1//===------ CGOpenMPRuntimeGPU.h - Interface to OpenMP GPU Runtimes ------===//
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// This provides a generalized class for OpenMP runtime code generation
10// specialized by GPU targets NVPTX and AMDGCN.
11//
12//===----------------------------------------------------------------------===//
13
14#ifndef LLVM_CLANG_LIB_CODEGEN_CGOPENMPRUNTIMEGPU_H
15#define LLVM_CLANG_LIB_CODEGEN_CGOPENMPRUNTIMEGPU_H
16
17#include "CGOpenMPRuntime.h"
18#include "CodeGenFunction.h"
19#include "clang/AST/StmtOpenMP.h"
20
21namespace clang {
22namespace CodeGen {
23
24class CGOpenMPRuntimeGPU : public CGOpenMPRuntime {
25public:
26 /// Defines the execution mode.
27 enum ExecutionMode {
28 /// SPMD execution mode (all threads are worker threads).
29 EM_SPMD,
30 /// Non-SPMD execution mode (1 master thread, others are workers).
31 EM_NonSPMD,
32 /// Unknown execution mode (orphaned directive).
33 EM_Unknown,
34 };
35
36 /// Target codegen is specialized based on two data-sharing modes: CUDA, in
37 /// which the local variables are actually global threadlocal, and Generic, in
38 /// which the local variables are placed in global memory if they may escape
39 /// their declaration context.
40 enum DataSharingMode {
41 /// CUDA data sharing mode.
42 DS_CUDA,
43 /// Generic data-sharing mode.
44 DS_Generic,
45 };
46
47private:
48 /// Parallel outlined function work for workers to execute.
49 llvm::SmallVector<llvm::Function *, 16> Work;
50
51 struct EntryFunctionState {
52 SourceLocation Loc;
53 };
54
55 ExecutionMode getExecutionMode() const;
56
57 DataSharingMode getDataSharingMode() const;
58
59 /// Get barrier to synchronize all threads in a block.
60 void syncCTAThreads(CodeGenFunction &CGF);
61
62 /// Helper for target directive initialization.
63 void emitKernelInit(const OMPExecutableDirective &D, CodeGenFunction &CGF,
64 EntryFunctionState &EST, bool IsSPMD);
65
66 /// Helper for target directive finalization.
67 void emitKernelDeinit(CodeGenFunction &CGF, EntryFunctionState &EST,
68 bool IsSPMD);
69
70 /// Helper for generic variables globalization prolog.
71 void emitGenericVarsProlog(CodeGenFunction &CGF, SourceLocation Loc);
72
73 /// Helper for generic variables globalization epilog.
74 void emitGenericVarsEpilog(CodeGenFunction &CGF);
75
76 //
77 // Base class overrides.
78 //
79
80 /// Emit outlined function specialized for the Fork-Join
81 /// programming model for applicable target directives on the NVPTX device.
82 /// \param D Directive to emit.
83 /// \param ParentName Name of the function that encloses the target region.
84 /// \param OutlinedFn Outlined function value to be defined by this call.
85 /// \param OutlinedFnID Outlined function ID value to be defined by this call.
86 /// \param IsOffloadEntry True if the outlined function is an offload entry.
87 /// An outlined function may not be an entry if, e.g. the if clause always
88 /// evaluates to false.
89 void emitNonSPMDKernel(const OMPExecutableDirective &D, StringRef ParentName,
90 llvm::Function *&OutlinedFn,
91 llvm::Constant *&OutlinedFnID, bool IsOffloadEntry,
92 const RegionCodeGenTy &CodeGen);
93
94 /// Emit outlined function specialized for the Single Program
95 /// Multiple Data programming model for applicable target directives on the
96 /// NVPTX device.
97 /// \param D Directive to emit.
98 /// \param ParentName Name of the function that encloses the target region.
99 /// \param OutlinedFn Outlined function value to be defined by this call.
100 /// \param OutlinedFnID Outlined function ID value to be defined by this call.
101 /// \param IsOffloadEntry True if the outlined function is an offload entry.
102 /// \param CodeGen Object containing the target statements.
103 /// An outlined function may not be an entry if, e.g. the if clause always
104 /// evaluates to false.
105 void emitSPMDKernel(const OMPExecutableDirective &D, StringRef ParentName,
106 llvm::Function *&OutlinedFn,
107 llvm::Constant *&OutlinedFnID, bool IsOffloadEntry,
108 const RegionCodeGenTy &CodeGen);
109
110 /// Emit outlined function for 'target' directive on the NVPTX
111 /// device.
112 /// \param D Directive to emit.
113 /// \param ParentName Name of the function that encloses the target region.
114 /// \param OutlinedFn Outlined function value to be defined by this call.
115 /// \param OutlinedFnID Outlined function ID value to be defined by this call.
116 /// \param IsOffloadEntry True if the outlined function is an offload entry.
117 /// An outlined function may not be an entry if, e.g. the if clause always
118 /// evaluates to false.
119 void emitTargetOutlinedFunction(const OMPExecutableDirective &D,
120 StringRef ParentName,
121 llvm::Function *&OutlinedFn,
122 llvm::Constant *&OutlinedFnID,
123 bool IsOffloadEntry,
124 const RegionCodeGenTy &CodeGen) override;
125
126protected:
127 /// Check if the default location must be constant.
128 /// Constant for NVPTX for better optimization.
129 bool isDefaultLocationConstant() const override { return true; }
130
131public:
132 explicit CGOpenMPRuntimeGPU(CodeGenModule &CGM);
133
134 bool isGPU() const override { return true; };
135
136 /// Declare generalized virtual functions which need to be defined
137 /// by all specializations of OpenMPGPURuntime Targets like AMDGCN
138 /// and NVPTX.
139
140 /// Check if the variable length declaration is delayed:
141 bool isDelayedVariableLengthDecl(CodeGenFunction &CGF,
142 const VarDecl *VD) const override;
143
144 /// Get call to __kmpc_alloc_shared
145 std::pair<llvm::Value *, llvm::Value *>
146 getKmpcAllocShared(CodeGenFunction &CGF, const VarDecl *VD) override;
147
148 /// Get call to __kmpc_free_shared
149 void getKmpcFreeShared(
150 CodeGenFunction &CGF,
151 const std::pair<llvm::Value *, llvm::Value *> &AddrSizePair) override;
152
153 /// Get the id of the current thread on the GPU.
154 llvm::Value *getGPUThreadID(CodeGenFunction &CGF);
155
156 /// Get the maximum number of threads in a block of the GPU.
157 llvm::Value *getGPUNumThreads(CodeGenFunction &CGF);
158
159 /// Emit call to void __kmpc_push_proc_bind(ident_t *loc, kmp_int32
160 /// global_tid, int proc_bind) to generate code for 'proc_bind' clause.
161 void emitProcBindClause(CodeGenFunction &CGF,
162 llvm::omp::ProcBindKind ProcBind,
163 SourceLocation Loc) override;
164
165 /// Emits call to void __kmpc_push_num_threads(ident_t *loc, kmp_int32
166 /// global_tid, kmp_int32 num_threads) to generate code for 'num_threads'
167 /// clause.
168 /// \param NumThreads An integer value of threads.
169 void emitNumThreadsClause(CodeGenFunction &CGF, llvm::Value *NumThreads,
170 SourceLocation Loc) override;
171
172 /// This function ought to emit, in the general case, a call to
173 // the openmp runtime kmpc_push_num_teams. In NVPTX backend it is not needed
174 // as these numbers are obtained through the PTX grid and block configuration.
175 /// \param NumTeams An integer expression of teams.
176 /// \param ThreadLimit An integer expression of threads.
177 void emitNumTeamsClause(CodeGenFunction &CGF, const Expr *NumTeams,
178 const Expr *ThreadLimit, SourceLocation Loc) override;
179
180 /// Emits inlined function for the specified OpenMP parallel
181 // directive.
182 /// \a D. This outlined function has type void(*)(kmp_int32 *ThreadID,
183 /// kmp_int32 BoundID, struct context_vars*).
184 /// \param CGF Reference to current CodeGenFunction.
185 /// \param D OpenMP directive.
186 /// \param ThreadIDVar Variable for thread id in the current OpenMP region.
187 /// \param InnermostKind Kind of innermost directive (for simple directives it
188 /// is a directive itself, for combined - its innermost directive).
189 /// \param CodeGen Code generation sequence for the \a D directive.
190 llvm::Function *emitParallelOutlinedFunction(
191 CodeGenFunction &CGF, const OMPExecutableDirective &D,
192 const VarDecl *ThreadIDVar, OpenMPDirectiveKind InnermostKind,
193 const RegionCodeGenTy &CodeGen) override;
194
195 /// Emits inlined function for the specified OpenMP teams
196 // directive.
197 /// \a D. This outlined function has type void(*)(kmp_int32 *ThreadID,
198 /// kmp_int32 BoundID, struct context_vars*).
199 /// \param CGF Reference to current CodeGenFunction.
200 /// \param D OpenMP directive.
201 /// \param ThreadIDVar Variable for thread id in the current OpenMP region.
202 /// \param InnermostKind Kind of innermost directive (for simple directives it
203 /// is a directive itself, for combined - its innermost directive).
204 /// \param CodeGen Code generation sequence for the \a D directive.
205 llvm::Function *emitTeamsOutlinedFunction(
206 CodeGenFunction &CGF, const OMPExecutableDirective &D,
207 const VarDecl *ThreadIDVar, OpenMPDirectiveKind InnermostKind,
208 const RegionCodeGenTy &CodeGen) override;
209
210 /// Emits code for teams call of the \a OutlinedFn with
211 /// variables captured in a record which address is stored in \a
212 /// CapturedStruct.
213 /// \param OutlinedFn Outlined function to be run by team masters. Type of
214 /// this function is void(*)(kmp_int32 *, kmp_int32, struct context_vars*).
215 /// \param CapturedVars A pointer to the record with the references to
216 /// variables used in \a OutlinedFn function.
217 ///
218 void emitTeamsCall(CodeGenFunction &CGF, const OMPExecutableDirective &D,
219 SourceLocation Loc, llvm::Function *OutlinedFn,
220 ArrayRef<llvm::Value *> CapturedVars) override;
221
222 /// Emits code for parallel or serial call of the \a OutlinedFn with
223 /// variables captured in a record which address is stored in \a
224 /// CapturedStruct.
225 /// \param OutlinedFn Outlined function to be run in parallel threads. Type of
226 /// this function is void(*)(kmp_int32 *, kmp_int32, struct context_vars*).
227 /// \param CapturedVars A pointer to the record with the references to
228 /// variables used in \a OutlinedFn function.
229 /// \param IfCond Condition in the associated 'if' clause, if it was
230 /// specified, nullptr otherwise.
231 /// \param NumThreads The value corresponding to the num_threads clause, if
232 /// any,
233 /// or nullptr.
234 void emitParallelCall(CodeGenFunction &CGF, SourceLocation Loc,
235 llvm::Function *OutlinedFn,
236 ArrayRef<llvm::Value *> CapturedVars,
237 const Expr *IfCond, llvm::Value *NumThreads) override;
238
239 /// Emit an implicit/explicit barrier for OpenMP threads.
240 /// \param Kind Directive for which this implicit barrier call must be
241 /// generated. Must be OMPD_barrier for explicit barrier generation.
242 /// \param EmitChecks true if need to emit checks for cancellation barriers.
243 /// \param ForceSimpleCall true simple barrier call must be emitted, false if
244 /// runtime class decides which one to emit (simple or with cancellation
245 /// checks).
246 ///
247 void emitBarrierCall(CodeGenFunction &CGF, SourceLocation Loc,
248 OpenMPDirectiveKind Kind, bool EmitChecks = true,
249 bool ForceSimpleCall = false) override;
250
251 /// Emits a critical region.
252 /// \param CriticalName Name of the critical region.
253 /// \param CriticalOpGen Generator for the statement associated with the given
254 /// critical region.
255 /// \param Hint Value of the 'hint' clause (optional).
256 void emitCriticalRegion(CodeGenFunction &CGF, StringRef CriticalName,
257 const RegionCodeGenTy &CriticalOpGen,
258 SourceLocation Loc,
259 const Expr *Hint = nullptr) override;
260
261 /// Emit a code for reduction clause.
262 ///
263 /// \param Privates List of private copies for original reduction arguments.
264 /// \param LHSExprs List of LHS in \a ReductionOps reduction operations.
265 /// \param RHSExprs List of RHS in \a ReductionOps reduction operations.
266 /// \param ReductionOps List of reduction operations in form 'LHS binop RHS'
267 /// or 'operator binop(LHS, RHS)'.
268 /// \param Options List of options for reduction codegen:
269 /// WithNowait true if parent directive has also nowait clause, false
270 /// otherwise.
271 /// SimpleReduction Emit reduction operation only. Used for omp simd
272 /// directive on the host.
273 /// ReductionKind The kind of reduction to perform.
274 void emitReduction(CodeGenFunction &CGF, SourceLocation Loc,
275 ArrayRef<const Expr *> Privates,
276 ArrayRef<const Expr *> LHSExprs,
277 ArrayRef<const Expr *> RHSExprs,
278 ArrayRef<const Expr *> ReductionOps,
279 ReductionOptionsTy Options) override;
280
281 /// Translates the native parameter of outlined function if this is required
282 /// for target.
283 /// \param FD Field decl from captured record for the parameter.
284 /// \param NativeParam Parameter itself.
285 const VarDecl *translateParameter(const FieldDecl *FD,
286 const VarDecl *NativeParam) const override;
287
288 /// Gets the address of the native argument basing on the address of the
289 /// target-specific parameter.
290 /// \param NativeParam Parameter itself.
291 /// \param TargetParam Corresponding target-specific parameter.
292 Address getParameterAddress(CodeGenFunction &CGF, const VarDecl *NativeParam,
293 const VarDecl *TargetParam) const override;
294
295 /// Emits call of the outlined function with the provided arguments,
296 /// translating these arguments to correct target-specific arguments.
297 void emitOutlinedFunctionCall(
298 CodeGenFunction &CGF, SourceLocation Loc, llvm::FunctionCallee OutlinedFn,
299 ArrayRef<llvm::Value *> Args = std::nullopt) const override;
300
301 /// Emits OpenMP-specific function prolog.
302 /// Required for device constructs.
303 void emitFunctionProlog(CodeGenFunction &CGF, const Decl *D) override;
304
305 /// Gets the OpenMP-specific address of the local variable.
306 Address getAddressOfLocalVariable(CodeGenFunction &CGF,
307 const VarDecl *VD) override;
308
309 /// Cleans up references to the objects in finished function.
310 ///
311 void functionFinished(CodeGenFunction &CGF) override;
312
313 /// Choose a default value for the dist_schedule clause.
314 void getDefaultDistScheduleAndChunk(CodeGenFunction &CGF,
315 const OMPLoopDirective &S, OpenMPDistScheduleClauseKind &ScheduleKind,
316 llvm::Value *&Chunk) const override;
317
318 /// Choose a default value for the schedule clause.
319 void getDefaultScheduleAndChunk(CodeGenFunction &CGF,
320 const OMPLoopDirective &S, OpenMPScheduleClauseKind &ScheduleKind,
321 const Expr *&ChunkExpr) const override;
322
323 /// Adjust some parameters for the target-based directives, like addresses of
324 /// the variables captured by reference in lambdas.
325 void adjustTargetSpecificDataForLambdas(
326 CodeGenFunction &CGF, const OMPExecutableDirective &D) const override;
327
328 /// Perform check on requires decl to ensure that target architecture
329 /// supports unified addressing
330 void processRequiresDirective(const OMPRequiresDecl *D) override;
331
332 /// Checks if the variable has associated OMPAllocateDeclAttr attribute with
333 /// the predefined allocator and translates it into the corresponding address
334 /// space.
335 bool hasAllocateAttributeForGlobalVar(const VarDecl *VD, LangAS &AS) override;
336
337private:
338 /// Track the execution mode when codegening directives within a target
339 /// region. The appropriate mode (SPMD/NON-SPMD) is set on entry to the
340 /// target region and used by containing directives such as 'parallel'
341 /// to emit optimized code.
342 ExecutionMode CurrentExecutionMode = EM_Unknown;
343
344 /// Track the data sharing mode when codegening directives within a target
345 /// region.
346 DataSharingMode CurrentDataSharingMode = DataSharingMode::DS_Generic;
347
348 /// true if currently emitting code for target/teams/distribute region, false
349 /// - otherwise.
350 bool IsInTTDRegion = false;
351
352 /// Map between an outlined function and its wrapper.
353 llvm::DenseMap<llvm::Function *, llvm::Function *> WrapperFunctionsMap;
354
355 /// Emit function which wraps the outline parallel region
356 /// and controls the parameters which are passed to this function.
357 /// The wrapper ensures that the outlined function is called
358 /// with the correct arguments when data is shared.
359 llvm::Function *createParallelDataSharingWrapper(
360 llvm::Function *OutlinedParallelFn, const OMPExecutableDirective &D);
361
362 /// The data for the single globalized variable.
363 struct MappedVarData {
364 /// Corresponding field in the global record.
365 llvm::Value *GlobalizedVal = nullptr;
366 /// Corresponding address.
367 Address PrivateAddr = Address::invalid();
368 };
369 /// The map of local variables to their addresses in the global memory.
370 using DeclToAddrMapTy = llvm::MapVector<const Decl *, MappedVarData>;
371 /// Set of the parameters passed by value escaping OpenMP context.
372 using EscapedParamsTy = llvm::SmallPtrSet<const Decl *, 4>;
373 struct FunctionData {
374 DeclToAddrMapTy LocalVarData;
375 EscapedParamsTy EscapedParameters;
376 llvm::SmallVector<const ValueDecl*, 4> EscapedVariableLengthDecls;
377 llvm::SmallVector<const ValueDecl *, 4> DelayedVariableLengthDecls;
378 llvm::SmallVector<std::pair<llvm::Value *, llvm::Value *>, 4>
379 EscapedVariableLengthDeclsAddrs;
380 std::unique_ptr<CodeGenFunction::OMPMapVars> MappedParams;
381 };
382 /// Maps the function to the list of the globalized variables with their
383 /// addresses.
384 llvm::SmallDenseMap<llvm::Function *, FunctionData> FunctionGlobalizedDecls;
385 /// List of the records with the list of fields for the reductions across the
386 /// teams. Used to build the intermediate buffer for the fast teams
387 /// reductions.
388 /// All the records are gathered into a union `union.type` is created.
389 llvm::SmallVector<const RecordDecl *, 4> TeamsReductions;
390 /// Pair of the Non-SPMD team and all reductions variables in this team
391 /// region.
392 std::pair<const Decl *, llvm::SmallVector<const ValueDecl *, 4>>
393 TeamAndReductions;
394};
395
396} // CodeGen namespace.
397} // clang namespace.
398
399#endif // LLVM_CLANG_LIB_CODEGEN_CGOPENMPRUNTIMEGPU_H
400