File: | clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp |
Warning: | line 4811, column 8 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===---- CGOpenMPRuntimeNVPTX.cpp - Interface to OpenMP NVPTX 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 class for OpenMP runtime code generation specialized to NVPTX | |||
10 | // targets. | |||
11 | // | |||
12 | //===----------------------------------------------------------------------===// | |||
13 | ||||
14 | #include "CGOpenMPRuntimeNVPTX.h" | |||
15 | #include "CodeGenFunction.h" | |||
16 | #include "clang/AST/Attr.h" | |||
17 | #include "clang/AST/DeclOpenMP.h" | |||
18 | #include "clang/AST/StmtOpenMP.h" | |||
19 | #include "clang/AST/StmtVisitor.h" | |||
20 | #include "clang/Basic/Cuda.h" | |||
21 | #include "llvm/ADT/SmallPtrSet.h" | |||
22 | #include "llvm/IR/IntrinsicsNVPTX.h" | |||
23 | ||||
24 | using namespace clang; | |||
25 | using namespace CodeGen; | |||
26 | using namespace llvm::omp; | |||
27 | ||||
28 | namespace { | |||
29 | enum OpenMPRTLFunctionNVPTX { | |||
30 | /// Call to void __kmpc_kernel_init(kmp_int32 thread_limit, | |||
31 | /// int16_t RequiresOMPRuntime); | |||
32 | OMPRTL_NVPTX__kmpc_kernel_init, | |||
33 | /// Call to void __kmpc_kernel_deinit(int16_t IsOMPRuntimeInitialized); | |||
34 | OMPRTL_NVPTX__kmpc_kernel_deinit, | |||
35 | /// Call to void __kmpc_spmd_kernel_init(kmp_int32 thread_limit, | |||
36 | /// int16_t RequiresOMPRuntime, int16_t RequiresDataSharing); | |||
37 | OMPRTL_NVPTX__kmpc_spmd_kernel_init, | |||
38 | /// Call to void __kmpc_spmd_kernel_deinit_v2(int16_t RequiresOMPRuntime); | |||
39 | OMPRTL_NVPTX__kmpc_spmd_kernel_deinit_v2, | |||
40 | /// Call to void __kmpc_kernel_prepare_parallel(void | |||
41 | /// *outlined_function, int16_t | |||
42 | /// IsOMPRuntimeInitialized); | |||
43 | OMPRTL_NVPTX__kmpc_kernel_prepare_parallel, | |||
44 | /// Call to bool __kmpc_kernel_parallel(void **outlined_function, | |||
45 | /// int16_t IsOMPRuntimeInitialized); | |||
46 | OMPRTL_NVPTX__kmpc_kernel_parallel, | |||
47 | /// Call to void __kmpc_kernel_end_parallel(); | |||
48 | OMPRTL_NVPTX__kmpc_kernel_end_parallel, | |||
49 | /// Call to void __kmpc_serialized_parallel(ident_t *loc, kmp_int32 | |||
50 | /// global_tid); | |||
51 | OMPRTL_NVPTX__kmpc_serialized_parallel, | |||
52 | /// Call to void __kmpc_end_serialized_parallel(ident_t *loc, kmp_int32 | |||
53 | /// global_tid); | |||
54 | OMPRTL_NVPTX__kmpc_end_serialized_parallel, | |||
55 | /// Call to int32_t __kmpc_shuffle_int32(int32_t element, | |||
56 | /// int16_t lane_offset, int16_t warp_size); | |||
57 | OMPRTL_NVPTX__kmpc_shuffle_int32, | |||
58 | /// Call to int64_t __kmpc_shuffle_int64(int64_t element, | |||
59 | /// int16_t lane_offset, int16_t warp_size); | |||
60 | OMPRTL_NVPTX__kmpc_shuffle_int64, | |||
61 | /// Call to __kmpc_nvptx_parallel_reduce_nowait_v2(ident_t *loc, kmp_int32 | |||
62 | /// global_tid, kmp_int32 num_vars, size_t reduce_size, void* reduce_data, | |||
63 | /// void (*kmp_ShuffleReductFctPtr)(void *rhsData, int16_t lane_id, int16_t | |||
64 | /// lane_offset, int16_t shortCircuit), | |||
65 | /// void (*kmp_InterWarpCopyFctPtr)(void* src, int32_t warp_num)); | |||
66 | OMPRTL_NVPTX__kmpc_nvptx_parallel_reduce_nowait_v2, | |||
67 | /// Call to __kmpc_nvptx_teams_reduce_nowait_v2(ident_t *loc, kmp_int32 | |||
68 | /// global_tid, void *global_buffer, int32_t num_of_records, void* | |||
69 | /// reduce_data, | |||
70 | /// void (*kmp_ShuffleReductFctPtr)(void *rhsData, int16_t lane_id, int16_t | |||
71 | /// lane_offset, int16_t shortCircuit), | |||
72 | /// void (*kmp_InterWarpCopyFctPtr)(void* src, int32_t warp_num), void | |||
73 | /// (*kmp_ListToGlobalCpyFctPtr)(void *buffer, int idx, void *reduce_data), | |||
74 | /// void (*kmp_GlobalToListCpyFctPtr)(void *buffer, int idx, | |||
75 | /// void *reduce_data), void (*kmp_GlobalToListCpyPtrsFctPtr)(void *buffer, | |||
76 | /// int idx, void *reduce_data), void (*kmp_GlobalToListRedFctPtr)(void | |||
77 | /// *buffer, int idx, void *reduce_data)); | |||
78 | OMPRTL_NVPTX__kmpc_nvptx_teams_reduce_nowait_v2, | |||
79 | /// Call to __kmpc_nvptx_end_reduce_nowait(int32_t global_tid); | |||
80 | OMPRTL_NVPTX__kmpc_end_reduce_nowait, | |||
81 | /// Call to void __kmpc_data_sharing_init_stack(); | |||
82 | OMPRTL_NVPTX__kmpc_data_sharing_init_stack, | |||
83 | /// Call to void __kmpc_data_sharing_init_stack_spmd(); | |||
84 | OMPRTL_NVPTX__kmpc_data_sharing_init_stack_spmd, | |||
85 | /// Call to void* __kmpc_data_sharing_coalesced_push_stack(size_t size, | |||
86 | /// int16_t UseSharedMemory); | |||
87 | OMPRTL_NVPTX__kmpc_data_sharing_coalesced_push_stack, | |||
88 | /// Call to void __kmpc_data_sharing_pop_stack(void *a); | |||
89 | OMPRTL_NVPTX__kmpc_data_sharing_pop_stack, | |||
90 | /// Call to void __kmpc_begin_sharing_variables(void ***args, | |||
91 | /// size_t n_args); | |||
92 | OMPRTL_NVPTX__kmpc_begin_sharing_variables, | |||
93 | /// Call to void __kmpc_end_sharing_variables(); | |||
94 | OMPRTL_NVPTX__kmpc_end_sharing_variables, | |||
95 | /// Call to void __kmpc_get_shared_variables(void ***GlobalArgs) | |||
96 | OMPRTL_NVPTX__kmpc_get_shared_variables, | |||
97 | /// Call to uint16_t __kmpc_parallel_level(ident_t *loc, kmp_int32 | |||
98 | /// global_tid); | |||
99 | OMPRTL_NVPTX__kmpc_parallel_level, | |||
100 | /// Call to int8_t __kmpc_is_spmd_exec_mode(); | |||
101 | OMPRTL_NVPTX__kmpc_is_spmd_exec_mode, | |||
102 | /// Call to void __kmpc_get_team_static_memory(int16_t isSPMDExecutionMode, | |||
103 | /// const void *buf, size_t size, int16_t is_shared, const void **res); | |||
104 | OMPRTL_NVPTX__kmpc_get_team_static_memory, | |||
105 | /// Call to void __kmpc_restore_team_static_memory(int16_t | |||
106 | /// isSPMDExecutionMode, int16_t is_shared); | |||
107 | OMPRTL_NVPTX__kmpc_restore_team_static_memory, | |||
108 | /// Call to void __kmpc_barrier(ident_t *loc, kmp_int32 global_tid); | |||
109 | OMPRTL__kmpc_barrier, | |||
110 | /// Call to void __kmpc_barrier_simple_spmd(ident_t *loc, kmp_int32 | |||
111 | /// global_tid); | |||
112 | OMPRTL__kmpc_barrier_simple_spmd, | |||
113 | /// Call to int32_t __kmpc_warp_active_thread_mask(void); | |||
114 | OMPRTL_NVPTX__kmpc_warp_active_thread_mask, | |||
115 | /// Call to void __kmpc_syncwarp(int32_t Mask); | |||
116 | OMPRTL_NVPTX__kmpc_syncwarp, | |||
117 | }; | |||
118 | ||||
119 | /// Pre(post)-action for different OpenMP constructs specialized for NVPTX. | |||
120 | class NVPTXActionTy final : public PrePostActionTy { | |||
121 | llvm::FunctionCallee EnterCallee = nullptr; | |||
122 | ArrayRef<llvm::Value *> EnterArgs; | |||
123 | llvm::FunctionCallee ExitCallee = nullptr; | |||
124 | ArrayRef<llvm::Value *> ExitArgs; | |||
125 | bool Conditional = false; | |||
126 | llvm::BasicBlock *ContBlock = nullptr; | |||
127 | ||||
128 | public: | |||
129 | NVPTXActionTy(llvm::FunctionCallee EnterCallee, | |||
130 | ArrayRef<llvm::Value *> EnterArgs, | |||
131 | llvm::FunctionCallee ExitCallee, | |||
132 | ArrayRef<llvm::Value *> ExitArgs, bool Conditional = false) | |||
133 | : EnterCallee(EnterCallee), EnterArgs(EnterArgs), ExitCallee(ExitCallee), | |||
134 | ExitArgs(ExitArgs), Conditional(Conditional) {} | |||
135 | void Enter(CodeGenFunction &CGF) override { | |||
136 | llvm::Value *EnterRes = CGF.EmitRuntimeCall(EnterCallee, EnterArgs); | |||
137 | if (Conditional) { | |||
138 | llvm::Value *CallBool = CGF.Builder.CreateIsNotNull(EnterRes); | |||
139 | auto *ThenBlock = CGF.createBasicBlock("omp_if.then"); | |||
140 | ContBlock = CGF.createBasicBlock("omp_if.end"); | |||
141 | // Generate the branch (If-stmt) | |||
142 | CGF.Builder.CreateCondBr(CallBool, ThenBlock, ContBlock); | |||
143 | CGF.EmitBlock(ThenBlock); | |||
144 | } | |||
145 | } | |||
146 | void Done(CodeGenFunction &CGF) { | |||
147 | // Emit the rest of blocks/branches | |||
148 | CGF.EmitBranch(ContBlock); | |||
149 | CGF.EmitBlock(ContBlock, true); | |||
150 | } | |||
151 | void Exit(CodeGenFunction &CGF) override { | |||
152 | CGF.EmitRuntimeCall(ExitCallee, ExitArgs); | |||
153 | } | |||
154 | }; | |||
155 | ||||
156 | /// A class to track the execution mode when codegening directives within | |||
157 | /// a target region. The appropriate mode (SPMD|NON-SPMD) is set on entry | |||
158 | /// to the target region and used by containing directives such as 'parallel' | |||
159 | /// to emit optimized code. | |||
160 | class ExecutionRuntimeModesRAII { | |||
161 | private: | |||
162 | CGOpenMPRuntimeNVPTX::ExecutionMode SavedExecMode = | |||
163 | CGOpenMPRuntimeNVPTX::EM_Unknown; | |||
164 | CGOpenMPRuntimeNVPTX::ExecutionMode &ExecMode; | |||
165 | bool SavedRuntimeMode = false; | |||
166 | bool *RuntimeMode = nullptr; | |||
167 | ||||
168 | public: | |||
169 | /// Constructor for Non-SPMD mode. | |||
170 | ExecutionRuntimeModesRAII(CGOpenMPRuntimeNVPTX::ExecutionMode &ExecMode) | |||
171 | : ExecMode(ExecMode) { | |||
172 | SavedExecMode = ExecMode; | |||
173 | ExecMode = CGOpenMPRuntimeNVPTX::EM_NonSPMD; | |||
174 | } | |||
175 | /// Constructor for SPMD mode. | |||
176 | ExecutionRuntimeModesRAII(CGOpenMPRuntimeNVPTX::ExecutionMode &ExecMode, | |||
177 | bool &RuntimeMode, bool FullRuntimeMode) | |||
178 | : ExecMode(ExecMode), RuntimeMode(&RuntimeMode) { | |||
179 | SavedExecMode = ExecMode; | |||
180 | SavedRuntimeMode = RuntimeMode; | |||
181 | ExecMode = CGOpenMPRuntimeNVPTX::EM_SPMD; | |||
182 | RuntimeMode = FullRuntimeMode; | |||
183 | } | |||
184 | ~ExecutionRuntimeModesRAII() { | |||
185 | ExecMode = SavedExecMode; | |||
186 | if (RuntimeMode) | |||
187 | *RuntimeMode = SavedRuntimeMode; | |||
188 | } | |||
189 | }; | |||
190 | ||||
191 | /// GPU Configuration: This information can be derived from cuda registers, | |||
192 | /// however, providing compile time constants helps generate more efficient | |||
193 | /// code. For all practical purposes this is fine because the configuration | |||
194 | /// is the same for all known NVPTX architectures. | |||
195 | enum MachineConfiguration : unsigned { | |||
196 | WarpSize = 32, | |||
197 | /// Number of bits required to represent a lane identifier, which is | |||
198 | /// computed as log_2(WarpSize). | |||
199 | LaneIDBits = 5, | |||
200 | LaneIDMask = WarpSize - 1, | |||
201 | ||||
202 | /// Global memory alignment for performance. | |||
203 | GlobalMemoryAlignment = 128, | |||
204 | ||||
205 | /// Maximal size of the shared memory buffer. | |||
206 | SharedMemorySize = 128, | |||
207 | }; | |||
208 | ||||
209 | static const ValueDecl *getPrivateItem(const Expr *RefExpr) { | |||
210 | RefExpr = RefExpr->IgnoreParens(); | |||
211 | if (const auto *ASE = dyn_cast<ArraySubscriptExpr>(RefExpr)) { | |||
212 | const Expr *Base = ASE->getBase()->IgnoreParenImpCasts(); | |||
213 | while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Base)) | |||
214 | Base = TempASE->getBase()->IgnoreParenImpCasts(); | |||
215 | RefExpr = Base; | |||
216 | } else if (auto *OASE = dyn_cast<OMPArraySectionExpr>(RefExpr)) { | |||
217 | const Expr *Base = OASE->getBase()->IgnoreParenImpCasts(); | |||
218 | while (const auto *TempOASE = dyn_cast<OMPArraySectionExpr>(Base)) | |||
219 | Base = TempOASE->getBase()->IgnoreParenImpCasts(); | |||
220 | while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Base)) | |||
221 | Base = TempASE->getBase()->IgnoreParenImpCasts(); | |||
222 | RefExpr = Base; | |||
223 | } | |||
224 | RefExpr = RefExpr->IgnoreParenImpCasts(); | |||
225 | if (const auto *DE = dyn_cast<DeclRefExpr>(RefExpr)) | |||
226 | return cast<ValueDecl>(DE->getDecl()->getCanonicalDecl()); | |||
227 | const auto *ME = cast<MemberExpr>(RefExpr); | |||
228 | return cast<ValueDecl>(ME->getMemberDecl()->getCanonicalDecl()); | |||
229 | } | |||
230 | ||||
231 | ||||
232 | static RecordDecl *buildRecordForGlobalizedVars( | |||
233 | ASTContext &C, ArrayRef<const ValueDecl *> EscapedDecls, | |||
234 | ArrayRef<const ValueDecl *> EscapedDeclsForTeams, | |||
235 | llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> | |||
236 | &MappedDeclsFields, int BufSize) { | |||
237 | using VarsDataTy = std::pair<CharUnits /*Align*/, const ValueDecl *>; | |||
238 | if (EscapedDecls.empty() && EscapedDeclsForTeams.empty()) | |||
239 | return nullptr; | |||
240 | SmallVector<VarsDataTy, 4> GlobalizedVars; | |||
241 | for (const ValueDecl *D : EscapedDecls) | |||
242 | GlobalizedVars.emplace_back( | |||
243 | CharUnits::fromQuantity(std::max( | |||
244 | C.getDeclAlign(D).getQuantity(), | |||
245 | static_cast<CharUnits::QuantityType>(GlobalMemoryAlignment))), | |||
246 | D); | |||
247 | for (const ValueDecl *D : EscapedDeclsForTeams) | |||
248 | GlobalizedVars.emplace_back(C.getDeclAlign(D), D); | |||
249 | llvm::stable_sort(GlobalizedVars, [](VarsDataTy L, VarsDataTy R) { | |||
250 | return L.first > R.first; | |||
251 | }); | |||
252 | ||||
253 | // Build struct _globalized_locals_ty { | |||
254 | // /* globalized vars */[WarSize] align (max(decl_align, | |||
255 | // GlobalMemoryAlignment)) | |||
256 | // /* globalized vars */ for EscapedDeclsForTeams | |||
257 | // }; | |||
258 | RecordDecl *GlobalizedRD = C.buildImplicitRecord("_globalized_locals_ty"); | |||
259 | GlobalizedRD->startDefinition(); | |||
260 | llvm::SmallPtrSet<const ValueDecl *, 16> SingleEscaped( | |||
261 | EscapedDeclsForTeams.begin(), EscapedDeclsForTeams.end()); | |||
262 | for (const auto &Pair : GlobalizedVars) { | |||
263 | const ValueDecl *VD = Pair.second; | |||
264 | QualType Type = VD->getType(); | |||
265 | if (Type->isLValueReferenceType()) | |||
266 | Type = C.getPointerType(Type.getNonReferenceType()); | |||
267 | else | |||
268 | Type = Type.getNonReferenceType(); | |||
269 | SourceLocation Loc = VD->getLocation(); | |||
270 | FieldDecl *Field; | |||
271 | if (SingleEscaped.count(VD)) { | |||
272 | Field = FieldDecl::Create( | |||
273 | C, GlobalizedRD, Loc, Loc, VD->getIdentifier(), Type, | |||
274 | C.getTrivialTypeSourceInfo(Type, SourceLocation()), | |||
275 | /*BW=*/nullptr, /*Mutable=*/false, | |||
276 | /*InitStyle=*/ICIS_NoInit); | |||
277 | Field->setAccess(AS_public); | |||
278 | if (VD->hasAttrs()) { | |||
279 | for (specific_attr_iterator<AlignedAttr> I(VD->getAttrs().begin()), | |||
280 | E(VD->getAttrs().end()); | |||
281 | I != E; ++I) | |||
282 | Field->addAttr(*I); | |||
283 | } | |||
284 | } else { | |||
285 | llvm::APInt ArraySize(32, BufSize); | |||
286 | Type = C.getConstantArrayType(Type, ArraySize, nullptr, ArrayType::Normal, | |||
287 | 0); | |||
288 | Field = FieldDecl::Create( | |||
289 | C, GlobalizedRD, Loc, Loc, VD->getIdentifier(), Type, | |||
290 | C.getTrivialTypeSourceInfo(Type, SourceLocation()), | |||
291 | /*BW=*/nullptr, /*Mutable=*/false, | |||
292 | /*InitStyle=*/ICIS_NoInit); | |||
293 | Field->setAccess(AS_public); | |||
294 | llvm::APInt Align(32, std::max(C.getDeclAlign(VD).getQuantity(), | |||
295 | static_cast<CharUnits::QuantityType>( | |||
296 | GlobalMemoryAlignment))); | |||
297 | Field->addAttr(AlignedAttr::CreateImplicit( | |||
298 | C, /*IsAlignmentExpr=*/true, | |||
299 | IntegerLiteral::Create(C, Align, | |||
300 | C.getIntTypeForBitwidth(32, /*Signed=*/0), | |||
301 | SourceLocation()), | |||
302 | {}, AttributeCommonInfo::AS_GNU, AlignedAttr::GNU_aligned)); | |||
303 | } | |||
304 | GlobalizedRD->addDecl(Field); | |||
305 | MappedDeclsFields.try_emplace(VD, Field); | |||
306 | } | |||
307 | GlobalizedRD->completeDefinition(); | |||
308 | return GlobalizedRD; | |||
309 | } | |||
310 | ||||
311 | /// Get the list of variables that can escape their declaration context. | |||
312 | class CheckVarsEscapingDeclContext final | |||
313 | : public ConstStmtVisitor<CheckVarsEscapingDeclContext> { | |||
314 | CodeGenFunction &CGF; | |||
315 | llvm::SetVector<const ValueDecl *> EscapedDecls; | |||
316 | llvm::SetVector<const ValueDecl *> EscapedVariableLengthDecls; | |||
317 | llvm::SmallPtrSet<const Decl *, 4> EscapedParameters; | |||
318 | RecordDecl *GlobalizedRD = nullptr; | |||
319 | llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> MappedDeclsFields; | |||
320 | bool AllEscaped = false; | |||
321 | bool IsForCombinedParallelRegion = false; | |||
322 | ||||
323 | void markAsEscaped(const ValueDecl *VD) { | |||
324 | // Do not globalize declare target variables. | |||
325 | if (!isa<VarDecl>(VD) || | |||
326 | OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD)) | |||
327 | return; | |||
328 | VD = cast<ValueDecl>(VD->getCanonicalDecl()); | |||
329 | // Use user-specified allocation. | |||
330 | if (VD->hasAttrs() && VD->hasAttr<OMPAllocateDeclAttr>()) | |||
331 | return; | |||
332 | // Variables captured by value must be globalized. | |||
333 | if (auto *CSI = CGF.CapturedStmtInfo) { | |||
334 | if (const FieldDecl *FD = CSI->lookup(cast<VarDecl>(VD))) { | |||
335 | // Check if need to capture the variable that was already captured by | |||
336 | // value in the outer region. | |||
337 | if (!IsForCombinedParallelRegion) { | |||
338 | if (!FD->hasAttrs()) | |||
339 | return; | |||
340 | const auto *Attr = FD->getAttr<OMPCaptureKindAttr>(); | |||
341 | if (!Attr) | |||
342 | return; | |||
343 | if (((Attr->getCaptureKind() != OMPC_map) && | |||
344 | !isOpenMPPrivate( | |||
345 | static_cast<OpenMPClauseKind>(Attr->getCaptureKind()))) || | |||
346 | ((Attr->getCaptureKind() == OMPC_map) && | |||
347 | !FD->getType()->isAnyPointerType())) | |||
348 | return; | |||
349 | } | |||
350 | if (!FD->getType()->isReferenceType()) { | |||
351 | assert(!VD->getType()->isVariablyModifiedType() &&((!VD->getType()->isVariablyModifiedType() && "Parameter captured by value with variably modified type" ) ? static_cast<void> (0) : __assert_fail ("!VD->getType()->isVariablyModifiedType() && \"Parameter captured by value with variably modified type\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 352, __PRETTY_FUNCTION__)) | |||
352 | "Parameter captured by value with variably modified type")((!VD->getType()->isVariablyModifiedType() && "Parameter captured by value with variably modified type" ) ? static_cast<void> (0) : __assert_fail ("!VD->getType()->isVariablyModifiedType() && \"Parameter captured by value with variably modified type\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 352, __PRETTY_FUNCTION__)); | |||
353 | EscapedParameters.insert(VD); | |||
354 | } else if (!IsForCombinedParallelRegion) { | |||
355 | return; | |||
356 | } | |||
357 | } | |||
358 | } | |||
359 | if ((!CGF.CapturedStmtInfo || | |||
360 | (IsForCombinedParallelRegion && CGF.CapturedStmtInfo)) && | |||
361 | VD->getType()->isReferenceType()) | |||
362 | // Do not globalize variables with reference type. | |||
363 | return; | |||
364 | if (VD->getType()->isVariablyModifiedType()) | |||
365 | EscapedVariableLengthDecls.insert(VD); | |||
366 | else | |||
367 | EscapedDecls.insert(VD); | |||
368 | } | |||
369 | ||||
370 | void VisitValueDecl(const ValueDecl *VD) { | |||
371 | if (VD->getType()->isLValueReferenceType()) | |||
372 | markAsEscaped(VD); | |||
373 | if (const auto *VarD = dyn_cast<VarDecl>(VD)) { | |||
374 | if (!isa<ParmVarDecl>(VarD) && VarD->hasInit()) { | |||
375 | const bool SavedAllEscaped = AllEscaped; | |||
376 | AllEscaped = VD->getType()->isLValueReferenceType(); | |||
377 | Visit(VarD->getInit()); | |||
378 | AllEscaped = SavedAllEscaped; | |||
379 | } | |||
380 | } | |||
381 | } | |||
382 | void VisitOpenMPCapturedStmt(const CapturedStmt *S, | |||
383 | ArrayRef<OMPClause *> Clauses, | |||
384 | bool IsCombinedParallelRegion) { | |||
385 | if (!S) | |||
386 | return; | |||
387 | for (const CapturedStmt::Capture &C : S->captures()) { | |||
388 | if (C.capturesVariable() && !C.capturesVariableByCopy()) { | |||
389 | const ValueDecl *VD = C.getCapturedVar(); | |||
390 | bool SavedIsForCombinedParallelRegion = IsForCombinedParallelRegion; | |||
391 | if (IsCombinedParallelRegion) { | |||
392 | // Check if the variable is privatized in the combined construct and | |||
393 | // those private copies must be shared in the inner parallel | |||
394 | // directive. | |||
395 | IsForCombinedParallelRegion = false; | |||
396 | for (const OMPClause *C : Clauses) { | |||
397 | if (!isOpenMPPrivate(C->getClauseKind()) || | |||
398 | C->getClauseKind() == OMPC_reduction || | |||
399 | C->getClauseKind() == OMPC_linear || | |||
400 | C->getClauseKind() == OMPC_private) | |||
401 | continue; | |||
402 | ArrayRef<const Expr *> Vars; | |||
403 | if (const auto *PC = dyn_cast<OMPFirstprivateClause>(C)) | |||
404 | Vars = PC->getVarRefs(); | |||
405 | else if (const auto *PC = dyn_cast<OMPLastprivateClause>(C)) | |||
406 | Vars = PC->getVarRefs(); | |||
407 | else | |||
408 | llvm_unreachable("Unexpected clause.")::llvm::llvm_unreachable_internal("Unexpected clause.", "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 408); | |||
409 | for (const auto *E : Vars) { | |||
410 | const Decl *D = | |||
411 | cast<DeclRefExpr>(E)->getDecl()->getCanonicalDecl(); | |||
412 | if (D == VD->getCanonicalDecl()) { | |||
413 | IsForCombinedParallelRegion = true; | |||
414 | break; | |||
415 | } | |||
416 | } | |||
417 | if (IsForCombinedParallelRegion) | |||
418 | break; | |||
419 | } | |||
420 | } | |||
421 | markAsEscaped(VD); | |||
422 | if (isa<OMPCapturedExprDecl>(VD)) | |||
423 | VisitValueDecl(VD); | |||
424 | IsForCombinedParallelRegion = SavedIsForCombinedParallelRegion; | |||
425 | } | |||
426 | } | |||
427 | } | |||
428 | ||||
429 | void buildRecordForGlobalizedVars(bool IsInTTDRegion) { | |||
430 | assert(!GlobalizedRD &&((!GlobalizedRD && "Record for globalized variables is built already." ) ? static_cast<void> (0) : __assert_fail ("!GlobalizedRD && \"Record for globalized variables is built already.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 431, __PRETTY_FUNCTION__)) | |||
431 | "Record for globalized variables is built already.")((!GlobalizedRD && "Record for globalized variables is built already." ) ? static_cast<void> (0) : __assert_fail ("!GlobalizedRD && \"Record for globalized variables is built already.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 431, __PRETTY_FUNCTION__)); | |||
432 | ArrayRef<const ValueDecl *> EscapedDeclsForParallel, EscapedDeclsForTeams; | |||
433 | if (IsInTTDRegion) | |||
434 | EscapedDeclsForTeams = EscapedDecls.getArrayRef(); | |||
435 | else | |||
436 | EscapedDeclsForParallel = EscapedDecls.getArrayRef(); | |||
437 | GlobalizedRD = ::buildRecordForGlobalizedVars( | |||
438 | CGF.getContext(), EscapedDeclsForParallel, EscapedDeclsForTeams, | |||
439 | MappedDeclsFields, WarpSize); | |||
440 | } | |||
441 | ||||
442 | public: | |||
443 | CheckVarsEscapingDeclContext(CodeGenFunction &CGF, | |||
444 | ArrayRef<const ValueDecl *> TeamsReductions) | |||
445 | : CGF(CGF), EscapedDecls(TeamsReductions.begin(), TeamsReductions.end()) { | |||
446 | } | |||
447 | virtual ~CheckVarsEscapingDeclContext() = default; | |||
448 | void VisitDeclStmt(const DeclStmt *S) { | |||
449 | if (!S) | |||
450 | return; | |||
451 | for (const Decl *D : S->decls()) | |||
452 | if (const auto *VD = dyn_cast_or_null<ValueDecl>(D)) | |||
453 | VisitValueDecl(VD); | |||
454 | } | |||
455 | void VisitOMPExecutableDirective(const OMPExecutableDirective *D) { | |||
456 | if (!D) | |||
457 | return; | |||
458 | if (!D->hasAssociatedStmt()) | |||
459 | return; | |||
460 | if (const auto *S = | |||
461 | dyn_cast_or_null<CapturedStmt>(D->getAssociatedStmt())) { | |||
462 | // Do not analyze directives that do not actually require capturing, | |||
463 | // like `omp for` or `omp simd` directives. | |||
464 | llvm::SmallVector<OpenMPDirectiveKind, 4> CaptureRegions; | |||
465 | getOpenMPCaptureRegions(CaptureRegions, D->getDirectiveKind()); | |||
466 | if (CaptureRegions.size() == 1 && CaptureRegions.back() == OMPD_unknown) { | |||
467 | VisitStmt(S->getCapturedStmt()); | |||
468 | return; | |||
469 | } | |||
470 | VisitOpenMPCapturedStmt( | |||
471 | S, D->clauses(), | |||
472 | CaptureRegions.back() == OMPD_parallel && | |||
473 | isOpenMPDistributeDirective(D->getDirectiveKind())); | |||
474 | } | |||
475 | } | |||
476 | void VisitCapturedStmt(const CapturedStmt *S) { | |||
477 | if (!S) | |||
478 | return; | |||
479 | for (const CapturedStmt::Capture &C : S->captures()) { | |||
480 | if (C.capturesVariable() && !C.capturesVariableByCopy()) { | |||
481 | const ValueDecl *VD = C.getCapturedVar(); | |||
482 | markAsEscaped(VD); | |||
483 | if (isa<OMPCapturedExprDecl>(VD)) | |||
484 | VisitValueDecl(VD); | |||
485 | } | |||
486 | } | |||
487 | } | |||
488 | void VisitLambdaExpr(const LambdaExpr *E) { | |||
489 | if (!E) | |||
490 | return; | |||
491 | for (const LambdaCapture &C : E->captures()) { | |||
492 | if (C.capturesVariable()) { | |||
493 | if (C.getCaptureKind() == LCK_ByRef) { | |||
494 | const ValueDecl *VD = C.getCapturedVar(); | |||
495 | markAsEscaped(VD); | |||
496 | if (E->isInitCapture(&C) || isa<OMPCapturedExprDecl>(VD)) | |||
497 | VisitValueDecl(VD); | |||
498 | } | |||
499 | } | |||
500 | } | |||
501 | } | |||
502 | void VisitBlockExpr(const BlockExpr *E) { | |||
503 | if (!E) | |||
504 | return; | |||
505 | for (const BlockDecl::Capture &C : E->getBlockDecl()->captures()) { | |||
506 | if (C.isByRef()) { | |||
507 | const VarDecl *VD = C.getVariable(); | |||
508 | markAsEscaped(VD); | |||
509 | if (isa<OMPCapturedExprDecl>(VD) || VD->isInitCapture()) | |||
510 | VisitValueDecl(VD); | |||
511 | } | |||
512 | } | |||
513 | } | |||
514 | void VisitCallExpr(const CallExpr *E) { | |||
515 | if (!E) | |||
516 | return; | |||
517 | for (const Expr *Arg : E->arguments()) { | |||
518 | if (!Arg) | |||
519 | continue; | |||
520 | if (Arg->isLValue()) { | |||
521 | const bool SavedAllEscaped = AllEscaped; | |||
522 | AllEscaped = true; | |||
523 | Visit(Arg); | |||
524 | AllEscaped = SavedAllEscaped; | |||
525 | } else { | |||
526 | Visit(Arg); | |||
527 | } | |||
528 | } | |||
529 | Visit(E->getCallee()); | |||
530 | } | |||
531 | void VisitDeclRefExpr(const DeclRefExpr *E) { | |||
532 | if (!E) | |||
533 | return; | |||
534 | const ValueDecl *VD = E->getDecl(); | |||
535 | if (AllEscaped) | |||
536 | markAsEscaped(VD); | |||
537 | if (isa<OMPCapturedExprDecl>(VD)) | |||
538 | VisitValueDecl(VD); | |||
539 | else if (const auto *VarD = dyn_cast<VarDecl>(VD)) | |||
540 | if (VarD->isInitCapture()) | |||
541 | VisitValueDecl(VD); | |||
542 | } | |||
543 | void VisitUnaryOperator(const UnaryOperator *E) { | |||
544 | if (!E) | |||
545 | return; | |||
546 | if (E->getOpcode() == UO_AddrOf) { | |||
547 | const bool SavedAllEscaped = AllEscaped; | |||
548 | AllEscaped = true; | |||
549 | Visit(E->getSubExpr()); | |||
550 | AllEscaped = SavedAllEscaped; | |||
551 | } else { | |||
552 | Visit(E->getSubExpr()); | |||
553 | } | |||
554 | } | |||
555 | void VisitImplicitCastExpr(const ImplicitCastExpr *E) { | |||
556 | if (!E) | |||
557 | return; | |||
558 | if (E->getCastKind() == CK_ArrayToPointerDecay) { | |||
559 | const bool SavedAllEscaped = AllEscaped; | |||
560 | AllEscaped = true; | |||
561 | Visit(E->getSubExpr()); | |||
562 | AllEscaped = SavedAllEscaped; | |||
563 | } else { | |||
564 | Visit(E->getSubExpr()); | |||
565 | } | |||
566 | } | |||
567 | void VisitExpr(const Expr *E) { | |||
568 | if (!E) | |||
569 | return; | |||
570 | bool SavedAllEscaped = AllEscaped; | |||
571 | if (!E->isLValue()) | |||
572 | AllEscaped = false; | |||
573 | for (const Stmt *Child : E->children()) | |||
574 | if (Child) | |||
575 | Visit(Child); | |||
576 | AllEscaped = SavedAllEscaped; | |||
577 | } | |||
578 | void VisitStmt(const Stmt *S) { | |||
579 | if (!S) | |||
580 | return; | |||
581 | for (const Stmt *Child : S->children()) | |||
582 | if (Child) | |||
583 | Visit(Child); | |||
584 | } | |||
585 | ||||
586 | /// Returns the record that handles all the escaped local variables and used | |||
587 | /// instead of their original storage. | |||
588 | const RecordDecl *getGlobalizedRecord(bool IsInTTDRegion) { | |||
589 | if (!GlobalizedRD) | |||
590 | buildRecordForGlobalizedVars(IsInTTDRegion); | |||
591 | return GlobalizedRD; | |||
592 | } | |||
593 | ||||
594 | /// Returns the field in the globalized record for the escaped variable. | |||
595 | const FieldDecl *getFieldForGlobalizedVar(const ValueDecl *VD) const { | |||
596 | assert(GlobalizedRD &&((GlobalizedRD && "Record for globalized variables must be generated already." ) ? static_cast<void> (0) : __assert_fail ("GlobalizedRD && \"Record for globalized variables must be generated already.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 597, __PRETTY_FUNCTION__)) | |||
597 | "Record for globalized variables must be generated already.")((GlobalizedRD && "Record for globalized variables must be generated already." ) ? static_cast<void> (0) : __assert_fail ("GlobalizedRD && \"Record for globalized variables must be generated already.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 597, __PRETTY_FUNCTION__)); | |||
598 | auto I = MappedDeclsFields.find(VD); | |||
599 | if (I == MappedDeclsFields.end()) | |||
600 | return nullptr; | |||
601 | return I->getSecond(); | |||
602 | } | |||
603 | ||||
604 | /// Returns the list of the escaped local variables/parameters. | |||
605 | ArrayRef<const ValueDecl *> getEscapedDecls() const { | |||
606 | return EscapedDecls.getArrayRef(); | |||
607 | } | |||
608 | ||||
609 | /// Checks if the escaped local variable is actually a parameter passed by | |||
610 | /// value. | |||
611 | const llvm::SmallPtrSetImpl<const Decl *> &getEscapedParameters() const { | |||
612 | return EscapedParameters; | |||
613 | } | |||
614 | ||||
615 | /// Returns the list of the escaped variables with the variably modified | |||
616 | /// types. | |||
617 | ArrayRef<const ValueDecl *> getEscapedVariableLengthDecls() const { | |||
618 | return EscapedVariableLengthDecls.getArrayRef(); | |||
619 | } | |||
620 | }; | |||
621 | } // anonymous namespace | |||
622 | ||||
623 | /// Get the GPU warp size. | |||
624 | static llvm::Value *getNVPTXWarpSize(CodeGenFunction &CGF) { | |||
625 | return CGF.EmitRuntimeCall( | |||
626 | llvm::Intrinsic::getDeclaration( | |||
627 | &CGF.CGM.getModule(), llvm::Intrinsic::nvvm_read_ptx_sreg_warpsize), | |||
628 | "nvptx_warp_size"); | |||
629 | } | |||
630 | ||||
631 | /// Get the id of the current thread on the GPU. | |||
632 | static llvm::Value *getNVPTXThreadID(CodeGenFunction &CGF) { | |||
633 | return CGF.EmitRuntimeCall( | |||
634 | llvm::Intrinsic::getDeclaration( | |||
635 | &CGF.CGM.getModule(), llvm::Intrinsic::nvvm_read_ptx_sreg_tid_x), | |||
636 | "nvptx_tid"); | |||
637 | } | |||
638 | ||||
639 | /// Get the id of the warp in the block. | |||
640 | /// We assume that the warp size is 32, which is always the case | |||
641 | /// on the NVPTX device, to generate more efficient code. | |||
642 | static llvm::Value *getNVPTXWarpID(CodeGenFunction &CGF) { | |||
643 | CGBuilderTy &Bld = CGF.Builder; | |||
644 | return Bld.CreateAShr(getNVPTXThreadID(CGF), LaneIDBits, "nvptx_warp_id"); | |||
645 | } | |||
646 | ||||
647 | /// Get the id of the current lane in the Warp. | |||
648 | /// We assume that the warp size is 32, which is always the case | |||
649 | /// on the NVPTX device, to generate more efficient code. | |||
650 | static llvm::Value *getNVPTXLaneID(CodeGenFunction &CGF) { | |||
651 | CGBuilderTy &Bld = CGF.Builder; | |||
652 | return Bld.CreateAnd(getNVPTXThreadID(CGF), Bld.getInt32(LaneIDMask), | |||
653 | "nvptx_lane_id"); | |||
654 | } | |||
655 | ||||
656 | /// Get the maximum number of threads in a block of the GPU. | |||
657 | static llvm::Value *getNVPTXNumThreads(CodeGenFunction &CGF) { | |||
658 | return CGF.EmitRuntimeCall( | |||
659 | llvm::Intrinsic::getDeclaration( | |||
660 | &CGF.CGM.getModule(), llvm::Intrinsic::nvvm_read_ptx_sreg_ntid_x), | |||
661 | "nvptx_num_threads"); | |||
662 | } | |||
663 | ||||
664 | /// Get the value of the thread_limit clause in the teams directive. | |||
665 | /// For the 'generic' execution mode, the runtime encodes thread_limit in | |||
666 | /// the launch parameters, always starting thread_limit+warpSize threads per | |||
667 | /// CTA. The threads in the last warp are reserved for master execution. | |||
668 | /// For the 'spmd' execution mode, all threads in a CTA are part of the team. | |||
669 | static llvm::Value *getThreadLimit(CodeGenFunction &CGF, | |||
670 | bool IsInSPMDExecutionMode = false) { | |||
671 | CGBuilderTy &Bld = CGF.Builder; | |||
672 | return IsInSPMDExecutionMode | |||
673 | ? getNVPTXNumThreads(CGF) | |||
674 | : Bld.CreateNUWSub(getNVPTXNumThreads(CGF), getNVPTXWarpSize(CGF), | |||
675 | "thread_limit"); | |||
676 | } | |||
677 | ||||
678 | /// Get the thread id of the OMP master thread. | |||
679 | /// The master thread id is the first thread (lane) of the last warp in the | |||
680 | /// GPU block. Warp size is assumed to be some power of 2. | |||
681 | /// Thread id is 0 indexed. | |||
682 | /// E.g: If NumThreads is 33, master id is 32. | |||
683 | /// If NumThreads is 64, master id is 32. | |||
684 | /// If NumThreads is 1024, master id is 992. | |||
685 | static llvm::Value *getMasterThreadID(CodeGenFunction &CGF) { | |||
686 | CGBuilderTy &Bld = CGF.Builder; | |||
687 | llvm::Value *NumThreads = getNVPTXNumThreads(CGF); | |||
688 | ||||
689 | // We assume that the warp size is a power of 2. | |||
690 | llvm::Value *Mask = Bld.CreateNUWSub(getNVPTXWarpSize(CGF), Bld.getInt32(1)); | |||
691 | ||||
692 | return Bld.CreateAnd(Bld.CreateNUWSub(NumThreads, Bld.getInt32(1)), | |||
693 | Bld.CreateNot(Mask), "master_tid"); | |||
694 | } | |||
695 | ||||
696 | CGOpenMPRuntimeNVPTX::WorkerFunctionState::WorkerFunctionState( | |||
697 | CodeGenModule &CGM, SourceLocation Loc) | |||
698 | : WorkerFn(nullptr), CGFI(CGM.getTypes().arrangeNullaryFunction()), | |||
699 | Loc(Loc) { | |||
700 | createWorkerFunction(CGM); | |||
701 | } | |||
702 | ||||
703 | void CGOpenMPRuntimeNVPTX::WorkerFunctionState::createWorkerFunction( | |||
704 | CodeGenModule &CGM) { | |||
705 | // Create an worker function with no arguments. | |||
706 | ||||
707 | WorkerFn = llvm::Function::Create( | |||
708 | CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, | |||
709 | /*placeholder=*/"_worker", &CGM.getModule()); | |||
710 | CGM.SetInternalFunctionAttributes(GlobalDecl(), WorkerFn, CGFI); | |||
711 | WorkerFn->setDoesNotRecurse(); | |||
712 | } | |||
713 | ||||
714 | CGOpenMPRuntimeNVPTX::ExecutionMode | |||
715 | CGOpenMPRuntimeNVPTX::getExecutionMode() const { | |||
716 | return CurrentExecutionMode; | |||
717 | } | |||
718 | ||||
719 | static CGOpenMPRuntimeNVPTX::DataSharingMode | |||
720 | getDataSharingMode(CodeGenModule &CGM) { | |||
721 | return CGM.getLangOpts().OpenMPCUDAMode ? CGOpenMPRuntimeNVPTX::CUDA | |||
722 | : CGOpenMPRuntimeNVPTX::Generic; | |||
723 | } | |||
724 | ||||
725 | /// Check for inner (nested) SPMD construct, if any | |||
726 | static bool hasNestedSPMDDirective(ASTContext &Ctx, | |||
727 | const OMPExecutableDirective &D) { | |||
728 | const auto *CS = D.getInnermostCapturedStmt(); | |||
729 | const auto *Body = | |||
730 | CS->getCapturedStmt()->IgnoreContainers(/*IgnoreCaptured=*/true); | |||
731 | const Stmt *ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); | |||
732 | ||||
733 | if (const auto *NestedDir = | |||
734 | dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { | |||
735 | OpenMPDirectiveKind DKind = NestedDir->getDirectiveKind(); | |||
736 | switch (D.getDirectiveKind()) { | |||
737 | case OMPD_target: | |||
738 | if (isOpenMPParallelDirective(DKind)) | |||
739 | return true; | |||
740 | if (DKind == OMPD_teams) { | |||
741 | Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers( | |||
742 | /*IgnoreCaptured=*/true); | |||
743 | if (!Body) | |||
744 | return false; | |||
745 | ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); | |||
746 | if (const auto *NND = | |||
747 | dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { | |||
748 | DKind = NND->getDirectiveKind(); | |||
749 | if (isOpenMPParallelDirective(DKind)) | |||
750 | return true; | |||
751 | } | |||
752 | } | |||
753 | return false; | |||
754 | case OMPD_target_teams: | |||
755 | return isOpenMPParallelDirective(DKind); | |||
756 | case OMPD_target_simd: | |||
757 | case OMPD_target_parallel: | |||
758 | case OMPD_target_parallel_for: | |||
759 | case OMPD_target_parallel_for_simd: | |||
760 | case OMPD_target_teams_distribute: | |||
761 | case OMPD_target_teams_distribute_simd: | |||
762 | case OMPD_target_teams_distribute_parallel_for: | |||
763 | case OMPD_target_teams_distribute_parallel_for_simd: | |||
764 | case OMPD_parallel: | |||
765 | case OMPD_for: | |||
766 | case OMPD_parallel_for: | |||
767 | case OMPD_parallel_master: | |||
768 | case OMPD_parallel_sections: | |||
769 | case OMPD_for_simd: | |||
770 | case OMPD_parallel_for_simd: | |||
771 | case OMPD_cancel: | |||
772 | case OMPD_cancellation_point: | |||
773 | case OMPD_ordered: | |||
774 | case OMPD_threadprivate: | |||
775 | case OMPD_allocate: | |||
776 | case OMPD_task: | |||
777 | case OMPD_simd: | |||
778 | case OMPD_sections: | |||
779 | case OMPD_section: | |||
780 | case OMPD_single: | |||
781 | case OMPD_master: | |||
782 | case OMPD_critical: | |||
783 | case OMPD_taskyield: | |||
784 | case OMPD_barrier: | |||
785 | case OMPD_taskwait: | |||
786 | case OMPD_taskgroup: | |||
787 | case OMPD_atomic: | |||
788 | case OMPD_flush: | |||
789 | case OMPD_depobj: | |||
790 | case OMPD_teams: | |||
791 | case OMPD_target_data: | |||
792 | case OMPD_target_exit_data: | |||
793 | case OMPD_target_enter_data: | |||
794 | case OMPD_distribute: | |||
795 | case OMPD_distribute_simd: | |||
796 | case OMPD_distribute_parallel_for: | |||
797 | case OMPD_distribute_parallel_for_simd: | |||
798 | case OMPD_teams_distribute: | |||
799 | case OMPD_teams_distribute_simd: | |||
800 | case OMPD_teams_distribute_parallel_for: | |||
801 | case OMPD_teams_distribute_parallel_for_simd: | |||
802 | case OMPD_target_update: | |||
803 | case OMPD_declare_simd: | |||
804 | case OMPD_declare_variant: | |||
805 | case OMPD_declare_target: | |||
806 | case OMPD_end_declare_target: | |||
807 | case OMPD_declare_reduction: | |||
808 | case OMPD_declare_mapper: | |||
809 | case OMPD_taskloop: | |||
810 | case OMPD_taskloop_simd: | |||
811 | case OMPD_master_taskloop: | |||
812 | case OMPD_master_taskloop_simd: | |||
813 | case OMPD_parallel_master_taskloop: | |||
814 | case OMPD_parallel_master_taskloop_simd: | |||
815 | case OMPD_requires: | |||
816 | case OMPD_unknown: | |||
817 | llvm_unreachable("Unexpected directive.")::llvm::llvm_unreachable_internal("Unexpected directive.", "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 817); | |||
818 | } | |||
819 | } | |||
820 | ||||
821 | return false; | |||
822 | } | |||
823 | ||||
824 | static bool supportsSPMDExecutionMode(ASTContext &Ctx, | |||
825 | const OMPExecutableDirective &D) { | |||
826 | OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind(); | |||
827 | switch (DirectiveKind) { | |||
828 | case OMPD_target: | |||
829 | case OMPD_target_teams: | |||
830 | return hasNestedSPMDDirective(Ctx, D); | |||
831 | case OMPD_target_parallel: | |||
832 | case OMPD_target_parallel_for: | |||
833 | case OMPD_target_parallel_for_simd: | |||
834 | case OMPD_target_teams_distribute_parallel_for: | |||
835 | case OMPD_target_teams_distribute_parallel_for_simd: | |||
836 | case OMPD_target_simd: | |||
837 | case OMPD_target_teams_distribute_simd: | |||
838 | return true; | |||
839 | case OMPD_target_teams_distribute: | |||
840 | return false; | |||
841 | case OMPD_parallel: | |||
842 | case OMPD_for: | |||
843 | case OMPD_parallel_for: | |||
844 | case OMPD_parallel_master: | |||
845 | case OMPD_parallel_sections: | |||
846 | case OMPD_for_simd: | |||
847 | case OMPD_parallel_for_simd: | |||
848 | case OMPD_cancel: | |||
849 | case OMPD_cancellation_point: | |||
850 | case OMPD_ordered: | |||
851 | case OMPD_threadprivate: | |||
852 | case OMPD_allocate: | |||
853 | case OMPD_task: | |||
854 | case OMPD_simd: | |||
855 | case OMPD_sections: | |||
856 | case OMPD_section: | |||
857 | case OMPD_single: | |||
858 | case OMPD_master: | |||
859 | case OMPD_critical: | |||
860 | case OMPD_taskyield: | |||
861 | case OMPD_barrier: | |||
862 | case OMPD_taskwait: | |||
863 | case OMPD_taskgroup: | |||
864 | case OMPD_atomic: | |||
865 | case OMPD_flush: | |||
866 | case OMPD_depobj: | |||
867 | case OMPD_teams: | |||
868 | case OMPD_target_data: | |||
869 | case OMPD_target_exit_data: | |||
870 | case OMPD_target_enter_data: | |||
871 | case OMPD_distribute: | |||
872 | case OMPD_distribute_simd: | |||
873 | case OMPD_distribute_parallel_for: | |||
874 | case OMPD_distribute_parallel_for_simd: | |||
875 | case OMPD_teams_distribute: | |||
876 | case OMPD_teams_distribute_simd: | |||
877 | case OMPD_teams_distribute_parallel_for: | |||
878 | case OMPD_teams_distribute_parallel_for_simd: | |||
879 | case OMPD_target_update: | |||
880 | case OMPD_declare_simd: | |||
881 | case OMPD_declare_variant: | |||
882 | case OMPD_declare_target: | |||
883 | case OMPD_end_declare_target: | |||
884 | case OMPD_declare_reduction: | |||
885 | case OMPD_declare_mapper: | |||
886 | case OMPD_taskloop: | |||
887 | case OMPD_taskloop_simd: | |||
888 | case OMPD_master_taskloop: | |||
889 | case OMPD_master_taskloop_simd: | |||
890 | case OMPD_parallel_master_taskloop: | |||
891 | case OMPD_parallel_master_taskloop_simd: | |||
892 | case OMPD_requires: | |||
893 | case OMPD_unknown: | |||
894 | break; | |||
895 | } | |||
896 | llvm_unreachable(::llvm::llvm_unreachable_internal("Unknown programming model for OpenMP directive on NVPTX target." , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 897) | |||
897 | "Unknown programming model for OpenMP directive on NVPTX target.")::llvm::llvm_unreachable_internal("Unknown programming model for OpenMP directive on NVPTX target." , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 897); | |||
898 | } | |||
899 | ||||
900 | /// Check if the directive is loops based and has schedule clause at all or has | |||
901 | /// static scheduling. | |||
902 | static bool hasStaticScheduling(const OMPExecutableDirective &D) { | |||
903 | assert(isOpenMPWorksharingDirective(D.getDirectiveKind()) &&((isOpenMPWorksharingDirective(D.getDirectiveKind()) && isOpenMPLoopDirective(D.getDirectiveKind()) && "Expected loop-based directive." ) ? static_cast<void> (0) : __assert_fail ("isOpenMPWorksharingDirective(D.getDirectiveKind()) && isOpenMPLoopDirective(D.getDirectiveKind()) && \"Expected loop-based directive.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 905, __PRETTY_FUNCTION__)) | |||
904 | isOpenMPLoopDirective(D.getDirectiveKind()) &&((isOpenMPWorksharingDirective(D.getDirectiveKind()) && isOpenMPLoopDirective(D.getDirectiveKind()) && "Expected loop-based directive." ) ? static_cast<void> (0) : __assert_fail ("isOpenMPWorksharingDirective(D.getDirectiveKind()) && isOpenMPLoopDirective(D.getDirectiveKind()) && \"Expected loop-based directive.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 905, __PRETTY_FUNCTION__)) | |||
905 | "Expected loop-based directive.")((isOpenMPWorksharingDirective(D.getDirectiveKind()) && isOpenMPLoopDirective(D.getDirectiveKind()) && "Expected loop-based directive." ) ? static_cast<void> (0) : __assert_fail ("isOpenMPWorksharingDirective(D.getDirectiveKind()) && isOpenMPLoopDirective(D.getDirectiveKind()) && \"Expected loop-based directive.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 905, __PRETTY_FUNCTION__)); | |||
906 | return !D.hasClausesOfKind<OMPOrderedClause>() && | |||
907 | (!D.hasClausesOfKind<OMPScheduleClause>() || | |||
908 | llvm::any_of(D.getClausesOfKind<OMPScheduleClause>(), | |||
909 | [](const OMPScheduleClause *C) { | |||
910 | return C->getScheduleKind() == OMPC_SCHEDULE_static; | |||
911 | })); | |||
912 | } | |||
913 | ||||
914 | /// Check for inner (nested) lightweight runtime construct, if any | |||
915 | static bool hasNestedLightweightDirective(ASTContext &Ctx, | |||
916 | const OMPExecutableDirective &D) { | |||
917 | assert(supportsSPMDExecutionMode(Ctx, D) && "Expected SPMD mode directive.")((supportsSPMDExecutionMode(Ctx, D) && "Expected SPMD mode directive." ) ? static_cast<void> (0) : __assert_fail ("supportsSPMDExecutionMode(Ctx, D) && \"Expected SPMD mode directive.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 917, __PRETTY_FUNCTION__)); | |||
918 | const auto *CS = D.getInnermostCapturedStmt(); | |||
919 | const auto *Body = | |||
920 | CS->getCapturedStmt()->IgnoreContainers(/*IgnoreCaptured=*/true); | |||
921 | const Stmt *ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); | |||
922 | ||||
923 | if (const auto *NestedDir = | |||
924 | dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { | |||
925 | OpenMPDirectiveKind DKind = NestedDir->getDirectiveKind(); | |||
926 | switch (D.getDirectiveKind()) { | |||
927 | case OMPD_target: | |||
928 | if (isOpenMPParallelDirective(DKind) && | |||
929 | isOpenMPWorksharingDirective(DKind) && isOpenMPLoopDirective(DKind) && | |||
930 | hasStaticScheduling(*NestedDir)) | |||
931 | return true; | |||
932 | if (DKind == OMPD_teams_distribute_simd || DKind == OMPD_simd) | |||
933 | return true; | |||
934 | if (DKind == OMPD_parallel) { | |||
935 | Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers( | |||
936 | /*IgnoreCaptured=*/true); | |||
937 | if (!Body) | |||
938 | return false; | |||
939 | ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); | |||
940 | if (const auto *NND = | |||
941 | dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { | |||
942 | DKind = NND->getDirectiveKind(); | |||
943 | if (isOpenMPWorksharingDirective(DKind) && | |||
944 | isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NND)) | |||
945 | return true; | |||
946 | } | |||
947 | } else if (DKind == OMPD_teams) { | |||
948 | Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers( | |||
949 | /*IgnoreCaptured=*/true); | |||
950 | if (!Body) | |||
951 | return false; | |||
952 | ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); | |||
953 | if (const auto *NND = | |||
954 | dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { | |||
955 | DKind = NND->getDirectiveKind(); | |||
956 | if (isOpenMPParallelDirective(DKind) && | |||
957 | isOpenMPWorksharingDirective(DKind) && | |||
958 | isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NND)) | |||
959 | return true; | |||
960 | if (DKind == OMPD_parallel) { | |||
961 | Body = NND->getInnermostCapturedStmt()->IgnoreContainers( | |||
962 | /*IgnoreCaptured=*/true); | |||
963 | if (!Body) | |||
964 | return false; | |||
965 | ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); | |||
966 | if (const auto *NND = | |||
967 | dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { | |||
968 | DKind = NND->getDirectiveKind(); | |||
969 | if (isOpenMPWorksharingDirective(DKind) && | |||
970 | isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NND)) | |||
971 | return true; | |||
972 | } | |||
973 | } | |||
974 | } | |||
975 | } | |||
976 | return false; | |||
977 | case OMPD_target_teams: | |||
978 | if (isOpenMPParallelDirective(DKind) && | |||
979 | isOpenMPWorksharingDirective(DKind) && isOpenMPLoopDirective(DKind) && | |||
980 | hasStaticScheduling(*NestedDir)) | |||
981 | return true; | |||
982 | if (DKind == OMPD_distribute_simd || DKind == OMPD_simd) | |||
983 | return true; | |||
984 | if (DKind == OMPD_parallel) { | |||
985 | Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers( | |||
986 | /*IgnoreCaptured=*/true); | |||
987 | if (!Body) | |||
988 | return false; | |||
989 | ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); | |||
990 | if (const auto *NND = | |||
991 | dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { | |||
992 | DKind = NND->getDirectiveKind(); | |||
993 | if (isOpenMPWorksharingDirective(DKind) && | |||
994 | isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NND)) | |||
995 | return true; | |||
996 | } | |||
997 | } | |||
998 | return false; | |||
999 | case OMPD_target_parallel: | |||
1000 | if (DKind == OMPD_simd) | |||
1001 | return true; | |||
1002 | return isOpenMPWorksharingDirective(DKind) && | |||
1003 | isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NestedDir); | |||
1004 | case OMPD_target_teams_distribute: | |||
1005 | case OMPD_target_simd: | |||
1006 | case OMPD_target_parallel_for: | |||
1007 | case OMPD_target_parallel_for_simd: | |||
1008 | case OMPD_target_teams_distribute_simd: | |||
1009 | case OMPD_target_teams_distribute_parallel_for: | |||
1010 | case OMPD_target_teams_distribute_parallel_for_simd: | |||
1011 | case OMPD_parallel: | |||
1012 | case OMPD_for: | |||
1013 | case OMPD_parallel_for: | |||
1014 | case OMPD_parallel_master: | |||
1015 | case OMPD_parallel_sections: | |||
1016 | case OMPD_for_simd: | |||
1017 | case OMPD_parallel_for_simd: | |||
1018 | case OMPD_cancel: | |||
1019 | case OMPD_cancellation_point: | |||
1020 | case OMPD_ordered: | |||
1021 | case OMPD_threadprivate: | |||
1022 | case OMPD_allocate: | |||
1023 | case OMPD_task: | |||
1024 | case OMPD_simd: | |||
1025 | case OMPD_sections: | |||
1026 | case OMPD_section: | |||
1027 | case OMPD_single: | |||
1028 | case OMPD_master: | |||
1029 | case OMPD_critical: | |||
1030 | case OMPD_taskyield: | |||
1031 | case OMPD_barrier: | |||
1032 | case OMPD_taskwait: | |||
1033 | case OMPD_taskgroup: | |||
1034 | case OMPD_atomic: | |||
1035 | case OMPD_flush: | |||
1036 | case OMPD_depobj: | |||
1037 | case OMPD_teams: | |||
1038 | case OMPD_target_data: | |||
1039 | case OMPD_target_exit_data: | |||
1040 | case OMPD_target_enter_data: | |||
1041 | case OMPD_distribute: | |||
1042 | case OMPD_distribute_simd: | |||
1043 | case OMPD_distribute_parallel_for: | |||
1044 | case OMPD_distribute_parallel_for_simd: | |||
1045 | case OMPD_teams_distribute: | |||
1046 | case OMPD_teams_distribute_simd: | |||
1047 | case OMPD_teams_distribute_parallel_for: | |||
1048 | case OMPD_teams_distribute_parallel_for_simd: | |||
1049 | case OMPD_target_update: | |||
1050 | case OMPD_declare_simd: | |||
1051 | case OMPD_declare_variant: | |||
1052 | case OMPD_declare_target: | |||
1053 | case OMPD_end_declare_target: | |||
1054 | case OMPD_declare_reduction: | |||
1055 | case OMPD_declare_mapper: | |||
1056 | case OMPD_taskloop: | |||
1057 | case OMPD_taskloop_simd: | |||
1058 | case OMPD_master_taskloop: | |||
1059 | case OMPD_master_taskloop_simd: | |||
1060 | case OMPD_parallel_master_taskloop: | |||
1061 | case OMPD_parallel_master_taskloop_simd: | |||
1062 | case OMPD_requires: | |||
1063 | case OMPD_unknown: | |||
1064 | llvm_unreachable("Unexpected directive.")::llvm::llvm_unreachable_internal("Unexpected directive.", "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 1064); | |||
1065 | } | |||
1066 | } | |||
1067 | ||||
1068 | return false; | |||
1069 | } | |||
1070 | ||||
1071 | /// Checks if the construct supports lightweight runtime. It must be SPMD | |||
1072 | /// construct + inner loop-based construct with static scheduling. | |||
1073 | static bool supportsLightweightRuntime(ASTContext &Ctx, | |||
1074 | const OMPExecutableDirective &D) { | |||
1075 | if (!supportsSPMDExecutionMode(Ctx, D)) | |||
1076 | return false; | |||
1077 | OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind(); | |||
1078 | switch (DirectiveKind) { | |||
1079 | case OMPD_target: | |||
1080 | case OMPD_target_teams: | |||
1081 | case OMPD_target_parallel: | |||
1082 | return hasNestedLightweightDirective(Ctx, D); | |||
1083 | case OMPD_target_parallel_for: | |||
1084 | case OMPD_target_parallel_for_simd: | |||
1085 | case OMPD_target_teams_distribute_parallel_for: | |||
1086 | case OMPD_target_teams_distribute_parallel_for_simd: | |||
1087 | // (Last|First)-privates must be shared in parallel region. | |||
1088 | return hasStaticScheduling(D); | |||
1089 | case OMPD_target_simd: | |||
1090 | case OMPD_target_teams_distribute_simd: | |||
1091 | return true; | |||
1092 | case OMPD_target_teams_distribute: | |||
1093 | return false; | |||
1094 | case OMPD_parallel: | |||
1095 | case OMPD_for: | |||
1096 | case OMPD_parallel_for: | |||
1097 | case OMPD_parallel_master: | |||
1098 | case OMPD_parallel_sections: | |||
1099 | case OMPD_for_simd: | |||
1100 | case OMPD_parallel_for_simd: | |||
1101 | case OMPD_cancel: | |||
1102 | case OMPD_cancellation_point: | |||
1103 | case OMPD_ordered: | |||
1104 | case OMPD_threadprivate: | |||
1105 | case OMPD_allocate: | |||
1106 | case OMPD_task: | |||
1107 | case OMPD_simd: | |||
1108 | case OMPD_sections: | |||
1109 | case OMPD_section: | |||
1110 | case OMPD_single: | |||
1111 | case OMPD_master: | |||
1112 | case OMPD_critical: | |||
1113 | case OMPD_taskyield: | |||
1114 | case OMPD_barrier: | |||
1115 | case OMPD_taskwait: | |||
1116 | case OMPD_taskgroup: | |||
1117 | case OMPD_atomic: | |||
1118 | case OMPD_flush: | |||
1119 | case OMPD_depobj: | |||
1120 | case OMPD_teams: | |||
1121 | case OMPD_target_data: | |||
1122 | case OMPD_target_exit_data: | |||
1123 | case OMPD_target_enter_data: | |||
1124 | case OMPD_distribute: | |||
1125 | case OMPD_distribute_simd: | |||
1126 | case OMPD_distribute_parallel_for: | |||
1127 | case OMPD_distribute_parallel_for_simd: | |||
1128 | case OMPD_teams_distribute: | |||
1129 | case OMPD_teams_distribute_simd: | |||
1130 | case OMPD_teams_distribute_parallel_for: | |||
1131 | case OMPD_teams_distribute_parallel_for_simd: | |||
1132 | case OMPD_target_update: | |||
1133 | case OMPD_declare_simd: | |||
1134 | case OMPD_declare_variant: | |||
1135 | case OMPD_declare_target: | |||
1136 | case OMPD_end_declare_target: | |||
1137 | case OMPD_declare_reduction: | |||
1138 | case OMPD_declare_mapper: | |||
1139 | case OMPD_taskloop: | |||
1140 | case OMPD_taskloop_simd: | |||
1141 | case OMPD_master_taskloop: | |||
1142 | case OMPD_master_taskloop_simd: | |||
1143 | case OMPD_parallel_master_taskloop: | |||
1144 | case OMPD_parallel_master_taskloop_simd: | |||
1145 | case OMPD_requires: | |||
1146 | case OMPD_unknown: | |||
1147 | break; | |||
1148 | } | |||
1149 | llvm_unreachable(::llvm::llvm_unreachable_internal("Unknown programming model for OpenMP directive on NVPTX target." , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 1150) | |||
1150 | "Unknown programming model for OpenMP directive on NVPTX target.")::llvm::llvm_unreachable_internal("Unknown programming model for OpenMP directive on NVPTX target." , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 1150); | |||
1151 | } | |||
1152 | ||||
1153 | void CGOpenMPRuntimeNVPTX::emitNonSPMDKernel(const OMPExecutableDirective &D, | |||
1154 | StringRef ParentName, | |||
1155 | llvm::Function *&OutlinedFn, | |||
1156 | llvm::Constant *&OutlinedFnID, | |||
1157 | bool IsOffloadEntry, | |||
1158 | const RegionCodeGenTy &CodeGen) { | |||
1159 | ExecutionRuntimeModesRAII ModeRAII(CurrentExecutionMode); | |||
1160 | EntryFunctionState EST; | |||
1161 | WorkerFunctionState WST(CGM, D.getBeginLoc()); | |||
1162 | Work.clear(); | |||
1163 | WrapperFunctionsMap.clear(); | |||
1164 | ||||
1165 | // Emit target region as a standalone region. | |||
1166 | class NVPTXPrePostActionTy : public PrePostActionTy { | |||
1167 | CGOpenMPRuntimeNVPTX::EntryFunctionState &EST; | |||
1168 | CGOpenMPRuntimeNVPTX::WorkerFunctionState &WST; | |||
1169 | ||||
1170 | public: | |||
1171 | NVPTXPrePostActionTy(CGOpenMPRuntimeNVPTX::EntryFunctionState &EST, | |||
1172 | CGOpenMPRuntimeNVPTX::WorkerFunctionState &WST) | |||
1173 | : EST(EST), WST(WST) {} | |||
1174 | void Enter(CodeGenFunction &CGF) override { | |||
1175 | auto &RT = | |||
1176 | static_cast<CGOpenMPRuntimeNVPTX &>(CGF.CGM.getOpenMPRuntime()); | |||
1177 | RT.emitNonSPMDEntryHeader(CGF, EST, WST); | |||
1178 | // Skip target region initialization. | |||
1179 | RT.setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true); | |||
1180 | } | |||
1181 | void Exit(CodeGenFunction &CGF) override { | |||
1182 | auto &RT = | |||
1183 | static_cast<CGOpenMPRuntimeNVPTX &>(CGF.CGM.getOpenMPRuntime()); | |||
1184 | RT.clearLocThreadIdInsertPt(CGF); | |||
1185 | RT.emitNonSPMDEntryFooter(CGF, EST); | |||
1186 | } | |||
1187 | } Action(EST, WST); | |||
1188 | CodeGen.setAction(Action); | |||
1189 | IsInTTDRegion = true; | |||
1190 | // Reserve place for the globalized memory. | |||
1191 | GlobalizedRecords.emplace_back(); | |||
1192 | if (!KernelStaticGlobalized) { | |||
1193 | KernelStaticGlobalized = new llvm::GlobalVariable( | |||
1194 | CGM.getModule(), CGM.VoidPtrTy, /*isConstant=*/false, | |||
1195 | llvm::GlobalValue::InternalLinkage, | |||
1196 | llvm::ConstantPointerNull::get(CGM.VoidPtrTy), | |||
1197 | "_openmp_kernel_static_glob_rd$ptr", /*InsertBefore=*/nullptr, | |||
1198 | llvm::GlobalValue::NotThreadLocal, | |||
1199 | CGM.getContext().getTargetAddressSpace(LangAS::cuda_shared)); | |||
1200 | } | |||
1201 | emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID, | |||
1202 | IsOffloadEntry, CodeGen); | |||
1203 | IsInTTDRegion = false; | |||
1204 | ||||
1205 | // Now change the name of the worker function to correspond to this target | |||
1206 | // region's entry function. | |||
1207 | WST.WorkerFn->setName(Twine(OutlinedFn->getName(), "_worker")); | |||
1208 | ||||
1209 | // Create the worker function | |||
1210 | emitWorkerFunction(WST); | |||
1211 | } | |||
1212 | ||||
1213 | // Setup NVPTX threads for master-worker OpenMP scheme. | |||
1214 | void CGOpenMPRuntimeNVPTX::emitNonSPMDEntryHeader(CodeGenFunction &CGF, | |||
1215 | EntryFunctionState &EST, | |||
1216 | WorkerFunctionState &WST) { | |||
1217 | CGBuilderTy &Bld = CGF.Builder; | |||
1218 | ||||
1219 | llvm::BasicBlock *WorkerBB = CGF.createBasicBlock(".worker"); | |||
1220 | llvm::BasicBlock *MasterCheckBB = CGF.createBasicBlock(".mastercheck"); | |||
1221 | llvm::BasicBlock *MasterBB = CGF.createBasicBlock(".master"); | |||
1222 | EST.ExitBB = CGF.createBasicBlock(".exit"); | |||
1223 | ||||
1224 | llvm::Value *IsWorker = | |||
1225 | Bld.CreateICmpULT(getNVPTXThreadID(CGF), getThreadLimit(CGF)); | |||
1226 | Bld.CreateCondBr(IsWorker, WorkerBB, MasterCheckBB); | |||
1227 | ||||
1228 | CGF.EmitBlock(WorkerBB); | |||
1229 | emitCall(CGF, WST.Loc, WST.WorkerFn); | |||
1230 | CGF.EmitBranch(EST.ExitBB); | |||
1231 | ||||
1232 | CGF.EmitBlock(MasterCheckBB); | |||
1233 | llvm::Value *IsMaster = | |||
1234 | Bld.CreateICmpEQ(getNVPTXThreadID(CGF), getMasterThreadID(CGF)); | |||
1235 | Bld.CreateCondBr(IsMaster, MasterBB, EST.ExitBB); | |||
1236 | ||||
1237 | CGF.EmitBlock(MasterBB); | |||
1238 | IsInTargetMasterThreadRegion = true; | |||
1239 | // SEQUENTIAL (MASTER) REGION START | |||
1240 | // First action in sequential region: | |||
1241 | // Initialize the state of the OpenMP runtime library on the GPU. | |||
1242 | // TODO: Optimize runtime initialization and pass in correct value. | |||
1243 | llvm::Value *Args[] = {getThreadLimit(CGF), | |||
1244 | Bld.getInt16(/*RequiresOMPRuntime=*/1)}; | |||
1245 | CGF.EmitRuntimeCall( | |||
1246 | createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_kernel_init), Args); | |||
1247 | ||||
1248 | // For data sharing, we need to initialize the stack. | |||
1249 | CGF.EmitRuntimeCall( | |||
1250 | createNVPTXRuntimeFunction( | |||
1251 | OMPRTL_NVPTX__kmpc_data_sharing_init_stack)); | |||
1252 | ||||
1253 | emitGenericVarsProlog(CGF, WST.Loc); | |||
1254 | } | |||
1255 | ||||
1256 | void CGOpenMPRuntimeNVPTX::emitNonSPMDEntryFooter(CodeGenFunction &CGF, | |||
1257 | EntryFunctionState &EST) { | |||
1258 | IsInTargetMasterThreadRegion = false; | |||
1259 | if (!CGF.HaveInsertPoint()) | |||
1260 | return; | |||
1261 | ||||
1262 | emitGenericVarsEpilog(CGF); | |||
1263 | ||||
1264 | if (!EST.ExitBB) | |||
1265 | EST.ExitBB = CGF.createBasicBlock(".exit"); | |||
1266 | ||||
1267 | llvm::BasicBlock *TerminateBB = CGF.createBasicBlock(".termination.notifier"); | |||
1268 | CGF.EmitBranch(TerminateBB); | |||
1269 | ||||
1270 | CGF.EmitBlock(TerminateBB); | |||
1271 | // Signal termination condition. | |||
1272 | // TODO: Optimize runtime initialization and pass in correct value. | |||
1273 | llvm::Value *Args[] = {CGF.Builder.getInt16(/*IsOMPRuntimeInitialized=*/1)}; | |||
1274 | CGF.EmitRuntimeCall( | |||
1275 | createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_kernel_deinit), Args); | |||
1276 | // Barrier to terminate worker threads. | |||
1277 | syncCTAThreads(CGF); | |||
1278 | // Master thread jumps to exit point. | |||
1279 | CGF.EmitBranch(EST.ExitBB); | |||
1280 | ||||
1281 | CGF.EmitBlock(EST.ExitBB); | |||
1282 | EST.ExitBB = nullptr; | |||
1283 | } | |||
1284 | ||||
1285 | void CGOpenMPRuntimeNVPTX::emitSPMDKernel(const OMPExecutableDirective &D, | |||
1286 | StringRef ParentName, | |||
1287 | llvm::Function *&OutlinedFn, | |||
1288 | llvm::Constant *&OutlinedFnID, | |||
1289 | bool IsOffloadEntry, | |||
1290 | const RegionCodeGenTy &CodeGen) { | |||
1291 | ExecutionRuntimeModesRAII ModeRAII( | |||
1292 | CurrentExecutionMode, RequiresFullRuntime, | |||
1293 | CGM.getLangOpts().OpenMPCUDAForceFullRuntime || | |||
1294 | !supportsLightweightRuntime(CGM.getContext(), D)); | |||
1295 | EntryFunctionState EST; | |||
1296 | ||||
1297 | // Emit target region as a standalone region. | |||
1298 | class NVPTXPrePostActionTy : public PrePostActionTy { | |||
1299 | CGOpenMPRuntimeNVPTX &RT; | |||
1300 | CGOpenMPRuntimeNVPTX::EntryFunctionState &EST; | |||
1301 | const OMPExecutableDirective &D; | |||
1302 | ||||
1303 | public: | |||
1304 | NVPTXPrePostActionTy(CGOpenMPRuntimeNVPTX &RT, | |||
1305 | CGOpenMPRuntimeNVPTX::EntryFunctionState &EST, | |||
1306 | const OMPExecutableDirective &D) | |||
1307 | : RT(RT), EST(EST), D(D) {} | |||
1308 | void Enter(CodeGenFunction &CGF) override { | |||
1309 | RT.emitSPMDEntryHeader(CGF, EST, D); | |||
1310 | // Skip target region initialization. | |||
1311 | RT.setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true); | |||
1312 | } | |||
1313 | void Exit(CodeGenFunction &CGF) override { | |||
1314 | RT.clearLocThreadIdInsertPt(CGF); | |||
1315 | RT.emitSPMDEntryFooter(CGF, EST); | |||
1316 | } | |||
1317 | } Action(*this, EST, D); | |||
1318 | CodeGen.setAction(Action); | |||
1319 | IsInTTDRegion = true; | |||
1320 | // Reserve place for the globalized memory. | |||
1321 | GlobalizedRecords.emplace_back(); | |||
1322 | if (!KernelStaticGlobalized) { | |||
1323 | KernelStaticGlobalized = new llvm::GlobalVariable( | |||
1324 | CGM.getModule(), CGM.VoidPtrTy, /*isConstant=*/false, | |||
1325 | llvm::GlobalValue::InternalLinkage, | |||
1326 | llvm::ConstantPointerNull::get(CGM.VoidPtrTy), | |||
1327 | "_openmp_kernel_static_glob_rd$ptr", /*InsertBefore=*/nullptr, | |||
1328 | llvm::GlobalValue::NotThreadLocal, | |||
1329 | CGM.getContext().getTargetAddressSpace(LangAS::cuda_shared)); | |||
1330 | } | |||
1331 | emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID, | |||
1332 | IsOffloadEntry, CodeGen); | |||
1333 | IsInTTDRegion = false; | |||
1334 | } | |||
1335 | ||||
1336 | void CGOpenMPRuntimeNVPTX::emitSPMDEntryHeader( | |||
1337 | CodeGenFunction &CGF, EntryFunctionState &EST, | |||
1338 | const OMPExecutableDirective &D) { | |||
1339 | CGBuilderTy &Bld = CGF.Builder; | |||
1340 | ||||
1341 | // Setup BBs in entry function. | |||
1342 | llvm::BasicBlock *ExecuteBB = CGF.createBasicBlock(".execute"); | |||
1343 | EST.ExitBB = CGF.createBasicBlock(".exit"); | |||
1344 | ||||
1345 | llvm::Value *Args[] = {getThreadLimit(CGF, /*IsInSPMDExecutionMode=*/true), | |||
1346 | /*RequiresOMPRuntime=*/ | |||
1347 | Bld.getInt16(RequiresFullRuntime ? 1 : 0), | |||
1348 | /*RequiresDataSharing=*/Bld.getInt16(0)}; | |||
1349 | CGF.EmitRuntimeCall( | |||
1350 | createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_spmd_kernel_init), Args); | |||
1351 | ||||
1352 | if (RequiresFullRuntime) { | |||
1353 | // For data sharing, we need to initialize the stack. | |||
1354 | CGF.EmitRuntimeCall(createNVPTXRuntimeFunction( | |||
1355 | OMPRTL_NVPTX__kmpc_data_sharing_init_stack_spmd)); | |||
1356 | } | |||
1357 | ||||
1358 | CGF.EmitBranch(ExecuteBB); | |||
1359 | ||||
1360 | CGF.EmitBlock(ExecuteBB); | |||
1361 | ||||
1362 | IsInTargetMasterThreadRegion = true; | |||
1363 | } | |||
1364 | ||||
1365 | void CGOpenMPRuntimeNVPTX::emitSPMDEntryFooter(CodeGenFunction &CGF, | |||
1366 | EntryFunctionState &EST) { | |||
1367 | IsInTargetMasterThreadRegion = false; | |||
1368 | if (!CGF.HaveInsertPoint()) | |||
1369 | return; | |||
1370 | ||||
1371 | if (!EST.ExitBB) | |||
1372 | EST.ExitBB = CGF.createBasicBlock(".exit"); | |||
1373 | ||||
1374 | llvm::BasicBlock *OMPDeInitBB = CGF.createBasicBlock(".omp.deinit"); | |||
1375 | CGF.EmitBranch(OMPDeInitBB); | |||
1376 | ||||
1377 | CGF.EmitBlock(OMPDeInitBB); | |||
1378 | // DeInitialize the OMP state in the runtime; called by all active threads. | |||
1379 | llvm::Value *Args[] = {/*RequiresOMPRuntime=*/ | |||
1380 | CGF.Builder.getInt16(RequiresFullRuntime ? 1 : 0)}; | |||
1381 | CGF.EmitRuntimeCall( | |||
1382 | createNVPTXRuntimeFunction( | |||
1383 | OMPRTL_NVPTX__kmpc_spmd_kernel_deinit_v2), Args); | |||
1384 | CGF.EmitBranch(EST.ExitBB); | |||
1385 | ||||
1386 | CGF.EmitBlock(EST.ExitBB); | |||
1387 | EST.ExitBB = nullptr; | |||
1388 | } | |||
1389 | ||||
1390 | // Create a unique global variable to indicate the execution mode of this target | |||
1391 | // region. The execution mode is either 'generic', or 'spmd' depending on the | |||
1392 | // target directive. This variable is picked up by the offload library to setup | |||
1393 | // the device appropriately before kernel launch. If the execution mode is | |||
1394 | // 'generic', the runtime reserves one warp for the master, otherwise, all | |||
1395 | // warps participate in parallel work. | |||
1396 | static void setPropertyExecutionMode(CodeGenModule &CGM, StringRef Name, | |||
1397 | bool Mode) { | |||
1398 | auto *GVMode = | |||
1399 | new llvm::GlobalVariable(CGM.getModule(), CGM.Int8Ty, /*isConstant=*/true, | |||
1400 | llvm::GlobalValue::WeakAnyLinkage, | |||
1401 | llvm::ConstantInt::get(CGM.Int8Ty, Mode ? 0 : 1), | |||
1402 | Twine(Name, "_exec_mode")); | |||
1403 | CGM.addCompilerUsedGlobal(GVMode); | |||
1404 | } | |||
1405 | ||||
1406 | void CGOpenMPRuntimeNVPTX::emitWorkerFunction(WorkerFunctionState &WST) { | |||
1407 | ASTContext &Ctx = CGM.getContext(); | |||
1408 | ||||
1409 | CodeGenFunction CGF(CGM, /*suppressNewContext=*/true); | |||
1410 | CGF.StartFunction(GlobalDecl(), Ctx.VoidTy, WST.WorkerFn, WST.CGFI, {}, | |||
1411 | WST.Loc, WST.Loc); | |||
1412 | emitWorkerLoop(CGF, WST); | |||
1413 | CGF.FinishFunction(); | |||
1414 | } | |||
1415 | ||||
1416 | void CGOpenMPRuntimeNVPTX::emitWorkerLoop(CodeGenFunction &CGF, | |||
1417 | WorkerFunctionState &WST) { | |||
1418 | // | |||
1419 | // The workers enter this loop and wait for parallel work from the master. | |||
1420 | // When the master encounters a parallel region it sets up the work + variable | |||
1421 | // arguments, and wakes up the workers. The workers first check to see if | |||
1422 | // they are required for the parallel region, i.e., within the # of requested | |||
1423 | // parallel threads. The activated workers load the variable arguments and | |||
1424 | // execute the parallel work. | |||
1425 | // | |||
1426 | ||||
1427 | CGBuilderTy &Bld = CGF.Builder; | |||
1428 | ||||
1429 | llvm::BasicBlock *AwaitBB = CGF.createBasicBlock(".await.work"); | |||
1430 | llvm::BasicBlock *SelectWorkersBB = CGF.createBasicBlock(".select.workers"); | |||
1431 | llvm::BasicBlock *ExecuteBB = CGF.createBasicBlock(".execute.parallel"); | |||
1432 | llvm::BasicBlock *TerminateBB = CGF.createBasicBlock(".terminate.parallel"); | |||
1433 | llvm::BasicBlock *BarrierBB = CGF.createBasicBlock(".barrier.parallel"); | |||
1434 | llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".exit"); | |||
1435 | ||||
1436 | CGF.EmitBranch(AwaitBB); | |||
1437 | ||||
1438 | // Workers wait for work from master. | |||
1439 | CGF.EmitBlock(AwaitBB); | |||
1440 | // Wait for parallel work | |||
1441 | syncCTAThreads(CGF); | |||
1442 | ||||
1443 | Address WorkFn = | |||
1444 | CGF.CreateDefaultAlignTempAlloca(CGF.Int8PtrTy, /*Name=*/"work_fn"); | |||
1445 | Address ExecStatus = | |||
1446 | CGF.CreateDefaultAlignTempAlloca(CGF.Int8Ty, /*Name=*/"exec_status"); | |||
1447 | CGF.InitTempAlloca(ExecStatus, Bld.getInt8(/*C=*/0)); | |||
1448 | CGF.InitTempAlloca(WorkFn, llvm::Constant::getNullValue(CGF.Int8PtrTy)); | |||
1449 | ||||
1450 | // TODO: Optimize runtime initialization and pass in correct value. | |||
1451 | llvm::Value *Args[] = {WorkFn.getPointer(), | |||
1452 | /*RequiresOMPRuntime=*/Bld.getInt16(1)}; | |||
1453 | llvm::Value *Ret = CGF.EmitRuntimeCall( | |||
1454 | createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_kernel_parallel), Args); | |||
1455 | Bld.CreateStore(Bld.CreateZExt(Ret, CGF.Int8Ty), ExecStatus); | |||
1456 | ||||
1457 | // On termination condition (workid == 0), exit loop. | |||
1458 | llvm::Value *WorkID = Bld.CreateLoad(WorkFn); | |||
1459 | llvm::Value *ShouldTerminate = Bld.CreateIsNull(WorkID, "should_terminate"); | |||
1460 | Bld.CreateCondBr(ShouldTerminate, ExitBB, SelectWorkersBB); | |||
1461 | ||||
1462 | // Activate requested workers. | |||
1463 | CGF.EmitBlock(SelectWorkersBB); | |||
1464 | llvm::Value *IsActive = | |||
1465 | Bld.CreateIsNotNull(Bld.CreateLoad(ExecStatus), "is_active"); | |||
1466 | Bld.CreateCondBr(IsActive, ExecuteBB, BarrierBB); | |||
1467 | ||||
1468 | // Signal start of parallel region. | |||
1469 | CGF.EmitBlock(ExecuteBB); | |||
1470 | // Skip initialization. | |||
1471 | setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true); | |||
1472 | ||||
1473 | // Process work items: outlined parallel functions. | |||
1474 | for (llvm::Function *W : Work) { | |||
1475 | // Try to match this outlined function. | |||
1476 | llvm::Value *ID = Bld.CreatePointerBitCastOrAddrSpaceCast(W, CGM.Int8PtrTy); | |||
1477 | ||||
1478 | llvm::Value *WorkFnMatch = | |||
1479 | Bld.CreateICmpEQ(Bld.CreateLoad(WorkFn), ID, "work_match"); | |||
1480 | ||||
1481 | llvm::BasicBlock *ExecuteFNBB = CGF.createBasicBlock(".execute.fn"); | |||
1482 | llvm::BasicBlock *CheckNextBB = CGF.createBasicBlock(".check.next"); | |||
1483 | Bld.CreateCondBr(WorkFnMatch, ExecuteFNBB, CheckNextBB); | |||
1484 | ||||
1485 | // Execute this outlined function. | |||
1486 | CGF.EmitBlock(ExecuteFNBB); | |||
1487 | ||||
1488 | // Insert call to work function via shared wrapper. The shared | |||
1489 | // wrapper takes two arguments: | |||
1490 | // - the parallelism level; | |||
1491 | // - the thread ID; | |||
1492 | emitCall(CGF, WST.Loc, W, | |||
1493 | {Bld.getInt16(/*ParallelLevel=*/0), getThreadID(CGF, WST.Loc)}); | |||
1494 | ||||
1495 | // Go to end of parallel region. | |||
1496 | CGF.EmitBranch(TerminateBB); | |||
1497 | ||||
1498 | CGF.EmitBlock(CheckNextBB); | |||
1499 | } | |||
1500 | // Default case: call to outlined function through pointer if the target | |||
1501 | // region makes a declare target call that may contain an orphaned parallel | |||
1502 | // directive. | |||
1503 | auto *ParallelFnTy = | |||
1504 | llvm::FunctionType::get(CGM.VoidTy, {CGM.Int16Ty, CGM.Int32Ty}, | |||
1505 | /*isVarArg=*/false); | |||
1506 | llvm::Value *WorkFnCast = | |||
1507 | Bld.CreateBitCast(WorkID, ParallelFnTy->getPointerTo()); | |||
1508 | // Insert call to work function via shared wrapper. The shared | |||
1509 | // wrapper takes two arguments: | |||
1510 | // - the parallelism level; | |||
1511 | // - the thread ID; | |||
1512 | emitCall(CGF, WST.Loc, {ParallelFnTy, WorkFnCast}, | |||
1513 | {Bld.getInt16(/*ParallelLevel=*/0), getThreadID(CGF, WST.Loc)}); | |||
1514 | // Go to end of parallel region. | |||
1515 | CGF.EmitBranch(TerminateBB); | |||
1516 | ||||
1517 | // Signal end of parallel region. | |||
1518 | CGF.EmitBlock(TerminateBB); | |||
1519 | CGF.EmitRuntimeCall( | |||
1520 | createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_kernel_end_parallel), | |||
1521 | llvm::None); | |||
1522 | CGF.EmitBranch(BarrierBB); | |||
1523 | ||||
1524 | // All active and inactive workers wait at a barrier after parallel region. | |||
1525 | CGF.EmitBlock(BarrierBB); | |||
1526 | // Barrier after parallel region. | |||
1527 | syncCTAThreads(CGF); | |||
1528 | CGF.EmitBranch(AwaitBB); | |||
1529 | ||||
1530 | // Exit target region. | |||
1531 | CGF.EmitBlock(ExitBB); | |||
1532 | // Skip initialization. | |||
1533 | clearLocThreadIdInsertPt(CGF); | |||
1534 | } | |||
1535 | ||||
1536 | /// Returns specified OpenMP runtime function for the current OpenMP | |||
1537 | /// implementation. Specialized for the NVPTX device. | |||
1538 | /// \param Function OpenMP runtime function. | |||
1539 | /// \return Specified function. | |||
1540 | llvm::FunctionCallee | |||
1541 | CGOpenMPRuntimeNVPTX::createNVPTXRuntimeFunction(unsigned Function) { | |||
1542 | llvm::FunctionCallee RTLFn = nullptr; | |||
1543 | switch (static_cast<OpenMPRTLFunctionNVPTX>(Function)) { | |||
1544 | case OMPRTL_NVPTX__kmpc_kernel_init: { | |||
1545 | // Build void __kmpc_kernel_init(kmp_int32 thread_limit, int16_t | |||
1546 | // RequiresOMPRuntime); | |||
1547 | llvm::Type *TypeParams[] = {CGM.Int32Ty, CGM.Int16Ty}; | |||
1548 | auto *FnTy = | |||
1549 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); | |||
1550 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_kernel_init"); | |||
1551 | break; | |||
1552 | } | |||
1553 | case OMPRTL_NVPTX__kmpc_kernel_deinit: { | |||
1554 | // Build void __kmpc_kernel_deinit(int16_t IsOMPRuntimeInitialized); | |||
1555 | llvm::Type *TypeParams[] = {CGM.Int16Ty}; | |||
1556 | auto *FnTy = | |||
1557 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); | |||
1558 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_kernel_deinit"); | |||
1559 | break; | |||
1560 | } | |||
1561 | case OMPRTL_NVPTX__kmpc_spmd_kernel_init: { | |||
1562 | // Build void __kmpc_spmd_kernel_init(kmp_int32 thread_limit, | |||
1563 | // int16_t RequiresOMPRuntime, int16_t RequiresDataSharing); | |||
1564 | llvm::Type *TypeParams[] = {CGM.Int32Ty, CGM.Int16Ty, CGM.Int16Ty}; | |||
1565 | auto *FnTy = | |||
1566 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); | |||
1567 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_spmd_kernel_init"); | |||
1568 | break; | |||
1569 | } | |||
1570 | case OMPRTL_NVPTX__kmpc_spmd_kernel_deinit_v2: { | |||
1571 | // Build void __kmpc_spmd_kernel_deinit_v2(int16_t RequiresOMPRuntime); | |||
1572 | llvm::Type *TypeParams[] = {CGM.Int16Ty}; | |||
1573 | auto *FnTy = | |||
1574 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); | |||
1575 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_spmd_kernel_deinit_v2"); | |||
1576 | break; | |||
1577 | } | |||
1578 | case OMPRTL_NVPTX__kmpc_kernel_prepare_parallel: { | |||
1579 | /// Build void __kmpc_kernel_prepare_parallel( | |||
1580 | /// void *outlined_function, int16_t IsOMPRuntimeInitialized); | |||
1581 | llvm::Type *TypeParams[] = {CGM.Int8PtrTy, CGM.Int16Ty}; | |||
1582 | auto *FnTy = | |||
1583 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); | |||
1584 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_kernel_prepare_parallel"); | |||
1585 | break; | |||
1586 | } | |||
1587 | case OMPRTL_NVPTX__kmpc_kernel_parallel: { | |||
1588 | /// Build bool __kmpc_kernel_parallel(void **outlined_function, | |||
1589 | /// int16_t IsOMPRuntimeInitialized); | |||
1590 | llvm::Type *TypeParams[] = {CGM.Int8PtrPtrTy, CGM.Int16Ty}; | |||
1591 | llvm::Type *RetTy = CGM.getTypes().ConvertType(CGM.getContext().BoolTy); | |||
1592 | auto *FnTy = | |||
1593 | llvm::FunctionType::get(RetTy, TypeParams, /*isVarArg*/ false); | |||
1594 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_kernel_parallel"); | |||
1595 | break; | |||
1596 | } | |||
1597 | case OMPRTL_NVPTX__kmpc_kernel_end_parallel: { | |||
1598 | /// Build void __kmpc_kernel_end_parallel(); | |||
1599 | auto *FnTy = | |||
1600 | llvm::FunctionType::get(CGM.VoidTy, llvm::None, /*isVarArg*/ false); | |||
1601 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_kernel_end_parallel"); | |||
1602 | break; | |||
1603 | } | |||
1604 | case OMPRTL_NVPTX__kmpc_serialized_parallel: { | |||
1605 | // Build void __kmpc_serialized_parallel(ident_t *loc, kmp_int32 | |||
1606 | // global_tid); | |||
1607 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty}; | |||
1608 | auto *FnTy = | |||
1609 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); | |||
1610 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_serialized_parallel"); | |||
1611 | break; | |||
1612 | } | |||
1613 | case OMPRTL_NVPTX__kmpc_end_serialized_parallel: { | |||
1614 | // Build void __kmpc_end_serialized_parallel(ident_t *loc, kmp_int32 | |||
1615 | // global_tid); | |||
1616 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty}; | |||
1617 | auto *FnTy = | |||
1618 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); | |||
1619 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_end_serialized_parallel"); | |||
1620 | break; | |||
1621 | } | |||
1622 | case OMPRTL_NVPTX__kmpc_shuffle_int32: { | |||
1623 | // Build int32_t __kmpc_shuffle_int32(int32_t element, | |||
1624 | // int16_t lane_offset, int16_t warp_size); | |||
1625 | llvm::Type *TypeParams[] = {CGM.Int32Ty, CGM.Int16Ty, CGM.Int16Ty}; | |||
1626 | auto *FnTy = | |||
1627 | llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg*/ false); | |||
1628 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_shuffle_int32"); | |||
1629 | break; | |||
1630 | } | |||
1631 | case OMPRTL_NVPTX__kmpc_shuffle_int64: { | |||
1632 | // Build int64_t __kmpc_shuffle_int64(int64_t element, | |||
1633 | // int16_t lane_offset, int16_t warp_size); | |||
1634 | llvm::Type *TypeParams[] = {CGM.Int64Ty, CGM.Int16Ty, CGM.Int16Ty}; | |||
1635 | auto *FnTy = | |||
1636 | llvm::FunctionType::get(CGM.Int64Ty, TypeParams, /*isVarArg*/ false); | |||
1637 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_shuffle_int64"); | |||
1638 | break; | |||
1639 | } | |||
1640 | case OMPRTL_NVPTX__kmpc_nvptx_parallel_reduce_nowait_v2: { | |||
1641 | // Build int32_t kmpc_nvptx_parallel_reduce_nowait_v2(ident_t *loc, | |||
1642 | // kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size, void* | |||
1643 | // reduce_data, void (*kmp_ShuffleReductFctPtr)(void *rhsData, int16_t | |||
1644 | // lane_id, int16_t lane_offset, int16_t Algorithm Version), void | |||
1645 | // (*kmp_InterWarpCopyFctPtr)(void* src, int warp_num)); | |||
1646 | llvm::Type *ShuffleReduceTypeParams[] = {CGM.VoidPtrTy, CGM.Int16Ty, | |||
1647 | CGM.Int16Ty, CGM.Int16Ty}; | |||
1648 | auto *ShuffleReduceFnTy = | |||
1649 | llvm::FunctionType::get(CGM.VoidTy, ShuffleReduceTypeParams, | |||
1650 | /*isVarArg=*/false); | |||
1651 | llvm::Type *InterWarpCopyTypeParams[] = {CGM.VoidPtrTy, CGM.Int32Ty}; | |||
1652 | auto *InterWarpCopyFnTy = | |||
1653 | llvm::FunctionType::get(CGM.VoidTy, InterWarpCopyTypeParams, | |||
1654 | /*isVarArg=*/false); | |||
1655 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), | |||
1656 | CGM.Int32Ty, | |||
1657 | CGM.Int32Ty, | |||
1658 | CGM.SizeTy, | |||
1659 | CGM.VoidPtrTy, | |||
1660 | ShuffleReduceFnTy->getPointerTo(), | |||
1661 | InterWarpCopyFnTy->getPointerTo()}; | |||
1662 | auto *FnTy = | |||
1663 | llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg=*/false); | |||
1664 | RTLFn = CGM.CreateRuntimeFunction( | |||
1665 | FnTy, /*Name=*/"__kmpc_nvptx_parallel_reduce_nowait_v2"); | |||
1666 | break; | |||
1667 | } | |||
1668 | case OMPRTL_NVPTX__kmpc_end_reduce_nowait: { | |||
1669 | // Build __kmpc_end_reduce_nowait(kmp_int32 global_tid); | |||
1670 | llvm::Type *TypeParams[] = {CGM.Int32Ty}; | |||
1671 | auto *FnTy = | |||
1672 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false); | |||
1673 | RTLFn = CGM.CreateRuntimeFunction( | |||
1674 | FnTy, /*Name=*/"__kmpc_nvptx_end_reduce_nowait"); | |||
1675 | break; | |||
1676 | } | |||
1677 | case OMPRTL_NVPTX__kmpc_nvptx_teams_reduce_nowait_v2: { | |||
1678 | // Build int32_t __kmpc_nvptx_teams_reduce_nowait_v2(ident_t *loc, kmp_int32 | |||
1679 | // global_tid, void *global_buffer, int32_t num_of_records, void* | |||
1680 | // reduce_data, | |||
1681 | // void (*kmp_ShuffleReductFctPtr)(void *rhsData, int16_t lane_id, int16_t | |||
1682 | // lane_offset, int16_t shortCircuit), | |||
1683 | // void (*kmp_InterWarpCopyFctPtr)(void* src, int32_t warp_num), void | |||
1684 | // (*kmp_ListToGlobalCpyFctPtr)(void *buffer, int idx, void *reduce_data), | |||
1685 | // void (*kmp_GlobalToListCpyFctPtr)(void *buffer, int idx, | |||
1686 | // void *reduce_data), void (*kmp_GlobalToListCpyPtrsFctPtr)(void *buffer, | |||
1687 | // int idx, void *reduce_data), void (*kmp_GlobalToListRedFctPtr)(void | |||
1688 | // *buffer, int idx, void *reduce_data)); | |||
1689 | llvm::Type *ShuffleReduceTypeParams[] = {CGM.VoidPtrTy, CGM.Int16Ty, | |||
1690 | CGM.Int16Ty, CGM.Int16Ty}; | |||
1691 | auto *ShuffleReduceFnTy = | |||
1692 | llvm::FunctionType::get(CGM.VoidTy, ShuffleReduceTypeParams, | |||
1693 | /*isVarArg=*/false); | |||
1694 | llvm::Type *InterWarpCopyTypeParams[] = {CGM.VoidPtrTy, CGM.Int32Ty}; | |||
1695 | auto *InterWarpCopyFnTy = | |||
1696 | llvm::FunctionType::get(CGM.VoidTy, InterWarpCopyTypeParams, | |||
1697 | /*isVarArg=*/false); | |||
1698 | llvm::Type *GlobalListTypeParams[] = {CGM.VoidPtrTy, CGM.IntTy, | |||
1699 | CGM.VoidPtrTy}; | |||
1700 | auto *GlobalListFnTy = | |||
1701 | llvm::FunctionType::get(CGM.VoidTy, GlobalListTypeParams, | |||
1702 | /*isVarArg=*/false); | |||
1703 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), | |||
1704 | CGM.Int32Ty, | |||
1705 | CGM.VoidPtrTy, | |||
1706 | CGM.Int32Ty, | |||
1707 | CGM.VoidPtrTy, | |||
1708 | ShuffleReduceFnTy->getPointerTo(), | |||
1709 | InterWarpCopyFnTy->getPointerTo(), | |||
1710 | GlobalListFnTy->getPointerTo(), | |||
1711 | GlobalListFnTy->getPointerTo(), | |||
1712 | GlobalListFnTy->getPointerTo(), | |||
1713 | GlobalListFnTy->getPointerTo()}; | |||
1714 | auto *FnTy = | |||
1715 | llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg=*/false); | |||
1716 | RTLFn = CGM.CreateRuntimeFunction( | |||
1717 | FnTy, /*Name=*/"__kmpc_nvptx_teams_reduce_nowait_v2"); | |||
1718 | break; | |||
1719 | } | |||
1720 | case OMPRTL_NVPTX__kmpc_data_sharing_init_stack: { | |||
1721 | /// Build void __kmpc_data_sharing_init_stack(); | |||
1722 | auto *FnTy = | |||
1723 | llvm::FunctionType::get(CGM.VoidTy, llvm::None, /*isVarArg*/ false); | |||
1724 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_data_sharing_init_stack"); | |||
1725 | break; | |||
1726 | } | |||
1727 | case OMPRTL_NVPTX__kmpc_data_sharing_init_stack_spmd: { | |||
1728 | /// Build void __kmpc_data_sharing_init_stack_spmd(); | |||
1729 | auto *FnTy = | |||
1730 | llvm::FunctionType::get(CGM.VoidTy, llvm::None, /*isVarArg*/ false); | |||
1731 | RTLFn = | |||
1732 | CGM.CreateRuntimeFunction(FnTy, "__kmpc_data_sharing_init_stack_spmd"); | |||
1733 | break; | |||
1734 | } | |||
1735 | case OMPRTL_NVPTX__kmpc_data_sharing_coalesced_push_stack: { | |||
1736 | // Build void *__kmpc_data_sharing_coalesced_push_stack(size_t size, | |||
1737 | // int16_t UseSharedMemory); | |||
1738 | llvm::Type *TypeParams[] = {CGM.SizeTy, CGM.Int16Ty}; | |||
1739 | auto *FnTy = | |||
1740 | llvm::FunctionType::get(CGM.VoidPtrTy, TypeParams, /*isVarArg=*/false); | |||
1741 | RTLFn = CGM.CreateRuntimeFunction( | |||
1742 | FnTy, /*Name=*/"__kmpc_data_sharing_coalesced_push_stack"); | |||
1743 | break; | |||
1744 | } | |||
1745 | case OMPRTL_NVPTX__kmpc_data_sharing_pop_stack: { | |||
1746 | // Build void __kmpc_data_sharing_pop_stack(void *a); | |||
1747 | llvm::Type *TypeParams[] = {CGM.VoidPtrTy}; | |||
1748 | auto *FnTy = | |||
1749 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false); | |||
1750 | RTLFn = CGM.CreateRuntimeFunction(FnTy, | |||
1751 | /*Name=*/"__kmpc_data_sharing_pop_stack"); | |||
1752 | break; | |||
1753 | } | |||
1754 | case OMPRTL_NVPTX__kmpc_begin_sharing_variables: { | |||
1755 | /// Build void __kmpc_begin_sharing_variables(void ***args, | |||
1756 | /// size_t n_args); | |||
1757 | llvm::Type *TypeParams[] = {CGM.Int8PtrPtrTy->getPointerTo(), CGM.SizeTy}; | |||
1758 | auto *FnTy = | |||
1759 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); | |||
1760 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_begin_sharing_variables"); | |||
1761 | break; | |||
1762 | } | |||
1763 | case OMPRTL_NVPTX__kmpc_end_sharing_variables: { | |||
1764 | /// Build void __kmpc_end_sharing_variables(); | |||
1765 | auto *FnTy = | |||
1766 | llvm::FunctionType::get(CGM.VoidTy, llvm::None, /*isVarArg*/ false); | |||
1767 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_end_sharing_variables"); | |||
1768 | break; | |||
1769 | } | |||
1770 | case OMPRTL_NVPTX__kmpc_get_shared_variables: { | |||
1771 | /// Build void __kmpc_get_shared_variables(void ***GlobalArgs); | |||
1772 | llvm::Type *TypeParams[] = {CGM.Int8PtrPtrTy->getPointerTo()}; | |||
1773 | auto *FnTy = | |||
1774 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); | |||
1775 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_get_shared_variables"); | |||
1776 | break; | |||
1777 | } | |||
1778 | case OMPRTL_NVPTX__kmpc_parallel_level: { | |||
1779 | // Build uint16_t __kmpc_parallel_level(ident_t *loc, kmp_int32 global_tid); | |||
1780 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty}; | |||
1781 | auto *FnTy = | |||
1782 | llvm::FunctionType::get(CGM.Int16Ty, TypeParams, /*isVarArg*/ false); | |||
1783 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_parallel_level"); | |||
1784 | break; | |||
1785 | } | |||
1786 | case OMPRTL_NVPTX__kmpc_is_spmd_exec_mode: { | |||
1787 | // Build int8_t __kmpc_is_spmd_exec_mode(); | |||
1788 | auto *FnTy = llvm::FunctionType::get(CGM.Int8Ty, /*isVarArg=*/false); | |||
1789 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_is_spmd_exec_mode"); | |||
1790 | break; | |||
1791 | } | |||
1792 | case OMPRTL_NVPTX__kmpc_get_team_static_memory: { | |||
1793 | // Build void __kmpc_get_team_static_memory(int16_t isSPMDExecutionMode, | |||
1794 | // const void *buf, size_t size, int16_t is_shared, const void **res); | |||
1795 | llvm::Type *TypeParams[] = {CGM.Int16Ty, CGM.VoidPtrTy, CGM.SizeTy, | |||
1796 | CGM.Int16Ty, CGM.VoidPtrPtrTy}; | |||
1797 | auto *FnTy = | |||
1798 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); | |||
1799 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_get_team_static_memory"); | |||
1800 | break; | |||
1801 | } | |||
1802 | case OMPRTL_NVPTX__kmpc_restore_team_static_memory: { | |||
1803 | // Build void __kmpc_restore_team_static_memory(int16_t isSPMDExecutionMode, | |||
1804 | // int16_t is_shared); | |||
1805 | llvm::Type *TypeParams[] = {CGM.Int16Ty, CGM.Int16Ty}; | |||
1806 | auto *FnTy = | |||
1807 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false); | |||
1808 | RTLFn = | |||
1809 | CGM.CreateRuntimeFunction(FnTy, "__kmpc_restore_team_static_memory"); | |||
1810 | break; | |||
1811 | } | |||
1812 | case OMPRTL__kmpc_barrier: { | |||
1813 | // Build void __kmpc_barrier(ident_t *loc, kmp_int32 global_tid); | |||
1814 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty}; | |||
1815 | auto *FnTy = | |||
1816 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); | |||
1817 | RTLFn = | |||
1818 | CGM.CreateConvergentRuntimeFunction(FnTy, /*Name*/ "__kmpc_barrier"); | |||
1819 | break; | |||
1820 | } | |||
1821 | case OMPRTL__kmpc_barrier_simple_spmd: { | |||
1822 | // Build void __kmpc_barrier_simple_spmd(ident_t *loc, kmp_int32 | |||
1823 | // global_tid); | |||
1824 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty}; | |||
1825 | auto *FnTy = | |||
1826 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); | |||
1827 | RTLFn = CGM.CreateConvergentRuntimeFunction( | |||
1828 | FnTy, /*Name*/ "__kmpc_barrier_simple_spmd"); | |||
1829 | break; | |||
1830 | } | |||
1831 | case OMPRTL_NVPTX__kmpc_warp_active_thread_mask: { | |||
1832 | // Build int32_t __kmpc_warp_active_thread_mask(void); | |||
1833 | auto *FnTy = | |||
1834 | llvm::FunctionType::get(CGM.Int32Ty, llvm::None, /*isVarArg=*/false); | |||
1835 | RTLFn = CGM.CreateConvergentRuntimeFunction(FnTy, "__kmpc_warp_active_thread_mask"); | |||
1836 | break; | |||
1837 | } | |||
1838 | case OMPRTL_NVPTX__kmpc_syncwarp: { | |||
1839 | // Build void __kmpc_syncwarp(kmp_int32 Mask); | |||
1840 | auto *FnTy = | |||
1841 | llvm::FunctionType::get(CGM.VoidTy, CGM.Int32Ty, /*isVarArg=*/false); | |||
1842 | RTLFn = CGM.CreateConvergentRuntimeFunction(FnTy, "__kmpc_syncwarp"); | |||
1843 | break; | |||
1844 | } | |||
1845 | } | |||
1846 | return RTLFn; | |||
1847 | } | |||
1848 | ||||
1849 | void CGOpenMPRuntimeNVPTX::createOffloadEntry(llvm::Constant *ID, | |||
1850 | llvm::Constant *Addr, | |||
1851 | uint64_t Size, int32_t, | |||
1852 | llvm::GlobalValue::LinkageTypes) { | |||
1853 | // TODO: Add support for global variables on the device after declare target | |||
1854 | // support. | |||
1855 | if (!isa<llvm::Function>(Addr)) | |||
1856 | return; | |||
1857 | llvm::Module &M = CGM.getModule(); | |||
1858 | llvm::LLVMContext &Ctx = CGM.getLLVMContext(); | |||
1859 | ||||
1860 | // Get "nvvm.annotations" metadata node | |||
1861 | llvm::NamedMDNode *MD = M.getOrInsertNamedMetadata("nvvm.annotations"); | |||
1862 | ||||
1863 | llvm::Metadata *MDVals[] = { | |||
1864 | llvm::ConstantAsMetadata::get(Addr), llvm::MDString::get(Ctx, "kernel"), | |||
1865 | llvm::ConstantAsMetadata::get( | |||
1866 | llvm::ConstantInt::get(llvm::Type::getInt32Ty(Ctx), 1))}; | |||
1867 | // Append metadata to nvvm.annotations | |||
1868 | MD->addOperand(llvm::MDNode::get(Ctx, MDVals)); | |||
1869 | } | |||
1870 | ||||
1871 | void CGOpenMPRuntimeNVPTX::emitTargetOutlinedFunction( | |||
1872 | const OMPExecutableDirective &D, StringRef ParentName, | |||
1873 | llvm::Function *&OutlinedFn, llvm::Constant *&OutlinedFnID, | |||
1874 | bool IsOffloadEntry, const RegionCodeGenTy &CodeGen) { | |||
1875 | if (!IsOffloadEntry) // Nothing to do. | |||
1876 | return; | |||
1877 | ||||
1878 | assert(!ParentName.empty() && "Invalid target region parent name!")((!ParentName.empty() && "Invalid target region parent name!" ) ? static_cast<void> (0) : __assert_fail ("!ParentName.empty() && \"Invalid target region parent name!\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 1878, __PRETTY_FUNCTION__)); | |||
1879 | ||||
1880 | bool Mode = supportsSPMDExecutionMode(CGM.getContext(), D); | |||
1881 | if (Mode) | |||
1882 | emitSPMDKernel(D, ParentName, OutlinedFn, OutlinedFnID, IsOffloadEntry, | |||
1883 | CodeGen); | |||
1884 | else | |||
1885 | emitNonSPMDKernel(D, ParentName, OutlinedFn, OutlinedFnID, IsOffloadEntry, | |||
1886 | CodeGen); | |||
1887 | ||||
1888 | setPropertyExecutionMode(CGM, OutlinedFn->getName(), Mode); | |||
1889 | } | |||
1890 | ||||
1891 | namespace { | |||
1892 | LLVM_ENABLE_BITMASK_ENUMS_IN_NAMESPACE()using ::llvm::BitmaskEnumDetail::operator~; using ::llvm::BitmaskEnumDetail ::operator|; using ::llvm::BitmaskEnumDetail::operator&; using ::llvm::BitmaskEnumDetail::operator^; using ::llvm::BitmaskEnumDetail ::operator|=; using ::llvm::BitmaskEnumDetail::operator&= ; using ::llvm::BitmaskEnumDetail::operator^=; | |||
1893 | /// Enum for accesseing the reserved_2 field of the ident_t struct. | |||
1894 | enum ModeFlagsTy : unsigned { | |||
1895 | /// Bit set to 1 when in SPMD mode. | |||
1896 | KMP_IDENT_SPMD_MODE = 0x01, | |||
1897 | /// Bit set to 1 when a simplified runtime is used. | |||
1898 | KMP_IDENT_SIMPLE_RT_MODE = 0x02, | |||
1899 | LLVM_MARK_AS_BITMASK_ENUM(/*LargestValue=*/KMP_IDENT_SIMPLE_RT_MODE)LLVM_BITMASK_LARGEST_ENUMERATOR = KMP_IDENT_SIMPLE_RT_MODE | |||
1900 | }; | |||
1901 | ||||
1902 | /// Special mode Undefined. Is the combination of Non-SPMD mode + SimpleRuntime. | |||
1903 | static const ModeFlagsTy UndefinedMode = | |||
1904 | (~KMP_IDENT_SPMD_MODE) & KMP_IDENT_SIMPLE_RT_MODE; | |||
1905 | } // anonymous namespace | |||
1906 | ||||
1907 | unsigned CGOpenMPRuntimeNVPTX::getDefaultLocationReserved2Flags() const { | |||
1908 | switch (getExecutionMode()) { | |||
1909 | case EM_SPMD: | |||
1910 | if (requiresFullRuntime()) | |||
1911 | return KMP_IDENT_SPMD_MODE & (~KMP_IDENT_SIMPLE_RT_MODE); | |||
1912 | return KMP_IDENT_SPMD_MODE | KMP_IDENT_SIMPLE_RT_MODE; | |||
1913 | case EM_NonSPMD: | |||
1914 | assert(requiresFullRuntime() && "Expected full runtime.")((requiresFullRuntime() && "Expected full runtime.") ? static_cast<void> (0) : __assert_fail ("requiresFullRuntime() && \"Expected full runtime.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 1914, __PRETTY_FUNCTION__)); | |||
1915 | return (~KMP_IDENT_SPMD_MODE) & (~KMP_IDENT_SIMPLE_RT_MODE); | |||
1916 | case EM_Unknown: | |||
1917 | return UndefinedMode; | |||
1918 | } | |||
1919 | llvm_unreachable("Unknown flags are requested.")::llvm::llvm_unreachable_internal("Unknown flags are requested." , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 1919); | |||
1920 | } | |||
1921 | ||||
1922 | bool CGOpenMPRuntimeNVPTX::tryEmitDeclareVariant(const GlobalDecl &NewGD, | |||
1923 | const GlobalDecl &OldGD, | |||
1924 | llvm::GlobalValue *OrigAddr, | |||
1925 | bool IsForDefinition) { | |||
1926 | // Emit the function in OldGD with the body from NewGD, if NewGD is defined. | |||
1927 | auto *NewFD = cast<FunctionDecl>(NewGD.getDecl()); | |||
1928 | if (NewFD->isDefined()) { | |||
1929 | CGM.emitOpenMPDeviceFunctionRedefinition(OldGD, NewGD, OrigAddr); | |||
1930 | return true; | |||
1931 | } | |||
1932 | return false; | |||
1933 | } | |||
1934 | ||||
1935 | CGOpenMPRuntimeNVPTX::CGOpenMPRuntimeNVPTX(CodeGenModule &CGM) | |||
1936 | : CGOpenMPRuntime(CGM, "_", "$") { | |||
1937 | if (!CGM.getLangOpts().OpenMPIsDevice) | |||
1938 | llvm_unreachable("OpenMP NVPTX can only handle device code.")::llvm::llvm_unreachable_internal("OpenMP NVPTX can only handle device code." , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 1938); | |||
1939 | } | |||
1940 | ||||
1941 | void CGOpenMPRuntimeNVPTX::emitProcBindClause(CodeGenFunction &CGF, | |||
1942 | ProcBindKind ProcBind, | |||
1943 | SourceLocation Loc) { | |||
1944 | // Do nothing in case of SPMD mode and L0 parallel. | |||
1945 | if (getExecutionMode() == CGOpenMPRuntimeNVPTX::EM_SPMD) | |||
1946 | return; | |||
1947 | ||||
1948 | CGOpenMPRuntime::emitProcBindClause(CGF, ProcBind, Loc); | |||
1949 | } | |||
1950 | ||||
1951 | void CGOpenMPRuntimeNVPTX::emitNumThreadsClause(CodeGenFunction &CGF, | |||
1952 | llvm::Value *NumThreads, | |||
1953 | SourceLocation Loc) { | |||
1954 | // Do nothing in case of SPMD mode and L0 parallel. | |||
1955 | if (getExecutionMode() == CGOpenMPRuntimeNVPTX::EM_SPMD) | |||
1956 | return; | |||
1957 | ||||
1958 | CGOpenMPRuntime::emitNumThreadsClause(CGF, NumThreads, Loc); | |||
1959 | } | |||
1960 | ||||
1961 | void CGOpenMPRuntimeNVPTX::emitNumTeamsClause(CodeGenFunction &CGF, | |||
1962 | const Expr *NumTeams, | |||
1963 | const Expr *ThreadLimit, | |||
1964 | SourceLocation Loc) {} | |||
1965 | ||||
1966 | llvm::Function *CGOpenMPRuntimeNVPTX::emitParallelOutlinedFunction( | |||
1967 | const OMPExecutableDirective &D, const VarDecl *ThreadIDVar, | |||
1968 | OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) { | |||
1969 | // Emit target region as a standalone region. | |||
1970 | class NVPTXPrePostActionTy : public PrePostActionTy { | |||
1971 | bool &IsInParallelRegion; | |||
1972 | bool PrevIsInParallelRegion; | |||
1973 | ||||
1974 | public: | |||
1975 | NVPTXPrePostActionTy(bool &IsInParallelRegion) | |||
1976 | : IsInParallelRegion(IsInParallelRegion) {} | |||
1977 | void Enter(CodeGenFunction &CGF) override { | |||
1978 | PrevIsInParallelRegion = IsInParallelRegion; | |||
1979 | IsInParallelRegion = true; | |||
1980 | } | |||
1981 | void Exit(CodeGenFunction &CGF) override { | |||
1982 | IsInParallelRegion = PrevIsInParallelRegion; | |||
1983 | } | |||
1984 | } Action(IsInParallelRegion); | |||
1985 | CodeGen.setAction(Action); | |||
1986 | bool PrevIsInTTDRegion = IsInTTDRegion; | |||
1987 | IsInTTDRegion = false; | |||
1988 | bool PrevIsInTargetMasterThreadRegion = IsInTargetMasterThreadRegion; | |||
1989 | IsInTargetMasterThreadRegion = false; | |||
1990 | auto *OutlinedFun = | |||
1991 | cast<llvm::Function>(CGOpenMPRuntime::emitParallelOutlinedFunction( | |||
1992 | D, ThreadIDVar, InnermostKind, CodeGen)); | |||
1993 | if (CGM.getLangOpts().Optimize) { | |||
1994 | OutlinedFun->removeFnAttr(llvm::Attribute::NoInline); | |||
1995 | OutlinedFun->removeFnAttr(llvm::Attribute::OptimizeNone); | |||
1996 | OutlinedFun->addFnAttr(llvm::Attribute::AlwaysInline); | |||
1997 | } | |||
1998 | IsInTargetMasterThreadRegion = PrevIsInTargetMasterThreadRegion; | |||
1999 | IsInTTDRegion = PrevIsInTTDRegion; | |||
2000 | if (getExecutionMode() != CGOpenMPRuntimeNVPTX::EM_SPMD && | |||
2001 | !IsInParallelRegion) { | |||
2002 | llvm::Function *WrapperFun = | |||
2003 | createParallelDataSharingWrapper(OutlinedFun, D); | |||
2004 | WrapperFunctionsMap[OutlinedFun] = WrapperFun; | |||
2005 | } | |||
2006 | ||||
2007 | return OutlinedFun; | |||
2008 | } | |||
2009 | ||||
2010 | /// Get list of lastprivate variables from the teams distribute ... or | |||
2011 | /// teams {distribute ...} directives. | |||
2012 | static void | |||
2013 | getDistributeLastprivateVars(ASTContext &Ctx, const OMPExecutableDirective &D, | |||
2014 | llvm::SmallVectorImpl<const ValueDecl *> &Vars) { | |||
2015 | assert(isOpenMPTeamsDirective(D.getDirectiveKind()) &&((isOpenMPTeamsDirective(D.getDirectiveKind()) && "expected teams directive." ) ? static_cast<void> (0) : __assert_fail ("isOpenMPTeamsDirective(D.getDirectiveKind()) && \"expected teams directive.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 2016, __PRETTY_FUNCTION__)) | |||
2016 | "expected teams directive.")((isOpenMPTeamsDirective(D.getDirectiveKind()) && "expected teams directive." ) ? static_cast<void> (0) : __assert_fail ("isOpenMPTeamsDirective(D.getDirectiveKind()) && \"expected teams directive.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 2016, __PRETTY_FUNCTION__)); | |||
2017 | const OMPExecutableDirective *Dir = &D; | |||
2018 | if (!isOpenMPDistributeDirective(D.getDirectiveKind())) { | |||
2019 | if (const Stmt *S = CGOpenMPRuntime::getSingleCompoundChild( | |||
2020 | Ctx, | |||
2021 | D.getInnermostCapturedStmt()->getCapturedStmt()->IgnoreContainers( | |||
2022 | /*IgnoreCaptured=*/true))) { | |||
2023 | Dir = dyn_cast_or_null<OMPExecutableDirective>(S); | |||
2024 | if (Dir && !isOpenMPDistributeDirective(Dir->getDirectiveKind())) | |||
2025 | Dir = nullptr; | |||
2026 | } | |||
2027 | } | |||
2028 | if (!Dir) | |||
2029 | return; | |||
2030 | for (const auto *C : Dir->getClausesOfKind<OMPLastprivateClause>()) { | |||
2031 | for (const Expr *E : C->getVarRefs()) | |||
2032 | Vars.push_back(getPrivateItem(E)); | |||
2033 | } | |||
2034 | } | |||
2035 | ||||
2036 | /// Get list of reduction variables from the teams ... directives. | |||
2037 | static void | |||
2038 | getTeamsReductionVars(ASTContext &Ctx, const OMPExecutableDirective &D, | |||
2039 | llvm::SmallVectorImpl<const ValueDecl *> &Vars) { | |||
2040 | assert(isOpenMPTeamsDirective(D.getDirectiveKind()) &&((isOpenMPTeamsDirective(D.getDirectiveKind()) && "expected teams directive." ) ? static_cast<void> (0) : __assert_fail ("isOpenMPTeamsDirective(D.getDirectiveKind()) && \"expected teams directive.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 2041, __PRETTY_FUNCTION__)) | |||
2041 | "expected teams directive.")((isOpenMPTeamsDirective(D.getDirectiveKind()) && "expected teams directive." ) ? static_cast<void> (0) : __assert_fail ("isOpenMPTeamsDirective(D.getDirectiveKind()) && \"expected teams directive.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 2041, __PRETTY_FUNCTION__)); | |||
2042 | for (const auto *C : D.getClausesOfKind<OMPReductionClause>()) { | |||
2043 | for (const Expr *E : C->privates()) | |||
2044 | Vars.push_back(getPrivateItem(E)); | |||
2045 | } | |||
2046 | } | |||
2047 | ||||
2048 | llvm::Function *CGOpenMPRuntimeNVPTX::emitTeamsOutlinedFunction( | |||
2049 | const OMPExecutableDirective &D, const VarDecl *ThreadIDVar, | |||
2050 | OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) { | |||
2051 | SourceLocation Loc = D.getBeginLoc(); | |||
2052 | ||||
2053 | const RecordDecl *GlobalizedRD = nullptr; | |||
2054 | llvm::SmallVector<const ValueDecl *, 4> LastPrivatesReductions; | |||
2055 | llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> MappedDeclsFields; | |||
2056 | // Globalize team reductions variable unconditionally in all modes. | |||
2057 | if (getExecutionMode() != CGOpenMPRuntimeNVPTX::EM_SPMD) | |||
2058 | getTeamsReductionVars(CGM.getContext(), D, LastPrivatesReductions); | |||
2059 | if (getExecutionMode() == CGOpenMPRuntimeNVPTX::EM_SPMD) { | |||
2060 | getDistributeLastprivateVars(CGM.getContext(), D, LastPrivatesReductions); | |||
2061 | if (!LastPrivatesReductions.empty()) { | |||
2062 | GlobalizedRD = ::buildRecordForGlobalizedVars( | |||
2063 | CGM.getContext(), llvm::None, LastPrivatesReductions, | |||
2064 | MappedDeclsFields, WarpSize); | |||
2065 | } | |||
2066 | } else if (!LastPrivatesReductions.empty()) { | |||
2067 | assert(!TeamAndReductions.first &&((!TeamAndReductions.first && "Previous team declaration is not expected." ) ? static_cast<void> (0) : __assert_fail ("!TeamAndReductions.first && \"Previous team declaration is not expected.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 2068, __PRETTY_FUNCTION__)) | |||
2068 | "Previous team declaration is not expected.")((!TeamAndReductions.first && "Previous team declaration is not expected." ) ? static_cast<void> (0) : __assert_fail ("!TeamAndReductions.first && \"Previous team declaration is not expected.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 2068, __PRETTY_FUNCTION__)); | |||
2069 | TeamAndReductions.first = D.getCapturedStmt(OMPD_teams)->getCapturedDecl(); | |||
2070 | std::swap(TeamAndReductions.second, LastPrivatesReductions); | |||
2071 | } | |||
2072 | ||||
2073 | // Emit target region as a standalone region. | |||
2074 | class NVPTXPrePostActionTy : public PrePostActionTy { | |||
2075 | SourceLocation &Loc; | |||
2076 | const RecordDecl *GlobalizedRD; | |||
2077 | llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> | |||
2078 | &MappedDeclsFields; | |||
2079 | ||||
2080 | public: | |||
2081 | NVPTXPrePostActionTy( | |||
2082 | SourceLocation &Loc, const RecordDecl *GlobalizedRD, | |||
2083 | llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> | |||
2084 | &MappedDeclsFields) | |||
2085 | : Loc(Loc), GlobalizedRD(GlobalizedRD), | |||
2086 | MappedDeclsFields(MappedDeclsFields) {} | |||
2087 | void Enter(CodeGenFunction &CGF) override { | |||
2088 | auto &Rt = | |||
2089 | static_cast<CGOpenMPRuntimeNVPTX &>(CGF.CGM.getOpenMPRuntime()); | |||
2090 | if (GlobalizedRD) { | |||
2091 | auto I = Rt.FunctionGlobalizedDecls.try_emplace(CGF.CurFn).first; | |||
2092 | I->getSecond().GlobalRecord = GlobalizedRD; | |||
2093 | I->getSecond().MappedParams = | |||
2094 | std::make_unique<CodeGenFunction::OMPMapVars>(); | |||
2095 | DeclToAddrMapTy &Data = I->getSecond().LocalVarData; | |||
2096 | for (const auto &Pair : MappedDeclsFields) { | |||
2097 | assert(Pair.getFirst()->isCanonicalDecl() &&((Pair.getFirst()->isCanonicalDecl() && "Expected canonical declaration" ) ? static_cast<void> (0) : __assert_fail ("Pair.getFirst()->isCanonicalDecl() && \"Expected canonical declaration\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 2098, __PRETTY_FUNCTION__)) | |||
2098 | "Expected canonical declaration")((Pair.getFirst()->isCanonicalDecl() && "Expected canonical declaration" ) ? static_cast<void> (0) : __assert_fail ("Pair.getFirst()->isCanonicalDecl() && \"Expected canonical declaration\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 2098, __PRETTY_FUNCTION__)); | |||
2099 | Data.insert(std::make_pair(Pair.getFirst(), | |||
2100 | MappedVarData(Pair.getSecond(), | |||
2101 | /*IsOnePerTeam=*/true))); | |||
2102 | } | |||
2103 | } | |||
2104 | Rt.emitGenericVarsProlog(CGF, Loc); | |||
2105 | } | |||
2106 | void Exit(CodeGenFunction &CGF) override { | |||
2107 | static_cast<CGOpenMPRuntimeNVPTX &>(CGF.CGM.getOpenMPRuntime()) | |||
2108 | .emitGenericVarsEpilog(CGF); | |||
2109 | } | |||
2110 | } Action(Loc, GlobalizedRD, MappedDeclsFields); | |||
2111 | CodeGen.setAction(Action); | |||
2112 | llvm::Function *OutlinedFun = CGOpenMPRuntime::emitTeamsOutlinedFunction( | |||
2113 | D, ThreadIDVar, InnermostKind, CodeGen); | |||
2114 | if (CGM.getLangOpts().Optimize) { | |||
2115 | OutlinedFun->removeFnAttr(llvm::Attribute::NoInline); | |||
2116 | OutlinedFun->removeFnAttr(llvm::Attribute::OptimizeNone); | |||
2117 | OutlinedFun->addFnAttr(llvm::Attribute::AlwaysInline); | |||
2118 | } | |||
2119 | ||||
2120 | return OutlinedFun; | |||
2121 | } | |||
2122 | ||||
2123 | void CGOpenMPRuntimeNVPTX::emitGenericVarsProlog(CodeGenFunction &CGF, | |||
2124 | SourceLocation Loc, | |||
2125 | bool WithSPMDCheck) { | |||
2126 | if (getDataSharingMode(CGM) != CGOpenMPRuntimeNVPTX::Generic && | |||
2127 | getExecutionMode() != CGOpenMPRuntimeNVPTX::EM_SPMD) | |||
2128 | return; | |||
2129 | ||||
2130 | CGBuilderTy &Bld = CGF.Builder; | |||
2131 | ||||
2132 | const auto I = FunctionGlobalizedDecls.find(CGF.CurFn); | |||
2133 | if (I == FunctionGlobalizedDecls.end()) | |||
2134 | return; | |||
2135 | if (const RecordDecl *GlobalizedVarsRecord = I->getSecond().GlobalRecord) { | |||
2136 | QualType GlobalRecTy = CGM.getContext().getRecordType(GlobalizedVarsRecord); | |||
2137 | QualType SecGlobalRecTy; | |||
2138 | ||||
2139 | // Recover pointer to this function's global record. The runtime will | |||
2140 | // handle the specifics of the allocation of the memory. | |||
2141 | // Use actual memory size of the record including the padding | |||
2142 | // for alignment purposes. | |||
2143 | unsigned Alignment = | |||
2144 | CGM.getContext().getTypeAlignInChars(GlobalRecTy).getQuantity(); | |||
2145 | unsigned GlobalRecordSize = | |||
2146 | CGM.getContext().getTypeSizeInChars(GlobalRecTy).getQuantity(); | |||
2147 | GlobalRecordSize = llvm::alignTo(GlobalRecordSize, Alignment); | |||
2148 | ||||
2149 | llvm::PointerType *GlobalRecPtrTy = | |||
2150 | CGF.ConvertTypeForMem(GlobalRecTy)->getPointerTo(); | |||
2151 | llvm::Value *GlobalRecCastAddr; | |||
2152 | llvm::Value *IsTTD = nullptr; | |||
2153 | if (!IsInTTDRegion && | |||
2154 | (WithSPMDCheck || | |||
2155 | getExecutionMode() == CGOpenMPRuntimeNVPTX::EM_Unknown)) { | |||
2156 | llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".exit"); | |||
2157 | llvm::BasicBlock *SPMDBB = CGF.createBasicBlock(".spmd"); | |||
2158 | llvm::BasicBlock *NonSPMDBB = CGF.createBasicBlock(".non-spmd"); | |||
2159 | if (I->getSecond().SecondaryGlobalRecord.hasValue()) { | |||
2160 | llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc); | |||
2161 | llvm::Value *ThreadID = getThreadID(CGF, Loc); | |||
2162 | llvm::Value *PL = CGF.EmitRuntimeCall( | |||
2163 | createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_parallel_level), | |||
2164 | {RTLoc, ThreadID}); | |||
2165 | IsTTD = Bld.CreateIsNull(PL); | |||
2166 | } | |||
2167 | llvm::Value *IsSPMD = Bld.CreateIsNotNull(CGF.EmitNounwindRuntimeCall( | |||
2168 | createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_is_spmd_exec_mode))); | |||
2169 | Bld.CreateCondBr(IsSPMD, SPMDBB, NonSPMDBB); | |||
2170 | // There is no need to emit line number for unconditional branch. | |||
2171 | (void)ApplyDebugLocation::CreateEmpty(CGF); | |||
2172 | CGF.EmitBlock(SPMDBB); | |||
2173 | Address RecPtr = Address(llvm::ConstantPointerNull::get(GlobalRecPtrTy), | |||
2174 | CharUnits::fromQuantity(Alignment)); | |||
2175 | CGF.EmitBranch(ExitBB); | |||
2176 | // There is no need to emit line number for unconditional branch. | |||
2177 | (void)ApplyDebugLocation::CreateEmpty(CGF); | |||
2178 | CGF.EmitBlock(NonSPMDBB); | |||
2179 | llvm::Value *Size = llvm::ConstantInt::get(CGM.SizeTy, GlobalRecordSize); | |||
2180 | if (const RecordDecl *SecGlobalizedVarsRecord = | |||
2181 | I->getSecond().SecondaryGlobalRecord.getValueOr(nullptr)) { | |||
2182 | SecGlobalRecTy = | |||
2183 | CGM.getContext().getRecordType(SecGlobalizedVarsRecord); | |||
2184 | ||||
2185 | // Recover pointer to this function's global record. The runtime will | |||
2186 | // handle the specifics of the allocation of the memory. | |||
2187 | // Use actual memory size of the record including the padding | |||
2188 | // for alignment purposes. | |||
2189 | unsigned Alignment = | |||
2190 | CGM.getContext().getTypeAlignInChars(SecGlobalRecTy).getQuantity(); | |||
2191 | unsigned GlobalRecordSize = | |||
2192 | CGM.getContext().getTypeSizeInChars(SecGlobalRecTy).getQuantity(); | |||
2193 | GlobalRecordSize = llvm::alignTo(GlobalRecordSize, Alignment); | |||
2194 | Size = Bld.CreateSelect( | |||
2195 | IsTTD, llvm::ConstantInt::get(CGM.SizeTy, GlobalRecordSize), Size); | |||
2196 | } | |||
2197 | // TODO: allow the usage of shared memory to be controlled by | |||
2198 | // the user, for now, default to global. | |||
2199 | llvm::Value *GlobalRecordSizeArg[] = { | |||
2200 | Size, CGF.Builder.getInt16(/*UseSharedMemory=*/0)}; | |||
2201 | llvm::Value *GlobalRecValue = CGF.EmitRuntimeCall( | |||
2202 | createNVPTXRuntimeFunction( | |||
2203 | OMPRTL_NVPTX__kmpc_data_sharing_coalesced_push_stack), | |||
2204 | GlobalRecordSizeArg); | |||
2205 | GlobalRecCastAddr = Bld.CreatePointerBitCastOrAddrSpaceCast( | |||
2206 | GlobalRecValue, GlobalRecPtrTy); | |||
2207 | CGF.EmitBlock(ExitBB); | |||
2208 | auto *Phi = Bld.CreatePHI(GlobalRecPtrTy, | |||
2209 | /*NumReservedValues=*/2, "_select_stack"); | |||
2210 | Phi->addIncoming(RecPtr.getPointer(), SPMDBB); | |||
2211 | Phi->addIncoming(GlobalRecCastAddr, NonSPMDBB); | |||
2212 | GlobalRecCastAddr = Phi; | |||
2213 | I->getSecond().GlobalRecordAddr = Phi; | |||
2214 | I->getSecond().IsInSPMDModeFlag = IsSPMD; | |||
2215 | } else if (IsInTTDRegion) { | |||
2216 | assert(GlobalizedRecords.back().Records.size() < 2 &&((GlobalizedRecords.back().Records.size() < 2 && "Expected less than 2 globalized records: one for target and one " "for teams.") ? static_cast<void> (0) : __assert_fail ( "GlobalizedRecords.back().Records.size() < 2 && \"Expected less than 2 globalized records: one for target and one \" \"for teams.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 2218, __PRETTY_FUNCTION__)) | |||
2217 | "Expected less than 2 globalized records: one for target and one "((GlobalizedRecords.back().Records.size() < 2 && "Expected less than 2 globalized records: one for target and one " "for teams.") ? static_cast<void> (0) : __assert_fail ( "GlobalizedRecords.back().Records.size() < 2 && \"Expected less than 2 globalized records: one for target and one \" \"for teams.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 2218, __PRETTY_FUNCTION__)) | |||
2218 | "for teams.")((GlobalizedRecords.back().Records.size() < 2 && "Expected less than 2 globalized records: one for target and one " "for teams.") ? static_cast<void> (0) : __assert_fail ( "GlobalizedRecords.back().Records.size() < 2 && \"Expected less than 2 globalized records: one for target and one \" \"for teams.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 2218, __PRETTY_FUNCTION__)); | |||
2219 | unsigned Offset = 0; | |||
2220 | for (const RecordDecl *RD : GlobalizedRecords.back().Records) { | |||
2221 | QualType RDTy = CGM.getContext().getRecordType(RD); | |||
2222 | unsigned Alignment = | |||
2223 | CGM.getContext().getTypeAlignInChars(RDTy).getQuantity(); | |||
2224 | unsigned Size = CGM.getContext().getTypeSizeInChars(RDTy).getQuantity(); | |||
2225 | Offset = | |||
2226 | llvm::alignTo(llvm::alignTo(Offset, Alignment) + Size, Alignment); | |||
2227 | } | |||
2228 | unsigned Alignment = | |||
2229 | CGM.getContext().getTypeAlignInChars(GlobalRecTy).getQuantity(); | |||
2230 | Offset = llvm::alignTo(Offset, Alignment); | |||
2231 | GlobalizedRecords.back().Records.push_back(GlobalizedVarsRecord); | |||
2232 | ++GlobalizedRecords.back().RegionCounter; | |||
2233 | if (GlobalizedRecords.back().Records.size() == 1) { | |||
2234 | assert(KernelStaticGlobalized &&((KernelStaticGlobalized && "Kernel static pointer must be initialized already." ) ? static_cast<void> (0) : __assert_fail ("KernelStaticGlobalized && \"Kernel static pointer must be initialized already.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 2235, __PRETTY_FUNCTION__)) | |||
2235 | "Kernel static pointer must be initialized already.")((KernelStaticGlobalized && "Kernel static pointer must be initialized already." ) ? static_cast<void> (0) : __assert_fail ("KernelStaticGlobalized && \"Kernel static pointer must be initialized already.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 2235, __PRETTY_FUNCTION__)); | |||
2236 | auto *UseSharedMemory = new llvm::GlobalVariable( | |||
2237 | CGM.getModule(), CGM.Int16Ty, /*isConstant=*/true, | |||
2238 | llvm::GlobalValue::InternalLinkage, nullptr, | |||
2239 | "_openmp_static_kernel$is_shared"); | |||
2240 | UseSharedMemory->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global); | |||
2241 | QualType Int16Ty = CGM.getContext().getIntTypeForBitwidth( | |||
2242 | /*DestWidth=*/16, /*Signed=*/0); | |||
2243 | llvm::Value *IsInSharedMemory = CGF.EmitLoadOfScalar( | |||
2244 | Address(UseSharedMemory, | |||
2245 | CGM.getContext().getTypeAlignInChars(Int16Ty)), | |||
2246 | /*Volatile=*/false, Int16Ty, Loc); | |||
2247 | auto *StaticGlobalized = new llvm::GlobalVariable( | |||
2248 | CGM.getModule(), CGM.Int8Ty, /*isConstant=*/false, | |||
2249 | llvm::GlobalValue::CommonLinkage, nullptr); | |||
2250 | auto *RecSize = new llvm::GlobalVariable( | |||
2251 | CGM.getModule(), CGM.SizeTy, /*isConstant=*/true, | |||
2252 | llvm::GlobalValue::InternalLinkage, nullptr, | |||
2253 | "_openmp_static_kernel$size"); | |||
2254 | RecSize->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global); | |||
2255 | llvm::Value *Ld = CGF.EmitLoadOfScalar( | |||
2256 | Address(RecSize, CGM.getSizeAlign()), /*Volatile=*/false, | |||
2257 | CGM.getContext().getSizeType(), Loc); | |||
2258 | llvm::Value *ResAddr = Bld.CreatePointerBitCastOrAddrSpaceCast( | |||
2259 | KernelStaticGlobalized, CGM.VoidPtrPtrTy); | |||
2260 | llvm::Value *GlobalRecordSizeArg[] = { | |||
2261 | llvm::ConstantInt::get( | |||
2262 | CGM.Int16Ty, | |||
2263 | getExecutionMode() == CGOpenMPRuntimeNVPTX::EM_SPMD ? 1 : 0), | |||
2264 | StaticGlobalized, Ld, IsInSharedMemory, ResAddr}; | |||
2265 | CGF.EmitRuntimeCall(createNVPTXRuntimeFunction( | |||
2266 | OMPRTL_NVPTX__kmpc_get_team_static_memory), | |||
2267 | GlobalRecordSizeArg); | |||
2268 | GlobalizedRecords.back().Buffer = StaticGlobalized; | |||
2269 | GlobalizedRecords.back().RecSize = RecSize; | |||
2270 | GlobalizedRecords.back().UseSharedMemory = UseSharedMemory; | |||
2271 | GlobalizedRecords.back().Loc = Loc; | |||
2272 | } | |||
2273 | assert(KernelStaticGlobalized && "Global address must be set already.")((KernelStaticGlobalized && "Global address must be set already." ) ? static_cast<void> (0) : __assert_fail ("KernelStaticGlobalized && \"Global address must be set already.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 2273, __PRETTY_FUNCTION__)); | |||
2274 | Address FrameAddr = CGF.EmitLoadOfPointer( | |||
2275 | Address(KernelStaticGlobalized, CGM.getPointerAlign()), | |||
2276 | CGM.getContext() | |||
2277 | .getPointerType(CGM.getContext().VoidPtrTy) | |||
2278 | .castAs<PointerType>()); | |||
2279 | llvm::Value *GlobalRecValue = | |||
2280 | Bld.CreateConstInBoundsGEP(FrameAddr, Offset).getPointer(); | |||
2281 | I->getSecond().GlobalRecordAddr = GlobalRecValue; | |||
2282 | I->getSecond().IsInSPMDModeFlag = nullptr; | |||
2283 | GlobalRecCastAddr = Bld.CreatePointerBitCastOrAddrSpaceCast( | |||
2284 | GlobalRecValue, CGF.ConvertTypeForMem(GlobalRecTy)->getPointerTo()); | |||
2285 | } else { | |||
2286 | // TODO: allow the usage of shared memory to be controlled by | |||
2287 | // the user, for now, default to global. | |||
2288 | llvm::Value *GlobalRecordSizeArg[] = { | |||
2289 | llvm::ConstantInt::get(CGM.SizeTy, GlobalRecordSize), | |||
2290 | CGF.Builder.getInt16(/*UseSharedMemory=*/0)}; | |||
2291 | llvm::Value *GlobalRecValue = CGF.EmitRuntimeCall( | |||
2292 | createNVPTXRuntimeFunction( | |||
2293 | OMPRTL_NVPTX__kmpc_data_sharing_coalesced_push_stack), | |||
2294 | GlobalRecordSizeArg); | |||
2295 | GlobalRecCastAddr = Bld.CreatePointerBitCastOrAddrSpaceCast( | |||
2296 | GlobalRecValue, GlobalRecPtrTy); | |||
2297 | I->getSecond().GlobalRecordAddr = GlobalRecValue; | |||
2298 | I->getSecond().IsInSPMDModeFlag = nullptr; | |||
2299 | } | |||
2300 | LValue Base = | |||
2301 | CGF.MakeNaturalAlignPointeeAddrLValue(GlobalRecCastAddr, GlobalRecTy); | |||
2302 | ||||
2303 | // Emit the "global alloca" which is a GEP from the global declaration | |||
2304 | // record using the pointer returned by the runtime. | |||
2305 | LValue SecBase; | |||
2306 | decltype(I->getSecond().LocalVarData)::const_iterator SecIt; | |||
2307 | if (IsTTD) { | |||
2308 | SecIt = I->getSecond().SecondaryLocalVarData->begin(); | |||
2309 | llvm::PointerType *SecGlobalRecPtrTy = | |||
2310 | CGF.ConvertTypeForMem(SecGlobalRecTy)->getPointerTo(); | |||
2311 | SecBase = CGF.MakeNaturalAlignPointeeAddrLValue( | |||
2312 | Bld.CreatePointerBitCastOrAddrSpaceCast( | |||
2313 | I->getSecond().GlobalRecordAddr, SecGlobalRecPtrTy), | |||
2314 | SecGlobalRecTy); | |||
2315 | } | |||
2316 | for (auto &Rec : I->getSecond().LocalVarData) { | |||
2317 | bool EscapedParam = I->getSecond().EscapedParameters.count(Rec.first); | |||
2318 | llvm::Value *ParValue; | |||
2319 | if (EscapedParam) { | |||
2320 | const auto *VD = cast<VarDecl>(Rec.first); | |||
2321 | LValue ParLVal = | |||
2322 | CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(VD), VD->getType()); | |||
2323 | ParValue = CGF.EmitLoadOfScalar(ParLVal, Loc); | |||
2324 | } | |||
2325 | LValue VarAddr = CGF.EmitLValueForField(Base, Rec.second.FD); | |||
2326 | // Emit VarAddr basing on lane-id if required. | |||
2327 | QualType VarTy; | |||
2328 | if (Rec.second.IsOnePerTeam) { | |||
2329 | VarTy = Rec.second.FD->getType(); | |||
2330 | } else { | |||
2331 | llvm::Value *Ptr = CGF.Builder.CreateInBoundsGEP( | |||
2332 | VarAddr.getAddress(CGF).getPointer(), | |||
2333 | {Bld.getInt32(0), getNVPTXLaneID(CGF)}); | |||
2334 | VarTy = | |||
2335 | Rec.second.FD->getType()->castAsArrayTypeUnsafe()->getElementType(); | |||
2336 | VarAddr = CGF.MakeAddrLValue( | |||
2337 | Address(Ptr, CGM.getContext().getDeclAlign(Rec.first)), VarTy, | |||
2338 | AlignmentSource::Decl); | |||
2339 | } | |||
2340 | Rec.second.PrivateAddr = VarAddr.getAddress(CGF); | |||
2341 | if (!IsInTTDRegion && | |||
2342 | (WithSPMDCheck || | |||
2343 | getExecutionMode() == CGOpenMPRuntimeNVPTX::EM_Unknown)) { | |||
2344 | assert(I->getSecond().IsInSPMDModeFlag &&((I->getSecond().IsInSPMDModeFlag && "Expected unknown execution mode or required SPMD check." ) ? static_cast<void> (0) : __assert_fail ("I->getSecond().IsInSPMDModeFlag && \"Expected unknown execution mode or required SPMD check.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 2345, __PRETTY_FUNCTION__)) | |||
2345 | "Expected unknown execution mode or required SPMD check.")((I->getSecond().IsInSPMDModeFlag && "Expected unknown execution mode or required SPMD check." ) ? static_cast<void> (0) : __assert_fail ("I->getSecond().IsInSPMDModeFlag && \"Expected unknown execution mode or required SPMD check.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 2345, __PRETTY_FUNCTION__)); | |||
2346 | if (IsTTD) { | |||
2347 | assert(SecIt->second.IsOnePerTeam &&((SecIt->second.IsOnePerTeam && "Secondary glob data must be one per team." ) ? static_cast<void> (0) : __assert_fail ("SecIt->second.IsOnePerTeam && \"Secondary glob data must be one per team.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 2348, __PRETTY_FUNCTION__)) | |||
2348 | "Secondary glob data must be one per team.")((SecIt->second.IsOnePerTeam && "Secondary glob data must be one per team." ) ? static_cast<void> (0) : __assert_fail ("SecIt->second.IsOnePerTeam && \"Secondary glob data must be one per team.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 2348, __PRETTY_FUNCTION__)); | |||
2349 | LValue SecVarAddr = CGF.EmitLValueForField(SecBase, SecIt->second.FD); | |||
2350 | VarAddr.setAddress( | |||
2351 | Address(Bld.CreateSelect(IsTTD, SecVarAddr.getPointer(CGF), | |||
2352 | VarAddr.getPointer(CGF)), | |||
2353 | VarAddr.getAlignment())); | |||
2354 | Rec.second.PrivateAddr = VarAddr.getAddress(CGF); | |||
2355 | } | |||
2356 | Address GlobalPtr = Rec.second.PrivateAddr; | |||
2357 | Address LocalAddr = CGF.CreateMemTemp(VarTy, Rec.second.FD->getName()); | |||
2358 | Rec.second.PrivateAddr = Address( | |||
2359 | Bld.CreateSelect(I->getSecond().IsInSPMDModeFlag, | |||
2360 | LocalAddr.getPointer(), GlobalPtr.getPointer()), | |||
2361 | LocalAddr.getAlignment()); | |||
2362 | } | |||
2363 | if (EscapedParam) { | |||
2364 | const auto *VD = cast<VarDecl>(Rec.first); | |||
2365 | CGF.EmitStoreOfScalar(ParValue, VarAddr); | |||
2366 | I->getSecond().MappedParams->setVarAddr(CGF, VD, | |||
2367 | VarAddr.getAddress(CGF)); | |||
2368 | } | |||
2369 | if (IsTTD) | |||
2370 | ++SecIt; | |||
2371 | } | |||
2372 | } | |||
2373 | for (const ValueDecl *VD : I->getSecond().EscapedVariableLengthDecls) { | |||
2374 | // Recover pointer to this function's global record. The runtime will | |||
2375 | // handle the specifics of the allocation of the memory. | |||
2376 | // Use actual memory size of the record including the padding | |||
2377 | // for alignment purposes. | |||
2378 | CGBuilderTy &Bld = CGF.Builder; | |||
2379 | llvm::Value *Size = CGF.getTypeSize(VD->getType()); | |||
2380 | CharUnits Align = CGM.getContext().getDeclAlign(VD); | |||
2381 | Size = Bld.CreateNUWAdd( | |||
2382 | Size, llvm::ConstantInt::get(CGF.SizeTy, Align.getQuantity() - 1)); | |||
2383 | llvm::Value *AlignVal = | |||
2384 | llvm::ConstantInt::get(CGF.SizeTy, Align.getQuantity()); | |||
2385 | Size = Bld.CreateUDiv(Size, AlignVal); | |||
2386 | Size = Bld.CreateNUWMul(Size, AlignVal); | |||
2387 | // TODO: allow the usage of shared memory to be controlled by | |||
2388 | // the user, for now, default to global. | |||
2389 | llvm::Value *GlobalRecordSizeArg[] = { | |||
2390 | Size, CGF.Builder.getInt16(/*UseSharedMemory=*/0)}; | |||
2391 | llvm::Value *GlobalRecValue = CGF.EmitRuntimeCall( | |||
2392 | createNVPTXRuntimeFunction( | |||
2393 | OMPRTL_NVPTX__kmpc_data_sharing_coalesced_push_stack), | |||
2394 | GlobalRecordSizeArg); | |||
2395 | llvm::Value *GlobalRecCastAddr = Bld.CreatePointerBitCastOrAddrSpaceCast( | |||
2396 | GlobalRecValue, CGF.ConvertTypeForMem(VD->getType())->getPointerTo()); | |||
2397 | LValue Base = CGF.MakeAddrLValue(GlobalRecCastAddr, VD->getType(), | |||
2398 | CGM.getContext().getDeclAlign(VD), | |||
2399 | AlignmentSource::Decl); | |||
2400 | I->getSecond().MappedParams->setVarAddr(CGF, cast<VarDecl>(VD), | |||
2401 | Base.getAddress(CGF)); | |||
2402 | I->getSecond().EscapedVariableLengthDeclsAddrs.emplace_back(GlobalRecValue); | |||
2403 | } | |||
2404 | I->getSecond().MappedParams->apply(CGF); | |||
2405 | } | |||
2406 | ||||
2407 | void CGOpenMPRuntimeNVPTX::emitGenericVarsEpilog(CodeGenFunction &CGF, | |||
2408 | bool WithSPMDCheck) { | |||
2409 | if (getDataSharingMode(CGM) != CGOpenMPRuntimeNVPTX::Generic && | |||
2410 | getExecutionMode() != CGOpenMPRuntimeNVPTX::EM_SPMD) | |||
2411 | return; | |||
2412 | ||||
2413 | const auto I = FunctionGlobalizedDecls.find(CGF.CurFn); | |||
2414 | if (I != FunctionGlobalizedDecls.end()) { | |||
2415 | I->getSecond().MappedParams->restore(CGF); | |||
2416 | if (!CGF.HaveInsertPoint()) | |||
2417 | return; | |||
2418 | for (llvm::Value *Addr : | |||
2419 | llvm::reverse(I->getSecond().EscapedVariableLengthDeclsAddrs)) { | |||
2420 | CGF.EmitRuntimeCall( | |||
2421 | createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_data_sharing_pop_stack), | |||
2422 | Addr); | |||
2423 | } | |||
2424 | if (I->getSecond().GlobalRecordAddr) { | |||
2425 | if (!IsInTTDRegion && | |||
2426 | (WithSPMDCheck || | |||
2427 | getExecutionMode() == CGOpenMPRuntimeNVPTX::EM_Unknown)) { | |||
2428 | CGBuilderTy &Bld = CGF.Builder; | |||
2429 | llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".exit"); | |||
2430 | llvm::BasicBlock *NonSPMDBB = CGF.createBasicBlock(".non-spmd"); | |||
2431 | Bld.CreateCondBr(I->getSecond().IsInSPMDModeFlag, ExitBB, NonSPMDBB); | |||
2432 | // There is no need to emit line number for unconditional branch. | |||
2433 | (void)ApplyDebugLocation::CreateEmpty(CGF); | |||
2434 | CGF.EmitBlock(NonSPMDBB); | |||
2435 | CGF.EmitRuntimeCall( | |||
2436 | createNVPTXRuntimeFunction( | |||
2437 | OMPRTL_NVPTX__kmpc_data_sharing_pop_stack), | |||
2438 | CGF.EmitCastToVoidPtr(I->getSecond().GlobalRecordAddr)); | |||
2439 | CGF.EmitBlock(ExitBB); | |||
2440 | } else if (IsInTTDRegion) { | |||
2441 | assert(GlobalizedRecords.back().RegionCounter > 0 &&((GlobalizedRecords.back().RegionCounter > 0 && "region counter must be > 0." ) ? static_cast<void> (0) : __assert_fail ("GlobalizedRecords.back().RegionCounter > 0 && \"region counter must be > 0.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 2442, __PRETTY_FUNCTION__)) | |||
2442 | "region counter must be > 0.")((GlobalizedRecords.back().RegionCounter > 0 && "region counter must be > 0." ) ? static_cast<void> (0) : __assert_fail ("GlobalizedRecords.back().RegionCounter > 0 && \"region counter must be > 0.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 2442, __PRETTY_FUNCTION__)); | |||
2443 | --GlobalizedRecords.back().RegionCounter; | |||
2444 | // Emit the restore function only in the target region. | |||
2445 | if (GlobalizedRecords.back().RegionCounter == 0) { | |||
2446 | QualType Int16Ty = CGM.getContext().getIntTypeForBitwidth( | |||
2447 | /*DestWidth=*/16, /*Signed=*/0); | |||
2448 | llvm::Value *IsInSharedMemory = CGF.EmitLoadOfScalar( | |||
2449 | Address(GlobalizedRecords.back().UseSharedMemory, | |||
2450 | CGM.getContext().getTypeAlignInChars(Int16Ty)), | |||
2451 | /*Volatile=*/false, Int16Ty, GlobalizedRecords.back().Loc); | |||
2452 | llvm::Value *Args[] = { | |||
2453 | llvm::ConstantInt::get( | |||
2454 | CGM.Int16Ty, | |||
2455 | getExecutionMode() == CGOpenMPRuntimeNVPTX::EM_SPMD ? 1 : 0), | |||
2456 | IsInSharedMemory}; | |||
2457 | CGF.EmitRuntimeCall( | |||
2458 | createNVPTXRuntimeFunction( | |||
2459 | OMPRTL_NVPTX__kmpc_restore_team_static_memory), | |||
2460 | Args); | |||
2461 | } | |||
2462 | } else { | |||
2463 | CGF.EmitRuntimeCall(createNVPTXRuntimeFunction( | |||
2464 | OMPRTL_NVPTX__kmpc_data_sharing_pop_stack), | |||
2465 | I->getSecond().GlobalRecordAddr); | |||
2466 | } | |||
2467 | } | |||
2468 | } | |||
2469 | } | |||
2470 | ||||
2471 | void CGOpenMPRuntimeNVPTX::emitTeamsCall(CodeGenFunction &CGF, | |||
2472 | const OMPExecutableDirective &D, | |||
2473 | SourceLocation Loc, | |||
2474 | llvm::Function *OutlinedFn, | |||
2475 | ArrayRef<llvm::Value *> CapturedVars) { | |||
2476 | if (!CGF.HaveInsertPoint()) | |||
2477 | return; | |||
2478 | ||||
2479 | Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty, | |||
2480 | /*Name=*/".zero.addr"); | |||
2481 | CGF.InitTempAlloca(ZeroAddr, CGF.Builder.getInt32(/*C*/ 0)); | |||
2482 | llvm::SmallVector<llvm::Value *, 16> OutlinedFnArgs; | |||
2483 | OutlinedFnArgs.push_back(emitThreadIDAddress(CGF, Loc).getPointer()); | |||
2484 | OutlinedFnArgs.push_back(ZeroAddr.getPointer()); | |||
2485 | OutlinedFnArgs.append(CapturedVars.begin(), CapturedVars.end()); | |||
2486 | emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, OutlinedFnArgs); | |||
2487 | } | |||
2488 | ||||
2489 | void CGOpenMPRuntimeNVPTX::emitParallelCall( | |||
2490 | CodeGenFunction &CGF, SourceLocation Loc, llvm::Function *OutlinedFn, | |||
2491 | ArrayRef<llvm::Value *> CapturedVars, const Expr *IfCond) { | |||
2492 | if (!CGF.HaveInsertPoint()) | |||
2493 | return; | |||
2494 | ||||
2495 | if (getExecutionMode() == CGOpenMPRuntimeNVPTX::EM_SPMD) | |||
2496 | emitSPMDParallelCall(CGF, Loc, OutlinedFn, CapturedVars, IfCond); | |||
2497 | else | |||
2498 | emitNonSPMDParallelCall(CGF, Loc, OutlinedFn, CapturedVars, IfCond); | |||
2499 | } | |||
2500 | ||||
2501 | void CGOpenMPRuntimeNVPTX::emitNonSPMDParallelCall( | |||
2502 | CodeGenFunction &CGF, SourceLocation Loc, llvm::Value *OutlinedFn, | |||
2503 | ArrayRef<llvm::Value *> CapturedVars, const Expr *IfCond) { | |||
2504 | llvm::Function *Fn = cast<llvm::Function>(OutlinedFn); | |||
2505 | ||||
2506 | // Force inline this outlined function at its call site. | |||
2507 | Fn->setLinkage(llvm::GlobalValue::InternalLinkage); | |||
2508 | ||||
2509 | Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty, | |||
2510 | /*Name=*/".zero.addr"); | |||
2511 | CGF.InitTempAlloca(ZeroAddr, CGF.Builder.getInt32(/*C*/ 0)); | |||
2512 | // ThreadId for serialized parallels is 0. | |||
2513 | Address ThreadIDAddr = ZeroAddr; | |||
2514 | auto &&CodeGen = [this, Fn, CapturedVars, Loc, &ThreadIDAddr]( | |||
2515 | CodeGenFunction &CGF, PrePostActionTy &Action) { | |||
2516 | Action.Enter(CGF); | |||
2517 | ||||
2518 | Address ZeroAddr = | |||
2519 | CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty, | |||
2520 | /*Name=*/".bound.zero.addr"); | |||
2521 | CGF.InitTempAlloca(ZeroAddr, CGF.Builder.getInt32(/*C*/ 0)); | |||
2522 | llvm::SmallVector<llvm::Value *, 16> OutlinedFnArgs; | |||
2523 | OutlinedFnArgs.push_back(ThreadIDAddr.getPointer()); | |||
2524 | OutlinedFnArgs.push_back(ZeroAddr.getPointer()); | |||
2525 | OutlinedFnArgs.append(CapturedVars.begin(), CapturedVars.end()); | |||
2526 | emitOutlinedFunctionCall(CGF, Loc, Fn, OutlinedFnArgs); | |||
2527 | }; | |||
2528 | auto &&SeqGen = [this, &CodeGen, Loc](CodeGenFunction &CGF, | |||
2529 | PrePostActionTy &) { | |||
2530 | ||||
2531 | RegionCodeGenTy RCG(CodeGen); | |||
2532 | llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc); | |||
2533 | llvm::Value *ThreadID = getThreadID(CGF, Loc); | |||
2534 | llvm::Value *Args[] = {RTLoc, ThreadID}; | |||
2535 | ||||
2536 | NVPTXActionTy Action( | |||
2537 | createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_serialized_parallel), | |||
2538 | Args, | |||
2539 | createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_end_serialized_parallel), | |||
2540 | Args); | |||
2541 | RCG.setAction(Action); | |||
2542 | RCG(CGF); | |||
2543 | }; | |||
2544 | ||||
2545 | auto &&L0ParallelGen = [this, CapturedVars, Fn](CodeGenFunction &CGF, | |||
2546 | PrePostActionTy &Action) { | |||
2547 | CGBuilderTy &Bld = CGF.Builder; | |||
2548 | llvm::Function *WFn = WrapperFunctionsMap[Fn]; | |||
2549 | assert(WFn && "Wrapper function does not exist!")((WFn && "Wrapper function does not exist!") ? static_cast <void> (0) : __assert_fail ("WFn && \"Wrapper function does not exist!\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 2549, __PRETTY_FUNCTION__)); | |||
2550 | llvm::Value *ID = Bld.CreateBitOrPointerCast(WFn, CGM.Int8PtrTy); | |||
2551 | ||||
2552 | // Prepare for parallel region. Indicate the outlined function. | |||
2553 | llvm::Value *Args[] = {ID, /*RequiresOMPRuntime=*/Bld.getInt16(1)}; | |||
2554 | CGF.EmitRuntimeCall( | |||
2555 | createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_kernel_prepare_parallel), | |||
2556 | Args); | |||
2557 | ||||
2558 | // Create a private scope that will globalize the arguments | |||
2559 | // passed from the outside of the target region. | |||
2560 | CodeGenFunction::OMPPrivateScope PrivateArgScope(CGF); | |||
2561 | ||||
2562 | // There's something to share. | |||
2563 | if (!CapturedVars.empty()) { | |||
2564 | // Prepare for parallel region. Indicate the outlined function. | |||
2565 | Address SharedArgs = | |||
2566 | CGF.CreateDefaultAlignTempAlloca(CGF.VoidPtrPtrTy, "shared_arg_refs"); | |||
2567 | llvm::Value *SharedArgsPtr = SharedArgs.getPointer(); | |||
2568 | ||||
2569 | llvm::Value *DataSharingArgs[] = { | |||
2570 | SharedArgsPtr, | |||
2571 | llvm::ConstantInt::get(CGM.SizeTy, CapturedVars.size())}; | |||
2572 | CGF.EmitRuntimeCall(createNVPTXRuntimeFunction( | |||
2573 | OMPRTL_NVPTX__kmpc_begin_sharing_variables), | |||
2574 | DataSharingArgs); | |||
2575 | ||||
2576 | // Store variable address in a list of references to pass to workers. | |||
2577 | unsigned Idx = 0; | |||
2578 | ASTContext &Ctx = CGF.getContext(); | |||
2579 | Address SharedArgListAddress = CGF.EmitLoadOfPointer( | |||
2580 | SharedArgs, Ctx.getPointerType(Ctx.getPointerType(Ctx.VoidPtrTy)) | |||
2581 | .castAs<PointerType>()); | |||
2582 | for (llvm::Value *V : CapturedVars) { | |||
2583 | Address Dst = Bld.CreateConstInBoundsGEP(SharedArgListAddress, Idx); | |||
2584 | llvm::Value *PtrV; | |||
2585 | if (V->getType()->isIntegerTy()) | |||
2586 | PtrV = Bld.CreateIntToPtr(V, CGF.VoidPtrTy); | |||
2587 | else | |||
2588 | PtrV = Bld.CreatePointerBitCastOrAddrSpaceCast(V, CGF.VoidPtrTy); | |||
2589 | CGF.EmitStoreOfScalar(PtrV, Dst, /*Volatile=*/false, | |||
2590 | Ctx.getPointerType(Ctx.VoidPtrTy)); | |||
2591 | ++Idx; | |||
2592 | } | |||
2593 | } | |||
2594 | ||||
2595 | // Activate workers. This barrier is used by the master to signal | |||
2596 | // work for the workers. | |||
2597 | syncCTAThreads(CGF); | |||
2598 | ||||
2599 | // OpenMP [2.5, Parallel Construct, p.49] | |||
2600 | // There is an implied barrier at the end of a parallel region. After the | |||
2601 | // end of a parallel region, only the master thread of the team resumes | |||
2602 | // execution of the enclosing task region. | |||
2603 | // | |||
2604 | // The master waits at this barrier until all workers are done. | |||
2605 | syncCTAThreads(CGF); | |||
2606 | ||||
2607 | if (!CapturedVars.empty()) | |||
2608 | CGF.EmitRuntimeCall( | |||
2609 | createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_end_sharing_variables)); | |||
2610 | ||||
2611 | // Remember for post-processing in worker loop. | |||
2612 | Work.emplace_back(WFn); | |||
2613 | }; | |||
2614 | ||||
2615 | auto &&LNParallelGen = [this, Loc, &SeqGen, &L0ParallelGen]( | |||
2616 | CodeGenFunction &CGF, PrePostActionTy &Action) { | |||
2617 | if (IsInParallelRegion) { | |||
2618 | SeqGen(CGF, Action); | |||
2619 | } else if (IsInTargetMasterThreadRegion) { | |||
2620 | L0ParallelGen(CGF, Action); | |||
2621 | } else { | |||
2622 | // Check for master and then parallelism: | |||
2623 | // if (__kmpc_is_spmd_exec_mode() || __kmpc_parallel_level(loc, gtid)) { | |||
2624 | // Serialized execution. | |||
2625 | // } else { | |||
2626 | // Worker call. | |||
2627 | // } | |||
2628 | CGBuilderTy &Bld = CGF.Builder; | |||
2629 | llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".exit"); | |||
2630 | llvm::BasicBlock *SeqBB = CGF.createBasicBlock(".sequential"); | |||
2631 | llvm::BasicBlock *ParallelCheckBB = CGF.createBasicBlock(".parcheck"); | |||
2632 | llvm::BasicBlock *MasterBB = CGF.createBasicBlock(".master"); | |||
2633 | llvm::Value *IsSPMD = Bld.CreateIsNotNull(CGF.EmitNounwindRuntimeCall( | |||
2634 | createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_is_spmd_exec_mode))); | |||
2635 | Bld.CreateCondBr(IsSPMD, SeqBB, ParallelCheckBB); | |||
2636 | // There is no need to emit line number for unconditional branch. | |||
2637 | (void)ApplyDebugLocation::CreateEmpty(CGF); | |||
2638 | CGF.EmitBlock(ParallelCheckBB); | |||
2639 | llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc); | |||
2640 | llvm::Value *ThreadID = getThreadID(CGF, Loc); | |||
2641 | llvm::Value *PL = CGF.EmitRuntimeCall( | |||
2642 | createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_parallel_level), | |||
2643 | {RTLoc, ThreadID}); | |||
2644 | llvm::Value *Res = Bld.CreateIsNotNull(PL); | |||
2645 | Bld.CreateCondBr(Res, SeqBB, MasterBB); | |||
2646 | CGF.EmitBlock(SeqBB); | |||
2647 | SeqGen(CGF, Action); | |||
2648 | CGF.EmitBranch(ExitBB); | |||
2649 | // There is no need to emit line number for unconditional branch. | |||
2650 | (void)ApplyDebugLocation::CreateEmpty(CGF); | |||
2651 | CGF.EmitBlock(MasterBB); | |||
2652 | L0ParallelGen(CGF, Action); | |||
2653 | CGF.EmitBranch(ExitBB); | |||
2654 | // There is no need to emit line number for unconditional branch. | |||
2655 | (void)ApplyDebugLocation::CreateEmpty(CGF); | |||
2656 | // Emit the continuation block for code after the if. | |||
2657 | CGF.EmitBlock(ExitBB, /*IsFinished=*/true); | |||
2658 | } | |||
2659 | }; | |||
2660 | ||||
2661 | if (IfCond) { | |||
2662 | emitIfClause(CGF, IfCond, LNParallelGen, SeqGen); | |||
2663 | } else { | |||
2664 | CodeGenFunction::RunCleanupsScope Scope(CGF); | |||
2665 | RegionCodeGenTy ThenRCG(LNParallelGen); | |||
2666 | ThenRCG(CGF); | |||
2667 | } | |||
2668 | } | |||
2669 | ||||
2670 | void CGOpenMPRuntimeNVPTX::emitSPMDParallelCall( | |||
2671 | CodeGenFunction &CGF, SourceLocation Loc, llvm::Function *OutlinedFn, | |||
2672 | ArrayRef<llvm::Value *> CapturedVars, const Expr *IfCond) { | |||
2673 | // Just call the outlined function to execute the parallel region. | |||
2674 | // OutlinedFn(>id, &zero, CapturedStruct); | |||
2675 | // | |||
2676 | llvm::SmallVector<llvm::Value *, 16> OutlinedFnArgs; | |||
2677 | ||||
2678 | Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty, | |||
2679 | /*Name=*/".zero.addr"); | |||
2680 | CGF.InitTempAlloca(ZeroAddr, CGF.Builder.getInt32(/*C*/ 0)); | |||
2681 | // ThreadId for serialized parallels is 0. | |||
2682 | Address ThreadIDAddr = ZeroAddr; | |||
2683 | auto &&CodeGen = [this, OutlinedFn, CapturedVars, Loc, &ThreadIDAddr]( | |||
2684 | CodeGenFunction &CGF, PrePostActionTy &Action) { | |||
2685 | Action.Enter(CGF); | |||
2686 | ||||
2687 | Address ZeroAddr = | |||
2688 | CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty, | |||
2689 | /*Name=*/".bound.zero.addr"); | |||
2690 | CGF.InitTempAlloca(ZeroAddr, CGF.Builder.getInt32(/*C*/ 0)); | |||
2691 | llvm::SmallVector<llvm::Value *, 16> OutlinedFnArgs; | |||
2692 | OutlinedFnArgs.push_back(ThreadIDAddr.getPointer()); | |||
2693 | OutlinedFnArgs.push_back(ZeroAddr.getPointer()); | |||
2694 | OutlinedFnArgs.append(CapturedVars.begin(), CapturedVars.end()); | |||
2695 | emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, OutlinedFnArgs); | |||
2696 | }; | |||
2697 | auto &&SeqGen = [this, &CodeGen, Loc](CodeGenFunction &CGF, | |||
2698 | PrePostActionTy &) { | |||
2699 | ||||
2700 | RegionCodeGenTy RCG(CodeGen); | |||
2701 | llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc); | |||
2702 | llvm::Value *ThreadID = getThreadID(CGF, Loc); | |||
2703 | llvm::Value *Args[] = {RTLoc, ThreadID}; | |||
2704 | ||||
2705 | NVPTXActionTy Action( | |||
2706 | createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_serialized_parallel), | |||
2707 | Args, | |||
2708 | createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_end_serialized_parallel), | |||
2709 | Args); | |||
2710 | RCG.setAction(Action); | |||
2711 | RCG(CGF); | |||
2712 | }; | |||
2713 | ||||
2714 | if (IsInTargetMasterThreadRegion) { | |||
2715 | // In the worker need to use the real thread id. | |||
2716 | ThreadIDAddr = emitThreadIDAddress(CGF, Loc); | |||
2717 | RegionCodeGenTy RCG(CodeGen); | |||
2718 | RCG(CGF); | |||
2719 | } else { | |||
2720 | // If we are not in the target region, it is definitely L2 parallelism or | |||
2721 | // more, because for SPMD mode we always has L1 parallel level, sowe don't | |||
2722 | // need to check for orphaned directives. | |||
2723 | RegionCodeGenTy RCG(SeqGen); | |||
2724 | RCG(CGF); | |||
2725 | } | |||
2726 | } | |||
2727 | ||||
2728 | void CGOpenMPRuntimeNVPTX::syncCTAThreads(CodeGenFunction &CGF) { | |||
2729 | // Always emit simple barriers! | |||
2730 | if (!CGF.HaveInsertPoint()) | |||
2731 | return; | |||
2732 | // Build call __kmpc_barrier_simple_spmd(nullptr, 0); | |||
2733 | // This function does not use parameters, so we can emit just default values. | |||
2734 | llvm::Value *Args[] = { | |||
2735 | llvm::ConstantPointerNull::get( | |||
2736 | cast<llvm::PointerType>(getIdentTyPointerTy())), | |||
2737 | llvm::ConstantInt::get(CGF.Int32Ty, /*V=*/0, /*isSigned=*/true)}; | |||
2738 | llvm::CallInst *Call = CGF.EmitRuntimeCall( | |||
2739 | createNVPTXRuntimeFunction(OMPRTL__kmpc_barrier_simple_spmd), Args); | |||
2740 | Call->setConvergent(); | |||
2741 | } | |||
2742 | ||||
2743 | void CGOpenMPRuntimeNVPTX::emitBarrierCall(CodeGenFunction &CGF, | |||
2744 | SourceLocation Loc, | |||
2745 | OpenMPDirectiveKind Kind, bool, | |||
2746 | bool) { | |||
2747 | // Always emit simple barriers! | |||
2748 | if (!CGF.HaveInsertPoint()) | |||
2749 | return; | |||
2750 | // Build call __kmpc_cancel_barrier(loc, thread_id); | |||
2751 | unsigned Flags = getDefaultFlagsForBarriers(Kind); | |||
2752 | llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc, Flags), | |||
2753 | getThreadID(CGF, Loc)}; | |||
2754 | llvm::CallInst *Call = CGF.EmitRuntimeCall( | |||
2755 | createNVPTXRuntimeFunction(OMPRTL__kmpc_barrier), Args); | |||
2756 | Call->setConvergent(); | |||
2757 | } | |||
2758 | ||||
2759 | void CGOpenMPRuntimeNVPTX::emitCriticalRegion( | |||
2760 | CodeGenFunction &CGF, StringRef CriticalName, | |||
2761 | const RegionCodeGenTy &CriticalOpGen, SourceLocation Loc, | |||
2762 | const Expr *Hint) { | |||
2763 | llvm::BasicBlock *LoopBB = CGF.createBasicBlock("omp.critical.loop"); | |||
2764 | llvm::BasicBlock *TestBB = CGF.createBasicBlock("omp.critical.test"); | |||
2765 | llvm::BasicBlock *SyncBB = CGF.createBasicBlock("omp.critical.sync"); | |||
2766 | llvm::BasicBlock *BodyBB = CGF.createBasicBlock("omp.critical.body"); | |||
2767 | llvm::BasicBlock *ExitBB = CGF.createBasicBlock("omp.critical.exit"); | |||
2768 | ||||
2769 | // Get the mask of active threads in the warp. | |||
2770 | llvm::Value *Mask = CGF.EmitRuntimeCall( | |||
2771 | createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_warp_active_thread_mask)); | |||
2772 | // Fetch team-local id of the thread. | |||
2773 | llvm::Value *ThreadID = getNVPTXThreadID(CGF); | |||
2774 | ||||
2775 | // Get the width of the team. | |||
2776 | llvm::Value *TeamWidth = getNVPTXNumThreads(CGF); | |||
2777 | ||||
2778 | // Initialize the counter variable for the loop. | |||
2779 | QualType Int32Ty = | |||
2780 | CGF.getContext().getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/0); | |||
2781 | Address Counter = CGF.CreateMemTemp(Int32Ty, "critical_counter"); | |||
2782 | LValue CounterLVal = CGF.MakeAddrLValue(Counter, Int32Ty); | |||
2783 | CGF.EmitStoreOfScalar(llvm::Constant::getNullValue(CGM.Int32Ty), CounterLVal, | |||
2784 | /*isInit=*/true); | |||
2785 | ||||
2786 | // Block checks if loop counter exceeds upper bound. | |||
2787 | CGF.EmitBlock(LoopBB); | |||
2788 | llvm::Value *CounterVal = CGF.EmitLoadOfScalar(CounterLVal, Loc); | |||
2789 | llvm::Value *CmpLoopBound = CGF.Builder.CreateICmpSLT(CounterVal, TeamWidth); | |||
2790 | CGF.Builder.CreateCondBr(CmpLoopBound, TestBB, ExitBB); | |||
2791 | ||||
2792 | // Block tests which single thread should execute region, and which threads | |||
2793 | // should go straight to synchronisation point. | |||
2794 | CGF.EmitBlock(TestBB); | |||
2795 | CounterVal = CGF.EmitLoadOfScalar(CounterLVal, Loc); | |||
2796 | llvm::Value *CmpThreadToCounter = | |||
2797 | CGF.Builder.CreateICmpEQ(ThreadID, CounterVal); | |||
2798 | CGF.Builder.CreateCondBr(CmpThreadToCounter, BodyBB, SyncBB); | |||
2799 | ||||
2800 | // Block emits the body of the critical region. | |||
2801 | CGF.EmitBlock(BodyBB); | |||
2802 | ||||
2803 | // Output the critical statement. | |||
2804 | CGOpenMPRuntime::emitCriticalRegion(CGF, CriticalName, CriticalOpGen, Loc, | |||
2805 | Hint); | |||
2806 | ||||
2807 | // After the body surrounded by the critical region, the single executing | |||
2808 | // thread will jump to the synchronisation point. | |||
2809 | // Block waits for all threads in current team to finish then increments the | |||
2810 | // counter variable and returns to the loop. | |||
2811 | CGF.EmitBlock(SyncBB); | |||
2812 | // Reconverge active threads in the warp. | |||
2813 | (void)CGF.EmitRuntimeCall( | |||
2814 | createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_syncwarp), Mask); | |||
2815 | ||||
2816 | llvm::Value *IncCounterVal = | |||
2817 | CGF.Builder.CreateNSWAdd(CounterVal, CGF.Builder.getInt32(1)); | |||
2818 | CGF.EmitStoreOfScalar(IncCounterVal, CounterLVal); | |||
2819 | CGF.EmitBranch(LoopBB); | |||
2820 | ||||
2821 | // Block that is reached when all threads in the team complete the region. | |||
2822 | CGF.EmitBlock(ExitBB, /*IsFinished=*/true); | |||
2823 | } | |||
2824 | ||||
2825 | /// Cast value to the specified type. | |||
2826 | static llvm::Value *castValueToType(CodeGenFunction &CGF, llvm::Value *Val, | |||
2827 | QualType ValTy, QualType CastTy, | |||
2828 | SourceLocation Loc) { | |||
2829 | assert(!CGF.getContext().getTypeSizeInChars(CastTy).isZero() &&((!CGF.getContext().getTypeSizeInChars(CastTy).isZero() && "Cast type must sized.") ? static_cast<void> (0) : __assert_fail ("!CGF.getContext().getTypeSizeInChars(CastTy).isZero() && \"Cast type must sized.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 2830, __PRETTY_FUNCTION__)) | |||
2830 | "Cast type must sized.")((!CGF.getContext().getTypeSizeInChars(CastTy).isZero() && "Cast type must sized.") ? static_cast<void> (0) : __assert_fail ("!CGF.getContext().getTypeSizeInChars(CastTy).isZero() && \"Cast type must sized.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 2830, __PRETTY_FUNCTION__)); | |||
2831 | assert(!CGF.getContext().getTypeSizeInChars(ValTy).isZero() &&((!CGF.getContext().getTypeSizeInChars(ValTy).isZero() && "Val type must sized.") ? static_cast<void> (0) : __assert_fail ("!CGF.getContext().getTypeSizeInChars(ValTy).isZero() && \"Val type must sized.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 2832, __PRETTY_FUNCTION__)) | |||
2832 | "Val type must sized.")((!CGF.getContext().getTypeSizeInChars(ValTy).isZero() && "Val type must sized.") ? static_cast<void> (0) : __assert_fail ("!CGF.getContext().getTypeSizeInChars(ValTy).isZero() && \"Val type must sized.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 2832, __PRETTY_FUNCTION__)); | |||
2833 | llvm::Type *LLVMCastTy = CGF.ConvertTypeForMem(CastTy); | |||
2834 | if (ValTy == CastTy) | |||
2835 | return Val; | |||
2836 | if (CGF.getContext().getTypeSizeInChars(ValTy) == | |||
2837 | CGF.getContext().getTypeSizeInChars(CastTy)) | |||
2838 | return CGF.Builder.CreateBitCast(Val, LLVMCastTy); | |||
2839 | if (CastTy->isIntegerType() && ValTy->isIntegerType()) | |||
2840 | return CGF.Builder.CreateIntCast(Val, LLVMCastTy, | |||
2841 | CastTy->hasSignedIntegerRepresentation()); | |||
2842 | Address CastItem = CGF.CreateMemTemp(CastTy); | |||
2843 | Address ValCastItem = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
2844 | CastItem, Val->getType()->getPointerTo(CastItem.getAddressSpace())); | |||
2845 | CGF.EmitStoreOfScalar(Val, ValCastItem, /*Volatile=*/false, ValTy); | |||
2846 | return CGF.EmitLoadOfScalar(CastItem, /*Volatile=*/false, CastTy, Loc); | |||
2847 | } | |||
2848 | ||||
2849 | /// This function creates calls to one of two shuffle functions to copy | |||
2850 | /// variables between lanes in a warp. | |||
2851 | static llvm::Value *createRuntimeShuffleFunction(CodeGenFunction &CGF, | |||
2852 | llvm::Value *Elem, | |||
2853 | QualType ElemType, | |||
2854 | llvm::Value *Offset, | |||
2855 | SourceLocation Loc) { | |||
2856 | CodeGenModule &CGM = CGF.CGM; | |||
2857 | CGBuilderTy &Bld = CGF.Builder; | |||
2858 | CGOpenMPRuntimeNVPTX &RT = | |||
2859 | *(static_cast<CGOpenMPRuntimeNVPTX *>(&CGM.getOpenMPRuntime())); | |||
2860 | ||||
2861 | CharUnits Size = CGF.getContext().getTypeSizeInChars(ElemType); | |||
2862 | assert(Size.getQuantity() <= 8 &&((Size.getQuantity() <= 8 && "Unsupported bitwidth in shuffle instruction." ) ? static_cast<void> (0) : __assert_fail ("Size.getQuantity() <= 8 && \"Unsupported bitwidth in shuffle instruction.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 2863, __PRETTY_FUNCTION__)) | |||
2863 | "Unsupported bitwidth in shuffle instruction.")((Size.getQuantity() <= 8 && "Unsupported bitwidth in shuffle instruction." ) ? static_cast<void> (0) : __assert_fail ("Size.getQuantity() <= 8 && \"Unsupported bitwidth in shuffle instruction.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 2863, __PRETTY_FUNCTION__)); | |||
2864 | ||||
2865 | OpenMPRTLFunctionNVPTX ShuffleFn = Size.getQuantity() <= 4 | |||
2866 | ? OMPRTL_NVPTX__kmpc_shuffle_int32 | |||
2867 | : OMPRTL_NVPTX__kmpc_shuffle_int64; | |||
2868 | ||||
2869 | // Cast all types to 32- or 64-bit values before calling shuffle routines. | |||
2870 | QualType CastTy = CGF.getContext().getIntTypeForBitwidth( | |||
2871 | Size.getQuantity() <= 4 ? 32 : 64, /*Signed=*/1); | |||
2872 | llvm::Value *ElemCast = castValueToType(CGF, Elem, ElemType, CastTy, Loc); | |||
2873 | llvm::Value *WarpSize = | |||
2874 | Bld.CreateIntCast(getNVPTXWarpSize(CGF), CGM.Int16Ty, /*isSigned=*/true); | |||
2875 | ||||
2876 | llvm::Value *ShuffledVal = CGF.EmitRuntimeCall( | |||
2877 | RT.createNVPTXRuntimeFunction(ShuffleFn), {ElemCast, Offset, WarpSize}); | |||
2878 | ||||
2879 | return castValueToType(CGF, ShuffledVal, CastTy, ElemType, Loc); | |||
2880 | } | |||
2881 | ||||
2882 | static void shuffleAndStore(CodeGenFunction &CGF, Address SrcAddr, | |||
2883 | Address DestAddr, QualType ElemType, | |||
2884 | llvm::Value *Offset, SourceLocation Loc) { | |||
2885 | CGBuilderTy &Bld = CGF.Builder; | |||
2886 | ||||
2887 | CharUnits Size = CGF.getContext().getTypeSizeInChars(ElemType); | |||
2888 | // Create the loop over the big sized data. | |||
2889 | // ptr = (void*)Elem; | |||
2890 | // ptrEnd = (void*) Elem + 1; | |||
2891 | // Step = 8; | |||
2892 | // while (ptr + Step < ptrEnd) | |||
2893 | // shuffle((int64_t)*ptr); | |||
2894 | // Step = 4; | |||
2895 | // while (ptr + Step < ptrEnd) | |||
2896 | // shuffle((int32_t)*ptr); | |||
2897 | // ... | |||
2898 | Address ElemPtr = DestAddr; | |||
2899 | Address Ptr = SrcAddr; | |||
2900 | Address PtrEnd = Bld.CreatePointerBitCastOrAddrSpaceCast( | |||
2901 | Bld.CreateConstGEP(SrcAddr, 1), CGF.VoidPtrTy); | |||
2902 | for (int IntSize = 8; IntSize >= 1; IntSize /= 2) { | |||
2903 | if (Size < CharUnits::fromQuantity(IntSize)) | |||
2904 | continue; | |||
2905 | QualType IntType = CGF.getContext().getIntTypeForBitwidth( | |||
2906 | CGF.getContext().toBits(CharUnits::fromQuantity(IntSize)), | |||
2907 | /*Signed=*/1); | |||
2908 | llvm::Type *IntTy = CGF.ConvertTypeForMem(IntType); | |||
2909 | Ptr = Bld.CreatePointerBitCastOrAddrSpaceCast(Ptr, IntTy->getPointerTo()); | |||
2910 | ElemPtr = | |||
2911 | Bld.CreatePointerBitCastOrAddrSpaceCast(ElemPtr, IntTy->getPointerTo()); | |||
2912 | if (Size.getQuantity() / IntSize > 1) { | |||
2913 | llvm::BasicBlock *PreCondBB = CGF.createBasicBlock(".shuffle.pre_cond"); | |||
2914 | llvm::BasicBlock *ThenBB = CGF.createBasicBlock(".shuffle.then"); | |||
2915 | llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".shuffle.exit"); | |||
2916 | llvm::BasicBlock *CurrentBB = Bld.GetInsertBlock(); | |||
2917 | CGF.EmitBlock(PreCondBB); | |||
2918 | llvm::PHINode *PhiSrc = | |||
2919 | Bld.CreatePHI(Ptr.getType(), /*NumReservedValues=*/2); | |||
2920 | PhiSrc->addIncoming(Ptr.getPointer(), CurrentBB); | |||
2921 | llvm::PHINode *PhiDest = | |||
2922 | Bld.CreatePHI(ElemPtr.getType(), /*NumReservedValues=*/2); | |||
2923 | PhiDest->addIncoming(ElemPtr.getPointer(), CurrentBB); | |||
2924 | Ptr = Address(PhiSrc, Ptr.getAlignment()); | |||
2925 | ElemPtr = Address(PhiDest, ElemPtr.getAlignment()); | |||
2926 | llvm::Value *PtrDiff = Bld.CreatePtrDiff( | |||
2927 | PtrEnd.getPointer(), Bld.CreatePointerBitCastOrAddrSpaceCast( | |||
2928 | Ptr.getPointer(), CGF.VoidPtrTy)); | |||
2929 | Bld.CreateCondBr(Bld.CreateICmpSGT(PtrDiff, Bld.getInt64(IntSize - 1)), | |||
2930 | ThenBB, ExitBB); | |||
2931 | CGF.EmitBlock(ThenBB); | |||
2932 | llvm::Value *Res = createRuntimeShuffleFunction( | |||
2933 | CGF, CGF.EmitLoadOfScalar(Ptr, /*Volatile=*/false, IntType, Loc), | |||
2934 | IntType, Offset, Loc); | |||
2935 | CGF.EmitStoreOfScalar(Res, ElemPtr, /*Volatile=*/false, IntType); | |||
2936 | Address LocalPtr = Bld.CreateConstGEP(Ptr, 1); | |||
2937 | Address LocalElemPtr = Bld.CreateConstGEP(ElemPtr, 1); | |||
2938 | PhiSrc->addIncoming(LocalPtr.getPointer(), ThenBB); | |||
2939 | PhiDest->addIncoming(LocalElemPtr.getPointer(), ThenBB); | |||
2940 | CGF.EmitBranch(PreCondBB); | |||
2941 | CGF.EmitBlock(ExitBB); | |||
2942 | } else { | |||
2943 | llvm::Value *Res = createRuntimeShuffleFunction( | |||
2944 | CGF, CGF.EmitLoadOfScalar(Ptr, /*Volatile=*/false, IntType, Loc), | |||
2945 | IntType, Offset, Loc); | |||
2946 | CGF.EmitStoreOfScalar(Res, ElemPtr, /*Volatile=*/false, IntType); | |||
2947 | Ptr = Bld.CreateConstGEP(Ptr, 1); | |||
2948 | ElemPtr = Bld.CreateConstGEP(ElemPtr, 1); | |||
2949 | } | |||
2950 | Size = Size % IntSize; | |||
2951 | } | |||
2952 | } | |||
2953 | ||||
2954 | namespace { | |||
2955 | enum CopyAction : unsigned { | |||
2956 | // RemoteLaneToThread: Copy over a Reduce list from a remote lane in | |||
2957 | // the warp using shuffle instructions. | |||
2958 | RemoteLaneToThread, | |||
2959 | // ThreadCopy: Make a copy of a Reduce list on the thread's stack. | |||
2960 | ThreadCopy, | |||
2961 | // ThreadToScratchpad: Copy a team-reduced array to the scratchpad. | |||
2962 | ThreadToScratchpad, | |||
2963 | // ScratchpadToThread: Copy from a scratchpad array in global memory | |||
2964 | // containing team-reduced data to a thread's stack. | |||
2965 | ScratchpadToThread, | |||
2966 | }; | |||
2967 | } // namespace | |||
2968 | ||||
2969 | struct CopyOptionsTy { | |||
2970 | llvm::Value *RemoteLaneOffset; | |||
2971 | llvm::Value *ScratchpadIndex; | |||
2972 | llvm::Value *ScratchpadWidth; | |||
2973 | }; | |||
2974 | ||||
2975 | /// Emit instructions to copy a Reduce list, which contains partially | |||
2976 | /// aggregated values, in the specified direction. | |||
2977 | static void emitReductionListCopy( | |||
2978 | CopyAction Action, CodeGenFunction &CGF, QualType ReductionArrayTy, | |||
2979 | ArrayRef<const Expr *> Privates, Address SrcBase, Address DestBase, | |||
2980 | CopyOptionsTy CopyOptions = {nullptr, nullptr, nullptr}) { | |||
2981 | ||||
2982 | CodeGenModule &CGM = CGF.CGM; | |||
2983 | ASTContext &C = CGM.getContext(); | |||
2984 | CGBuilderTy &Bld = CGF.Builder; | |||
2985 | ||||
2986 | llvm::Value *RemoteLaneOffset = CopyOptions.RemoteLaneOffset; | |||
2987 | llvm::Value *ScratchpadIndex = CopyOptions.ScratchpadIndex; | |||
2988 | llvm::Value *ScratchpadWidth = CopyOptions.ScratchpadWidth; | |||
2989 | ||||
2990 | // Iterates, element-by-element, through the source Reduce list and | |||
2991 | // make a copy. | |||
2992 | unsigned Idx = 0; | |||
2993 | unsigned Size = Privates.size(); | |||
2994 | for (const Expr *Private : Privates) { | |||
2995 | Address SrcElementAddr = Address::invalid(); | |||
2996 | Address DestElementAddr = Address::invalid(); | |||
2997 | Address DestElementPtrAddr = Address::invalid(); | |||
2998 | // Should we shuffle in an element from a remote lane? | |||
2999 | bool ShuffleInElement = false; | |||
3000 | // Set to true to update the pointer in the dest Reduce list to a | |||
3001 | // newly created element. | |||
3002 | bool UpdateDestListPtr = false; | |||
3003 | // Increment the src or dest pointer to the scratchpad, for each | |||
3004 | // new element. | |||
3005 | bool IncrScratchpadSrc = false; | |||
3006 | bool IncrScratchpadDest = false; | |||
3007 | ||||
3008 | switch (Action) { | |||
3009 | case RemoteLaneToThread: { | |||
3010 | // Step 1.1: Get the address for the src element in the Reduce list. | |||
3011 | Address SrcElementPtrAddr = Bld.CreateConstArrayGEP(SrcBase, Idx); | |||
3012 | SrcElementAddr = CGF.EmitLoadOfPointer( | |||
3013 | SrcElementPtrAddr, | |||
3014 | C.getPointerType(Private->getType())->castAs<PointerType>()); | |||
3015 | ||||
3016 | // Step 1.2: Create a temporary to store the element in the destination | |||
3017 | // Reduce list. | |||
3018 | DestElementPtrAddr = Bld.CreateConstArrayGEP(DestBase, Idx); | |||
3019 | DestElementAddr = | |||
3020 | CGF.CreateMemTemp(Private->getType(), ".omp.reduction.element"); | |||
3021 | ShuffleInElement = true; | |||
3022 | UpdateDestListPtr = true; | |||
3023 | break; | |||
3024 | } | |||
3025 | case ThreadCopy: { | |||
3026 | // Step 1.1: Get the address for the src element in the Reduce list. | |||
3027 | Address SrcElementPtrAddr = Bld.CreateConstArrayGEP(SrcBase, Idx); | |||
3028 | SrcElementAddr = CGF.EmitLoadOfPointer( | |||
3029 | SrcElementPtrAddr, | |||
3030 | C.getPointerType(Private->getType())->castAs<PointerType>()); | |||
3031 | ||||
3032 | // Step 1.2: Get the address for dest element. The destination | |||
3033 | // element has already been created on the thread's stack. | |||
3034 | DestElementPtrAddr = Bld.CreateConstArrayGEP(DestBase, Idx); | |||
3035 | DestElementAddr = CGF.EmitLoadOfPointer( | |||
3036 | DestElementPtrAddr, | |||
3037 | C.getPointerType(Private->getType())->castAs<PointerType>()); | |||
3038 | break; | |||
3039 | } | |||
3040 | case ThreadToScratchpad: { | |||
3041 | // Step 1.1: Get the address for the src element in the Reduce list. | |||
3042 | Address SrcElementPtrAddr = Bld.CreateConstArrayGEP(SrcBase, Idx); | |||
3043 | SrcElementAddr = CGF.EmitLoadOfPointer( | |||
3044 | SrcElementPtrAddr, | |||
3045 | C.getPointerType(Private->getType())->castAs<PointerType>()); | |||
3046 | ||||
3047 | // Step 1.2: Get the address for dest element: | |||
3048 | // address = base + index * ElementSizeInChars. | |||
3049 | llvm::Value *ElementSizeInChars = CGF.getTypeSize(Private->getType()); | |||
3050 | llvm::Value *CurrentOffset = | |||
3051 | Bld.CreateNUWMul(ElementSizeInChars, ScratchpadIndex); | |||
3052 | llvm::Value *ScratchPadElemAbsolutePtrVal = | |||
3053 | Bld.CreateNUWAdd(DestBase.getPointer(), CurrentOffset); | |||
3054 | ScratchPadElemAbsolutePtrVal = | |||
3055 | Bld.CreateIntToPtr(ScratchPadElemAbsolutePtrVal, CGF.VoidPtrTy); | |||
3056 | DestElementAddr = Address(ScratchPadElemAbsolutePtrVal, | |||
3057 | C.getTypeAlignInChars(Private->getType())); | |||
3058 | IncrScratchpadDest = true; | |||
3059 | break; | |||
3060 | } | |||
3061 | case ScratchpadToThread: { | |||
3062 | // Step 1.1: Get the address for the src element in the scratchpad. | |||
3063 | // address = base + index * ElementSizeInChars. | |||
3064 | llvm::Value *ElementSizeInChars = CGF.getTypeSize(Private->getType()); | |||
3065 | llvm::Value *CurrentOffset = | |||
3066 | Bld.CreateNUWMul(ElementSizeInChars, ScratchpadIndex); | |||
3067 | llvm::Value *ScratchPadElemAbsolutePtrVal = | |||
3068 | Bld.CreateNUWAdd(SrcBase.getPointer(), CurrentOffset); | |||
3069 | ScratchPadElemAbsolutePtrVal = | |||
3070 | Bld.CreateIntToPtr(ScratchPadElemAbsolutePtrVal, CGF.VoidPtrTy); | |||
3071 | SrcElementAddr = Address(ScratchPadElemAbsolutePtrVal, | |||
3072 | C.getTypeAlignInChars(Private->getType())); | |||
3073 | IncrScratchpadSrc = true; | |||
3074 | ||||
3075 | // Step 1.2: Create a temporary to store the element in the destination | |||
3076 | // Reduce list. | |||
3077 | DestElementPtrAddr = Bld.CreateConstArrayGEP(DestBase, Idx); | |||
3078 | DestElementAddr = | |||
3079 | CGF.CreateMemTemp(Private->getType(), ".omp.reduction.element"); | |||
3080 | UpdateDestListPtr = true; | |||
3081 | break; | |||
3082 | } | |||
3083 | } | |||
3084 | ||||
3085 | // Regardless of src and dest of copy, we emit the load of src | |||
3086 | // element as this is required in all directions | |||
3087 | SrcElementAddr = Bld.CreateElementBitCast( | |||
3088 | SrcElementAddr, CGF.ConvertTypeForMem(Private->getType())); | |||
3089 | DestElementAddr = Bld.CreateElementBitCast(DestElementAddr, | |||
3090 | SrcElementAddr.getElementType()); | |||
3091 | ||||
3092 | // Now that all active lanes have read the element in the | |||
3093 | // Reduce list, shuffle over the value from the remote lane. | |||
3094 | if (ShuffleInElement) { | |||
3095 | shuffleAndStore(CGF, SrcElementAddr, DestElementAddr, Private->getType(), | |||
3096 | RemoteLaneOffset, Private->getExprLoc()); | |||
3097 | } else { | |||
3098 | switch (CGF.getEvaluationKind(Private->getType())) { | |||
3099 | case TEK_Scalar: { | |||
3100 | llvm::Value *Elem = | |||
3101 | CGF.EmitLoadOfScalar(SrcElementAddr, /*Volatile=*/false, | |||
3102 | Private->getType(), Private->getExprLoc()); | |||
3103 | // Store the source element value to the dest element address. | |||
3104 | CGF.EmitStoreOfScalar(Elem, DestElementAddr, /*Volatile=*/false, | |||
3105 | Private->getType()); | |||
3106 | break; | |||
3107 | } | |||
3108 | case TEK_Complex: { | |||
3109 | CodeGenFunction::ComplexPairTy Elem = CGF.EmitLoadOfComplex( | |||
3110 | CGF.MakeAddrLValue(SrcElementAddr, Private->getType()), | |||
3111 | Private->getExprLoc()); | |||
3112 | CGF.EmitStoreOfComplex( | |||
3113 | Elem, CGF.MakeAddrLValue(DestElementAddr, Private->getType()), | |||
3114 | /*isInit=*/false); | |||
3115 | break; | |||
3116 | } | |||
3117 | case TEK_Aggregate: | |||
3118 | CGF.EmitAggregateCopy( | |||
3119 | CGF.MakeAddrLValue(DestElementAddr, Private->getType()), | |||
3120 | CGF.MakeAddrLValue(SrcElementAddr, Private->getType()), | |||
3121 | Private->getType(), AggValueSlot::DoesNotOverlap); | |||
3122 | break; | |||
3123 | } | |||
3124 | } | |||
3125 | ||||
3126 | // Step 3.1: Modify reference in dest Reduce list as needed. | |||
3127 | // Modifying the reference in Reduce list to point to the newly | |||
3128 | // created element. The element is live in the current function | |||
3129 | // scope and that of functions it invokes (i.e., reduce_function). | |||
3130 | // RemoteReduceData[i] = (void*)&RemoteElem | |||
3131 | if (UpdateDestListPtr) { | |||
3132 | CGF.EmitStoreOfScalar(Bld.CreatePointerBitCastOrAddrSpaceCast( | |||
3133 | DestElementAddr.getPointer(), CGF.VoidPtrTy), | |||
3134 | DestElementPtrAddr, /*Volatile=*/false, | |||
3135 | C.VoidPtrTy); | |||
3136 | } | |||
3137 | ||||
3138 | // Step 4.1: Increment SrcBase/DestBase so that it points to the starting | |||
3139 | // address of the next element in scratchpad memory, unless we're currently | |||
3140 | // processing the last one. Memory alignment is also taken care of here. | |||
3141 | if ((IncrScratchpadDest || IncrScratchpadSrc) && (Idx + 1 < Size)) { | |||
3142 | llvm::Value *ScratchpadBasePtr = | |||
3143 | IncrScratchpadDest ? DestBase.getPointer() : SrcBase.getPointer(); | |||
3144 | llvm::Value *ElementSizeInChars = CGF.getTypeSize(Private->getType()); | |||
3145 | ScratchpadBasePtr = Bld.CreateNUWAdd( | |||
3146 | ScratchpadBasePtr, | |||
3147 | Bld.CreateNUWMul(ScratchpadWidth, ElementSizeInChars)); | |||
3148 | ||||
3149 | // Take care of global memory alignment for performance | |||
3150 | ScratchpadBasePtr = Bld.CreateNUWSub( | |||
3151 | ScratchpadBasePtr, llvm::ConstantInt::get(CGM.SizeTy, 1)); | |||
3152 | ScratchpadBasePtr = Bld.CreateUDiv( | |||
3153 | ScratchpadBasePtr, | |||
3154 | llvm::ConstantInt::get(CGM.SizeTy, GlobalMemoryAlignment)); | |||
3155 | ScratchpadBasePtr = Bld.CreateNUWAdd( | |||
3156 | ScratchpadBasePtr, llvm::ConstantInt::get(CGM.SizeTy, 1)); | |||
3157 | ScratchpadBasePtr = Bld.CreateNUWMul( | |||
3158 | ScratchpadBasePtr, | |||
3159 | llvm::ConstantInt::get(CGM.SizeTy, GlobalMemoryAlignment)); | |||
3160 | ||||
3161 | if (IncrScratchpadDest) | |||
3162 | DestBase = Address(ScratchpadBasePtr, CGF.getPointerAlign()); | |||
3163 | else /* IncrScratchpadSrc = true */ | |||
3164 | SrcBase = Address(ScratchpadBasePtr, CGF.getPointerAlign()); | |||
3165 | } | |||
3166 | ||||
3167 | ++Idx; | |||
3168 | } | |||
3169 | } | |||
3170 | ||||
3171 | /// This function emits a helper that gathers Reduce lists from the first | |||
3172 | /// lane of every active warp to lanes in the first warp. | |||
3173 | /// | |||
3174 | /// void inter_warp_copy_func(void* reduce_data, num_warps) | |||
3175 | /// shared smem[warp_size]; | |||
3176 | /// For all data entries D in reduce_data: | |||
3177 | /// sync | |||
3178 | /// If (I am the first lane in each warp) | |||
3179 | /// Copy my local D to smem[warp_id] | |||
3180 | /// sync | |||
3181 | /// if (I am the first warp) | |||
3182 | /// Copy smem[thread_id] to my local D | |||
3183 | static llvm::Value *emitInterWarpCopyFunction(CodeGenModule &CGM, | |||
3184 | ArrayRef<const Expr *> Privates, | |||
3185 | QualType ReductionArrayTy, | |||
3186 | SourceLocation Loc) { | |||
3187 | ASTContext &C = CGM.getContext(); | |||
3188 | llvm::Module &M = CGM.getModule(); | |||
3189 | ||||
3190 | // ReduceList: thread local Reduce list. | |||
3191 | // At the stage of the computation when this function is called, partially | |||
3192 | // aggregated values reside in the first lane of every active warp. | |||
3193 | ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
3194 | C.VoidPtrTy, ImplicitParamDecl::Other); | |||
3195 | // NumWarps: number of warps active in the parallel region. This could | |||
3196 | // be smaller than 32 (max warps in a CTA) for partial block reduction. | |||
3197 | ImplicitParamDecl NumWarpsArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
3198 | C.getIntTypeForBitwidth(32, /* Signed */ true), | |||
3199 | ImplicitParamDecl::Other); | |||
3200 | FunctionArgList Args; | |||
3201 | Args.push_back(&ReduceListArg); | |||
3202 | Args.push_back(&NumWarpsArg); | |||
3203 | ||||
3204 | const CGFunctionInfo &CGFI = | |||
3205 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | |||
3206 | auto *Fn = llvm::Function::Create(CGM.getTypes().GetFunctionType(CGFI), | |||
3207 | llvm::GlobalValue::InternalLinkage, | |||
3208 | "_omp_reduction_inter_warp_copy_func", &M); | |||
3209 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); | |||
3210 | Fn->setDoesNotRecurse(); | |||
3211 | CodeGenFunction CGF(CGM); | |||
3212 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); | |||
3213 | ||||
3214 | CGBuilderTy &Bld = CGF.Builder; | |||
3215 | ||||
3216 | // This array is used as a medium to transfer, one reduce element at a time, | |||
3217 | // the data from the first lane of every warp to lanes in the first warp | |||
3218 | // in order to perform the final step of a reduction in a parallel region | |||
3219 | // (reduction across warps). The array is placed in NVPTX __shared__ memory | |||
3220 | // for reduced latency, as well as to have a distinct copy for concurrently | |||
3221 | // executing target regions. The array is declared with common linkage so | |||
3222 | // as to be shared across compilation units. | |||
3223 | StringRef TransferMediumName = | |||
3224 | "__openmp_nvptx_data_transfer_temporary_storage"; | |||
3225 | llvm::GlobalVariable *TransferMedium = | |||
3226 | M.getGlobalVariable(TransferMediumName); | |||
3227 | if (!TransferMedium) { | |||
3228 | auto *Ty = llvm::ArrayType::get(CGM.Int32Ty, WarpSize); | |||
3229 | unsigned SharedAddressSpace = C.getTargetAddressSpace(LangAS::cuda_shared); | |||
3230 | TransferMedium = new llvm::GlobalVariable( | |||
3231 | M, Ty, /*isConstant=*/false, llvm::GlobalVariable::CommonLinkage, | |||
3232 | llvm::Constant::getNullValue(Ty), TransferMediumName, | |||
3233 | /*InsertBefore=*/nullptr, llvm::GlobalVariable::NotThreadLocal, | |||
3234 | SharedAddressSpace); | |||
3235 | CGM.addCompilerUsedGlobal(TransferMedium); | |||
3236 | } | |||
3237 | ||||
3238 | // Get the CUDA thread id of the current OpenMP thread on the GPU. | |||
3239 | llvm::Value *ThreadID = getNVPTXThreadID(CGF); | |||
3240 | // nvptx_lane_id = nvptx_id % warpsize | |||
3241 | llvm::Value *LaneID = getNVPTXLaneID(CGF); | |||
3242 | // nvptx_warp_id = nvptx_id / warpsize | |||
3243 | llvm::Value *WarpID = getNVPTXWarpID(CGF); | |||
3244 | ||||
3245 | Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); | |||
3246 | Address LocalReduceList( | |||
3247 | Bld.CreatePointerBitCastOrAddrSpaceCast( | |||
3248 | CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false, | |||
3249 | C.VoidPtrTy, Loc), | |||
3250 | CGF.ConvertTypeForMem(ReductionArrayTy)->getPointerTo()), | |||
3251 | CGF.getPointerAlign()); | |||
3252 | ||||
3253 | unsigned Idx = 0; | |||
3254 | for (const Expr *Private : Privates) { | |||
3255 | // | |||
3256 | // Warp master copies reduce element to transfer medium in __shared__ | |||
3257 | // memory. | |||
3258 | // | |||
3259 | unsigned RealTySize = | |||
3260 | C.getTypeSizeInChars(Private->getType()) | |||
3261 | .alignTo(C.getTypeAlignInChars(Private->getType())) | |||
3262 | .getQuantity(); | |||
3263 | for (unsigned TySize = 4; TySize > 0 && RealTySize > 0; TySize /=2) { | |||
3264 | unsigned NumIters = RealTySize / TySize; | |||
3265 | if (NumIters == 0) | |||
3266 | continue; | |||
3267 | QualType CType = C.getIntTypeForBitwidth( | |||
3268 | C.toBits(CharUnits::fromQuantity(TySize)), /*Signed=*/1); | |||
3269 | llvm::Type *CopyType = CGF.ConvertTypeForMem(CType); | |||
3270 | CharUnits Align = CharUnits::fromQuantity(TySize); | |||
3271 | llvm::Value *Cnt = nullptr; | |||
3272 | Address CntAddr = Address::invalid(); | |||
3273 | llvm::BasicBlock *PrecondBB = nullptr; | |||
3274 | llvm::BasicBlock *ExitBB = nullptr; | |||
3275 | if (NumIters > 1) { | |||
3276 | CntAddr = CGF.CreateMemTemp(C.IntTy, ".cnt.addr"); | |||
3277 | CGF.EmitStoreOfScalar(llvm::Constant::getNullValue(CGM.IntTy), CntAddr, | |||
3278 | /*Volatile=*/false, C.IntTy); | |||
3279 | PrecondBB = CGF.createBasicBlock("precond"); | |||
3280 | ExitBB = CGF.createBasicBlock("exit"); | |||
3281 | llvm::BasicBlock *BodyBB = CGF.createBasicBlock("body"); | |||
3282 | // There is no need to emit line number for unconditional branch. | |||
3283 | (void)ApplyDebugLocation::CreateEmpty(CGF); | |||
3284 | CGF.EmitBlock(PrecondBB); | |||
3285 | Cnt = CGF.EmitLoadOfScalar(CntAddr, /*Volatile=*/false, C.IntTy, Loc); | |||
3286 | llvm::Value *Cmp = | |||
3287 | Bld.CreateICmpULT(Cnt, llvm::ConstantInt::get(CGM.IntTy, NumIters)); | |||
3288 | Bld.CreateCondBr(Cmp, BodyBB, ExitBB); | |||
3289 | CGF.EmitBlock(BodyBB); | |||
3290 | } | |||
3291 | // kmpc_barrier. | |||
3292 | CGM.getOpenMPRuntime().emitBarrierCall(CGF, Loc, OMPD_unknown, | |||
3293 | /*EmitChecks=*/false, | |||
3294 | /*ForceSimpleCall=*/true); | |||
3295 | llvm::BasicBlock *ThenBB = CGF.createBasicBlock("then"); | |||
3296 | llvm::BasicBlock *ElseBB = CGF.createBasicBlock("else"); | |||
3297 | llvm::BasicBlock *MergeBB = CGF.createBasicBlock("ifcont"); | |||
3298 | ||||
3299 | // if (lane_id == 0) | |||
3300 | llvm::Value *IsWarpMaster = Bld.CreateIsNull(LaneID, "warp_master"); | |||
3301 | Bld.CreateCondBr(IsWarpMaster, ThenBB, ElseBB); | |||
3302 | CGF.EmitBlock(ThenBB); | |||
3303 | ||||
3304 | // Reduce element = LocalReduceList[i] | |||
3305 | Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx); | |||
3306 | llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar( | |||
3307 | ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation()); | |||
3308 | // elemptr = ((CopyType*)(elemptrptr)) + I | |||
3309 | Address ElemPtr = Address(ElemPtrPtr, Align); | |||
3310 | ElemPtr = Bld.CreateElementBitCast(ElemPtr, CopyType); | |||
3311 | if (NumIters > 1) { | |||
3312 | ElemPtr = Address(Bld.CreateGEP(ElemPtr.getPointer(), Cnt), | |||
3313 | ElemPtr.getAlignment()); | |||
3314 | } | |||
3315 | ||||
3316 | // Get pointer to location in transfer medium. | |||
3317 | // MediumPtr = &medium[warp_id] | |||
3318 | llvm::Value *MediumPtrVal = Bld.CreateInBoundsGEP( | |||
3319 | TransferMedium, {llvm::Constant::getNullValue(CGM.Int64Ty), WarpID}); | |||
3320 | Address MediumPtr(MediumPtrVal, Align); | |||
3321 | // Casting to actual data type. | |||
3322 | // MediumPtr = (CopyType*)MediumPtrAddr; | |||
3323 | MediumPtr = Bld.CreateElementBitCast(MediumPtr, CopyType); | |||
3324 | ||||
3325 | // elem = *elemptr | |||
3326 | //*MediumPtr = elem | |||
3327 | llvm::Value *Elem = | |||
3328 | CGF.EmitLoadOfScalar(ElemPtr, /*Volatile=*/false, CType, Loc); | |||
3329 | // Store the source element value to the dest element address. | |||
3330 | CGF.EmitStoreOfScalar(Elem, MediumPtr, /*Volatile=*/true, CType); | |||
3331 | ||||
3332 | Bld.CreateBr(MergeBB); | |||
3333 | ||||
3334 | CGF.EmitBlock(ElseBB); | |||
3335 | Bld.CreateBr(MergeBB); | |||
3336 | ||||
3337 | CGF.EmitBlock(MergeBB); | |||
3338 | ||||
3339 | // kmpc_barrier. | |||
3340 | CGM.getOpenMPRuntime().emitBarrierCall(CGF, Loc, OMPD_unknown, | |||
3341 | /*EmitChecks=*/false, | |||
3342 | /*ForceSimpleCall=*/true); | |||
3343 | ||||
3344 | // | |||
3345 | // Warp 0 copies reduce element from transfer medium. | |||
3346 | // | |||
3347 | llvm::BasicBlock *W0ThenBB = CGF.createBasicBlock("then"); | |||
3348 | llvm::BasicBlock *W0ElseBB = CGF.createBasicBlock("else"); | |||
3349 | llvm::BasicBlock *W0MergeBB = CGF.createBasicBlock("ifcont"); | |||
3350 | ||||
3351 | Address AddrNumWarpsArg = CGF.GetAddrOfLocalVar(&NumWarpsArg); | |||
3352 | llvm::Value *NumWarpsVal = CGF.EmitLoadOfScalar( | |||
3353 | AddrNumWarpsArg, /*Volatile=*/false, C.IntTy, Loc); | |||
3354 | ||||
3355 | // Up to 32 threads in warp 0 are active. | |||
3356 | llvm::Value *IsActiveThread = | |||
3357 | Bld.CreateICmpULT(ThreadID, NumWarpsVal, "is_active_thread"); | |||
3358 | Bld.CreateCondBr(IsActiveThread, W0ThenBB, W0ElseBB); | |||
3359 | ||||
3360 | CGF.EmitBlock(W0ThenBB); | |||
3361 | ||||
3362 | // SrcMediumPtr = &medium[tid] | |||
3363 | llvm::Value *SrcMediumPtrVal = Bld.CreateInBoundsGEP( | |||
3364 | TransferMedium, | |||
3365 | {llvm::Constant::getNullValue(CGM.Int64Ty), ThreadID}); | |||
3366 | Address SrcMediumPtr(SrcMediumPtrVal, Align); | |||
3367 | // SrcMediumVal = *SrcMediumPtr; | |||
3368 | SrcMediumPtr = Bld.CreateElementBitCast(SrcMediumPtr, CopyType); | |||
3369 | ||||
3370 | // TargetElemPtr = (CopyType*)(SrcDataAddr[i]) + I | |||
3371 | Address TargetElemPtrPtr = Bld.CreateConstArrayGEP(LocalReduceList, Idx); | |||
3372 | llvm::Value *TargetElemPtrVal = CGF.EmitLoadOfScalar( | |||
3373 | TargetElemPtrPtr, /*Volatile=*/false, C.VoidPtrTy, Loc); | |||
3374 | Address TargetElemPtr = Address(TargetElemPtrVal, Align); | |||
3375 | TargetElemPtr = Bld.CreateElementBitCast(TargetElemPtr, CopyType); | |||
3376 | if (NumIters > 1) { | |||
3377 | TargetElemPtr = Address(Bld.CreateGEP(TargetElemPtr.getPointer(), Cnt), | |||
3378 | TargetElemPtr.getAlignment()); | |||
3379 | } | |||
3380 | ||||
3381 | // *TargetElemPtr = SrcMediumVal; | |||
3382 | llvm::Value *SrcMediumValue = | |||
3383 | CGF.EmitLoadOfScalar(SrcMediumPtr, /*Volatile=*/true, CType, Loc); | |||
3384 | CGF.EmitStoreOfScalar(SrcMediumValue, TargetElemPtr, /*Volatile=*/false, | |||
3385 | CType); | |||
3386 | Bld.CreateBr(W0MergeBB); | |||
3387 | ||||
3388 | CGF.EmitBlock(W0ElseBB); | |||
3389 | Bld.CreateBr(W0MergeBB); | |||
3390 | ||||
3391 | CGF.EmitBlock(W0MergeBB); | |||
3392 | ||||
3393 | if (NumIters > 1) { | |||
3394 | Cnt = Bld.CreateNSWAdd(Cnt, llvm::ConstantInt::get(CGM.IntTy, /*V=*/1)); | |||
3395 | CGF.EmitStoreOfScalar(Cnt, CntAddr, /*Volatile=*/false, C.IntTy); | |||
3396 | CGF.EmitBranch(PrecondBB); | |||
3397 | (void)ApplyDebugLocation::CreateEmpty(CGF); | |||
3398 | CGF.EmitBlock(ExitBB); | |||
3399 | } | |||
3400 | RealTySize %= TySize; | |||
3401 | } | |||
3402 | ++Idx; | |||
3403 | } | |||
3404 | ||||
3405 | CGF.FinishFunction(); | |||
3406 | return Fn; | |||
3407 | } | |||
3408 | ||||
3409 | /// Emit a helper that reduces data across two OpenMP threads (lanes) | |||
3410 | /// in the same warp. It uses shuffle instructions to copy over data from | |||
3411 | /// a remote lane's stack. The reduction algorithm performed is specified | |||
3412 | /// by the fourth parameter. | |||
3413 | /// | |||
3414 | /// Algorithm Versions. | |||
3415 | /// Full Warp Reduce (argument value 0): | |||
3416 | /// This algorithm assumes that all 32 lanes are active and gathers | |||
3417 | /// data from these 32 lanes, producing a single resultant value. | |||
3418 | /// Contiguous Partial Warp Reduce (argument value 1): | |||
3419 | /// This algorithm assumes that only a *contiguous* subset of lanes | |||
3420 | /// are active. This happens for the last warp in a parallel region | |||
3421 | /// when the user specified num_threads is not an integer multiple of | |||
3422 | /// 32. This contiguous subset always starts with the zeroth lane. | |||
3423 | /// Partial Warp Reduce (argument value 2): | |||
3424 | /// This algorithm gathers data from any number of lanes at any position. | |||
3425 | /// All reduced values are stored in the lowest possible lane. The set | |||
3426 | /// of problems every algorithm addresses is a super set of those | |||
3427 | /// addressable by algorithms with a lower version number. Overhead | |||
3428 | /// increases as algorithm version increases. | |||
3429 | /// | |||
3430 | /// Terminology | |||
3431 | /// Reduce element: | |||
3432 | /// Reduce element refers to the individual data field with primitive | |||
3433 | /// data types to be combined and reduced across threads. | |||
3434 | /// Reduce list: | |||
3435 | /// Reduce list refers to a collection of local, thread-private | |||
3436 | /// reduce elements. | |||
3437 | /// Remote Reduce list: | |||
3438 | /// Remote Reduce list refers to a collection of remote (relative to | |||
3439 | /// the current thread) reduce elements. | |||
3440 | /// | |||
3441 | /// We distinguish between three states of threads that are important to | |||
3442 | /// the implementation of this function. | |||
3443 | /// Alive threads: | |||
3444 | /// Threads in a warp executing the SIMT instruction, as distinguished from | |||
3445 | /// threads that are inactive due to divergent control flow. | |||
3446 | /// Active threads: | |||
3447 | /// The minimal set of threads that has to be alive upon entry to this | |||
3448 | /// function. The computation is correct iff active threads are alive. | |||
3449 | /// Some threads are alive but they are not active because they do not | |||
3450 | /// contribute to the computation in any useful manner. Turning them off | |||
3451 | /// may introduce control flow overheads without any tangible benefits. | |||
3452 | /// Effective threads: | |||
3453 | /// In order to comply with the argument requirements of the shuffle | |||
3454 | /// function, we must keep all lanes holding data alive. But at most | |||
3455 | /// half of them perform value aggregation; we refer to this half of | |||
3456 | /// threads as effective. The other half is simply handing off their | |||
3457 | /// data. | |||
3458 | /// | |||
3459 | /// Procedure | |||
3460 | /// Value shuffle: | |||
3461 | /// In this step active threads transfer data from higher lane positions | |||
3462 | /// in the warp to lower lane positions, creating Remote Reduce list. | |||
3463 | /// Value aggregation: | |||
3464 | /// In this step, effective threads combine their thread local Reduce list | |||
3465 | /// with Remote Reduce list and store the result in the thread local | |||
3466 | /// Reduce list. | |||
3467 | /// Value copy: | |||
3468 | /// In this step, we deal with the assumption made by algorithm 2 | |||
3469 | /// (i.e. contiguity assumption). When we have an odd number of lanes | |||
3470 | /// active, say 2k+1, only k threads will be effective and therefore k | |||
3471 | /// new values will be produced. However, the Reduce list owned by the | |||
3472 | /// (2k+1)th thread is ignored in the value aggregation. Therefore | |||
3473 | /// we copy the Reduce list from the (2k+1)th lane to (k+1)th lane so | |||
3474 | /// that the contiguity assumption still holds. | |||
3475 | static llvm::Function *emitShuffleAndReduceFunction( | |||
3476 | CodeGenModule &CGM, ArrayRef<const Expr *> Privates, | |||
3477 | QualType ReductionArrayTy, llvm::Function *ReduceFn, SourceLocation Loc) { | |||
3478 | ASTContext &C = CGM.getContext(); | |||
3479 | ||||
3480 | // Thread local Reduce list used to host the values of data to be reduced. | |||
3481 | ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
3482 | C.VoidPtrTy, ImplicitParamDecl::Other); | |||
3483 | // Current lane id; could be logical. | |||
3484 | ImplicitParamDecl LaneIDArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.ShortTy, | |||
3485 | ImplicitParamDecl::Other); | |||
3486 | // Offset of the remote source lane relative to the current lane. | |||
3487 | ImplicitParamDecl RemoteLaneOffsetArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
3488 | C.ShortTy, ImplicitParamDecl::Other); | |||
3489 | // Algorithm version. This is expected to be known at compile time. | |||
3490 | ImplicitParamDecl AlgoVerArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
3491 | C.ShortTy, ImplicitParamDecl::Other); | |||
3492 | FunctionArgList Args; | |||
3493 | Args.push_back(&ReduceListArg); | |||
3494 | Args.push_back(&LaneIDArg); | |||
3495 | Args.push_back(&RemoteLaneOffsetArg); | |||
3496 | Args.push_back(&AlgoVerArg); | |||
3497 | ||||
3498 | const CGFunctionInfo &CGFI = | |||
3499 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | |||
3500 | auto *Fn = llvm::Function::Create( | |||
3501 | CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, | |||
3502 | "_omp_reduction_shuffle_and_reduce_func", &CGM.getModule()); | |||
3503 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); | |||
3504 | Fn->setDoesNotRecurse(); | |||
3505 | if (CGM.getLangOpts().Optimize) { | |||
3506 | Fn->removeFnAttr(llvm::Attribute::NoInline); | |||
3507 | Fn->removeFnAttr(llvm::Attribute::OptimizeNone); | |||
3508 | Fn->addFnAttr(llvm::Attribute::AlwaysInline); | |||
3509 | } | |||
3510 | ||||
3511 | CodeGenFunction CGF(CGM); | |||
3512 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); | |||
3513 | ||||
3514 | CGBuilderTy &Bld = CGF.Builder; | |||
3515 | ||||
3516 | Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); | |||
3517 | Address LocalReduceList( | |||
3518 | Bld.CreatePointerBitCastOrAddrSpaceCast( | |||
3519 | CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false, | |||
3520 | C.VoidPtrTy, SourceLocation()), | |||
3521 | CGF.ConvertTypeForMem(ReductionArrayTy)->getPointerTo()), | |||
3522 | CGF.getPointerAlign()); | |||
3523 | ||||
3524 | Address AddrLaneIDArg = CGF.GetAddrOfLocalVar(&LaneIDArg); | |||
3525 | llvm::Value *LaneIDArgVal = CGF.EmitLoadOfScalar( | |||
3526 | AddrLaneIDArg, /*Volatile=*/false, C.ShortTy, SourceLocation()); | |||
3527 | ||||
3528 | Address AddrRemoteLaneOffsetArg = CGF.GetAddrOfLocalVar(&RemoteLaneOffsetArg); | |||
3529 | llvm::Value *RemoteLaneOffsetArgVal = CGF.EmitLoadOfScalar( | |||
3530 | AddrRemoteLaneOffsetArg, /*Volatile=*/false, C.ShortTy, SourceLocation()); | |||
3531 | ||||
3532 | Address AddrAlgoVerArg = CGF.GetAddrOfLocalVar(&AlgoVerArg); | |||
3533 | llvm::Value *AlgoVerArgVal = CGF.EmitLoadOfScalar( | |||
3534 | AddrAlgoVerArg, /*Volatile=*/false, C.ShortTy, SourceLocation()); | |||
3535 | ||||
3536 | // Create a local thread-private variable to host the Reduce list | |||
3537 | // from a remote lane. | |||
3538 | Address RemoteReduceList = | |||
3539 | CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.remote_reduce_list"); | |||
3540 | ||||
3541 | // This loop iterates through the list of reduce elements and copies, | |||
3542 | // element by element, from a remote lane in the warp to RemoteReduceList, | |||
3543 | // hosted on the thread's stack. | |||
3544 | emitReductionListCopy(RemoteLaneToThread, CGF, ReductionArrayTy, Privates, | |||
3545 | LocalReduceList, RemoteReduceList, | |||
3546 | {/*RemoteLaneOffset=*/RemoteLaneOffsetArgVal, | |||
3547 | /*ScratchpadIndex=*/nullptr, | |||
3548 | /*ScratchpadWidth=*/nullptr}); | |||
3549 | ||||
3550 | // The actions to be performed on the Remote Reduce list is dependent | |||
3551 | // on the algorithm version. | |||
3552 | // | |||
3553 | // if (AlgoVer==0) || (AlgoVer==1 && (LaneId < Offset)) || (AlgoVer==2 && | |||
3554 | // LaneId % 2 == 0 && Offset > 0): | |||
3555 | // do the reduction value aggregation | |||
3556 | // | |||
3557 | // The thread local variable Reduce list is mutated in place to host the | |||
3558 | // reduced data, which is the aggregated value produced from local and | |||
3559 | // remote lanes. | |||
3560 | // | |||
3561 | // Note that AlgoVer is expected to be a constant integer known at compile | |||
3562 | // time. | |||
3563 | // When AlgoVer==0, the first conjunction evaluates to true, making | |||
3564 | // the entire predicate true during compile time. | |||
3565 | // When AlgoVer==1, the second conjunction has only the second part to be | |||
3566 | // evaluated during runtime. Other conjunctions evaluates to false | |||
3567 | // during compile time. | |||
3568 | // When AlgoVer==2, the third conjunction has only the second part to be | |||
3569 | // evaluated during runtime. Other conjunctions evaluates to false | |||
3570 | // during compile time. | |||
3571 | llvm::Value *CondAlgo0 = Bld.CreateIsNull(AlgoVerArgVal); | |||
3572 | ||||
3573 | llvm::Value *Algo1 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(1)); | |||
3574 | llvm::Value *CondAlgo1 = Bld.CreateAnd( | |||
3575 | Algo1, Bld.CreateICmpULT(LaneIDArgVal, RemoteLaneOffsetArgVal)); | |||
3576 | ||||
3577 | llvm::Value *Algo2 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(2)); | |||
3578 | llvm::Value *CondAlgo2 = Bld.CreateAnd( | |||
3579 | Algo2, Bld.CreateIsNull(Bld.CreateAnd(LaneIDArgVal, Bld.getInt16(1)))); | |||
3580 | CondAlgo2 = Bld.CreateAnd( | |||
3581 | CondAlgo2, Bld.CreateICmpSGT(RemoteLaneOffsetArgVal, Bld.getInt16(0))); | |||
3582 | ||||
3583 | llvm::Value *CondReduce = Bld.CreateOr(CondAlgo0, CondAlgo1); | |||
3584 | CondReduce = Bld.CreateOr(CondReduce, CondAlgo2); | |||
3585 | ||||
3586 | llvm::BasicBlock *ThenBB = CGF.createBasicBlock("then"); | |||
3587 | llvm::BasicBlock *ElseBB = CGF.createBasicBlock("else"); | |||
3588 | llvm::BasicBlock *MergeBB = CGF.createBasicBlock("ifcont"); | |||
3589 | Bld.CreateCondBr(CondReduce, ThenBB, ElseBB); | |||
3590 | ||||
3591 | CGF.EmitBlock(ThenBB); | |||
3592 | // reduce_function(LocalReduceList, RemoteReduceList) | |||
3593 | llvm::Value *LocalReduceListPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( | |||
3594 | LocalReduceList.getPointer(), CGF.VoidPtrTy); | |||
3595 | llvm::Value *RemoteReduceListPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( | |||
3596 | RemoteReduceList.getPointer(), CGF.VoidPtrTy); | |||
3597 | CGM.getOpenMPRuntime().emitOutlinedFunctionCall( | |||
3598 | CGF, Loc, ReduceFn, {LocalReduceListPtr, RemoteReduceListPtr}); | |||
3599 | Bld.CreateBr(MergeBB); | |||
3600 | ||||
3601 | CGF.EmitBlock(ElseBB); | |||
3602 | Bld.CreateBr(MergeBB); | |||
3603 | ||||
3604 | CGF.EmitBlock(MergeBB); | |||
3605 | ||||
3606 | // if (AlgoVer==1 && (LaneId >= Offset)) copy Remote Reduce list to local | |||
3607 | // Reduce list. | |||
3608 | Algo1 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(1)); | |||
3609 | llvm::Value *CondCopy = Bld.CreateAnd( | |||
3610 | Algo1, Bld.CreateICmpUGE(LaneIDArgVal, RemoteLaneOffsetArgVal)); | |||
3611 | ||||
3612 | llvm::BasicBlock *CpyThenBB = CGF.createBasicBlock("then"); | |||
3613 | llvm::BasicBlock *CpyElseBB = CGF.createBasicBlock("else"); | |||
3614 | llvm::BasicBlock *CpyMergeBB = CGF.createBasicBlock("ifcont"); | |||
3615 | Bld.CreateCondBr(CondCopy, CpyThenBB, CpyElseBB); | |||
3616 | ||||
3617 | CGF.EmitBlock(CpyThenBB); | |||
3618 | emitReductionListCopy(ThreadCopy, CGF, ReductionArrayTy, Privates, | |||
3619 | RemoteReduceList, LocalReduceList); | |||
3620 | Bld.CreateBr(CpyMergeBB); | |||
3621 | ||||
3622 | CGF.EmitBlock(CpyElseBB); | |||
3623 | Bld.CreateBr(CpyMergeBB); | |||
3624 | ||||
3625 | CGF.EmitBlock(CpyMergeBB); | |||
3626 | ||||
3627 | CGF.FinishFunction(); | |||
3628 | return Fn; | |||
3629 | } | |||
3630 | ||||
3631 | /// This function emits a helper that copies all the reduction variables from | |||
3632 | /// the team into the provided global buffer for the reduction variables. | |||
3633 | /// | |||
3634 | /// void list_to_global_copy_func(void *buffer, int Idx, void *reduce_data) | |||
3635 | /// For all data entries D in reduce_data: | |||
3636 | /// Copy local D to buffer.D[Idx] | |||
3637 | static llvm::Value *emitListToGlobalCopyFunction( | |||
3638 | CodeGenModule &CGM, ArrayRef<const Expr *> Privates, | |||
3639 | QualType ReductionArrayTy, SourceLocation Loc, | |||
3640 | const RecordDecl *TeamReductionRec, | |||
3641 | const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> | |||
3642 | &VarFieldMap) { | |||
3643 | ASTContext &C = CGM.getContext(); | |||
3644 | ||||
3645 | // Buffer: global reduction buffer. | |||
3646 | ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
3647 | C.VoidPtrTy, ImplicitParamDecl::Other); | |||
3648 | // Idx: index of the buffer. | |||
3649 | ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy, | |||
3650 | ImplicitParamDecl::Other); | |||
3651 | // ReduceList: thread local Reduce list. | |||
3652 | ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
3653 | C.VoidPtrTy, ImplicitParamDecl::Other); | |||
3654 | FunctionArgList Args; | |||
3655 | Args.push_back(&BufferArg); | |||
3656 | Args.push_back(&IdxArg); | |||
3657 | Args.push_back(&ReduceListArg); | |||
3658 | ||||
3659 | const CGFunctionInfo &CGFI = | |||
3660 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | |||
3661 | auto *Fn = llvm::Function::Create( | |||
3662 | CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, | |||
3663 | "_omp_reduction_list_to_global_copy_func", &CGM.getModule()); | |||
3664 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); | |||
3665 | Fn->setDoesNotRecurse(); | |||
3666 | CodeGenFunction CGF(CGM); | |||
3667 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); | |||
3668 | ||||
3669 | CGBuilderTy &Bld = CGF.Builder; | |||
3670 | ||||
3671 | Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); | |||
3672 | Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg); | |||
3673 | Address LocalReduceList( | |||
3674 | Bld.CreatePointerBitCastOrAddrSpaceCast( | |||
3675 | CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false, | |||
3676 | C.VoidPtrTy, Loc), | |||
3677 | CGF.ConvertTypeForMem(ReductionArrayTy)->getPointerTo()), | |||
3678 | CGF.getPointerAlign()); | |||
3679 | QualType StaticTy = C.getRecordType(TeamReductionRec); | |||
3680 | llvm::Type *LLVMReductionsBufferTy = | |||
3681 | CGM.getTypes().ConvertTypeForMem(StaticTy); | |||
3682 | llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( | |||
3683 | CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc), | |||
3684 | LLVMReductionsBufferTy->getPointerTo()); | |||
3685 | llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty), | |||
3686 | CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg), | |||
3687 | /*Volatile=*/false, C.IntTy, | |||
3688 | Loc)}; | |||
3689 | unsigned Idx = 0; | |||
3690 | for (const Expr *Private : Privates) { | |||
3691 | // Reduce element = LocalReduceList[i] | |||
3692 | Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx); | |||
3693 | llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar( | |||
3694 | ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation()); | |||
3695 | // elemptr = ((CopyType*)(elemptrptr)) + I | |||
3696 | ElemPtrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( | |||
3697 | ElemPtrPtr, CGF.ConvertTypeForMem(Private->getType())->getPointerTo()); | |||
3698 | Address ElemPtr = | |||
3699 | Address(ElemPtrPtr, C.getTypeAlignInChars(Private->getType())); | |||
3700 | const ValueDecl *VD = cast<DeclRefExpr>(Private)->getDecl(); | |||
3701 | // Global = Buffer.VD[Idx]; | |||
3702 | const FieldDecl *FD = VarFieldMap.lookup(VD); | |||
3703 | LValue GlobLVal = CGF.EmitLValueForField( | |||
3704 | CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD); | |||
3705 | llvm::Value *BufferPtr = | |||
3706 | Bld.CreateInBoundsGEP(GlobLVal.getPointer(CGF), Idxs); | |||
3707 | GlobLVal.setAddress(Address(BufferPtr, GlobLVal.getAlignment())); | |||
3708 | switch (CGF.getEvaluationKind(Private->getType())) { | |||
3709 | case TEK_Scalar: { | |||
3710 | llvm::Value *V = CGF.EmitLoadOfScalar(ElemPtr, /*Volatile=*/false, | |||
3711 | Private->getType(), Loc); | |||
3712 | CGF.EmitStoreOfScalar(V, GlobLVal); | |||
3713 | break; | |||
3714 | } | |||
3715 | case TEK_Complex: { | |||
3716 | CodeGenFunction::ComplexPairTy V = CGF.EmitLoadOfComplex( | |||
3717 | CGF.MakeAddrLValue(ElemPtr, Private->getType()), Loc); | |||
3718 | CGF.EmitStoreOfComplex(V, GlobLVal, /*isInit=*/false); | |||
3719 | break; | |||
3720 | } | |||
3721 | case TEK_Aggregate: | |||
3722 | CGF.EmitAggregateCopy(GlobLVal, | |||
3723 | CGF.MakeAddrLValue(ElemPtr, Private->getType()), | |||
3724 | Private->getType(), AggValueSlot::DoesNotOverlap); | |||
3725 | break; | |||
3726 | } | |||
3727 | ++Idx; | |||
3728 | } | |||
3729 | ||||
3730 | CGF.FinishFunction(); | |||
3731 | return Fn; | |||
3732 | } | |||
3733 | ||||
3734 | /// This function emits a helper that reduces all the reduction variables from | |||
3735 | /// the team into the provided global buffer for the reduction variables. | |||
3736 | /// | |||
3737 | /// void list_to_global_reduce_func(void *buffer, int Idx, void *reduce_data) | |||
3738 | /// void *GlobPtrs[]; | |||
3739 | /// GlobPtrs[0] = (void*)&buffer.D0[Idx]; | |||
3740 | /// ... | |||
3741 | /// GlobPtrs[N] = (void*)&buffer.DN[Idx]; | |||
3742 | /// reduce_function(GlobPtrs, reduce_data); | |||
3743 | static llvm::Value *emitListToGlobalReduceFunction( | |||
3744 | CodeGenModule &CGM, ArrayRef<const Expr *> Privates, | |||
3745 | QualType ReductionArrayTy, SourceLocation Loc, | |||
3746 | const RecordDecl *TeamReductionRec, | |||
3747 | const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> | |||
3748 | &VarFieldMap, | |||
3749 | llvm::Function *ReduceFn) { | |||
3750 | ASTContext &C = CGM.getContext(); | |||
3751 | ||||
3752 | // Buffer: global reduction buffer. | |||
3753 | ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
3754 | C.VoidPtrTy, ImplicitParamDecl::Other); | |||
3755 | // Idx: index of the buffer. | |||
3756 | ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy, | |||
3757 | ImplicitParamDecl::Other); | |||
3758 | // ReduceList: thread local Reduce list. | |||
3759 | ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
3760 | C.VoidPtrTy, ImplicitParamDecl::Other); | |||
3761 | FunctionArgList Args; | |||
3762 | Args.push_back(&BufferArg); | |||
3763 | Args.push_back(&IdxArg); | |||
3764 | Args.push_back(&ReduceListArg); | |||
3765 | ||||
3766 | const CGFunctionInfo &CGFI = | |||
3767 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | |||
3768 | auto *Fn = llvm::Function::Create( | |||
3769 | CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, | |||
3770 | "_omp_reduction_list_to_global_reduce_func", &CGM.getModule()); | |||
3771 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); | |||
3772 | Fn->setDoesNotRecurse(); | |||
3773 | CodeGenFunction CGF(CGM); | |||
3774 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); | |||
3775 | ||||
3776 | CGBuilderTy &Bld = CGF.Builder; | |||
3777 | ||||
3778 | Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg); | |||
3779 | QualType StaticTy = C.getRecordType(TeamReductionRec); | |||
3780 | llvm::Type *LLVMReductionsBufferTy = | |||
3781 | CGM.getTypes().ConvertTypeForMem(StaticTy); | |||
3782 | llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( | |||
3783 | CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc), | |||
3784 | LLVMReductionsBufferTy->getPointerTo()); | |||
3785 | ||||
3786 | // 1. Build a list of reduction variables. | |||
3787 | // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]}; | |||
3788 | Address ReductionList = | |||
3789 | CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list"); | |||
3790 | auto IPriv = Privates.begin(); | |||
3791 | llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty), | |||
3792 | CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg), | |||
3793 | /*Volatile=*/false, C.IntTy, | |||
3794 | Loc)}; | |||
3795 | unsigned Idx = 0; | |||
3796 | for (unsigned I = 0, E = Privates.size(); I < E; ++I, ++IPriv, ++Idx) { | |||
3797 | Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); | |||
3798 | // Global = Buffer.VD[Idx]; | |||
3799 | const ValueDecl *VD = cast<DeclRefExpr>(*IPriv)->getDecl(); | |||
3800 | const FieldDecl *FD = VarFieldMap.lookup(VD); | |||
3801 | LValue GlobLVal = CGF.EmitLValueForField( | |||
3802 | CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD); | |||
3803 | llvm::Value *BufferPtr = | |||
3804 | Bld.CreateInBoundsGEP(GlobLVal.getPointer(CGF), Idxs); | |||
3805 | llvm::Value *Ptr = CGF.EmitCastToVoidPtr(BufferPtr); | |||
3806 | CGF.EmitStoreOfScalar(Ptr, Elem, /*Volatile=*/false, C.VoidPtrTy); | |||
3807 | if ((*IPriv)->getType()->isVariablyModifiedType()) { | |||
3808 | // Store array size. | |||
3809 | ++Idx; | |||
3810 | Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); | |||
3811 | llvm::Value *Size = CGF.Builder.CreateIntCast( | |||
3812 | CGF.getVLASize( | |||
3813 | CGF.getContext().getAsVariableArrayType((*IPriv)->getType())) | |||
3814 | .NumElts, | |||
3815 | CGF.SizeTy, /*isSigned=*/false); | |||
3816 | CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy), | |||
3817 | Elem); | |||
3818 | } | |||
3819 | } | |||
3820 | ||||
3821 | // Call reduce_function(GlobalReduceList, ReduceList) | |||
3822 | llvm::Value *GlobalReduceList = | |||
3823 | CGF.EmitCastToVoidPtr(ReductionList.getPointer()); | |||
3824 | Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); | |||
3825 | llvm::Value *ReducedPtr = CGF.EmitLoadOfScalar( | |||
3826 | AddrReduceListArg, /*Volatile=*/false, C.VoidPtrTy, Loc); | |||
3827 | CGM.getOpenMPRuntime().emitOutlinedFunctionCall( | |||
3828 | CGF, Loc, ReduceFn, {GlobalReduceList, ReducedPtr}); | |||
3829 | CGF.FinishFunction(); | |||
3830 | return Fn; | |||
3831 | } | |||
3832 | ||||
3833 | /// This function emits a helper that copies all the reduction variables from | |||
3834 | /// the team into the provided global buffer for the reduction variables. | |||
3835 | /// | |||
3836 | /// void list_to_global_copy_func(void *buffer, int Idx, void *reduce_data) | |||
3837 | /// For all data entries D in reduce_data: | |||
3838 | /// Copy buffer.D[Idx] to local D; | |||
3839 | static llvm::Value *emitGlobalToListCopyFunction( | |||
3840 | CodeGenModule &CGM, ArrayRef<const Expr *> Privates, | |||
3841 | QualType ReductionArrayTy, SourceLocation Loc, | |||
3842 | const RecordDecl *TeamReductionRec, | |||
3843 | const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> | |||
3844 | &VarFieldMap) { | |||
3845 | ASTContext &C = CGM.getContext(); | |||
3846 | ||||
3847 | // Buffer: global reduction buffer. | |||
3848 | ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
3849 | C.VoidPtrTy, ImplicitParamDecl::Other); | |||
3850 | // Idx: index of the buffer. | |||
3851 | ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy, | |||
3852 | ImplicitParamDecl::Other); | |||
3853 | // ReduceList: thread local Reduce list. | |||
3854 | ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
3855 | C.VoidPtrTy, ImplicitParamDecl::Other); | |||
3856 | FunctionArgList Args; | |||
3857 | Args.push_back(&BufferArg); | |||
3858 | Args.push_back(&IdxArg); | |||
3859 | Args.push_back(&ReduceListArg); | |||
3860 | ||||
3861 | const CGFunctionInfo &CGFI = | |||
3862 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | |||
3863 | auto *Fn = llvm::Function::Create( | |||
3864 | CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, | |||
3865 | "_omp_reduction_global_to_list_copy_func", &CGM.getModule()); | |||
3866 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); | |||
3867 | Fn->setDoesNotRecurse(); | |||
3868 | CodeGenFunction CGF(CGM); | |||
3869 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); | |||
3870 | ||||
3871 | CGBuilderTy &Bld = CGF.Builder; | |||
3872 | ||||
3873 | Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); | |||
3874 | Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg); | |||
3875 | Address LocalReduceList( | |||
3876 | Bld.CreatePointerBitCastOrAddrSpaceCast( | |||
3877 | CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false, | |||
3878 | C.VoidPtrTy, Loc), | |||
3879 | CGF.ConvertTypeForMem(ReductionArrayTy)->getPointerTo()), | |||
3880 | CGF.getPointerAlign()); | |||
3881 | QualType StaticTy = C.getRecordType(TeamReductionRec); | |||
3882 | llvm::Type *LLVMReductionsBufferTy = | |||
3883 | CGM.getTypes().ConvertTypeForMem(StaticTy); | |||
3884 | llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( | |||
3885 | CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc), | |||
3886 | LLVMReductionsBufferTy->getPointerTo()); | |||
3887 | ||||
3888 | llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty), | |||
3889 | CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg), | |||
3890 | /*Volatile=*/false, C.IntTy, | |||
3891 | Loc)}; | |||
3892 | unsigned Idx = 0; | |||
3893 | for (const Expr *Private : Privates) { | |||
3894 | // Reduce element = LocalReduceList[i] | |||
3895 | Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx); | |||
3896 | llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar( | |||
3897 | ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation()); | |||
3898 | // elemptr = ((CopyType*)(elemptrptr)) + I | |||
3899 | ElemPtrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( | |||
3900 | ElemPtrPtr, CGF.ConvertTypeForMem(Private->getType())->getPointerTo()); | |||
3901 | Address ElemPtr = | |||
3902 | Address(ElemPtrPtr, C.getTypeAlignInChars(Private->getType())); | |||
3903 | const ValueDecl *VD = cast<DeclRefExpr>(Private)->getDecl(); | |||
3904 | // Global = Buffer.VD[Idx]; | |||
3905 | const FieldDecl *FD = VarFieldMap.lookup(VD); | |||
3906 | LValue GlobLVal = CGF.EmitLValueForField( | |||
3907 | CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD); | |||
3908 | llvm::Value *BufferPtr = | |||
3909 | Bld.CreateInBoundsGEP(GlobLVal.getPointer(CGF), Idxs); | |||
3910 | GlobLVal.setAddress(Address(BufferPtr, GlobLVal.getAlignment())); | |||
3911 | switch (CGF.getEvaluationKind(Private->getType())) { | |||
3912 | case TEK_Scalar: { | |||
3913 | llvm::Value *V = CGF.EmitLoadOfScalar(GlobLVal, Loc); | |||
3914 | CGF.EmitStoreOfScalar(V, ElemPtr, /*Volatile=*/false, Private->getType()); | |||
3915 | break; | |||
3916 | } | |||
3917 | case TEK_Complex: { | |||
3918 | CodeGenFunction::ComplexPairTy V = CGF.EmitLoadOfComplex(GlobLVal, Loc); | |||
3919 | CGF.EmitStoreOfComplex(V, CGF.MakeAddrLValue(ElemPtr, Private->getType()), | |||
3920 | /*isInit=*/false); | |||
3921 | break; | |||
3922 | } | |||
3923 | case TEK_Aggregate: | |||
3924 | CGF.EmitAggregateCopy(CGF.MakeAddrLValue(ElemPtr, Private->getType()), | |||
3925 | GlobLVal, Private->getType(), | |||
3926 | AggValueSlot::DoesNotOverlap); | |||
3927 | break; | |||
3928 | } | |||
3929 | ++Idx; | |||
3930 | } | |||
3931 | ||||
3932 | CGF.FinishFunction(); | |||
3933 | return Fn; | |||
3934 | } | |||
3935 | ||||
3936 | /// This function emits a helper that reduces all the reduction variables from | |||
3937 | /// the team into the provided global buffer for the reduction variables. | |||
3938 | /// | |||
3939 | /// void global_to_list_reduce_func(void *buffer, int Idx, void *reduce_data) | |||
3940 | /// void *GlobPtrs[]; | |||
3941 | /// GlobPtrs[0] = (void*)&buffer.D0[Idx]; | |||
3942 | /// ... | |||
3943 | /// GlobPtrs[N] = (void*)&buffer.DN[Idx]; | |||
3944 | /// reduce_function(reduce_data, GlobPtrs); | |||
3945 | static llvm::Value *emitGlobalToListReduceFunction( | |||
3946 | CodeGenModule &CGM, ArrayRef<const Expr *> Privates, | |||
3947 | QualType ReductionArrayTy, SourceLocation Loc, | |||
3948 | const RecordDecl *TeamReductionRec, | |||
3949 | const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> | |||
3950 | &VarFieldMap, | |||
3951 | llvm::Function *ReduceFn) { | |||
3952 | ASTContext &C = CGM.getContext(); | |||
3953 | ||||
3954 | // Buffer: global reduction buffer. | |||
3955 | ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
3956 | C.VoidPtrTy, ImplicitParamDecl::Other); | |||
3957 | // Idx: index of the buffer. | |||
3958 | ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy, | |||
3959 | ImplicitParamDecl::Other); | |||
3960 | // ReduceList: thread local Reduce list. | |||
3961 | ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
3962 | C.VoidPtrTy, ImplicitParamDecl::Other); | |||
3963 | FunctionArgList Args; | |||
3964 | Args.push_back(&BufferArg); | |||
3965 | Args.push_back(&IdxArg); | |||
3966 | Args.push_back(&ReduceListArg); | |||
3967 | ||||
3968 | const CGFunctionInfo &CGFI = | |||
3969 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | |||
3970 | auto *Fn = llvm::Function::Create( | |||
3971 | CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, | |||
3972 | "_omp_reduction_global_to_list_reduce_func", &CGM.getModule()); | |||
3973 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); | |||
3974 | Fn->setDoesNotRecurse(); | |||
3975 | CodeGenFunction CGF(CGM); | |||
3976 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); | |||
3977 | ||||
3978 | CGBuilderTy &Bld = CGF.Builder; | |||
3979 | ||||
3980 | Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg); | |||
3981 | QualType StaticTy = C.getRecordType(TeamReductionRec); | |||
3982 | llvm::Type *LLVMReductionsBufferTy = | |||
3983 | CGM.getTypes().ConvertTypeForMem(StaticTy); | |||
3984 | llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( | |||
3985 | CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc), | |||
3986 | LLVMReductionsBufferTy->getPointerTo()); | |||
3987 | ||||
3988 | // 1. Build a list of reduction variables. | |||
3989 | // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]}; | |||
3990 | Address ReductionList = | |||
3991 | CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list"); | |||
3992 | auto IPriv = Privates.begin(); | |||
3993 | llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty), | |||
3994 | CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg), | |||
3995 | /*Volatile=*/false, C.IntTy, | |||
3996 | Loc)}; | |||
3997 | unsigned Idx = 0; | |||
3998 | for (unsigned I = 0, E = Privates.size(); I < E; ++I, ++IPriv, ++Idx) { | |||
3999 | Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); | |||
4000 | // Global = Buffer.VD[Idx]; | |||
4001 | const ValueDecl *VD = cast<DeclRefExpr>(*IPriv)->getDecl(); | |||
4002 | const FieldDecl *FD = VarFieldMap.lookup(VD); | |||
4003 | LValue GlobLVal = CGF.EmitLValueForField( | |||
4004 | CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD); | |||
4005 | llvm::Value *BufferPtr = | |||
4006 | Bld.CreateInBoundsGEP(GlobLVal.getPointer(CGF), Idxs); | |||
4007 | llvm::Value *Ptr = CGF.EmitCastToVoidPtr(BufferPtr); | |||
4008 | CGF.EmitStoreOfScalar(Ptr, Elem, /*Volatile=*/false, C.VoidPtrTy); | |||
4009 | if ((*IPriv)->getType()->isVariablyModifiedType()) { | |||
4010 | // Store array size. | |||
4011 | ++Idx; | |||
4012 | Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); | |||
4013 | llvm::Value *Size = CGF.Builder.CreateIntCast( | |||
4014 | CGF.getVLASize( | |||
4015 | CGF.getContext().getAsVariableArrayType((*IPriv)->getType())) | |||
4016 | .NumElts, | |||
4017 | CGF.SizeTy, /*isSigned=*/false); | |||
4018 | CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy), | |||
4019 | Elem); | |||
4020 | } | |||
4021 | } | |||
4022 | ||||
4023 | // Call reduce_function(ReduceList, GlobalReduceList) | |||
4024 | llvm::Value *GlobalReduceList = | |||
4025 | CGF.EmitCastToVoidPtr(ReductionList.getPointer()); | |||
4026 | Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); | |||
4027 | llvm::Value *ReducedPtr = CGF.EmitLoadOfScalar( | |||
4028 | AddrReduceListArg, /*Volatile=*/false, C.VoidPtrTy, Loc); | |||
4029 | CGM.getOpenMPRuntime().emitOutlinedFunctionCall( | |||
4030 | CGF, Loc, ReduceFn, {ReducedPtr, GlobalReduceList}); | |||
4031 | CGF.FinishFunction(); | |||
4032 | return Fn; | |||
4033 | } | |||
4034 | ||||
4035 | /// | |||
4036 | /// Design of OpenMP reductions on the GPU | |||
4037 | /// | |||
4038 | /// Consider a typical OpenMP program with one or more reduction | |||
4039 | /// clauses: | |||
4040 | /// | |||
4041 | /// float foo; | |||
4042 | /// double bar; | |||
4043 | /// #pragma omp target teams distribute parallel for \ | |||
4044 | /// reduction(+:foo) reduction(*:bar) | |||
4045 | /// for (int i = 0; i < N; i++) { | |||
4046 | /// foo += A[i]; bar *= B[i]; | |||
4047 | /// } | |||
4048 | /// | |||
4049 | /// where 'foo' and 'bar' are reduced across all OpenMP threads in | |||
4050 | /// all teams. In our OpenMP implementation on the NVPTX device an | |||
4051 | /// OpenMP team is mapped to a CUDA threadblock and OpenMP threads | |||
4052 | /// within a team are mapped to CUDA threads within a threadblock. | |||
4053 | /// Our goal is to efficiently aggregate values across all OpenMP | |||
4054 | /// threads such that: | |||
4055 | /// | |||
4056 | /// - the compiler and runtime are logically concise, and | |||
4057 | /// - the reduction is performed efficiently in a hierarchical | |||
4058 | /// manner as follows: within OpenMP threads in the same warp, | |||
4059 | /// across warps in a threadblock, and finally across teams on | |||
4060 | /// the NVPTX device. | |||
4061 | /// | |||
4062 | /// Introduction to Decoupling | |||
4063 | /// | |||
4064 | /// We would like to decouple the compiler and the runtime so that the | |||
4065 | /// latter is ignorant of the reduction variables (number, data types) | |||
4066 | /// and the reduction operators. This allows a simpler interface | |||
4067 | /// and implementation while still attaining good performance. | |||
4068 | /// | |||
4069 | /// Pseudocode for the aforementioned OpenMP program generated by the | |||
4070 | /// compiler is as follows: | |||
4071 | /// | |||
4072 | /// 1. Create private copies of reduction variables on each OpenMP | |||
4073 | /// thread: 'foo_private', 'bar_private' | |||
4074 | /// 2. Each OpenMP thread reduces the chunk of 'A' and 'B' assigned | |||
4075 | /// to it and writes the result in 'foo_private' and 'bar_private' | |||
4076 | /// respectively. | |||
4077 | /// 3. Call the OpenMP runtime on the GPU to reduce within a team | |||
4078 | /// and store the result on the team master: | |||
4079 | /// | |||
4080 | /// __kmpc_nvptx_parallel_reduce_nowait_v2(..., | |||
4081 | /// reduceData, shuffleReduceFn, interWarpCpyFn) | |||
4082 | /// | |||
4083 | /// where: | |||
4084 | /// struct ReduceData { | |||
4085 | /// double *foo; | |||
4086 | /// double *bar; | |||
4087 | /// } reduceData | |||
4088 | /// reduceData.foo = &foo_private | |||
4089 | /// reduceData.bar = &bar_private | |||
4090 | /// | |||
4091 | /// 'shuffleReduceFn' and 'interWarpCpyFn' are pointers to two | |||
4092 | /// auxiliary functions generated by the compiler that operate on | |||
4093 | /// variables of type 'ReduceData'. They aid the runtime perform | |||
4094 | /// algorithmic steps in a data agnostic manner. | |||
4095 | /// | |||
4096 | /// 'shuffleReduceFn' is a pointer to a function that reduces data | |||
4097 | /// of type 'ReduceData' across two OpenMP threads (lanes) in the | |||
4098 | /// same warp. It takes the following arguments as input: | |||
4099 | /// | |||
4100 | /// a. variable of type 'ReduceData' on the calling lane, | |||
4101 | /// b. its lane_id, | |||
4102 | /// c. an offset relative to the current lane_id to generate a | |||
4103 | /// remote_lane_id. The remote lane contains the second | |||
4104 | /// variable of type 'ReduceData' that is to be reduced. | |||
4105 | /// d. an algorithm version parameter determining which reduction | |||
4106 | /// algorithm to use. | |||
4107 | /// | |||
4108 | /// 'shuffleReduceFn' retrieves data from the remote lane using | |||
4109 | /// efficient GPU shuffle intrinsics and reduces, using the | |||
4110 | /// algorithm specified by the 4th parameter, the two operands | |||
4111 | /// element-wise. The result is written to the first operand. | |||
4112 | /// | |||
4113 | /// Different reduction algorithms are implemented in different | |||
4114 | /// runtime functions, all calling 'shuffleReduceFn' to perform | |||
4115 | /// the essential reduction step. Therefore, based on the 4th | |||
4116 | /// parameter, this function behaves slightly differently to | |||
4117 | /// cooperate with the runtime to ensure correctness under | |||
4118 | /// different circumstances. | |||
4119 | /// | |||
4120 | /// 'InterWarpCpyFn' is a pointer to a function that transfers | |||
4121 | /// reduced variables across warps. It tunnels, through CUDA | |||
4122 | /// shared memory, the thread-private data of type 'ReduceData' | |||
4123 | /// from lane 0 of each warp to a lane in the first warp. | |||
4124 | /// 4. Call the OpenMP runtime on the GPU to reduce across teams. | |||
4125 | /// The last team writes the global reduced value to memory. | |||
4126 | /// | |||
4127 | /// ret = __kmpc_nvptx_teams_reduce_nowait(..., | |||
4128 | /// reduceData, shuffleReduceFn, interWarpCpyFn, | |||
4129 | /// scratchpadCopyFn, loadAndReduceFn) | |||
4130 | /// | |||
4131 | /// 'scratchpadCopyFn' is a helper that stores reduced | |||
4132 | /// data from the team master to a scratchpad array in | |||
4133 | /// global memory. | |||
4134 | /// | |||
4135 | /// 'loadAndReduceFn' is a helper that loads data from | |||
4136 | /// the scratchpad array and reduces it with the input | |||
4137 | /// operand. | |||
4138 | /// | |||
4139 | /// These compiler generated functions hide address | |||
4140 | /// calculation and alignment information from the runtime. | |||
4141 | /// 5. if ret == 1: | |||
4142 | /// The team master of the last team stores the reduced | |||
4143 | /// result to the globals in memory. | |||
4144 | /// foo += reduceData.foo; bar *= reduceData.bar | |||
4145 | /// | |||
4146 | /// | |||
4147 | /// Warp Reduction Algorithms | |||
4148 | /// | |||
4149 | /// On the warp level, we have three algorithms implemented in the | |||
4150 | /// OpenMP runtime depending on the number of active lanes: | |||
4151 | /// | |||
4152 | /// Full Warp Reduction | |||
4153 | /// | |||
4154 | /// The reduce algorithm within a warp where all lanes are active | |||
4155 | /// is implemented in the runtime as follows: | |||
4156 | /// | |||
4157 | /// full_warp_reduce(void *reduce_data, | |||
4158 | /// kmp_ShuffleReductFctPtr ShuffleReduceFn) { | |||
4159 | /// for (int offset = WARPSIZE/2; offset > 0; offset /= 2) | |||
4160 | /// ShuffleReduceFn(reduce_data, 0, offset, 0); | |||
4161 | /// } | |||
4162 | /// | |||
4163 | /// The algorithm completes in log(2, WARPSIZE) steps. | |||
4164 | /// | |||
4165 | /// 'ShuffleReduceFn' is used here with lane_id set to 0 because it is | |||
4166 | /// not used therefore we save instructions by not retrieving lane_id | |||
4167 | /// from the corresponding special registers. The 4th parameter, which | |||
4168 | /// represents the version of the algorithm being used, is set to 0 to | |||
4169 | /// signify full warp reduction. | |||
4170 | /// | |||
4171 | /// In this version, 'ShuffleReduceFn' behaves, per element, as follows: | |||
4172 | /// | |||
4173 | /// #reduce_elem refers to an element in the local lane's data structure | |||
4174 | /// #remote_elem is retrieved from a remote lane | |||
4175 | /// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE); | |||
4176 | /// reduce_elem = reduce_elem REDUCE_OP remote_elem; | |||
4177 | /// | |||
4178 | /// Contiguous Partial Warp Reduction | |||
4179 | /// | |||
4180 | /// This reduce algorithm is used within a warp where only the first | |||
4181 | /// 'n' (n <= WARPSIZE) lanes are active. It is typically used when the | |||
4182 | /// number of OpenMP threads in a parallel region is not a multiple of | |||
4183 | /// WARPSIZE. The algorithm is implemented in the runtime as follows: | |||
4184 | /// | |||
4185 | /// void | |||
4186 | /// contiguous_partial_reduce(void *reduce_data, | |||
4187 | /// kmp_ShuffleReductFctPtr ShuffleReduceFn, | |||
4188 | /// int size, int lane_id) { | |||
4189 | /// int curr_size; | |||
4190 | /// int offset; | |||
4191 | /// curr_size = size; | |||
4192 | /// mask = curr_size/2; | |||
4193 | /// while (offset>0) { | |||
4194 | /// ShuffleReduceFn(reduce_data, lane_id, offset, 1); | |||
4195 | /// curr_size = (curr_size+1)/2; | |||
4196 | /// offset = curr_size/2; | |||
4197 | /// } | |||
4198 | /// } | |||
4199 | /// | |||
4200 | /// In this version, 'ShuffleReduceFn' behaves, per element, as follows: | |||
4201 | /// | |||
4202 | /// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE); | |||
4203 | /// if (lane_id < offset) | |||
4204 | /// reduce_elem = reduce_elem REDUCE_OP remote_elem | |||
4205 | /// else | |||
4206 | /// reduce_elem = remote_elem | |||
4207 | /// | |||
4208 | /// This algorithm assumes that the data to be reduced are located in a | |||
4209 | /// contiguous subset of lanes starting from the first. When there is | |||
4210 | /// an odd number of active lanes, the data in the last lane is not | |||
4211 | /// aggregated with any other lane's dat but is instead copied over. | |||
4212 | /// | |||
4213 | /// Dispersed Partial Warp Reduction | |||
4214 | /// | |||
4215 | /// This algorithm is used within a warp when any discontiguous subset of | |||
4216 | /// lanes are active. It is used to implement the reduction operation | |||
4217 | /// across lanes in an OpenMP simd region or in a nested parallel region. | |||
4218 | /// | |||
4219 | /// void | |||
4220 | /// dispersed_partial_reduce(void *reduce_data, | |||
4221 | /// kmp_ShuffleReductFctPtr ShuffleReduceFn) { | |||
4222 | /// int size, remote_id; | |||
4223 | /// int logical_lane_id = number_of_active_lanes_before_me() * 2; | |||
4224 | /// do { | |||
4225 | /// remote_id = next_active_lane_id_right_after_me(); | |||
4226 | /// # the above function returns 0 of no active lane | |||
4227 | /// # is present right after the current lane. | |||
4228 | /// size = number_of_active_lanes_in_this_warp(); | |||
4229 | /// logical_lane_id /= 2; | |||
4230 | /// ShuffleReduceFn(reduce_data, logical_lane_id, | |||
4231 | /// remote_id-1-threadIdx.x, 2); | |||
4232 | /// } while (logical_lane_id % 2 == 0 && size > 1); | |||
4233 | /// } | |||
4234 | /// | |||
4235 | /// There is no assumption made about the initial state of the reduction. | |||
4236 | /// Any number of lanes (>=1) could be active at any position. The reduction | |||
4237 | /// result is returned in the first active lane. | |||
4238 | /// | |||
4239 | /// In this version, 'ShuffleReduceFn' behaves, per element, as follows: | |||
4240 | /// | |||
4241 | /// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE); | |||
4242 | /// if (lane_id % 2 == 0 && offset > 0) | |||
4243 | /// reduce_elem = reduce_elem REDUCE_OP remote_elem | |||
4244 | /// else | |||
4245 | /// reduce_elem = remote_elem | |||
4246 | /// | |||
4247 | /// | |||
4248 | /// Intra-Team Reduction | |||
4249 | /// | |||
4250 | /// This function, as implemented in the runtime call | |||
4251 | /// '__kmpc_nvptx_parallel_reduce_nowait_v2', aggregates data across OpenMP | |||
4252 | /// threads in a team. It first reduces within a warp using the | |||
4253 | /// aforementioned algorithms. We then proceed to gather all such | |||
4254 | /// reduced values at the first warp. | |||
4255 | /// | |||
4256 | /// The runtime makes use of the function 'InterWarpCpyFn', which copies | |||
4257 | /// data from each of the "warp master" (zeroth lane of each warp, where | |||
4258 | /// warp-reduced data is held) to the zeroth warp. This step reduces (in | |||
4259 | /// a mathematical sense) the problem of reduction across warp masters in | |||
4260 | /// a block to the problem of warp reduction. | |||
4261 | /// | |||
4262 | /// | |||
4263 | /// Inter-Team Reduction | |||
4264 | /// | |||
4265 | /// Once a team has reduced its data to a single value, it is stored in | |||
4266 | /// a global scratchpad array. Since each team has a distinct slot, this | |||
4267 | /// can be done without locking. | |||
4268 | /// | |||
4269 | /// The last team to write to the scratchpad array proceeds to reduce the | |||
4270 | /// scratchpad array. One or more workers in the last team use the helper | |||
4271 | /// 'loadAndReduceDataFn' to load and reduce values from the array, i.e., | |||
4272 | /// the k'th worker reduces every k'th element. | |||
4273 | /// | |||
4274 | /// Finally, a call is made to '__kmpc_nvptx_parallel_reduce_nowait_v2' to | |||
4275 | /// reduce across workers and compute a globally reduced value. | |||
4276 | /// | |||
4277 | void CGOpenMPRuntimeNVPTX::emitReduction( | |||
4278 | CodeGenFunction &CGF, SourceLocation Loc, ArrayRef<const Expr *> Privates, | |||
4279 | ArrayRef<const Expr *> LHSExprs, ArrayRef<const Expr *> RHSExprs, | |||
4280 | ArrayRef<const Expr *> ReductionOps, ReductionOptionsTy Options) { | |||
4281 | if (!CGF.HaveInsertPoint()) | |||
4282 | return; | |||
4283 | ||||
4284 | bool ParallelReduction = isOpenMPParallelDirective(Options.ReductionKind); | |||
4285 | #ifndef NDEBUG | |||
4286 | bool TeamsReduction = isOpenMPTeamsDirective(Options.ReductionKind); | |||
4287 | #endif | |||
4288 | ||||
4289 | if (Options.SimpleReduction) { | |||
4290 | assert(!TeamsReduction && !ParallelReduction &&((!TeamsReduction && !ParallelReduction && "Invalid reduction selection in emitReduction." ) ? static_cast<void> (0) : __assert_fail ("!TeamsReduction && !ParallelReduction && \"Invalid reduction selection in emitReduction.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 4291, __PRETTY_FUNCTION__)) | |||
4291 | "Invalid reduction selection in emitReduction.")((!TeamsReduction && !ParallelReduction && "Invalid reduction selection in emitReduction." ) ? static_cast<void> (0) : __assert_fail ("!TeamsReduction && !ParallelReduction && \"Invalid reduction selection in emitReduction.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 4291, __PRETTY_FUNCTION__)); | |||
4292 | CGOpenMPRuntime::emitReduction(CGF, Loc, Privates, LHSExprs, RHSExprs, | |||
4293 | ReductionOps, Options); | |||
4294 | return; | |||
4295 | } | |||
4296 | ||||
4297 | assert((TeamsReduction || ParallelReduction) &&(((TeamsReduction || ParallelReduction) && "Invalid reduction selection in emitReduction." ) ? static_cast<void> (0) : __assert_fail ("(TeamsReduction || ParallelReduction) && \"Invalid reduction selection in emitReduction.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 4298, __PRETTY_FUNCTION__)) | |||
4298 | "Invalid reduction selection in emitReduction.")(((TeamsReduction || ParallelReduction) && "Invalid reduction selection in emitReduction." ) ? static_cast<void> (0) : __assert_fail ("(TeamsReduction || ParallelReduction) && \"Invalid reduction selection in emitReduction.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 4298, __PRETTY_FUNCTION__)); | |||
4299 | ||||
4300 | // Build res = __kmpc_reduce{_nowait}(<gtid>, <n>, sizeof(RedList), | |||
4301 | // RedList, shuffle_reduce_func, interwarp_copy_func); | |||
4302 | // or | |||
4303 | // Build res = __kmpc_reduce_teams_nowait_simple(<loc>, <gtid>, <lck>); | |||
4304 | llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc); | |||
4305 | llvm::Value *ThreadId = getThreadID(CGF, Loc); | |||
4306 | ||||
4307 | llvm::Value *Res; | |||
4308 | ASTContext &C = CGM.getContext(); | |||
4309 | // 1. Build a list of reduction variables. | |||
4310 | // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]}; | |||
4311 | auto Size = RHSExprs.size(); | |||
4312 | for (const Expr *E : Privates) { | |||
4313 | if (E->getType()->isVariablyModifiedType()) | |||
4314 | // Reserve place for array size. | |||
4315 | ++Size; | |||
4316 | } | |||
4317 | llvm::APInt ArraySize(/*unsigned int numBits=*/32, Size); | |||
4318 | QualType ReductionArrayTy = | |||
4319 | C.getConstantArrayType(C.VoidPtrTy, ArraySize, nullptr, ArrayType::Normal, | |||
4320 | /*IndexTypeQuals=*/0); | |||
4321 | Address ReductionList = | |||
4322 | CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list"); | |||
4323 | auto IPriv = Privates.begin(); | |||
4324 | unsigned Idx = 0; | |||
4325 | for (unsigned I = 0, E = RHSExprs.size(); I < E; ++I, ++IPriv, ++Idx) { | |||
4326 | Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); | |||
4327 | CGF.Builder.CreateStore( | |||
4328 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
4329 | CGF.EmitLValue(RHSExprs[I]).getPointer(CGF), CGF.VoidPtrTy), | |||
4330 | Elem); | |||
4331 | if ((*IPriv)->getType()->isVariablyModifiedType()) { | |||
4332 | // Store array size. | |||
4333 | ++Idx; | |||
4334 | Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); | |||
4335 | llvm::Value *Size = CGF.Builder.CreateIntCast( | |||
4336 | CGF.getVLASize( | |||
4337 | CGF.getContext().getAsVariableArrayType((*IPriv)->getType())) | |||
4338 | .NumElts, | |||
4339 | CGF.SizeTy, /*isSigned=*/false); | |||
4340 | CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy), | |||
4341 | Elem); | |||
4342 | } | |||
4343 | } | |||
4344 | ||||
4345 | llvm::Value *RL = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
4346 | ReductionList.getPointer(), CGF.VoidPtrTy); | |||
4347 | llvm::Function *ReductionFn = emitReductionFunction( | |||
4348 | Loc, CGF.ConvertTypeForMem(ReductionArrayTy)->getPointerTo(), Privates, | |||
4349 | LHSExprs, RHSExprs, ReductionOps); | |||
4350 | llvm::Value *ReductionArrayTySize = CGF.getTypeSize(ReductionArrayTy); | |||
4351 | llvm::Function *ShuffleAndReduceFn = emitShuffleAndReduceFunction( | |||
4352 | CGM, Privates, ReductionArrayTy, ReductionFn, Loc); | |||
4353 | llvm::Value *InterWarpCopyFn = | |||
4354 | emitInterWarpCopyFunction(CGM, Privates, ReductionArrayTy, Loc); | |||
4355 | ||||
4356 | if (ParallelReduction) { | |||
4357 | llvm::Value *Args[] = {RTLoc, | |||
4358 | ThreadId, | |||
4359 | CGF.Builder.getInt32(RHSExprs.size()), | |||
4360 | ReductionArrayTySize, | |||
4361 | RL, | |||
4362 | ShuffleAndReduceFn, | |||
4363 | InterWarpCopyFn}; | |||
4364 | ||||
4365 | Res = CGF.EmitRuntimeCall( | |||
4366 | createNVPTXRuntimeFunction( | |||
4367 | OMPRTL_NVPTX__kmpc_nvptx_parallel_reduce_nowait_v2), | |||
4368 | Args); | |||
4369 | } else { | |||
4370 | assert(TeamsReduction && "expected teams reduction.")((TeamsReduction && "expected teams reduction.") ? static_cast <void> (0) : __assert_fail ("TeamsReduction && \"expected teams reduction.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 4370, __PRETTY_FUNCTION__)); | |||
4371 | llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> VarFieldMap; | |||
4372 | llvm::SmallVector<const ValueDecl *, 4> PrivatesReductions(Privates.size()); | |||
4373 | int Cnt = 0; | |||
4374 | for (const Expr *DRE : Privates) { | |||
4375 | PrivatesReductions[Cnt] = cast<DeclRefExpr>(DRE)->getDecl(); | |||
4376 | ++Cnt; | |||
4377 | } | |||
4378 | const RecordDecl *TeamReductionRec = ::buildRecordForGlobalizedVars( | |||
4379 | CGM.getContext(), PrivatesReductions, llvm::None, VarFieldMap, | |||
4380 | C.getLangOpts().OpenMPCUDAReductionBufNum); | |||
4381 | TeamsReductions.push_back(TeamReductionRec); | |||
4382 | if (!KernelTeamsReductionPtr) { | |||
4383 | KernelTeamsReductionPtr = new llvm::GlobalVariable( | |||
4384 | CGM.getModule(), CGM.VoidPtrTy, /*isConstant=*/true, | |||
4385 | llvm::GlobalValue::InternalLinkage, nullptr, | |||
4386 | "_openmp_teams_reductions_buffer_$_$ptr"); | |||
4387 | } | |||
4388 | llvm::Value *GlobalBufferPtr = CGF.EmitLoadOfScalar( | |||
4389 | Address(KernelTeamsReductionPtr, CGM.getPointerAlign()), | |||
4390 | /*Volatile=*/false, C.getPointerType(C.VoidPtrTy), Loc); | |||
4391 | llvm::Value *GlobalToBufferCpyFn = ::emitListToGlobalCopyFunction( | |||
4392 | CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap); | |||
4393 | llvm::Value *GlobalToBufferRedFn = ::emitListToGlobalReduceFunction( | |||
4394 | CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap, | |||
4395 | ReductionFn); | |||
4396 | llvm::Value *BufferToGlobalCpyFn = ::emitGlobalToListCopyFunction( | |||
4397 | CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap); | |||
4398 | llvm::Value *BufferToGlobalRedFn = ::emitGlobalToListReduceFunction( | |||
4399 | CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap, | |||
4400 | ReductionFn); | |||
4401 | ||||
4402 | llvm::Value *Args[] = { | |||
4403 | RTLoc, | |||
4404 | ThreadId, | |||
4405 | GlobalBufferPtr, | |||
4406 | CGF.Builder.getInt32(C.getLangOpts().OpenMPCUDAReductionBufNum), | |||
4407 | RL, | |||
4408 | ShuffleAndReduceFn, | |||
4409 | InterWarpCopyFn, | |||
4410 | GlobalToBufferCpyFn, | |||
4411 | GlobalToBufferRedFn, | |||
4412 | BufferToGlobalCpyFn, | |||
4413 | BufferToGlobalRedFn}; | |||
4414 | ||||
4415 | Res = CGF.EmitRuntimeCall( | |||
4416 | createNVPTXRuntimeFunction( | |||
4417 | OMPRTL_NVPTX__kmpc_nvptx_teams_reduce_nowait_v2), | |||
4418 | Args); | |||
4419 | } | |||
4420 | ||||
4421 | // 5. Build if (res == 1) | |||
4422 | llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".omp.reduction.done"); | |||
4423 | llvm::BasicBlock *ThenBB = CGF.createBasicBlock(".omp.reduction.then"); | |||
4424 | llvm::Value *Cond = CGF.Builder.CreateICmpEQ( | |||
4425 | Res, llvm::ConstantInt::get(CGM.Int32Ty, /*V=*/1)); | |||
4426 | CGF.Builder.CreateCondBr(Cond, ThenBB, ExitBB); | |||
4427 | ||||
4428 | // 6. Build then branch: where we have reduced values in the master | |||
4429 | // thread in each team. | |||
4430 | // __kmpc_end_reduce{_nowait}(<gtid>); | |||
4431 | // break; | |||
4432 | CGF.EmitBlock(ThenBB); | |||
4433 | ||||
4434 | // Add emission of __kmpc_end_reduce{_nowait}(<gtid>); | |||
4435 | auto &&CodeGen = [Privates, LHSExprs, RHSExprs, ReductionOps, | |||
4436 | this](CodeGenFunction &CGF, PrePostActionTy &Action) { | |||
4437 | auto IPriv = Privates.begin(); | |||
4438 | auto ILHS = LHSExprs.begin(); | |||
4439 | auto IRHS = RHSExprs.begin(); | |||
4440 | for (const Expr *E : ReductionOps) { | |||
4441 | emitSingleReductionCombiner(CGF, E, *IPriv, cast<DeclRefExpr>(*ILHS), | |||
4442 | cast<DeclRefExpr>(*IRHS)); | |||
4443 | ++IPriv; | |||
4444 | ++ILHS; | |||
4445 | ++IRHS; | |||
4446 | } | |||
4447 | }; | |||
4448 | llvm::Value *EndArgs[] = {ThreadId}; | |||
4449 | RegionCodeGenTy RCG(CodeGen); | |||
4450 | NVPTXActionTy Action( | |||
4451 | nullptr, llvm::None, | |||
4452 | createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_end_reduce_nowait), | |||
4453 | EndArgs); | |||
4454 | RCG.setAction(Action); | |||
4455 | RCG(CGF); | |||
4456 | // There is no need to emit line number for unconditional branch. | |||
4457 | (void)ApplyDebugLocation::CreateEmpty(CGF); | |||
4458 | CGF.EmitBlock(ExitBB, /*IsFinished=*/true); | |||
4459 | } | |||
4460 | ||||
4461 | const VarDecl * | |||
4462 | CGOpenMPRuntimeNVPTX::translateParameter(const FieldDecl *FD, | |||
4463 | const VarDecl *NativeParam) const { | |||
4464 | if (!NativeParam->getType()->isReferenceType()) | |||
4465 | return NativeParam; | |||
4466 | QualType ArgType = NativeParam->getType(); | |||
4467 | QualifierCollector QC; | |||
4468 | const Type *NonQualTy = QC.strip(ArgType); | |||
4469 | QualType PointeeTy = cast<ReferenceType>(NonQualTy)->getPointeeType(); | |||
4470 | if (const auto *Attr = FD->getAttr<OMPCaptureKindAttr>()) { | |||
4471 | if (Attr->getCaptureKind() == OMPC_map) { | |||
4472 | PointeeTy = CGM.getContext().getAddrSpaceQualType(PointeeTy, | |||
4473 | LangAS::opencl_global); | |||
4474 | } else if (Attr->getCaptureKind() == OMPC_firstprivate && | |||
4475 | PointeeTy.isConstant(CGM.getContext())) { | |||
4476 | PointeeTy = CGM.getContext().getAddrSpaceQualType(PointeeTy, | |||
4477 | LangAS::opencl_generic); | |||
4478 | } | |||
4479 | } | |||
4480 | ArgType = CGM.getContext().getPointerType(PointeeTy); | |||
4481 | QC.addRestrict(); | |||
4482 | enum { NVPTX_local_addr = 5 }; | |||
4483 | QC.addAddressSpace(getLangASFromTargetAS(NVPTX_local_addr)); | |||
4484 | ArgType = QC.apply(CGM.getContext(), ArgType); | |||
4485 | if (isa<ImplicitParamDecl>(NativeParam)) | |||
4486 | return ImplicitParamDecl::Create( | |||
4487 | CGM.getContext(), /*DC=*/nullptr, NativeParam->getLocation(), | |||
4488 | NativeParam->getIdentifier(), ArgType, ImplicitParamDecl::Other); | |||
4489 | return ParmVarDecl::Create( | |||
4490 | CGM.getContext(), | |||
4491 | const_cast<DeclContext *>(NativeParam->getDeclContext()), | |||
4492 | NativeParam->getBeginLoc(), NativeParam->getLocation(), | |||
4493 | NativeParam->getIdentifier(), ArgType, | |||
4494 | /*TInfo=*/nullptr, SC_None, /*DefArg=*/nullptr); | |||
4495 | } | |||
4496 | ||||
4497 | Address | |||
4498 | CGOpenMPRuntimeNVPTX::getParameterAddress(CodeGenFunction &CGF, | |||
4499 | const VarDecl *NativeParam, | |||
4500 | const VarDecl *TargetParam) const { | |||
4501 | assert(NativeParam != TargetParam &&((NativeParam != TargetParam && NativeParam->getType ()->isReferenceType() && "Native arg must not be the same as target arg." ) ? static_cast<void> (0) : __assert_fail ("NativeParam != TargetParam && NativeParam->getType()->isReferenceType() && \"Native arg must not be the same as target arg.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 4503, __PRETTY_FUNCTION__)) | |||
4502 | NativeParam->getType()->isReferenceType() &&((NativeParam != TargetParam && NativeParam->getType ()->isReferenceType() && "Native arg must not be the same as target arg." ) ? static_cast<void> (0) : __assert_fail ("NativeParam != TargetParam && NativeParam->getType()->isReferenceType() && \"Native arg must not be the same as target arg.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 4503, __PRETTY_FUNCTION__)) | |||
4503 | "Native arg must not be the same as target arg.")((NativeParam != TargetParam && NativeParam->getType ()->isReferenceType() && "Native arg must not be the same as target arg." ) ? static_cast<void> (0) : __assert_fail ("NativeParam != TargetParam && NativeParam->getType()->isReferenceType() && \"Native arg must not be the same as target arg.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 4503, __PRETTY_FUNCTION__)); | |||
4504 | Address LocalAddr = CGF.GetAddrOfLocalVar(TargetParam); | |||
4505 | QualType NativeParamType = NativeParam->getType(); | |||
4506 | QualifierCollector QC; | |||
4507 | const Type *NonQualTy = QC.strip(NativeParamType); | |||
4508 | QualType NativePointeeTy = cast<ReferenceType>(NonQualTy)->getPointeeType(); | |||
4509 | unsigned NativePointeeAddrSpace = | |||
4510 | CGF.getContext().getTargetAddressSpace(NativePointeeTy); | |||
4511 | QualType TargetTy = TargetParam->getType(); | |||
4512 | llvm::Value *TargetAddr = CGF.EmitLoadOfScalar( | |||
4513 | LocalAddr, /*Volatile=*/false, TargetTy, SourceLocation()); | |||
4514 | // First cast to generic. | |||
4515 | TargetAddr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
4516 | TargetAddr, TargetAddr->getType()->getPointerElementType()->getPointerTo( | |||
4517 | /*AddrSpace=*/0)); | |||
4518 | // Cast from generic to native address space. | |||
4519 | TargetAddr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
4520 | TargetAddr, TargetAddr->getType()->getPointerElementType()->getPointerTo( | |||
4521 | NativePointeeAddrSpace)); | |||
4522 | Address NativeParamAddr = CGF.CreateMemTemp(NativeParamType); | |||
4523 | CGF.EmitStoreOfScalar(TargetAddr, NativeParamAddr, /*Volatile=*/false, | |||
4524 | NativeParamType); | |||
4525 | return NativeParamAddr; | |||
4526 | } | |||
4527 | ||||
4528 | void CGOpenMPRuntimeNVPTX::emitOutlinedFunctionCall( | |||
4529 | CodeGenFunction &CGF, SourceLocation Loc, llvm::FunctionCallee OutlinedFn, | |||
4530 | ArrayRef<llvm::Value *> Args) const { | |||
4531 | SmallVector<llvm::Value *, 4> TargetArgs; | |||
4532 | TargetArgs.reserve(Args.size()); | |||
4533 | auto *FnType = OutlinedFn.getFunctionType(); | |||
4534 | for (unsigned I = 0, E = Args.size(); I < E; ++I) { | |||
4535 | if (FnType->isVarArg() && FnType->getNumParams() <= I) { | |||
4536 | TargetArgs.append(std::next(Args.begin(), I), Args.end()); | |||
4537 | break; | |||
4538 | } | |||
4539 | llvm::Type *TargetType = FnType->getParamType(I); | |||
4540 | llvm::Value *NativeArg = Args[I]; | |||
4541 | if (!TargetType->isPointerTy()) { | |||
4542 | TargetArgs.emplace_back(NativeArg); | |||
4543 | continue; | |||
4544 | } | |||
4545 | llvm::Value *TargetArg = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
4546 | NativeArg, | |||
4547 | NativeArg->getType()->getPointerElementType()->getPointerTo()); | |||
4548 | TargetArgs.emplace_back( | |||
4549 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(TargetArg, TargetType)); | |||
4550 | } | |||
4551 | CGOpenMPRuntime::emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, TargetArgs); | |||
4552 | } | |||
4553 | ||||
4554 | /// Emit function which wraps the outline parallel region | |||
4555 | /// and controls the arguments which are passed to this function. | |||
4556 | /// The wrapper ensures that the outlined function is called | |||
4557 | /// with the correct arguments when data is shared. | |||
4558 | llvm::Function *CGOpenMPRuntimeNVPTX::createParallelDataSharingWrapper( | |||
4559 | llvm::Function *OutlinedParallelFn, const OMPExecutableDirective &D) { | |||
4560 | ASTContext &Ctx = CGM.getContext(); | |||
4561 | const auto &CS = *D.getCapturedStmt(OMPD_parallel); | |||
4562 | ||||
4563 | // Create a function that takes as argument the source thread. | |||
4564 | FunctionArgList WrapperArgs; | |||
4565 | QualType Int16QTy = | |||
4566 | Ctx.getIntTypeForBitwidth(/*DestWidth=*/16, /*Signed=*/false); | |||
4567 | QualType Int32QTy = | |||
4568 | Ctx.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/false); | |||
4569 | ImplicitParamDecl ParallelLevelArg(Ctx, /*DC=*/nullptr, D.getBeginLoc(), | |||
4570 | /*Id=*/nullptr, Int16QTy, | |||
4571 | ImplicitParamDecl::Other); | |||
4572 | ImplicitParamDecl WrapperArg(Ctx, /*DC=*/nullptr, D.getBeginLoc(), | |||
4573 | /*Id=*/nullptr, Int32QTy, | |||
4574 | ImplicitParamDecl::Other); | |||
4575 | WrapperArgs.emplace_back(&ParallelLevelArg); | |||
4576 | WrapperArgs.emplace_back(&WrapperArg); | |||
4577 | ||||
4578 | const CGFunctionInfo &CGFI = | |||
4579 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(Ctx.VoidTy, WrapperArgs); | |||
4580 | ||||
4581 | auto *Fn = llvm::Function::Create( | |||
4582 | CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, | |||
4583 | Twine(OutlinedParallelFn->getName(), "_wrapper"), &CGM.getModule()); | |||
4584 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); | |||
4585 | Fn->setLinkage(llvm::GlobalValue::InternalLinkage); | |||
4586 | Fn->setDoesNotRecurse(); | |||
4587 | ||||
4588 | CodeGenFunction CGF(CGM, /*suppressNewContext=*/true); | |||
4589 | CGF.StartFunction(GlobalDecl(), Ctx.VoidTy, Fn, CGFI, WrapperArgs, | |||
4590 | D.getBeginLoc(), D.getBeginLoc()); | |||
4591 | ||||
4592 | const auto *RD = CS.getCapturedRecordDecl(); | |||
4593 | auto CurField = RD->field_begin(); | |||
4594 | ||||
4595 | Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty, | |||
4596 | /*Name=*/".zero.addr"); | |||
4597 | CGF.InitTempAlloca(ZeroAddr, CGF.Builder.getInt32(/*C*/ 0)); | |||
4598 | // Get the array of arguments. | |||
4599 | SmallVector<llvm::Value *, 8> Args; | |||
4600 | ||||
4601 | Args.emplace_back(CGF.GetAddrOfLocalVar(&WrapperArg).getPointer()); | |||
4602 | Args.emplace_back(ZeroAddr.getPointer()); | |||
4603 | ||||
4604 | CGBuilderTy &Bld = CGF.Builder; | |||
4605 | auto CI = CS.capture_begin(); | |||
4606 | ||||
4607 | // Use global memory for data sharing. | |||
4608 | // Handle passing of global args to workers. | |||
4609 | Address GlobalArgs = | |||
4610 | CGF.CreateDefaultAlignTempAlloca(CGF.VoidPtrPtrTy, "global_args"); | |||
4611 | llvm::Value *GlobalArgsPtr = GlobalArgs.getPointer(); | |||
4612 | llvm::Value *DataSharingArgs[] = {GlobalArgsPtr}; | |||
4613 | CGF.EmitRuntimeCall( | |||
4614 | createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_get_shared_variables), | |||
4615 | DataSharingArgs); | |||
4616 | ||||
4617 | // Retrieve the shared variables from the list of references returned | |||
4618 | // by the runtime. Pass the variables to the outlined function. | |||
4619 | Address SharedArgListAddress = Address::invalid(); | |||
4620 | if (CS.capture_size() > 0 || | |||
4621 | isOpenMPLoopBoundSharingDirective(D.getDirectiveKind())) { | |||
4622 | SharedArgListAddress = CGF.EmitLoadOfPointer( | |||
4623 | GlobalArgs, CGF.getContext() | |||
4624 | .getPointerType(CGF.getContext().getPointerType( | |||
4625 | CGF.getContext().VoidPtrTy)) | |||
4626 | .castAs<PointerType>()); | |||
4627 | } | |||
4628 | unsigned Idx = 0; | |||
4629 | if (isOpenMPLoopBoundSharingDirective(D.getDirectiveKind())) { | |||
4630 | Address Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, Idx); | |||
4631 | Address TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast( | |||
4632 | Src, CGF.SizeTy->getPointerTo()); | |||
4633 | llvm::Value *LB = CGF.EmitLoadOfScalar( | |||
4634 | TypedAddress, | |||
4635 | /*Volatile=*/false, | |||
4636 | CGF.getContext().getPointerType(CGF.getContext().getSizeType()), | |||
4637 | cast<OMPLoopDirective>(D).getLowerBoundVariable()->getExprLoc()); | |||
4638 | Args.emplace_back(LB); | |||
4639 | ++Idx; | |||
4640 | Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, Idx); | |||
4641 | TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast( | |||
4642 | Src, CGF.SizeTy->getPointerTo()); | |||
4643 | llvm::Value *UB = CGF.EmitLoadOfScalar( | |||
4644 | TypedAddress, | |||
4645 | /*Volatile=*/false, | |||
4646 | CGF.getContext().getPointerType(CGF.getContext().getSizeType()), | |||
4647 | cast<OMPLoopDirective>(D).getUpperBoundVariable()->getExprLoc()); | |||
4648 | Args.emplace_back(UB); | |||
4649 | ++Idx; | |||
4650 | } | |||
4651 | if (CS.capture_size() > 0) { | |||
4652 | ASTContext &CGFContext = CGF.getContext(); | |||
4653 | for (unsigned I = 0, E = CS.capture_size(); I < E; ++I, ++CI, ++CurField) { | |||
4654 | QualType ElemTy = CurField->getType(); | |||
4655 | Address Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, I + Idx); | |||
4656 | Address TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast( | |||
4657 | Src, CGF.ConvertTypeForMem(CGFContext.getPointerType(ElemTy))); | |||
4658 | llvm::Value *Arg = CGF.EmitLoadOfScalar(TypedAddress, | |||
4659 | /*Volatile=*/false, | |||
4660 | CGFContext.getPointerType(ElemTy), | |||
4661 | CI->getLocation()); | |||
4662 | if (CI->capturesVariableByCopy() && | |||
4663 | !CI->getCapturedVar()->getType()->isAnyPointerType()) { | |||
4664 | Arg = castValueToType(CGF, Arg, ElemTy, CGFContext.getUIntPtrType(), | |||
4665 | CI->getLocation()); | |||
4666 | } | |||
4667 | Args.emplace_back(Arg); | |||
4668 | } | |||
4669 | } | |||
4670 | ||||
4671 | emitOutlinedFunctionCall(CGF, D.getBeginLoc(), OutlinedParallelFn, Args); | |||
4672 | CGF.FinishFunction(); | |||
4673 | return Fn; | |||
4674 | } | |||
4675 | ||||
4676 | void CGOpenMPRuntimeNVPTX::emitFunctionProlog(CodeGenFunction &CGF, | |||
4677 | const Decl *D) { | |||
4678 | if (getDataSharingMode(CGM) != CGOpenMPRuntimeNVPTX::Generic) | |||
4679 | return; | |||
4680 | ||||
4681 | assert(D && "Expected function or captured|block decl.")((D && "Expected function or captured|block decl.") ? static_cast<void> (0) : __assert_fail ("D && \"Expected function or captured|block decl.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 4681, __PRETTY_FUNCTION__)); | |||
4682 | assert(FunctionGlobalizedDecls.count(CGF.CurFn) == 0 &&((FunctionGlobalizedDecls.count(CGF.CurFn) == 0 && "Function is registered already." ) ? static_cast<void> (0) : __assert_fail ("FunctionGlobalizedDecls.count(CGF.CurFn) == 0 && \"Function is registered already.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 4683, __PRETTY_FUNCTION__)) | |||
4683 | "Function is registered already.")((FunctionGlobalizedDecls.count(CGF.CurFn) == 0 && "Function is registered already." ) ? static_cast<void> (0) : __assert_fail ("FunctionGlobalizedDecls.count(CGF.CurFn) == 0 && \"Function is registered already.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 4683, __PRETTY_FUNCTION__)); | |||
4684 | assert((!TeamAndReductions.first || TeamAndReductions.first == D) &&(((!TeamAndReductions.first || TeamAndReductions.first == D) && "Team is set but not processed.") ? static_cast<void> ( 0) : __assert_fail ("(!TeamAndReductions.first || TeamAndReductions.first == D) && \"Team is set but not processed.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 4685, __PRETTY_FUNCTION__)) | |||
4685 | "Team is set but not processed.")(((!TeamAndReductions.first || TeamAndReductions.first == D) && "Team is set but not processed.") ? static_cast<void> ( 0) : __assert_fail ("(!TeamAndReductions.first || TeamAndReductions.first == D) && \"Team is set but not processed.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 4685, __PRETTY_FUNCTION__)); | |||
4686 | const Stmt *Body = nullptr; | |||
4687 | bool NeedToDelayGlobalization = false; | |||
4688 | if (const auto *FD = dyn_cast<FunctionDecl>(D)) { | |||
4689 | Body = FD->getBody(); | |||
4690 | } else if (const auto *BD = dyn_cast<BlockDecl>(D)) { | |||
4691 | Body = BD->getBody(); | |||
4692 | } else if (const auto *CD = dyn_cast<CapturedDecl>(D)) { | |||
4693 | Body = CD->getBody(); | |||
4694 | NeedToDelayGlobalization = CGF.CapturedStmtInfo->getKind() == CR_OpenMP; | |||
4695 | if (NeedToDelayGlobalization && | |||
4696 | getExecutionMode() == CGOpenMPRuntimeNVPTX::EM_SPMD) | |||
4697 | return; | |||
4698 | } | |||
4699 | if (!Body) | |||
4700 | return; | |||
4701 | CheckVarsEscapingDeclContext VarChecker(CGF, TeamAndReductions.second); | |||
4702 | VarChecker.Visit(Body); | |||
4703 | const RecordDecl *GlobalizedVarsRecord = | |||
4704 | VarChecker.getGlobalizedRecord(IsInTTDRegion); | |||
4705 | TeamAndReductions.first = nullptr; | |||
4706 | TeamAndReductions.second.clear(); | |||
4707 | ArrayRef<const ValueDecl *> EscapedVariableLengthDecls = | |||
4708 | VarChecker.getEscapedVariableLengthDecls(); | |||
4709 | if (!GlobalizedVarsRecord && EscapedVariableLengthDecls.empty()) | |||
4710 | return; | |||
4711 | auto I = FunctionGlobalizedDecls.try_emplace(CGF.CurFn).first; | |||
4712 | I->getSecond().MappedParams = | |||
4713 | std::make_unique<CodeGenFunction::OMPMapVars>(); | |||
4714 | I->getSecond().GlobalRecord = GlobalizedVarsRecord; | |||
4715 | I->getSecond().EscapedParameters.insert( | |||
4716 | VarChecker.getEscapedParameters().begin(), | |||
4717 | VarChecker.getEscapedParameters().end()); | |||
4718 | I->getSecond().EscapedVariableLengthDecls.append( | |||
4719 | EscapedVariableLengthDecls.begin(), EscapedVariableLengthDecls.end()); | |||
4720 | DeclToAddrMapTy &Data = I->getSecond().LocalVarData; | |||
4721 | for (const ValueDecl *VD : VarChecker.getEscapedDecls()) { | |||
4722 | assert(VD->isCanonicalDecl() && "Expected canonical declaration")((VD->isCanonicalDecl() && "Expected canonical declaration" ) ? static_cast<void> (0) : __assert_fail ("VD->isCanonicalDecl() && \"Expected canonical declaration\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 4722, __PRETTY_FUNCTION__)); | |||
4723 | const FieldDecl *FD = VarChecker.getFieldForGlobalizedVar(VD); | |||
4724 | Data.insert(std::make_pair(VD, MappedVarData(FD, IsInTTDRegion))); | |||
4725 | } | |||
4726 | if (!IsInTTDRegion && !NeedToDelayGlobalization && !IsInParallelRegion) { | |||
4727 | CheckVarsEscapingDeclContext VarChecker(CGF, llvm::None); | |||
4728 | VarChecker.Visit(Body); | |||
4729 | I->getSecond().SecondaryGlobalRecord = | |||
4730 | VarChecker.getGlobalizedRecord(/*IsInTTDRegion=*/true); | |||
4731 | I->getSecond().SecondaryLocalVarData.emplace(); | |||
4732 | DeclToAddrMapTy &Data = I->getSecond().SecondaryLocalVarData.getValue(); | |||
4733 | for (const ValueDecl *VD : VarChecker.getEscapedDecls()) { | |||
4734 | assert(VD->isCanonicalDecl() && "Expected canonical declaration")((VD->isCanonicalDecl() && "Expected canonical declaration" ) ? static_cast<void> (0) : __assert_fail ("VD->isCanonicalDecl() && \"Expected canonical declaration\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 4734, __PRETTY_FUNCTION__)); | |||
4735 | const FieldDecl *FD = VarChecker.getFieldForGlobalizedVar(VD); | |||
4736 | Data.insert( | |||
4737 | std::make_pair(VD, MappedVarData(FD, /*IsInTTDRegion=*/true))); | |||
4738 | } | |||
4739 | } | |||
4740 | if (!NeedToDelayGlobalization) { | |||
4741 | emitGenericVarsProlog(CGF, D->getBeginLoc(), /*WithSPMDCheck=*/true); | |||
4742 | struct GlobalizationScope final : EHScopeStack::Cleanup { | |||
4743 | GlobalizationScope() = default; | |||
4744 | ||||
4745 | void Emit(CodeGenFunction &CGF, Flags flags) override { | |||
4746 | static_cast<CGOpenMPRuntimeNVPTX &>(CGF.CGM.getOpenMPRuntime()) | |||
4747 | .emitGenericVarsEpilog(CGF, /*WithSPMDCheck=*/true); | |||
4748 | } | |||
4749 | }; | |||
4750 | CGF.EHStack.pushCleanup<GlobalizationScope>(NormalAndEHCleanup); | |||
4751 | } | |||
4752 | } | |||
4753 | ||||
4754 | Address CGOpenMPRuntimeNVPTX::getAddressOfLocalVariable(CodeGenFunction &CGF, | |||
4755 | const VarDecl *VD) { | |||
4756 | if (VD && VD->hasAttr<OMPAllocateDeclAttr>()) { | |||
| ||||
4757 | const auto *A = VD->getAttr<OMPAllocateDeclAttr>(); | |||
4758 | switch (A->getAllocatorType()) { | |||
4759 | // Use the default allocator here as by default local vars are | |||
4760 | // threadlocal. | |||
4761 | case OMPAllocateDeclAttr::OMPDefaultMemAlloc: | |||
4762 | case OMPAllocateDeclAttr::OMPThreadMemAlloc: | |||
4763 | case OMPAllocateDeclAttr::OMPHighBWMemAlloc: | |||
4764 | case OMPAllocateDeclAttr::OMPLowLatMemAlloc: | |||
4765 | // Follow the user decision - use default allocation. | |||
4766 | return Address::invalid(); | |||
4767 | case OMPAllocateDeclAttr::OMPUserDefinedMemAlloc: | |||
4768 | // TODO: implement aupport for user-defined allocators. | |||
4769 | return Address::invalid(); | |||
4770 | case OMPAllocateDeclAttr::OMPConstMemAlloc: { | |||
4771 | llvm::Type *VarTy = CGF.ConvertTypeForMem(VD->getType()); | |||
4772 | auto *GV = new llvm::GlobalVariable( | |||
4773 | CGM.getModule(), VarTy, /*isConstant=*/false, | |||
4774 | llvm::GlobalValue::InternalLinkage, | |||
4775 | llvm::Constant::getNullValue(VarTy), VD->getName(), | |||
4776 | /*InsertBefore=*/nullptr, llvm::GlobalValue::NotThreadLocal, | |||
4777 | CGM.getContext().getTargetAddressSpace(LangAS::cuda_constant)); | |||
4778 | CharUnits Align = CGM.getContext().getDeclAlign(VD); | |||
4779 | GV->setAlignment(Align.getAsAlign()); | |||
4780 | return Address(GV, Align); | |||
4781 | } | |||
4782 | case OMPAllocateDeclAttr::OMPPTeamMemAlloc: { | |||
4783 | llvm::Type *VarTy = CGF.ConvertTypeForMem(VD->getType()); | |||
4784 | auto *GV = new llvm::GlobalVariable( | |||
4785 | CGM.getModule(), VarTy, /*isConstant=*/false, | |||
4786 | llvm::GlobalValue::InternalLinkage, | |||
4787 | llvm::Constant::getNullValue(VarTy), VD->getName(), | |||
4788 | /*InsertBefore=*/nullptr, llvm::GlobalValue::NotThreadLocal, | |||
4789 | CGM.getContext().getTargetAddressSpace(LangAS::cuda_shared)); | |||
4790 | CharUnits Align = CGM.getContext().getDeclAlign(VD); | |||
4791 | GV->setAlignment(Align.getAsAlign()); | |||
4792 | return Address(GV, Align); | |||
4793 | } | |||
4794 | case OMPAllocateDeclAttr::OMPLargeCapMemAlloc: | |||
4795 | case OMPAllocateDeclAttr::OMPCGroupMemAlloc: { | |||
4796 | llvm::Type *VarTy = CGF.ConvertTypeForMem(VD->getType()); | |||
4797 | auto *GV = new llvm::GlobalVariable( | |||
4798 | CGM.getModule(), VarTy, /*isConstant=*/false, | |||
4799 | llvm::GlobalValue::InternalLinkage, | |||
4800 | llvm::Constant::getNullValue(VarTy), VD->getName()); | |||
4801 | CharUnits Align = CGM.getContext().getDeclAlign(VD); | |||
4802 | GV->setAlignment(Align.getAsAlign()); | |||
4803 | return Address(GV, Align); | |||
4804 | } | |||
4805 | } | |||
4806 | } | |||
4807 | ||||
4808 | if (getDataSharingMode(CGM) != CGOpenMPRuntimeNVPTX::Generic) | |||
4809 | return Address::invalid(); | |||
4810 | ||||
4811 | VD = VD->getCanonicalDecl(); | |||
| ||||
4812 | auto I = FunctionGlobalizedDecls.find(CGF.CurFn); | |||
4813 | if (I == FunctionGlobalizedDecls.end()) | |||
4814 | return Address::invalid(); | |||
4815 | auto VDI = I->getSecond().LocalVarData.find(VD); | |||
4816 | if (VDI != I->getSecond().LocalVarData.end()) | |||
4817 | return VDI->second.PrivateAddr; | |||
4818 | if (VD->hasAttrs()) { | |||
4819 | for (specific_attr_iterator<OMPReferencedVarAttr> IT(VD->attr_begin()), | |||
4820 | E(VD->attr_end()); | |||
4821 | IT != E; ++IT) { | |||
4822 | auto VDI = I->getSecond().LocalVarData.find( | |||
4823 | cast<VarDecl>(cast<DeclRefExpr>(IT->getRef())->getDecl()) | |||
4824 | ->getCanonicalDecl()); | |||
4825 | if (VDI != I->getSecond().LocalVarData.end()) | |||
4826 | return VDI->second.PrivateAddr; | |||
4827 | } | |||
4828 | } | |||
4829 | ||||
4830 | return Address::invalid(); | |||
4831 | } | |||
4832 | ||||
4833 | void CGOpenMPRuntimeNVPTX::functionFinished(CodeGenFunction &CGF) { | |||
4834 | FunctionGlobalizedDecls.erase(CGF.CurFn); | |||
4835 | CGOpenMPRuntime::functionFinished(CGF); | |||
4836 | } | |||
4837 | ||||
4838 | void CGOpenMPRuntimeNVPTX::getDefaultDistScheduleAndChunk( | |||
4839 | CodeGenFunction &CGF, const OMPLoopDirective &S, | |||
4840 | OpenMPDistScheduleClauseKind &ScheduleKind, | |||
4841 | llvm::Value *&Chunk) const { | |||
4842 | if (getExecutionMode() == CGOpenMPRuntimeNVPTX::EM_SPMD) { | |||
4843 | ScheduleKind = OMPC_DIST_SCHEDULE_static; | |||
4844 | Chunk = CGF.EmitScalarConversion(getNVPTXNumThreads(CGF), | |||
4845 | CGF.getContext().getIntTypeForBitwidth(32, /*Signed=*/0), | |||
4846 | S.getIterationVariable()->getType(), S.getBeginLoc()); | |||
4847 | return; | |||
4848 | } | |||
4849 | CGOpenMPRuntime::getDefaultDistScheduleAndChunk( | |||
4850 | CGF, S, ScheduleKind, Chunk); | |||
4851 | } | |||
4852 | ||||
4853 | void CGOpenMPRuntimeNVPTX::getDefaultScheduleAndChunk( | |||
4854 | CodeGenFunction &CGF, const OMPLoopDirective &S, | |||
4855 | OpenMPScheduleClauseKind &ScheduleKind, | |||
4856 | const Expr *&ChunkExpr) const { | |||
4857 | ScheduleKind = OMPC_SCHEDULE_static; | |||
4858 | // Chunk size is 1 in this case. | |||
4859 | llvm::APInt ChunkSize(32, 1); | |||
4860 | ChunkExpr = IntegerLiteral::Create(CGF.getContext(), ChunkSize, | |||
4861 | CGF.getContext().getIntTypeForBitwidth(32, /*Signed=*/0), | |||
4862 | SourceLocation()); | |||
4863 | } | |||
4864 | ||||
4865 | void CGOpenMPRuntimeNVPTX::adjustTargetSpecificDataForLambdas( | |||
4866 | CodeGenFunction &CGF, const OMPExecutableDirective &D) const { | |||
4867 | assert(isOpenMPTargetExecutionDirective(D.getDirectiveKind()) &&((isOpenMPTargetExecutionDirective(D.getDirectiveKind()) && " Expected target-based directive.") ? static_cast<void> (0) : __assert_fail ("isOpenMPTargetExecutionDirective(D.getDirectiveKind()) && \" Expected target-based directive.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 4868, __PRETTY_FUNCTION__)) | |||
4868 | " Expected target-based directive.")((isOpenMPTargetExecutionDirective(D.getDirectiveKind()) && " Expected target-based directive.") ? static_cast<void> (0) : __assert_fail ("isOpenMPTargetExecutionDirective(D.getDirectiveKind()) && \" Expected target-based directive.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 4868, __PRETTY_FUNCTION__)); | |||
4869 | const CapturedStmt *CS = D.getCapturedStmt(OMPD_target); | |||
4870 | for (const CapturedStmt::Capture &C : CS->captures()) { | |||
4871 | // Capture variables captured by reference in lambdas for target-based | |||
4872 | // directives. | |||
4873 | if (!C.capturesVariable()) | |||
4874 | continue; | |||
4875 | const VarDecl *VD = C.getCapturedVar(); | |||
4876 | const auto *RD = VD->getType() | |||
4877 | .getCanonicalType() | |||
4878 | .getNonReferenceType() | |||
4879 | ->getAsCXXRecordDecl(); | |||
4880 | if (!RD || !RD->isLambda()) | |||
4881 | continue; | |||
4882 | Address VDAddr = CGF.GetAddrOfLocalVar(VD); | |||
4883 | LValue VDLVal; | |||
4884 | if (VD->getType().getCanonicalType()->isReferenceType()) | |||
4885 | VDLVal = CGF.EmitLoadOfReferenceLValue(VDAddr, VD->getType()); | |||
4886 | else | |||
4887 | VDLVal = CGF.MakeAddrLValue( | |||
4888 | VDAddr, VD->getType().getCanonicalType().getNonReferenceType()); | |||
4889 | llvm::DenseMap<const VarDecl *, FieldDecl *> Captures; | |||
4890 | FieldDecl *ThisCapture = nullptr; | |||
4891 | RD->getCaptureFields(Captures, ThisCapture); | |||
4892 | if (ThisCapture && CGF.CapturedStmtInfo->isCXXThisExprCaptured()) { | |||
4893 | LValue ThisLVal = | |||
4894 | CGF.EmitLValueForFieldInitialization(VDLVal, ThisCapture); | |||
4895 | llvm::Value *CXXThis = CGF.LoadCXXThis(); | |||
4896 | CGF.EmitStoreOfScalar(CXXThis, ThisLVal); | |||
4897 | } | |||
4898 | for (const LambdaCapture &LC : RD->captures()) { | |||
4899 | if (LC.getCaptureKind() != LCK_ByRef) | |||
4900 | continue; | |||
4901 | const VarDecl *VD = LC.getCapturedVar(); | |||
4902 | if (!CS->capturesVariable(VD)) | |||
4903 | continue; | |||
4904 | auto It = Captures.find(VD); | |||
4905 | assert(It != Captures.end() && "Found lambda capture without field.")((It != Captures.end() && "Found lambda capture without field." ) ? static_cast<void> (0) : __assert_fail ("It != Captures.end() && \"Found lambda capture without field.\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 4905, __PRETTY_FUNCTION__)); | |||
4906 | LValue VarLVal = CGF.EmitLValueForFieldInitialization(VDLVal, It->second); | |||
4907 | Address VDAddr = CGF.GetAddrOfLocalVar(VD); | |||
4908 | if (VD->getType().getCanonicalType()->isReferenceType()) | |||
4909 | VDAddr = CGF.EmitLoadOfReferenceLValue(VDAddr, | |||
4910 | VD->getType().getCanonicalType()) | |||
4911 | .getAddress(CGF); | |||
4912 | CGF.EmitStoreOfScalar(VDAddr.getPointer(), VarLVal); | |||
4913 | } | |||
4914 | } | |||
4915 | } | |||
4916 | ||||
4917 | unsigned CGOpenMPRuntimeNVPTX::getDefaultFirstprivateAddressSpace() const { | |||
4918 | return CGM.getContext().getTargetAddressSpace(LangAS::cuda_constant); | |||
4919 | } | |||
4920 | ||||
4921 | bool CGOpenMPRuntimeNVPTX::hasAllocateAttributeForGlobalVar(const VarDecl *VD, | |||
4922 | LangAS &AS) { | |||
4923 | if (!VD || !VD->hasAttr<OMPAllocateDeclAttr>()) | |||
4924 | return false; | |||
4925 | const auto *A = VD->getAttr<OMPAllocateDeclAttr>(); | |||
4926 | switch(A->getAllocatorType()) { | |||
4927 | case OMPAllocateDeclAttr::OMPDefaultMemAlloc: | |||
4928 | // Not supported, fallback to the default mem space. | |||
4929 | case OMPAllocateDeclAttr::OMPThreadMemAlloc: | |||
4930 | case OMPAllocateDeclAttr::OMPLargeCapMemAlloc: | |||
4931 | case OMPAllocateDeclAttr::OMPCGroupMemAlloc: | |||
4932 | case OMPAllocateDeclAttr::OMPHighBWMemAlloc: | |||
4933 | case OMPAllocateDeclAttr::OMPLowLatMemAlloc: | |||
4934 | AS = LangAS::Default; | |||
4935 | return true; | |||
4936 | case OMPAllocateDeclAttr::OMPConstMemAlloc: | |||
4937 | AS = LangAS::cuda_constant; | |||
4938 | return true; | |||
4939 | case OMPAllocateDeclAttr::OMPPTeamMemAlloc: | |||
4940 | AS = LangAS::cuda_shared; | |||
4941 | return true; | |||
4942 | case OMPAllocateDeclAttr::OMPUserDefinedMemAlloc: | |||
4943 | llvm_unreachable("Expected predefined allocator for the variables with the "::llvm::llvm_unreachable_internal("Expected predefined allocator for the variables with the " "static storage.", "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 4944) | |||
4944 | "static storage.")::llvm::llvm_unreachable_internal("Expected predefined allocator for the variables with the " "static storage.", "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 4944); | |||
4945 | } | |||
4946 | return false; | |||
4947 | } | |||
4948 | ||||
4949 | // Get current CudaArch and ignore any unknown values | |||
4950 | static CudaArch getCudaArch(CodeGenModule &CGM) { | |||
4951 | if (!CGM.getTarget().hasFeature("ptx")) | |||
4952 | return CudaArch::UNKNOWN; | |||
4953 | llvm::StringMap<bool> Features; | |||
4954 | CGM.getTarget().initFeatureMap(Features, CGM.getDiags(), | |||
4955 | CGM.getTarget().getTargetOpts().CPU, | |||
4956 | CGM.getTarget().getTargetOpts().Features); | |||
4957 | for (const auto &Feature : Features) { | |||
4958 | if (Feature.getValue()) { | |||
4959 | CudaArch Arch = StringToCudaArch(Feature.getKey()); | |||
4960 | if (Arch != CudaArch::UNKNOWN) | |||
4961 | return Arch; | |||
4962 | } | |||
4963 | } | |||
4964 | return CudaArch::UNKNOWN; | |||
4965 | } | |||
4966 | ||||
4967 | /// Check to see if target architecture supports unified addressing which is | |||
4968 | /// a restriction for OpenMP requires clause "unified_shared_memory". | |||
4969 | void CGOpenMPRuntimeNVPTX::processRequiresDirective( | |||
4970 | const OMPRequiresDecl *D) { | |||
4971 | for (const OMPClause *Clause : D->clauselists()) { | |||
4972 | if (Clause->getClauseKind() == OMPC_unified_shared_memory) { | |||
4973 | CudaArch Arch = getCudaArch(CGM); | |||
4974 | switch (Arch) { | |||
4975 | case CudaArch::SM_20: | |||
4976 | case CudaArch::SM_21: | |||
4977 | case CudaArch::SM_30: | |||
4978 | case CudaArch::SM_32: | |||
4979 | case CudaArch::SM_35: | |||
4980 | case CudaArch::SM_37: | |||
4981 | case CudaArch::SM_50: | |||
4982 | case CudaArch::SM_52: | |||
4983 | case CudaArch::SM_53: | |||
4984 | case CudaArch::SM_60: | |||
4985 | case CudaArch::SM_61: | |||
4986 | case CudaArch::SM_62: { | |||
4987 | SmallString<256> Buffer; | |||
4988 | llvm::raw_svector_ostream Out(Buffer); | |||
4989 | Out << "Target architecture " << CudaArchToString(Arch) | |||
4990 | << " does not support unified addressing"; | |||
4991 | CGM.Error(Clause->getBeginLoc(), Out.str()); | |||
4992 | return; | |||
4993 | } | |||
4994 | case CudaArch::SM_70: | |||
4995 | case CudaArch::SM_72: | |||
4996 | case CudaArch::SM_75: | |||
4997 | case CudaArch::GFX600: | |||
4998 | case CudaArch::GFX601: | |||
4999 | case CudaArch::GFX700: | |||
5000 | case CudaArch::GFX701: | |||
5001 | case CudaArch::GFX702: | |||
5002 | case CudaArch::GFX703: | |||
5003 | case CudaArch::GFX704: | |||
5004 | case CudaArch::GFX801: | |||
5005 | case CudaArch::GFX802: | |||
5006 | case CudaArch::GFX803: | |||
5007 | case CudaArch::GFX810: | |||
5008 | case CudaArch::GFX900: | |||
5009 | case CudaArch::GFX902: | |||
5010 | case CudaArch::GFX904: | |||
5011 | case CudaArch::GFX906: | |||
5012 | case CudaArch::GFX908: | |||
5013 | case CudaArch::GFX909: | |||
5014 | case CudaArch::GFX1010: | |||
5015 | case CudaArch::GFX1011: | |||
5016 | case CudaArch::GFX1012: | |||
5017 | case CudaArch::UNKNOWN: | |||
5018 | break; | |||
5019 | case CudaArch::LAST: | |||
5020 | llvm_unreachable("Unexpected Cuda arch.")::llvm::llvm_unreachable_internal("Unexpected Cuda arch.", "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 5020); | |||
5021 | } | |||
5022 | } | |||
5023 | } | |||
5024 | CGOpenMPRuntime::processRequiresDirective(D); | |||
5025 | } | |||
5026 | ||||
5027 | /// Get number of SMs and number of blocks per SM. | |||
5028 | static std::pair<unsigned, unsigned> getSMsBlocksPerSM(CodeGenModule &CGM) { | |||
5029 | std::pair<unsigned, unsigned> Data; | |||
5030 | if (CGM.getLangOpts().OpenMPCUDANumSMs) | |||
5031 | Data.first = CGM.getLangOpts().OpenMPCUDANumSMs; | |||
5032 | if (CGM.getLangOpts().OpenMPCUDABlocksPerSM) | |||
5033 | Data.second = CGM.getLangOpts().OpenMPCUDABlocksPerSM; | |||
5034 | if (Data.first && Data.second) | |||
5035 | return Data; | |||
5036 | switch (getCudaArch(CGM)) { | |||
5037 | case CudaArch::SM_20: | |||
5038 | case CudaArch::SM_21: | |||
5039 | case CudaArch::SM_30: | |||
5040 | case CudaArch::SM_32: | |||
5041 | case CudaArch::SM_35: | |||
5042 | case CudaArch::SM_37: | |||
5043 | case CudaArch::SM_50: | |||
5044 | case CudaArch::SM_52: | |||
5045 | case CudaArch::SM_53: | |||
5046 | return {16, 16}; | |||
5047 | case CudaArch::SM_60: | |||
5048 | case CudaArch::SM_61: | |||
5049 | case CudaArch::SM_62: | |||
5050 | return {56, 32}; | |||
5051 | case CudaArch::SM_70: | |||
5052 | case CudaArch::SM_72: | |||
5053 | case CudaArch::SM_75: | |||
5054 | return {84, 32}; | |||
5055 | case CudaArch::GFX600: | |||
5056 | case CudaArch::GFX601: | |||
5057 | case CudaArch::GFX700: | |||
5058 | case CudaArch::GFX701: | |||
5059 | case CudaArch::GFX702: | |||
5060 | case CudaArch::GFX703: | |||
5061 | case CudaArch::GFX704: | |||
5062 | case CudaArch::GFX801: | |||
5063 | case CudaArch::GFX802: | |||
5064 | case CudaArch::GFX803: | |||
5065 | case CudaArch::GFX810: | |||
5066 | case CudaArch::GFX900: | |||
5067 | case CudaArch::GFX902: | |||
5068 | case CudaArch::GFX904: | |||
5069 | case CudaArch::GFX906: | |||
5070 | case CudaArch::GFX908: | |||
5071 | case CudaArch::GFX909: | |||
5072 | case CudaArch::GFX1010: | |||
5073 | case CudaArch::GFX1011: | |||
5074 | case CudaArch::GFX1012: | |||
5075 | case CudaArch::UNKNOWN: | |||
5076 | break; | |||
5077 | case CudaArch::LAST: | |||
5078 | llvm_unreachable("Unexpected Cuda arch.")::llvm::llvm_unreachable_internal("Unexpected Cuda arch.", "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 5078); | |||
5079 | } | |||
5080 | llvm_unreachable("Unexpected NVPTX target without ptx feature.")::llvm::llvm_unreachable_internal("Unexpected NVPTX target without ptx feature." , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp" , 5080); | |||
5081 | } | |||
5082 | ||||
5083 | void CGOpenMPRuntimeNVPTX::clear() { | |||
5084 | if (!GlobalizedRecords.empty()) { | |||
5085 | ASTContext &C = CGM.getContext(); | |||
5086 | llvm::SmallVector<const GlobalPtrSizeRecsTy *, 4> GlobalRecs; | |||
5087 | llvm::SmallVector<const GlobalPtrSizeRecsTy *, 4> SharedRecs; | |||
5088 | RecordDecl *StaticRD = C.buildImplicitRecord( | |||
5089 | "_openmp_static_memory_type_$_", RecordDecl::TagKind::TTK_Union); | |||
5090 | StaticRD->startDefinition(); | |||
5091 | RecordDecl *SharedStaticRD = C.buildImplicitRecord( | |||
5092 | "_shared_openmp_static_memory_type_$_", RecordDecl::TagKind::TTK_Union); | |||
5093 | SharedStaticRD->startDefinition(); | |||
5094 | for (const GlobalPtrSizeRecsTy &Records : GlobalizedRecords) { | |||
5095 | if (Records.Records.empty()) | |||
5096 | continue; | |||
5097 | unsigned Size = 0; | |||
5098 | unsigned RecAlignment = 0; | |||
5099 | for (const RecordDecl *RD : Records.Records) { | |||
5100 | QualType RDTy = C.getRecordType(RD); | |||
5101 | unsigned Alignment = C.getTypeAlignInChars(RDTy).getQuantity(); | |||
5102 | RecAlignment = std::max(RecAlignment, Alignment); | |||
5103 | unsigned RecSize = C.getTypeSizeInChars(RDTy).getQuantity(); | |||
5104 | Size = | |||
5105 | llvm::alignTo(llvm::alignTo(Size, Alignment) + RecSize, Alignment); | |||
5106 | } | |||
5107 | Size = llvm::alignTo(Size, RecAlignment); | |||
5108 | llvm::APInt ArySize(/*numBits=*/64, Size); | |||
5109 | QualType SubTy = C.getConstantArrayType( | |||
5110 | C.CharTy, ArySize, nullptr, ArrayType::Normal, /*IndexTypeQuals=*/0); | |||
5111 | const bool UseSharedMemory = Size <= SharedMemorySize; | |||
5112 | auto *Field = | |||
5113 | FieldDecl::Create(C, UseSharedMemory ? SharedStaticRD : StaticRD, | |||
5114 | SourceLocation(), SourceLocation(), nullptr, SubTy, | |||
5115 | C.getTrivialTypeSourceInfo(SubTy, SourceLocation()), | |||
5116 | /*BW=*/nullptr, /*Mutable=*/false, | |||
5117 | /*InitStyle=*/ICIS_NoInit); | |||
5118 | Field->setAccess(AS_public); | |||
5119 | if (UseSharedMemory) { | |||
5120 | SharedStaticRD->addDecl(Field); | |||
5121 | SharedRecs.push_back(&Records); | |||
5122 | } else { | |||
5123 | StaticRD->addDecl(Field); | |||
5124 | GlobalRecs.push_back(&Records); | |||
5125 | } | |||
5126 | Records.RecSize->setInitializer(llvm::ConstantInt::get(CGM.SizeTy, Size)); | |||
5127 | Records.UseSharedMemory->setInitializer( | |||
5128 | llvm::ConstantInt::get(CGM.Int16Ty, UseSharedMemory ? 1 : 0)); | |||
5129 | } | |||
5130 | // Allocate SharedMemorySize buffer for the shared memory. | |||
5131 | // FIXME: nvlink does not handle weak linkage correctly (object with the | |||
5132 | // different size are reported as erroneous). | |||
5133 | // Restore this code as sson as nvlink is fixed. | |||
5134 | if (!SharedStaticRD->field_empty()) { | |||
5135 | llvm::APInt ArySize(/*numBits=*/64, SharedMemorySize); | |||
5136 | QualType SubTy = C.getConstantArrayType( | |||
5137 | C.CharTy, ArySize, nullptr, ArrayType::Normal, /*IndexTypeQuals=*/0); | |||
5138 | auto *Field = FieldDecl::Create( | |||
5139 | C, SharedStaticRD, SourceLocation(), SourceLocation(), nullptr, SubTy, | |||
5140 | C.getTrivialTypeSourceInfo(SubTy, SourceLocation()), | |||
5141 | /*BW=*/nullptr, /*Mutable=*/false, | |||
5142 | /*InitStyle=*/ICIS_NoInit); | |||
5143 | Field->setAccess(AS_public); | |||
5144 | SharedStaticRD->addDecl(Field); | |||
5145 | } | |||
5146 | SharedStaticRD->completeDefinition(); | |||
5147 | if (!SharedStaticRD->field_empty()) { | |||
5148 | QualType StaticTy = C.getRecordType(SharedStaticRD); | |||
5149 | llvm::Type *LLVMStaticTy = CGM.getTypes().ConvertTypeForMem(StaticTy); | |||
5150 | auto *GV = new llvm::GlobalVariable( | |||
5151 | CGM.getModule(), LLVMStaticTy, | |||
5152 | /*isConstant=*/false, llvm::GlobalValue::CommonLinkage, | |||
5153 | llvm::Constant::getNullValue(LLVMStaticTy), | |||
5154 | "_openmp_shared_static_glob_rd_$_", /*InsertBefore=*/nullptr, | |||
5155 | llvm::GlobalValue::NotThreadLocal, | |||
5156 | C.getTargetAddressSpace(LangAS::cuda_shared)); | |||
5157 | auto *Replacement = llvm::ConstantExpr::getPointerBitCastOrAddrSpaceCast( | |||
5158 | GV, CGM.VoidPtrTy); | |||
5159 | for (const GlobalPtrSizeRecsTy *Rec : SharedRecs) { | |||
5160 | Rec->Buffer->replaceAllUsesWith(Replacement); | |||
5161 | Rec->Buffer->eraseFromParent(); | |||
5162 | } | |||
5163 | } | |||
5164 | StaticRD->completeDefinition(); | |||
5165 | if (!StaticRD->field_empty()) { | |||
5166 | QualType StaticTy = C.getRecordType(StaticRD); | |||
5167 | std::pair<unsigned, unsigned> SMsBlockPerSM = getSMsBlocksPerSM(CGM); | |||
5168 | llvm::APInt Size1(32, SMsBlockPerSM.second); | |||
5169 | QualType Arr1Ty = | |||
5170 | C.getConstantArrayType(StaticTy, Size1, nullptr, ArrayType::Normal, | |||
5171 | /*IndexTypeQuals=*/0); | |||
5172 | llvm::APInt Size2(32, SMsBlockPerSM.first); | |||
5173 | QualType Arr2Ty = | |||
5174 | C.getConstantArrayType(Arr1Ty, Size2, nullptr, ArrayType::Normal, | |||
5175 | /*IndexTypeQuals=*/0); | |||
5176 | llvm::Type *LLVMArr2Ty = CGM.getTypes().ConvertTypeForMem(Arr2Ty); | |||
5177 | // FIXME: nvlink does not handle weak linkage correctly (object with the | |||
5178 | // different size are reported as erroneous). | |||
5179 | // Restore CommonLinkage as soon as nvlink is fixed. | |||
5180 | auto *GV = new llvm::GlobalVariable( | |||
5181 | CGM.getModule(), LLVMArr2Ty, | |||
5182 | /*isConstant=*/false, llvm::GlobalValue::InternalLinkage, | |||
5183 | llvm::Constant::getNullValue(LLVMArr2Ty), | |||
5184 | "_openmp_static_glob_rd_$_"); | |||
5185 | auto *Replacement = llvm::ConstantExpr::getPointerBitCastOrAddrSpaceCast( | |||
5186 | GV, CGM.VoidPtrTy); | |||
5187 | for (const GlobalPtrSizeRecsTy *Rec : GlobalRecs) { | |||
5188 | Rec->Buffer->replaceAllUsesWith(Replacement); | |||
5189 | Rec->Buffer->eraseFromParent(); | |||
5190 | } | |||
5191 | } | |||
5192 | } | |||
5193 | if (!TeamsReductions.empty()) { | |||
5194 | ASTContext &C = CGM.getContext(); | |||
5195 | RecordDecl *StaticRD = C.buildImplicitRecord( | |||
5196 | "_openmp_teams_reduction_type_$_", RecordDecl::TagKind::TTK_Union); | |||
5197 | StaticRD->startDefinition(); | |||
5198 | for (const RecordDecl *TeamReductionRec : TeamsReductions) { | |||
5199 | QualType RecTy = C.getRecordType(TeamReductionRec); | |||
5200 | auto *Field = FieldDecl::Create( | |||
5201 | C, StaticRD, SourceLocation(), SourceLocation(), nullptr, RecTy, | |||
5202 | C.getTrivialTypeSourceInfo(RecTy, SourceLocation()), | |||
5203 | /*BW=*/nullptr, /*Mutable=*/false, | |||
5204 | /*InitStyle=*/ICIS_NoInit); | |||
5205 | Field->setAccess(AS_public); | |||
5206 | StaticRD->addDecl(Field); | |||
5207 | } | |||
5208 | StaticRD->completeDefinition(); | |||
5209 | QualType StaticTy = C.getRecordType(StaticRD); | |||
5210 | llvm::Type *LLVMReductionsBufferTy = | |||
5211 | CGM.getTypes().ConvertTypeForMem(StaticTy); | |||
5212 | // FIXME: nvlink does not handle weak linkage correctly (object with the | |||
5213 | // different size are reported as erroneous). | |||
5214 | // Restore CommonLinkage as soon as nvlink is fixed. | |||
5215 | auto *GV = new llvm::GlobalVariable( | |||
5216 | CGM.getModule(), LLVMReductionsBufferTy, | |||
5217 | /*isConstant=*/false, llvm::GlobalValue::InternalLinkage, | |||
5218 | llvm::Constant::getNullValue(LLVMReductionsBufferTy), | |||
5219 | "_openmp_teams_reductions_buffer_$_"); | |||
5220 | KernelTeamsReductionPtr->setInitializer( | |||
5221 | llvm::ConstantExpr::getPointerBitCastOrAddrSpaceCast(GV, | |||
5222 | CGM.VoidPtrTy)); | |||
5223 | } | |||
5224 | CGOpenMPRuntime::clear(); | |||
5225 | } |