1 | //===- X86.cpp ------------------------------------------------------------===// |
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 | #include "ABIInfoImpl.h" |
10 | #include "TargetInfo.h" |
11 | #include "clang/Basic/DiagnosticFrontend.h" |
12 | #include "llvm/ADT/SmallBitVector.h" |
13 | |
14 | using namespace clang; |
15 | using namespace clang::CodeGen; |
16 | |
17 | namespace { |
18 | |
19 | /// IsX86_MMXType - Return true if this is an MMX type. |
20 | bool IsX86_MMXType(llvm::Type *IRType) { |
21 | // Return true if the type is an MMX type <2 x i32>, <4 x i16>, or <8 x i8>. |
22 | return IRType->isVectorTy() && IRType->getPrimitiveSizeInBits() == 64 && |
23 | cast<llvm::VectorType>(Val: IRType)->getElementType()->isIntegerTy() && |
24 | IRType->getScalarSizeInBits() != 64; |
25 | } |
26 | |
27 | static llvm::Type *X86AdjustInlineAsmType(CodeGen::CodeGenFunction &CGF, |
28 | StringRef Constraint, |
29 | llvm::Type *Ty) { |
30 | if (Constraint == "k" ) { |
31 | llvm::Type *Int1Ty = llvm::Type::getInt1Ty(C&: CGF.getLLVMContext()); |
32 | return llvm::FixedVectorType::get(ElementType: Int1Ty, NumElts: Ty->getScalarSizeInBits()); |
33 | } |
34 | |
35 | // No operation needed |
36 | return Ty; |
37 | } |
38 | |
39 | /// Returns true if this type can be passed in SSE registers with the |
40 | /// X86_VectorCall calling convention. Shared between x86_32 and x86_64. |
41 | static bool isX86VectorTypeForVectorCall(ASTContext &Context, QualType Ty) { |
42 | if (const BuiltinType *BT = Ty->getAs<BuiltinType>()) { |
43 | if (BT->isFloatingPoint() && BT->getKind() != BuiltinType::Half) { |
44 | if (BT->getKind() == BuiltinType::LongDouble) { |
45 | if (&Context.getTargetInfo().getLongDoubleFormat() == |
46 | &llvm::APFloat::x87DoubleExtended()) |
47 | return false; |
48 | } |
49 | return true; |
50 | } |
51 | } else if (const VectorType *VT = Ty->getAs<VectorType>()) { |
52 | // vectorcall can pass XMM, YMM, and ZMM vectors. We don't pass SSE1 MMX |
53 | // registers specially. |
54 | unsigned VecSize = Context.getTypeSize(T: VT); |
55 | if (VecSize == 128 || VecSize == 256 || VecSize == 512) |
56 | return true; |
57 | } |
58 | return false; |
59 | } |
60 | |
61 | /// Returns true if this aggregate is small enough to be passed in SSE registers |
62 | /// in the X86_VectorCall calling convention. Shared between x86_32 and x86_64. |
63 | static bool isX86VectorCallAggregateSmallEnough(uint64_t NumMembers) { |
64 | return NumMembers <= 4; |
65 | } |
66 | |
67 | /// Returns a Homogeneous Vector Aggregate ABIArgInfo, used in X86. |
68 | static ABIArgInfo getDirectX86Hva(llvm::Type* T = nullptr) { |
69 | auto AI = ABIArgInfo::getDirect(T); |
70 | AI.setInReg(true); |
71 | AI.setCanBeFlattened(false); |
72 | return AI; |
73 | } |
74 | |
75 | //===----------------------------------------------------------------------===// |
76 | // X86-32 ABI Implementation |
77 | //===----------------------------------------------------------------------===// |
78 | |
79 | /// Similar to llvm::CCState, but for Clang. |
80 | struct CCState { |
81 | CCState(CGFunctionInfo &FI) |
82 | : IsPreassigned(FI.arg_size()), CC(FI.getCallingConvention()), |
83 | Required(FI.getRequiredArgs()), IsDelegateCall(FI.isDelegateCall()) {} |
84 | |
85 | llvm::SmallBitVector IsPreassigned; |
86 | unsigned CC = CallingConv::CC_C; |
87 | unsigned FreeRegs = 0; |
88 | unsigned FreeSSERegs = 0; |
89 | RequiredArgs Required; |
90 | bool IsDelegateCall = false; |
91 | }; |
92 | |
93 | /// X86_32ABIInfo - The X86-32 ABI information. |
94 | class X86_32ABIInfo : public ABIInfo { |
95 | enum Class { |
96 | Integer, |
97 | Float |
98 | }; |
99 | |
100 | static const unsigned MinABIStackAlignInBytes = 4; |
101 | |
102 | bool IsDarwinVectorABI; |
103 | bool IsRetSmallStructInRegABI; |
104 | bool IsWin32StructABI; |
105 | bool IsSoftFloatABI; |
106 | bool IsMCUABI; |
107 | bool IsLinuxABI; |
108 | unsigned DefaultNumRegisterParameters; |
109 | |
110 | static bool isRegisterSize(unsigned Size) { |
111 | return (Size == 8 || Size == 16 || Size == 32 || Size == 64); |
112 | } |
113 | |
114 | bool isHomogeneousAggregateBaseType(QualType Ty) const override { |
115 | // FIXME: Assumes vectorcall is in use. |
116 | return isX86VectorTypeForVectorCall(Context&: getContext(), Ty); |
117 | } |
118 | |
119 | bool isHomogeneousAggregateSmallEnough(const Type *Ty, |
120 | uint64_t NumMembers) const override { |
121 | // FIXME: Assumes vectorcall is in use. |
122 | return isX86VectorCallAggregateSmallEnough(NumMembers); |
123 | } |
124 | |
125 | bool shouldReturnTypeInRegister(QualType Ty, ASTContext &Context) const; |
126 | |
127 | /// getIndirectResult - Give a source type \arg Ty, return a suitable result |
128 | /// such that the argument will be passed in memory. |
129 | ABIArgInfo getIndirectResult(QualType Ty, bool ByVal, CCState &State) const; |
130 | |
131 | ABIArgInfo getIndirectReturnResult(QualType Ty, CCState &State) const; |
132 | |
133 | /// Return the alignment to use for the given type on the stack. |
134 | unsigned getTypeStackAlignInBytes(QualType Ty, unsigned Align) const; |
135 | |
136 | Class classify(QualType Ty) const; |
137 | ABIArgInfo classifyReturnType(QualType RetTy, CCState &State) const; |
138 | ABIArgInfo classifyArgumentType(QualType RetTy, CCState &State, |
139 | unsigned ArgIndex) const; |
140 | |
141 | /// Updates the number of available free registers, returns |
142 | /// true if any registers were allocated. |
143 | bool updateFreeRegs(QualType Ty, CCState &State) const; |
144 | |
145 | bool shouldAggregateUseDirect(QualType Ty, CCState &State, bool &InReg, |
146 | bool &NeedsPadding) const; |
147 | bool shouldPrimitiveUseInReg(QualType Ty, CCState &State) const; |
148 | |
149 | bool canExpandIndirectArgument(QualType Ty) const; |
150 | |
151 | /// Rewrite the function info so that all memory arguments use |
152 | /// inalloca. |
153 | void rewriteWithInAlloca(CGFunctionInfo &FI) const; |
154 | |
155 | void addFieldToArgStruct(SmallVector<llvm::Type *, 6> &FrameFields, |
156 | CharUnits &StackOffset, ABIArgInfo &Info, |
157 | QualType Type) const; |
158 | void runVectorCallFirstPass(CGFunctionInfo &FI, CCState &State) const; |
159 | |
160 | public: |
161 | |
162 | void computeInfo(CGFunctionInfo &FI) const override; |
163 | RValue EmitVAArg(CodeGenFunction &CGF, Address VAListAddr, QualType Ty, |
164 | AggValueSlot Slot) const override; |
165 | |
166 | X86_32ABIInfo(CodeGen::CodeGenTypes &CGT, bool DarwinVectorABI, |
167 | bool RetSmallStructInRegABI, bool Win32StructABI, |
168 | unsigned NumRegisterParameters, bool SoftFloatABI) |
169 | : ABIInfo(CGT), IsDarwinVectorABI(DarwinVectorABI), |
170 | IsRetSmallStructInRegABI(RetSmallStructInRegABI), |
171 | IsWin32StructABI(Win32StructABI), IsSoftFloatABI(SoftFloatABI), |
172 | IsMCUABI(CGT.getTarget().getTriple().isOSIAMCU()), |
173 | IsLinuxABI(CGT.getTarget().getTriple().isOSLinux() || |
174 | CGT.getTarget().getTriple().isOSCygMing()), |
175 | DefaultNumRegisterParameters(NumRegisterParameters) {} |
176 | }; |
177 | |
178 | class X86_32SwiftABIInfo : public SwiftABIInfo { |
179 | public: |
180 | explicit X86_32SwiftABIInfo(CodeGenTypes &CGT) |
181 | : SwiftABIInfo(CGT, /*SwiftErrorInRegister=*/false) {} |
182 | |
183 | bool shouldPassIndirectly(ArrayRef<llvm::Type *> ComponentTys, |
184 | bool AsReturnValue) const override { |
185 | // LLVM's x86-32 lowering currently only assigns up to three |
186 | // integer registers and three fp registers. Oddly, it'll use up to |
187 | // four vector registers for vectors, but those can overlap with the |
188 | // scalar registers. |
189 | return occupiesMoreThan(scalarTypes: ComponentTys, /*total=*/maxAllRegisters: 3); |
190 | } |
191 | }; |
192 | |
193 | class X86_32TargetCodeGenInfo : public TargetCodeGenInfo { |
194 | public: |
195 | X86_32TargetCodeGenInfo(CodeGen::CodeGenTypes &CGT, bool DarwinVectorABI, |
196 | bool RetSmallStructInRegABI, bool Win32StructABI, |
197 | unsigned NumRegisterParameters, bool SoftFloatABI) |
198 | : TargetCodeGenInfo(std::make_unique<X86_32ABIInfo>( |
199 | args&: CGT, args&: DarwinVectorABI, args&: RetSmallStructInRegABI, args&: Win32StructABI, |
200 | args&: NumRegisterParameters, args&: SoftFloatABI)) { |
201 | SwiftInfo = std::make_unique<X86_32SwiftABIInfo>(args&: CGT); |
202 | } |
203 | |
204 | static bool isStructReturnInRegABI( |
205 | const llvm::Triple &Triple, const CodeGenOptions &Opts); |
206 | |
207 | void setTargetAttributes(const Decl *D, llvm::GlobalValue *GV, |
208 | CodeGen::CodeGenModule &CGM) const override; |
209 | |
210 | int getDwarfEHStackPointer(CodeGen::CodeGenModule &CGM) const override { |
211 | // Darwin uses different dwarf register numbers for EH. |
212 | if (CGM.getTarget().getTriple().isOSDarwin()) return 5; |
213 | return 4; |
214 | } |
215 | |
216 | bool initDwarfEHRegSizeTable(CodeGen::CodeGenFunction &CGF, |
217 | llvm::Value *Address) const override; |
218 | |
219 | llvm::Type* adjustInlineAsmType(CodeGen::CodeGenFunction &CGF, |
220 | StringRef Constraint, |
221 | llvm::Type* Ty) const override { |
222 | return X86AdjustInlineAsmType(CGF, Constraint, Ty); |
223 | } |
224 | |
225 | void addReturnRegisterOutputs(CodeGenFunction &CGF, LValue ReturnValue, |
226 | std::string &Constraints, |
227 | std::vector<llvm::Type *> &ResultRegTypes, |
228 | std::vector<llvm::Type *> &ResultTruncRegTypes, |
229 | std::vector<LValue> &ResultRegDests, |
230 | std::string &AsmString, |
231 | unsigned NumOutputs) const override; |
232 | |
233 | StringRef getARCRetainAutoreleasedReturnValueMarker() const override { |
234 | return "movl\t%ebp, %ebp" |
235 | "\t\t// marker for objc_retainAutoreleaseReturnValue" ; |
236 | } |
237 | }; |
238 | |
239 | } |
240 | |
241 | /// Rewrite input constraint references after adding some output constraints. |
242 | /// In the case where there is one output and one input and we add one output, |
243 | /// we need to replace all operand references greater than or equal to 1: |
244 | /// mov $0, $1 |
245 | /// mov eax, $1 |
246 | /// The result will be: |
247 | /// mov $0, $2 |
248 | /// mov eax, $2 |
249 | static void rewriteInputConstraintReferences(unsigned FirstIn, |
250 | unsigned NumNewOuts, |
251 | std::string &AsmString) { |
252 | std::string Buf; |
253 | llvm::raw_string_ostream OS(Buf); |
254 | size_t Pos = 0; |
255 | while (Pos < AsmString.size()) { |
256 | size_t DollarStart = AsmString.find(c: '$', pos: Pos); |
257 | if (DollarStart == std::string::npos) |
258 | DollarStart = AsmString.size(); |
259 | size_t DollarEnd = AsmString.find_first_not_of(c: '$', pos: DollarStart); |
260 | if (DollarEnd == std::string::npos) |
261 | DollarEnd = AsmString.size(); |
262 | OS << StringRef(&AsmString[Pos], DollarEnd - Pos); |
263 | Pos = DollarEnd; |
264 | size_t NumDollars = DollarEnd - DollarStart; |
265 | if (NumDollars % 2 != 0 && Pos < AsmString.size()) { |
266 | // We have an operand reference. |
267 | size_t DigitStart = Pos; |
268 | if (AsmString[DigitStart] == '{') { |
269 | OS << '{'; |
270 | ++DigitStart; |
271 | } |
272 | size_t DigitEnd = AsmString.find_first_not_of(s: "0123456789" , pos: DigitStart); |
273 | if (DigitEnd == std::string::npos) |
274 | DigitEnd = AsmString.size(); |
275 | StringRef OperandStr(&AsmString[DigitStart], DigitEnd - DigitStart); |
276 | unsigned OperandIndex; |
277 | if (!OperandStr.getAsInteger(Radix: 10, Result&: OperandIndex)) { |
278 | if (OperandIndex >= FirstIn) |
279 | OperandIndex += NumNewOuts; |
280 | OS << OperandIndex; |
281 | } else { |
282 | OS << OperandStr; |
283 | } |
284 | Pos = DigitEnd; |
285 | } |
286 | } |
287 | AsmString = std::move(Buf); |
288 | } |
289 | |
290 | /// Add output constraints for EAX:EDX because they are return registers. |
291 | void X86_32TargetCodeGenInfo::addReturnRegisterOutputs( |
292 | CodeGenFunction &CGF, LValue ReturnSlot, std::string &Constraints, |
293 | std::vector<llvm::Type *> &ResultRegTypes, |
294 | std::vector<llvm::Type *> &ResultTruncRegTypes, |
295 | std::vector<LValue> &ResultRegDests, std::string &AsmString, |
296 | unsigned NumOutputs) const { |
297 | uint64_t RetWidth = CGF.getContext().getTypeSize(T: ReturnSlot.getType()); |
298 | |
299 | // Use the EAX constraint if the width is 32 or smaller and EAX:EDX if it is |
300 | // larger. |
301 | if (!Constraints.empty()) |
302 | Constraints += ','; |
303 | if (RetWidth <= 32) { |
304 | Constraints += "={eax}" ; |
305 | ResultRegTypes.push_back(x: CGF.Int32Ty); |
306 | } else { |
307 | // Use the 'A' constraint for EAX:EDX. |
308 | Constraints += "=A" ; |
309 | ResultRegTypes.push_back(x: CGF.Int64Ty); |
310 | } |
311 | |
312 | // Truncate EAX or EAX:EDX to an integer of the appropriate size. |
313 | llvm::Type *CoerceTy = llvm::IntegerType::get(C&: CGF.getLLVMContext(), NumBits: RetWidth); |
314 | ResultTruncRegTypes.push_back(x: CoerceTy); |
315 | |
316 | // Coerce the integer by bitcasting the return slot pointer. |
317 | ReturnSlot.setAddress(ReturnSlot.getAddress().withElementType(ElemTy: CoerceTy)); |
318 | ResultRegDests.push_back(x: ReturnSlot); |
319 | |
320 | rewriteInputConstraintReferences(FirstIn: NumOutputs, NumNewOuts: 1, AsmString); |
321 | } |
322 | |
323 | /// shouldReturnTypeInRegister - Determine if the given type should be |
324 | /// returned in a register (for the Darwin and MCU ABI). |
325 | bool X86_32ABIInfo::shouldReturnTypeInRegister(QualType Ty, |
326 | ASTContext &Context) const { |
327 | uint64_t Size = Context.getTypeSize(T: Ty); |
328 | |
329 | // For i386, type must be register sized. |
330 | // For the MCU ABI, it only needs to be <= 8-byte |
331 | if ((IsMCUABI && Size > 64) || (!IsMCUABI && !isRegisterSize(Size))) |
332 | return false; |
333 | |
334 | if (Ty->isVectorType()) { |
335 | // 64- and 128- bit vectors inside structures are not returned in |
336 | // registers. |
337 | if (Size == 64 || Size == 128) |
338 | return false; |
339 | |
340 | return true; |
341 | } |
342 | |
343 | // If this is a builtin, pointer, enum, complex type, member pointer, or |
344 | // member function pointer it is ok. |
345 | if (Ty->getAs<BuiltinType>() || Ty->hasPointerRepresentation() || |
346 | Ty->isAnyComplexType() || Ty->isEnumeralType() || |
347 | Ty->isBlockPointerType() || Ty->isMemberPointerType()) |
348 | return true; |
349 | |
350 | // Arrays are treated like records. |
351 | if (const ConstantArrayType *AT = Context.getAsConstantArrayType(T: Ty)) |
352 | return shouldReturnTypeInRegister(Ty: AT->getElementType(), Context); |
353 | |
354 | // Otherwise, it must be a record type. |
355 | const RecordType *RT = Ty->getAs<RecordType>(); |
356 | if (!RT) return false; |
357 | |
358 | // FIXME: Traverse bases here too. |
359 | |
360 | // Structure types are passed in register if all fields would be |
361 | // passed in a register. |
362 | for (const auto *FD : RT->getDecl()->fields()) { |
363 | // Empty fields are ignored. |
364 | if (isEmptyField(Context, FD, AllowArrays: true)) |
365 | continue; |
366 | |
367 | // Check fields recursively. |
368 | if (!shouldReturnTypeInRegister(Ty: FD->getType(), Context)) |
369 | return false; |
370 | } |
371 | return true; |
372 | } |
373 | |
374 | static bool is32Or64BitBasicType(QualType Ty, ASTContext &Context) { |
375 | // Treat complex types as the element type. |
376 | if (const ComplexType *CTy = Ty->getAs<ComplexType>()) |
377 | Ty = CTy->getElementType(); |
378 | |
379 | // Check for a type which we know has a simple scalar argument-passing |
380 | // convention without any padding. (We're specifically looking for 32 |
381 | // and 64-bit integer and integer-equivalents, float, and double.) |
382 | if (!Ty->getAs<BuiltinType>() && !Ty->hasPointerRepresentation() && |
383 | !Ty->isEnumeralType() && !Ty->isBlockPointerType()) |
384 | return false; |
385 | |
386 | uint64_t Size = Context.getTypeSize(T: Ty); |
387 | return Size == 32 || Size == 64; |
388 | } |
389 | |
390 | static bool addFieldSizes(ASTContext &Context, const RecordDecl *RD, |
391 | uint64_t &Size) { |
392 | for (const auto *FD : RD->fields()) { |
393 | // Scalar arguments on the stack get 4 byte alignment on x86. If the |
394 | // argument is smaller than 32-bits, expanding the struct will create |
395 | // alignment padding. |
396 | if (!is32Or64BitBasicType(Ty: FD->getType(), Context)) |
397 | return false; |
398 | |
399 | // FIXME: Reject bit-fields wholesale; there are two problems, we don't know |
400 | // how to expand them yet, and the predicate for telling if a bitfield still |
401 | // counts as "basic" is more complicated than what we were doing previously. |
402 | if (FD->isBitField()) |
403 | return false; |
404 | |
405 | Size += Context.getTypeSize(T: FD->getType()); |
406 | } |
407 | return true; |
408 | } |
409 | |
410 | static bool addBaseAndFieldSizes(ASTContext &Context, const CXXRecordDecl *RD, |
411 | uint64_t &Size) { |
412 | // Don't do this if there are any non-empty bases. |
413 | for (const CXXBaseSpecifier &Base : RD->bases()) { |
414 | if (!addBaseAndFieldSizes(Context, RD: Base.getType()->getAsCXXRecordDecl(), |
415 | Size)) |
416 | return false; |
417 | } |
418 | if (!addFieldSizes(Context, RD, Size)) |
419 | return false; |
420 | return true; |
421 | } |
422 | |
423 | /// Test whether an argument type which is to be passed indirectly (on the |
424 | /// stack) would have the equivalent layout if it was expanded into separate |
425 | /// arguments. If so, we prefer to do the latter to avoid inhibiting |
426 | /// optimizations. |
427 | bool X86_32ABIInfo::canExpandIndirectArgument(QualType Ty) const { |
428 | // We can only expand structure types. |
429 | const RecordType *RT = Ty->getAs<RecordType>(); |
430 | if (!RT) |
431 | return false; |
432 | const RecordDecl *RD = RT->getDecl(); |
433 | uint64_t Size = 0; |
434 | if (const CXXRecordDecl *CXXRD = dyn_cast<CXXRecordDecl>(Val: RD)) { |
435 | if (!IsWin32StructABI) { |
436 | // On non-Windows, we have to conservatively match our old bitcode |
437 | // prototypes in order to be ABI-compatible at the bitcode level. |
438 | if (!CXXRD->isCLike()) |
439 | return false; |
440 | } else { |
441 | // Don't do this for dynamic classes. |
442 | if (CXXRD->isDynamicClass()) |
443 | return false; |
444 | } |
445 | if (!addBaseAndFieldSizes(Context&: getContext(), RD: CXXRD, Size)) |
446 | return false; |
447 | } else { |
448 | if (!addFieldSizes(Context&: getContext(), RD, Size)) |
449 | return false; |
450 | } |
451 | |
452 | // We can do this if there was no alignment padding. |
453 | return Size == getContext().getTypeSize(T: Ty); |
454 | } |
455 | |
456 | ABIArgInfo X86_32ABIInfo::getIndirectReturnResult(QualType RetTy, CCState &State) const { |
457 | // If the return value is indirect, then the hidden argument is consuming one |
458 | // integer register. |
459 | if (State.CC != llvm::CallingConv::X86_FastCall && |
460 | State.CC != llvm::CallingConv::X86_VectorCall && State.FreeRegs) { |
461 | --State.FreeRegs; |
462 | if (!IsMCUABI) |
463 | return getNaturalAlignIndirectInReg(Ty: RetTy); |
464 | } |
465 | return getNaturalAlignIndirect( |
466 | Ty: RetTy, /*AddrSpace=*/getDataLayout().getAllocaAddrSpace(), |
467 | /*ByVal=*/false); |
468 | } |
469 | |
470 | ABIArgInfo X86_32ABIInfo::classifyReturnType(QualType RetTy, |
471 | CCState &State) const { |
472 | if (RetTy->isVoidType()) |
473 | return ABIArgInfo::getIgnore(); |
474 | |
475 | const Type *Base = nullptr; |
476 | uint64_t NumElts = 0; |
477 | if ((State.CC == llvm::CallingConv::X86_VectorCall || |
478 | State.CC == llvm::CallingConv::X86_RegCall) && |
479 | isHomogeneousAggregate(Ty: RetTy, Base, Members&: NumElts)) { |
480 | // The LLVM struct type for such an aggregate should lower properly. |
481 | return ABIArgInfo::getDirect(); |
482 | } |
483 | |
484 | if (const VectorType *VT = RetTy->getAs<VectorType>()) { |
485 | // On Darwin, some vectors are returned in registers. |
486 | if (IsDarwinVectorABI) { |
487 | uint64_t Size = getContext().getTypeSize(T: RetTy); |
488 | |
489 | // 128-bit vectors are a special case; they are returned in |
490 | // registers and we need to make sure to pick a type the LLVM |
491 | // backend will like. |
492 | if (Size == 128) |
493 | return ABIArgInfo::getDirect(T: llvm::FixedVectorType::get( |
494 | ElementType: llvm::Type::getInt64Ty(C&: getVMContext()), NumElts: 2)); |
495 | |
496 | // Always return in register if it fits in a general purpose |
497 | // register, or if it is 64 bits and has a single element. |
498 | if ((Size == 8 || Size == 16 || Size == 32) || |
499 | (Size == 64 && VT->getNumElements() == 1)) |
500 | return ABIArgInfo::getDirect(T: llvm::IntegerType::get(C&: getVMContext(), |
501 | NumBits: Size)); |
502 | |
503 | return getIndirectReturnResult(RetTy, State); |
504 | } |
505 | |
506 | return ABIArgInfo::getDirect(); |
507 | } |
508 | |
509 | if (isAggregateTypeForABI(T: RetTy)) { |
510 | if (const RecordType *RT = RetTy->getAs<RecordType>()) { |
511 | // Structures with flexible arrays are always indirect. |
512 | if (RT->getDecl()->hasFlexibleArrayMember()) |
513 | return getIndirectReturnResult(RetTy, State); |
514 | } |
515 | |
516 | // If specified, structs and unions are always indirect. |
517 | if (!IsRetSmallStructInRegABI && !RetTy->isAnyComplexType()) |
518 | return getIndirectReturnResult(RetTy, State); |
519 | |
520 | // Ignore empty structs/unions. |
521 | if (isEmptyRecord(Context&: getContext(), T: RetTy, AllowArrays: true)) |
522 | return ABIArgInfo::getIgnore(); |
523 | |
524 | // Return complex of _Float16 as <2 x half> so the backend will use xmm0. |
525 | if (const ComplexType *CT = RetTy->getAs<ComplexType>()) { |
526 | QualType ET = getContext().getCanonicalType(T: CT->getElementType()); |
527 | if (ET->isFloat16Type()) |
528 | return ABIArgInfo::getDirect(T: llvm::FixedVectorType::get( |
529 | ElementType: llvm::Type::getHalfTy(C&: getVMContext()), NumElts: 2)); |
530 | } |
531 | |
532 | // Small structures which are register sized are generally returned |
533 | // in a register. |
534 | if (shouldReturnTypeInRegister(Ty: RetTy, Context&: getContext())) { |
535 | uint64_t Size = getContext().getTypeSize(T: RetTy); |
536 | |
537 | // As a special-case, if the struct is a "single-element" struct, and |
538 | // the field is of type "float" or "double", return it in a |
539 | // floating-point register. (MSVC does not apply this special case.) |
540 | // We apply a similar transformation for pointer types to improve the |
541 | // quality of the generated IR. |
542 | if (const Type *SeltTy = isSingleElementStruct(T: RetTy, Context&: getContext())) |
543 | if ((!IsWin32StructABI && SeltTy->isRealFloatingType()) |
544 | || SeltTy->hasPointerRepresentation()) |
545 | return ABIArgInfo::getDirect(T: CGT.ConvertType(T: QualType(SeltTy, 0))); |
546 | |
547 | // FIXME: We should be able to narrow this integer in cases with dead |
548 | // padding. |
549 | return ABIArgInfo::getDirect(T: llvm::IntegerType::get(C&: getVMContext(),NumBits: Size)); |
550 | } |
551 | |
552 | return getIndirectReturnResult(RetTy, State); |
553 | } |
554 | |
555 | // Treat an enum type as its underlying type. |
556 | if (const EnumType *EnumTy = RetTy->getAs<EnumType>()) |
557 | RetTy = EnumTy->getDecl()->getIntegerType(); |
558 | |
559 | if (const auto *EIT = RetTy->getAs<BitIntType>()) |
560 | if (EIT->getNumBits() > 64) |
561 | return getIndirectReturnResult(RetTy, State); |
562 | |
563 | return (isPromotableIntegerTypeForABI(Ty: RetTy) ? ABIArgInfo::getExtend(Ty: RetTy) |
564 | : ABIArgInfo::getDirect()); |
565 | } |
566 | |
567 | unsigned X86_32ABIInfo::getTypeStackAlignInBytes(QualType Ty, |
568 | unsigned Align) const { |
569 | // Otherwise, if the alignment is less than or equal to the minimum ABI |
570 | // alignment, just use the default; the backend will handle this. |
571 | if (Align <= MinABIStackAlignInBytes) |
572 | return 0; // Use default alignment. |
573 | |
574 | if (IsLinuxABI) { |
575 | // Exclude other System V OS (e.g Darwin, PS4 and FreeBSD) since we don't |
576 | // want to spend any effort dealing with the ramifications of ABI breaks. |
577 | // |
578 | // If the vector type is __m128/__m256/__m512, return the default alignment. |
579 | if (Ty->isVectorType() && (Align == 16 || Align == 32 || Align == 64)) |
580 | return Align; |
581 | } |
582 | // On non-Darwin, the stack type alignment is always 4. |
583 | if (!IsDarwinVectorABI) { |
584 | // Set explicit alignment, since we may need to realign the top. |
585 | return MinABIStackAlignInBytes; |
586 | } |
587 | |
588 | // Otherwise, if the type contains an SSE vector type, the alignment is 16. |
589 | if (Align >= 16 && (isSIMDVectorType(Context&: getContext(), Ty) || |
590 | isRecordWithSIMDVectorType(Context&: getContext(), Ty))) |
591 | return 16; |
592 | |
593 | return MinABIStackAlignInBytes; |
594 | } |
595 | |
596 | ABIArgInfo X86_32ABIInfo::getIndirectResult(QualType Ty, bool ByVal, |
597 | CCState &State) const { |
598 | if (!ByVal) { |
599 | if (State.FreeRegs) { |
600 | --State.FreeRegs; // Non-byval indirects just use one pointer. |
601 | if (!IsMCUABI) |
602 | return getNaturalAlignIndirectInReg(Ty); |
603 | } |
604 | return getNaturalAlignIndirect(Ty, AddrSpace: getDataLayout().getAllocaAddrSpace(), |
605 | ByVal: false); |
606 | } |
607 | |
608 | // Compute the byval alignment. |
609 | unsigned TypeAlign = getContext().getTypeAlign(T: Ty) / 8; |
610 | unsigned StackAlign = getTypeStackAlignInBytes(Ty, Align: TypeAlign); |
611 | if (StackAlign == 0) |
612 | return ABIArgInfo::getIndirect( |
613 | Alignment: CharUnits::fromQuantity(Quantity: 4), |
614 | /*AddrSpace=*/getDataLayout().getAllocaAddrSpace(), |
615 | /*ByVal=*/true); |
616 | |
617 | // If the stack alignment is less than the type alignment, realign the |
618 | // argument. |
619 | bool Realign = TypeAlign > StackAlign; |
620 | return ABIArgInfo::getIndirect( |
621 | Alignment: CharUnits::fromQuantity(Quantity: StackAlign), |
622 | /*AddrSpace=*/getDataLayout().getAllocaAddrSpace(), /*ByVal=*/true, |
623 | Realign); |
624 | } |
625 | |
626 | X86_32ABIInfo::Class X86_32ABIInfo::classify(QualType Ty) const { |
627 | const Type *T = isSingleElementStruct(T: Ty, Context&: getContext()); |
628 | if (!T) |
629 | T = Ty.getTypePtr(); |
630 | |
631 | if (const BuiltinType *BT = T->getAs<BuiltinType>()) { |
632 | BuiltinType::Kind K = BT->getKind(); |
633 | if (K == BuiltinType::Float || K == BuiltinType::Double) |
634 | return Float; |
635 | } |
636 | return Integer; |
637 | } |
638 | |
639 | bool X86_32ABIInfo::updateFreeRegs(QualType Ty, CCState &State) const { |
640 | if (!IsSoftFloatABI) { |
641 | Class C = classify(Ty); |
642 | if (C == Float) |
643 | return false; |
644 | } |
645 | |
646 | unsigned Size = getContext().getTypeSize(T: Ty); |
647 | unsigned SizeInRegs = (Size + 31) / 32; |
648 | |
649 | if (SizeInRegs == 0) |
650 | return false; |
651 | |
652 | if (!IsMCUABI) { |
653 | if (SizeInRegs > State.FreeRegs) { |
654 | State.FreeRegs = 0; |
655 | return false; |
656 | } |
657 | } else { |
658 | // The MCU psABI allows passing parameters in-reg even if there are |
659 | // earlier parameters that are passed on the stack. Also, |
660 | // it does not allow passing >8-byte structs in-register, |
661 | // even if there are 3 free registers available. |
662 | if (SizeInRegs > State.FreeRegs || SizeInRegs > 2) |
663 | return false; |
664 | } |
665 | |
666 | State.FreeRegs -= SizeInRegs; |
667 | return true; |
668 | } |
669 | |
670 | bool X86_32ABIInfo::shouldAggregateUseDirect(QualType Ty, CCState &State, |
671 | bool &InReg, |
672 | bool &NeedsPadding) const { |
673 | // On Windows, aggregates other than HFAs are never passed in registers, and |
674 | // they do not consume register slots. Homogenous floating-point aggregates |
675 | // (HFAs) have already been dealt with at this point. |
676 | if (IsWin32StructABI && isAggregateTypeForABI(T: Ty)) |
677 | return false; |
678 | |
679 | NeedsPadding = false; |
680 | InReg = !IsMCUABI; |
681 | |
682 | if (!updateFreeRegs(Ty, State)) |
683 | return false; |
684 | |
685 | if (IsMCUABI) |
686 | return true; |
687 | |
688 | if (State.CC == llvm::CallingConv::X86_FastCall || |
689 | State.CC == llvm::CallingConv::X86_VectorCall || |
690 | State.CC == llvm::CallingConv::X86_RegCall) { |
691 | if (getContext().getTypeSize(T: Ty) <= 32 && State.FreeRegs) |
692 | NeedsPadding = true; |
693 | |
694 | return false; |
695 | } |
696 | |
697 | return true; |
698 | } |
699 | |
700 | bool X86_32ABIInfo::shouldPrimitiveUseInReg(QualType Ty, CCState &State) const { |
701 | bool IsPtrOrInt = (getContext().getTypeSize(T: Ty) <= 32) && |
702 | (Ty->isIntegralOrEnumerationType() || Ty->isPointerType() || |
703 | Ty->isReferenceType()); |
704 | |
705 | if (!IsPtrOrInt && (State.CC == llvm::CallingConv::X86_FastCall || |
706 | State.CC == llvm::CallingConv::X86_VectorCall)) |
707 | return false; |
708 | |
709 | if (!updateFreeRegs(Ty, State)) |
710 | return false; |
711 | |
712 | if (!IsPtrOrInt && State.CC == llvm::CallingConv::X86_RegCall) |
713 | return false; |
714 | |
715 | // Return true to apply inreg to all legal parameters except for MCU targets. |
716 | return !IsMCUABI; |
717 | } |
718 | |
719 | void X86_32ABIInfo::runVectorCallFirstPass(CGFunctionInfo &FI, CCState &State) const { |
720 | // Vectorcall x86 works subtly different than in x64, so the format is |
721 | // a bit different than the x64 version. First, all vector types (not HVAs) |
722 | // are assigned, with the first 6 ending up in the [XYZ]MM0-5 registers. |
723 | // This differs from the x64 implementation, where the first 6 by INDEX get |
724 | // registers. |
725 | // In the second pass over the arguments, HVAs are passed in the remaining |
726 | // vector registers if possible, or indirectly by address. The address will be |
727 | // passed in ECX/EDX if available. Any other arguments are passed according to |
728 | // the usual fastcall rules. |
729 | MutableArrayRef<CGFunctionInfoArgInfo> Args = FI.arguments(); |
730 | for (int I = 0, E = Args.size(); I < E; ++I) { |
731 | const Type *Base = nullptr; |
732 | uint64_t NumElts = 0; |
733 | const QualType &Ty = Args[I].type; |
734 | if ((Ty->isVectorType() || Ty->isBuiltinType()) && |
735 | isHomogeneousAggregate(Ty, Base, Members&: NumElts)) { |
736 | if (State.FreeSSERegs >= NumElts) { |
737 | State.FreeSSERegs -= NumElts; |
738 | Args[I].info = ABIArgInfo::getDirectInReg(); |
739 | State.IsPreassigned.set(I); |
740 | } |
741 | } |
742 | } |
743 | } |
744 | |
745 | ABIArgInfo X86_32ABIInfo::classifyArgumentType(QualType Ty, CCState &State, |
746 | unsigned ArgIndex) const { |
747 | // FIXME: Set alignment on indirect arguments. |
748 | bool IsFastCall = State.CC == llvm::CallingConv::X86_FastCall; |
749 | bool IsRegCall = State.CC == llvm::CallingConv::X86_RegCall; |
750 | bool IsVectorCall = State.CC == llvm::CallingConv::X86_VectorCall; |
751 | |
752 | Ty = useFirstFieldIfTransparentUnion(Ty); |
753 | TypeInfo TI = getContext().getTypeInfo(T: Ty); |
754 | |
755 | // Check with the C++ ABI first. |
756 | const RecordType *RT = Ty->getAs<RecordType>(); |
757 | if (RT) { |
758 | CGCXXABI::RecordArgABI RAA = getRecordArgABI(RT, CXXABI&: getCXXABI()); |
759 | if (RAA == CGCXXABI::RAA_Indirect) { |
760 | return getIndirectResult(Ty, ByVal: false, State); |
761 | } else if (State.IsDelegateCall) { |
762 | // Avoid having different alignments on delegate call args by always |
763 | // setting the alignment to 4, which is what we do for inallocas. |
764 | ABIArgInfo Res = getIndirectResult(Ty, ByVal: false, State); |
765 | Res.setIndirectAlign(CharUnits::fromQuantity(Quantity: 4)); |
766 | return Res; |
767 | } else if (RAA == CGCXXABI::RAA_DirectInMemory) { |
768 | // The field index doesn't matter, we'll fix it up later. |
769 | return ABIArgInfo::getInAlloca(/*FieldIndex=*/0); |
770 | } |
771 | } |
772 | |
773 | // Regcall uses the concept of a homogenous vector aggregate, similar |
774 | // to other targets. |
775 | const Type *Base = nullptr; |
776 | uint64_t NumElts = 0; |
777 | if ((IsRegCall || IsVectorCall) && |
778 | isHomogeneousAggregate(Ty, Base, Members&: NumElts)) { |
779 | if (State.FreeSSERegs >= NumElts) { |
780 | State.FreeSSERegs -= NumElts; |
781 | |
782 | // Vectorcall passes HVAs directly and does not flatten them, but regcall |
783 | // does. |
784 | if (IsVectorCall) |
785 | return getDirectX86Hva(); |
786 | |
787 | if (Ty->isBuiltinType() || Ty->isVectorType()) |
788 | return ABIArgInfo::getDirect(); |
789 | return ABIArgInfo::getExpand(); |
790 | } |
791 | if (IsVectorCall && Ty->isBuiltinType()) |
792 | return ABIArgInfo::getDirect(); |
793 | return getIndirectResult(Ty, /*ByVal=*/false, State); |
794 | } |
795 | |
796 | if (isAggregateTypeForABI(T: Ty)) { |
797 | // Structures with flexible arrays are always indirect. |
798 | // FIXME: This should not be byval! |
799 | if (RT && RT->getDecl()->hasFlexibleArrayMember()) |
800 | return getIndirectResult(Ty, ByVal: true, State); |
801 | |
802 | // Ignore empty structs/unions on non-Windows. |
803 | if (!IsWin32StructABI && isEmptyRecord(Context&: getContext(), T: Ty, AllowArrays: true)) |
804 | return ABIArgInfo::getIgnore(); |
805 | |
806 | // Ignore 0 sized structs. |
807 | if (TI.Width == 0) |
808 | return ABIArgInfo::getIgnore(); |
809 | |
810 | llvm::LLVMContext &LLVMContext = getVMContext(); |
811 | llvm::IntegerType *Int32 = llvm::Type::getInt32Ty(C&: LLVMContext); |
812 | bool NeedsPadding = false; |
813 | bool InReg; |
814 | if (shouldAggregateUseDirect(Ty, State, InReg, NeedsPadding)) { |
815 | unsigned SizeInRegs = (TI.Width + 31) / 32; |
816 | SmallVector<llvm::Type*, 3> Elements(SizeInRegs, Int32); |
817 | llvm::Type *Result = llvm::StructType::get(Context&: LLVMContext, Elements); |
818 | if (InReg) |
819 | return ABIArgInfo::getDirectInReg(T: Result); |
820 | else |
821 | return ABIArgInfo::getDirect(T: Result); |
822 | } |
823 | llvm::IntegerType *PaddingType = NeedsPadding ? Int32 : nullptr; |
824 | |
825 | // Pass over-aligned aggregates to non-variadic functions on Windows |
826 | // indirectly. This behavior was added in MSVC 2015. Use the required |
827 | // alignment from the record layout, since that may be less than the |
828 | // regular type alignment, and types with required alignment of less than 4 |
829 | // bytes are not passed indirectly. |
830 | if (IsWin32StructABI && State.Required.isRequiredArg(argIdx: ArgIndex)) { |
831 | unsigned AlignInBits = 0; |
832 | if (RT) { |
833 | const ASTRecordLayout &Layout = |
834 | getContext().getASTRecordLayout(D: RT->getDecl()); |
835 | AlignInBits = getContext().toBits(CharSize: Layout.getRequiredAlignment()); |
836 | } else if (TI.isAlignRequired()) { |
837 | AlignInBits = TI.Align; |
838 | } |
839 | if (AlignInBits > 32) |
840 | return getIndirectResult(Ty, /*ByVal=*/false, State); |
841 | } |
842 | |
843 | // Expand small (<= 128-bit) record types when we know that the stack layout |
844 | // of those arguments will match the struct. This is important because the |
845 | // LLVM backend isn't smart enough to remove byval, which inhibits many |
846 | // optimizations. |
847 | // Don't do this for the MCU if there are still free integer registers |
848 | // (see X86_64 ABI for full explanation). |
849 | if (TI.Width <= 4 * 32 && (!IsMCUABI || State.FreeRegs == 0) && |
850 | canExpandIndirectArgument(Ty)) |
851 | return ABIArgInfo::getExpandWithPadding( |
852 | PaddingInReg: IsFastCall || IsVectorCall || IsRegCall, Padding: PaddingType); |
853 | |
854 | return getIndirectResult(Ty, ByVal: true, State); |
855 | } |
856 | |
857 | if (const VectorType *VT = Ty->getAs<VectorType>()) { |
858 | // On Windows, vectors are passed directly if registers are available, or |
859 | // indirectly if not. This avoids the need to align argument memory. Pass |
860 | // user-defined vector types larger than 512 bits indirectly for simplicity. |
861 | if (IsWin32StructABI) { |
862 | if (TI.Width <= 512 && State.FreeSSERegs > 0) { |
863 | --State.FreeSSERegs; |
864 | return ABIArgInfo::getDirectInReg(); |
865 | } |
866 | return getIndirectResult(Ty, /*ByVal=*/false, State); |
867 | } |
868 | |
869 | // On Darwin, some vectors are passed in memory, we handle this by passing |
870 | // it as an i8/i16/i32/i64. |
871 | if (IsDarwinVectorABI) { |
872 | if ((TI.Width == 8 || TI.Width == 16 || TI.Width == 32) || |
873 | (TI.Width == 64 && VT->getNumElements() == 1)) |
874 | return ABIArgInfo::getDirect( |
875 | T: llvm::IntegerType::get(C&: getVMContext(), NumBits: TI.Width)); |
876 | } |
877 | |
878 | if (IsX86_MMXType(IRType: CGT.ConvertType(T: Ty))) |
879 | return ABIArgInfo::getDirect(T: llvm::IntegerType::get(C&: getVMContext(), NumBits: 64)); |
880 | |
881 | return ABIArgInfo::getDirect(); |
882 | } |
883 | |
884 | |
885 | if (const EnumType *EnumTy = Ty->getAs<EnumType>()) |
886 | Ty = EnumTy->getDecl()->getIntegerType(); |
887 | |
888 | bool InReg = shouldPrimitiveUseInReg(Ty, State); |
889 | |
890 | if (isPromotableIntegerTypeForABI(Ty)) { |
891 | if (InReg) |
892 | return ABIArgInfo::getExtendInReg(Ty, T: CGT.ConvertType(T: Ty)); |
893 | return ABIArgInfo::getExtend(Ty, T: CGT.ConvertType(T: Ty)); |
894 | } |
895 | |
896 | if (const auto *EIT = Ty->getAs<BitIntType>()) { |
897 | if (EIT->getNumBits() <= 64) { |
898 | if (InReg) |
899 | return ABIArgInfo::getDirectInReg(); |
900 | return ABIArgInfo::getDirect(); |
901 | } |
902 | return getIndirectResult(Ty, /*ByVal=*/false, State); |
903 | } |
904 | |
905 | if (InReg) |
906 | return ABIArgInfo::getDirectInReg(); |
907 | return ABIArgInfo::getDirect(); |
908 | } |
909 | |
910 | void X86_32ABIInfo::computeInfo(CGFunctionInfo &FI) const { |
911 | CCState State(FI); |
912 | if (IsMCUABI) |
913 | State.FreeRegs = 3; |
914 | else if (State.CC == llvm::CallingConv::X86_FastCall) { |
915 | State.FreeRegs = 2; |
916 | State.FreeSSERegs = 3; |
917 | } else if (State.CC == llvm::CallingConv::X86_VectorCall) { |
918 | State.FreeRegs = 2; |
919 | State.FreeSSERegs = 6; |
920 | } else if (FI.getHasRegParm()) |
921 | State.FreeRegs = FI.getRegParm(); |
922 | else if (State.CC == llvm::CallingConv::X86_RegCall) { |
923 | State.FreeRegs = 5; |
924 | State.FreeSSERegs = 8; |
925 | } else if (IsWin32StructABI) { |
926 | // Since MSVC 2015, the first three SSE vectors have been passed in |
927 | // registers. The rest are passed indirectly. |
928 | State.FreeRegs = DefaultNumRegisterParameters; |
929 | State.FreeSSERegs = 3; |
930 | } else |
931 | State.FreeRegs = DefaultNumRegisterParameters; |
932 | |
933 | if (!::classifyReturnType(CXXABI: getCXXABI(), FI, Info: *this)) { |
934 | FI.getReturnInfo() = classifyReturnType(RetTy: FI.getReturnType(), State); |
935 | } else if (FI.getReturnInfo().isIndirect()) { |
936 | // The C++ ABI is not aware of register usage, so we have to check if the |
937 | // return value was sret and put it in a register ourselves if appropriate. |
938 | if (State.FreeRegs) { |
939 | --State.FreeRegs; // The sret parameter consumes a register. |
940 | if (!IsMCUABI) |
941 | FI.getReturnInfo().setInReg(true); |
942 | } |
943 | } |
944 | |
945 | // The chain argument effectively gives us another free register. |
946 | if (FI.isChainCall()) |
947 | ++State.FreeRegs; |
948 | |
949 | // For vectorcall, do a first pass over the arguments, assigning FP and vector |
950 | // arguments to XMM registers as available. |
951 | if (State.CC == llvm::CallingConv::X86_VectorCall) |
952 | runVectorCallFirstPass(FI, State); |
953 | |
954 | bool UsedInAlloca = false; |
955 | MutableArrayRef<CGFunctionInfoArgInfo> Args = FI.arguments(); |
956 | for (unsigned I = 0, E = Args.size(); I < E; ++I) { |
957 | // Skip arguments that have already been assigned. |
958 | if (State.IsPreassigned.test(Idx: I)) |
959 | continue; |
960 | |
961 | Args[I].info = |
962 | classifyArgumentType(Ty: Args[I].type, State, ArgIndex: I); |
963 | UsedInAlloca |= (Args[I].info.getKind() == ABIArgInfo::InAlloca); |
964 | } |
965 | |
966 | // If we needed to use inalloca for any argument, do a second pass and rewrite |
967 | // all the memory arguments to use inalloca. |
968 | if (UsedInAlloca) |
969 | rewriteWithInAlloca(FI); |
970 | } |
971 | |
972 | void |
973 | X86_32ABIInfo::addFieldToArgStruct(SmallVector<llvm::Type *, 6> &FrameFields, |
974 | CharUnits &StackOffset, ABIArgInfo &Info, |
975 | QualType Type) const { |
976 | // Arguments are always 4-byte-aligned. |
977 | CharUnits WordSize = CharUnits::fromQuantity(Quantity: 4); |
978 | assert(StackOffset.isMultipleOf(WordSize) && "unaligned inalloca struct" ); |
979 | |
980 | // sret pointers and indirect things will require an extra pointer |
981 | // indirection, unless they are byval. Most things are byval, and will not |
982 | // require this indirection. |
983 | bool IsIndirect = false; |
984 | if (Info.isIndirect() && !Info.getIndirectByVal()) |
985 | IsIndirect = true; |
986 | Info = ABIArgInfo::getInAlloca(FieldIndex: FrameFields.size(), Indirect: IsIndirect); |
987 | llvm::Type *LLTy = CGT.ConvertTypeForMem(T: Type); |
988 | if (IsIndirect) |
989 | LLTy = llvm::PointerType::getUnqual(C&: getVMContext()); |
990 | FrameFields.push_back(Elt: LLTy); |
991 | StackOffset += IsIndirect ? WordSize : getContext().getTypeSizeInChars(T: Type); |
992 | |
993 | // Insert padding bytes to respect alignment. |
994 | CharUnits FieldEnd = StackOffset; |
995 | StackOffset = FieldEnd.alignTo(Align: WordSize); |
996 | if (StackOffset != FieldEnd) { |
997 | CharUnits NumBytes = StackOffset - FieldEnd; |
998 | llvm::Type *Ty = llvm::Type::getInt8Ty(C&: getVMContext()); |
999 | Ty = llvm::ArrayType::get(ElementType: Ty, NumElements: NumBytes.getQuantity()); |
1000 | FrameFields.push_back(Elt: Ty); |
1001 | } |
1002 | } |
1003 | |
1004 | static bool isArgInAlloca(const ABIArgInfo &Info) { |
1005 | // Leave ignored and inreg arguments alone. |
1006 | switch (Info.getKind()) { |
1007 | case ABIArgInfo::InAlloca: |
1008 | return true; |
1009 | case ABIArgInfo::Ignore: |
1010 | case ABIArgInfo::IndirectAliased: |
1011 | return false; |
1012 | case ABIArgInfo::Indirect: |
1013 | case ABIArgInfo::Direct: |
1014 | case ABIArgInfo::Extend: |
1015 | return !Info.getInReg(); |
1016 | case ABIArgInfo::Expand: |
1017 | case ABIArgInfo::CoerceAndExpand: |
1018 | // These are aggregate types which are never passed in registers when |
1019 | // inalloca is involved. |
1020 | return true; |
1021 | } |
1022 | llvm_unreachable("invalid enum" ); |
1023 | } |
1024 | |
1025 | void X86_32ABIInfo::rewriteWithInAlloca(CGFunctionInfo &FI) const { |
1026 | assert(IsWin32StructABI && "inalloca only supported on win32" ); |
1027 | |
1028 | // Build a packed struct type for all of the arguments in memory. |
1029 | SmallVector<llvm::Type *, 6> FrameFields; |
1030 | |
1031 | // The stack alignment is always 4. |
1032 | CharUnits StackAlign = CharUnits::fromQuantity(Quantity: 4); |
1033 | |
1034 | CharUnits StackOffset; |
1035 | CGFunctionInfo::arg_iterator I = FI.arg_begin(), E = FI.arg_end(); |
1036 | |
1037 | // Put 'this' into the struct before 'sret', if necessary. |
1038 | bool IsThisCall = |
1039 | FI.getCallingConvention() == llvm::CallingConv::X86_ThisCall; |
1040 | ABIArgInfo &Ret = FI.getReturnInfo(); |
1041 | if (Ret.isIndirect() && Ret.isSRetAfterThis() && !IsThisCall && |
1042 | isArgInAlloca(Info: I->info)) { |
1043 | addFieldToArgStruct(FrameFields, StackOffset, Info&: I->info, Type: I->type); |
1044 | ++I; |
1045 | } |
1046 | |
1047 | // Put the sret parameter into the inalloca struct if it's in memory. |
1048 | if (Ret.isIndirect() && !Ret.getInReg()) { |
1049 | addFieldToArgStruct(FrameFields, StackOffset, Info&: Ret, Type: FI.getReturnType()); |
1050 | // On Windows, the hidden sret parameter is always returned in eax. |
1051 | Ret.setInAllocaSRet(IsWin32StructABI); |
1052 | } |
1053 | |
1054 | // Skip the 'this' parameter in ecx. |
1055 | if (IsThisCall) |
1056 | ++I; |
1057 | |
1058 | // Put arguments passed in memory into the struct. |
1059 | for (; I != E; ++I) { |
1060 | if (isArgInAlloca(Info: I->info)) |
1061 | addFieldToArgStruct(FrameFields, StackOffset, Info&: I->info, Type: I->type); |
1062 | } |
1063 | |
1064 | FI.setArgStruct(Ty: llvm::StructType::get(Context&: getVMContext(), Elements: FrameFields, |
1065 | /*isPacked=*/true), |
1066 | Align: StackAlign); |
1067 | } |
1068 | |
1069 | RValue X86_32ABIInfo::EmitVAArg(CodeGenFunction &CGF, Address VAListAddr, |
1070 | QualType Ty, AggValueSlot Slot) const { |
1071 | |
1072 | auto TypeInfo = getContext().getTypeInfoInChars(T: Ty); |
1073 | |
1074 | CCState State(*const_cast<CGFunctionInfo *>(CGF.CurFnInfo)); |
1075 | ABIArgInfo AI = classifyArgumentType(Ty, State, /*ArgIndex*/ 0); |
1076 | // Empty records are ignored for parameter passing purposes. |
1077 | if (AI.isIgnore()) |
1078 | return Slot.asRValue(); |
1079 | |
1080 | // x86-32 changes the alignment of certain arguments on the stack. |
1081 | // |
1082 | // Just messing with TypeInfo like this works because we never pass |
1083 | // anything indirectly. |
1084 | TypeInfo.Align = CharUnits::fromQuantity( |
1085 | Quantity: getTypeStackAlignInBytes(Ty, Align: TypeInfo.Align.getQuantity())); |
1086 | |
1087 | return emitVoidPtrVAArg(CGF, VAListAddr, ValueTy: Ty, /*Indirect*/ IsIndirect: false, ValueInfo: TypeInfo, |
1088 | SlotSizeAndAlign: CharUnits::fromQuantity(Quantity: 4), |
1089 | /*AllowHigherAlign*/ true, Slot); |
1090 | } |
1091 | |
1092 | bool X86_32TargetCodeGenInfo::isStructReturnInRegABI( |
1093 | const llvm::Triple &Triple, const CodeGenOptions &Opts) { |
1094 | assert(Triple.getArch() == llvm::Triple::x86); |
1095 | |
1096 | switch (Opts.getStructReturnConvention()) { |
1097 | case CodeGenOptions::SRCK_Default: |
1098 | break; |
1099 | case CodeGenOptions::SRCK_OnStack: // -fpcc-struct-return |
1100 | return false; |
1101 | case CodeGenOptions::SRCK_InRegs: // -freg-struct-return |
1102 | return true; |
1103 | } |
1104 | |
1105 | if (Triple.isOSDarwin() || Triple.isOSIAMCU()) |
1106 | return true; |
1107 | |
1108 | switch (Triple.getOS()) { |
1109 | case llvm::Triple::DragonFly: |
1110 | case llvm::Triple::FreeBSD: |
1111 | case llvm::Triple::OpenBSD: |
1112 | case llvm::Triple::Win32: |
1113 | return true; |
1114 | default: |
1115 | return false; |
1116 | } |
1117 | } |
1118 | |
1119 | static void addX86InterruptAttrs(const FunctionDecl *FD, llvm::GlobalValue *GV, |
1120 | CodeGen::CodeGenModule &CGM) { |
1121 | if (!FD->hasAttr<AnyX86InterruptAttr>()) |
1122 | return; |
1123 | |
1124 | llvm::Function *Fn = cast<llvm::Function>(Val: GV); |
1125 | Fn->setCallingConv(llvm::CallingConv::X86_INTR); |
1126 | if (FD->getNumParams() == 0) |
1127 | return; |
1128 | |
1129 | auto PtrTy = cast<PointerType>(Val: FD->getParamDecl(i: 0)->getType()); |
1130 | llvm::Type *ByValTy = CGM.getTypes().ConvertType(T: PtrTy->getPointeeType()); |
1131 | llvm::Attribute NewAttr = llvm::Attribute::getWithByValType( |
1132 | Context&: Fn->getContext(), Ty: ByValTy); |
1133 | Fn->addParamAttr(ArgNo: 0, Attr: NewAttr); |
1134 | } |
1135 | |
1136 | void X86_32TargetCodeGenInfo::setTargetAttributes( |
1137 | const Decl *D, llvm::GlobalValue *GV, CodeGen::CodeGenModule &CGM) const { |
1138 | if (GV->isDeclaration()) |
1139 | return; |
1140 | if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(Val: D)) { |
1141 | if (FD->hasAttr<X86ForceAlignArgPointerAttr>()) { |
1142 | llvm::Function *Fn = cast<llvm::Function>(Val: GV); |
1143 | Fn->addFnAttr(Kind: "stackrealign" ); |
1144 | } |
1145 | |
1146 | addX86InterruptAttrs(FD, GV, CGM); |
1147 | } |
1148 | } |
1149 | |
1150 | bool X86_32TargetCodeGenInfo::initDwarfEHRegSizeTable( |
1151 | CodeGen::CodeGenFunction &CGF, |
1152 | llvm::Value *Address) const { |
1153 | CodeGen::CGBuilderTy &Builder = CGF.Builder; |
1154 | |
1155 | llvm::Value *Four8 = llvm::ConstantInt::get(Ty: CGF.Int8Ty, V: 4); |
1156 | |
1157 | // 0-7 are the eight integer registers; the order is different |
1158 | // on Darwin (for EH), but the range is the same. |
1159 | // 8 is %eip. |
1160 | AssignToArrayRange(Builder, Array: Address, Value: Four8, FirstIndex: 0, LastIndex: 8); |
1161 | |
1162 | if (CGF.CGM.getTarget().getTriple().isOSDarwin()) { |
1163 | // 12-16 are st(0..4). Not sure why we stop at 4. |
1164 | // These have size 16, which is sizeof(long double) on |
1165 | // platforms with 8-byte alignment for that type. |
1166 | llvm::Value *Sixteen8 = llvm::ConstantInt::get(Ty: CGF.Int8Ty, V: 16); |
1167 | AssignToArrayRange(Builder, Array: Address, Value: Sixteen8, FirstIndex: 12, LastIndex: 16); |
1168 | |
1169 | } else { |
1170 | // 9 is %eflags, which doesn't get a size on Darwin for some |
1171 | // reason. |
1172 | Builder.CreateAlignedStore( |
1173 | Val: Four8, Addr: Builder.CreateConstInBoundsGEP1_32(Ty: CGF.Int8Ty, Ptr: Address, Idx0: 9), |
1174 | Align: CharUnits::One()); |
1175 | |
1176 | // 11-16 are st(0..5). Not sure why we stop at 5. |
1177 | // These have size 12, which is sizeof(long double) on |
1178 | // platforms with 4-byte alignment for that type. |
1179 | llvm::Value *Twelve8 = llvm::ConstantInt::get(Ty: CGF.Int8Ty, V: 12); |
1180 | AssignToArrayRange(Builder, Array: Address, Value: Twelve8, FirstIndex: 11, LastIndex: 16); |
1181 | } |
1182 | |
1183 | return false; |
1184 | } |
1185 | |
1186 | //===----------------------------------------------------------------------===// |
1187 | // X86-64 ABI Implementation |
1188 | //===----------------------------------------------------------------------===// |
1189 | |
1190 | |
1191 | namespace { |
1192 | |
1193 | /// \p returns the size in bits of the largest (native) vector for \p AVXLevel. |
1194 | static unsigned getNativeVectorSizeForAVXABI(X86AVXABILevel AVXLevel) { |
1195 | switch (AVXLevel) { |
1196 | case X86AVXABILevel::AVX512: |
1197 | return 512; |
1198 | case X86AVXABILevel::AVX: |
1199 | return 256; |
1200 | case X86AVXABILevel::None: |
1201 | return 128; |
1202 | } |
1203 | llvm_unreachable("Unknown AVXLevel" ); |
1204 | } |
1205 | |
1206 | /// X86_64ABIInfo - The X86_64 ABI information. |
1207 | class X86_64ABIInfo : public ABIInfo { |
1208 | enum Class { |
1209 | Integer = 0, |
1210 | SSE, |
1211 | SSEUp, |
1212 | X87, |
1213 | X87Up, |
1214 | ComplexX87, |
1215 | NoClass, |
1216 | Memory |
1217 | }; |
1218 | |
1219 | /// merge - Implement the X86_64 ABI merging algorithm. |
1220 | /// |
1221 | /// Merge an accumulating classification \arg Accum with a field |
1222 | /// classification \arg Field. |
1223 | /// |
1224 | /// \param Accum - The accumulating classification. This should |
1225 | /// always be either NoClass or the result of a previous merge |
1226 | /// call. In addition, this should never be Memory (the caller |
1227 | /// should just return Memory for the aggregate). |
1228 | static Class merge(Class Accum, Class Field); |
1229 | |
1230 | /// postMerge - Implement the X86_64 ABI post merging algorithm. |
1231 | /// |
1232 | /// Post merger cleanup, reduces a malformed Hi and Lo pair to |
1233 | /// final MEMORY or SSE classes when necessary. |
1234 | /// |
1235 | /// \param AggregateSize - The size of the current aggregate in |
1236 | /// the classification process. |
1237 | /// |
1238 | /// \param Lo - The classification for the parts of the type |
1239 | /// residing in the low word of the containing object. |
1240 | /// |
1241 | /// \param Hi - The classification for the parts of the type |
1242 | /// residing in the higher words of the containing object. |
1243 | /// |
1244 | void postMerge(unsigned AggregateSize, Class &Lo, Class &Hi) const; |
1245 | |
1246 | /// classify - Determine the x86_64 register classes in which the |
1247 | /// given type T should be passed. |
1248 | /// |
1249 | /// \param Lo - The classification for the parts of the type |
1250 | /// residing in the low word of the containing object. |
1251 | /// |
1252 | /// \param Hi - The classification for the parts of the type |
1253 | /// residing in the high word of the containing object. |
1254 | /// |
1255 | /// \param OffsetBase - The bit offset of this type in the |
1256 | /// containing object. Some parameters are classified different |
1257 | /// depending on whether they straddle an eightbyte boundary. |
1258 | /// |
1259 | /// \param isNamedArg - Whether the argument in question is a "named" |
1260 | /// argument, as used in AMD64-ABI 3.5.7. |
1261 | /// |
1262 | /// \param IsRegCall - Whether the calling conversion is regcall. |
1263 | /// |
1264 | /// If a word is unused its result will be NoClass; if a type should |
1265 | /// be passed in Memory then at least the classification of \arg Lo |
1266 | /// will be Memory. |
1267 | /// |
1268 | /// The \arg Lo class will be NoClass iff the argument is ignored. |
1269 | /// |
1270 | /// If the \arg Lo class is ComplexX87, then the \arg Hi class will |
1271 | /// also be ComplexX87. |
1272 | void classify(QualType T, uint64_t OffsetBase, Class &Lo, Class &Hi, |
1273 | bool isNamedArg, bool IsRegCall = false) const; |
1274 | |
1275 | llvm::Type *GetByteVectorType(QualType Ty) const; |
1276 | llvm::Type *GetSSETypeAtOffset(llvm::Type *IRType, |
1277 | unsigned IROffset, QualType SourceTy, |
1278 | unsigned SourceOffset) const; |
1279 | llvm::Type *GetINTEGERTypeAtOffset(llvm::Type *IRType, |
1280 | unsigned IROffset, QualType SourceTy, |
1281 | unsigned SourceOffset) const; |
1282 | |
1283 | /// getIndirectResult - Give a source type \arg Ty, return a suitable result |
1284 | /// such that the argument will be returned in memory. |
1285 | ABIArgInfo getIndirectReturnResult(QualType Ty) const; |
1286 | |
1287 | /// getIndirectResult - Give a source type \arg Ty, return a suitable result |
1288 | /// such that the argument will be passed in memory. |
1289 | /// |
1290 | /// \param freeIntRegs - The number of free integer registers remaining |
1291 | /// available. |
1292 | ABIArgInfo getIndirectResult(QualType Ty, unsigned freeIntRegs) const; |
1293 | |
1294 | ABIArgInfo classifyReturnType(QualType RetTy) const; |
1295 | |
1296 | ABIArgInfo classifyArgumentType(QualType Ty, unsigned freeIntRegs, |
1297 | unsigned &neededInt, unsigned &neededSSE, |
1298 | bool isNamedArg, |
1299 | bool IsRegCall = false) const; |
1300 | |
1301 | ABIArgInfo classifyRegCallStructType(QualType Ty, unsigned &NeededInt, |
1302 | unsigned &NeededSSE, |
1303 | unsigned &MaxVectorWidth) const; |
1304 | |
1305 | ABIArgInfo classifyRegCallStructTypeImpl(QualType Ty, unsigned &NeededInt, |
1306 | unsigned &NeededSSE, |
1307 | unsigned &MaxVectorWidth) const; |
1308 | |
1309 | bool IsIllegalVectorType(QualType Ty) const; |
1310 | |
1311 | /// The 0.98 ABI revision clarified a lot of ambiguities, |
1312 | /// unfortunately in ways that were not always consistent with |
1313 | /// certain previous compilers. In particular, platforms which |
1314 | /// required strict binary compatibility with older versions of GCC |
1315 | /// may need to exempt themselves. |
1316 | bool honorsRevision0_98() const { |
1317 | return !getTarget().getTriple().isOSDarwin(); |
1318 | } |
1319 | |
1320 | /// GCC classifies <1 x long long> as SSE but some platform ABIs choose to |
1321 | /// classify it as INTEGER (for compatibility with older clang compilers). |
1322 | bool classifyIntegerMMXAsSSE() const { |
1323 | // Clang <= 3.8 did not do this. |
1324 | if (getContext().getLangOpts().getClangABICompat() <= |
1325 | LangOptions::ClangABI::Ver3_8) |
1326 | return false; |
1327 | |
1328 | const llvm::Triple &Triple = getTarget().getTriple(); |
1329 | if (Triple.isOSDarwin() || Triple.isPS() || Triple.isOSFreeBSD()) |
1330 | return false; |
1331 | return true; |
1332 | } |
1333 | |
1334 | // GCC classifies vectors of __int128 as memory. |
1335 | bool passInt128VectorsInMem() const { |
1336 | // Clang <= 9.0 did not do this. |
1337 | if (getContext().getLangOpts().getClangABICompat() <= |
1338 | LangOptions::ClangABI::Ver9) |
1339 | return false; |
1340 | |
1341 | const llvm::Triple &T = getTarget().getTriple(); |
1342 | return T.isOSLinux() || T.isOSNetBSD(); |
1343 | } |
1344 | |
1345 | bool returnCXXRecordGreaterThan128InMem() const { |
1346 | // Clang <= 20.0 did not do this. |
1347 | if (getContext().getLangOpts().getClangABICompat() <= |
1348 | LangOptions::ClangABI::Ver20) |
1349 | return false; |
1350 | |
1351 | return true; |
1352 | } |
1353 | |
1354 | X86AVXABILevel AVXLevel; |
1355 | // Some ABIs (e.g. X32 ABI and Native Client OS) use 32 bit pointers on |
1356 | // 64-bit hardware. |
1357 | bool Has64BitPointers; |
1358 | |
1359 | public: |
1360 | X86_64ABIInfo(CodeGen::CodeGenTypes &CGT, X86AVXABILevel AVXLevel) |
1361 | : ABIInfo(CGT), AVXLevel(AVXLevel), |
1362 | Has64BitPointers(CGT.getDataLayout().getPointerSize(AS: 0) == 8) {} |
1363 | |
1364 | bool isPassedUsingAVXType(QualType type) const { |
1365 | unsigned neededInt, neededSSE; |
1366 | // The freeIntRegs argument doesn't matter here. |
1367 | ABIArgInfo info = classifyArgumentType(Ty: type, freeIntRegs: 0, neededInt, neededSSE, |
1368 | /*isNamedArg*/true); |
1369 | if (info.isDirect()) { |
1370 | llvm::Type *ty = info.getCoerceToType(); |
1371 | if (llvm::VectorType *vectorTy = dyn_cast_or_null<llvm::VectorType>(Val: ty)) |
1372 | return vectorTy->getPrimitiveSizeInBits().getFixedValue() > 128; |
1373 | } |
1374 | return false; |
1375 | } |
1376 | |
1377 | void computeInfo(CGFunctionInfo &FI) const override; |
1378 | |
1379 | RValue EmitVAArg(CodeGenFunction &CGF, Address VAListAddr, QualType Ty, |
1380 | AggValueSlot Slot) const override; |
1381 | RValue EmitMSVAArg(CodeGenFunction &CGF, Address VAListAddr, QualType Ty, |
1382 | AggValueSlot Slot) const override; |
1383 | |
1384 | bool has64BitPointers() const { |
1385 | return Has64BitPointers; |
1386 | } |
1387 | }; |
1388 | |
1389 | /// WinX86_64ABIInfo - The Windows X86_64 ABI information. |
1390 | class WinX86_64ABIInfo : public ABIInfo { |
1391 | public: |
1392 | WinX86_64ABIInfo(CodeGen::CodeGenTypes &CGT, X86AVXABILevel AVXLevel) |
1393 | : ABIInfo(CGT), AVXLevel(AVXLevel), |
1394 | IsMingw64(getTarget().getTriple().isWindowsGNUEnvironment()) {} |
1395 | |
1396 | void computeInfo(CGFunctionInfo &FI) const override; |
1397 | |
1398 | RValue EmitVAArg(CodeGenFunction &CGF, Address VAListAddr, QualType Ty, |
1399 | AggValueSlot Slot) const override; |
1400 | |
1401 | bool isHomogeneousAggregateBaseType(QualType Ty) const override { |
1402 | // FIXME: Assumes vectorcall is in use. |
1403 | return isX86VectorTypeForVectorCall(Context&: getContext(), Ty); |
1404 | } |
1405 | |
1406 | bool isHomogeneousAggregateSmallEnough(const Type *Ty, |
1407 | uint64_t NumMembers) const override { |
1408 | // FIXME: Assumes vectorcall is in use. |
1409 | return isX86VectorCallAggregateSmallEnough(NumMembers); |
1410 | } |
1411 | |
1412 | private: |
1413 | ABIArgInfo classify(QualType Ty, unsigned &FreeSSERegs, bool IsReturnType, |
1414 | bool IsVectorCall, bool IsRegCall) const; |
1415 | ABIArgInfo reclassifyHvaArgForVectorCall(QualType Ty, unsigned &FreeSSERegs, |
1416 | const ABIArgInfo ¤t) const; |
1417 | |
1418 | X86AVXABILevel AVXLevel; |
1419 | |
1420 | bool IsMingw64; |
1421 | }; |
1422 | |
1423 | class X86_64TargetCodeGenInfo : public TargetCodeGenInfo { |
1424 | public: |
1425 | X86_64TargetCodeGenInfo(CodeGen::CodeGenTypes &CGT, X86AVXABILevel AVXLevel) |
1426 | : TargetCodeGenInfo(std::make_unique<X86_64ABIInfo>(args&: CGT, args&: AVXLevel)) { |
1427 | SwiftInfo = |
1428 | std::make_unique<SwiftABIInfo>(args&: CGT, /*SwiftErrorInRegister=*/args: true); |
1429 | } |
1430 | |
1431 | /// Disable tail call on x86-64. The epilogue code before the tail jump blocks |
1432 | /// autoreleaseRV/retainRV and autoreleaseRV/unsafeClaimRV optimizations. |
1433 | bool markARCOptimizedReturnCallsAsNoTail() const override { return true; } |
1434 | |
1435 | int getDwarfEHStackPointer(CodeGen::CodeGenModule &CGM) const override { |
1436 | return 7; |
1437 | } |
1438 | |
1439 | bool initDwarfEHRegSizeTable(CodeGen::CodeGenFunction &CGF, |
1440 | llvm::Value *Address) const override { |
1441 | llvm::Value *Eight8 = llvm::ConstantInt::get(Ty: CGF.Int8Ty, V: 8); |
1442 | |
1443 | // 0-15 are the 16 integer registers. |
1444 | // 16 is %rip. |
1445 | AssignToArrayRange(Builder&: CGF.Builder, Array: Address, Value: Eight8, FirstIndex: 0, LastIndex: 16); |
1446 | return false; |
1447 | } |
1448 | |
1449 | llvm::Type* adjustInlineAsmType(CodeGen::CodeGenFunction &CGF, |
1450 | StringRef Constraint, |
1451 | llvm::Type* Ty) const override { |
1452 | return X86AdjustInlineAsmType(CGF, Constraint, Ty); |
1453 | } |
1454 | |
1455 | bool isNoProtoCallVariadic(const CallArgList &args, |
1456 | const FunctionNoProtoType *fnType) const override { |
1457 | // The default CC on x86-64 sets %al to the number of SSA |
1458 | // registers used, and GCC sets this when calling an unprototyped |
1459 | // function, so we override the default behavior. However, don't do |
1460 | // that when AVX types are involved: the ABI explicitly states it is |
1461 | // undefined, and it doesn't work in practice because of how the ABI |
1462 | // defines varargs anyway. |
1463 | if (fnType->getCallConv() == CC_C) { |
1464 | bool HasAVXType = false; |
1465 | for (const CallArg &arg : args) { |
1466 | if (getABIInfo<X86_64ABIInfo>().isPassedUsingAVXType(type: arg.Ty)) { |
1467 | HasAVXType = true; |
1468 | break; |
1469 | } |
1470 | } |
1471 | |
1472 | if (!HasAVXType) |
1473 | return true; |
1474 | } |
1475 | |
1476 | return TargetCodeGenInfo::isNoProtoCallVariadic(args, fnType); |
1477 | } |
1478 | |
1479 | void setTargetAttributes(const Decl *D, llvm::GlobalValue *GV, |
1480 | CodeGen::CodeGenModule &CGM) const override { |
1481 | if (GV->isDeclaration()) |
1482 | return; |
1483 | if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(Val: D)) { |
1484 | if (FD->hasAttr<X86ForceAlignArgPointerAttr>()) { |
1485 | llvm::Function *Fn = cast<llvm::Function>(Val: GV); |
1486 | Fn->addFnAttr(Kind: "stackrealign" ); |
1487 | } |
1488 | |
1489 | addX86InterruptAttrs(FD, GV, CGM); |
1490 | } |
1491 | } |
1492 | |
1493 | void checkFunctionCallABI(CodeGenModule &CGM, SourceLocation CallLoc, |
1494 | const FunctionDecl *Caller, |
1495 | const FunctionDecl *Callee, const CallArgList &Args, |
1496 | QualType ReturnType) const override; |
1497 | }; |
1498 | } // namespace |
1499 | |
1500 | static void initFeatureMaps(const ASTContext &Ctx, |
1501 | llvm::StringMap<bool> &CallerMap, |
1502 | const FunctionDecl *Caller, |
1503 | llvm::StringMap<bool> &CalleeMap, |
1504 | const FunctionDecl *Callee) { |
1505 | if (CalleeMap.empty() && CallerMap.empty()) { |
1506 | // The caller is potentially nullptr in the case where the call isn't in a |
1507 | // function. In this case, the getFunctionFeatureMap ensures we just get |
1508 | // the TU level setting (since it cannot be modified by 'target'.. |
1509 | Ctx.getFunctionFeatureMap(FeatureMap&: CallerMap, Caller); |
1510 | Ctx.getFunctionFeatureMap(FeatureMap&: CalleeMap, Callee); |
1511 | } |
1512 | } |
1513 | |
1514 | static bool checkAVXParamFeature(DiagnosticsEngine &Diag, |
1515 | SourceLocation CallLoc, |
1516 | const llvm::StringMap<bool> &CallerMap, |
1517 | const llvm::StringMap<bool> &CalleeMap, |
1518 | QualType Ty, StringRef Feature, |
1519 | bool IsArgument) { |
1520 | bool CallerHasFeat = CallerMap.lookup(Key: Feature); |
1521 | bool CalleeHasFeat = CalleeMap.lookup(Key: Feature); |
1522 | if (!CallerHasFeat && !CalleeHasFeat) |
1523 | return Diag.Report(Loc: CallLoc, DiagID: diag::warn_avx_calling_convention) |
1524 | << IsArgument << Ty << Feature; |
1525 | |
1526 | // Mixing calling conventions here is very clearly an error. |
1527 | if (!CallerHasFeat || !CalleeHasFeat) |
1528 | return Diag.Report(Loc: CallLoc, DiagID: diag::err_avx_calling_convention) |
1529 | << IsArgument << Ty << Feature; |
1530 | |
1531 | // Else, both caller and callee have the required feature, so there is no need |
1532 | // to diagnose. |
1533 | return false; |
1534 | } |
1535 | |
1536 | static bool checkAVX512ParamFeature(DiagnosticsEngine &Diag, |
1537 | SourceLocation CallLoc, |
1538 | const llvm::StringMap<bool> &CallerMap, |
1539 | const llvm::StringMap<bool> &CalleeMap, |
1540 | QualType Ty, bool IsArgument) { |
1541 | bool Caller256 = CallerMap.lookup(Key: "avx512f" ) && !CallerMap.lookup(Key: "evex512" ); |
1542 | bool Callee256 = CalleeMap.lookup(Key: "avx512f" ) && !CalleeMap.lookup(Key: "evex512" ); |
1543 | |
1544 | // Forbid 512-bit or larger vector pass or return when we disabled ZMM |
1545 | // instructions. |
1546 | if (Caller256 || Callee256) |
1547 | return Diag.Report(Loc: CallLoc, DiagID: diag::err_avx_calling_convention) |
1548 | << IsArgument << Ty << "evex512" ; |
1549 | |
1550 | return checkAVXParamFeature(Diag, CallLoc, CallerMap, CalleeMap, Ty, |
1551 | Feature: "avx512f" , IsArgument); |
1552 | } |
1553 | |
1554 | static bool checkAVXParam(DiagnosticsEngine &Diag, ASTContext &Ctx, |
1555 | SourceLocation CallLoc, |
1556 | const llvm::StringMap<bool> &CallerMap, |
1557 | const llvm::StringMap<bool> &CalleeMap, QualType Ty, |
1558 | bool IsArgument) { |
1559 | uint64_t Size = Ctx.getTypeSize(T: Ty); |
1560 | if (Size > 256) |
1561 | return checkAVX512ParamFeature(Diag, CallLoc, CallerMap, CalleeMap, Ty, |
1562 | IsArgument); |
1563 | |
1564 | if (Size > 128) |
1565 | return checkAVXParamFeature(Diag, CallLoc, CallerMap, CalleeMap, Ty, Feature: "avx" , |
1566 | IsArgument); |
1567 | |
1568 | return false; |
1569 | } |
1570 | |
1571 | void X86_64TargetCodeGenInfo::checkFunctionCallABI(CodeGenModule &CGM, |
1572 | SourceLocation CallLoc, |
1573 | const FunctionDecl *Caller, |
1574 | const FunctionDecl *Callee, |
1575 | const CallArgList &Args, |
1576 | QualType ReturnType) const { |
1577 | if (!Callee) |
1578 | return; |
1579 | |
1580 | llvm::StringMap<bool> CallerMap; |
1581 | llvm::StringMap<bool> CalleeMap; |
1582 | unsigned ArgIndex = 0; |
1583 | |
1584 | // We need to loop through the actual call arguments rather than the |
1585 | // function's parameters, in case this variadic. |
1586 | for (const CallArg &Arg : Args) { |
1587 | // The "avx" feature changes how vectors >128 in size are passed. "avx512f" |
1588 | // additionally changes how vectors >256 in size are passed. Like GCC, we |
1589 | // warn when a function is called with an argument where this will change. |
1590 | // Unlike GCC, we also error when it is an obvious ABI mismatch, that is, |
1591 | // the caller and callee features are mismatched. |
1592 | // Unfortunately, we cannot do this diagnostic in SEMA, since the callee can |
1593 | // change its ABI with attribute-target after this call. |
1594 | if (Arg.getType()->isVectorType() && |
1595 | CGM.getContext().getTypeSize(T: Arg.getType()) > 128) { |
1596 | initFeatureMaps(Ctx: CGM.getContext(), CallerMap, Caller, CalleeMap, Callee); |
1597 | QualType Ty = Arg.getType(); |
1598 | // The CallArg seems to have desugared the type already, so for clearer |
1599 | // diagnostics, replace it with the type in the FunctionDecl if possible. |
1600 | if (ArgIndex < Callee->getNumParams()) |
1601 | Ty = Callee->getParamDecl(i: ArgIndex)->getType(); |
1602 | |
1603 | if (checkAVXParam(Diag&: CGM.getDiags(), Ctx&: CGM.getContext(), CallLoc, CallerMap, |
1604 | CalleeMap, Ty, /*IsArgument*/ true)) |
1605 | return; |
1606 | } |
1607 | ++ArgIndex; |
1608 | } |
1609 | |
1610 | // Check return always, as we don't have a good way of knowing in codegen |
1611 | // whether this value is used, tail-called, etc. |
1612 | if (Callee->getReturnType()->isVectorType() && |
1613 | CGM.getContext().getTypeSize(T: Callee->getReturnType()) > 128) { |
1614 | initFeatureMaps(Ctx: CGM.getContext(), CallerMap, Caller, CalleeMap, Callee); |
1615 | checkAVXParam(Diag&: CGM.getDiags(), Ctx&: CGM.getContext(), CallLoc, CallerMap, |
1616 | CalleeMap, Ty: Callee->getReturnType(), |
1617 | /*IsArgument*/ false); |
1618 | } |
1619 | } |
1620 | |
1621 | std::string TargetCodeGenInfo::qualifyWindowsLibrary(StringRef Lib) { |
1622 | // If the argument does not end in .lib, automatically add the suffix. |
1623 | // If the argument contains a space, enclose it in quotes. |
1624 | // This matches the behavior of MSVC. |
1625 | bool Quote = Lib.contains(C: ' '); |
1626 | std::string ArgStr = Quote ? "\"" : "" ; |
1627 | ArgStr += Lib; |
1628 | if (!Lib.ends_with_insensitive(Suffix: ".lib" ) && !Lib.ends_with_insensitive(Suffix: ".a" )) |
1629 | ArgStr += ".lib" ; |
1630 | ArgStr += Quote ? "\"" : "" ; |
1631 | return ArgStr; |
1632 | } |
1633 | |
1634 | namespace { |
1635 | class WinX86_32TargetCodeGenInfo : public X86_32TargetCodeGenInfo { |
1636 | public: |
1637 | WinX86_32TargetCodeGenInfo(CodeGen::CodeGenTypes &CGT, |
1638 | bool DarwinVectorABI, bool RetSmallStructInRegABI, bool Win32StructABI, |
1639 | unsigned NumRegisterParameters) |
1640 | : X86_32TargetCodeGenInfo(CGT, DarwinVectorABI, RetSmallStructInRegABI, |
1641 | Win32StructABI, NumRegisterParameters, false) {} |
1642 | |
1643 | void setTargetAttributes(const Decl *D, llvm::GlobalValue *GV, |
1644 | CodeGen::CodeGenModule &CGM) const override; |
1645 | |
1646 | void getDependentLibraryOption(llvm::StringRef Lib, |
1647 | llvm::SmallString<24> &Opt) const override { |
1648 | Opt = "/DEFAULTLIB:" ; |
1649 | Opt += qualifyWindowsLibrary(Lib); |
1650 | } |
1651 | |
1652 | void getDetectMismatchOption(llvm::StringRef Name, |
1653 | llvm::StringRef Value, |
1654 | llvm::SmallString<32> &Opt) const override { |
1655 | Opt = "/FAILIFMISMATCH:\"" + Name.str() + "=" + Value.str() + "\"" ; |
1656 | } |
1657 | }; |
1658 | } // namespace |
1659 | |
1660 | void WinX86_32TargetCodeGenInfo::setTargetAttributes( |
1661 | const Decl *D, llvm::GlobalValue *GV, CodeGen::CodeGenModule &CGM) const { |
1662 | X86_32TargetCodeGenInfo::setTargetAttributes(D, GV, CGM); |
1663 | if (GV->isDeclaration()) |
1664 | return; |
1665 | addStackProbeTargetAttributes(D, GV, CGM); |
1666 | } |
1667 | |
1668 | namespace { |
1669 | class WinX86_64TargetCodeGenInfo : public TargetCodeGenInfo { |
1670 | public: |
1671 | WinX86_64TargetCodeGenInfo(CodeGen::CodeGenTypes &CGT, |
1672 | X86AVXABILevel AVXLevel) |
1673 | : TargetCodeGenInfo(std::make_unique<WinX86_64ABIInfo>(args&: CGT, args&: AVXLevel)) { |
1674 | SwiftInfo = |
1675 | std::make_unique<SwiftABIInfo>(args&: CGT, /*SwiftErrorInRegister=*/args: true); |
1676 | } |
1677 | |
1678 | void setTargetAttributes(const Decl *D, llvm::GlobalValue *GV, |
1679 | CodeGen::CodeGenModule &CGM) const override; |
1680 | |
1681 | int getDwarfEHStackPointer(CodeGen::CodeGenModule &CGM) const override { |
1682 | return 7; |
1683 | } |
1684 | |
1685 | bool initDwarfEHRegSizeTable(CodeGen::CodeGenFunction &CGF, |
1686 | llvm::Value *Address) const override { |
1687 | llvm::Value *Eight8 = llvm::ConstantInt::get(Ty: CGF.Int8Ty, V: 8); |
1688 | |
1689 | // 0-15 are the 16 integer registers. |
1690 | // 16 is %rip. |
1691 | AssignToArrayRange(Builder&: CGF.Builder, Array: Address, Value: Eight8, FirstIndex: 0, LastIndex: 16); |
1692 | return false; |
1693 | } |
1694 | |
1695 | void getDependentLibraryOption(llvm::StringRef Lib, |
1696 | llvm::SmallString<24> &Opt) const override { |
1697 | Opt = "/DEFAULTLIB:" ; |
1698 | Opt += qualifyWindowsLibrary(Lib); |
1699 | } |
1700 | |
1701 | void getDetectMismatchOption(llvm::StringRef Name, |
1702 | llvm::StringRef Value, |
1703 | llvm::SmallString<32> &Opt) const override { |
1704 | Opt = "/FAILIFMISMATCH:\"" + Name.str() + "=" + Value.str() + "\"" ; |
1705 | } |
1706 | }; |
1707 | } // namespace |
1708 | |
1709 | void WinX86_64TargetCodeGenInfo::setTargetAttributes( |
1710 | const Decl *D, llvm::GlobalValue *GV, CodeGen::CodeGenModule &CGM) const { |
1711 | TargetCodeGenInfo::setTargetAttributes(D, GV, M&: CGM); |
1712 | if (GV->isDeclaration()) |
1713 | return; |
1714 | if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(Val: D)) { |
1715 | if (FD->hasAttr<X86ForceAlignArgPointerAttr>()) { |
1716 | llvm::Function *Fn = cast<llvm::Function>(Val: GV); |
1717 | Fn->addFnAttr(Kind: "stackrealign" ); |
1718 | } |
1719 | |
1720 | addX86InterruptAttrs(FD, GV, CGM); |
1721 | } |
1722 | |
1723 | addStackProbeTargetAttributes(D, GV, CGM); |
1724 | } |
1725 | |
1726 | void X86_64ABIInfo::postMerge(unsigned AggregateSize, Class &Lo, |
1727 | Class &Hi) const { |
1728 | // AMD64-ABI 3.2.3p2: Rule 5. Then a post merger cleanup is done: |
1729 | // |
1730 | // (a) If one of the classes is Memory, the whole argument is passed in |
1731 | // memory. |
1732 | // |
1733 | // (b) If X87UP is not preceded by X87, the whole argument is passed in |
1734 | // memory. |
1735 | // |
1736 | // (c) If the size of the aggregate exceeds two eightbytes and the first |
1737 | // eightbyte isn't SSE or any other eightbyte isn't SSEUP, the whole |
1738 | // argument is passed in memory. NOTE: This is necessary to keep the |
1739 | // ABI working for processors that don't support the __m256 type. |
1740 | // |
1741 | // (d) If SSEUP is not preceded by SSE or SSEUP, it is converted to SSE. |
1742 | // |
1743 | // Some of these are enforced by the merging logic. Others can arise |
1744 | // only with unions; for example: |
1745 | // union { _Complex double; unsigned; } |
1746 | // |
1747 | // Note that clauses (b) and (c) were added in 0.98. |
1748 | // |
1749 | if (Hi == Memory) |
1750 | Lo = Memory; |
1751 | if (Hi == X87Up && Lo != X87 && honorsRevision0_98()) |
1752 | Lo = Memory; |
1753 | if (AggregateSize > 128 && (Lo != SSE || Hi != SSEUp)) |
1754 | Lo = Memory; |
1755 | if (Hi == SSEUp && Lo != SSE) |
1756 | Hi = SSE; |
1757 | } |
1758 | |
1759 | X86_64ABIInfo::Class X86_64ABIInfo::merge(Class Accum, Class Field) { |
1760 | // AMD64-ABI 3.2.3p2: Rule 4. Each field of an object is |
1761 | // classified recursively so that always two fields are |
1762 | // considered. The resulting class is calculated according to |
1763 | // the classes of the fields in the eightbyte: |
1764 | // |
1765 | // (a) If both classes are equal, this is the resulting class. |
1766 | // |
1767 | // (b) If one of the classes is NO_CLASS, the resulting class is |
1768 | // the other class. |
1769 | // |
1770 | // (c) If one of the classes is MEMORY, the result is the MEMORY |
1771 | // class. |
1772 | // |
1773 | // (d) If one of the classes is INTEGER, the result is the |
1774 | // INTEGER. |
1775 | // |
1776 | // (e) If one of the classes is X87, X87UP, COMPLEX_X87 class, |
1777 | // MEMORY is used as class. |
1778 | // |
1779 | // (f) Otherwise class SSE is used. |
1780 | |
1781 | // Accum should never be memory (we should have returned) or |
1782 | // ComplexX87 (because this cannot be passed in a structure). |
1783 | assert((Accum != Memory && Accum != ComplexX87) && |
1784 | "Invalid accumulated classification during merge." ); |
1785 | if (Accum == Field || Field == NoClass) |
1786 | return Accum; |
1787 | if (Field == Memory) |
1788 | return Memory; |
1789 | if (Accum == NoClass) |
1790 | return Field; |
1791 | if (Accum == Integer || Field == Integer) |
1792 | return Integer; |
1793 | if (Field == X87 || Field == X87Up || Field == ComplexX87 || |
1794 | Accum == X87 || Accum == X87Up) |
1795 | return Memory; |
1796 | return SSE; |
1797 | } |
1798 | |
1799 | void X86_64ABIInfo::classify(QualType Ty, uint64_t OffsetBase, Class &Lo, |
1800 | Class &Hi, bool isNamedArg, bool IsRegCall) const { |
1801 | // FIXME: This code can be simplified by introducing a simple value class for |
1802 | // Class pairs with appropriate constructor methods for the various |
1803 | // situations. |
1804 | |
1805 | // FIXME: Some of the split computations are wrong; unaligned vectors |
1806 | // shouldn't be passed in registers for example, so there is no chance they |
1807 | // can straddle an eightbyte. Verify & simplify. |
1808 | |
1809 | Lo = Hi = NoClass; |
1810 | |
1811 | Class &Current = OffsetBase < 64 ? Lo : Hi; |
1812 | Current = Memory; |
1813 | |
1814 | if (const BuiltinType *BT = Ty->getAs<BuiltinType>()) { |
1815 | BuiltinType::Kind k = BT->getKind(); |
1816 | |
1817 | if (k == BuiltinType::Void) { |
1818 | Current = NoClass; |
1819 | } else if (k == BuiltinType::Int128 || k == BuiltinType::UInt128) { |
1820 | Lo = Integer; |
1821 | Hi = Integer; |
1822 | } else if (k >= BuiltinType::Bool && k <= BuiltinType::LongLong) { |
1823 | Current = Integer; |
1824 | } else if (k == BuiltinType::Float || k == BuiltinType::Double || |
1825 | k == BuiltinType::Float16 || k == BuiltinType::BFloat16) { |
1826 | Current = SSE; |
1827 | } else if (k == BuiltinType::Float128) { |
1828 | Lo = SSE; |
1829 | Hi = SSEUp; |
1830 | } else if (k == BuiltinType::LongDouble) { |
1831 | const llvm::fltSemantics *LDF = &getTarget().getLongDoubleFormat(); |
1832 | if (LDF == &llvm::APFloat::IEEEquad()) { |
1833 | Lo = SSE; |
1834 | Hi = SSEUp; |
1835 | } else if (LDF == &llvm::APFloat::x87DoubleExtended()) { |
1836 | Lo = X87; |
1837 | Hi = X87Up; |
1838 | } else if (LDF == &llvm::APFloat::IEEEdouble()) { |
1839 | Current = SSE; |
1840 | } else |
1841 | llvm_unreachable("unexpected long double representation!" ); |
1842 | } |
1843 | // FIXME: _Decimal32 and _Decimal64 are SSE. |
1844 | // FIXME: _float128 and _Decimal128 are (SSE, SSEUp). |
1845 | return; |
1846 | } |
1847 | |
1848 | if (const EnumType *ET = Ty->getAs<EnumType>()) { |
1849 | // Classify the underlying integer type. |
1850 | classify(Ty: ET->getDecl()->getIntegerType(), OffsetBase, Lo, Hi, isNamedArg); |
1851 | return; |
1852 | } |
1853 | |
1854 | if (Ty->hasPointerRepresentation()) { |
1855 | Current = Integer; |
1856 | return; |
1857 | } |
1858 | |
1859 | if (Ty->isMemberPointerType()) { |
1860 | if (Ty->isMemberFunctionPointerType()) { |
1861 | if (Has64BitPointers) { |
1862 | // If Has64BitPointers, this is an {i64, i64}, so classify both |
1863 | // Lo and Hi now. |
1864 | Lo = Hi = Integer; |
1865 | } else { |
1866 | // Otherwise, with 32-bit pointers, this is an {i32, i32}. If that |
1867 | // straddles an eightbyte boundary, Hi should be classified as well. |
1868 | uint64_t EB_FuncPtr = (OffsetBase) / 64; |
1869 | uint64_t EB_ThisAdj = (OffsetBase + 64 - 1) / 64; |
1870 | if (EB_FuncPtr != EB_ThisAdj) { |
1871 | Lo = Hi = Integer; |
1872 | } else { |
1873 | Current = Integer; |
1874 | } |
1875 | } |
1876 | } else { |
1877 | Current = Integer; |
1878 | } |
1879 | return; |
1880 | } |
1881 | |
1882 | if (const VectorType *VT = Ty->getAs<VectorType>()) { |
1883 | uint64_t Size = getContext().getTypeSize(T: VT); |
1884 | if (Size == 1 || Size == 8 || Size == 16 || Size == 32) { |
1885 | // gcc passes the following as integer: |
1886 | // 4 bytes - <4 x char>, <2 x short>, <1 x int>, <1 x float> |
1887 | // 2 bytes - <2 x char>, <1 x short> |
1888 | // 1 byte - <1 x char> |
1889 | Current = Integer; |
1890 | |
1891 | // If this type crosses an eightbyte boundary, it should be |
1892 | // split. |
1893 | uint64_t EB_Lo = (OffsetBase) / 64; |
1894 | uint64_t EB_Hi = (OffsetBase + Size - 1) / 64; |
1895 | if (EB_Lo != EB_Hi) |
1896 | Hi = Lo; |
1897 | } else if (Size == 64) { |
1898 | QualType ElementType = VT->getElementType(); |
1899 | |
1900 | // gcc passes <1 x double> in memory. :( |
1901 | if (ElementType->isSpecificBuiltinType(K: BuiltinType::Double)) |
1902 | return; |
1903 | |
1904 | // gcc passes <1 x long long> as SSE but clang used to unconditionally |
1905 | // pass them as integer. For platforms where clang is the de facto |
1906 | // platform compiler, we must continue to use integer. |
1907 | if (!classifyIntegerMMXAsSSE() && |
1908 | (ElementType->isSpecificBuiltinType(K: BuiltinType::LongLong) || |
1909 | ElementType->isSpecificBuiltinType(K: BuiltinType::ULongLong) || |
1910 | ElementType->isSpecificBuiltinType(K: BuiltinType::Long) || |
1911 | ElementType->isSpecificBuiltinType(K: BuiltinType::ULong))) |
1912 | Current = Integer; |
1913 | else |
1914 | Current = SSE; |
1915 | |
1916 | // If this type crosses an eightbyte boundary, it should be |
1917 | // split. |
1918 | if (OffsetBase && OffsetBase != 64) |
1919 | Hi = Lo; |
1920 | } else if (Size == 128 || |
1921 | (isNamedArg && Size <= getNativeVectorSizeForAVXABI(AVXLevel))) { |
1922 | QualType ElementType = VT->getElementType(); |
1923 | |
1924 | // gcc passes 256 and 512 bit <X x __int128> vectors in memory. :( |
1925 | if (passInt128VectorsInMem() && Size != 128 && |
1926 | (ElementType->isSpecificBuiltinType(K: BuiltinType::Int128) || |
1927 | ElementType->isSpecificBuiltinType(K: BuiltinType::UInt128))) |
1928 | return; |
1929 | |
1930 | // Arguments of 256-bits are split into four eightbyte chunks. The |
1931 | // least significant one belongs to class SSE and all the others to class |
1932 | // SSEUP. The original Lo and Hi design considers that types can't be |
1933 | // greater than 128-bits, so a 64-bit split in Hi and Lo makes sense. |
1934 | // This design isn't correct for 256-bits, but since there're no cases |
1935 | // where the upper parts would need to be inspected, avoid adding |
1936 | // complexity and just consider Hi to match the 64-256 part. |
1937 | // |
1938 | // Note that per 3.5.7 of AMD64-ABI, 256-bit args are only passed in |
1939 | // registers if they are "named", i.e. not part of the "..." of a |
1940 | // variadic function. |
1941 | // |
1942 | // Similarly, per 3.2.3. of the AVX512 draft, 512-bits ("named") args are |
1943 | // split into eight eightbyte chunks, one SSE and seven SSEUP. |
1944 | Lo = SSE; |
1945 | Hi = SSEUp; |
1946 | } |
1947 | return; |
1948 | } |
1949 | |
1950 | if (const ComplexType *CT = Ty->getAs<ComplexType>()) { |
1951 | QualType ET = getContext().getCanonicalType(T: CT->getElementType()); |
1952 | |
1953 | uint64_t Size = getContext().getTypeSize(T: Ty); |
1954 | if (ET->isIntegralOrEnumerationType()) { |
1955 | if (Size <= 64) |
1956 | Current = Integer; |
1957 | else if (Size <= 128) |
1958 | Lo = Hi = Integer; |
1959 | } else if (ET->isFloat16Type() || ET == getContext().FloatTy || |
1960 | ET->isBFloat16Type()) { |
1961 | Current = SSE; |
1962 | } else if (ET == getContext().DoubleTy) { |
1963 | Lo = Hi = SSE; |
1964 | } else if (ET == getContext().LongDoubleTy) { |
1965 | const llvm::fltSemantics *LDF = &getTarget().getLongDoubleFormat(); |
1966 | if (LDF == &llvm::APFloat::IEEEquad()) |
1967 | Current = Memory; |
1968 | else if (LDF == &llvm::APFloat::x87DoubleExtended()) |
1969 | Current = ComplexX87; |
1970 | else if (LDF == &llvm::APFloat::IEEEdouble()) |
1971 | Lo = Hi = SSE; |
1972 | else |
1973 | llvm_unreachable("unexpected long double representation!" ); |
1974 | } |
1975 | |
1976 | // If this complex type crosses an eightbyte boundary then it |
1977 | // should be split. |
1978 | uint64_t EB_Real = (OffsetBase) / 64; |
1979 | uint64_t EB_Imag = (OffsetBase + getContext().getTypeSize(T: ET)) / 64; |
1980 | if (Hi == NoClass && EB_Real != EB_Imag) |
1981 | Hi = Lo; |
1982 | |
1983 | return; |
1984 | } |
1985 | |
1986 | if (const auto *EITy = Ty->getAs<BitIntType>()) { |
1987 | if (EITy->getNumBits() <= 64) |
1988 | Current = Integer; |
1989 | else if (EITy->getNumBits() <= 128) |
1990 | Lo = Hi = Integer; |
1991 | // Larger values need to get passed in memory. |
1992 | return; |
1993 | } |
1994 | |
1995 | if (const ConstantArrayType *AT = getContext().getAsConstantArrayType(T: Ty)) { |
1996 | // Arrays are treated like structures. |
1997 | |
1998 | uint64_t Size = getContext().getTypeSize(T: Ty); |
1999 | |
2000 | // AMD64-ABI 3.2.3p2: Rule 1. If the size of an object is larger |
2001 | // than eight eightbytes, ..., it has class MEMORY. |
2002 | // regcall ABI doesn't have limitation to an object. The only limitation |
2003 | // is the free registers, which will be checked in computeInfo. |
2004 | if (!IsRegCall && Size > 512) |
2005 | return; |
2006 | |
2007 | // AMD64-ABI 3.2.3p2: Rule 1. If ..., or it contains unaligned |
2008 | // fields, it has class MEMORY. |
2009 | // |
2010 | // Only need to check alignment of array base. |
2011 | if (OffsetBase % getContext().getTypeAlign(T: AT->getElementType())) |
2012 | return; |
2013 | |
2014 | // Otherwise implement simplified merge. We could be smarter about |
2015 | // this, but it isn't worth it and would be harder to verify. |
2016 | Current = NoClass; |
2017 | uint64_t EltSize = getContext().getTypeSize(T: AT->getElementType()); |
2018 | uint64_t ArraySize = AT->getZExtSize(); |
2019 | |
2020 | // The only case a 256-bit wide vector could be used is when the array |
2021 | // contains a single 256-bit element. Since Lo and Hi logic isn't extended |
2022 | // to work for sizes wider than 128, early check and fallback to memory. |
2023 | // |
2024 | if (Size > 128 && |
2025 | (Size != EltSize || Size > getNativeVectorSizeForAVXABI(AVXLevel))) |
2026 | return; |
2027 | |
2028 | for (uint64_t i=0, Offset=OffsetBase; i<ArraySize; ++i, Offset += EltSize) { |
2029 | Class FieldLo, FieldHi; |
2030 | classify(Ty: AT->getElementType(), OffsetBase: Offset, Lo&: FieldLo, Hi&: FieldHi, isNamedArg); |
2031 | Lo = merge(Accum: Lo, Field: FieldLo); |
2032 | Hi = merge(Accum: Hi, Field: FieldHi); |
2033 | if (Lo == Memory || Hi == Memory) |
2034 | break; |
2035 | } |
2036 | |
2037 | postMerge(AggregateSize: Size, Lo, Hi); |
2038 | assert((Hi != SSEUp || Lo == SSE) && "Invalid SSEUp array classification." ); |
2039 | return; |
2040 | } |
2041 | |
2042 | if (const RecordType *RT = Ty->getAs<RecordType>()) { |
2043 | uint64_t Size = getContext().getTypeSize(T: Ty); |
2044 | |
2045 | // AMD64-ABI 3.2.3p2: Rule 1. If the size of an object is larger |
2046 | // than eight eightbytes, ..., it has class MEMORY. |
2047 | if (Size > 512) |
2048 | return; |
2049 | |
2050 | // AMD64-ABI 3.2.3p2: Rule 2. If a C++ object has either a non-trivial |
2051 | // copy constructor or a non-trivial destructor, it is passed by invisible |
2052 | // reference. |
2053 | if (getRecordArgABI(RT, CXXABI&: getCXXABI())) |
2054 | return; |
2055 | |
2056 | const RecordDecl *RD = RT->getDecl(); |
2057 | |
2058 | // Assume variable sized types are passed in memory. |
2059 | if (RD->hasFlexibleArrayMember()) |
2060 | return; |
2061 | |
2062 | const ASTRecordLayout &Layout = getContext().getASTRecordLayout(D: RD); |
2063 | |
2064 | // Reset Lo class, this will be recomputed. |
2065 | Current = NoClass; |
2066 | |
2067 | // If this is a C++ record, classify the bases first. |
2068 | if (const CXXRecordDecl *CXXRD = dyn_cast<CXXRecordDecl>(Val: RD)) { |
2069 | for (const auto &I : CXXRD->bases()) { |
2070 | assert(!I.isVirtual() && !I.getType()->isDependentType() && |
2071 | "Unexpected base class!" ); |
2072 | const auto *Base = |
2073 | cast<CXXRecordDecl>(Val: I.getType()->castAs<RecordType>()->getDecl()); |
2074 | |
2075 | // Classify this field. |
2076 | // |
2077 | // AMD64-ABI 3.2.3p2: Rule 3. If the size of the aggregate exceeds a |
2078 | // single eightbyte, each is classified separately. Each eightbyte gets |
2079 | // initialized to class NO_CLASS. |
2080 | Class FieldLo, FieldHi; |
2081 | uint64_t Offset = |
2082 | OffsetBase + getContext().toBits(CharSize: Layout.getBaseClassOffset(Base)); |
2083 | classify(Ty: I.getType(), OffsetBase: Offset, Lo&: FieldLo, Hi&: FieldHi, isNamedArg); |
2084 | Lo = merge(Accum: Lo, Field: FieldLo); |
2085 | Hi = merge(Accum: Hi, Field: FieldHi); |
2086 | if (returnCXXRecordGreaterThan128InMem() && |
2087 | (Size > 128 && (Size != getContext().getTypeSize(T: I.getType()) || |
2088 | Size > getNativeVectorSizeForAVXABI(AVXLevel)))) { |
2089 | // The only case a 256(or 512)-bit wide vector could be used to return |
2090 | // is when CXX record contains a single 256(or 512)-bit element. |
2091 | Lo = Memory; |
2092 | } |
2093 | if (Lo == Memory || Hi == Memory) { |
2094 | postMerge(AggregateSize: Size, Lo, Hi); |
2095 | return; |
2096 | } |
2097 | } |
2098 | } |
2099 | |
2100 | // Classify the fields one at a time, merging the results. |
2101 | unsigned idx = 0; |
2102 | bool UseClang11Compat = getContext().getLangOpts().getClangABICompat() <= |
2103 | LangOptions::ClangABI::Ver11 || |
2104 | getContext().getTargetInfo().getTriple().isPS(); |
2105 | bool IsUnion = RT->isUnionType() && !UseClang11Compat; |
2106 | |
2107 | for (RecordDecl::field_iterator i = RD->field_begin(), e = RD->field_end(); |
2108 | i != e; ++i, ++idx) { |
2109 | uint64_t Offset = OffsetBase + Layout.getFieldOffset(FieldNo: idx); |
2110 | bool BitField = i->isBitField(); |
2111 | |
2112 | // Ignore padding bit-fields. |
2113 | if (BitField && i->isUnnamedBitField()) |
2114 | continue; |
2115 | |
2116 | // AMD64-ABI 3.2.3p2: Rule 1. If the size of an object is larger than |
2117 | // eight eightbytes, or it contains unaligned fields, it has class MEMORY. |
2118 | // |
2119 | // The only case a 256-bit or a 512-bit wide vector could be used is when |
2120 | // the struct contains a single 256-bit or 512-bit element. Early check |
2121 | // and fallback to memory. |
2122 | // |
2123 | // FIXME: Extended the Lo and Hi logic properly to work for size wider |
2124 | // than 128. |
2125 | if (Size > 128 && |
2126 | ((!IsUnion && Size != getContext().getTypeSize(T: i->getType())) || |
2127 | Size > getNativeVectorSizeForAVXABI(AVXLevel))) { |
2128 | Lo = Memory; |
2129 | postMerge(AggregateSize: Size, Lo, Hi); |
2130 | return; |
2131 | } |
2132 | |
2133 | bool IsInMemory = |
2134 | Offset % getContext().getTypeAlign(T: i->getType().getCanonicalType()); |
2135 | // Note, skip this test for bit-fields, see below. |
2136 | if (!BitField && IsInMemory) { |
2137 | Lo = Memory; |
2138 | postMerge(AggregateSize: Size, Lo, Hi); |
2139 | return; |
2140 | } |
2141 | |
2142 | // Classify this field. |
2143 | // |
2144 | // AMD64-ABI 3.2.3p2: Rule 3. If the size of the aggregate |
2145 | // exceeds a single eightbyte, each is classified |
2146 | // separately. Each eightbyte gets initialized to class |
2147 | // NO_CLASS. |
2148 | Class FieldLo, FieldHi; |
2149 | |
2150 | // Bit-fields require special handling, they do not force the |
2151 | // structure to be passed in memory even if unaligned, and |
2152 | // therefore they can straddle an eightbyte. |
2153 | if (BitField) { |
2154 | assert(!i->isUnnamedBitField()); |
2155 | uint64_t Offset = OffsetBase + Layout.getFieldOffset(FieldNo: idx); |
2156 | uint64_t Size = i->getBitWidthValue(); |
2157 | |
2158 | uint64_t EB_Lo = Offset / 64; |
2159 | uint64_t EB_Hi = (Offset + Size - 1) / 64; |
2160 | |
2161 | if (EB_Lo) { |
2162 | assert(EB_Hi == EB_Lo && "Invalid classification, type > 16 bytes." ); |
2163 | FieldLo = NoClass; |
2164 | FieldHi = Integer; |
2165 | } else { |
2166 | FieldLo = Integer; |
2167 | FieldHi = EB_Hi ? Integer : NoClass; |
2168 | } |
2169 | } else |
2170 | classify(Ty: i->getType(), OffsetBase: Offset, Lo&: FieldLo, Hi&: FieldHi, isNamedArg); |
2171 | Lo = merge(Accum: Lo, Field: FieldLo); |
2172 | Hi = merge(Accum: Hi, Field: FieldHi); |
2173 | if (Lo == Memory || Hi == Memory) |
2174 | break; |
2175 | } |
2176 | |
2177 | postMerge(AggregateSize: Size, Lo, Hi); |
2178 | } |
2179 | } |
2180 | |
2181 | ABIArgInfo X86_64ABIInfo::getIndirectReturnResult(QualType Ty) const { |
2182 | // If this is a scalar LLVM value then assume LLVM will pass it in the right |
2183 | // place naturally. |
2184 | if (!isAggregateTypeForABI(T: Ty)) { |
2185 | // Treat an enum type as its underlying type. |
2186 | if (const EnumType *EnumTy = Ty->getAs<EnumType>()) |
2187 | Ty = EnumTy->getDecl()->getIntegerType(); |
2188 | |
2189 | if (Ty->isBitIntType()) |
2190 | return getNaturalAlignIndirect(Ty, AddrSpace: getDataLayout().getAllocaAddrSpace()); |
2191 | |
2192 | return (isPromotableIntegerTypeForABI(Ty) ? ABIArgInfo::getExtend(Ty) |
2193 | : ABIArgInfo::getDirect()); |
2194 | } |
2195 | |
2196 | return getNaturalAlignIndirect(Ty, AddrSpace: getDataLayout().getAllocaAddrSpace()); |
2197 | } |
2198 | |
2199 | bool X86_64ABIInfo::IsIllegalVectorType(QualType Ty) const { |
2200 | if (const VectorType *VecTy = Ty->getAs<VectorType>()) { |
2201 | uint64_t Size = getContext().getTypeSize(T: VecTy); |
2202 | unsigned LargestVector = getNativeVectorSizeForAVXABI(AVXLevel); |
2203 | if (Size <= 64 || Size > LargestVector) |
2204 | return true; |
2205 | QualType EltTy = VecTy->getElementType(); |
2206 | if (passInt128VectorsInMem() && |
2207 | (EltTy->isSpecificBuiltinType(K: BuiltinType::Int128) || |
2208 | EltTy->isSpecificBuiltinType(K: BuiltinType::UInt128))) |
2209 | return true; |
2210 | } |
2211 | |
2212 | return false; |
2213 | } |
2214 | |
2215 | ABIArgInfo X86_64ABIInfo::getIndirectResult(QualType Ty, |
2216 | unsigned freeIntRegs) const { |
2217 | // If this is a scalar LLVM value then assume LLVM will pass it in the right |
2218 | // place naturally. |
2219 | // |
2220 | // This assumption is optimistic, as there could be free registers available |
2221 | // when we need to pass this argument in memory, and LLVM could try to pass |
2222 | // the argument in the free register. This does not seem to happen currently, |
2223 | // but this code would be much safer if we could mark the argument with |
2224 | // 'onstack'. See PR12193. |
2225 | if (!isAggregateTypeForABI(T: Ty) && !IsIllegalVectorType(Ty) && |
2226 | !Ty->isBitIntType()) { |
2227 | // Treat an enum type as its underlying type. |
2228 | if (const EnumType *EnumTy = Ty->getAs<EnumType>()) |
2229 | Ty = EnumTy->getDecl()->getIntegerType(); |
2230 | |
2231 | return (isPromotableIntegerTypeForABI(Ty) ? ABIArgInfo::getExtend(Ty) |
2232 | : ABIArgInfo::getDirect()); |
2233 | } |
2234 | |
2235 | if (CGCXXABI::RecordArgABI RAA = getRecordArgABI(T: Ty, CXXABI&: getCXXABI())) |
2236 | return getNaturalAlignIndirect(Ty, AddrSpace: getDataLayout().getAllocaAddrSpace(), |
2237 | ByVal: RAA == CGCXXABI::RAA_DirectInMemory); |
2238 | |
2239 | // Compute the byval alignment. We specify the alignment of the byval in all |
2240 | // cases so that the mid-level optimizer knows the alignment of the byval. |
2241 | unsigned Align = std::max(a: getContext().getTypeAlign(T: Ty) / 8, b: 8U); |
2242 | |
2243 | // Attempt to avoid passing indirect results using byval when possible. This |
2244 | // is important for good codegen. |
2245 | // |
2246 | // We do this by coercing the value into a scalar type which the backend can |
2247 | // handle naturally (i.e., without using byval). |
2248 | // |
2249 | // For simplicity, we currently only do this when we have exhausted all of the |
2250 | // free integer registers. Doing this when there are free integer registers |
2251 | // would require more care, as we would have to ensure that the coerced value |
2252 | // did not claim the unused register. That would require either reording the |
2253 | // arguments to the function (so that any subsequent inreg values came first), |
2254 | // or only doing this optimization when there were no following arguments that |
2255 | // might be inreg. |
2256 | // |
2257 | // We currently expect it to be rare (particularly in well written code) for |
2258 | // arguments to be passed on the stack when there are still free integer |
2259 | // registers available (this would typically imply large structs being passed |
2260 | // by value), so this seems like a fair tradeoff for now. |
2261 | // |
2262 | // We can revisit this if the backend grows support for 'onstack' parameter |
2263 | // attributes. See PR12193. |
2264 | if (freeIntRegs == 0) { |
2265 | uint64_t Size = getContext().getTypeSize(T: Ty); |
2266 | |
2267 | // If this type fits in an eightbyte, coerce it into the matching integral |
2268 | // type, which will end up on the stack (with alignment 8). |
2269 | if (Align == 8 && Size <= 64) |
2270 | return ABIArgInfo::getDirect(T: llvm::IntegerType::get(C&: getVMContext(), |
2271 | NumBits: Size)); |
2272 | } |
2273 | |
2274 | return ABIArgInfo::getIndirect(Alignment: CharUnits::fromQuantity(Quantity: Align), |
2275 | AddrSpace: getDataLayout().getAllocaAddrSpace()); |
2276 | } |
2277 | |
2278 | /// The ABI specifies that a value should be passed in a full vector XMM/YMM |
2279 | /// register. Pick an LLVM IR type that will be passed as a vector register. |
2280 | llvm::Type *X86_64ABIInfo::GetByteVectorType(QualType Ty) const { |
2281 | // Wrapper structs/arrays that only contain vectors are passed just like |
2282 | // vectors; strip them off if present. |
2283 | if (const Type *InnerTy = isSingleElementStruct(T: Ty, Context&: getContext())) |
2284 | Ty = QualType(InnerTy, 0); |
2285 | |
2286 | llvm::Type *IRType = CGT.ConvertType(T: Ty); |
2287 | if (isa<llvm::VectorType>(Val: IRType)) { |
2288 | // Don't pass vXi128 vectors in their native type, the backend can't |
2289 | // legalize them. |
2290 | if (passInt128VectorsInMem() && |
2291 | cast<llvm::VectorType>(Val: IRType)->getElementType()->isIntegerTy(Bitwidth: 128)) { |
2292 | // Use a vXi64 vector. |
2293 | uint64_t Size = getContext().getTypeSize(T: Ty); |
2294 | return llvm::FixedVectorType::get(ElementType: llvm::Type::getInt64Ty(C&: getVMContext()), |
2295 | NumElts: Size / 64); |
2296 | } |
2297 | |
2298 | return IRType; |
2299 | } |
2300 | |
2301 | if (IRType->getTypeID() == llvm::Type::FP128TyID) |
2302 | return IRType; |
2303 | |
2304 | // We couldn't find the preferred IR vector type for 'Ty'. |
2305 | uint64_t Size = getContext().getTypeSize(T: Ty); |
2306 | assert((Size == 128 || Size == 256 || Size == 512) && "Invalid type found!" ); |
2307 | |
2308 | |
2309 | // Return a LLVM IR vector type based on the size of 'Ty'. |
2310 | return llvm::FixedVectorType::get(ElementType: llvm::Type::getDoubleTy(C&: getVMContext()), |
2311 | NumElts: Size / 64); |
2312 | } |
2313 | |
2314 | /// BitsContainNoUserData - Return true if the specified [start,end) bit range |
2315 | /// is known to either be off the end of the specified type or being in |
2316 | /// alignment padding. The user type specified is known to be at most 128 bits |
2317 | /// in size, and have passed through X86_64ABIInfo::classify with a successful |
2318 | /// classification that put one of the two halves in the INTEGER class. |
2319 | /// |
2320 | /// It is conservatively correct to return false. |
2321 | static bool BitsContainNoUserData(QualType Ty, unsigned StartBit, |
2322 | unsigned EndBit, ASTContext &Context) { |
2323 | // If the bytes being queried are off the end of the type, there is no user |
2324 | // data hiding here. This handles analysis of builtins, vectors and other |
2325 | // types that don't contain interesting padding. |
2326 | unsigned TySize = (unsigned)Context.getTypeSize(T: Ty); |
2327 | if (TySize <= StartBit) |
2328 | return true; |
2329 | |
2330 | if (const ConstantArrayType *AT = Context.getAsConstantArrayType(T: Ty)) { |
2331 | unsigned EltSize = (unsigned)Context.getTypeSize(T: AT->getElementType()); |
2332 | unsigned NumElts = (unsigned)AT->getZExtSize(); |
2333 | |
2334 | // Check each element to see if the element overlaps with the queried range. |
2335 | for (unsigned i = 0; i != NumElts; ++i) { |
2336 | // If the element is after the span we care about, then we're done.. |
2337 | unsigned EltOffset = i*EltSize; |
2338 | if (EltOffset >= EndBit) break; |
2339 | |
2340 | unsigned EltStart = EltOffset < StartBit ? StartBit-EltOffset :0; |
2341 | if (!BitsContainNoUserData(Ty: AT->getElementType(), StartBit: EltStart, |
2342 | EndBit: EndBit-EltOffset, Context)) |
2343 | return false; |
2344 | } |
2345 | // If it overlaps no elements, then it is safe to process as padding. |
2346 | return true; |
2347 | } |
2348 | |
2349 | if (const RecordType *RT = Ty->getAs<RecordType>()) { |
2350 | const RecordDecl *RD = RT->getDecl(); |
2351 | const ASTRecordLayout &Layout = Context.getASTRecordLayout(D: RD); |
2352 | |
2353 | // If this is a C++ record, check the bases first. |
2354 | if (const CXXRecordDecl *CXXRD = dyn_cast<CXXRecordDecl>(Val: RD)) { |
2355 | for (const auto &I : CXXRD->bases()) { |
2356 | assert(!I.isVirtual() && !I.getType()->isDependentType() && |
2357 | "Unexpected base class!" ); |
2358 | const auto *Base = |
2359 | cast<CXXRecordDecl>(Val: I.getType()->castAs<RecordType>()->getDecl()); |
2360 | |
2361 | // If the base is after the span we care about, ignore it. |
2362 | unsigned BaseOffset = Context.toBits(CharSize: Layout.getBaseClassOffset(Base)); |
2363 | if (BaseOffset >= EndBit) continue; |
2364 | |
2365 | unsigned BaseStart = BaseOffset < StartBit ? StartBit-BaseOffset :0; |
2366 | if (!BitsContainNoUserData(Ty: I.getType(), StartBit: BaseStart, |
2367 | EndBit: EndBit-BaseOffset, Context)) |
2368 | return false; |
2369 | } |
2370 | } |
2371 | |
2372 | // Verify that no field has data that overlaps the region of interest. Yes |
2373 | // this could be sped up a lot by being smarter about queried fields, |
2374 | // however we're only looking at structs up to 16 bytes, so we don't care |
2375 | // much. |
2376 | unsigned idx = 0; |
2377 | for (RecordDecl::field_iterator i = RD->field_begin(), e = RD->field_end(); |
2378 | i != e; ++i, ++idx) { |
2379 | unsigned FieldOffset = (unsigned)Layout.getFieldOffset(FieldNo: idx); |
2380 | |
2381 | // If we found a field after the region we care about, then we're done. |
2382 | if (FieldOffset >= EndBit) break; |
2383 | |
2384 | unsigned FieldStart = FieldOffset < StartBit ? StartBit-FieldOffset :0; |
2385 | if (!BitsContainNoUserData(Ty: i->getType(), StartBit: FieldStart, EndBit: EndBit-FieldOffset, |
2386 | Context)) |
2387 | return false; |
2388 | } |
2389 | |
2390 | // If nothing in this record overlapped the area of interest, then we're |
2391 | // clean. |
2392 | return true; |
2393 | } |
2394 | |
2395 | return false; |
2396 | } |
2397 | |
2398 | /// getFPTypeAtOffset - Return a floating point type at the specified offset. |
2399 | static llvm::Type *getFPTypeAtOffset(llvm::Type *IRType, unsigned IROffset, |
2400 | const llvm::DataLayout &TD) { |
2401 | if (IROffset == 0 && IRType->isFloatingPointTy()) |
2402 | return IRType; |
2403 | |
2404 | // If this is a struct, recurse into the field at the specified offset. |
2405 | if (llvm::StructType *STy = dyn_cast<llvm::StructType>(Val: IRType)) { |
2406 | if (!STy->getNumContainedTypes()) |
2407 | return nullptr; |
2408 | |
2409 | const llvm::StructLayout *SL = TD.getStructLayout(Ty: STy); |
2410 | unsigned Elt = SL->getElementContainingOffset(FixedOffset: IROffset); |
2411 | IROffset -= SL->getElementOffset(Idx: Elt); |
2412 | return getFPTypeAtOffset(IRType: STy->getElementType(N: Elt), IROffset, TD); |
2413 | } |
2414 | |
2415 | // If this is an array, recurse into the field at the specified offset. |
2416 | if (llvm::ArrayType *ATy = dyn_cast<llvm::ArrayType>(Val: IRType)) { |
2417 | llvm::Type *EltTy = ATy->getElementType(); |
2418 | unsigned EltSize = TD.getTypeAllocSize(Ty: EltTy); |
2419 | IROffset -= IROffset / EltSize * EltSize; |
2420 | return getFPTypeAtOffset(IRType: EltTy, IROffset, TD); |
2421 | } |
2422 | |
2423 | return nullptr; |
2424 | } |
2425 | |
2426 | /// GetSSETypeAtOffset - Return a type that will be passed by the backend in the |
2427 | /// low 8 bytes of an XMM register, corresponding to the SSE class. |
2428 | llvm::Type *X86_64ABIInfo:: |
2429 | GetSSETypeAtOffset(llvm::Type *IRType, unsigned IROffset, |
2430 | QualType SourceTy, unsigned SourceOffset) const { |
2431 | const llvm::DataLayout &TD = getDataLayout(); |
2432 | unsigned SourceSize = |
2433 | (unsigned)getContext().getTypeSize(T: SourceTy) / 8 - SourceOffset; |
2434 | llvm::Type *T0 = getFPTypeAtOffset(IRType, IROffset, TD); |
2435 | if (!T0 || T0->isDoubleTy()) |
2436 | return llvm::Type::getDoubleTy(C&: getVMContext()); |
2437 | |
2438 | // Get the adjacent FP type. |
2439 | llvm::Type *T1 = nullptr; |
2440 | unsigned T0Size = TD.getTypeAllocSize(Ty: T0); |
2441 | if (SourceSize > T0Size) |
2442 | T1 = getFPTypeAtOffset(IRType, IROffset: IROffset + T0Size, TD); |
2443 | if (T1 == nullptr) { |
2444 | // Check if IRType is a half/bfloat + float. float type will be in IROffset+4 due |
2445 | // to its alignment. |
2446 | if (T0->is16bitFPTy() && SourceSize > 4) |
2447 | T1 = getFPTypeAtOffset(IRType, IROffset: IROffset + 4, TD); |
2448 | // If we can't get a second FP type, return a simple half or float. |
2449 | // avx512fp16-abi.c:pr51813_2 shows it works to return float for |
2450 | // {float, i8} too. |
2451 | if (T1 == nullptr) |
2452 | return T0; |
2453 | } |
2454 | |
2455 | if (T0->isFloatTy() && T1->isFloatTy()) |
2456 | return llvm::FixedVectorType::get(ElementType: T0, NumElts: 2); |
2457 | |
2458 | if (T0->is16bitFPTy() && T1->is16bitFPTy()) { |
2459 | llvm::Type *T2 = nullptr; |
2460 | if (SourceSize > 4) |
2461 | T2 = getFPTypeAtOffset(IRType, IROffset: IROffset + 4, TD); |
2462 | if (T2 == nullptr) |
2463 | return llvm::FixedVectorType::get(ElementType: T0, NumElts: 2); |
2464 | return llvm::FixedVectorType::get(ElementType: T0, NumElts: 4); |
2465 | } |
2466 | |
2467 | if (T0->is16bitFPTy() || T1->is16bitFPTy()) |
2468 | return llvm::FixedVectorType::get(ElementType: llvm::Type::getHalfTy(C&: getVMContext()), NumElts: 4); |
2469 | |
2470 | return llvm::Type::getDoubleTy(C&: getVMContext()); |
2471 | } |
2472 | |
2473 | |
2474 | /// GetINTEGERTypeAtOffset - The ABI specifies that a value should be passed in |
2475 | /// an 8-byte GPR. This means that we either have a scalar or we are talking |
2476 | /// about the high or low part of an up-to-16-byte struct. This routine picks |
2477 | /// the best LLVM IR type to represent this, which may be i64 or may be anything |
2478 | /// else that the backend will pass in a GPR that works better (e.g. i8, %foo*, |
2479 | /// etc). |
2480 | /// |
2481 | /// PrefType is an LLVM IR type that corresponds to (part of) the IR type for |
2482 | /// the source type. IROffset is an offset in bytes into the LLVM IR type that |
2483 | /// the 8-byte value references. PrefType may be null. |
2484 | /// |
2485 | /// SourceTy is the source-level type for the entire argument. SourceOffset is |
2486 | /// an offset into this that we're processing (which is always either 0 or 8). |
2487 | /// |
2488 | llvm::Type *X86_64ABIInfo:: |
2489 | GetINTEGERTypeAtOffset(llvm::Type *IRType, unsigned IROffset, |
2490 | QualType SourceTy, unsigned SourceOffset) const { |
2491 | // If we're dealing with an un-offset LLVM IR type, then it means that we're |
2492 | // returning an 8-byte unit starting with it. See if we can safely use it. |
2493 | if (IROffset == 0) { |
2494 | // Pointers and int64's always fill the 8-byte unit. |
2495 | if ((isa<llvm::PointerType>(Val: IRType) && Has64BitPointers) || |
2496 | IRType->isIntegerTy(Bitwidth: 64)) |
2497 | return IRType; |
2498 | |
2499 | // If we have a 1/2/4-byte integer, we can use it only if the rest of the |
2500 | // goodness in the source type is just tail padding. This is allowed to |
2501 | // kick in for struct {double,int} on the int, but not on |
2502 | // struct{double,int,int} because we wouldn't return the second int. We |
2503 | // have to do this analysis on the source type because we can't depend on |
2504 | // unions being lowered a specific way etc. |
2505 | if (IRType->isIntegerTy(Bitwidth: 8) || IRType->isIntegerTy(Bitwidth: 16) || |
2506 | IRType->isIntegerTy(Bitwidth: 32) || |
2507 | (isa<llvm::PointerType>(Val: IRType) && !Has64BitPointers)) { |
2508 | unsigned BitWidth = isa<llvm::PointerType>(Val: IRType) ? 32 : |
2509 | cast<llvm::IntegerType>(Val: IRType)->getBitWidth(); |
2510 | |
2511 | if (BitsContainNoUserData(Ty: SourceTy, StartBit: SourceOffset*8+BitWidth, |
2512 | EndBit: SourceOffset*8+64, Context&: getContext())) |
2513 | return IRType; |
2514 | } |
2515 | } |
2516 | |
2517 | if (llvm::StructType *STy = dyn_cast<llvm::StructType>(Val: IRType)) { |
2518 | // If this is a struct, recurse into the field at the specified offset. |
2519 | const llvm::StructLayout *SL = getDataLayout().getStructLayout(Ty: STy); |
2520 | if (IROffset < SL->getSizeInBytes()) { |
2521 | unsigned FieldIdx = SL->getElementContainingOffset(FixedOffset: IROffset); |
2522 | IROffset -= SL->getElementOffset(Idx: FieldIdx); |
2523 | |
2524 | return GetINTEGERTypeAtOffset(IRType: STy->getElementType(N: FieldIdx), IROffset, |
2525 | SourceTy, SourceOffset); |
2526 | } |
2527 | } |
2528 | |
2529 | if (llvm::ArrayType *ATy = dyn_cast<llvm::ArrayType>(Val: IRType)) { |
2530 | llvm::Type *EltTy = ATy->getElementType(); |
2531 | unsigned EltSize = getDataLayout().getTypeAllocSize(Ty: EltTy); |
2532 | unsigned EltOffset = IROffset/EltSize*EltSize; |
2533 | return GetINTEGERTypeAtOffset(IRType: EltTy, IROffset: IROffset-EltOffset, SourceTy, |
2534 | SourceOffset); |
2535 | } |
2536 | |
2537 | // Okay, we don't have any better idea of what to pass, so we pass this in an |
2538 | // integer register that isn't too big to fit the rest of the struct. |
2539 | unsigned TySizeInBytes = |
2540 | (unsigned)getContext().getTypeSizeInChars(T: SourceTy).getQuantity(); |
2541 | |
2542 | assert(TySizeInBytes != SourceOffset && "Empty field?" ); |
2543 | |
2544 | // It is always safe to classify this as an integer type up to i64 that |
2545 | // isn't larger than the structure. |
2546 | return llvm::IntegerType::get(C&: getVMContext(), |
2547 | NumBits: std::min(a: TySizeInBytes-SourceOffset, b: 8U)*8); |
2548 | } |
2549 | |
2550 | |
2551 | /// GetX86_64ByValArgumentPair - Given a high and low type that can ideally |
2552 | /// be used as elements of a two register pair to pass or return, return a |
2553 | /// first class aggregate to represent them. For example, if the low part of |
2554 | /// a by-value argument should be passed as i32* and the high part as float, |
2555 | /// return {i32*, float}. |
2556 | static llvm::Type * |
2557 | GetX86_64ByValArgumentPair(llvm::Type *Lo, llvm::Type *Hi, |
2558 | const llvm::DataLayout &TD) { |
2559 | // In order to correctly satisfy the ABI, we need to the high part to start |
2560 | // at offset 8. If the high and low parts we inferred are both 4-byte types |
2561 | // (e.g. i32 and i32) then the resultant struct type ({i32,i32}) won't have |
2562 | // the second element at offset 8. Check for this: |
2563 | unsigned LoSize = (unsigned)TD.getTypeAllocSize(Ty: Lo); |
2564 | llvm::Align HiAlign = TD.getABITypeAlign(Ty: Hi); |
2565 | unsigned HiStart = llvm::alignTo(Size: LoSize, A: HiAlign); |
2566 | assert(HiStart != 0 && HiStart <= 8 && "Invalid x86-64 argument pair!" ); |
2567 | |
2568 | // To handle this, we have to increase the size of the low part so that the |
2569 | // second element will start at an 8 byte offset. We can't increase the size |
2570 | // of the second element because it might make us access off the end of the |
2571 | // struct. |
2572 | if (HiStart != 8) { |
2573 | // There are usually two sorts of types the ABI generation code can produce |
2574 | // for the low part of a pair that aren't 8 bytes in size: half, float or |
2575 | // i8/i16/i32. This can also include pointers when they are 32-bit (X32 and |
2576 | // NaCl). |
2577 | // Promote these to a larger type. |
2578 | if (Lo->isHalfTy() || Lo->isFloatTy()) |
2579 | Lo = llvm::Type::getDoubleTy(C&: Lo->getContext()); |
2580 | else { |
2581 | assert((Lo->isIntegerTy() || Lo->isPointerTy()) |
2582 | && "Invalid/unknown lo type" ); |
2583 | Lo = llvm::Type::getInt64Ty(C&: Lo->getContext()); |
2584 | } |
2585 | } |
2586 | |
2587 | llvm::StructType *Result = llvm::StructType::get(elt1: Lo, elts: Hi); |
2588 | |
2589 | // Verify that the second element is at an 8-byte offset. |
2590 | assert(TD.getStructLayout(Result)->getElementOffset(1) == 8 && |
2591 | "Invalid x86-64 argument pair!" ); |
2592 | return Result; |
2593 | } |
2594 | |
2595 | ABIArgInfo X86_64ABIInfo:: |
2596 | classifyReturnType(QualType RetTy) const { |
2597 | // AMD64-ABI 3.2.3p4: Rule 1. Classify the return type with the |
2598 | // classification algorithm. |
2599 | X86_64ABIInfo::Class Lo, Hi; |
2600 | classify(Ty: RetTy, OffsetBase: 0, Lo, Hi, /*isNamedArg*/ true); |
2601 | |
2602 | // Check some invariants. |
2603 | assert((Hi != Memory || Lo == Memory) && "Invalid memory classification." ); |
2604 | assert((Hi != SSEUp || Lo == SSE) && "Invalid SSEUp classification." ); |
2605 | |
2606 | llvm::Type *ResType = nullptr; |
2607 | switch (Lo) { |
2608 | case NoClass: |
2609 | if (Hi == NoClass) |
2610 | return ABIArgInfo::getIgnore(); |
2611 | // If the low part is just padding, it takes no register, leave ResType |
2612 | // null. |
2613 | assert((Hi == SSE || Hi == Integer || Hi == X87Up) && |
2614 | "Unknown missing lo part" ); |
2615 | break; |
2616 | |
2617 | case SSEUp: |
2618 | case X87Up: |
2619 | llvm_unreachable("Invalid classification for lo word." ); |
2620 | |
2621 | // AMD64-ABI 3.2.3p4: Rule 2. Types of class memory are returned via |
2622 | // hidden argument. |
2623 | case Memory: |
2624 | return getIndirectReturnResult(Ty: RetTy); |
2625 | |
2626 | // AMD64-ABI 3.2.3p4: Rule 3. If the class is INTEGER, the next |
2627 | // available register of the sequence %rax, %rdx is used. |
2628 | case Integer: |
2629 | ResType = GetINTEGERTypeAtOffset(IRType: CGT.ConvertType(T: RetTy), IROffset: 0, SourceTy: RetTy, SourceOffset: 0); |
2630 | |
2631 | // If we have a sign or zero extended integer, make sure to return Extend |
2632 | // so that the parameter gets the right LLVM IR attributes. |
2633 | if (Hi == NoClass && isa<llvm::IntegerType>(Val: ResType)) { |
2634 | // Treat an enum type as its underlying type. |
2635 | if (const EnumType *EnumTy = RetTy->getAs<EnumType>()) |
2636 | RetTy = EnumTy->getDecl()->getIntegerType(); |
2637 | |
2638 | if (RetTy->isIntegralOrEnumerationType() && |
2639 | isPromotableIntegerTypeForABI(Ty: RetTy)) |
2640 | return ABIArgInfo::getExtend(Ty: RetTy); |
2641 | } |
2642 | break; |
2643 | |
2644 | // AMD64-ABI 3.2.3p4: Rule 4. If the class is SSE, the next |
2645 | // available SSE register of the sequence %xmm0, %xmm1 is used. |
2646 | case SSE: |
2647 | ResType = GetSSETypeAtOffset(IRType: CGT.ConvertType(T: RetTy), IROffset: 0, SourceTy: RetTy, SourceOffset: 0); |
2648 | break; |
2649 | |
2650 | // AMD64-ABI 3.2.3p4: Rule 6. If the class is X87, the value is |
2651 | // returned on the X87 stack in %st0 as 80-bit x87 number. |
2652 | case X87: |
2653 | ResType = llvm::Type::getX86_FP80Ty(C&: getVMContext()); |
2654 | break; |
2655 | |
2656 | // AMD64-ABI 3.2.3p4: Rule 8. If the class is COMPLEX_X87, the real |
2657 | // part of the value is returned in %st0 and the imaginary part in |
2658 | // %st1. |
2659 | case ComplexX87: |
2660 | assert(Hi == ComplexX87 && "Unexpected ComplexX87 classification." ); |
2661 | ResType = llvm::StructType::get(elt1: llvm::Type::getX86_FP80Ty(C&: getVMContext()), |
2662 | elts: llvm::Type::getX86_FP80Ty(C&: getVMContext())); |
2663 | break; |
2664 | } |
2665 | |
2666 | llvm::Type *HighPart = nullptr; |
2667 | switch (Hi) { |
2668 | // Memory was handled previously and X87 should |
2669 | // never occur as a hi class. |
2670 | case Memory: |
2671 | case X87: |
2672 | llvm_unreachable("Invalid classification for hi word." ); |
2673 | |
2674 | case ComplexX87: // Previously handled. |
2675 | case NoClass: |
2676 | break; |
2677 | |
2678 | case Integer: |
2679 | HighPart = GetINTEGERTypeAtOffset(IRType: CGT.ConvertType(T: RetTy), IROffset: 8, SourceTy: RetTy, SourceOffset: 8); |
2680 | if (Lo == NoClass) // Return HighPart at offset 8 in memory. |
2681 | return ABIArgInfo::getDirect(T: HighPart, Offset: 8); |
2682 | break; |
2683 | case SSE: |
2684 | HighPart = GetSSETypeAtOffset(IRType: CGT.ConvertType(T: RetTy), IROffset: 8, SourceTy: RetTy, SourceOffset: 8); |
2685 | if (Lo == NoClass) // Return HighPart at offset 8 in memory. |
2686 | return ABIArgInfo::getDirect(T: HighPart, Offset: 8); |
2687 | break; |
2688 | |
2689 | // AMD64-ABI 3.2.3p4: Rule 5. If the class is SSEUP, the eightbyte |
2690 | // is passed in the next available eightbyte chunk if the last used |
2691 | // vector register. |
2692 | // |
2693 | // SSEUP should always be preceded by SSE, just widen. |
2694 | case SSEUp: |
2695 | assert(Lo == SSE && "Unexpected SSEUp classification." ); |
2696 | ResType = GetByteVectorType(Ty: RetTy); |
2697 | break; |
2698 | |
2699 | // AMD64-ABI 3.2.3p4: Rule 7. If the class is X87UP, the value is |
2700 | // returned together with the previous X87 value in %st0. |
2701 | case X87Up: |
2702 | // If X87Up is preceded by X87, we don't need to do |
2703 | // anything. However, in some cases with unions it may not be |
2704 | // preceded by X87. In such situations we follow gcc and pass the |
2705 | // extra bits in an SSE reg. |
2706 | if (Lo != X87) { |
2707 | HighPart = GetSSETypeAtOffset(IRType: CGT.ConvertType(T: RetTy), IROffset: 8, SourceTy: RetTy, SourceOffset: 8); |
2708 | if (Lo == NoClass) // Return HighPart at offset 8 in memory. |
2709 | return ABIArgInfo::getDirect(T: HighPart, Offset: 8); |
2710 | } |
2711 | break; |
2712 | } |
2713 | |
2714 | // If a high part was specified, merge it together with the low part. It is |
2715 | // known to pass in the high eightbyte of the result. We do this by forming a |
2716 | // first class struct aggregate with the high and low part: {low, high} |
2717 | if (HighPart) |
2718 | ResType = GetX86_64ByValArgumentPair(Lo: ResType, Hi: HighPart, TD: getDataLayout()); |
2719 | |
2720 | return ABIArgInfo::getDirect(T: ResType); |
2721 | } |
2722 | |
2723 | ABIArgInfo |
2724 | X86_64ABIInfo::classifyArgumentType(QualType Ty, unsigned freeIntRegs, |
2725 | unsigned &neededInt, unsigned &neededSSE, |
2726 | bool isNamedArg, bool IsRegCall) const { |
2727 | Ty = useFirstFieldIfTransparentUnion(Ty); |
2728 | |
2729 | X86_64ABIInfo::Class Lo, Hi; |
2730 | classify(Ty, OffsetBase: 0, Lo, Hi, isNamedArg, IsRegCall); |
2731 | |
2732 | // Check some invariants. |
2733 | // FIXME: Enforce these by construction. |
2734 | assert((Hi != Memory || Lo == Memory) && "Invalid memory classification." ); |
2735 | assert((Hi != SSEUp || Lo == SSE) && "Invalid SSEUp classification." ); |
2736 | |
2737 | neededInt = 0; |
2738 | neededSSE = 0; |
2739 | llvm::Type *ResType = nullptr; |
2740 | switch (Lo) { |
2741 | case NoClass: |
2742 | if (Hi == NoClass) |
2743 | return ABIArgInfo::getIgnore(); |
2744 | // If the low part is just padding, it takes no register, leave ResType |
2745 | // null. |
2746 | assert((Hi == SSE || Hi == Integer || Hi == X87Up) && |
2747 | "Unknown missing lo part" ); |
2748 | break; |
2749 | |
2750 | // AMD64-ABI 3.2.3p3: Rule 1. If the class is MEMORY, pass the argument |
2751 | // on the stack. |
2752 | case Memory: |
2753 | |
2754 | // AMD64-ABI 3.2.3p3: Rule 5. If the class is X87, X87UP or |
2755 | // COMPLEX_X87, it is passed in memory. |
2756 | case X87: |
2757 | case ComplexX87: |
2758 | if (getRecordArgABI(T: Ty, CXXABI&: getCXXABI()) == CGCXXABI::RAA_Indirect) |
2759 | ++neededInt; |
2760 | return getIndirectResult(Ty, freeIntRegs); |
2761 | |
2762 | case SSEUp: |
2763 | case X87Up: |
2764 | llvm_unreachable("Invalid classification for lo word." ); |
2765 | |
2766 | // AMD64-ABI 3.2.3p3: Rule 2. If the class is INTEGER, the next |
2767 | // available register of the sequence %rdi, %rsi, %rdx, %rcx, %r8 |
2768 | // and %r9 is used. |
2769 | case Integer: |
2770 | ++neededInt; |
2771 | |
2772 | // Pick an 8-byte type based on the preferred type. |
2773 | ResType = GetINTEGERTypeAtOffset(IRType: CGT.ConvertType(T: Ty), IROffset: 0, SourceTy: Ty, SourceOffset: 0); |
2774 | |
2775 | // If we have a sign or zero extended integer, make sure to return Extend |
2776 | // so that the parameter gets the right LLVM IR attributes. |
2777 | if (Hi == NoClass && isa<llvm::IntegerType>(Val: ResType)) { |
2778 | // Treat an enum type as its underlying type. |
2779 | if (const EnumType *EnumTy = Ty->getAs<EnumType>()) |
2780 | Ty = EnumTy->getDecl()->getIntegerType(); |
2781 | |
2782 | if (Ty->isIntegralOrEnumerationType() && |
2783 | isPromotableIntegerTypeForABI(Ty)) |
2784 | return ABIArgInfo::getExtend(Ty, T: CGT.ConvertType(T: Ty)); |
2785 | } |
2786 | |
2787 | break; |
2788 | |
2789 | // AMD64-ABI 3.2.3p3: Rule 3. If the class is SSE, the next |
2790 | // available SSE register is used, the registers are taken in the |
2791 | // order from %xmm0 to %xmm7. |
2792 | case SSE: { |
2793 | llvm::Type *IRType = CGT.ConvertType(T: Ty); |
2794 | ResType = GetSSETypeAtOffset(IRType, IROffset: 0, SourceTy: Ty, SourceOffset: 0); |
2795 | ++neededSSE; |
2796 | break; |
2797 | } |
2798 | } |
2799 | |
2800 | llvm::Type *HighPart = nullptr; |
2801 | switch (Hi) { |
2802 | // Memory was handled previously, ComplexX87 and X87 should |
2803 | // never occur as hi classes, and X87Up must be preceded by X87, |
2804 | // which is passed in memory. |
2805 | case Memory: |
2806 | case X87: |
2807 | case ComplexX87: |
2808 | llvm_unreachable("Invalid classification for hi word." ); |
2809 | |
2810 | case NoClass: break; |
2811 | |
2812 | case Integer: |
2813 | ++neededInt; |
2814 | // Pick an 8-byte type based on the preferred type. |
2815 | HighPart = GetINTEGERTypeAtOffset(IRType: CGT.ConvertType(T: Ty), IROffset: 8, SourceTy: Ty, SourceOffset: 8); |
2816 | |
2817 | if (Lo == NoClass) // Pass HighPart at offset 8 in memory. |
2818 | return ABIArgInfo::getDirect(T: HighPart, Offset: 8); |
2819 | break; |
2820 | |
2821 | // X87Up generally doesn't occur here (long double is passed in |
2822 | // memory), except in situations involving unions. |
2823 | case X87Up: |
2824 | case SSE: |
2825 | ++neededSSE; |
2826 | HighPart = GetSSETypeAtOffset(IRType: CGT.ConvertType(T: Ty), IROffset: 8, SourceTy: Ty, SourceOffset: 8); |
2827 | |
2828 | if (Lo == NoClass) // Pass HighPart at offset 8 in memory. |
2829 | return ABIArgInfo::getDirect(T: HighPart, Offset: 8); |
2830 | break; |
2831 | |
2832 | // AMD64-ABI 3.2.3p3: Rule 4. If the class is SSEUP, the |
2833 | // eightbyte is passed in the upper half of the last used SSE |
2834 | // register. This only happens when 128-bit vectors are passed. |
2835 | case SSEUp: |
2836 | assert(Lo == SSE && "Unexpected SSEUp classification" ); |
2837 | ResType = GetByteVectorType(Ty); |
2838 | break; |
2839 | } |
2840 | |
2841 | // If a high part was specified, merge it together with the low part. It is |
2842 | // known to pass in the high eightbyte of the result. We do this by forming a |
2843 | // first class struct aggregate with the high and low part: {low, high} |
2844 | if (HighPart) |
2845 | ResType = GetX86_64ByValArgumentPair(Lo: ResType, Hi: HighPart, TD: getDataLayout()); |
2846 | |
2847 | return ABIArgInfo::getDirect(T: ResType); |
2848 | } |
2849 | |
2850 | ABIArgInfo |
2851 | X86_64ABIInfo::classifyRegCallStructTypeImpl(QualType Ty, unsigned &NeededInt, |
2852 | unsigned &NeededSSE, |
2853 | unsigned &MaxVectorWidth) const { |
2854 | auto RT = Ty->getAs<RecordType>(); |
2855 | assert(RT && "classifyRegCallStructType only valid with struct types" ); |
2856 | |
2857 | if (RT->getDecl()->hasFlexibleArrayMember()) |
2858 | return getIndirectReturnResult(Ty); |
2859 | |
2860 | // Sum up bases |
2861 | if (auto CXXRD = dyn_cast<CXXRecordDecl>(Val: RT->getDecl())) { |
2862 | if (CXXRD->isDynamicClass()) { |
2863 | NeededInt = NeededSSE = 0; |
2864 | return getIndirectReturnResult(Ty); |
2865 | } |
2866 | |
2867 | for (const auto &I : CXXRD->bases()) |
2868 | if (classifyRegCallStructTypeImpl(Ty: I.getType(), NeededInt, NeededSSE, |
2869 | MaxVectorWidth) |
2870 | .isIndirect()) { |
2871 | NeededInt = NeededSSE = 0; |
2872 | return getIndirectReturnResult(Ty); |
2873 | } |
2874 | } |
2875 | |
2876 | // Sum up members |
2877 | for (const auto *FD : RT->getDecl()->fields()) { |
2878 | QualType MTy = FD->getType(); |
2879 | if (MTy->isRecordType() && !MTy->isUnionType()) { |
2880 | if (classifyRegCallStructTypeImpl(Ty: MTy, NeededInt, NeededSSE, |
2881 | MaxVectorWidth) |
2882 | .isIndirect()) { |
2883 | NeededInt = NeededSSE = 0; |
2884 | return getIndirectReturnResult(Ty); |
2885 | } |
2886 | } else { |
2887 | unsigned LocalNeededInt, LocalNeededSSE; |
2888 | if (classifyArgumentType(Ty: MTy, UINT_MAX, neededInt&: LocalNeededInt, neededSSE&: LocalNeededSSE, |
2889 | isNamedArg: true, IsRegCall: true) |
2890 | .isIndirect()) { |
2891 | NeededInt = NeededSSE = 0; |
2892 | return getIndirectReturnResult(Ty); |
2893 | } |
2894 | if (const auto *AT = getContext().getAsConstantArrayType(T: MTy)) |
2895 | MTy = AT->getElementType(); |
2896 | if (const auto *VT = MTy->getAs<VectorType>()) |
2897 | if (getContext().getTypeSize(T: VT) > MaxVectorWidth) |
2898 | MaxVectorWidth = getContext().getTypeSize(T: VT); |
2899 | NeededInt += LocalNeededInt; |
2900 | NeededSSE += LocalNeededSSE; |
2901 | } |
2902 | } |
2903 | |
2904 | return ABIArgInfo::getDirect(); |
2905 | } |
2906 | |
2907 | ABIArgInfo |
2908 | X86_64ABIInfo::classifyRegCallStructType(QualType Ty, unsigned &NeededInt, |
2909 | unsigned &NeededSSE, |
2910 | unsigned &MaxVectorWidth) const { |
2911 | |
2912 | NeededInt = 0; |
2913 | NeededSSE = 0; |
2914 | MaxVectorWidth = 0; |
2915 | |
2916 | return classifyRegCallStructTypeImpl(Ty, NeededInt, NeededSSE, |
2917 | MaxVectorWidth); |
2918 | } |
2919 | |
2920 | void X86_64ABIInfo::computeInfo(CGFunctionInfo &FI) const { |
2921 | |
2922 | const unsigned CallingConv = FI.getCallingConvention(); |
2923 | // It is possible to force Win64 calling convention on any x86_64 target by |
2924 | // using __attribute__((ms_abi)). In such case to correctly emit Win64 |
2925 | // compatible code delegate this call to WinX86_64ABIInfo::computeInfo. |
2926 | if (CallingConv == llvm::CallingConv::Win64) { |
2927 | WinX86_64ABIInfo Win64ABIInfo(CGT, AVXLevel); |
2928 | Win64ABIInfo.computeInfo(FI); |
2929 | return; |
2930 | } |
2931 | |
2932 | bool IsRegCall = CallingConv == llvm::CallingConv::X86_RegCall; |
2933 | |
2934 | // Keep track of the number of assigned registers. |
2935 | unsigned FreeIntRegs = IsRegCall ? 11 : 6; |
2936 | unsigned FreeSSERegs = IsRegCall ? 16 : 8; |
2937 | unsigned NeededInt = 0, NeededSSE = 0, MaxVectorWidth = 0; |
2938 | |
2939 | if (!::classifyReturnType(CXXABI: getCXXABI(), FI, Info: *this)) { |
2940 | if (IsRegCall && FI.getReturnType()->getTypePtr()->isRecordType() && |
2941 | !FI.getReturnType()->getTypePtr()->isUnionType()) { |
2942 | FI.getReturnInfo() = classifyRegCallStructType( |
2943 | Ty: FI.getReturnType(), NeededInt, NeededSSE, MaxVectorWidth); |
2944 | if (FreeIntRegs >= NeededInt && FreeSSERegs >= NeededSSE) { |
2945 | FreeIntRegs -= NeededInt; |
2946 | FreeSSERegs -= NeededSSE; |
2947 | } else { |
2948 | FI.getReturnInfo() = getIndirectReturnResult(Ty: FI.getReturnType()); |
2949 | } |
2950 | } else if (IsRegCall && FI.getReturnType()->getAs<ComplexType>() && |
2951 | getContext().getCanonicalType(T: FI.getReturnType() |
2952 | ->getAs<ComplexType>() |
2953 | ->getElementType()) == |
2954 | getContext().LongDoubleTy) |
2955 | // Complex Long Double Type is passed in Memory when Regcall |
2956 | // calling convention is used. |
2957 | FI.getReturnInfo() = getIndirectReturnResult(Ty: FI.getReturnType()); |
2958 | else |
2959 | FI.getReturnInfo() = classifyReturnType(RetTy: FI.getReturnType()); |
2960 | } |
2961 | |
2962 | // If the return value is indirect, then the hidden argument is consuming one |
2963 | // integer register. |
2964 | if (FI.getReturnInfo().isIndirect()) |
2965 | --FreeIntRegs; |
2966 | else if (NeededSSE && MaxVectorWidth > 0) |
2967 | FI.setMaxVectorWidth(MaxVectorWidth); |
2968 | |
2969 | // The chain argument effectively gives us another free register. |
2970 | if (FI.isChainCall()) |
2971 | ++FreeIntRegs; |
2972 | |
2973 | unsigned NumRequiredArgs = FI.getNumRequiredArgs(); |
2974 | // AMD64-ABI 3.2.3p3: Once arguments are classified, the registers |
2975 | // get assigned (in left-to-right order) for passing as follows... |
2976 | unsigned ArgNo = 0; |
2977 | for (CGFunctionInfo::arg_iterator it = FI.arg_begin(), ie = FI.arg_end(); |
2978 | it != ie; ++it, ++ArgNo) { |
2979 | bool IsNamedArg = ArgNo < NumRequiredArgs; |
2980 | |
2981 | if (IsRegCall && it->type->isStructureOrClassType()) |
2982 | it->info = classifyRegCallStructType(Ty: it->type, NeededInt, NeededSSE, |
2983 | MaxVectorWidth); |
2984 | else |
2985 | it->info = classifyArgumentType(Ty: it->type, freeIntRegs: FreeIntRegs, neededInt&: NeededInt, |
2986 | neededSSE&: NeededSSE, isNamedArg: IsNamedArg); |
2987 | |
2988 | // AMD64-ABI 3.2.3p3: If there are no registers available for any |
2989 | // eightbyte of an argument, the whole argument is passed on the |
2990 | // stack. If registers have already been assigned for some |
2991 | // eightbytes of such an argument, the assignments get reverted. |
2992 | if (FreeIntRegs >= NeededInt && FreeSSERegs >= NeededSSE) { |
2993 | FreeIntRegs -= NeededInt; |
2994 | FreeSSERegs -= NeededSSE; |
2995 | if (MaxVectorWidth > FI.getMaxVectorWidth()) |
2996 | FI.setMaxVectorWidth(MaxVectorWidth); |
2997 | } else { |
2998 | it->info = getIndirectResult(Ty: it->type, freeIntRegs: FreeIntRegs); |
2999 | } |
3000 | } |
3001 | } |
3002 | |
3003 | static Address EmitX86_64VAArgFromMemory(CodeGenFunction &CGF, |
3004 | Address VAListAddr, QualType Ty) { |
3005 | Address overflow_arg_area_p = |
3006 | CGF.Builder.CreateStructGEP(Addr: VAListAddr, Index: 2, Name: "overflow_arg_area_p" ); |
3007 | llvm::Value *overflow_arg_area = |
3008 | CGF.Builder.CreateLoad(Addr: overflow_arg_area_p, Name: "overflow_arg_area" ); |
3009 | |
3010 | // AMD64-ABI 3.5.7p5: Step 7. Align l->overflow_arg_area upwards to a 16 |
3011 | // byte boundary if alignment needed by type exceeds 8 byte boundary. |
3012 | // It isn't stated explicitly in the standard, but in practice we use |
3013 | // alignment greater than 16 where necessary. |
3014 | CharUnits Align = CGF.getContext().getTypeAlignInChars(T: Ty); |
3015 | if (Align > CharUnits::fromQuantity(Quantity: 8)) { |
3016 | overflow_arg_area = emitRoundPointerUpToAlignment(CGF, Ptr: overflow_arg_area, |
3017 | Align); |
3018 | } |
3019 | |
3020 | // AMD64-ABI 3.5.7p5: Step 8. Fetch type from l->overflow_arg_area. |
3021 | llvm::Type *LTy = CGF.ConvertTypeForMem(T: Ty); |
3022 | llvm::Value *Res = overflow_arg_area; |
3023 | |
3024 | // AMD64-ABI 3.5.7p5: Step 9. Set l->overflow_arg_area to: |
3025 | // l->overflow_arg_area + sizeof(type). |
3026 | // AMD64-ABI 3.5.7p5: Step 10. Align l->overflow_arg_area upwards to |
3027 | // an 8 byte boundary. |
3028 | |
3029 | uint64_t SizeInBytes = (CGF.getContext().getTypeSize(T: Ty) + 7) / 8; |
3030 | llvm::Value *Offset = |
3031 | llvm::ConstantInt::get(Ty: CGF.Int32Ty, V: (SizeInBytes + 7) & ~7); |
3032 | overflow_arg_area = CGF.Builder.CreateGEP(Ty: CGF.Int8Ty, Ptr: overflow_arg_area, |
3033 | IdxList: Offset, Name: "overflow_arg_area.next" ); |
3034 | CGF.Builder.CreateStore(Val: overflow_arg_area, Addr: overflow_arg_area_p); |
3035 | |
3036 | // AMD64-ABI 3.5.7p5: Step 11. Return the fetched type. |
3037 | return Address(Res, LTy, Align); |
3038 | } |
3039 | |
3040 | RValue X86_64ABIInfo::EmitVAArg(CodeGenFunction &CGF, Address VAListAddr, |
3041 | QualType Ty, AggValueSlot Slot) const { |
3042 | // Assume that va_list type is correct; should be pointer to LLVM type: |
3043 | // struct { |
3044 | // i32 gp_offset; |
3045 | // i32 fp_offset; |
3046 | // i8* overflow_arg_area; |
3047 | // i8* reg_save_area; |
3048 | // }; |
3049 | unsigned neededInt, neededSSE; |
3050 | |
3051 | Ty = getContext().getCanonicalType(T: Ty); |
3052 | ABIArgInfo AI = classifyArgumentType(Ty, freeIntRegs: 0, neededInt, neededSSE, |
3053 | /*isNamedArg*/false); |
3054 | |
3055 | // Empty records are ignored for parameter passing purposes. |
3056 | if (AI.isIgnore()) |
3057 | return Slot.asRValue(); |
3058 | |
3059 | // AMD64-ABI 3.5.7p5: Step 1. Determine whether type may be passed |
3060 | // in the registers. If not go to step 7. |
3061 | if (!neededInt && !neededSSE) |
3062 | return CGF.EmitLoadOfAnyValue( |
3063 | V: CGF.MakeAddrLValue(Addr: EmitX86_64VAArgFromMemory(CGF, VAListAddr, Ty), T: Ty), |
3064 | Slot); |
3065 | |
3066 | // AMD64-ABI 3.5.7p5: Step 2. Compute num_gp to hold the number of |
3067 | // general purpose registers needed to pass type and num_fp to hold |
3068 | // the number of floating point registers needed. |
3069 | |
3070 | // AMD64-ABI 3.5.7p5: Step 3. Verify whether arguments fit into |
3071 | // registers. In the case: l->gp_offset > 48 - num_gp * 8 or |
3072 | // l->fp_offset > 304 - num_fp * 16 go to step 7. |
3073 | // |
3074 | // NOTE: 304 is a typo, there are (6 * 8 + 8 * 16) = 176 bytes of |
3075 | // register save space). |
3076 | |
3077 | llvm::Value *InRegs = nullptr; |
3078 | Address gp_offset_p = Address::invalid(), fp_offset_p = Address::invalid(); |
3079 | llvm::Value *gp_offset = nullptr, *fp_offset = nullptr; |
3080 | if (neededInt) { |
3081 | gp_offset_p = CGF.Builder.CreateStructGEP(Addr: VAListAddr, Index: 0, Name: "gp_offset_p" ); |
3082 | gp_offset = CGF.Builder.CreateLoad(Addr: gp_offset_p, Name: "gp_offset" ); |
3083 | InRegs = llvm::ConstantInt::get(Ty: CGF.Int32Ty, V: 48 - neededInt * 8); |
3084 | InRegs = CGF.Builder.CreateICmpULE(LHS: gp_offset, RHS: InRegs, Name: "fits_in_gp" ); |
3085 | } |
3086 | |
3087 | if (neededSSE) { |
3088 | fp_offset_p = CGF.Builder.CreateStructGEP(Addr: VAListAddr, Index: 1, Name: "fp_offset_p" ); |
3089 | fp_offset = CGF.Builder.CreateLoad(Addr: fp_offset_p, Name: "fp_offset" ); |
3090 | llvm::Value *FitsInFP = |
3091 | llvm::ConstantInt::get(Ty: CGF.Int32Ty, V: 176 - neededSSE * 16); |
3092 | FitsInFP = CGF.Builder.CreateICmpULE(LHS: fp_offset, RHS: FitsInFP, Name: "fits_in_fp" ); |
3093 | InRegs = InRegs ? CGF.Builder.CreateAnd(LHS: InRegs, RHS: FitsInFP) : FitsInFP; |
3094 | } |
3095 | |
3096 | llvm::BasicBlock *InRegBlock = CGF.createBasicBlock(name: "vaarg.in_reg" ); |
3097 | llvm::BasicBlock *InMemBlock = CGF.createBasicBlock(name: "vaarg.in_mem" ); |
3098 | llvm::BasicBlock *ContBlock = CGF.createBasicBlock(name: "vaarg.end" ); |
3099 | CGF.Builder.CreateCondBr(Cond: InRegs, True: InRegBlock, False: InMemBlock); |
3100 | |
3101 | // Emit code to load the value if it was passed in registers. |
3102 | |
3103 | CGF.EmitBlock(BB: InRegBlock); |
3104 | |
3105 | // AMD64-ABI 3.5.7p5: Step 4. Fetch type from l->reg_save_area with |
3106 | // an offset of l->gp_offset and/or l->fp_offset. This may require |
3107 | // copying to a temporary location in case the parameter is passed |
3108 | // in different register classes or requires an alignment greater |
3109 | // than 8 for general purpose registers and 16 for XMM registers. |
3110 | // |
3111 | // FIXME: This really results in shameful code when we end up needing to |
3112 | // collect arguments from different places; often what should result in a |
3113 | // simple assembling of a structure from scattered addresses has many more |
3114 | // loads than necessary. Can we clean this up? |
3115 | llvm::Type *LTy = CGF.ConvertTypeForMem(T: Ty); |
3116 | llvm::Value *RegSaveArea = CGF.Builder.CreateLoad( |
3117 | Addr: CGF.Builder.CreateStructGEP(Addr: VAListAddr, Index: 3), Name: "reg_save_area" ); |
3118 | |
3119 | Address RegAddr = Address::invalid(); |
3120 | if (neededInt && neededSSE) { |
3121 | // FIXME: Cleanup. |
3122 | assert(AI.isDirect() && "Unexpected ABI info for mixed regs" ); |
3123 | llvm::StructType *ST = cast<llvm::StructType>(Val: AI.getCoerceToType()); |
3124 | Address Tmp = CGF.CreateMemTemp(T: Ty); |
3125 | Tmp = Tmp.withElementType(ElemTy: ST); |
3126 | assert(ST->getNumElements() == 2 && "Unexpected ABI info for mixed regs" ); |
3127 | llvm::Type *TyLo = ST->getElementType(N: 0); |
3128 | llvm::Type *TyHi = ST->getElementType(N: 1); |
3129 | assert((TyLo->isFPOrFPVectorTy() ^ TyHi->isFPOrFPVectorTy()) && |
3130 | "Unexpected ABI info for mixed regs" ); |
3131 | llvm::Value *GPAddr = |
3132 | CGF.Builder.CreateGEP(Ty: CGF.Int8Ty, Ptr: RegSaveArea, IdxList: gp_offset); |
3133 | llvm::Value *FPAddr = |
3134 | CGF.Builder.CreateGEP(Ty: CGF.Int8Ty, Ptr: RegSaveArea, IdxList: fp_offset); |
3135 | llvm::Value *RegLoAddr = TyLo->isFPOrFPVectorTy() ? FPAddr : GPAddr; |
3136 | llvm::Value *RegHiAddr = TyLo->isFPOrFPVectorTy() ? GPAddr : FPAddr; |
3137 | |
3138 | // Copy the first element. |
3139 | // FIXME: Our choice of alignment here and below is probably pessimistic. |
3140 | llvm::Value *V = CGF.Builder.CreateAlignedLoad( |
3141 | Ty: TyLo, Addr: RegLoAddr, |
3142 | Align: CharUnits::fromQuantity(Quantity: getDataLayout().getABITypeAlign(Ty: TyLo))); |
3143 | CGF.Builder.CreateStore(Val: V, Addr: CGF.Builder.CreateStructGEP(Addr: Tmp, Index: 0)); |
3144 | |
3145 | // Copy the second element. |
3146 | V = CGF.Builder.CreateAlignedLoad( |
3147 | Ty: TyHi, Addr: RegHiAddr, |
3148 | Align: CharUnits::fromQuantity(Quantity: getDataLayout().getABITypeAlign(Ty: TyHi))); |
3149 | CGF.Builder.CreateStore(Val: V, Addr: CGF.Builder.CreateStructGEP(Addr: Tmp, Index: 1)); |
3150 | |
3151 | RegAddr = Tmp.withElementType(ElemTy: LTy); |
3152 | } else if (neededInt || neededSSE == 1) { |
3153 | // Copy to a temporary if necessary to ensure the appropriate alignment. |
3154 | auto TInfo = getContext().getTypeInfoInChars(T: Ty); |
3155 | uint64_t TySize = TInfo.Width.getQuantity(); |
3156 | CharUnits TyAlign = TInfo.Align; |
3157 | llvm::Type *CoTy = nullptr; |
3158 | if (AI.isDirect()) |
3159 | CoTy = AI.getCoerceToType(); |
3160 | |
3161 | llvm::Value *GpOrFpOffset = neededInt ? gp_offset : fp_offset; |
3162 | uint64_t Alignment = neededInt ? 8 : 16; |
3163 | uint64_t RegSize = neededInt ? neededInt * 8 : 16; |
3164 | // There are two cases require special handling: |
3165 | // 1) |
3166 | // ``` |
3167 | // struct { |
3168 | // struct {} a[8]; |
3169 | // int b; |
3170 | // }; |
3171 | // ``` |
3172 | // The lower 8 bytes of the structure are not stored, |
3173 | // so an 8-byte offset is needed when accessing the structure. |
3174 | // 2) |
3175 | // ``` |
3176 | // struct { |
3177 | // long long a; |
3178 | // struct {} b; |
3179 | // }; |
3180 | // ``` |
3181 | // The stored size of this structure is smaller than its actual size, |
3182 | // which may lead to reading past the end of the register save area. |
3183 | if (CoTy && (AI.getDirectOffset() == 8 || RegSize < TySize)) { |
3184 | Address Tmp = CGF.CreateMemTemp(T: Ty); |
3185 | llvm::Value *Addr = |
3186 | CGF.Builder.CreateGEP(Ty: CGF.Int8Ty, Ptr: RegSaveArea, IdxList: GpOrFpOffset); |
3187 | llvm::Value *Src = CGF.Builder.CreateAlignedLoad(Ty: CoTy, Addr, Align: TyAlign); |
3188 | llvm::Value *PtrOffset = |
3189 | llvm::ConstantInt::get(Ty: CGF.Int32Ty, V: AI.getDirectOffset()); |
3190 | Address Dst = Address( |
3191 | CGF.Builder.CreateGEP(Ty: CGF.Int8Ty, Ptr: Tmp.getBasePointer(), IdxList: PtrOffset), |
3192 | LTy, TyAlign); |
3193 | CGF.Builder.CreateStore(Val: Src, Addr: Dst); |
3194 | RegAddr = Tmp.withElementType(ElemTy: LTy); |
3195 | } else { |
3196 | RegAddr = |
3197 | Address(CGF.Builder.CreateGEP(Ty: CGF.Int8Ty, Ptr: RegSaveArea, IdxList: GpOrFpOffset), |
3198 | LTy, CharUnits::fromQuantity(Quantity: Alignment)); |
3199 | |
3200 | // Copy into a temporary if the type is more aligned than the |
3201 | // register save area. |
3202 | if (neededInt && TyAlign.getQuantity() > 8) { |
3203 | Address Tmp = CGF.CreateMemTemp(T: Ty); |
3204 | CGF.Builder.CreateMemCpy(Dest: Tmp, Src: RegAddr, Size: TySize, IsVolatile: false); |
3205 | RegAddr = Tmp; |
3206 | } |
3207 | } |
3208 | |
3209 | } else { |
3210 | assert(neededSSE == 2 && "Invalid number of needed registers!" ); |
3211 | // SSE registers are spaced 16 bytes apart in the register save |
3212 | // area, we need to collect the two eightbytes together. |
3213 | // The ABI isn't explicit about this, but it seems reasonable |
3214 | // to assume that the slots are 16-byte aligned, since the stack is |
3215 | // naturally 16-byte aligned and the prologue is expected to store |
3216 | // all the SSE registers to the RSA. |
3217 | Address RegAddrLo = Address(CGF.Builder.CreateGEP(Ty: CGF.Int8Ty, Ptr: RegSaveArea, |
3218 | IdxList: fp_offset), |
3219 | CGF.Int8Ty, CharUnits::fromQuantity(Quantity: 16)); |
3220 | Address RegAddrHi = |
3221 | CGF.Builder.CreateConstInBoundsByteGEP(Addr: RegAddrLo, |
3222 | Offset: CharUnits::fromQuantity(Quantity: 16)); |
3223 | llvm::Type *ST = AI.canHaveCoerceToType() |
3224 | ? AI.getCoerceToType() |
3225 | : llvm::StructType::get(elt1: CGF.DoubleTy, elts: CGF.DoubleTy); |
3226 | llvm::Value *V; |
3227 | Address Tmp = CGF.CreateMemTemp(T: Ty); |
3228 | Tmp = Tmp.withElementType(ElemTy: ST); |
3229 | V = CGF.Builder.CreateLoad( |
3230 | Addr: RegAddrLo.withElementType(ElemTy: ST->getStructElementType(N: 0))); |
3231 | CGF.Builder.CreateStore(Val: V, Addr: CGF.Builder.CreateStructGEP(Addr: Tmp, Index: 0)); |
3232 | V = CGF.Builder.CreateLoad( |
3233 | Addr: RegAddrHi.withElementType(ElemTy: ST->getStructElementType(N: 1))); |
3234 | CGF.Builder.CreateStore(Val: V, Addr: CGF.Builder.CreateStructGEP(Addr: Tmp, Index: 1)); |
3235 | |
3236 | RegAddr = Tmp.withElementType(ElemTy: LTy); |
3237 | } |
3238 | |
3239 | // AMD64-ABI 3.5.7p5: Step 5. Set: |
3240 | // l->gp_offset = l->gp_offset + num_gp * 8 |
3241 | // l->fp_offset = l->fp_offset + num_fp * 16. |
3242 | if (neededInt) { |
3243 | llvm::Value *Offset = llvm::ConstantInt::get(Ty: CGF.Int32Ty, V: neededInt * 8); |
3244 | CGF.Builder.CreateStore(Val: CGF.Builder.CreateAdd(LHS: gp_offset, RHS: Offset), |
3245 | Addr: gp_offset_p); |
3246 | } |
3247 | if (neededSSE) { |
3248 | llvm::Value *Offset = llvm::ConstantInt::get(Ty: CGF.Int32Ty, V: neededSSE * 16); |
3249 | CGF.Builder.CreateStore(Val: CGF.Builder.CreateAdd(LHS: fp_offset, RHS: Offset), |
3250 | Addr: fp_offset_p); |
3251 | } |
3252 | CGF.EmitBranch(Block: ContBlock); |
3253 | |
3254 | // Emit code to load the value if it was passed in memory. |
3255 | |
3256 | CGF.EmitBlock(BB: InMemBlock); |
3257 | Address MemAddr = EmitX86_64VAArgFromMemory(CGF, VAListAddr, Ty); |
3258 | |
3259 | // Return the appropriate result. |
3260 | |
3261 | CGF.EmitBlock(BB: ContBlock); |
3262 | Address ResAddr = emitMergePHI(CGF, Addr1: RegAddr, Block1: InRegBlock, Addr2: MemAddr, Block2: InMemBlock, |
3263 | Name: "vaarg.addr" ); |
3264 | return CGF.EmitLoadOfAnyValue(V: CGF.MakeAddrLValue(Addr: ResAddr, T: Ty), Slot); |
3265 | } |
3266 | |
3267 | RValue X86_64ABIInfo::EmitMSVAArg(CodeGenFunction &CGF, Address VAListAddr, |
3268 | QualType Ty, AggValueSlot Slot) const { |
3269 | // MS x64 ABI requirement: "Any argument that doesn't fit in 8 bytes, or is |
3270 | // not 1, 2, 4, or 8 bytes, must be passed by reference." |
3271 | uint64_t Width = getContext().getTypeSize(T: Ty); |
3272 | bool IsIndirect = Width > 64 || !llvm::isPowerOf2_64(Value: Width); |
3273 | |
3274 | return emitVoidPtrVAArg(CGF, VAListAddr, ValueTy: Ty, IsIndirect, |
3275 | ValueInfo: CGF.getContext().getTypeInfoInChars(T: Ty), |
3276 | SlotSizeAndAlign: CharUnits::fromQuantity(Quantity: 8), |
3277 | /*allowHigherAlign*/ AllowHigherAlign: false, Slot); |
3278 | } |
3279 | |
3280 | ABIArgInfo WinX86_64ABIInfo::reclassifyHvaArgForVectorCall( |
3281 | QualType Ty, unsigned &FreeSSERegs, const ABIArgInfo ¤t) const { |
3282 | const Type *Base = nullptr; |
3283 | uint64_t NumElts = 0; |
3284 | |
3285 | if (!Ty->isBuiltinType() && !Ty->isVectorType() && |
3286 | isHomogeneousAggregate(Ty, Base, Members&: NumElts) && FreeSSERegs >= NumElts) { |
3287 | FreeSSERegs -= NumElts; |
3288 | return getDirectX86Hva(); |
3289 | } |
3290 | return current; |
3291 | } |
3292 | |
3293 | ABIArgInfo WinX86_64ABIInfo::classify(QualType Ty, unsigned &FreeSSERegs, |
3294 | bool IsReturnType, bool IsVectorCall, |
3295 | bool IsRegCall) const { |
3296 | |
3297 | if (Ty->isVoidType()) |
3298 | return ABIArgInfo::getIgnore(); |
3299 | |
3300 | if (const EnumType *EnumTy = Ty->getAs<EnumType>()) |
3301 | Ty = EnumTy->getDecl()->getIntegerType(); |
3302 | |
3303 | TypeInfo Info = getContext().getTypeInfo(T: Ty); |
3304 | uint64_t Width = Info.Width; |
3305 | CharUnits Align = getContext().toCharUnitsFromBits(BitSize: Info.Align); |
3306 | |
3307 | const RecordType *RT = Ty->getAs<RecordType>(); |
3308 | if (RT) { |
3309 | if (!IsReturnType) { |
3310 | if (CGCXXABI::RecordArgABI RAA = getRecordArgABI(RT, CXXABI&: getCXXABI())) |
3311 | return getNaturalAlignIndirect(Ty, AddrSpace: getDataLayout().getAllocaAddrSpace(), |
3312 | ByVal: RAA == CGCXXABI::RAA_DirectInMemory); |
3313 | } |
3314 | |
3315 | if (RT->getDecl()->hasFlexibleArrayMember()) |
3316 | return getNaturalAlignIndirect(Ty, AddrSpace: getDataLayout().getAllocaAddrSpace(), |
3317 | /*ByVal=*/false); |
3318 | } |
3319 | |
3320 | const Type *Base = nullptr; |
3321 | uint64_t NumElts = 0; |
3322 | // vectorcall adds the concept of a homogenous vector aggregate, similar to |
3323 | // other targets. |
3324 | if ((IsVectorCall || IsRegCall) && |
3325 | isHomogeneousAggregate(Ty, Base, Members&: NumElts)) { |
3326 | if (IsRegCall) { |
3327 | if (FreeSSERegs >= NumElts) { |
3328 | FreeSSERegs -= NumElts; |
3329 | if (IsReturnType || Ty->isBuiltinType() || Ty->isVectorType()) |
3330 | return ABIArgInfo::getDirect(); |
3331 | return ABIArgInfo::getExpand(); |
3332 | } |
3333 | return ABIArgInfo::getIndirect( |
3334 | Alignment: Align, /*AddrSpace=*/getDataLayout().getAllocaAddrSpace(), |
3335 | /*ByVal=*/false); |
3336 | } else if (IsVectorCall) { |
3337 | if (FreeSSERegs >= NumElts && |
3338 | (IsReturnType || Ty->isBuiltinType() || Ty->isVectorType())) { |
3339 | FreeSSERegs -= NumElts; |
3340 | return ABIArgInfo::getDirect(); |
3341 | } else if (IsReturnType) { |
3342 | return ABIArgInfo::getExpand(); |
3343 | } else if (!Ty->isBuiltinType() && !Ty->isVectorType()) { |
3344 | // HVAs are delayed and reclassified in the 2nd step. |
3345 | return ABIArgInfo::getIndirect( |
3346 | Alignment: Align, /*AddrSpace=*/getDataLayout().getAllocaAddrSpace(), |
3347 | /*ByVal=*/false); |
3348 | } |
3349 | } |
3350 | } |
3351 | |
3352 | if (Ty->isMemberPointerType()) { |
3353 | // If the member pointer is represented by an LLVM int or ptr, pass it |
3354 | // directly. |
3355 | llvm::Type *LLTy = CGT.ConvertType(T: Ty); |
3356 | if (LLTy->isPointerTy() || LLTy->isIntegerTy()) |
3357 | return ABIArgInfo::getDirect(); |
3358 | } |
3359 | |
3360 | if (RT || Ty->isAnyComplexType() || Ty->isMemberPointerType()) { |
3361 | // MS x64 ABI requirement: "Any argument that doesn't fit in 8 bytes, or is |
3362 | // not 1, 2, 4, or 8 bytes, must be passed by reference." |
3363 | if (Width > 64 || !llvm::isPowerOf2_64(Value: Width)) |
3364 | return getNaturalAlignIndirect(Ty, AddrSpace: getDataLayout().getAllocaAddrSpace(), |
3365 | /*ByVal=*/false); |
3366 | |
3367 | // Otherwise, coerce it to a small integer. |
3368 | return ABIArgInfo::getDirect(T: llvm::IntegerType::get(C&: getVMContext(), NumBits: Width)); |
3369 | } |
3370 | |
3371 | if (const BuiltinType *BT = Ty->getAs<BuiltinType>()) { |
3372 | switch (BT->getKind()) { |
3373 | case BuiltinType::Bool: |
3374 | // Bool type is always extended to the ABI, other builtin types are not |
3375 | // extended. |
3376 | return ABIArgInfo::getExtend(Ty); |
3377 | |
3378 | case BuiltinType::LongDouble: |
3379 | // Mingw64 GCC uses the old 80 bit extended precision floating point |
3380 | // unit. It passes them indirectly through memory. |
3381 | if (IsMingw64) { |
3382 | const llvm::fltSemantics *LDF = &getTarget().getLongDoubleFormat(); |
3383 | if (LDF == &llvm::APFloat::x87DoubleExtended()) |
3384 | return ABIArgInfo::getIndirect( |
3385 | Alignment: Align, /*AddrSpace=*/getDataLayout().getAllocaAddrSpace(), |
3386 | /*ByVal=*/false); |
3387 | } |
3388 | break; |
3389 | |
3390 | case BuiltinType::Int128: |
3391 | case BuiltinType::UInt128: |
3392 | case BuiltinType::Float128: |
3393 | // 128-bit float and integer types share the same ABI. |
3394 | |
3395 | // If it's a parameter type, the normal ABI rule is that arguments larger |
3396 | // than 8 bytes are passed indirectly. GCC follows it. We follow it too, |
3397 | // even though it isn't particularly efficient. |
3398 | if (!IsReturnType) |
3399 | return ABIArgInfo::getIndirect( |
3400 | Alignment: Align, /*AddrSpace=*/getDataLayout().getAllocaAddrSpace(), |
3401 | /*ByVal=*/false); |
3402 | |
3403 | // Mingw64 GCC returns i128 in XMM0. Coerce to v2i64 to handle that. |
3404 | // Clang matches them for compatibility. |
3405 | // NOTE: GCC actually returns f128 indirectly but will hopefully change. |
3406 | // See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115054#c8. |
3407 | return ABIArgInfo::getDirect(T: llvm::FixedVectorType::get( |
3408 | ElementType: llvm::Type::getInt64Ty(C&: getVMContext()), NumElts: 2)); |
3409 | |
3410 | default: |
3411 | break; |
3412 | } |
3413 | } |
3414 | |
3415 | if (Ty->isBitIntType()) { |
3416 | // MS x64 ABI requirement: "Any argument that doesn't fit in 8 bytes, or is |
3417 | // not 1, 2, 4, or 8 bytes, must be passed by reference." |
3418 | // However, non-power-of-two bit-precise integers will be passed as 1, 2, 4, |
3419 | // or 8 bytes anyway as long is it fits in them, so we don't have to check |
3420 | // the power of 2. |
3421 | if (Width <= 64) |
3422 | return ABIArgInfo::getDirect(); |
3423 | return ABIArgInfo::getIndirect( |
3424 | Alignment: Align, /*AddrSpace=*/getDataLayout().getAllocaAddrSpace(), |
3425 | /*ByVal=*/false); |
3426 | } |
3427 | |
3428 | return ABIArgInfo::getDirect(); |
3429 | } |
3430 | |
3431 | void WinX86_64ABIInfo::computeInfo(CGFunctionInfo &FI) const { |
3432 | const unsigned CC = FI.getCallingConvention(); |
3433 | bool IsVectorCall = CC == llvm::CallingConv::X86_VectorCall; |
3434 | bool IsRegCall = CC == llvm::CallingConv::X86_RegCall; |
3435 | |
3436 | // If __attribute__((sysv_abi)) is in use, use the SysV argument |
3437 | // classification rules. |
3438 | if (CC == llvm::CallingConv::X86_64_SysV) { |
3439 | X86_64ABIInfo SysVABIInfo(CGT, AVXLevel); |
3440 | SysVABIInfo.computeInfo(FI); |
3441 | return; |
3442 | } |
3443 | |
3444 | unsigned FreeSSERegs = 0; |
3445 | if (IsVectorCall) { |
3446 | // We can use up to 4 SSE return registers with vectorcall. |
3447 | FreeSSERegs = 4; |
3448 | } else if (IsRegCall) { |
3449 | // RegCall gives us 16 SSE registers. |
3450 | FreeSSERegs = 16; |
3451 | } |
3452 | |
3453 | if (!getCXXABI().classifyReturnType(FI)) |
3454 | FI.getReturnInfo() = classify(Ty: FI.getReturnType(), FreeSSERegs, IsReturnType: true, |
3455 | IsVectorCall, IsRegCall); |
3456 | |
3457 | if (IsVectorCall) { |
3458 | // We can use up to 6 SSE register parameters with vectorcall. |
3459 | FreeSSERegs = 6; |
3460 | } else if (IsRegCall) { |
3461 | // RegCall gives us 16 SSE registers, we can reuse the return registers. |
3462 | FreeSSERegs = 16; |
3463 | } |
3464 | |
3465 | unsigned ArgNum = 0; |
3466 | unsigned ZeroSSERegs = 0; |
3467 | for (auto &I : FI.arguments()) { |
3468 | // Vectorcall in x64 only permits the first 6 arguments to be passed as |
3469 | // XMM/YMM registers. After the sixth argument, pretend no vector |
3470 | // registers are left. |
3471 | unsigned *MaybeFreeSSERegs = |
3472 | (IsVectorCall && ArgNum >= 6) ? &ZeroSSERegs : &FreeSSERegs; |
3473 | I.info = |
3474 | classify(Ty: I.type, FreeSSERegs&: *MaybeFreeSSERegs, IsReturnType: false, IsVectorCall, IsRegCall); |
3475 | ++ArgNum; |
3476 | } |
3477 | |
3478 | if (IsVectorCall) { |
3479 | // For vectorcall, assign aggregate HVAs to any free vector registers in a |
3480 | // second pass. |
3481 | for (auto &I : FI.arguments()) |
3482 | I.info = reclassifyHvaArgForVectorCall(Ty: I.type, FreeSSERegs, current: I.info); |
3483 | } |
3484 | } |
3485 | |
3486 | RValue WinX86_64ABIInfo::EmitVAArg(CodeGenFunction &CGF, Address VAListAddr, |
3487 | QualType Ty, AggValueSlot Slot) const { |
3488 | // MS x64 ABI requirement: "Any argument that doesn't fit in 8 bytes, or is |
3489 | // not 1, 2, 4, or 8 bytes, must be passed by reference." |
3490 | uint64_t Width = getContext().getTypeSize(T: Ty); |
3491 | bool IsIndirect = Width > 64 || !llvm::isPowerOf2_64(Value: Width); |
3492 | |
3493 | return emitVoidPtrVAArg(CGF, VAListAddr, ValueTy: Ty, IsIndirect, |
3494 | ValueInfo: CGF.getContext().getTypeInfoInChars(T: Ty), |
3495 | SlotSizeAndAlign: CharUnits::fromQuantity(Quantity: 8), |
3496 | /*allowHigherAlign*/ AllowHigherAlign: false, Slot); |
3497 | } |
3498 | |
3499 | std::unique_ptr<TargetCodeGenInfo> CodeGen::createX86_32TargetCodeGenInfo( |
3500 | CodeGenModule &CGM, bool DarwinVectorABI, bool Win32StructABI, |
3501 | unsigned NumRegisterParameters, bool SoftFloatABI) { |
3502 | bool RetSmallStructInRegABI = X86_32TargetCodeGenInfo::isStructReturnInRegABI( |
3503 | Triple: CGM.getTriple(), Opts: CGM.getCodeGenOpts()); |
3504 | return std::make_unique<X86_32TargetCodeGenInfo>( |
3505 | args&: CGM.getTypes(), args&: DarwinVectorABI, args&: RetSmallStructInRegABI, args&: Win32StructABI, |
3506 | args&: NumRegisterParameters, args&: SoftFloatABI); |
3507 | } |
3508 | |
3509 | std::unique_ptr<TargetCodeGenInfo> CodeGen::createWinX86_32TargetCodeGenInfo( |
3510 | CodeGenModule &CGM, bool DarwinVectorABI, bool Win32StructABI, |
3511 | unsigned NumRegisterParameters) { |
3512 | bool RetSmallStructInRegABI = X86_32TargetCodeGenInfo::isStructReturnInRegABI( |
3513 | Triple: CGM.getTriple(), Opts: CGM.getCodeGenOpts()); |
3514 | return std::make_unique<WinX86_32TargetCodeGenInfo>( |
3515 | args&: CGM.getTypes(), args&: DarwinVectorABI, args&: RetSmallStructInRegABI, args&: Win32StructABI, |
3516 | args&: NumRegisterParameters); |
3517 | } |
3518 | |
3519 | std::unique_ptr<TargetCodeGenInfo> |
3520 | CodeGen::createX86_64TargetCodeGenInfo(CodeGenModule &CGM, |
3521 | X86AVXABILevel AVXLevel) { |
3522 | return std::make_unique<X86_64TargetCodeGenInfo>(args&: CGM.getTypes(), args&: AVXLevel); |
3523 | } |
3524 | |
3525 | std::unique_ptr<TargetCodeGenInfo> |
3526 | CodeGen::createWinX86_64TargetCodeGenInfo(CodeGenModule &CGM, |
3527 | X86AVXABILevel AVXLevel) { |
3528 | return std::make_unique<WinX86_64TargetCodeGenInfo>(args&: CGM.getTypes(), args&: AVXLevel); |
3529 | } |
3530 | |