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