| 1 | #include "TargetInfo.h" |
| 2 | #include "ABIInfo.h" |
| 3 | #include "CIRGenFunction.h" |
| 4 | #include "CIRGenModule.h" |
| 5 | #include "mlir/Dialect/Ptr/IR/MemorySpaceInterfaces.h" |
| 6 | #include "clang/Basic/AddressSpaces.h" |
| 7 | #include "clang/CIR/Dialect/IR/CIRAttrs.h" |
| 8 | #include "clang/CIR/Dialect/IR/CIRDialect.h" |
| 9 | #include "clang/CIR/MissingFeatures.h" |
| 10 | #include "clang/CodeGenUtils/TargetUtils.h" |
| 11 | |
| 12 | using namespace clang; |
| 13 | using namespace clang::CIRGen; |
| 14 | |
| 15 | bool clang::CIRGen::isEmptyRecordForABI(const ASTContext &context, QualType t) { |
| 16 | const auto *rd = t->getAsRecordDecl(); |
| 17 | if (!rd) |
| 18 | return false; |
| 19 | if (rd->hasFlexibleArrayMember()) |
| 20 | return false; |
| 21 | |
| 22 | if (const auto *cxxrd = dyn_cast<CXXRecordDecl>(Val: rd)) { |
| 23 | // A vtable pointer is neither a base nor a field, so clang's predicate |
| 24 | // calls a polymorphic class empty and leans on its callers rejecting one as |
| 25 | // non-trivially-copyable beforehand. This answer is read off the record |
| 26 | // type without that precondition, so rule it out here instead. |
| 27 | if (cxxrd->isDynamicClass()) |
| 28 | return false; |
| 29 | |
| 30 | for (const auto &i : cxxrd->bases()) |
| 31 | if (!isEmptyRecordForABI(context, t: i.getType())) |
| 32 | return false; |
| 33 | } |
| 34 | |
| 35 | for (const auto *i : rd->fields()) |
| 36 | if (!isEmptyFieldForABI(context, fd: i)) |
| 37 | return false; |
| 38 | return true; |
| 39 | } |
| 40 | |
| 41 | bool clang::CIRGen::isEmptyFieldForABI(const ASTContext &context, |
| 42 | const FieldDecl *fd) { |
| 43 | // A zero-width bit-field holds no bits, so it holds no data. Any other |
| 44 | // unnamed bit-field is real storage that the x86-64 classifier gives the |
| 45 | // eightbyte classes of a named bit-field, so it does hold data. |
| 46 | // |
| 47 | // FIXME: this needs an ABI compatibility check. Clang used to ignore every |
| 48 | // unnamed bit-field here, and the older compatibility levels have to keep |
| 49 | // that answer once there is a switch to ask. |
| 50 | if (fd->isUnnamedBitField()) |
| 51 | return fd->isZeroLengthBitField(); |
| 52 | |
| 53 | QualType ft = fd->getType(); |
| 54 | |
| 55 | // An array of empty records is empty, and a zero-length array always is. |
| 56 | bool wasArray = false; |
| 57 | while (const ConstantArrayType *at = context.getAsConstantArrayType(T: ft)) { |
| 58 | if (at->isZeroSize()) |
| 59 | return true; |
| 60 | ft = at->getElementType(); |
| 61 | wasArray = true; |
| 62 | } |
| 63 | |
| 64 | const auto *rt = ft->getAsCanonical<RecordType>(); |
| 65 | if (!rt) |
| 66 | return false; |
| 67 | |
| 68 | // A C++ record field is never empty under the Itanium ABI unless |
| 69 | // [[no_unique_address]] makes it so, and that exception covers a record |
| 70 | // rather than an array of them. |
| 71 | if (isa<CXXRecordDecl>(Val: rt->getDecl()) && |
| 72 | (wasArray || !fd->hasAttr<NoUniqueAddressAttr>())) |
| 73 | return false; |
| 74 | |
| 75 | return isEmptyRecordForABI(context, t: ft); |
| 76 | } |
| 77 | |
| 78 | namespace { |
| 79 | |
| 80 | class AMDGPUABIInfo : public ABIInfo { |
| 81 | public: |
| 82 | AMDGPUABIInfo(CIRGenTypes &cgt) : ABIInfo(cgt) {} |
| 83 | }; |
| 84 | |
| 85 | class AMDGPUTargetCIRGenInfo : public TargetCIRGenInfo { |
| 86 | public: |
| 87 | AMDGPUTargetCIRGenInfo(CIRGenTypes &cgt) |
| 88 | : TargetCIRGenInfo(std::make_unique<AMDGPUABIInfo>(args&: cgt)) {} |
| 89 | |
| 90 | bool supportsLibCall() const override { return false; } |
| 91 | |
| 92 | cir::CallingConv getDeviceKernelCallingConv() const override { |
| 93 | return cir::CallingConv::AMDGPUKernel; |
| 94 | } |
| 95 | |
| 96 | void setCUDAKernelCallingConvention(const FunctionType *&ft) const override { |
| 97 | ft = getABIInfo().cgt.getASTContext().adjustFunctionType( |
| 98 | Fn: ft, EInfo: ft->getExtInfo().withCallingConv(cc: CC_DeviceKernel)); |
| 99 | } |
| 100 | |
| 101 | void setTargetAttributes(const clang::Decl *decl, mlir::Operation *global, |
| 102 | CIRGenModule &cgm) const override { |
| 103 | if (auto func = mlir::dyn_cast<cir::FuncOp>(Val: global)) { |
| 104 | if (CodeGenUtils::requiresAMDGPUProtectedVisibility( |
| 105 | D: decl, |
| 106 | HasHiddenVisibility: func.getGlobalVisibility() == cir::VisibilityKind::Hidden)) { |
| 107 | func.setGlobalVisibility(cir::VisibilityKind::Protected); |
| 108 | func.setDSOLocal(true); |
| 109 | } |
| 110 | setAMDGPUTargetFunctionAttributes(decl, func, cgm); |
| 111 | } else if (auto gv = mlir::dyn_cast<cir::GlobalOp>(Val: global)) { |
| 112 | if (CodeGenUtils::requiresAMDGPUProtectedVisibility( |
| 113 | D: decl, HasHiddenVisibility: gv.getGlobalVisibility() == cir::VisibilityKind::Hidden)) { |
| 114 | gv.setGlobalVisibility(cir::VisibilityKind::Protected); |
| 115 | gv.setDSOLocal(true); |
| 116 | } |
| 117 | } |
| 118 | } |
| 119 | |
| 120 | clang::LangAS |
| 121 | getGlobalVarAddressSpace(CIRGenModule &cgm, |
| 122 | const clang::VarDecl *decl) const override { |
| 123 | using clang::LangAS; |
| 124 | assert(!cgm.getLangOpts().OpenCL && |
| 125 | !(cgm.getLangOpts().CUDA && cgm.getLangOpts().CUDAIsDevice) && |
| 126 | "Address space agnostic languages only" ); |
| 127 | LangAS defaultGlobalAS = LangAS::opencl_global; |
| 128 | if (!decl) |
| 129 | return defaultGlobalAS; |
| 130 | |
| 131 | LangAS addrSpace = decl->getType().getAddressSpace(); |
| 132 | if (addrSpace != LangAS::Default) |
| 133 | return addrSpace; |
| 134 | |
| 135 | // Only promote to address space 4 if VarDecl has constant initialization. |
| 136 | if (decl->getType().isConstantStorage(Ctx: cgm.getASTContext(), ExcludeCtor: false, ExcludeDtor: false) && |
| 137 | decl->hasConstantInitialization()) |
| 138 | return LangAS::opencl_constant; |
| 139 | |
| 140 | return defaultGlobalAS; |
| 141 | } |
| 142 | |
| 143 | mlir::ptr::MemorySpaceAttrInterface |
| 144 | getCIRAllocaAddressSpace() const override { |
| 145 | return cir::LangAddressSpaceAttr::get( |
| 146 | context: &getABIInfo().cgt.getMLIRContext(), |
| 147 | value: cir::LangAddressSpace::OffloadPrivate); |
| 148 | } |
| 149 | |
| 150 | // In some addrspaces for AMDGPU, null isn't 0 (such as local and private), |
| 151 | // thus we need an addrspacecast from Generic to emit correct null value for |
| 152 | // these addrspaces. |
| 153 | mlir::Value getNullPointer(CIRGenModule &cgm, cir::PointerType ptrTy, |
| 154 | QualType qt, mlir::Location loc) const override { |
| 155 | CIRGenBuilderTy &builder = cgm.getBuilder(); |
| 156 | if (cgm.getASTContext().getTargetNullPointerValue(QT: qt) == 0) |
| 157 | return builder.getNullPtr(ty: ptrTy, loc); |
| 158 | |
| 159 | cir::PointerType genericPtrTy = |
| 160 | builder.getPointerTo(ty: ptrTy.getPointee(), langAS: LangAS::opencl_generic); |
| 161 | return builder.createAddrSpaceCast( |
| 162 | loc, src: builder.getNullPtr(ty: genericPtrTy, loc), newTy: ptrTy); |
| 163 | } |
| 164 | }; |
| 165 | |
| 166 | } // namespace |
| 167 | |
| 168 | namespace { |
| 169 | |
| 170 | class X8664ABIInfo : public ABIInfo { |
| 171 | public: |
| 172 | X8664ABIInfo(CIRGenTypes &cgt) : ABIInfo(cgt) {} |
| 173 | }; |
| 174 | |
| 175 | class X8664TargetCIRGenInfo : public TargetCIRGenInfo { |
| 176 | public: |
| 177 | X8664TargetCIRGenInfo(CIRGenTypes &cgt) |
| 178 | : TargetCIRGenInfo(std::make_unique<X8664ABIInfo>(args&: cgt)) {} |
| 179 | }; |
| 180 | } // namespace |
| 181 | |
| 182 | std::unique_ptr<TargetCIRGenInfo> |
| 183 | clang::CIRGen::createAMDGPUTargetCIRGenInfo(CIRGenTypes &cgt) { |
| 184 | return std::make_unique<AMDGPUTargetCIRGenInfo>(args&: cgt); |
| 185 | } |
| 186 | |
| 187 | std::unique_ptr<TargetCIRGenInfo> |
| 188 | clang::CIRGen::createX8664TargetCIRGenInfo(CIRGenTypes &cgt) { |
| 189 | return std::make_unique<X8664TargetCIRGenInfo>(args&: cgt); |
| 190 | } |
| 191 | |
| 192 | ABIInfo::~ABIInfo() = default; |
| 193 | |
| 194 | bool TargetCIRGenInfo::isNoProtoCallVariadic( |
| 195 | const FunctionNoProtoType *fnType) const { |
| 196 | // The following conventions are known to require this to be false: |
| 197 | // x86_stdcall |
| 198 | // MIPS |
| 199 | // For everything else, we just prefer false unless we opt out. |
| 200 | return false; |
| 201 | } |
| 202 | |
| 203 | cir::CallingConv TargetCIRGenInfo::getDeviceKernelCallingConv() const { |
| 204 | // Device kernels are entered through a runtime API, not called as normal |
| 205 | // sub-functions, so a modified C calling convention is used. |
| 206 | assert(getABIInfo().cgt.getASTContext().getLangOpts().OpenCL && |
| 207 | "Kernel calling convention only defined for OpenCL" ); |
| 208 | return cir::CallingConv::C; |
| 209 | } |
| 210 | |
| 211 | mlir::Value TargetCIRGenInfo::getNullPointer(CIRGenModule &cgm, |
| 212 | cir::PointerType ptrTy, |
| 213 | QualType qt, |
| 214 | mlir::Location loc) const { |
| 215 | return cgm.getBuilder().getNullPtr(ty: ptrTy, loc); |
| 216 | } |
| 217 | |
| 218 | clang::LangAS |
| 219 | TargetCIRGenInfo::getGlobalVarAddressSpace(CIRGenModule &cgm, |
| 220 | const clang::VarDecl *d) const { |
| 221 | assert(!cgm.getLangOpts().OpenCL && |
| 222 | !(cgm.getLangOpts().CUDA && cgm.getLangOpts().CUDAIsDevice) && |
| 223 | "Address space agnostic languages only" ); |
| 224 | return d ? d->getType().getAddressSpace() : LangAS::Default; |
| 225 | } |
| 226 | |