |
13450 | case NVPTX::BI__nvvm_ldg_ull: |
13451 | case NVPTX::BI__nvvm_ldg_ull2: |
13452 | // PTX Interoperability section 2.2: "For a vector with an even number of |
13453 | // elements, its alignment is set to number of elements times the alignment |
13454 | // of its member: n*alignof(t)." |
13455 | return MakeLdg(Intrinsic::nvvm_ldg_global_i); |
13456 | case NVPTX::BI__nvvm_ldg_f: |
13457 | case NVPTX::BI__nvvm_ldg_f2: |
13458 | case NVPTX::BI__nvvm_ldg_f4: |
13459 | case NVPTX::BI__nvvm_ldg_d: |
13460 | case NVPTX::BI__nvvm_ldg_d2: |
13461 | return MakeLdg(Intrinsic::nvvm_ldg_global_f); |
13462 | |
13463 | case NVPTX::BI__nvvm_atom_cta_add_gen_i: |
13464 | case NVPTX::BI__nvvm_atom_cta_add_gen_l: |
13465 | case NVPTX::BI__nvvm_atom_cta_add_gen_ll: |
13466 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_add_gen_i_cta); |
13467 | case NVPTX::BI__nvvm_atom_sys_add_gen_i: |
13468 | case NVPTX::BI__nvvm_atom_sys_add_gen_l: |
13469 | case NVPTX::BI__nvvm_atom_sys_add_gen_ll: |
13470 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_add_gen_i_sys); |
13471 | case NVPTX::BI__nvvm_atom_cta_add_gen_f: |
13472 | case NVPTX::BI__nvvm_atom_cta_add_gen_d: |
13473 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_add_gen_f_cta); |
13474 | case NVPTX::BI__nvvm_atom_sys_add_gen_f: |
13475 | case NVPTX::BI__nvvm_atom_sys_add_gen_d: |
13476 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_add_gen_f_sys); |
13477 | case NVPTX::BI__nvvm_atom_cta_xchg_gen_i: |
13478 | case NVPTX::BI__nvvm_atom_cta_xchg_gen_l: |
13479 | case NVPTX::BI__nvvm_atom_cta_xchg_gen_ll: |
13480 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_exch_gen_i_cta); |
13481 | case NVPTX::BI__nvvm_atom_sys_xchg_gen_i: |
13482 | case NVPTX::BI__nvvm_atom_sys_xchg_gen_l: |
13483 | case NVPTX::BI__nvvm_atom_sys_xchg_gen_ll: |
13484 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_exch_gen_i_sys); |
13485 | case NVPTX::BI__nvvm_atom_cta_max_gen_i: |
13486 | case NVPTX::BI__nvvm_atom_cta_max_gen_ui: |
13487 | case NVPTX::BI__nvvm_atom_cta_max_gen_l: |
13488 | case NVPTX::BI__nvvm_atom_cta_max_gen_ul: |
13489 | case NVPTX::BI__nvvm_atom_cta_max_gen_ll: |
13490 | case NVPTX::BI__nvvm_atom_cta_max_gen_ull: |
13491 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_max_gen_i_cta); |
13492 | case NVPTX::BI__nvvm_atom_sys_max_gen_i: |
13493 | case NVPTX::BI__nvvm_atom_sys_max_gen_ui: |
13494 | case NVPTX::BI__nvvm_atom_sys_max_gen_l: |
13495 | case NVPTX::BI__nvvm_atom_sys_max_gen_ul: |
13496 | case NVPTX::BI__nvvm_atom_sys_max_gen_ll: |
13497 | case NVPTX::BI__nvvm_atom_sys_max_gen_ull: |
13498 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_max_gen_i_sys); |
13499 | case NVPTX::BI__nvvm_atom_cta_min_gen_i: |
13500 | case NVPTX::BI__nvvm_atom_cta_min_gen_ui: |
13501 | case NVPTX::BI__nvvm_atom_cta_min_gen_l: |
13502 | case NVPTX::BI__nvvm_atom_cta_min_gen_ul: |
13503 | case NVPTX::BI__nvvm_atom_cta_min_gen_ll: |
13504 | case NVPTX::BI__nvvm_atom_cta_min_gen_ull: |
13505 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_min_gen_i_cta); |
13506 | case NVPTX::BI__nvvm_atom_sys_min_gen_i: |
13507 | case NVPTX::BI__nvvm_atom_sys_min_gen_ui: |
13508 | case NVPTX::BI__nvvm_atom_sys_min_gen_l: |
13509 | case NVPTX::BI__nvvm_atom_sys_min_gen_ul: |
13510 | case NVPTX::BI__nvvm_atom_sys_min_gen_ll: |
13511 | case NVPTX::BI__nvvm_atom_sys_min_gen_ull: |
13512 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_min_gen_i_sys); |
13513 | case NVPTX::BI__nvvm_atom_cta_inc_gen_ui: |
13514 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_inc_gen_i_cta); |
13515 | case NVPTX::BI__nvvm_atom_cta_dec_gen_ui: |
13516 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_dec_gen_i_cta); |
13517 | case NVPTX::BI__nvvm_atom_sys_inc_gen_ui: |
13518 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_inc_gen_i_sys); |
13519 | case NVPTX::BI__nvvm_atom_sys_dec_gen_ui: |
13520 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_dec_gen_i_sys); |
13521 | case NVPTX::BI__nvvm_atom_cta_and_gen_i: |
13522 | case NVPTX::BI__nvvm_atom_cta_and_gen_l: |
13523 | case NVPTX::BI__nvvm_atom_cta_and_gen_ll: |
13524 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_and_gen_i_cta); |
13525 | case NVPTX::BI__nvvm_atom_sys_and_gen_i: |
13526 | case NVPTX::BI__nvvm_atom_sys_and_gen_l: |
13527 | case NVPTX::BI__nvvm_atom_sys_and_gen_ll: |
13528 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_and_gen_i_sys); |
13529 | case NVPTX::BI__nvvm_atom_cta_or_gen_i: |
13530 | case NVPTX::BI__nvvm_atom_cta_or_gen_l: |
13531 | case NVPTX::BI__nvvm_atom_cta_or_gen_ll: |
13532 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_or_gen_i_cta); |
13533 | case NVPTX::BI__nvvm_atom_sys_or_gen_i: |
13534 | case NVPTX::BI__nvvm_atom_sys_or_gen_l: |
13535 | case NVPTX::BI__nvvm_atom_sys_or_gen_ll: |
13536 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_or_gen_i_sys); |
13537 | case NVPTX::BI__nvvm_atom_cta_xor_gen_i: |
13538 | case NVPTX::BI__nvvm_atom_cta_xor_gen_l: |
13539 | case NVPTX::BI__nvvm_atom_cta_xor_gen_ll: |
13540 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_xor_gen_i_cta); |
13541 | case NVPTX::BI__nvvm_atom_sys_xor_gen_i: |
13542 | case NVPTX::BI__nvvm_atom_sys_xor_gen_l: |
13543 | case NVPTX::BI__nvvm_atom_sys_xor_gen_ll: |
13544 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_xor_gen_i_sys); |
13545 | case NVPTX::BI__nvvm_atom_cta_cas_gen_i: |
13546 | case NVPTX::BI__nvvm_atom_cta_cas_gen_l: |
13547 | case NVPTX::BI__nvvm_atom_cta_cas_gen_ll: { |
13548 | Value *Ptr = EmitScalarExpr(E->getArg(0)); |
13549 | return Builder.CreateCall( |
13550 | CGM.getIntrinsic( |
13551 | Intrinsic::nvvm_atomic_cas_gen_i_cta, |
13552 | {Ptr->getType()->getPointerElementType(), Ptr->getType()}), |
13553 | {Ptr, EmitScalarExpr(E->getArg(1)), EmitScalarExpr(E->getArg(2))}); |
13554 | } |
13555 | case NVPTX::BI__nvvm_atom_sys_cas_gen_i: |
13556 | case NVPTX::BI__nvvm_atom_sys_cas_gen_l: |
13557 | case NVPTX::BI__nvvm_atom_sys_cas_gen_ll: { |
13558 | Value *Ptr = EmitScalarExpr(E->getArg(0)); |
13559 | return Builder.CreateCall( |
13560 | CGM.getIntrinsic( |
13561 | Intrinsic::nvvm_atomic_cas_gen_i_sys, |
13562 | {Ptr->getType()->getPointerElementType(), Ptr->getType()}), |
13563 | {Ptr, EmitScalarExpr(E->getArg(1)), EmitScalarExpr(E->getArg(2))}); |
13564 | } |
13565 | case NVPTX::BI__nvvm_match_all_sync_i32p: |
13566 | case NVPTX::BI__nvvm_match_all_sync_i64p: { |
13567 | Value *Mask = EmitScalarExpr(E->getArg(0)); |
13568 | Value *Val = EmitScalarExpr(E->getArg(1)); |
13569 | Address PredOutPtr = EmitPointerWithAlignment(E->getArg(2)); |
13570 | Value *ResultPair = Builder.CreateCall( |
13571 | CGM.getIntrinsic(BuiltinID == NVPTX::BI__nvvm_match_all_sync_i32p |
13572 | ? Intrinsic::nvvm_match_all_sync_i32p |
13573 | : Intrinsic::nvvm_match_all_sync_i64p), |
13574 | {Mask, Val}); |
13575 | Value *Pred = Builder.CreateZExt(Builder.CreateExtractValue(ResultPair, 1), |
13576 | PredOutPtr.getElementType()); |
13577 | Builder.CreateStore(Pred, PredOutPtr); |
13578 | return Builder.CreateExtractValue(ResultPair, 0); |
13579 | } |
13580 | |
13581 | // FP MMA loads |
13582 | case NVPTX::BI__hmma_m16n16k16_ld_a: |
13583 | case NVPTX::BI__hmma_m16n16k16_ld_b: |
13584 | case NVPTX::BI__hmma_m16n16k16_ld_c_f16: |
13585 | case NVPTX::BI__hmma_m16n16k16_ld_c_f32: |
13586 | case NVPTX::BI__hmma_m32n8k16_ld_a: |
13587 | case NVPTX::BI__hmma_m32n8k16_ld_b: |
13588 | case NVPTX::BI__hmma_m32n8k16_ld_c_f16: |
13589 | case NVPTX::BI__hmma_m32n8k16_ld_c_f32: |
13590 | case NVPTX::BI__hmma_m8n32k16_ld_a: |
13591 | case NVPTX::BI__hmma_m8n32k16_ld_b: |
13592 | case NVPTX::BI__hmma_m8n32k16_ld_c_f16: |
13593 | case NVPTX::BI__hmma_m8n32k16_ld_c_f32: |
13594 | // Integer MMA loads. |
13595 | case NVPTX::BI__imma_m16n16k16_ld_a_s8: |
13596 | case NVPTX::BI__imma_m16n16k16_ld_a_u8: |
13597 | case NVPTX::BI__imma_m16n16k16_ld_b_s8: |
13598 | case NVPTX::BI__imma_m16n16k16_ld_b_u8: |
13599 | case NVPTX::BI__imma_m16n16k16_ld_c: |
13600 | case NVPTX::BI__imma_m32n8k16_ld_a_s8: |
13601 | case NVPTX::BI__imma_m32n8k16_ld_a_u8: |
13602 | case NVPTX::BI__imma_m32n8k16_ld_b_s8: |
13603 | case NVPTX::BI__imma_m32n8k16_ld_b_u8: |
13604 | case NVPTX::BI__imma_m32n8k16_ld_c: |
13605 | case NVPTX::BI__imma_m8n32k16_ld_a_s8: |
13606 | case NVPTX::BI__imma_m8n32k16_ld_a_u8: |
13607 | case NVPTX::BI__imma_m8n32k16_ld_b_s8: |
13608 | case NVPTX::BI__imma_m8n32k16_ld_b_u8: |
13609 | case NVPTX::BI__imma_m8n32k16_ld_c: |
13610 | // Sub-integer MMA loads. |
13611 | case NVPTX::BI__imma_m8n8k32_ld_a_s4: |
13612 | case NVPTX::BI__imma_m8n8k32_ld_a_u4: |
13613 | case NVPTX::BI__imma_m8n8k32_ld_b_s4: |
13614 | case NVPTX::BI__imma_m8n8k32_ld_b_u4: |
13615 | case NVPTX::BI__imma_m8n8k32_ld_c: |
13616 | case NVPTX::BI__bmma_m8n8k128_ld_a_b1: |
13617 | case NVPTX::BI__bmma_m8n8k128_ld_b_b1: |
13618 | case NVPTX::BI__bmma_m8n8k128_ld_c: |
13619 | { |
13620 | Address Dst = EmitPointerWithAlignment(E->getArg(0)); |
13621 | Value *Src = EmitScalarExpr(E->getArg(1)); |
13622 | Value *Ldm = EmitScalarExpr(E->getArg(2)); |
13623 | llvm::APSInt isColMajorArg; |
13624 | if (!E->getArg(3)->isIntegerConstantExpr(isColMajorArg, getContext())) |
13625 | return nullptr; |
13626 | bool isColMajor = isColMajorArg.getSExtValue(); |
13627 | NVPTXMmaLdstInfo II = getNVPTXMmaLdstInfo(BuiltinID); |
13628 | unsigned IID = isColMajor ? II.IID_col : II.IID_row; |
13629 | if (IID == 0) |
13630 | return nullptr; |
13631 | |
13632 | Value *Result = |
13633 | Builder.CreateCall(CGM.getIntrinsic(IID, Src->getType()), {Src, Ldm}); |
13634 | |
13635 | // Save returned values. |
13636 | assert(II.NumResults)((II.NumResults) ? static_cast<void> (0) : __assert_fail ("II.NumResults", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 13636, __PRETTY_FUNCTION__)); |
13637 | if (II.NumResults == 1) { |
13638 | Builder.CreateAlignedStore(Result, Dst.getPointer(), |
13639 | CharUnits::fromQuantity(4)); |
13640 | } else { |
13641 | for (unsigned i = 0; i < II.NumResults; ++i) { |
13642 | Builder.CreateAlignedStore( |
13643 | Builder.CreateBitCast(Builder.CreateExtractValue(Result, i), |
13644 | Dst.getElementType()), |
13645 | Builder.CreateGEP(Dst.getPointer(), |
13646 | llvm::ConstantInt::get(IntTy, i)), |
13647 | CharUnits::fromQuantity(4)); |
13648 | } |
13649 | } |
13650 | return Result; |
13651 | } |
13652 | |
13653 | case NVPTX::BI__hmma_m16n16k16_st_c_f16: |
13654 | case NVPTX::BI__hmma_m16n16k16_st_c_f32: |
13655 | case NVPTX::BI__hmma_m32n8k16_st_c_f16: |
13656 | case NVPTX::BI__hmma_m32n8k16_st_c_f32: |
13657 | case NVPTX::BI__hmma_m8n32k16_st_c_f16: |
13658 | case NVPTX::BI__hmma_m8n32k16_st_c_f32: |
13659 | case NVPTX::BI__imma_m16n16k16_st_c_i32: |
13660 | case NVPTX::BI__imma_m32n8k16_st_c_i32: |
13661 | case NVPTX::BI__imma_m8n32k16_st_c_i32: |
13662 | case NVPTX::BI__imma_m8n8k32_st_c_i32: |
13663 | case NVPTX::BI__bmma_m8n8k128_st_c_i32: { |
13664 | Value *Dst = EmitScalarExpr(E->getArg(0)); |
13665 | Address Src = EmitPointerWithAlignment(E->getArg(1)); |
13666 | Value *Ldm = EmitScalarExpr(E->getArg(2)); |
13667 | llvm::APSInt isColMajorArg; |
13668 | if (!E->getArg(3)->isIntegerConstantExpr(isColMajorArg, getContext())) |
13669 | return nullptr; |
13670 | bool isColMajor = isColMajorArg.getSExtValue(); |
13671 | NVPTXMmaLdstInfo II = getNVPTXMmaLdstInfo(BuiltinID); |
13672 | unsigned IID = isColMajor ? II.IID_col : II.IID_row; |
13673 | if (IID == 0) |
13674 | return nullptr; |
13675 | Function *Intrinsic = |
13676 | CGM.getIntrinsic(IID, Dst->getType()); |
13677 | llvm::Type *ParamType = Intrinsic->getFunctionType()->getParamType(1); |
13678 | SmallVector<Value *, 10> Values = {Dst}; |
13679 | for (unsigned i = 0; i < II.NumResults; ++i) { |
13680 | Value *V = Builder.CreateAlignedLoad( |
13681 | Builder.CreateGEP(Src.getPointer(), llvm::ConstantInt::get(IntTy, i)), |
13682 | CharUnits::fromQuantity(4)); |
13683 | Values.push_back(Builder.CreateBitCast(V, ParamType)); |
13684 | } |
13685 | Values.push_back(Ldm); |
13686 | Value *Result = Builder.CreateCall(Intrinsic, Values); |
13687 | return Result; |
13688 | } |
13689 | |
13690 | // BI__hmma_m16n16k16_mma_<Dtype><CType>(d, a, b, c, layout, satf) --> |
13691 | // Intrinsic::nvvm_wmma_m16n16k16_mma_sync<layout A,B><DType><CType><Satf> |
13692 | case NVPTX::BI__hmma_m16n16k16_mma_f16f16: |
13693 | case NVPTX::BI__hmma_m16n16k16_mma_f32f16: |
13694 | case NVPTX::BI__hmma_m16n16k16_mma_f32f32: |
13695 | case NVPTX::BI__hmma_m16n16k16_mma_f16f32: |
13696 | case NVPTX::BI__hmma_m32n8k16_mma_f16f16: |
13697 | case NVPTX::BI__hmma_m32n8k16_mma_f32f16: |
13698 | case NVPTX::BI__hmma_m32n8k16_mma_f32f32: |
13699 | case NVPTX::BI__hmma_m32n8k16_mma_f16f32: |
13700 | case NVPTX::BI__hmma_m8n32k16_mma_f16f16: |
13701 | case NVPTX::BI__hmma_m8n32k16_mma_f32f16: |
13702 | case NVPTX::BI__hmma_m8n32k16_mma_f32f32: |
13703 | case NVPTX::BI__hmma_m8n32k16_mma_f16f32: |
13704 | case NVPTX::BI__imma_m16n16k16_mma_s8: |
13705 | case NVPTX::BI__imma_m16n16k16_mma_u8: |
13706 | case NVPTX::BI__imma_m32n8k16_mma_s8: |
13707 | case NVPTX::BI__imma_m32n8k16_mma_u8: |
13708 | case NVPTX::BI__imma_m8n32k16_mma_s8: |
13709 | case NVPTX::BI__imma_m8n32k16_mma_u8: |
13710 | case NVPTX::BI__imma_m8n8k32_mma_s4: |
13711 | case NVPTX::BI__imma_m8n8k32_mma_u4: |
13712 | case NVPTX::BI__bmma_m8n8k128_mma_xor_popc_b1: { |
13713 | Address Dst = EmitPointerWithAlignment(E->getArg(0)); |
13714 | Address SrcA = EmitPointerWithAlignment(E->getArg(1)); |
13715 | Address SrcB = EmitPointerWithAlignment(E->getArg(2)); |
13716 | Address SrcC = EmitPointerWithAlignment(E->getArg(3)); |
13717 | llvm::APSInt LayoutArg; |
13718 | if (!E->getArg(4)->isIntegerConstantExpr(LayoutArg, getContext())) |
13719 | return nullptr; |
13720 | int Layout = LayoutArg.getSExtValue(); |
13721 | if (Layout < 0 || Layout > 3) |
13722 | return nullptr; |
13723 | llvm::APSInt SatfArg; |
13724 | if (BuiltinID == NVPTX::BI__bmma_m8n8k128_mma_xor_popc_b1) |
13725 | SatfArg = 0; // .b1 does not have satf argument. |
13726 | else if (!E->getArg(5)->isIntegerConstantExpr(SatfArg, getContext())) |
13727 | return nullptr; |
13728 | bool Satf = SatfArg.getSExtValue(); |
13729 | NVPTXMmaInfo MI = getNVPTXMmaInfo(BuiltinID); |
13730 | unsigned IID = MI.getMMAIntrinsic(Layout, Satf); |
13731 | if (IID == 0) // Unsupported combination of Layout/Satf. |
13732 | return nullptr; |
13733 | |
13734 | SmallVector<Value *, 24> Values; |
13735 | Function *Intrinsic = CGM.getIntrinsic(IID); |
13736 | llvm::Type *AType = Intrinsic->getFunctionType()->getParamType(0); |
13737 | // Load A |
13738 | for (unsigned i = 0; i < MI.NumEltsA; ++i) { |
13739 | Value *V = Builder.CreateAlignedLoad( |
13740 | Builder.CreateGEP(SrcA.getPointer(), |
13741 | llvm::ConstantInt::get(IntTy, i)), |
13742 | CharUnits::fromQuantity(4)); |
13743 | Values.push_back(Builder.CreateBitCast(V, AType)); |
13744 | } |
13745 | // Load B |
13746 | llvm::Type *BType = Intrinsic->getFunctionType()->getParamType(MI.NumEltsA); |
13747 | for (unsigned i = 0; i < MI.NumEltsB; ++i) { |
13748 | Value *V = Builder.CreateAlignedLoad( |
13749 | Builder.CreateGEP(SrcB.getPointer(), |
13750 | llvm::ConstantInt::get(IntTy, i)), |
13751 | CharUnits::fromQuantity(4)); |
13752 | Values.push_back(Builder.CreateBitCast(V, BType)); |
13753 | } |
13754 | // Load C |
13755 | llvm::Type *CType = |
13756 | Intrinsic->getFunctionType()->getParamType(MI.NumEltsA + MI.NumEltsB); |
13757 | for (unsigned i = 0; i < MI.NumEltsC; ++i) { |
13758 | Value *V = Builder.CreateAlignedLoad( |
13759 | Builder.CreateGEP(SrcC.getPointer(), |
13760 | llvm::ConstantInt::get(IntTy, i)), |
13761 | CharUnits::fromQuantity(4)); |
13762 | Values.push_back(Builder.CreateBitCast(V, CType)); |
13763 | } |
13764 | Value *Result = Builder.CreateCall(Intrinsic, Values); |
13765 | llvm::Type *DType = Dst.getElementType(); |
13766 | for (unsigned i = 0; i < MI.NumEltsD; ++i) |
13767 | Builder.CreateAlignedStore( |
13768 | Builder.CreateBitCast(Builder.CreateExtractValue(Result, i), DType), |
13769 | Builder.CreateGEP(Dst.getPointer(), llvm::ConstantInt::get(IntTy, i)), |
13770 | CharUnits::fromQuantity(4)); |
13771 | return Result; |
13772 | } |
13773 | default: |
13774 | return nullptr; |
13775 | } |
13776 | } |
13777 | |
13778 | Value *CodeGenFunction::EmitWebAssemblyBuiltinExpr(unsigned BuiltinID, |
13779 | const CallExpr *E) { |
13780 | switch (BuiltinID) { |
13781 | case WebAssembly::BI__builtin_wasm_memory_size: { |
13782 | llvm::Type *ResultType = ConvertType(E->getType()); |
13783 | Value *I = EmitScalarExpr(E->getArg(0)); |
13784 | Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_memory_size, ResultType); |
13785 | return Builder.CreateCall(Callee, I); |
13786 | } |
13787 | case WebAssembly::BI__builtin_wasm_memory_grow: { |
13788 | llvm::Type *ResultType = ConvertType(E->getType()); |
13789 | Value *Args[] = { |
13790 | EmitScalarExpr(E->getArg(0)), |
13791 | EmitScalarExpr(E->getArg(1)) |
13792 | }; |
13793 | Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_memory_grow, ResultType); |
13794 | return Builder.CreateCall(Callee, Args); |
13795 | } |
13796 | case WebAssembly::BI__builtin_wasm_memory_init: { |
13797 | llvm::APSInt SegConst; |
13798 | if (!E->getArg(0)->isIntegerConstantExpr(SegConst, getContext())) |
13799 | llvm_unreachable("Constant arg isn't actually constant?")::llvm::llvm_unreachable_internal("Constant arg isn't actually constant?" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 13799); |
13800 | llvm::APSInt MemConst; |
13801 | if (!E->getArg(1)->isIntegerConstantExpr(MemConst, getContext())) |
13802 | llvm_unreachable("Constant arg isn't actually constant?")::llvm::llvm_unreachable_internal("Constant arg isn't actually constant?" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 13802); |
13803 | if (!MemConst.isNullValue()) |
13804 | ErrorUnsupported(E, "non-zero memory index"); |
13805 | Value *Args[] = {llvm::ConstantInt::get(getLLVMContext(), SegConst), |
13806 | llvm::ConstantInt::get(getLLVMContext(), MemConst), |
13807 | EmitScalarExpr(E->getArg(2)), EmitScalarExpr(E->getArg(3)), |
13808 | EmitScalarExpr(E->getArg(4))}; |
13809 | Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_memory_init); |
13810 | return Builder.CreateCall(Callee, Args); |
13811 | } |
13812 | case WebAssembly::BI__builtin_wasm_data_drop: { |
13813 | llvm::APSInt SegConst; |
13814 | if (!E->getArg(0)->isIntegerConstantExpr(SegConst, getContext())) |
13815 | llvm_unreachable("Constant arg isn't actually constant?")::llvm::llvm_unreachable_internal("Constant arg isn't actually constant?" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 13815); |
13816 | Value *Arg = llvm::ConstantInt::get(getLLVMContext(), SegConst); |
13817 | Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_data_drop); |
13818 | return Builder.CreateCall(Callee, {Arg}); |
13819 | } |
13820 | case WebAssembly::BI__builtin_wasm_throw: { |
13821 | Value *Tag = EmitScalarExpr(E->getArg(0)); |
13822 | Value *Obj = EmitScalarExpr(E->getArg(1)); |
13823 | Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_throw); |
13824 | return Builder.CreateCall(Callee, {Tag, Obj}); |
13825 | } |
13826 | case WebAssembly::BI__builtin_wasm_rethrow_in_catch: { |
13827 | Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_rethrow_in_catch); |
13828 | return Builder.CreateCall(Callee); |
13829 | } |
13830 | case WebAssembly::BI__builtin_wasm_atomic_wait_i32: { |
13831 | Value *Addr = EmitScalarExpr(E->getArg(0)); |
13832 | Value *Expected = EmitScalarExpr(E->getArg(1)); |
13833 | Value *Timeout = EmitScalarExpr(E->getArg(2)); |
13834 | Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_atomic_wait_i32); |
13835 | return Builder.CreateCall(Callee, {Addr, Expected, Timeout}); |
13836 | } |
13837 | case WebAssembly::BI__builtin_wasm_atomic_wait_i64: { |
13838 | Value *Addr = EmitScalarExpr(E->getArg(0)); |
13839 | Value *Expected = EmitScalarExpr(E->getArg(1)); |
13840 | Value *Timeout = EmitScalarExpr(E->getArg(2)); |
13841 | Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_atomic_wait_i64); |
13842 | return Builder.CreateCall(Callee, {Addr, Expected, Timeout}); |
13843 | } |
13844 | case WebAssembly::BI__builtin_wasm_atomic_notify: { |
13845 | Value *Addr = EmitScalarExpr(E->getArg(0)); |
13846 | Value *Count = EmitScalarExpr(E->getArg(1)); |
13847 | Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_atomic_notify); |
13848 | return Builder.CreateCall(Callee, {Addr, Count}); |
13849 | } |
13850 | case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i32_f32: |
13851 | case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i32_f64: |
13852 | case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i64_f32: |
13853 | case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i64_f64: |
13854 | case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i32x4_f32x4: |
13855 | case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i64x2_f64x2: { |
13856 | Value *Src = EmitScalarExpr(E->getArg(0)); |
13857 | llvm::Type *ResT = ConvertType(E->getType()); |
13858 | Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_trunc_saturate_signed, |
13859 | {ResT, Src->getType()}); |
13860 | return Builder.CreateCall(Callee, {Src}); |
13861 | } |
13862 | case WebAssembly::BI__builtin_wasm_trunc_saturate_u_i32_f32: |
13863 | case WebAssembly::BI__builtin_wasm_trunc_saturate_u_i32_f64: |
13864 | case WebAssembly::BI__builtin_wasm_trunc_saturate_u_i64_f32: |
13865 | case WebAssembly::BI__builtin_wasm_trunc_saturate_u_i64_f64: |
13866 | case WebAssembly::BI__builtin_wasm_trunc_saturate_u_i32x4_f32x4: |
13867 | case WebAssembly::BI__builtin_wasm_trunc_saturate_u_i64x2_f64x2: { |
13868 | Value *Src = EmitScalarExpr(E->getArg(0)); |
13869 | llvm::Type *ResT = ConvertType(E->getType()); |
13870 | Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_trunc_saturate_unsigned, |
13871 | {ResT, Src->getType()}); |
13872 | return Builder.CreateCall(Callee, {Src}); |
13873 | } |
13874 | case WebAssembly::BI__builtin_wasm_min_f32: |
13875 | case WebAssembly::BI__builtin_wasm_min_f64: |
13876 | case WebAssembly::BI__builtin_wasm_min_f32x4: |
13877 | case WebAssembly::BI__builtin_wasm_min_f64x2: { |
13878 | Value *LHS = EmitScalarExpr(E->getArg(0)); |
13879 | Value *RHS = EmitScalarExpr(E->getArg(1)); |
13880 | Function *Callee = CGM.getIntrinsic(Intrinsic::minimum, |
13881 | ConvertType(E->getType())); |
13882 | return Builder.CreateCall(Callee, {LHS, RHS}); |
13883 | } |
13884 | case WebAssembly::BI__builtin_wasm_max_f32: |
13885 | case WebAssembly::BI__builtin_wasm_max_f64: |
13886 | case WebAssembly::BI__builtin_wasm_max_f32x4: |
13887 | case WebAssembly::BI__builtin_wasm_max_f64x2: { |
13888 | Value *LHS = EmitScalarExpr(E->getArg(0)); |
13889 | Value *RHS = EmitScalarExpr(E->getArg(1)); |
13890 | Function *Callee = CGM.getIntrinsic(Intrinsic::maximum, |
13891 | ConvertType(E->getType())); |
13892 | return Builder.CreateCall(Callee, {LHS, RHS}); |
13893 | } |
13894 | case WebAssembly::BI__builtin_wasm_extract_lane_s_i8x16: |
13895 | case WebAssembly::BI__builtin_wasm_extract_lane_u_i8x16: |
13896 | case WebAssembly::BI__builtin_wasm_extract_lane_s_i16x8: |
13897 | case WebAssembly::BI__builtin_wasm_extract_lane_u_i16x8: |
13898 | case WebAssembly::BI__builtin_wasm_extract_lane_i32x4: |
13899 | case WebAssembly::BI__builtin_wasm_extract_lane_i64x2: |
13900 | case WebAssembly::BI__builtin_wasm_extract_lane_f32x4: |
13901 | case WebAssembly::BI__builtin_wasm_extract_lane_f64x2: { |
13902 | llvm::APSInt LaneConst; |
13903 | if (!E->getArg(1)->isIntegerConstantExpr(LaneConst, getContext())) |
13904 | llvm_unreachable("Constant arg isn't actually constant?")::llvm::llvm_unreachable_internal("Constant arg isn't actually constant?" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 13904); |
13905 | Value *Vec = EmitScalarExpr(E->getArg(0)); |
13906 | Value *Lane = llvm::ConstantInt::get(getLLVMContext(), LaneConst); |
13907 | Value *Extract = Builder.CreateExtractElement(Vec, Lane); |
13908 | switch (BuiltinID) { |
13909 | case WebAssembly::BI__builtin_wasm_extract_lane_s_i8x16: |
13910 | case WebAssembly::BI__builtin_wasm_extract_lane_s_i16x8: |
13911 | return Builder.CreateSExt(Extract, ConvertType(E->getType())); |
13912 | case WebAssembly::BI__builtin_wasm_extract_lane_u_i8x16: |
13913 | case WebAssembly::BI__builtin_wasm_extract_lane_u_i16x8: |
13914 | return Builder.CreateZExt(Extract, ConvertType(E->getType())); |
13915 | case WebAssembly::BI__builtin_wasm_extract_lane_i32x4: |
13916 | case WebAssembly::BI__builtin_wasm_extract_lane_i64x2: |
13917 | case WebAssembly::BI__builtin_wasm_extract_lane_f32x4: |
13918 | case WebAssembly::BI__builtin_wasm_extract_lane_f64x2: |
13919 | return Extract; |
13920 | default: |
13921 | llvm_unreachable("unexpected builtin ID")::llvm::llvm_unreachable_internal("unexpected builtin ID", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 13921); |
13922 | } |
13923 | } |
13924 | case WebAssembly::BI__builtin_wasm_replace_lane_i8x16: |
13925 | case WebAssembly::BI__builtin_wasm_replace_lane_i16x8: |
13926 | case WebAssembly::BI__builtin_wasm_replace_lane_i32x4: |
13927 | case WebAssembly::BI__builtin_wasm_replace_lane_i64x2: |
13928 | case WebAssembly::BI__builtin_wasm_replace_lane_f32x4: |
13929 | case WebAssembly::BI__builtin_wasm_replace_lane_f64x2: { |
13930 | llvm::APSInt LaneConst; |
13931 | if (!E->getArg(1)->isIntegerConstantExpr(LaneConst, getContext())) |
13932 | llvm_unreachable("Constant arg isn't actually constant?")::llvm::llvm_unreachable_internal("Constant arg isn't actually constant?" , "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 13932); |
13933 | Value *Vec = EmitScalarExpr(E->getArg(0)); |
13934 | Value *Lane = llvm::ConstantInt::get(getLLVMContext(), LaneConst); |
13935 | Value *Val = EmitScalarExpr(E->getArg(2)); |
13936 | switch (BuiltinID) { |
13937 | case WebAssembly::BI__builtin_wasm_replace_lane_i8x16: |
13938 | case WebAssembly::BI__builtin_wasm_replace_lane_i16x8: { |
13939 | llvm::Type *ElemType = ConvertType(E->getType())->getVectorElementType(); |
13940 | Value *Trunc = Builder.CreateTrunc(Val, ElemType); |
13941 | return Builder.CreateInsertElement(Vec, Trunc, Lane); |
13942 | } |
13943 | case WebAssembly::BI__builtin_wasm_replace_lane_i32x4: |
13944 | case WebAssembly::BI__builtin_wasm_replace_lane_i64x2: |
13945 | case WebAssembly::BI__builtin_wasm_replace_lane_f32x4: |
13946 | case WebAssembly::BI__builtin_wasm_replace_lane_f64x2: |
13947 | return Builder.CreateInsertElement(Vec, Val, Lane); |
13948 | default: |
13949 | llvm_unreachable("unexpected builtin ID")::llvm::llvm_unreachable_internal("unexpected builtin ID", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 13949); |
13950 | } |
13951 | } |
13952 | case WebAssembly::BI__builtin_wasm_add_saturate_s_i8x16: |
13953 | case WebAssembly::BI__builtin_wasm_add_saturate_u_i8x16: |
13954 | case WebAssembly::BI__builtin_wasm_add_saturate_s_i16x8: |
13955 | case WebAssembly::BI__builtin_wasm_add_saturate_u_i16x8: |
13956 | case WebAssembly::BI__builtin_wasm_sub_saturate_s_i8x16: |
13957 | case WebAssembly::BI__builtin_wasm_sub_saturate_u_i8x16: |
13958 | case WebAssembly::BI__builtin_wasm_sub_saturate_s_i16x8: |
13959 | case WebAssembly::BI__builtin_wasm_sub_saturate_u_i16x8: { |
13960 | unsigned IntNo; |
13961 | switch (BuiltinID) { |
13962 | case WebAssembly::BI__builtin_wasm_add_saturate_s_i8x16: |
13963 | case WebAssembly::BI__builtin_wasm_add_saturate_s_i16x8: |
13964 | IntNo = Intrinsic::sadd_sat; |
13965 | break; |
13966 | case WebAssembly::BI__builtin_wasm_add_saturate_u_i8x16: |
13967 | case WebAssembly::BI__builtin_wasm_add_saturate_u_i16x8: |
13968 | IntNo = Intrinsic::uadd_sat; |
13969 | break; |
13970 | case WebAssembly::BI__builtin_wasm_sub_saturate_s_i8x16: |
13971 | case WebAssembly::BI__builtin_wasm_sub_saturate_s_i16x8: |
13972 | IntNo = Intrinsic::wasm_sub_saturate_signed; |
13973 | break; |
13974 | case WebAssembly::BI__builtin_wasm_sub_saturate_u_i8x16: |
13975 | case WebAssembly::BI__builtin_wasm_sub_saturate_u_i16x8: |
13976 | IntNo = Intrinsic::wasm_sub_saturate_unsigned; |
13977 | break; |
13978 | default: |
13979 | llvm_unreachable("unexpected builtin ID")::llvm::llvm_unreachable_internal("unexpected builtin ID", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 13979); |
13980 | } |
13981 | Value *LHS = EmitScalarExpr(E->getArg(0)); |
13982 | Value *RHS = EmitScalarExpr(E->getArg(1)); |
13983 | Function *Callee = CGM.getIntrinsic(IntNo, ConvertType(E->getType())); |
13984 | return Builder.CreateCall(Callee, {LHS, RHS}); |
13985 | } |
13986 | case WebAssembly::BI__builtin_wasm_bitselect: { |
13987 | Value *V1 = EmitScalarExpr(E->getArg(0)); |
13988 | Value *V2 = EmitScalarExpr(E->getArg(1)); |
13989 | Value *C = EmitScalarExpr(E->getArg(2)); |
13990 | Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_bitselect, |
13991 | ConvertType(E->getType())); |
13992 | return Builder.CreateCall(Callee, {V1, V2, C}); |
13993 | } |
13994 | case WebAssembly::BI__builtin_wasm_any_true_i8x16: |
13995 | case WebAssembly::BI__builtin_wasm_any_true_i16x8: |
13996 | case WebAssembly::BI__builtin_wasm_any_true_i32x4: |
13997 | case WebAssembly::BI__builtin_wasm_any_true_i64x2: |
13998 | case WebAssembly::BI__builtin_wasm_all_true_i8x16: |
13999 | case WebAssembly::BI__builtin_wasm_all_true_i16x8: |
14000 | case WebAssembly::BI__builtin_wasm_all_true_i32x4: |
14001 | case WebAssembly::BI__builtin_wasm_all_true_i64x2: { |
14002 | unsigned IntNo; |
14003 | switch (BuiltinID) { |
14004 | case WebAssembly::BI__builtin_wasm_any_true_i8x16: |
14005 | case WebAssembly::BI__builtin_wasm_any_true_i16x8: |
14006 | case WebAssembly::BI__builtin_wasm_any_true_i32x4: |
14007 | case WebAssembly::BI__builtin_wasm_any_true_i64x2: |
14008 | IntNo = Intrinsic::wasm_anytrue; |
14009 | break; |
14010 | case WebAssembly::BI__builtin_wasm_all_true_i8x16: |
14011 | case WebAssembly::BI__builtin_wasm_all_true_i16x8: |
14012 | case WebAssembly::BI__builtin_wasm_all_true_i32x4: |
14013 | case WebAssembly::BI__builtin_wasm_all_true_i64x2: |
14014 | IntNo = Intrinsic::wasm_alltrue; |
14015 | break; |
14016 | default: |
14017 | llvm_unreachable("unexpected builtin ID")::llvm::llvm_unreachable_internal("unexpected builtin ID", "/build/llvm-toolchain-snapshot-9~svn361194/tools/clang/lib/CodeGen/CGBuiltin.cpp" , 14017); |
14018 | } |
14019 | Value *Vec = EmitScalarExpr(E->getArg(0)); |
14020 | Function *Callee = CGM.getIntrinsic(IntNo, Vec->getType()); |
14021 | return Builder.CreateCall(Callee, {Vec}); |
14022 | } |
14023 | case WebAssembly::BI__builtin_wasm_abs_f32x4: |
14024 | case WebAssembly::BI__builtin_wasm_abs_f64x2: { |
14025 | Value *Vec = EmitScalarExpr(E->getArg(0)); |
14026 | Function *Callee = CGM.getIntrinsic(Intrinsic::fabs, Vec->getType()); |
14027 | return Builder.CreateCall(Callee, {Vec}); |
14028 | } |
14029 | case WebAssembly::BI__builtin_wasm_sqrt_f32x4: |
14030 | case WebAssembly::BI__builtin_wasm_sqrt_f64x2: { |
14031 | Value *Vec = EmitScalarExpr(E->getArg(0)); |
14032 | Function *Callee = CGM.getIntrinsic(Intrinsic::sqrt, Vec->getType()); |
14033 | return Builder.CreateCall(Callee, {Vec}); |
14034 | } |
14035 | |
14036 | default: |
14037 | return nullptr; |
14038 | } |
14039 | } |
14040 | |
14041 | Value *CodeGenFunction::EmitHexagonBuiltinExpr(unsigned BuiltinID, |
14042 | const CallExpr *E) { |
14043 | SmallVector<llvm::Value *, 4> Ops; |
14044 | Intrinsic::ID ID = Intrinsic::not_intrinsic; |
14045 | |
14046 | auto MakeCircLd = [&](unsigned IntID, bool HasImm) { |
14047 | // The base pointer is passed by address, so it needs to be loaded. |
14048 | Address BP = EmitPointerWithAlignment(E->getArg(0)); |
14049 | BP = Address(Builder.CreateBitCast(BP.getPointer(), Int8PtrPtrTy), |
14050 | BP.getAlignment()); |
14051 | llvm::Value *Base = Builder.CreateLoad(BP); |
14052 | // Operands are Base, Increment, Modifier, Start. |
14053 | if (HasImm) |
14054 | Ops = { Base, EmitScalarExpr(E->getArg(1)), EmitScalarExpr(E->getArg(2)), |
14055 | EmitScalarExpr(E->getArg(3)) }; |
14056 | else |
14057 | Ops = { Base, EmitScalarExpr(E->getArg(1)), |
14058 | EmitScalarExpr(E->getArg(2)) }; |
14059 | |
14060 | llvm::Value *Result = Builder.CreateCall(CGM.getIntrinsic(IntID), Ops); |
14061 | llvm::Value *NewBase = Builder.CreateExtractValue(Result, 1); |
14062 | llvm::Value *LV = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), |
14063 | NewBase->getType()->getPointerTo()); |
14064 | Address Dest = EmitPointerWithAlignment(E->getArg(0)); |
14065 | // The intrinsic generates two results. The new value for the base pointer |
14066 | // needs to be stored. |
14067 | Builder.CreateAlignedStore(NewBase, LV, Dest.getAlignment()); |
14068 | return Builder.CreateExtractValue(Result, 0); |
14069 | }; |
14070 | |
14071 | auto MakeCircSt = [&](unsigned IntID, bool HasImm) { |
14072 | // The base pointer is passed by address, so it needs to be loaded. |
14073 | Address BP = EmitPointerWithAlignment(E->getArg(0)); |
14074 | BP = Address(Builder.CreateBitCast(BP.getPointer(), Int8PtrPtrTy), |
14075 | BP.getAlignment()); |
14076 | llvm::Value *Base = Builder.CreateLoad(BP); |
14077 | // Operands are Base, Increment, Modifier, Value, Start. |
14078 | if (HasImm) |
14079 | Ops = { Base, EmitScalarExpr(E->getArg(1)), EmitScalarExpr(E->getArg(2)), |
14080 | EmitScalarExpr(E->getArg(3)), EmitScalarExpr(E->getArg(4)) }; |
14081 | else |
14082 | Ops = { Base, EmitScalarExpr(E->getArg(1)), |
14083 | EmitScalarExpr(E->getArg(2)), EmitScalarExpr(E->getArg(3)) }; |
14084 | |
14085 | llvm::Value *NewBase = Builder.CreateCall(CGM.getIntrinsic(IntID), Ops); |
14086 | llvm::Value *LV = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), |
14087 | NewBase->getType()->getPointerTo()); |
14088 | Address Dest = EmitPointerWithAlignment(E->getArg(0)); |
14089 | // The intrinsic generates one result, which is the new value for the base |
14090 | // pointer. It needs to be stored. |
14091 | return Builder.CreateAlignedStore(NewBase, LV, Dest.getAlignment()); |
14092 | }; |
14093 | |
14094 | // Handle the conversion of bit-reverse load intrinsics to bit code. |
14095 | // The intrinsic call after this function only reads from memory and the |
14096 | // write to memory is dealt by the store instruction. |
14097 | auto MakeBrevLd = [&](unsigned IntID, llvm::Type *DestTy) { |
14098 | // The intrinsic generates one result, which is the new value for the base |
14099 | // pointer. It needs to be returned. The result of the load instruction is |
14100 | // passed to intrinsic by address, so the value needs to be stored. |
14101 | llvm::Value *BaseAddress = |
14102 | Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int8PtrTy); |
14103 | |
14104 | // Expressions like &(*pt++) will be incremented per evaluation. |
14105 | // EmitPointerWithAlignment and EmitScalarExpr evaluates the expression |
14106 | // per call. |
14107 | Address DestAddr = EmitPointerWithAlignment(E->getArg(1)); |
14108 | DestAddr = Address(Builder.CreateBitCast(DestAddr.getPointer(), Int8PtrTy), |
14109 | DestAddr.getAlignment()); |
14110 | llvm::Value *DestAddress = DestAddr.getPointer(); |
14111 | |
14112 | // Operands are Base, Dest, Modifier. |
14113 | // The intrinsic format in LLVM IR is defined as |
14114 | // { ValueType, i8* } (i8*, i32). |
14115 | Ops = {BaseAddress, EmitScalarExpr(E->getArg(2))}; |
14116 | |
14117 | llvm::Value *Result = Builder.CreateCall(CGM.getIntrinsic(IntID), Ops); |
14118 | // The value needs to be stored as the variable is passed by reference. |
14119 | llvm::Value *DestVal = Builder.CreateExtractValue(Result, 0); |
14120 | |
14121 | // The store needs to be truncated to fit the destination type. |
14122 | // While i32 and i64 are natively supported on Hexagon, i8 and i16 needs |
14123 | // to be handled with stores of respective destination type. |
14124 | DestVal = Builder.CreateTrunc(DestVal, DestTy); |
14125 | |
14126 | llvm::Value *DestForStore = |
14127 | Builder.CreateBitCast(DestAddress, DestVal->getType()->getPointerTo()); |
14128 | Builder.CreateAlignedStore(DestVal, DestForStore, DestAddr.getAlignment()); |
14129 | // The updated value of the base pointer is returned. |
14130 | return Builder.CreateExtractValue(Result, 1); |
14131 | }; |
14132 | |
14133 | switch (BuiltinID) { |
14134 | case Hexagon::BI__builtin_HEXAGON_V6_vaddcarry: |
14135 | case Hexagon::BI__builtin_HEXAGON_V6_vaddcarry_128B: { |
14136 | Address Dest = EmitPointerWithAlignment(E->getArg(2)); |
14137 | unsigned Size; |
14138 | if (BuiltinID == Hexagon::BI__builtin_HEXAGON_V6_vaddcarry) { |
14139 | Size = 512; |
14140 | ID = Intrinsic::hexagon_V6_vaddcarry; |
14141 | } else { |
14142 | Size = 1024; |
14143 | ID = Intrinsic::hexagon_V6_vaddcarry_128B; |
14144 | } |
14145 | Dest = Builder.CreateBitCast(Dest, |
14146 | llvm::VectorType::get(Builder.getInt1Ty(), Size)->getPointerTo(0)); |
14147 | LoadInst *QLd = Builder.CreateLoad(Dest); |
14148 | Ops = { EmitScalarExpr(E->getArg(0)), EmitScalarExpr(E->getArg(1)), QLd }; |
14149 | llvm::Value *Result = Builder.CreateCall(CGM.getIntrinsic(ID), Ops); |
14150 | llvm::Value *Vprd = Builder.CreateExtractValue(Result, 1); |
14151 | llvm::Value *Base = Builder.CreateBitCast(EmitScalarExpr(E->getArg(2)), |
14152 | Vprd->getType()->getPointerTo(0)); |
14153 | Builder.CreateAlignedStore(Vprd, Base, Dest.getAlignment()); |
14154 | return Builder.CreateExtractValue(Result, 0); |
14155 | } |
14156 | case Hexagon::BI__builtin_HEXAGON_V6_vsubcarry: |
14157 | case Hexagon::BI__builtin_HEXAGON_V6_vsubcarry_128B: { |
14158 | Address Dest = EmitPointerWithAlignment(E->getArg(2)); |
14159 | unsigned Size; |
14160 | if (BuiltinID == Hexagon::BI__builtin_HEXAGON_V6_vsubcarry) { |
14161 | Size = 512; |
14162 | ID = Intrinsic::hexagon_V6_vsubcarry; |
14163 | } else { |
14164 | Size = 1024; |
14165 | ID = Intrinsic::hexagon_V6_vsubcarry_128B; |
14166 | } |
14167 | Dest = Builder.CreateBitCast(Dest, |
14168 | llvm::VectorType::get(Builder.getInt1Ty(), Size)->getPointerTo(0)); |
14169 | LoadInst *QLd = Builder.CreateLoad(Dest); |
14170 | Ops = { EmitScalarExpr(E->getArg(0)), EmitScalarExpr(E->getArg(1)), QLd }; |
14171 | llvm::Value *Result = Builder.CreateCall(CGM.getIntrinsic(ID), Ops); |
14172 | llvm::Value *Vprd = Builder.CreateExtractValue(Result, 1); |
14173 | llvm::Value *Base = Builder.CreateBitCast(EmitScalarExpr(E->getArg(2)), |
14174 | Vprd->getType()->getPointerTo(0)); |
14175 | Builder.CreateAlignedStore(Vprd, Base, Dest.getAlignment()); |
14176 | return Builder.CreateExtractValue(Result, 0); |
14177 | } |
14178 | case Hexagon::BI__builtin_HEXAGON_L2_loadrub_pci: |
14179 | return MakeCircLd(Intrinsic::hexagon_L2_loadrub_pci, /*HasImm*/true); |
14180 | case Hexagon::BI__builtin_HEXAGON_L2_loadrb_pci: |
14181 | return MakeCircLd(Intrinsic::hexagon_L2_loadrb_pci, /*HasImm*/true); |
14182 | case Hexagon::BI__builtin_HEXAGON_L2_loadruh_pci: |
14183 | return MakeCircLd(Intrinsic::hexagon_L2_loadruh_pci, /*HasImm*/true); |
14184 | case Hexagon::BI__builtin_HEXAGON_L2_loadrh_pci: |
14185 | return MakeCircLd(Intrinsic::hexagon_L2_loadrh_pci, /*HasImm*/true); |
14186 | case Hexagon::BI__builtin_HEXAGON_L2_loadri_pci: |
14187 | return MakeCircLd(Intrinsic::hexagon_L2_loadri_pci, /*HasImm*/true); |
14188 | case Hexagon::BI__builtin_HEXAGON_L2_loadrd_pci: |
14189 | return MakeCircLd(Intrinsic::hexagon_L2_loadrd_pci, /*HasImm*/true); |
14190 | case Hexagon::BI__builtin_HEXAGON_L2_loadrub_pcr: |
14191 | return MakeCircLd(Intrinsic::hexagon_L2_loadrub_pcr, /*HasImm*/false); |
14192 | case Hexagon::BI__builtin_HEXAGON_L2_loadrb_pcr: |
14193 | return MakeCircLd(Intrinsic::hexagon_L2_loadrb_pcr, /*HasImm*/false); |
14194 | case Hexagon::BI__builtin_HEXAGON_L2_loadruh_pcr: |
14195 | return MakeCircLd(Intrinsic::hexagon_L2_loadruh_pcr, /*HasImm*/false); |
14196 | case Hexagon::BI__builtin_HEXAGON_L2_loadrh_pcr: |
14197 | return MakeCircLd(Intrinsic::hexagon_L2_loadrh_pcr, /*HasImm*/false); |
14198 | case Hexagon::BI__builtin_HEXAGON_L2_loadri_pcr: |
14199 | return MakeCircLd(Intrinsic::hexagon_L2_loadri_pcr, /*HasImm*/false); |
14200 | case Hexagon::BI__builtin_HEXAGON_L2_loadrd_pcr: |
14201 | return MakeCircLd(Intrinsic::hexagon_L2_loadrd_pcr, /*HasImm*/false); |
14202 | case Hexagon::BI__builtin_HEXAGON_S2_storerb_pci: |
14203 | return MakeCircSt(Intrinsic::hexagon_S2_storerb_pci, /*HasImm*/true); |
14204 | case Hexagon::BI__builtin_HEXAGON_S2_storerh_pci: |
14205 | return MakeCircSt(Intrinsic::hexagon_S2_storerh_pci, /*HasImm*/true); |
14206 | case Hexagon::BI__builtin_HEXAGON_S2_storerf_pci: |
14207 | return MakeCircSt(Intrinsic::hexagon_S2_storerf_pci, /*HasImm*/true); |
14208 | case Hexagon::BI__builtin_HEXAGON_S2_storeri_pci: |
14209 | return MakeCircSt(Intrinsic::hexagon_S2_storeri_pci, /*HasImm*/true); |
14210 | case Hexagon::BI__builtin_HEXAGON_S2_storerd_pci: |
14211 | return MakeCircSt(Intrinsic::hexagon_S2_storerd_pci, /*HasImm*/true); |
14212 | case Hexagon::BI__builtin_HEXAGON_S2_storerb_pcr: |
14213 | return MakeCircSt(Intrinsic::hexagon_S2_storerb_pcr, /*HasImm*/false); |
14214 | case Hexagon::BI__builtin_HEXAGON_S2_storerh_pcr: |
14215 | return MakeCircSt(Intrinsic::hexagon_S2_storerh_pcr, /*HasImm*/false); |
14216 | case Hexagon::BI__builtin_HEXAGON_S2_storerf_pcr: |
14217 | return MakeCircSt(Intrinsic::hexagon_S2_storerf_pcr, /*HasImm*/false); |
14218 | case Hexagon::BI__builtin_HEXAGON_S2_storeri_pcr: |
14219 | return MakeCircSt(Intrinsic::hexagon_S2_storeri_pcr, /*HasImm*/false); |
14220 | case Hexagon::BI__builtin_HEXAGON_S2_storerd_pcr: |
14221 | return MakeCircSt(Intrinsic::hexagon_S2_storerd_pcr, /*HasImm*/false); |
14222 | case Hexagon::BI__builtin_brev_ldub: |
14223 | return MakeBrevLd(Intrinsic::hexagon_L2_loadrub_pbr, Int8Ty); |
14224 | case Hexagon::BI__builtin_brev_ldb: |
14225 | return MakeBrevLd(Intrinsic::hexagon_L2_loadrb_pbr, Int8Ty); |
14226 | case Hexagon::BI__builtin_brev_lduh: |
14227 | return MakeBrevLd(Intrinsic::hexagon_L2_loadruh_pbr, Int16Ty); |
14228 | case Hexagon::BI__builtin_brev_ldh: |
14229 | return MakeBrevLd(Intrinsic::hexagon_L2_loadrh_pbr, Int16Ty); |
14230 | case Hexagon::BI__builtin_brev_ldw: |
14231 | return MakeBrevLd(Intrinsic::hexagon_L2_loadri_pbr, Int32Ty); |
14232 | case Hexagon::BI__builtin_brev_ldd: |
14233 | return MakeBrevLd(Intrinsic::hexagon_L2_loadrd_pbr, Int64Ty); |
14234 | default: |
14235 | break; |
14236 | } // switch |
14237 | |
14238 | return nullptr; |
14239 | } |