/src/llvm-project/clang/lib/CodeGen/Targets/ARM.cpp
Line | Count | Source (jump to first uncovered line) |
1 | | //===- ARM.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 | | |
12 | | using namespace clang; |
13 | | using namespace clang::CodeGen; |
14 | | |
15 | | //===----------------------------------------------------------------------===// |
16 | | // ARM ABI Implementation |
17 | | //===----------------------------------------------------------------------===// |
18 | | |
19 | | namespace { |
20 | | |
21 | | class ARMABIInfo : public ABIInfo { |
22 | | ARMABIKind Kind; |
23 | | bool IsFloatABISoftFP; |
24 | | |
25 | | public: |
26 | 0 | ARMABIInfo(CodeGenTypes &CGT, ARMABIKind Kind) : ABIInfo(CGT), Kind(Kind) { |
27 | 0 | setCCs(); |
28 | 0 | IsFloatABISoftFP = CGT.getCodeGenOpts().FloatABI == "softfp" || |
29 | 0 | CGT.getCodeGenOpts().FloatABI == ""; // default |
30 | 0 | } |
31 | | |
32 | 0 | bool isEABI() const { |
33 | 0 | switch (getTarget().getTriple().getEnvironment()) { |
34 | 0 | case llvm::Triple::Android: |
35 | 0 | case llvm::Triple::EABI: |
36 | 0 | case llvm::Triple::EABIHF: |
37 | 0 | case llvm::Triple::GNUEABI: |
38 | 0 | case llvm::Triple::GNUEABIHF: |
39 | 0 | case llvm::Triple::MuslEABI: |
40 | 0 | case llvm::Triple::MuslEABIHF: |
41 | 0 | return true; |
42 | 0 | default: |
43 | 0 | return getTarget().getTriple().isOHOSFamily(); |
44 | 0 | } |
45 | 0 | } |
46 | | |
47 | 0 | bool isEABIHF() const { |
48 | 0 | switch (getTarget().getTriple().getEnvironment()) { |
49 | 0 | case llvm::Triple::EABIHF: |
50 | 0 | case llvm::Triple::GNUEABIHF: |
51 | 0 | case llvm::Triple::MuslEABIHF: |
52 | 0 | return true; |
53 | 0 | default: |
54 | 0 | return false; |
55 | 0 | } |
56 | 0 | } |
57 | | |
58 | 0 | ARMABIKind getABIKind() const { return Kind; } |
59 | | |
60 | 0 | bool allowBFloatArgsAndRet() const override { |
61 | 0 | return !IsFloatABISoftFP && getTarget().hasBFloat16Type(); |
62 | 0 | } |
63 | | |
64 | | private: |
65 | | ABIArgInfo classifyReturnType(QualType RetTy, bool isVariadic, |
66 | | unsigned functionCallConv) const; |
67 | | ABIArgInfo classifyArgumentType(QualType RetTy, bool isVariadic, |
68 | | unsigned functionCallConv) const; |
69 | | ABIArgInfo classifyHomogeneousAggregate(QualType Ty, const Type *Base, |
70 | | uint64_t Members) const; |
71 | | ABIArgInfo coerceIllegalVector(QualType Ty) const; |
72 | | bool isIllegalVectorType(QualType Ty) const; |
73 | | bool containsAnyFP16Vectors(QualType Ty) const; |
74 | | |
75 | | bool isHomogeneousAggregateBaseType(QualType Ty) const override; |
76 | | bool isHomogeneousAggregateSmallEnough(const Type *Ty, |
77 | | uint64_t Members) const override; |
78 | | bool isZeroLengthBitfieldPermittedInHomogeneousAggregate() const override; |
79 | | |
80 | | bool isEffectivelyAAPCS_VFP(unsigned callConvention, bool acceptHalf) const; |
81 | | |
82 | | void computeInfo(CGFunctionInfo &FI) const override; |
83 | | |
84 | | Address EmitVAArg(CodeGenFunction &CGF, Address VAListAddr, |
85 | | QualType Ty) const override; |
86 | | |
87 | | llvm::CallingConv::ID getLLVMDefaultCC() const; |
88 | | llvm::CallingConv::ID getABIDefaultCC() const; |
89 | | void setCCs(); |
90 | | }; |
91 | | |
92 | | class ARMSwiftABIInfo : public SwiftABIInfo { |
93 | | public: |
94 | | explicit ARMSwiftABIInfo(CodeGenTypes &CGT) |
95 | 0 | : SwiftABIInfo(CGT, /*SwiftErrorInRegister=*/true) {} |
96 | | |
97 | | bool isLegalVectorType(CharUnits VectorSize, llvm::Type *EltTy, |
98 | | unsigned NumElts) const override; |
99 | | }; |
100 | | |
101 | | class ARMTargetCodeGenInfo : public TargetCodeGenInfo { |
102 | | public: |
103 | | ARMTargetCodeGenInfo(CodeGenTypes &CGT, ARMABIKind K) |
104 | 0 | : TargetCodeGenInfo(std::make_unique<ARMABIInfo>(CGT, K)) { |
105 | 0 | SwiftInfo = std::make_unique<ARMSwiftABIInfo>(CGT); |
106 | 0 | } |
107 | | |
108 | 0 | int getDwarfEHStackPointer(CodeGen::CodeGenModule &M) const override { |
109 | 0 | return 13; |
110 | 0 | } |
111 | | |
112 | 0 | StringRef getARCRetainAutoreleasedReturnValueMarker() const override { |
113 | 0 | return "mov\tr7, r7\t\t// marker for objc_retainAutoreleaseReturnValue"; |
114 | 0 | } |
115 | | |
116 | | bool initDwarfEHRegSizeTable(CodeGen::CodeGenFunction &CGF, |
117 | 0 | llvm::Value *Address) const override { |
118 | 0 | llvm::Value *Four8 = llvm::ConstantInt::get(CGF.Int8Ty, 4); |
119 | | |
120 | | // 0-15 are the 16 integer registers. |
121 | 0 | AssignToArrayRange(CGF.Builder, Address, Four8, 0, 15); |
122 | 0 | return false; |
123 | 0 | } |
124 | | |
125 | 0 | unsigned getSizeOfUnwindException() const override { |
126 | 0 | if (getABIInfo<ARMABIInfo>().isEABI()) |
127 | 0 | return 88; |
128 | 0 | return TargetCodeGenInfo::getSizeOfUnwindException(); |
129 | 0 | } |
130 | | |
131 | | void setTargetAttributes(const Decl *D, llvm::GlobalValue *GV, |
132 | 0 | CodeGen::CodeGenModule &CGM) const override { |
133 | 0 | if (GV->isDeclaration()) |
134 | 0 | return; |
135 | 0 | const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D); |
136 | 0 | if (!FD) |
137 | 0 | return; |
138 | 0 | auto *Fn = cast<llvm::Function>(GV); |
139 | |
|
140 | 0 | if (const auto *TA = FD->getAttr<TargetAttr>()) { |
141 | 0 | ParsedTargetAttr Attr = |
142 | 0 | CGM.getTarget().parseTargetAttr(TA->getFeaturesStr()); |
143 | 0 | if (!Attr.BranchProtection.empty()) { |
144 | 0 | TargetInfo::BranchProtectionInfo BPI; |
145 | 0 | StringRef DiagMsg; |
146 | 0 | StringRef Arch = |
147 | 0 | Attr.CPU.empty() ? CGM.getTarget().getTargetOpts().CPU : Attr.CPU; |
148 | 0 | if (!CGM.getTarget().validateBranchProtection(Attr.BranchProtection, |
149 | 0 | Arch, BPI, DiagMsg)) { |
150 | 0 | CGM.getDiags().Report( |
151 | 0 | D->getLocation(), |
152 | 0 | diag::warn_target_unsupported_branch_protection_attribute) |
153 | 0 | << Arch; |
154 | 0 | } else { |
155 | 0 | static const char *SignReturnAddrStr[] = {"none", "non-leaf", "all"}; |
156 | 0 | assert(static_cast<unsigned>(BPI.SignReturnAddr) <= 2 && |
157 | 0 | "Unexpected SignReturnAddressScopeKind"); |
158 | 0 | Fn->addFnAttr( |
159 | 0 | "sign-return-address", |
160 | 0 | SignReturnAddrStr[static_cast<int>(BPI.SignReturnAddr)]); |
161 | |
|
162 | 0 | Fn->addFnAttr("branch-target-enforcement", |
163 | 0 | BPI.BranchTargetEnforcement ? "true" : "false"); |
164 | 0 | } |
165 | 0 | } else if (CGM.getLangOpts().BranchTargetEnforcement || |
166 | 0 | CGM.getLangOpts().hasSignReturnAddress()) { |
167 | | // If the Branch Protection attribute is missing, validate the target |
168 | | // Architecture attribute against Branch Protection command line |
169 | | // settings. |
170 | 0 | if (!CGM.getTarget().isBranchProtectionSupportedArch(Attr.CPU)) |
171 | 0 | CGM.getDiags().Report( |
172 | 0 | D->getLocation(), |
173 | 0 | diag::warn_target_unsupported_branch_protection_attribute) |
174 | 0 | << Attr.CPU; |
175 | 0 | } |
176 | 0 | } |
177 | | |
178 | 0 | const ARMInterruptAttr *Attr = FD->getAttr<ARMInterruptAttr>(); |
179 | 0 | if (!Attr) |
180 | 0 | return; |
181 | | |
182 | 0 | const char *Kind; |
183 | 0 | switch (Attr->getInterrupt()) { |
184 | 0 | case ARMInterruptAttr::Generic: Kind = ""; break; |
185 | 0 | case ARMInterruptAttr::IRQ: Kind = "IRQ"; break; |
186 | 0 | case ARMInterruptAttr::FIQ: Kind = "FIQ"; break; |
187 | 0 | case ARMInterruptAttr::SWI: Kind = "SWI"; break; |
188 | 0 | case ARMInterruptAttr::ABORT: Kind = "ABORT"; break; |
189 | 0 | case ARMInterruptAttr::UNDEF: Kind = "UNDEF"; break; |
190 | 0 | } |
191 | | |
192 | 0 | Fn->addFnAttr("interrupt", Kind); |
193 | |
|
194 | 0 | ARMABIKind ABI = getABIInfo<ARMABIInfo>().getABIKind(); |
195 | 0 | if (ABI == ARMABIKind::APCS) |
196 | 0 | return; |
197 | | |
198 | | // AAPCS guarantees that sp will be 8-byte aligned on any public interface, |
199 | | // however this is not necessarily true on taking any interrupt. Instruct |
200 | | // the backend to perform a realignment as part of the function prologue. |
201 | 0 | llvm::AttrBuilder B(Fn->getContext()); |
202 | 0 | B.addStackAlignmentAttr(8); |
203 | 0 | Fn->addFnAttrs(B); |
204 | 0 | } |
205 | | }; |
206 | | |
207 | | class WindowsARMTargetCodeGenInfo : public ARMTargetCodeGenInfo { |
208 | | public: |
209 | | WindowsARMTargetCodeGenInfo(CodeGenTypes &CGT, ARMABIKind K) |
210 | 0 | : ARMTargetCodeGenInfo(CGT, K) {} |
211 | | |
212 | | void setTargetAttributes(const Decl *D, llvm::GlobalValue *GV, |
213 | | CodeGen::CodeGenModule &CGM) const override; |
214 | | |
215 | | void getDependentLibraryOption(llvm::StringRef Lib, |
216 | 0 | llvm::SmallString<24> &Opt) const override { |
217 | 0 | Opt = "/DEFAULTLIB:" + qualifyWindowsLibrary(Lib); |
218 | 0 | } |
219 | | |
220 | | void getDetectMismatchOption(llvm::StringRef Name, llvm::StringRef Value, |
221 | 0 | llvm::SmallString<32> &Opt) const override { |
222 | 0 | Opt = "/FAILIFMISMATCH:\"" + Name.str() + "=" + Value.str() + "\""; |
223 | 0 | } |
224 | | }; |
225 | | |
226 | | void WindowsARMTargetCodeGenInfo::setTargetAttributes( |
227 | 0 | const Decl *D, llvm::GlobalValue *GV, CodeGen::CodeGenModule &CGM) const { |
228 | 0 | ARMTargetCodeGenInfo::setTargetAttributes(D, GV, CGM); |
229 | 0 | if (GV->isDeclaration()) |
230 | 0 | return; |
231 | 0 | addStackProbeTargetAttributes(D, GV, CGM); |
232 | 0 | } |
233 | | } |
234 | | |
235 | 0 | void ARMABIInfo::computeInfo(CGFunctionInfo &FI) const { |
236 | 0 | if (!::classifyReturnType(getCXXABI(), FI, *this)) |
237 | 0 | FI.getReturnInfo() = classifyReturnType(FI.getReturnType(), FI.isVariadic(), |
238 | 0 | FI.getCallingConvention()); |
239 | |
|
240 | 0 | for (auto &I : FI.arguments()) |
241 | 0 | I.info = classifyArgumentType(I.type, FI.isVariadic(), |
242 | 0 | FI.getCallingConvention()); |
243 | | |
244 | | |
245 | | // Always honor user-specified calling convention. |
246 | 0 | if (FI.getCallingConvention() != llvm::CallingConv::C) |
247 | 0 | return; |
248 | | |
249 | 0 | llvm::CallingConv::ID cc = getRuntimeCC(); |
250 | 0 | if (cc != llvm::CallingConv::C) |
251 | 0 | FI.setEffectiveCallingConvention(cc); |
252 | 0 | } |
253 | | |
254 | | /// Return the default calling convention that LLVM will use. |
255 | 0 | llvm::CallingConv::ID ARMABIInfo::getLLVMDefaultCC() const { |
256 | | // The default calling convention that LLVM will infer. |
257 | 0 | if (isEABIHF() || getTarget().getTriple().isWatchABI()) |
258 | 0 | return llvm::CallingConv::ARM_AAPCS_VFP; |
259 | 0 | else if (isEABI()) |
260 | 0 | return llvm::CallingConv::ARM_AAPCS; |
261 | 0 | else |
262 | 0 | return llvm::CallingConv::ARM_APCS; |
263 | 0 | } |
264 | | |
265 | | /// Return the calling convention that our ABI would like us to use |
266 | | /// as the C calling convention. |
267 | 0 | llvm::CallingConv::ID ARMABIInfo::getABIDefaultCC() const { |
268 | 0 | switch (getABIKind()) { |
269 | 0 | case ARMABIKind::APCS: |
270 | 0 | return llvm::CallingConv::ARM_APCS; |
271 | 0 | case ARMABIKind::AAPCS: |
272 | 0 | return llvm::CallingConv::ARM_AAPCS; |
273 | 0 | case ARMABIKind::AAPCS_VFP: |
274 | 0 | return llvm::CallingConv::ARM_AAPCS_VFP; |
275 | 0 | case ARMABIKind::AAPCS16_VFP: |
276 | 0 | return llvm::CallingConv::ARM_AAPCS_VFP; |
277 | 0 | } |
278 | 0 | llvm_unreachable("bad ABI kind"); |
279 | 0 | } |
280 | | |
281 | 0 | void ARMABIInfo::setCCs() { |
282 | 0 | assert(getRuntimeCC() == llvm::CallingConv::C); |
283 | | |
284 | | // Don't muddy up the IR with a ton of explicit annotations if |
285 | | // they'd just match what LLVM will infer from the triple. |
286 | 0 | llvm::CallingConv::ID abiCC = getABIDefaultCC(); |
287 | 0 | if (abiCC != getLLVMDefaultCC()) |
288 | 0 | RuntimeCC = abiCC; |
289 | 0 | } |
290 | | |
291 | 0 | ABIArgInfo ARMABIInfo::coerceIllegalVector(QualType Ty) const { |
292 | 0 | uint64_t Size = getContext().getTypeSize(Ty); |
293 | 0 | if (Size <= 32) { |
294 | 0 | llvm::Type *ResType = |
295 | 0 | llvm::Type::getInt32Ty(getVMContext()); |
296 | 0 | return ABIArgInfo::getDirect(ResType); |
297 | 0 | } |
298 | 0 | if (Size == 64 || Size == 128) { |
299 | 0 | auto *ResType = llvm::FixedVectorType::get( |
300 | 0 | llvm::Type::getInt32Ty(getVMContext()), Size / 32); |
301 | 0 | return ABIArgInfo::getDirect(ResType); |
302 | 0 | } |
303 | 0 | return getNaturalAlignIndirect(Ty, /*ByVal=*/false); |
304 | 0 | } |
305 | | |
306 | | ABIArgInfo ARMABIInfo::classifyHomogeneousAggregate(QualType Ty, |
307 | | const Type *Base, |
308 | 0 | uint64_t Members) const { |
309 | 0 | assert(Base && "Base class should be set for homogeneous aggregate"); |
310 | | // Base can be a floating-point or a vector. |
311 | 0 | if (const VectorType *VT = Base->getAs<VectorType>()) { |
312 | | // FP16 vectors should be converted to integer vectors |
313 | 0 | if (!getTarget().hasLegalHalfType() && containsAnyFP16Vectors(Ty)) { |
314 | 0 | uint64_t Size = getContext().getTypeSize(VT); |
315 | 0 | auto *NewVecTy = llvm::FixedVectorType::get( |
316 | 0 | llvm::Type::getInt32Ty(getVMContext()), Size / 32); |
317 | 0 | llvm::Type *Ty = llvm::ArrayType::get(NewVecTy, Members); |
318 | 0 | return ABIArgInfo::getDirect(Ty, 0, nullptr, false); |
319 | 0 | } |
320 | 0 | } |
321 | 0 | unsigned Align = 0; |
322 | 0 | if (getABIKind() == ARMABIKind::AAPCS || |
323 | 0 | getABIKind() == ARMABIKind::AAPCS_VFP) { |
324 | | // For alignment adjusted HFAs, cap the argument alignment to 8, leave it |
325 | | // default otherwise. |
326 | 0 | Align = getContext().getTypeUnadjustedAlignInChars(Ty).getQuantity(); |
327 | 0 | unsigned BaseAlign = getContext().getTypeAlignInChars(Base).getQuantity(); |
328 | 0 | Align = (Align > BaseAlign && Align >= 8) ? 8 : 0; |
329 | 0 | } |
330 | 0 | return ABIArgInfo::getDirect(nullptr, 0, nullptr, false, Align); |
331 | 0 | } |
332 | | |
333 | | ABIArgInfo ARMABIInfo::classifyArgumentType(QualType Ty, bool isVariadic, |
334 | 0 | unsigned functionCallConv) const { |
335 | | // 6.1.2.1 The following argument types are VFP CPRCs: |
336 | | // A single-precision floating-point type (including promoted |
337 | | // half-precision types); A double-precision floating-point type; |
338 | | // A 64-bit or 128-bit containerized vector type; Homogeneous Aggregate |
339 | | // with a Base Type of a single- or double-precision floating-point type, |
340 | | // 64-bit containerized vectors or 128-bit containerized vectors with one |
341 | | // to four Elements. |
342 | | // Variadic functions should always marshal to the base standard. |
343 | 0 | bool IsAAPCS_VFP = |
344 | 0 | !isVariadic && isEffectivelyAAPCS_VFP(functionCallConv, /* AAPCS16 */ false); |
345 | |
|
346 | 0 | Ty = useFirstFieldIfTransparentUnion(Ty); |
347 | | |
348 | | // Handle illegal vector types here. |
349 | 0 | if (isIllegalVectorType(Ty)) |
350 | 0 | return coerceIllegalVector(Ty); |
351 | | |
352 | 0 | if (!isAggregateTypeForABI(Ty)) { |
353 | | // Treat an enum type as its underlying type. |
354 | 0 | if (const EnumType *EnumTy = Ty->getAs<EnumType>()) { |
355 | 0 | Ty = EnumTy->getDecl()->getIntegerType(); |
356 | 0 | } |
357 | |
|
358 | 0 | if (const auto *EIT = Ty->getAs<BitIntType>()) |
359 | 0 | if (EIT->getNumBits() > 64) |
360 | 0 | return getNaturalAlignIndirect(Ty, /*ByVal=*/true); |
361 | | |
362 | 0 | return (isPromotableIntegerTypeForABI(Ty) ? ABIArgInfo::getExtend(Ty) |
363 | 0 | : ABIArgInfo::getDirect()); |
364 | 0 | } |
365 | | |
366 | 0 | if (CGCXXABI::RecordArgABI RAA = getRecordArgABI(Ty, getCXXABI())) { |
367 | 0 | return getNaturalAlignIndirect(Ty, RAA == CGCXXABI::RAA_DirectInMemory); |
368 | 0 | } |
369 | | |
370 | | // Ignore empty records. |
371 | 0 | if (isEmptyRecord(getContext(), Ty, true)) |
372 | 0 | return ABIArgInfo::getIgnore(); |
373 | | |
374 | 0 | if (IsAAPCS_VFP) { |
375 | | // Homogeneous Aggregates need to be expanded when we can fit the aggregate |
376 | | // into VFP registers. |
377 | 0 | const Type *Base = nullptr; |
378 | 0 | uint64_t Members = 0; |
379 | 0 | if (isHomogeneousAggregate(Ty, Base, Members)) |
380 | 0 | return classifyHomogeneousAggregate(Ty, Base, Members); |
381 | 0 | } else if (getABIKind() == ARMABIKind::AAPCS16_VFP) { |
382 | | // WatchOS does have homogeneous aggregates. Note that we intentionally use |
383 | | // this convention even for a variadic function: the backend will use GPRs |
384 | | // if needed. |
385 | 0 | const Type *Base = nullptr; |
386 | 0 | uint64_t Members = 0; |
387 | 0 | if (isHomogeneousAggregate(Ty, Base, Members)) { |
388 | 0 | assert(Base && Members <= 4 && "unexpected homogeneous aggregate"); |
389 | 0 | llvm::Type *Ty = |
390 | 0 | llvm::ArrayType::get(CGT.ConvertType(QualType(Base, 0)), Members); |
391 | 0 | return ABIArgInfo::getDirect(Ty, 0, nullptr, false); |
392 | 0 | } |
393 | 0 | } |
394 | | |
395 | 0 | if (getABIKind() == ARMABIKind::AAPCS16_VFP && |
396 | 0 | getContext().getTypeSizeInChars(Ty) > CharUnits::fromQuantity(16)) { |
397 | | // WatchOS is adopting the 64-bit AAPCS rule on composite types: if they're |
398 | | // bigger than 128-bits, they get placed in space allocated by the caller, |
399 | | // and a pointer is passed. |
400 | 0 | return ABIArgInfo::getIndirect( |
401 | 0 | CharUnits::fromQuantity(getContext().getTypeAlign(Ty) / 8), false); |
402 | 0 | } |
403 | | |
404 | | // Support byval for ARM. |
405 | | // The ABI alignment for APCS is 4-byte and for AAPCS at least 4-byte and at |
406 | | // most 8-byte. We realign the indirect argument if type alignment is bigger |
407 | | // than ABI alignment. |
408 | 0 | uint64_t ABIAlign = 4; |
409 | 0 | uint64_t TyAlign; |
410 | 0 | if (getABIKind() == ARMABIKind::AAPCS_VFP || |
411 | 0 | getABIKind() == ARMABIKind::AAPCS) { |
412 | 0 | TyAlign = getContext().getTypeUnadjustedAlignInChars(Ty).getQuantity(); |
413 | 0 | ABIAlign = std::clamp(TyAlign, (uint64_t)4, (uint64_t)8); |
414 | 0 | } else { |
415 | 0 | TyAlign = getContext().getTypeAlignInChars(Ty).getQuantity(); |
416 | 0 | } |
417 | 0 | if (getContext().getTypeSizeInChars(Ty) > CharUnits::fromQuantity(64)) { |
418 | 0 | assert(getABIKind() != ARMABIKind::AAPCS16_VFP && "unexpected byval"); |
419 | 0 | return ABIArgInfo::getIndirect(CharUnits::fromQuantity(ABIAlign), |
420 | 0 | /*ByVal=*/true, |
421 | 0 | /*Realign=*/TyAlign > ABIAlign); |
422 | 0 | } |
423 | | |
424 | | // On RenderScript, coerce Aggregates <= 64 bytes to an integer array of |
425 | | // same size and alignment. |
426 | 0 | if (getTarget().isRenderScriptTarget()) { |
427 | 0 | return coerceToIntArray(Ty, getContext(), getVMContext()); |
428 | 0 | } |
429 | | |
430 | | // Otherwise, pass by coercing to a structure of the appropriate size. |
431 | 0 | llvm::Type* ElemTy; |
432 | 0 | unsigned SizeRegs; |
433 | | // FIXME: Try to match the types of the arguments more accurately where |
434 | | // we can. |
435 | 0 | if (TyAlign <= 4) { |
436 | 0 | ElemTy = llvm::Type::getInt32Ty(getVMContext()); |
437 | 0 | SizeRegs = (getContext().getTypeSize(Ty) + 31) / 32; |
438 | 0 | } else { |
439 | 0 | ElemTy = llvm::Type::getInt64Ty(getVMContext()); |
440 | 0 | SizeRegs = (getContext().getTypeSize(Ty) + 63) / 64; |
441 | 0 | } |
442 | |
|
443 | 0 | return ABIArgInfo::getDirect(llvm::ArrayType::get(ElemTy, SizeRegs)); |
444 | 0 | } |
445 | | |
446 | | static bool isIntegerLikeType(QualType Ty, ASTContext &Context, |
447 | 0 | llvm::LLVMContext &VMContext) { |
448 | | // APCS, C Language Calling Conventions, Non-Simple Return Values: A structure |
449 | | // is called integer-like if its size is less than or equal to one word, and |
450 | | // the offset of each of its addressable sub-fields is zero. |
451 | |
|
452 | 0 | uint64_t Size = Context.getTypeSize(Ty); |
453 | | |
454 | | // Check that the type fits in a word. |
455 | 0 | if (Size > 32) |
456 | 0 | return false; |
457 | | |
458 | | // FIXME: Handle vector types! |
459 | 0 | if (Ty->isVectorType()) |
460 | 0 | return false; |
461 | | |
462 | | // Float types are never treated as "integer like". |
463 | 0 | if (Ty->isRealFloatingType()) |
464 | 0 | return false; |
465 | | |
466 | | // If this is a builtin or pointer type then it is ok. |
467 | 0 | if (Ty->getAs<BuiltinType>() || Ty->isPointerType()) |
468 | 0 | return true; |
469 | | |
470 | | // Small complex integer types are "integer like". |
471 | 0 | if (const ComplexType *CT = Ty->getAs<ComplexType>()) |
472 | 0 | return isIntegerLikeType(CT->getElementType(), Context, VMContext); |
473 | | |
474 | | // Single element and zero sized arrays should be allowed, by the definition |
475 | | // above, but they are not. |
476 | | |
477 | | // Otherwise, it must be a record type. |
478 | 0 | const RecordType *RT = Ty->getAs<RecordType>(); |
479 | 0 | if (!RT) return false; |
480 | | |
481 | | // Ignore records with flexible arrays. |
482 | 0 | const RecordDecl *RD = RT->getDecl(); |
483 | 0 | if (RD->hasFlexibleArrayMember()) |
484 | 0 | return false; |
485 | | |
486 | | // Check that all sub-fields are at offset 0, and are themselves "integer |
487 | | // like". |
488 | 0 | const ASTRecordLayout &Layout = Context.getASTRecordLayout(RD); |
489 | |
|
490 | 0 | bool HadField = false; |
491 | 0 | unsigned idx = 0; |
492 | 0 | for (RecordDecl::field_iterator i = RD->field_begin(), e = RD->field_end(); |
493 | 0 | i != e; ++i, ++idx) { |
494 | 0 | const FieldDecl *FD = *i; |
495 | | |
496 | | // Bit-fields are not addressable, we only need to verify they are "integer |
497 | | // like". We still have to disallow a subsequent non-bitfield, for example: |
498 | | // struct { int : 0; int x } |
499 | | // is non-integer like according to gcc. |
500 | 0 | if (FD->isBitField()) { |
501 | 0 | if (!RD->isUnion()) |
502 | 0 | HadField = true; |
503 | |
|
504 | 0 | if (!isIntegerLikeType(FD->getType(), Context, VMContext)) |
505 | 0 | return false; |
506 | | |
507 | 0 | continue; |
508 | 0 | } |
509 | | |
510 | | // Check if this field is at offset 0. |
511 | 0 | if (Layout.getFieldOffset(idx) != 0) |
512 | 0 | return false; |
513 | | |
514 | 0 | if (!isIntegerLikeType(FD->getType(), Context, VMContext)) |
515 | 0 | return false; |
516 | | |
517 | | // Only allow at most one field in a structure. This doesn't match the |
518 | | // wording above, but follows gcc in situations with a field following an |
519 | | // empty structure. |
520 | 0 | if (!RD->isUnion()) { |
521 | 0 | if (HadField) |
522 | 0 | return false; |
523 | | |
524 | 0 | HadField = true; |
525 | 0 | } |
526 | 0 | } |
527 | | |
528 | 0 | return true; |
529 | 0 | } |
530 | | |
531 | | ABIArgInfo ARMABIInfo::classifyReturnType(QualType RetTy, bool isVariadic, |
532 | 0 | unsigned functionCallConv) const { |
533 | | |
534 | | // Variadic functions should always marshal to the base standard. |
535 | 0 | bool IsAAPCS_VFP = |
536 | 0 | !isVariadic && isEffectivelyAAPCS_VFP(functionCallConv, /* AAPCS16 */ true); |
537 | |
|
538 | 0 | if (RetTy->isVoidType()) |
539 | 0 | return ABIArgInfo::getIgnore(); |
540 | | |
541 | 0 | if (const VectorType *VT = RetTy->getAs<VectorType>()) { |
542 | | // Large vector types should be returned via memory. |
543 | 0 | if (getContext().getTypeSize(RetTy) > 128) |
544 | 0 | return getNaturalAlignIndirect(RetTy); |
545 | | // TODO: FP16/BF16 vectors should be converted to integer vectors |
546 | | // This check is similar to isIllegalVectorType - refactor? |
547 | 0 | if ((!getTarget().hasLegalHalfType() && |
548 | 0 | (VT->getElementType()->isFloat16Type() || |
549 | 0 | VT->getElementType()->isHalfType())) || |
550 | 0 | (IsFloatABISoftFP && |
551 | 0 | VT->getElementType()->isBFloat16Type())) |
552 | 0 | return coerceIllegalVector(RetTy); |
553 | 0 | } |
554 | | |
555 | 0 | if (!isAggregateTypeForABI(RetTy)) { |
556 | | // Treat an enum type as its underlying type. |
557 | 0 | if (const EnumType *EnumTy = RetTy->getAs<EnumType>()) |
558 | 0 | RetTy = EnumTy->getDecl()->getIntegerType(); |
559 | |
|
560 | 0 | if (const auto *EIT = RetTy->getAs<BitIntType>()) |
561 | 0 | if (EIT->getNumBits() > 64) |
562 | 0 | return getNaturalAlignIndirect(RetTy, /*ByVal=*/false); |
563 | | |
564 | 0 | return isPromotableIntegerTypeForABI(RetTy) ? ABIArgInfo::getExtend(RetTy) |
565 | 0 | : ABIArgInfo::getDirect(); |
566 | 0 | } |
567 | | |
568 | | // Are we following APCS? |
569 | 0 | if (getABIKind() == ARMABIKind::APCS) { |
570 | 0 | if (isEmptyRecord(getContext(), RetTy, false)) |
571 | 0 | return ABIArgInfo::getIgnore(); |
572 | | |
573 | | // Complex types are all returned as packed integers. |
574 | | // |
575 | | // FIXME: Consider using 2 x vector types if the back end handles them |
576 | | // correctly. |
577 | 0 | if (RetTy->isAnyComplexType()) |
578 | 0 | return ABIArgInfo::getDirect(llvm::IntegerType::get( |
579 | 0 | getVMContext(), getContext().getTypeSize(RetTy))); |
580 | | |
581 | | // Integer like structures are returned in r0. |
582 | 0 | if (isIntegerLikeType(RetTy, getContext(), getVMContext())) { |
583 | | // Return in the smallest viable integer type. |
584 | 0 | uint64_t Size = getContext().getTypeSize(RetTy); |
585 | 0 | if (Size <= 8) |
586 | 0 | return ABIArgInfo::getDirect(llvm::Type::getInt8Ty(getVMContext())); |
587 | 0 | if (Size <= 16) |
588 | 0 | return ABIArgInfo::getDirect(llvm::Type::getInt16Ty(getVMContext())); |
589 | 0 | return ABIArgInfo::getDirect(llvm::Type::getInt32Ty(getVMContext())); |
590 | 0 | } |
591 | | |
592 | | // Otherwise return in memory. |
593 | 0 | return getNaturalAlignIndirect(RetTy); |
594 | 0 | } |
595 | | |
596 | | // Otherwise this is an AAPCS variant. |
597 | | |
598 | 0 | if (isEmptyRecord(getContext(), RetTy, true)) |
599 | 0 | return ABIArgInfo::getIgnore(); |
600 | | |
601 | | // Check for homogeneous aggregates with AAPCS-VFP. |
602 | 0 | if (IsAAPCS_VFP) { |
603 | 0 | const Type *Base = nullptr; |
604 | 0 | uint64_t Members = 0; |
605 | 0 | if (isHomogeneousAggregate(RetTy, Base, Members)) |
606 | 0 | return classifyHomogeneousAggregate(RetTy, Base, Members); |
607 | 0 | } |
608 | | |
609 | | // Aggregates <= 4 bytes are returned in r0; other aggregates |
610 | | // are returned indirectly. |
611 | 0 | uint64_t Size = getContext().getTypeSize(RetTy); |
612 | 0 | if (Size <= 32) { |
613 | | // On RenderScript, coerce Aggregates <= 4 bytes to an integer array of |
614 | | // same size and alignment. |
615 | 0 | if (getTarget().isRenderScriptTarget()) { |
616 | 0 | return coerceToIntArray(RetTy, getContext(), getVMContext()); |
617 | 0 | } |
618 | 0 | if (getDataLayout().isBigEndian()) |
619 | | // Return in 32 bit integer integer type (as if loaded by LDR, AAPCS 5.4) |
620 | 0 | return ABIArgInfo::getDirect(llvm::Type::getInt32Ty(getVMContext())); |
621 | | |
622 | | // Return in the smallest viable integer type. |
623 | 0 | if (Size <= 8) |
624 | 0 | return ABIArgInfo::getDirect(llvm::Type::getInt8Ty(getVMContext())); |
625 | 0 | if (Size <= 16) |
626 | 0 | return ABIArgInfo::getDirect(llvm::Type::getInt16Ty(getVMContext())); |
627 | 0 | return ABIArgInfo::getDirect(llvm::Type::getInt32Ty(getVMContext())); |
628 | 0 | } else if (Size <= 128 && getABIKind() == ARMABIKind::AAPCS16_VFP) { |
629 | 0 | llvm::Type *Int32Ty = llvm::Type::getInt32Ty(getVMContext()); |
630 | 0 | llvm::Type *CoerceTy = |
631 | 0 | llvm::ArrayType::get(Int32Ty, llvm::alignTo(Size, 32) / 32); |
632 | 0 | return ABIArgInfo::getDirect(CoerceTy); |
633 | 0 | } |
634 | | |
635 | 0 | return getNaturalAlignIndirect(RetTy); |
636 | 0 | } |
637 | | |
638 | | /// isIllegalVector - check whether Ty is an illegal vector type. |
639 | 0 | bool ARMABIInfo::isIllegalVectorType(QualType Ty) const { |
640 | 0 | if (const VectorType *VT = Ty->getAs<VectorType> ()) { |
641 | | // On targets that don't support half, fp16 or bfloat, they are expanded |
642 | | // into float, and we don't want the ABI to depend on whether or not they |
643 | | // are supported in hardware. Thus return false to coerce vectors of these |
644 | | // types into integer vectors. |
645 | | // We do not depend on hasLegalHalfType for bfloat as it is a |
646 | | // separate IR type. |
647 | 0 | if ((!getTarget().hasLegalHalfType() && |
648 | 0 | (VT->getElementType()->isFloat16Type() || |
649 | 0 | VT->getElementType()->isHalfType())) || |
650 | 0 | (IsFloatABISoftFP && |
651 | 0 | VT->getElementType()->isBFloat16Type())) |
652 | 0 | return true; |
653 | 0 | if (isAndroid()) { |
654 | | // Android shipped using Clang 3.1, which supported a slightly different |
655 | | // vector ABI. The primary differences were that 3-element vector types |
656 | | // were legal, and so were sub 32-bit vectors (i.e. <2 x i8>). This path |
657 | | // accepts that legacy behavior for Android only. |
658 | | // Check whether VT is legal. |
659 | 0 | unsigned NumElements = VT->getNumElements(); |
660 | | // NumElements should be power of 2 or equal to 3. |
661 | 0 | if (!llvm::isPowerOf2_32(NumElements) && NumElements != 3) |
662 | 0 | return true; |
663 | 0 | } else { |
664 | | // Check whether VT is legal. |
665 | 0 | unsigned NumElements = VT->getNumElements(); |
666 | 0 | uint64_t Size = getContext().getTypeSize(VT); |
667 | | // NumElements should be power of 2. |
668 | 0 | if (!llvm::isPowerOf2_32(NumElements)) |
669 | 0 | return true; |
670 | | // Size should be greater than 32 bits. |
671 | 0 | return Size <= 32; |
672 | 0 | } |
673 | 0 | } |
674 | 0 | return false; |
675 | 0 | } |
676 | | |
677 | | /// Return true if a type contains any 16-bit floating point vectors |
678 | 0 | bool ARMABIInfo::containsAnyFP16Vectors(QualType Ty) const { |
679 | 0 | if (const ConstantArrayType *AT = getContext().getAsConstantArrayType(Ty)) { |
680 | 0 | uint64_t NElements = AT->getSize().getZExtValue(); |
681 | 0 | if (NElements == 0) |
682 | 0 | return false; |
683 | 0 | return containsAnyFP16Vectors(AT->getElementType()); |
684 | 0 | } else if (const RecordType *RT = Ty->getAs<RecordType>()) { |
685 | 0 | const RecordDecl *RD = RT->getDecl(); |
686 | | |
687 | | // If this is a C++ record, check the bases first. |
688 | 0 | if (const CXXRecordDecl *CXXRD = dyn_cast<CXXRecordDecl>(RD)) |
689 | 0 | if (llvm::any_of(CXXRD->bases(), [this](const CXXBaseSpecifier &B) { |
690 | 0 | return containsAnyFP16Vectors(B.getType()); |
691 | 0 | })) |
692 | 0 | return true; |
693 | | |
694 | 0 | if (llvm::any_of(RD->fields(), [this](FieldDecl *FD) { |
695 | 0 | return FD && containsAnyFP16Vectors(FD->getType()); |
696 | 0 | })) |
697 | 0 | return true; |
698 | | |
699 | 0 | return false; |
700 | 0 | } else { |
701 | 0 | if (const VectorType *VT = Ty->getAs<VectorType>()) |
702 | 0 | return (VT->getElementType()->isFloat16Type() || |
703 | 0 | VT->getElementType()->isBFloat16Type() || |
704 | 0 | VT->getElementType()->isHalfType()); |
705 | 0 | return false; |
706 | 0 | } |
707 | 0 | } |
708 | | |
709 | | bool ARMSwiftABIInfo::isLegalVectorType(CharUnits VectorSize, llvm::Type *EltTy, |
710 | 0 | unsigned NumElts) const { |
711 | 0 | if (!llvm::isPowerOf2_32(NumElts)) |
712 | 0 | return false; |
713 | 0 | unsigned size = CGT.getDataLayout().getTypeStoreSizeInBits(EltTy); |
714 | 0 | if (size > 64) |
715 | 0 | return false; |
716 | 0 | if (VectorSize.getQuantity() != 8 && |
717 | 0 | (VectorSize.getQuantity() != 16 || NumElts == 1)) |
718 | 0 | return false; |
719 | 0 | return true; |
720 | 0 | } |
721 | | |
722 | 0 | bool ARMABIInfo::isHomogeneousAggregateBaseType(QualType Ty) const { |
723 | | // Homogeneous aggregates for AAPCS-VFP must have base types of float, |
724 | | // double, or 64-bit or 128-bit vectors. |
725 | 0 | if (const BuiltinType *BT = Ty->getAs<BuiltinType>()) { |
726 | 0 | if (BT->getKind() == BuiltinType::Float || |
727 | 0 | BT->getKind() == BuiltinType::Double || |
728 | 0 | BT->getKind() == BuiltinType::LongDouble) |
729 | 0 | return true; |
730 | 0 | } else if (const VectorType *VT = Ty->getAs<VectorType>()) { |
731 | 0 | unsigned VecSize = getContext().getTypeSize(VT); |
732 | 0 | if (VecSize == 64 || VecSize == 128) |
733 | 0 | return true; |
734 | 0 | } |
735 | 0 | return false; |
736 | 0 | } |
737 | | |
738 | | bool ARMABIInfo::isHomogeneousAggregateSmallEnough(const Type *Base, |
739 | 0 | uint64_t Members) const { |
740 | 0 | return Members <= 4; |
741 | 0 | } |
742 | | |
743 | 0 | bool ARMABIInfo::isZeroLengthBitfieldPermittedInHomogeneousAggregate() const { |
744 | | // AAPCS32 says that the rule for whether something is a homogeneous |
745 | | // aggregate is applied to the output of the data layout decision. So |
746 | | // anything that doesn't affect the data layout also does not affect |
747 | | // homogeneity. In particular, zero-length bitfields don't stop a struct |
748 | | // being homogeneous. |
749 | 0 | return true; |
750 | 0 | } |
751 | | |
752 | | bool ARMABIInfo::isEffectivelyAAPCS_VFP(unsigned callConvention, |
753 | 0 | bool acceptHalf) const { |
754 | | // Give precedence to user-specified calling conventions. |
755 | 0 | if (callConvention != llvm::CallingConv::C) |
756 | 0 | return (callConvention == llvm::CallingConv::ARM_AAPCS_VFP); |
757 | 0 | else |
758 | 0 | return (getABIKind() == ARMABIKind::AAPCS_VFP) || |
759 | 0 | (acceptHalf && (getABIKind() == ARMABIKind::AAPCS16_VFP)); |
760 | 0 | } |
761 | | |
762 | | Address ARMABIInfo::EmitVAArg(CodeGenFunction &CGF, Address VAListAddr, |
763 | 0 | QualType Ty) const { |
764 | 0 | CharUnits SlotSize = CharUnits::fromQuantity(4); |
765 | | |
766 | | // Empty records are ignored for parameter passing purposes. |
767 | 0 | if (isEmptyRecord(getContext(), Ty, true)) { |
768 | 0 | VAListAddr = VAListAddr.withElementType(CGF.Int8PtrTy); |
769 | 0 | auto *Load = CGF.Builder.CreateLoad(VAListAddr); |
770 | 0 | return Address(Load, CGF.ConvertTypeForMem(Ty), SlotSize); |
771 | 0 | } |
772 | | |
773 | 0 | CharUnits TySize = getContext().getTypeSizeInChars(Ty); |
774 | 0 | CharUnits TyAlignForABI = getContext().getTypeUnadjustedAlignInChars(Ty); |
775 | | |
776 | | // Use indirect if size of the illegal vector is bigger than 16 bytes. |
777 | 0 | bool IsIndirect = false; |
778 | 0 | const Type *Base = nullptr; |
779 | 0 | uint64_t Members = 0; |
780 | 0 | if (TySize > CharUnits::fromQuantity(16) && isIllegalVectorType(Ty)) { |
781 | 0 | IsIndirect = true; |
782 | | |
783 | | // ARMv7k passes structs bigger than 16 bytes indirectly, in space |
784 | | // allocated by the caller. |
785 | 0 | } else if (TySize > CharUnits::fromQuantity(16) && |
786 | 0 | getABIKind() == ARMABIKind::AAPCS16_VFP && |
787 | 0 | !isHomogeneousAggregate(Ty, Base, Members)) { |
788 | 0 | IsIndirect = true; |
789 | | |
790 | | // Otherwise, bound the type's ABI alignment. |
791 | | // The ABI alignment for 64-bit or 128-bit vectors is 8 for AAPCS and 4 for |
792 | | // APCS. For AAPCS, the ABI alignment is at least 4-byte and at most 8-byte. |
793 | | // Our callers should be prepared to handle an under-aligned address. |
794 | 0 | } else if (getABIKind() == ARMABIKind::AAPCS_VFP || |
795 | 0 | getABIKind() == ARMABIKind::AAPCS) { |
796 | 0 | TyAlignForABI = std::max(TyAlignForABI, CharUnits::fromQuantity(4)); |
797 | 0 | TyAlignForABI = std::min(TyAlignForABI, CharUnits::fromQuantity(8)); |
798 | 0 | } else if (getABIKind() == ARMABIKind::AAPCS16_VFP) { |
799 | | // ARMv7k allows type alignment up to 16 bytes. |
800 | 0 | TyAlignForABI = std::max(TyAlignForABI, CharUnits::fromQuantity(4)); |
801 | 0 | TyAlignForABI = std::min(TyAlignForABI, CharUnits::fromQuantity(16)); |
802 | 0 | } else { |
803 | 0 | TyAlignForABI = CharUnits::fromQuantity(4); |
804 | 0 | } |
805 | |
|
806 | 0 | TypeInfoChars TyInfo(TySize, TyAlignForABI, AlignRequirementKind::None); |
807 | 0 | return emitVoidPtrVAArg(CGF, VAListAddr, Ty, IsIndirect, TyInfo, |
808 | 0 | SlotSize, /*AllowHigherAlign*/ true); |
809 | 0 | } |
810 | | |
811 | | std::unique_ptr<TargetCodeGenInfo> |
812 | 0 | CodeGen::createARMTargetCodeGenInfo(CodeGenModule &CGM, ARMABIKind Kind) { |
813 | 0 | return std::make_unique<ARMTargetCodeGenInfo>(CGM.getTypes(), Kind); |
814 | 0 | } |
815 | | |
816 | | std::unique_ptr<TargetCodeGenInfo> |
817 | 0 | CodeGen::createWindowsARMTargetCodeGenInfo(CodeGenModule &CGM, ARMABIKind K) { |
818 | 0 | return std::make_unique<WindowsARMTargetCodeGenInfo>(CGM.getTypes(), K); |
819 | 0 | } |