| 1 | //===- AMDGPU.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 "llvm/ADT/StringExtras.h" |
| 12 | #include "llvm/Support/AMDGPUAddrSpace.h" |
| 13 | |
| 14 | using namespace clang; |
| 15 | using namespace clang::CodeGen; |
| 16 | |
| 17 | //===----------------------------------------------------------------------===// |
| 18 | // AMDGPU ABI Implementation |
| 19 | //===----------------------------------------------------------------------===// |
| 20 | |
| 21 | namespace { |
| 22 | |
| 23 | class AMDGPUABIInfo final : public DefaultABIInfo { |
| 24 | private: |
| 25 | static const unsigned MaxNumRegsForArgsRet = 16; |
| 26 | |
| 27 | uint64_t numRegsForType(QualType Ty) const; |
| 28 | |
| 29 | bool isHomogeneousAggregateBaseType(QualType Ty) const override; |
| 30 | bool isHomogeneousAggregateSmallEnough(const Type *Base, |
| 31 | uint64_t Members) const override; |
| 32 | |
| 33 | // Coerce HIP scalar pointer arguments from generic pointers to global ones. |
| 34 | llvm::Type *coerceKernelArgumentType(llvm::Type *Ty, unsigned FromAS, |
| 35 | unsigned ToAS) const { |
| 36 | // Single value types. |
| 37 | auto *PtrTy = llvm::dyn_cast<llvm::PointerType>(Val: Ty); |
| 38 | if (PtrTy && PtrTy->getAddressSpace() == FromAS) |
| 39 | return llvm::PointerType::get(C&: Ty->getContext(), AddressSpace: ToAS); |
| 40 | return Ty; |
| 41 | } |
| 42 | |
| 43 | public: |
| 44 | explicit AMDGPUABIInfo(CodeGen::CodeGenTypes &CGT) : |
| 45 | DefaultABIInfo(CGT) {} |
| 46 | |
| 47 | ABIArgInfo classifyReturnType(QualType RetTy) const; |
| 48 | ABIArgInfo classifyKernelArgumentType(QualType Ty) const; |
| 49 | ABIArgInfo classifyArgumentType(QualType Ty, bool Variadic, |
| 50 | unsigned &NumRegsLeft) const; |
| 51 | |
| 52 | void computeInfo(CGFunctionInfo &FI) const override; |
| 53 | RValue EmitVAArg(CodeGenFunction &CGF, Address VAListAddr, QualType Ty, |
| 54 | AggValueSlot Slot) const override; |
| 55 | |
| 56 | llvm::FixedVectorType * |
| 57 | getOptimalVectorMemoryType(llvm::FixedVectorType *T, |
| 58 | const LangOptions &Opt) const override { |
| 59 | // We have legal instructions for 96-bit so 3x32 can be supported. |
| 60 | // FIXME: This check should be a subtarget feature as technically SI doesn't |
| 61 | // support it. |
| 62 | if (T->getNumElements() == 3 && getDataLayout().getTypeSizeInBits(Ty: T) == 96) |
| 63 | return T; |
| 64 | return DefaultABIInfo::getOptimalVectorMemoryType(T, Opt); |
| 65 | } |
| 66 | }; |
| 67 | |
| 68 | bool AMDGPUABIInfo::isHomogeneousAggregateBaseType(QualType Ty) const { |
| 69 | return true; |
| 70 | } |
| 71 | |
| 72 | bool AMDGPUABIInfo::isHomogeneousAggregateSmallEnough( |
| 73 | const Type *Base, uint64_t Members) const { |
| 74 | uint32_t NumRegs = (getContext().getTypeSize(T: Base) + 31) / 32; |
| 75 | |
| 76 | // Homogeneous Aggregates may occupy at most 16 registers. |
| 77 | return Members * NumRegs <= MaxNumRegsForArgsRet; |
| 78 | } |
| 79 | |
| 80 | /// Estimate number of registers the type will use when passed in registers. |
| 81 | uint64_t AMDGPUABIInfo::numRegsForType(QualType Ty) const { |
| 82 | uint64_t NumRegs = 0; |
| 83 | |
| 84 | if (const VectorType *VT = Ty->getAs<VectorType>()) { |
| 85 | // Compute from the number of elements. The reported size is based on the |
| 86 | // in-memory size, which includes the padding 4th element for 3-vectors. |
| 87 | QualType EltTy = VT->getElementType(); |
| 88 | uint64_t EltSize = getContext().getTypeSize(T: EltTy); |
| 89 | |
| 90 | // 16-bit element vectors should be passed as packed. |
| 91 | if (EltSize == 16) |
| 92 | return (VT->getNumElements() + 1) / 2; |
| 93 | |
| 94 | uint64_t EltNumRegs = (EltSize + 31) / 32; |
| 95 | return EltNumRegs * VT->getNumElements(); |
| 96 | } |
| 97 | |
| 98 | if (const auto *RD = Ty->getAsRecordDecl()) { |
| 99 | assert(!RD->hasFlexibleArrayMember()); |
| 100 | |
| 101 | for (const FieldDecl *Field : RD->fields()) { |
| 102 | QualType FieldTy = Field->getType(); |
| 103 | NumRegs += numRegsForType(Ty: FieldTy); |
| 104 | } |
| 105 | |
| 106 | return NumRegs; |
| 107 | } |
| 108 | |
| 109 | return (getContext().getTypeSize(T: Ty) + 31) / 32; |
| 110 | } |
| 111 | |
| 112 | void AMDGPUABIInfo::computeInfo(CGFunctionInfo &FI) const { |
| 113 | llvm::CallingConv::ID CC = FI.getCallingConvention(); |
| 114 | |
| 115 | if (!getCXXABI().classifyReturnType(FI)) |
| 116 | FI.getReturnInfo() = classifyReturnType(RetTy: FI.getReturnType()); |
| 117 | |
| 118 | unsigned ArgumentIndex = 0; |
| 119 | const unsigned numFixedArguments = FI.getNumRequiredArgs(); |
| 120 | |
| 121 | unsigned NumRegsLeft = MaxNumRegsForArgsRet; |
| 122 | for (auto &Arg : FI.arguments()) { |
| 123 | if (CC == llvm::CallingConv::AMDGPU_KERNEL) { |
| 124 | Arg.info = classifyKernelArgumentType(Ty: Arg.type); |
| 125 | } else { |
| 126 | bool FixedArgument = ArgumentIndex++ < numFixedArguments; |
| 127 | Arg.info = classifyArgumentType(Ty: Arg.type, Variadic: !FixedArgument, NumRegsLeft); |
| 128 | } |
| 129 | } |
| 130 | } |
| 131 | |
| 132 | RValue AMDGPUABIInfo::EmitVAArg(CodeGenFunction &CGF, Address VAListAddr, |
| 133 | QualType Ty, AggValueSlot Slot) const { |
| 134 | const bool IsIndirect = false; |
| 135 | const bool AllowHigherAlign = false; |
| 136 | return emitVoidPtrVAArg(CGF, VAListAddr, ValueTy: Ty, IsIndirect, |
| 137 | ValueInfo: getContext().getTypeInfoInChars(T: Ty), |
| 138 | SlotSizeAndAlign: CharUnits::fromQuantity(Quantity: 4), AllowHigherAlign, Slot); |
| 139 | } |
| 140 | |
| 141 | ABIArgInfo AMDGPUABIInfo::classifyReturnType(QualType RetTy) const { |
| 142 | if (isAggregateTypeForABI(T: RetTy)) { |
| 143 | // Records with non-trivial destructors/copy-constructors should not be |
| 144 | // returned by value. |
| 145 | if (!getRecordArgABI(T: RetTy, CXXABI&: getCXXABI())) { |
| 146 | // Ignore empty structs/unions. |
| 147 | if (isEmptyRecord(Context&: getContext(), T: RetTy, AllowArrays: true)) |
| 148 | return ABIArgInfo::getIgnore(); |
| 149 | |
| 150 | // Lower single-element structs to just return a regular value. |
| 151 | if (const Type *SeltTy = isSingleElementStruct(T: RetTy, Context&: getContext())) |
| 152 | return ABIArgInfo::getDirect(T: CGT.ConvertType(T: QualType(SeltTy, 0))); |
| 153 | |
| 154 | if (const auto *RD = RetTy->getAsRecordDecl(); |
| 155 | RD && RD->hasFlexibleArrayMember()) |
| 156 | return DefaultABIInfo::classifyReturnType(RetTy); |
| 157 | |
| 158 | // Pack aggregates <= 4 bytes into single VGPR or pair. |
| 159 | uint64_t Size = getContext().getTypeSize(T: RetTy); |
| 160 | if (Size <= 16) |
| 161 | return ABIArgInfo::getDirect(T: llvm::Type::getInt16Ty(C&: getVMContext())); |
| 162 | |
| 163 | if (Size <= 32) |
| 164 | return ABIArgInfo::getDirect(T: llvm::Type::getInt32Ty(C&: getVMContext())); |
| 165 | |
| 166 | if (Size <= 64) { |
| 167 | llvm::Type *I32Ty = llvm::Type::getInt32Ty(C&: getVMContext()); |
| 168 | return ABIArgInfo::getDirect(T: llvm::ArrayType::get(ElementType: I32Ty, NumElements: 2)); |
| 169 | } |
| 170 | |
| 171 | if (numRegsForType(Ty: RetTy) <= MaxNumRegsForArgsRet) |
| 172 | return ABIArgInfo::getDirect(); |
| 173 | } |
| 174 | } |
| 175 | |
| 176 | // Otherwise just do the default thing. |
| 177 | return DefaultABIInfo::classifyReturnType(RetTy); |
| 178 | } |
| 179 | |
| 180 | /// For kernels all parameters are really passed in a special buffer. It doesn't |
| 181 | /// make sense to pass anything byval, so everything must be direct. |
| 182 | ABIArgInfo AMDGPUABIInfo::classifyKernelArgumentType(QualType Ty) const { |
| 183 | Ty = useFirstFieldIfTransparentUnion(Ty); |
| 184 | |
| 185 | // TODO: Can we omit empty structs? |
| 186 | |
| 187 | if (const Type *SeltTy = isSingleElementStruct(T: Ty, Context&: getContext())) |
| 188 | Ty = QualType(SeltTy, 0); |
| 189 | |
| 190 | llvm::Type *OrigLTy = CGT.ConvertType(T: Ty); |
| 191 | llvm::Type *LTy = OrigLTy; |
| 192 | if (getContext().getLangOpts().HIP) { |
| 193 | LTy = coerceKernelArgumentType( |
| 194 | Ty: OrigLTy, /*FromAS=*/getContext().getTargetAddressSpace(AS: LangAS::Default), |
| 195 | /*ToAS=*/getContext().getTargetAddressSpace(AS: LangAS::cuda_device)); |
| 196 | } |
| 197 | |
| 198 | // FIXME: This doesn't apply the optimization of coercing pointers in structs |
| 199 | // to global address space when using byref. This would require implementing a |
| 200 | // new kind of coercion of the in-memory type when for indirect arguments. |
| 201 | if (LTy == OrigLTy && isAggregateTypeForABI(T: Ty)) { |
| 202 | return ABIArgInfo::getIndirectAliased( |
| 203 | Alignment: getContext().getTypeAlignInChars(T: Ty), |
| 204 | AddrSpace: getContext().getTargetAddressSpace(AS: LangAS::opencl_constant), |
| 205 | Realign: false /*Realign*/, Padding: nullptr /*Padding*/); |
| 206 | } |
| 207 | |
| 208 | // If we set CanBeFlattened to true, CodeGen will expand the struct to its |
| 209 | // individual elements, which confuses the Clover OpenCL backend; therefore we |
| 210 | // have to set it to false here. Other args of getDirect() are just defaults. |
| 211 | return ABIArgInfo::getDirect(T: LTy, Offset: 0, Padding: nullptr, CanBeFlattened: false); |
| 212 | } |
| 213 | |
| 214 | ABIArgInfo AMDGPUABIInfo::classifyArgumentType(QualType Ty, bool Variadic, |
| 215 | unsigned &NumRegsLeft) const { |
| 216 | assert(NumRegsLeft <= MaxNumRegsForArgsRet && "register estimate underflow" ); |
| 217 | |
| 218 | Ty = useFirstFieldIfTransparentUnion(Ty); |
| 219 | |
| 220 | if (Variadic) { |
| 221 | return ABIArgInfo::getDirect(/*T=*/nullptr, |
| 222 | /*Offset=*/0, |
| 223 | /*Padding=*/nullptr, |
| 224 | /*CanBeFlattened=*/false, |
| 225 | /*Align=*/0); |
| 226 | } |
| 227 | |
| 228 | if (isAggregateTypeForABI(T: Ty)) { |
| 229 | // Records with non-trivial destructors/copy-constructors should not be |
| 230 | // passed by value. |
| 231 | if (auto RAA = getRecordArgABI(T: Ty, CXXABI&: getCXXABI())) |
| 232 | return getNaturalAlignIndirect(Ty, AddrSpace: getDataLayout().getAllocaAddrSpace(), |
| 233 | ByVal: RAA == CGCXXABI::RAA_DirectInMemory); |
| 234 | |
| 235 | // Ignore empty structs/unions. |
| 236 | if (isEmptyRecord(Context&: getContext(), T: Ty, AllowArrays: true)) |
| 237 | return ABIArgInfo::getIgnore(); |
| 238 | |
| 239 | // Lower single-element structs to just pass a regular value. TODO: We |
| 240 | // could do reasonable-size multiple-element structs too, using getExpand(), |
| 241 | // though watch out for things like bitfields. |
| 242 | if (const Type *SeltTy = isSingleElementStruct(T: Ty, Context&: getContext())) |
| 243 | return ABIArgInfo::getDirect(T: CGT.ConvertType(T: QualType(SeltTy, 0))); |
| 244 | |
| 245 | if (const auto *RD = Ty->getAsRecordDecl(); |
| 246 | RD && RD->hasFlexibleArrayMember()) |
| 247 | return DefaultABIInfo::classifyArgumentType(RetTy: Ty); |
| 248 | |
| 249 | // Pack aggregates <= 8 bytes into single VGPR or pair. |
| 250 | uint64_t Size = getContext().getTypeSize(T: Ty); |
| 251 | if (Size <= 64) { |
| 252 | unsigned NumRegs = (Size + 31) / 32; |
| 253 | NumRegsLeft -= std::min(a: NumRegsLeft, b: NumRegs); |
| 254 | |
| 255 | if (Size <= 16) |
| 256 | return ABIArgInfo::getDirect(T: llvm::Type::getInt16Ty(C&: getVMContext())); |
| 257 | |
| 258 | if (Size <= 32) |
| 259 | return ABIArgInfo::getDirect(T: llvm::Type::getInt32Ty(C&: getVMContext())); |
| 260 | |
| 261 | // XXX: Should this be i64 instead, and should the limit increase? |
| 262 | llvm::Type *I32Ty = llvm::Type::getInt32Ty(C&: getVMContext()); |
| 263 | return ABIArgInfo::getDirect(T: llvm::ArrayType::get(ElementType: I32Ty, NumElements: 2)); |
| 264 | } |
| 265 | |
| 266 | if (NumRegsLeft > 0) { |
| 267 | uint64_t NumRegs = numRegsForType(Ty); |
| 268 | if (NumRegsLeft >= NumRegs) { |
| 269 | NumRegsLeft -= NumRegs; |
| 270 | return ABIArgInfo::getDirect(); |
| 271 | } |
| 272 | } |
| 273 | |
| 274 | // Use pass-by-reference in stead of pass-by-value for struct arguments in |
| 275 | // function ABI. |
| 276 | return ABIArgInfo::getIndirectAliased( |
| 277 | Alignment: getContext().getTypeAlignInChars(T: Ty), |
| 278 | AddrSpace: getContext().getTargetAddressSpace(AS: LangAS::opencl_private)); |
| 279 | } |
| 280 | |
| 281 | // Otherwise just do the default thing. |
| 282 | ABIArgInfo ArgInfo = DefaultABIInfo::classifyArgumentType(RetTy: Ty); |
| 283 | if (!ArgInfo.isIndirect()) { |
| 284 | uint64_t NumRegs = numRegsForType(Ty); |
| 285 | NumRegsLeft -= std::min(a: NumRegs, b: uint64_t{NumRegsLeft}); |
| 286 | } |
| 287 | |
| 288 | return ArgInfo; |
| 289 | } |
| 290 | |
| 291 | class AMDGPUTargetCodeGenInfo : public TargetCodeGenInfo { |
| 292 | public: |
| 293 | AMDGPUTargetCodeGenInfo(CodeGenTypes &CGT) |
| 294 | : TargetCodeGenInfo(std::make_unique<AMDGPUABIInfo>(args&: CGT)) {} |
| 295 | |
| 296 | bool supportsLibCall() const override { return false; } |
| 297 | void setFunctionDeclAttributes(const FunctionDecl *FD, llvm::Function *F, |
| 298 | CodeGenModule &CGM) const; |
| 299 | |
| 300 | void setTargetAttributes(const Decl *D, llvm::GlobalValue *GV, |
| 301 | CodeGen::CodeGenModule &M) const override; |
| 302 | unsigned getDeviceKernelCallingConv() const override; |
| 303 | |
| 304 | llvm::Constant *getNullPointer(const CodeGen::CodeGenModule &CGM, |
| 305 | llvm::PointerType *T, QualType QT) const override; |
| 306 | |
| 307 | LangAS getASTAllocaAddressSpace() const override { |
| 308 | return getLangASFromTargetAS( |
| 309 | TargetAS: getABIInfo().getDataLayout().getAllocaAddrSpace()); |
| 310 | } |
| 311 | LangAS getGlobalVarAddressSpace(CodeGenModule &CGM, |
| 312 | const VarDecl *D) const override; |
| 313 | llvm::SyncScope::ID getLLVMSyncScopeID(const LangOptions &LangOpts, |
| 314 | SyncScope Scope, |
| 315 | llvm::AtomicOrdering Ordering, |
| 316 | llvm::LLVMContext &Ctx) const override; |
| 317 | void setTargetAtomicMetadata(CodeGenFunction &CGF, |
| 318 | llvm::Instruction &AtomicInst, |
| 319 | const AtomicExpr *Expr = nullptr) const override; |
| 320 | llvm::Value *createEnqueuedBlockKernel(CodeGenFunction &CGF, |
| 321 | llvm::Function *BlockInvokeFunc, |
| 322 | llvm::Type *BlockTy) const override; |
| 323 | bool shouldEmitStaticExternCAliases() const override; |
| 324 | bool shouldEmitDWARFBitFieldSeparators() const override; |
| 325 | void setCUDAKernelCallingConvention(const FunctionType *&FT) const override; |
| 326 | }; |
| 327 | } |
| 328 | |
| 329 | static bool requiresAMDGPUProtectedVisibility(const Decl *D, |
| 330 | llvm::GlobalValue *GV) { |
| 331 | if (GV->getVisibility() != llvm::GlobalValue::HiddenVisibility) |
| 332 | return false; |
| 333 | |
| 334 | return !D->hasAttr<OMPDeclareTargetDeclAttr>() && |
| 335 | (D->hasAttr<DeviceKernelAttr>() || |
| 336 | (isa<FunctionDecl>(Val: D) && D->hasAttr<CUDAGlobalAttr>()) || |
| 337 | (isa<VarDecl>(Val: D) && |
| 338 | (D->hasAttr<CUDADeviceAttr>() || D->hasAttr<CUDAConstantAttr>() || |
| 339 | cast<VarDecl>(Val: D)->getType()->isCUDADeviceBuiltinSurfaceType() || |
| 340 | cast<VarDecl>(Val: D)->getType()->isCUDADeviceBuiltinTextureType()))); |
| 341 | } |
| 342 | |
| 343 | void AMDGPUTargetCodeGenInfo::setFunctionDeclAttributes( |
| 344 | const FunctionDecl *FD, llvm::Function *F, CodeGenModule &M) const { |
| 345 | const auto *ReqdWGS = |
| 346 | M.getLangOpts().OpenCL ? FD->getAttr<ReqdWorkGroupSizeAttr>() : nullptr; |
| 347 | const bool IsOpenCLKernel = |
| 348 | M.getLangOpts().OpenCL && FD->hasAttr<DeviceKernelAttr>(); |
| 349 | const bool IsHIPKernel = M.getLangOpts().HIP && FD->hasAttr<CUDAGlobalAttr>(); |
| 350 | |
| 351 | const auto *FlatWGS = FD->getAttr<AMDGPUFlatWorkGroupSizeAttr>(); |
| 352 | if (ReqdWGS || FlatWGS) { |
| 353 | M.handleAMDGPUFlatWorkGroupSizeAttr(F, A: FlatWGS, ReqdWGS); |
| 354 | } else if (IsOpenCLKernel || IsHIPKernel) { |
| 355 | // By default, restrict the maximum size to a value specified by |
| 356 | // --gpu-max-threads-per-block=n or its default value for HIP. |
| 357 | const unsigned OpenCLDefaultMaxWorkGroupSize = 256; |
| 358 | const unsigned DefaultMaxWorkGroupSize = |
| 359 | IsOpenCLKernel ? OpenCLDefaultMaxWorkGroupSize |
| 360 | : M.getLangOpts().GPUMaxThreadsPerBlock; |
| 361 | std::string AttrVal = |
| 362 | std::string("1," ) + llvm::utostr(X: DefaultMaxWorkGroupSize); |
| 363 | F->addFnAttr(Kind: "amdgpu-flat-work-group-size" , Val: AttrVal); |
| 364 | } |
| 365 | |
| 366 | if (const auto *Attr = FD->getAttr<AMDGPUWavesPerEUAttr>()) |
| 367 | M.handleAMDGPUWavesPerEUAttr(F, A: Attr); |
| 368 | |
| 369 | if (const auto *Attr = FD->getAttr<AMDGPUNumSGPRAttr>()) { |
| 370 | unsigned NumSGPR = Attr->getNumSGPR(); |
| 371 | |
| 372 | if (NumSGPR != 0) |
| 373 | F->addFnAttr(Kind: "amdgpu-num-sgpr" , Val: llvm::utostr(X: NumSGPR)); |
| 374 | } |
| 375 | |
| 376 | if (const auto *Attr = FD->getAttr<AMDGPUNumVGPRAttr>()) { |
| 377 | uint32_t NumVGPR = Attr->getNumVGPR(); |
| 378 | |
| 379 | if (NumVGPR != 0) |
| 380 | F->addFnAttr(Kind: "amdgpu-num-vgpr" , Val: llvm::utostr(X: NumVGPR)); |
| 381 | } |
| 382 | |
| 383 | if (const auto *Attr = FD->getAttr<AMDGPUMaxNumWorkGroupsAttr>()) { |
| 384 | uint32_t X = Attr->getMaxNumWorkGroupsX() |
| 385 | ->EvaluateKnownConstInt(Ctx: M.getContext()) |
| 386 | .getExtValue(); |
| 387 | // Y and Z dimensions default to 1 if not specified |
| 388 | uint32_t Y = Attr->getMaxNumWorkGroupsY() |
| 389 | ? Attr->getMaxNumWorkGroupsY() |
| 390 | ->EvaluateKnownConstInt(Ctx: M.getContext()) |
| 391 | .getExtValue() |
| 392 | : 1; |
| 393 | uint32_t Z = Attr->getMaxNumWorkGroupsZ() |
| 394 | ? Attr->getMaxNumWorkGroupsZ() |
| 395 | ->EvaluateKnownConstInt(Ctx: M.getContext()) |
| 396 | .getExtValue() |
| 397 | : 1; |
| 398 | |
| 399 | llvm::SmallString<32> AttrVal; |
| 400 | llvm::raw_svector_ostream OS(AttrVal); |
| 401 | OS << X << ',' << Y << ',' << Z; |
| 402 | |
| 403 | F->addFnAttr(Kind: "amdgpu-max-num-workgroups" , Val: AttrVal.str()); |
| 404 | } |
| 405 | |
| 406 | if (auto *Attr = FD->getAttr<CUDAClusterDimsAttr>()) { |
| 407 | auto GetExprVal = [&](const auto &E) { |
| 408 | return E ? E->EvaluateKnownConstInt(M.getContext()).getExtValue() : 1; |
| 409 | }; |
| 410 | unsigned X = GetExprVal(Attr->getX()); |
| 411 | unsigned Y = GetExprVal(Attr->getY()); |
| 412 | unsigned Z = GetExprVal(Attr->getZ()); |
| 413 | llvm::SmallString<32> AttrVal; |
| 414 | llvm::raw_svector_ostream OS(AttrVal); |
| 415 | OS << X << ',' << Y << ',' << Z; |
| 416 | F->addFnAttr(Kind: "amdgpu-cluster-dims" , Val: AttrVal.str()); |
| 417 | } |
| 418 | |
| 419 | // OpenCL doesn't support cluster feature. |
| 420 | const TargetInfo &TTI = M.getContext().getTargetInfo(); |
| 421 | if ((IsOpenCLKernel && |
| 422 | TTI.hasFeatureEnabled(Features: TTI.getTargetOpts().FeatureMap, Name: "clusters" )) || |
| 423 | FD->hasAttr<CUDANoClusterAttr>()) |
| 424 | F->addFnAttr(Kind: "amdgpu-cluster-dims" , Val: "0,0,0" ); |
| 425 | } |
| 426 | |
| 427 | void AMDGPUTargetCodeGenInfo::setTargetAttributes( |
| 428 | const Decl *D, llvm::GlobalValue *GV, CodeGen::CodeGenModule &M) const { |
| 429 | if (requiresAMDGPUProtectedVisibility(D, GV)) { |
| 430 | GV->setVisibility(llvm::GlobalValue::ProtectedVisibility); |
| 431 | GV->setDSOLocal(true); |
| 432 | } |
| 433 | |
| 434 | if (GV->isDeclaration()) |
| 435 | return; |
| 436 | |
| 437 | llvm::Function *F = dyn_cast<llvm::Function>(Val: GV); |
| 438 | if (!F) |
| 439 | return; |
| 440 | |
| 441 | const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(Val: D); |
| 442 | if (FD) |
| 443 | setFunctionDeclAttributes(FD, F, M); |
| 444 | if (!getABIInfo().getCodeGenOpts().EmitIEEENaNCompliantInsts) |
| 445 | F->addFnAttr(Kind: "amdgpu-ieee" , Val: "false" ); |
| 446 | if (getABIInfo().getCodeGenOpts().AMDGPUExpandWaitcntProfiling) |
| 447 | F->addFnAttr(Kind: "amdgpu-expand-waitcnt-profiling" ); |
| 448 | } |
| 449 | |
| 450 | unsigned AMDGPUTargetCodeGenInfo::getDeviceKernelCallingConv() const { |
| 451 | return llvm::CallingConv::AMDGPU_KERNEL; |
| 452 | } |
| 453 | |
| 454 | // Currently LLVM assumes null pointers always have value 0, |
| 455 | // which results in incorrectly transformed IR. Therefore, instead of |
| 456 | // emitting null pointers in private and local address spaces, a null |
| 457 | // pointer in generic address space is emitted which is casted to a |
| 458 | // pointer in local or private address space. |
| 459 | llvm::Constant *AMDGPUTargetCodeGenInfo::getNullPointer( |
| 460 | const CodeGen::CodeGenModule &CGM, llvm::PointerType *PT, |
| 461 | QualType QT) const { |
| 462 | if (CGM.getContext().getTargetNullPointerValue(QT) == 0) |
| 463 | return llvm::ConstantPointerNull::get(T: PT); |
| 464 | |
| 465 | auto &Ctx = CGM.getContext(); |
| 466 | auto NPT = llvm::PointerType::get( |
| 467 | C&: PT->getContext(), AddressSpace: Ctx.getTargetAddressSpace(AS: LangAS::opencl_generic)); |
| 468 | return llvm::ConstantExpr::getAddrSpaceCast( |
| 469 | C: llvm::ConstantPointerNull::get(T: NPT), Ty: PT); |
| 470 | } |
| 471 | |
| 472 | LangAS |
| 473 | AMDGPUTargetCodeGenInfo::getGlobalVarAddressSpace(CodeGenModule &CGM, |
| 474 | const VarDecl *D) const { |
| 475 | assert(!CGM.getLangOpts().OpenCL && |
| 476 | !(CGM.getLangOpts().CUDA && CGM.getLangOpts().CUDAIsDevice) && |
| 477 | "Address space agnostic languages only" ); |
| 478 | LangAS DefaultGlobalAS = getLangASFromTargetAS( |
| 479 | TargetAS: CGM.getContext().getTargetAddressSpace(AS: LangAS::opencl_global)); |
| 480 | if (!D) |
| 481 | return DefaultGlobalAS; |
| 482 | |
| 483 | LangAS AddrSpace = D->getType().getAddressSpace(); |
| 484 | if (AddrSpace != LangAS::Default) |
| 485 | return AddrSpace; |
| 486 | |
| 487 | // Only promote to address space 4 if VarDecl has constant initialization. |
| 488 | if (D->getType().isConstantStorage(Ctx: CGM.getContext(), ExcludeCtor: false, ExcludeDtor: false) && |
| 489 | D->hasConstantInitialization()) { |
| 490 | if (auto ConstAS = CGM.getTarget().getConstantAddressSpace()) |
| 491 | return *ConstAS; |
| 492 | } |
| 493 | return DefaultGlobalAS; |
| 494 | } |
| 495 | |
| 496 | llvm::SyncScope::ID |
| 497 | AMDGPUTargetCodeGenInfo::getLLVMSyncScopeID(const LangOptions &LangOpts, |
| 498 | SyncScope Scope, |
| 499 | llvm::AtomicOrdering Ordering, |
| 500 | llvm::LLVMContext &Ctx) const { |
| 501 | std::string Name; |
| 502 | switch (Scope) { |
| 503 | case SyncScope::HIPSingleThread: |
| 504 | case SyncScope::SingleScope: |
| 505 | Name = "singlethread" ; |
| 506 | break; |
| 507 | case SyncScope::HIPWavefront: |
| 508 | case SyncScope::OpenCLSubGroup: |
| 509 | case SyncScope::WavefrontScope: |
| 510 | Name = "wavefront" ; |
| 511 | break; |
| 512 | case SyncScope::HIPCluster: |
| 513 | case SyncScope::ClusterScope: |
| 514 | Name = "cluster" ; |
| 515 | break; |
| 516 | case SyncScope::HIPWorkgroup: |
| 517 | case SyncScope::OpenCLWorkGroup: |
| 518 | case SyncScope::WorkgroupScope: |
| 519 | Name = "workgroup" ; |
| 520 | break; |
| 521 | case SyncScope::HIPAgent: |
| 522 | case SyncScope::OpenCLDevice: |
| 523 | case SyncScope::DeviceScope: |
| 524 | Name = "agent" ; |
| 525 | break; |
| 526 | case SyncScope::SystemScope: |
| 527 | case SyncScope::HIPSystem: |
| 528 | case SyncScope::OpenCLAllSVMDevices: |
| 529 | Name = "" ; |
| 530 | break; |
| 531 | } |
| 532 | |
| 533 | // OpenCL assumes by default that atomic scopes are per-address space for |
| 534 | // non-sequentially consistent operations. |
| 535 | if (Scope >= SyncScope::OpenCLWorkGroup && |
| 536 | Scope <= SyncScope::OpenCLSubGroup && |
| 537 | Ordering != llvm::AtomicOrdering::SequentiallyConsistent) { |
| 538 | if (!Name.empty()) |
| 539 | Name = Twine(Twine(Name) + Twine("-" )).str(); |
| 540 | |
| 541 | Name = Twine(Twine(Name) + Twine("one-as" )).str(); |
| 542 | } |
| 543 | |
| 544 | return Ctx.getOrInsertSyncScopeID(SSN: Name); |
| 545 | } |
| 546 | |
| 547 | void AMDGPUTargetCodeGenInfo::setTargetAtomicMetadata( |
| 548 | CodeGenFunction &CGF, llvm::Instruction &AtomicInst, |
| 549 | const AtomicExpr *AE) const { |
| 550 | auto *RMW = dyn_cast<llvm::AtomicRMWInst>(Val: &AtomicInst); |
| 551 | auto *CmpX = dyn_cast<llvm::AtomicCmpXchgInst>(Val: &AtomicInst); |
| 552 | |
| 553 | // OpenCL and old style HIP atomics consider atomics targeting thread private |
| 554 | // memory to be undefined. |
| 555 | // |
| 556 | // TODO: This is probably undefined for atomic load/store, but there's not |
| 557 | // much direct codegen benefit to knowing this. |
| 558 | if (((RMW && RMW->getPointerAddressSpace() == llvm::AMDGPUAS::FLAT_ADDRESS) || |
| 559 | (CmpX && |
| 560 | CmpX->getPointerAddressSpace() == llvm::AMDGPUAS::FLAT_ADDRESS)) && |
| 561 | AE && AE->threadPrivateMemoryAtomicsAreUndefined()) { |
| 562 | llvm::MDBuilder MDHelper(CGF.getLLVMContext()); |
| 563 | llvm::MDNode *ASRange = MDHelper.createRange( |
| 564 | Lo: llvm::APInt(32, llvm::AMDGPUAS::PRIVATE_ADDRESS), |
| 565 | Hi: llvm::APInt(32, llvm::AMDGPUAS::PRIVATE_ADDRESS + 1)); |
| 566 | AtomicInst.setMetadata(KindID: llvm::LLVMContext::MD_noalias_addrspace, Node: ASRange); |
| 567 | } |
| 568 | |
| 569 | if (!RMW) |
| 570 | return; |
| 571 | |
| 572 | AtomicOptions AO = CGF.CGM.getAtomicOpts(); |
| 573 | llvm::MDNode *Empty = llvm::MDNode::get(Context&: CGF.getLLVMContext(), MDs: {}); |
| 574 | if (!AO.getOption(Kind: clang::AtomicOptionKind::FineGrainedMemory)) |
| 575 | RMW->setMetadata(Kind: "amdgpu.no.fine.grained.memory" , Node: Empty); |
| 576 | if (!AO.getOption(Kind: clang::AtomicOptionKind::RemoteMemory)) |
| 577 | RMW->setMetadata(Kind: "amdgpu.no.remote.memory" , Node: Empty); |
| 578 | if (AO.getOption(Kind: clang::AtomicOptionKind::IgnoreDenormalMode) && |
| 579 | RMW->getOperation() == llvm::AtomicRMWInst::FAdd && |
| 580 | RMW->getType()->isFloatTy()) |
| 581 | RMW->setMetadata(Kind: "amdgpu.ignore.denormal.mode" , Node: Empty); |
| 582 | } |
| 583 | |
| 584 | bool AMDGPUTargetCodeGenInfo::shouldEmitStaticExternCAliases() const { |
| 585 | return false; |
| 586 | } |
| 587 | |
| 588 | bool AMDGPUTargetCodeGenInfo::shouldEmitDWARFBitFieldSeparators() const { |
| 589 | return true; |
| 590 | } |
| 591 | |
| 592 | void AMDGPUTargetCodeGenInfo::setCUDAKernelCallingConvention( |
| 593 | const FunctionType *&FT) const { |
| 594 | FT = getABIInfo().getContext().adjustFunctionType( |
| 595 | Fn: FT, EInfo: FT->getExtInfo().withCallingConv(cc: CC_DeviceKernel)); |
| 596 | } |
| 597 | |
| 598 | /// Return IR struct type for rtinfo struct in rocm-device-libs used for device |
| 599 | /// enqueue. |
| 600 | /// |
| 601 | /// ptr addrspace(1) kernel_object, i32 private_segment_size, |
| 602 | /// i32 group_segment_size |
| 603 | |
| 604 | static llvm::StructType * |
| 605 | getAMDGPURuntimeHandleType(llvm::LLVMContext &C, |
| 606 | llvm::Type *KernelDescriptorPtrTy) { |
| 607 | llvm::Type *Int32 = llvm::Type::getInt32Ty(C); |
| 608 | return llvm::StructType::create(Context&: C, Elements: {KernelDescriptorPtrTy, Int32, Int32}, |
| 609 | Name: "block.runtime.handle.t" ); |
| 610 | } |
| 611 | |
| 612 | /// Create an OpenCL kernel for an enqueued block. |
| 613 | /// |
| 614 | /// The type of the first argument (the block literal) is the struct type |
| 615 | /// of the block literal instead of a pointer type. The first argument |
| 616 | /// (block literal) is passed directly by value to the kernel. The kernel |
| 617 | /// allocates the same type of struct on stack and stores the block literal |
| 618 | /// to it and passes its pointer to the block invoke function. The kernel |
| 619 | /// has "enqueued-block" function attribute and kernel argument metadata. |
| 620 | llvm::Value *AMDGPUTargetCodeGenInfo::createEnqueuedBlockKernel( |
| 621 | CodeGenFunction &CGF, llvm::Function *Invoke, llvm::Type *BlockTy) const { |
| 622 | auto &Builder = CGF.Builder; |
| 623 | auto &C = CGF.getLLVMContext(); |
| 624 | |
| 625 | auto *InvokeFT = Invoke->getFunctionType(); |
| 626 | llvm::SmallVector<llvm::Type *, 2> ArgTys; |
| 627 | llvm::SmallVector<llvm::Metadata *, 8> AddressQuals; |
| 628 | llvm::SmallVector<llvm::Metadata *, 8> AccessQuals; |
| 629 | llvm::SmallVector<llvm::Metadata *, 8> ArgTypeNames; |
| 630 | llvm::SmallVector<llvm::Metadata *, 8> ArgBaseTypeNames; |
| 631 | llvm::SmallVector<llvm::Metadata *, 8> ArgTypeQuals; |
| 632 | llvm::SmallVector<llvm::Metadata *, 8> ArgNames; |
| 633 | |
| 634 | ArgTys.push_back(Elt: BlockTy); |
| 635 | ArgTypeNames.push_back(Elt: llvm::MDString::get(Context&: C, Str: "__block_literal" )); |
| 636 | AddressQuals.push_back(Elt: llvm::ConstantAsMetadata::get(C: Builder.getInt32(C: 0))); |
| 637 | ArgBaseTypeNames.push_back(Elt: llvm::MDString::get(Context&: C, Str: "__block_literal" )); |
| 638 | ArgTypeQuals.push_back(Elt: llvm::MDString::get(Context&: C, Str: "" )); |
| 639 | AccessQuals.push_back(Elt: llvm::MDString::get(Context&: C, Str: "none" )); |
| 640 | ArgNames.push_back(Elt: llvm::MDString::get(Context&: C, Str: "block_literal" )); |
| 641 | for (unsigned I = 1, E = InvokeFT->getNumParams(); I < E; ++I) { |
| 642 | ArgTys.push_back(Elt: InvokeFT->getParamType(i: I)); |
| 643 | ArgTypeNames.push_back(Elt: llvm::MDString::get(Context&: C, Str: "void*" )); |
| 644 | AddressQuals.push_back(Elt: llvm::ConstantAsMetadata::get(C: Builder.getInt32(C: 3))); |
| 645 | AccessQuals.push_back(Elt: llvm::MDString::get(Context&: C, Str: "none" )); |
| 646 | ArgBaseTypeNames.push_back(Elt: llvm::MDString::get(Context&: C, Str: "void*" )); |
| 647 | ArgTypeQuals.push_back(Elt: llvm::MDString::get(Context&: C, Str: "" )); |
| 648 | ArgNames.push_back( |
| 649 | Elt: llvm::MDString::get(Context&: C, Str: (Twine("local_arg" ) + Twine(I)).str())); |
| 650 | } |
| 651 | |
| 652 | llvm::Module &Mod = CGF.CGM.getModule(); |
| 653 | const llvm::DataLayout &DL = Mod.getDataLayout(); |
| 654 | |
| 655 | llvm::Twine Name = Invoke->getName() + "_kernel" ; |
| 656 | auto *FT = llvm::FunctionType::get(Result: llvm::Type::getVoidTy(C), Params: ArgTys, isVarArg: false); |
| 657 | |
| 658 | // The kernel itself can be internal, the runtime does not directly access the |
| 659 | // kernel address (only the kernel descriptor). |
| 660 | auto *F = llvm::Function::Create(Ty: FT, Linkage: llvm::GlobalValue::InternalLinkage, N: Name, |
| 661 | M: &Mod); |
| 662 | F->setCallingConv(getDeviceKernelCallingConv()); |
| 663 | |
| 664 | llvm::AttrBuilder KernelAttrs(C); |
| 665 | // FIXME: The invoke isn't applying the right attributes either |
| 666 | // FIXME: This is missing setTargetAttributes |
| 667 | CGF.CGM.addDefaultFunctionDefinitionAttributes(attrs&: KernelAttrs); |
| 668 | F->addFnAttrs(Attrs: KernelAttrs); |
| 669 | |
| 670 | auto IP = CGF.Builder.saveIP(); |
| 671 | auto *BB = llvm::BasicBlock::Create(Context&: C, Name: "entry" , Parent: F); |
| 672 | Builder.SetInsertPoint(BB); |
| 673 | const auto BlockAlign = DL.getPrefTypeAlign(Ty: BlockTy); |
| 674 | auto *BlockPtr = Builder.CreateAlloca(Ty: BlockTy, ArraySize: nullptr); |
| 675 | BlockPtr->setAlignment(BlockAlign); |
| 676 | Builder.CreateAlignedStore(Val: F->arg_begin(), Ptr: BlockPtr, Align: BlockAlign); |
| 677 | auto *Cast = Builder.CreatePointerCast(V: BlockPtr, DestTy: InvokeFT->getParamType(i: 0)); |
| 678 | llvm::SmallVector<llvm::Value *, 2> Args; |
| 679 | Args.push_back(Elt: Cast); |
| 680 | for (llvm::Argument &A : llvm::drop_begin(RangeOrContainer: F->args())) |
| 681 | Args.push_back(Elt: &A); |
| 682 | llvm::CallInst *call = Builder.CreateCall(Callee: Invoke, Args); |
| 683 | call->setCallingConv(Invoke->getCallingConv()); |
| 684 | Builder.CreateRetVoid(); |
| 685 | Builder.restoreIP(IP); |
| 686 | |
| 687 | F->setMetadata(Kind: "kernel_arg_addr_space" , Node: llvm::MDNode::get(Context&: C, MDs: AddressQuals)); |
| 688 | F->setMetadata(Kind: "kernel_arg_access_qual" , Node: llvm::MDNode::get(Context&: C, MDs: AccessQuals)); |
| 689 | F->setMetadata(Kind: "kernel_arg_type" , Node: llvm::MDNode::get(Context&: C, MDs: ArgTypeNames)); |
| 690 | F->setMetadata(Kind: "kernel_arg_base_type" , |
| 691 | Node: llvm::MDNode::get(Context&: C, MDs: ArgBaseTypeNames)); |
| 692 | F->setMetadata(Kind: "kernel_arg_type_qual" , Node: llvm::MDNode::get(Context&: C, MDs: ArgTypeQuals)); |
| 693 | if (CGF.CGM.getCodeGenOpts().EmitOpenCLArgMetadata) |
| 694 | F->setMetadata(Kind: "kernel_arg_name" , Node: llvm::MDNode::get(Context&: C, MDs: ArgNames)); |
| 695 | |
| 696 | llvm::StructType *HandleTy = getAMDGPURuntimeHandleType( |
| 697 | C, KernelDescriptorPtrTy: llvm::PointerType::get(C, AddressSpace: DL.getDefaultGlobalsAddressSpace())); |
| 698 | llvm::Constant *RuntimeHandleInitializer = |
| 699 | llvm::ConstantAggregateZero::get(Ty: HandleTy); |
| 700 | |
| 701 | llvm::Twine RuntimeHandleName = F->getName() + ".runtime.handle" ; |
| 702 | |
| 703 | // The runtime needs access to the runtime handle as an external symbol. The |
| 704 | // runtime handle will need to be made external later, in |
| 705 | // AMDGPUExportOpenCLEnqueuedBlocks. The kernel itself has a hidden reference |
| 706 | // inside the runtime handle, and is not directly referenced. |
| 707 | |
| 708 | // TODO: We would initialize the first field by declaring F->getName() + ".kd" |
| 709 | // to reference the kernel descriptor. The runtime wouldn't need to bother |
| 710 | // setting it. We would need to have a final symbol name though. |
| 711 | // TODO: Can we directly use an external symbol with getGlobalIdentifier? |
| 712 | auto *RuntimeHandle = new llvm::GlobalVariable( |
| 713 | Mod, HandleTy, |
| 714 | /*isConstant=*/true, llvm::GlobalValue::InternalLinkage, |
| 715 | /*Initializer=*/RuntimeHandleInitializer, RuntimeHandleName, |
| 716 | /*InsertBefore=*/nullptr, llvm::GlobalValue::NotThreadLocal, |
| 717 | DL.getDefaultGlobalsAddressSpace(), |
| 718 | /*isExternallyInitialized=*/true); |
| 719 | |
| 720 | llvm::MDNode *HandleAsMD = |
| 721 | llvm::MDNode::get(Context&: C, MDs: llvm::ValueAsMetadata::get(V: RuntimeHandle)); |
| 722 | F->setMetadata(KindID: llvm::LLVMContext::MD_associated, Node: HandleAsMD); |
| 723 | |
| 724 | RuntimeHandle->setSection(".amdgpu.kernel.runtime.handle" ); |
| 725 | |
| 726 | CGF.CGM.addUsedGlobal(GV: F); |
| 727 | CGF.CGM.addUsedGlobal(GV: RuntimeHandle); |
| 728 | return RuntimeHandle; |
| 729 | } |
| 730 | |
| 731 | void CodeGenModule::handleAMDGPUFlatWorkGroupSizeAttr( |
| 732 | llvm::Function *F, const AMDGPUFlatWorkGroupSizeAttr *FlatWGS, |
| 733 | const ReqdWorkGroupSizeAttr *ReqdWGS, int32_t *MinThreadsVal, |
| 734 | int32_t *MaxThreadsVal) { |
| 735 | unsigned Min = 0; |
| 736 | unsigned Max = 0; |
| 737 | auto Eval = [&](Expr *E) { |
| 738 | return E->EvaluateKnownConstInt(Ctx: getContext()).getExtValue(); |
| 739 | }; |
| 740 | if (FlatWGS) { |
| 741 | Min = Eval(FlatWGS->getMin()); |
| 742 | Max = Eval(FlatWGS->getMax()); |
| 743 | } |
| 744 | if (ReqdWGS && Min == 0 && Max == 0) |
| 745 | Min = Max = Eval(ReqdWGS->getXDim()) * Eval(ReqdWGS->getYDim()) * |
| 746 | Eval(ReqdWGS->getZDim()); |
| 747 | |
| 748 | if (Min != 0) { |
| 749 | assert(Min <= Max && "Min must be less than or equal Max" ); |
| 750 | |
| 751 | if (MinThreadsVal) |
| 752 | *MinThreadsVal = Min; |
| 753 | if (MaxThreadsVal) |
| 754 | *MaxThreadsVal = Max; |
| 755 | std::string AttrVal = llvm::utostr(X: Min) + "," + llvm::utostr(X: Max); |
| 756 | if (F) |
| 757 | F->addFnAttr(Kind: "amdgpu-flat-work-group-size" , Val: AttrVal); |
| 758 | } else |
| 759 | assert(Max == 0 && "Max must be zero" ); |
| 760 | } |
| 761 | |
| 762 | void CodeGenModule::handleAMDGPUWavesPerEUAttr( |
| 763 | llvm::Function *F, const AMDGPUWavesPerEUAttr *Attr) { |
| 764 | unsigned Min = |
| 765 | Attr->getMin()->EvaluateKnownConstInt(Ctx: getContext()).getExtValue(); |
| 766 | unsigned Max = |
| 767 | Attr->getMax() |
| 768 | ? Attr->getMax()->EvaluateKnownConstInt(Ctx: getContext()).getExtValue() |
| 769 | : 0; |
| 770 | |
| 771 | if (Min != 0) { |
| 772 | assert((Max == 0 || Min <= Max) && "Min must be less than or equal Max" ); |
| 773 | |
| 774 | std::string AttrVal = llvm::utostr(X: Min); |
| 775 | if (Max != 0) |
| 776 | AttrVal = AttrVal + "," + llvm::utostr(X: Max); |
| 777 | F->addFnAttr(Kind: "amdgpu-waves-per-eu" , Val: AttrVal); |
| 778 | } else |
| 779 | assert(Max == 0 && "Max must be zero" ); |
| 780 | } |
| 781 | |
| 782 | std::unique_ptr<TargetCodeGenInfo> |
| 783 | CodeGen::createAMDGPUTargetCodeGenInfo(CodeGenModule &CGM) { |
| 784 | return std::make_unique<AMDGPUTargetCodeGenInfo>(args&: CGM.getTypes()); |
| 785 | } |
| 786 | |