1//===- NVPTX.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/STLExtras.h"
12#include "llvm/ADT/StringExtras.h"
13#include "llvm/IR/CallingConv.h"
14#include "llvm/IR/IntrinsicsNVPTX.h"
15
16using namespace clang;
17using namespace clang::CodeGen;
18
19//===----------------------------------------------------------------------===//
20// NVPTX ABI Implementation
21//===----------------------------------------------------------------------===//
22
23namespace {
24
25class NVPTXTargetCodeGenInfo;
26
27class NVPTXABIInfo : public ABIInfo {
28 NVPTXTargetCodeGenInfo &CGInfo;
29
30public:
31 NVPTXABIInfo(CodeGenTypes &CGT, NVPTXTargetCodeGenInfo &Info)
32 : ABIInfo(CGT), CGInfo(Info) {}
33
34 ABIArgInfo classifyReturnType(QualType RetTy) const;
35 ABIArgInfo classifyArgumentType(QualType Ty) const;
36
37 void computeInfo(CGFunctionInfo &FI) const override;
38 RValue EmitVAArg(CodeGenFunction &CGF, Address VAListAddr, QualType Ty,
39 AggValueSlot Slot) const override;
40 bool isUnsupportedType(QualType T) const;
41 ABIArgInfo coerceToIntArrayWithLimit(QualType Ty, unsigned MaxSize) const;
42};
43
44class NVPTXTargetCodeGenInfo : public TargetCodeGenInfo {
45public:
46 NVPTXTargetCodeGenInfo(CodeGenTypes &CGT)
47 : TargetCodeGenInfo(std::make_unique<NVPTXABIInfo>(args&: CGT, args&: *this)) {}
48
49 void setTargetAttributes(const Decl *D, llvm::GlobalValue *GV,
50 CodeGen::CodeGenModule &M) const override;
51 bool shouldEmitStaticExternCAliases() const override;
52
53 llvm::Constant *getNullPointer(const CodeGen::CodeGenModule &CGM,
54 llvm::PointerType *T,
55 QualType QT) const override;
56
57 llvm::Type *getCUDADeviceBuiltinSurfaceDeviceType() const override {
58 // On the device side, surface reference is represented as an object handle
59 // in 64-bit integer.
60 return llvm::Type::getInt64Ty(C&: getABIInfo().getVMContext());
61 }
62
63 llvm::Type *getCUDADeviceBuiltinTextureDeviceType() const override {
64 // On the device side, texture reference is represented as an object handle
65 // in 64-bit integer.
66 return llvm::Type::getInt64Ty(C&: getABIInfo().getVMContext());
67 }
68
69 bool emitCUDADeviceBuiltinSurfaceDeviceCopy(CodeGenFunction &CGF, LValue Dst,
70 LValue Src) const override {
71 emitBuiltinSurfTexDeviceCopy(CGF, Dst, Src);
72 return true;
73 }
74
75 bool emitCUDADeviceBuiltinTextureDeviceCopy(CodeGenFunction &CGF, LValue Dst,
76 LValue Src) const override {
77 emitBuiltinSurfTexDeviceCopy(CGF, Dst, Src);
78 return true;
79 }
80
81 unsigned getDeviceKernelCallingConv() const override {
82 return llvm::CallingConv::PTX_Kernel;
83 }
84
85 // Adds a NamedMDNode with GV, Name, and Operand as operands, and adds the
86 // resulting MDNode to the nvvm.annotations MDNode.
87 static void addNVVMMetadata(llvm::GlobalValue *GV, StringRef Name,
88 int Operand);
89
90 static void
91 addGridConstantNVVMMetadata(llvm::GlobalValue *GV,
92 const SmallVectorImpl<int> &GridConstantArgs);
93
94private:
95 static void emitBuiltinSurfTexDeviceCopy(CodeGenFunction &CGF, LValue Dst,
96 LValue Src) {
97 llvm::Value *Handle = nullptr;
98 llvm::Constant *C =
99 llvm::dyn_cast<llvm::Constant>(Val: Src.getAddress().emitRawPointer(CGF));
100 // Lookup `addrspacecast` through the constant pointer if any.
101 if (auto *ASC = llvm::dyn_cast_or_null<llvm::AddrSpaceCastOperator>(Val: C))
102 C = llvm::cast<llvm::Constant>(Val: ASC->getPointerOperand());
103 if (auto *GV = llvm::dyn_cast_or_null<llvm::GlobalVariable>(Val: C)) {
104 // Load the handle from the specific global variable using
105 // `nvvm.texsurf.handle.internal` intrinsic.
106 Handle = CGF.EmitRuntimeCall(
107 callee: CGF.CGM.getIntrinsic(IID: llvm::Intrinsic::nvvm_texsurf_handle_internal,
108 Tys: {GV->getType()}),
109 args: {GV}, name: "texsurf_handle");
110 } else
111 Handle = CGF.EmitLoadOfScalar(lvalue: Src, Loc: SourceLocation());
112 CGF.EmitStoreOfScalar(value: Handle, lvalue: Dst);
113 }
114};
115
116/// Checks if the type is unsupported directly by the current target.
117bool NVPTXABIInfo::isUnsupportedType(QualType T) const {
118 ASTContext &Context = getContext();
119 if (!Context.getTargetInfo().hasFloat16Type() && T->isFloat16Type())
120 return true;
121 if (!Context.getTargetInfo().hasFloat128Type() &&
122 (T->isFloat128Type() ||
123 (T->isRealFloatingType() && Context.getTypeSize(T) == 128)))
124 return true;
125 if (const auto *EIT = T->getAs<BitIntType>())
126 return EIT->getNumBits() >
127 (Context.getTargetInfo().hasInt128Type() ? 128U : 64U);
128 if (!Context.getTargetInfo().hasInt128Type() && T->isIntegerType() &&
129 Context.getTypeSize(T) > 64U)
130 return true;
131 if (const auto *AT = T->getAsArrayTypeUnsafe())
132 return isUnsupportedType(T: AT->getElementType());
133 const auto *RT = T->getAs<RecordType>();
134 if (!RT)
135 return false;
136 const RecordDecl *RD = RT->getDecl();
137
138 // If this is a C++ record, check the bases first.
139 if (const CXXRecordDecl *CXXRD = dyn_cast<CXXRecordDecl>(Val: RD))
140 for (const CXXBaseSpecifier &I : CXXRD->bases())
141 if (isUnsupportedType(T: I.getType()))
142 return true;
143
144 for (const FieldDecl *I : RD->fields())
145 if (isUnsupportedType(T: I->getType()))
146 return true;
147 return false;
148}
149
150/// Coerce the given type into an array with maximum allowed size of elements.
151ABIArgInfo NVPTXABIInfo::coerceToIntArrayWithLimit(QualType Ty,
152 unsigned MaxSize) const {
153 // Alignment and Size are measured in bits.
154 const uint64_t Size = getContext().getTypeSize(T: Ty);
155 const uint64_t Alignment = getContext().getTypeAlign(T: Ty);
156 const unsigned Div = std::min<unsigned>(a: MaxSize, b: Alignment);
157 llvm::Type *IntType = llvm::Type::getIntNTy(C&: getVMContext(), N: Div);
158 const uint64_t NumElements = (Size + Div - 1) / Div;
159 return ABIArgInfo::getDirect(T: llvm::ArrayType::get(ElementType: IntType, NumElements));
160}
161
162ABIArgInfo NVPTXABIInfo::classifyReturnType(QualType RetTy) const {
163 if (RetTy->isVoidType())
164 return ABIArgInfo::getIgnore();
165
166 if (getContext().getLangOpts().OpenMP &&
167 getContext().getLangOpts().OpenMPIsTargetDevice &&
168 isUnsupportedType(T: RetTy))
169 return coerceToIntArrayWithLimit(Ty: RetTy, MaxSize: 64);
170
171 // note: this is different from default ABI
172 if (!RetTy->isScalarType())
173 return ABIArgInfo::getDirect();
174
175 // Treat an enum type as its underlying type.
176 if (const EnumType *EnumTy = RetTy->getAs<EnumType>())
177 RetTy = EnumTy->getDecl()->getIntegerType();
178
179 return (isPromotableIntegerTypeForABI(Ty: RetTy) ? ABIArgInfo::getExtend(Ty: RetTy)
180 : ABIArgInfo::getDirect());
181}
182
183ABIArgInfo NVPTXABIInfo::classifyArgumentType(QualType Ty) const {
184 // Treat an enum type as its underlying type.
185 if (const EnumType *EnumTy = Ty->getAs<EnumType>())
186 Ty = EnumTy->getDecl()->getIntegerType();
187
188 // Return aggregates type as indirect by value
189 if (isAggregateTypeForABI(T: Ty)) {
190 // Under CUDA device compilation, tex/surf builtin types are replaced with
191 // object types and passed directly.
192 if (getContext().getLangOpts().CUDAIsDevice) {
193 if (Ty->isCUDADeviceBuiltinSurfaceType())
194 return ABIArgInfo::getDirect(
195 T: CGInfo.getCUDADeviceBuiltinSurfaceDeviceType());
196 if (Ty->isCUDADeviceBuiltinTextureType())
197 return ABIArgInfo::getDirect(
198 T: CGInfo.getCUDADeviceBuiltinTextureDeviceType());
199 }
200 return getNaturalAlignIndirect(
201 Ty, /* AddrSpace */ getDataLayout().getAllocaAddrSpace(),
202 /* byval */ ByVal: true);
203 }
204
205 if (const auto *EIT = Ty->getAs<BitIntType>()) {
206 if ((EIT->getNumBits() > 128) ||
207 (!getContext().getTargetInfo().hasInt128Type() &&
208 EIT->getNumBits() > 64))
209 return getNaturalAlignIndirect(
210 Ty, /* AddrSpace */ getDataLayout().getAllocaAddrSpace(),
211 /* byval */ ByVal: true);
212 }
213
214 return (isPromotableIntegerTypeForABI(Ty) ? ABIArgInfo::getExtend(Ty)
215 : ABIArgInfo::getDirect());
216}
217
218void NVPTXABIInfo::computeInfo(CGFunctionInfo &FI) const {
219 if (!getCXXABI().classifyReturnType(FI))
220 FI.getReturnInfo() = classifyReturnType(RetTy: FI.getReturnType());
221
222 for (auto &&[ArgumentsCount, I] : llvm::enumerate(First: FI.arguments()))
223 I.info = ArgumentsCount < FI.getNumRequiredArgs()
224 ? classifyArgumentType(Ty: I.type)
225 : ABIArgInfo::getDirect();
226
227 // Always honor user-specified calling convention.
228 if (FI.getCallingConvention() != llvm::CallingConv::C)
229 return;
230
231 FI.setEffectiveCallingConvention(getRuntimeCC());
232}
233
234RValue NVPTXABIInfo::EmitVAArg(CodeGenFunction &CGF, Address VAListAddr,
235 QualType Ty, AggValueSlot Slot) const {
236 return emitVoidPtrVAArg(CGF, VAListAddr, ValueTy: Ty, /*IsIndirect=*/false,
237 ValueInfo: getContext().getTypeInfoInChars(T: Ty),
238 SlotSizeAndAlign: CharUnits::fromQuantity(Quantity: 1),
239 /*AllowHigherAlign=*/true, Slot);
240}
241
242void NVPTXTargetCodeGenInfo::setTargetAttributes(
243 const Decl *D, llvm::GlobalValue *GV, CodeGen::CodeGenModule &M) const {
244 if (GV->isDeclaration())
245 return;
246 const VarDecl *VD = dyn_cast_or_null<VarDecl>(Val: D);
247 if (VD) {
248 if (M.getLangOpts().CUDA) {
249 if (VD->getType()->isCUDADeviceBuiltinSurfaceType())
250 addNVVMMetadata(GV, Name: "surface", Operand: 1);
251 else if (VD->getType()->isCUDADeviceBuiltinTextureType())
252 addNVVMMetadata(GV, Name: "texture", Operand: 1);
253 return;
254 }
255 }
256
257 const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(Val: D);
258 if (!FD)
259 return;
260
261 llvm::Function *F = cast<llvm::Function>(Val: GV);
262
263 // Perform special handling in OpenCL/CUDA mode
264 if (M.getLangOpts().OpenCL || M.getLangOpts().CUDA) {
265 // Use function attributes to check for kernel functions
266 // By default, all functions are device functions
267 if (FD->hasAttr<DeviceKernelAttr>() || FD->hasAttr<CUDAGlobalAttr>()) {
268 // OpenCL/CUDA kernel functions get kernel metadata
269 // Create !{<func-ref>, metadata !"kernel", i32 1} node
270 // And kernel functions are not subject to inlining
271 F->addFnAttr(Kind: llvm::Attribute::NoInline);
272 if (FD->hasAttr<CUDAGlobalAttr>()) {
273 SmallVector<int, 10> GCI;
274 for (auto IV : llvm::enumerate(First: FD->parameters()))
275 if (IV.value()->hasAttr<CUDAGridConstantAttr>())
276 // For some reason arg indices are 1-based in NVVM
277 GCI.push_back(Elt: IV.index() + 1);
278 // Create !{<func-ref>, metadata !"kernel", i32 1} node
279 F->setCallingConv(llvm::CallingConv::PTX_Kernel);
280 addGridConstantNVVMMetadata(GV: F, GridConstantArgs: GCI);
281 }
282 if (CUDALaunchBoundsAttr *Attr = FD->getAttr<CUDALaunchBoundsAttr>())
283 M.handleCUDALaunchBoundsAttr(F, A: Attr);
284 }
285 }
286 // Attach kernel metadata directly if compiling for NVPTX.
287 if (FD->hasAttr<DeviceKernelAttr>()) {
288 F->setCallingConv(llvm::CallingConv::PTX_Kernel);
289 }
290}
291
292void NVPTXTargetCodeGenInfo::addNVVMMetadata(llvm::GlobalValue *GV,
293 StringRef Name, int Operand) {
294 llvm::Module *M = GV->getParent();
295 llvm::LLVMContext &Ctx = M->getContext();
296
297 // Get "nvvm.annotations" metadata node
298 llvm::NamedMDNode *MD = M->getOrInsertNamedMetadata(Name: "nvvm.annotations");
299
300 SmallVector<llvm::Metadata *, 5> MDVals = {
301 llvm::ConstantAsMetadata::get(C: GV), llvm::MDString::get(Context&: Ctx, Str: Name),
302 llvm::ConstantAsMetadata::get(
303 C: llvm::ConstantInt::get(Ty: llvm::Type::getInt32Ty(C&: Ctx), V: Operand))};
304
305 // Append metadata to nvvm.annotations
306 MD->addOperand(M: llvm::MDNode::get(Context&: Ctx, MDs: MDVals));
307}
308
309void NVPTXTargetCodeGenInfo::addGridConstantNVVMMetadata(
310 llvm::GlobalValue *GV, const SmallVectorImpl<int> &GridConstantArgs) {
311
312 llvm::Module *M = GV->getParent();
313 llvm::LLVMContext &Ctx = M->getContext();
314
315 // Get "nvvm.annotations" metadata node
316 llvm::NamedMDNode *MD = M->getOrInsertNamedMetadata(Name: "nvvm.annotations");
317
318 SmallVector<llvm::Metadata *, 5> MDVals = {llvm::ConstantAsMetadata::get(C: GV)};
319 if (!GridConstantArgs.empty()) {
320 SmallVector<llvm::Metadata *, 10> GCM;
321 for (int I : GridConstantArgs)
322 GCM.push_back(Elt: llvm::ConstantAsMetadata::get(
323 C: llvm::ConstantInt::get(Ty: llvm::Type::getInt32Ty(C&: Ctx), V: I)));
324 MDVals.append(IL: {llvm::MDString::get(Context&: Ctx, Str: "grid_constant"),
325 llvm::MDNode::get(Context&: Ctx, MDs: GCM)});
326 }
327
328 // Append metadata to nvvm.annotations
329 MD->addOperand(M: llvm::MDNode::get(Context&: Ctx, MDs: MDVals));
330}
331
332bool NVPTXTargetCodeGenInfo::shouldEmitStaticExternCAliases() const {
333 return false;
334}
335
336llvm::Constant *
337NVPTXTargetCodeGenInfo::getNullPointer(const CodeGen::CodeGenModule &CGM,
338 llvm::PointerType *PT,
339 QualType QT) const {
340 auto &Ctx = CGM.getContext();
341 if (PT->getAddressSpace() != Ctx.getTargetAddressSpace(AS: LangAS::opencl_local))
342 return llvm::ConstantPointerNull::get(T: PT);
343
344 auto NPT = llvm::PointerType::get(
345 C&: PT->getContext(), AddressSpace: Ctx.getTargetAddressSpace(AS: LangAS::opencl_generic));
346 return llvm::ConstantExpr::getAddrSpaceCast(
347 C: llvm::ConstantPointerNull::get(T: NPT), Ty: PT);
348}
349} // namespace
350
351void CodeGenModule::handleCUDALaunchBoundsAttr(llvm::Function *F,
352 const CUDALaunchBoundsAttr *Attr,
353 int32_t *MaxThreadsVal,
354 int32_t *MinBlocksVal,
355 int32_t *MaxClusterRankVal) {
356 llvm::APSInt MaxThreads(32);
357 MaxThreads = Attr->getMaxThreads()->EvaluateKnownConstInt(Ctx: getContext());
358 if (MaxThreads > 0) {
359 if (MaxThreadsVal)
360 *MaxThreadsVal = MaxThreads.getExtValue();
361 if (F)
362 F->addFnAttr(Kind: "nvvm.maxntid", Val: llvm::utostr(X: MaxThreads.getExtValue()));
363 }
364
365 // min and max blocks is an optional argument for CUDALaunchBoundsAttr. If it
366 // was not specified in __launch_bounds__ or if the user specified a 0 value,
367 // we don't have to add a PTX directive.
368 if (Attr->getMinBlocks()) {
369 llvm::APSInt MinBlocks(32);
370 MinBlocks = Attr->getMinBlocks()->EvaluateKnownConstInt(Ctx: getContext());
371 if (MinBlocks > 0) {
372 if (MinBlocksVal)
373 *MinBlocksVal = MinBlocks.getExtValue();
374 if (F)
375 F->addFnAttr(Kind: "nvvm.minctasm", Val: llvm::utostr(X: MinBlocks.getExtValue()));
376 }
377 }
378 if (Attr->getMaxBlocks()) {
379 llvm::APSInt MaxBlocks(32);
380 MaxBlocks = Attr->getMaxBlocks()->EvaluateKnownConstInt(Ctx: getContext());
381 if (MaxBlocks > 0) {
382 if (MaxClusterRankVal)
383 *MaxClusterRankVal = MaxBlocks.getExtValue();
384 if (F)
385 F->addFnAttr(Kind: "nvvm.maxclusterrank",
386 Val: llvm::utostr(X: MaxBlocks.getExtValue()));
387 }
388 }
389}
390
391std::unique_ptr<TargetCodeGenInfo>
392CodeGen::createNVPTXTargetCodeGenInfo(CodeGenModule &CGM) {
393 return std::make_unique<NVPTXTargetCodeGenInfo>(args&: CGM.getTypes());
394}
395