Bug Summary

File:clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
Warning:line 4807, column 8
Called C++ object pointer is null

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name CGOpenMPRuntimeNVPTX.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -setup-static-analyzer -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -mthread-model posix -mframe-pointer=none -relaxed-aliasing -fmath-errno -fno-rounding-math -masm-verbose -mconstructor-aliases -munwind-tables -target-cpu x86-64 -dwarf-column-info -fno-split-dwarf-inlining -debugger-tuning=gdb -ffunction-sections -fdata-sections -resource-dir /usr/lib/llvm-10/lib/clang/10.0.0 -D CLANG_VENDOR="Debian " -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/build-llvm/tools/clang/lib/CodeGen -I /build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen -I /build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/clang/include -I /build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/build-llvm/tools/clang/include -I /build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/build-llvm/include -I /build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/include -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0/backward -internal-isystem /usr/local/include -internal-isystem /usr/lib/llvm-10/lib/clang/10.0.0/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-comment -std=c++14 -fdeprecated-macro -fdebug-compilation-dir /build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/build-llvm/tools/clang/lib/CodeGen -fdebug-prefix-map=/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd=. -ferror-limit 19 -fmessage-length 0 -fvisibility-inlines-hidden -stack-protector 2 -fgnuc-version=4.2.1 -fobjc-runtime=gcc -fno-common -fdiagnostics-show-option -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -faddrsig -o /tmp/scan-build-2020-01-13-084841-49055-1 -x c++ /build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
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
24using namespace clang;
25using namespace CodeGen;
26using namespace llvm::omp;
27
28namespace {
29enum 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.
120class 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
128public:
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.
160class ExecutionRuntimeModesRAII {
161private:
162 CGOpenMPRuntimeNVPTX::ExecutionMode SavedExecMode =
163 CGOpenMPRuntimeNVPTX::EM_Unknown;
164 CGOpenMPRuntimeNVPTX::ExecutionMode &ExecMode;
165 bool SavedRuntimeMode = false;
166 bool *RuntimeMode = nullptr;
167
168public:
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.
195enum 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
209static 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
232static 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.
312class 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-10~++20200112100611+7fa5290d5bd/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-10~++20200112100611+7fa5290d5bd/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-10~++20200112100611+7fa5290d5bd/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-10~++20200112100611+7fa5290d5bd/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-10~++20200112100611+7fa5290d5bd/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
442public:
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-10~++20200112100611+7fa5290d5bd/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-10~++20200112100611+7fa5290d5bd/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.
624static 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.
632static 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.
642static 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.
650static 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.
657static 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.
669static 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.
685static 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
696CGOpenMPRuntimeNVPTX::WorkerFunctionState::WorkerFunctionState(
697 CodeGenModule &CGM, SourceLocation Loc)
698 : WorkerFn(nullptr), CGFI(CGM.getTypes().arrangeNullaryFunction()),
699 Loc(Loc) {
700 createWorkerFunction(CGM);
701}
702
703void 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
714CGOpenMPRuntimeNVPTX::ExecutionMode
715CGOpenMPRuntimeNVPTX::getExecutionMode() const {
716 return CurrentExecutionMode;
717}
718
719static CGOpenMPRuntimeNVPTX::DataSharingMode
720getDataSharingMode(CodeGenModule &CGM) {
721 return CGM.getLangOpts().OpenMPCUDAMode ? CGOpenMPRuntimeNVPTX::CUDA
722 : CGOpenMPRuntimeNVPTX::Generic;
723}
724
725/// Check for inner (nested) SPMD construct, if any
726static 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_teams:
790 case OMPD_target_data:
791 case OMPD_target_exit_data:
792 case OMPD_target_enter_data:
793 case OMPD_distribute:
794 case OMPD_distribute_simd:
795 case OMPD_distribute_parallel_for:
796 case OMPD_distribute_parallel_for_simd:
797 case OMPD_teams_distribute:
798 case OMPD_teams_distribute_simd:
799 case OMPD_teams_distribute_parallel_for:
800 case OMPD_teams_distribute_parallel_for_simd:
801 case OMPD_target_update:
802 case OMPD_declare_simd:
803 case OMPD_declare_variant:
804 case OMPD_declare_target:
805 case OMPD_end_declare_target:
806 case OMPD_declare_reduction:
807 case OMPD_declare_mapper:
808 case OMPD_taskloop:
809 case OMPD_taskloop_simd:
810 case OMPD_master_taskloop:
811 case OMPD_master_taskloop_simd:
812 case OMPD_parallel_master_taskloop:
813 case OMPD_parallel_master_taskloop_simd:
814 case OMPD_requires:
815 case OMPD_unknown:
816 llvm_unreachable("Unexpected directive.")::llvm::llvm_unreachable_internal("Unexpected directive.", "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 816)
;
817 }
818 }
819
820 return false;
821}
822
823static bool supportsSPMDExecutionMode(ASTContext &Ctx,
824 const OMPExecutableDirective &D) {
825 OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind();
826 switch (DirectiveKind) {
827 case OMPD_target:
828 case OMPD_target_teams:
829 return hasNestedSPMDDirective(Ctx, D);
830 case OMPD_target_parallel:
831 case OMPD_target_parallel_for:
832 case OMPD_target_parallel_for_simd:
833 case OMPD_target_teams_distribute_parallel_for:
834 case OMPD_target_teams_distribute_parallel_for_simd:
835 case OMPD_target_simd:
836 case OMPD_target_teams_distribute_simd:
837 return true;
838 case OMPD_target_teams_distribute:
839 return false;
840 case OMPD_parallel:
841 case OMPD_for:
842 case OMPD_parallel_for:
843 case OMPD_parallel_master:
844 case OMPD_parallel_sections:
845 case OMPD_for_simd:
846 case OMPD_parallel_for_simd:
847 case OMPD_cancel:
848 case OMPD_cancellation_point:
849 case OMPD_ordered:
850 case OMPD_threadprivate:
851 case OMPD_allocate:
852 case OMPD_task:
853 case OMPD_simd:
854 case OMPD_sections:
855 case OMPD_section:
856 case OMPD_single:
857 case OMPD_master:
858 case OMPD_critical:
859 case OMPD_taskyield:
860 case OMPD_barrier:
861 case OMPD_taskwait:
862 case OMPD_taskgroup:
863 case OMPD_atomic:
864 case OMPD_flush:
865 case OMPD_teams:
866 case OMPD_target_data:
867 case OMPD_target_exit_data:
868 case OMPD_target_enter_data:
869 case OMPD_distribute:
870 case OMPD_distribute_simd:
871 case OMPD_distribute_parallel_for:
872 case OMPD_distribute_parallel_for_simd:
873 case OMPD_teams_distribute:
874 case OMPD_teams_distribute_simd:
875 case OMPD_teams_distribute_parallel_for:
876 case OMPD_teams_distribute_parallel_for_simd:
877 case OMPD_target_update:
878 case OMPD_declare_simd:
879 case OMPD_declare_variant:
880 case OMPD_declare_target:
881 case OMPD_end_declare_target:
882 case OMPD_declare_reduction:
883 case OMPD_declare_mapper:
884 case OMPD_taskloop:
885 case OMPD_taskloop_simd:
886 case OMPD_master_taskloop:
887 case OMPD_master_taskloop_simd:
888 case OMPD_parallel_master_taskloop:
889 case OMPD_parallel_master_taskloop_simd:
890 case OMPD_requires:
891 case OMPD_unknown:
892 break;
893 }
894 llvm_unreachable(::llvm::llvm_unreachable_internal("Unknown programming model for OpenMP directive on NVPTX target."
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 895)
895 "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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 895)
;
896}
897
898/// Check if the directive is loops based and has schedule clause at all or has
899/// static scheduling.
900static bool hasStaticScheduling(const OMPExecutableDirective &D) {
901 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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 903, __PRETTY_FUNCTION__))
902 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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 903, __PRETTY_FUNCTION__))
903 "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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 903, __PRETTY_FUNCTION__))
;
904 return !D.hasClausesOfKind<OMPOrderedClause>() &&
905 (!D.hasClausesOfKind<OMPScheduleClause>() ||
906 llvm::any_of(D.getClausesOfKind<OMPScheduleClause>(),
907 [](const OMPScheduleClause *C) {
908 return C->getScheduleKind() == OMPC_SCHEDULE_static;
909 }));
910}
911
912/// Check for inner (nested) lightweight runtime construct, if any
913static bool hasNestedLightweightDirective(ASTContext &Ctx,
914 const OMPExecutableDirective &D) {
915 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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 915, __PRETTY_FUNCTION__))
;
916 const auto *CS = D.getInnermostCapturedStmt();
917 const auto *Body =
918 CS->getCapturedStmt()->IgnoreContainers(/*IgnoreCaptured=*/true);
919 const Stmt *ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body);
920
921 if (const auto *NestedDir =
922 dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
923 OpenMPDirectiveKind DKind = NestedDir->getDirectiveKind();
924 switch (D.getDirectiveKind()) {
925 case OMPD_target:
926 if (isOpenMPParallelDirective(DKind) &&
927 isOpenMPWorksharingDirective(DKind) && isOpenMPLoopDirective(DKind) &&
928 hasStaticScheduling(*NestedDir))
929 return true;
930 if (DKind == OMPD_teams_distribute_simd || DKind == OMPD_simd)
931 return true;
932 if (DKind == OMPD_parallel) {
933 Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers(
934 /*IgnoreCaptured=*/true);
935 if (!Body)
936 return false;
937 ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body);
938 if (const auto *NND =
939 dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
940 DKind = NND->getDirectiveKind();
941 if (isOpenMPWorksharingDirective(DKind) &&
942 isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NND))
943 return true;
944 }
945 } else if (DKind == OMPD_teams) {
946 Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers(
947 /*IgnoreCaptured=*/true);
948 if (!Body)
949 return false;
950 ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body);
951 if (const auto *NND =
952 dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
953 DKind = NND->getDirectiveKind();
954 if (isOpenMPParallelDirective(DKind) &&
955 isOpenMPWorksharingDirective(DKind) &&
956 isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NND))
957 return true;
958 if (DKind == OMPD_parallel) {
959 Body = NND->getInnermostCapturedStmt()->IgnoreContainers(
960 /*IgnoreCaptured=*/true);
961 if (!Body)
962 return false;
963 ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body);
964 if (const auto *NND =
965 dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
966 DKind = NND->getDirectiveKind();
967 if (isOpenMPWorksharingDirective(DKind) &&
968 isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NND))
969 return true;
970 }
971 }
972 }
973 }
974 return false;
975 case OMPD_target_teams:
976 if (isOpenMPParallelDirective(DKind) &&
977 isOpenMPWorksharingDirective(DKind) && isOpenMPLoopDirective(DKind) &&
978 hasStaticScheduling(*NestedDir))
979 return true;
980 if (DKind == OMPD_distribute_simd || DKind == OMPD_simd)
981 return true;
982 if (DKind == OMPD_parallel) {
983 Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers(
984 /*IgnoreCaptured=*/true);
985 if (!Body)
986 return false;
987 ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body);
988 if (const auto *NND =
989 dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
990 DKind = NND->getDirectiveKind();
991 if (isOpenMPWorksharingDirective(DKind) &&
992 isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NND))
993 return true;
994 }
995 }
996 return false;
997 case OMPD_target_parallel:
998 if (DKind == OMPD_simd)
999 return true;
1000 return isOpenMPWorksharingDirective(DKind) &&
1001 isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NestedDir);
1002 case OMPD_target_teams_distribute:
1003 case OMPD_target_simd:
1004 case OMPD_target_parallel_for:
1005 case OMPD_target_parallel_for_simd:
1006 case OMPD_target_teams_distribute_simd:
1007 case OMPD_target_teams_distribute_parallel_for:
1008 case OMPD_target_teams_distribute_parallel_for_simd:
1009 case OMPD_parallel:
1010 case OMPD_for:
1011 case OMPD_parallel_for:
1012 case OMPD_parallel_master:
1013 case OMPD_parallel_sections:
1014 case OMPD_for_simd:
1015 case OMPD_parallel_for_simd:
1016 case OMPD_cancel:
1017 case OMPD_cancellation_point:
1018 case OMPD_ordered:
1019 case OMPD_threadprivate:
1020 case OMPD_allocate:
1021 case OMPD_task:
1022 case OMPD_simd:
1023 case OMPD_sections:
1024 case OMPD_section:
1025 case OMPD_single:
1026 case OMPD_master:
1027 case OMPD_critical:
1028 case OMPD_taskyield:
1029 case OMPD_barrier:
1030 case OMPD_taskwait:
1031 case OMPD_taskgroup:
1032 case OMPD_atomic:
1033 case OMPD_flush:
1034 case OMPD_teams:
1035 case OMPD_target_data:
1036 case OMPD_target_exit_data:
1037 case OMPD_target_enter_data:
1038 case OMPD_distribute:
1039 case OMPD_distribute_simd:
1040 case OMPD_distribute_parallel_for:
1041 case OMPD_distribute_parallel_for_simd:
1042 case OMPD_teams_distribute:
1043 case OMPD_teams_distribute_simd:
1044 case OMPD_teams_distribute_parallel_for:
1045 case OMPD_teams_distribute_parallel_for_simd:
1046 case OMPD_target_update:
1047 case OMPD_declare_simd:
1048 case OMPD_declare_variant:
1049 case OMPD_declare_target:
1050 case OMPD_end_declare_target:
1051 case OMPD_declare_reduction:
1052 case OMPD_declare_mapper:
1053 case OMPD_taskloop:
1054 case OMPD_taskloop_simd:
1055 case OMPD_master_taskloop:
1056 case OMPD_master_taskloop_simd:
1057 case OMPD_parallel_master_taskloop:
1058 case OMPD_parallel_master_taskloop_simd:
1059 case OMPD_requires:
1060 case OMPD_unknown:
1061 llvm_unreachable("Unexpected directive.")::llvm::llvm_unreachable_internal("Unexpected directive.", "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 1061)
;
1062 }
1063 }
1064
1065 return false;
1066}
1067
1068/// Checks if the construct supports lightweight runtime. It must be SPMD
1069/// construct + inner loop-based construct with static scheduling.
1070static bool supportsLightweightRuntime(ASTContext &Ctx,
1071 const OMPExecutableDirective &D) {
1072 if (!supportsSPMDExecutionMode(Ctx, D))
1073 return false;
1074 OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind();
1075 switch (DirectiveKind) {
1076 case OMPD_target:
1077 case OMPD_target_teams:
1078 case OMPD_target_parallel:
1079 return hasNestedLightweightDirective(Ctx, D);
1080 case OMPD_target_parallel_for:
1081 case OMPD_target_parallel_for_simd:
1082 case OMPD_target_teams_distribute_parallel_for:
1083 case OMPD_target_teams_distribute_parallel_for_simd:
1084 // (Last|First)-privates must be shared in parallel region.
1085 return hasStaticScheduling(D);
1086 case OMPD_target_simd:
1087 case OMPD_target_teams_distribute_simd:
1088 return true;
1089 case OMPD_target_teams_distribute:
1090 return false;
1091 case OMPD_parallel:
1092 case OMPD_for:
1093 case OMPD_parallel_for:
1094 case OMPD_parallel_master:
1095 case OMPD_parallel_sections:
1096 case OMPD_for_simd:
1097 case OMPD_parallel_for_simd:
1098 case OMPD_cancel:
1099 case OMPD_cancellation_point:
1100 case OMPD_ordered:
1101 case OMPD_threadprivate:
1102 case OMPD_allocate:
1103 case OMPD_task:
1104 case OMPD_simd:
1105 case OMPD_sections:
1106 case OMPD_section:
1107 case OMPD_single:
1108 case OMPD_master:
1109 case OMPD_critical:
1110 case OMPD_taskyield:
1111 case OMPD_barrier:
1112 case OMPD_taskwait:
1113 case OMPD_taskgroup:
1114 case OMPD_atomic:
1115 case OMPD_flush:
1116 case OMPD_teams:
1117 case OMPD_target_data:
1118 case OMPD_target_exit_data:
1119 case OMPD_target_enter_data:
1120 case OMPD_distribute:
1121 case OMPD_distribute_simd:
1122 case OMPD_distribute_parallel_for:
1123 case OMPD_distribute_parallel_for_simd:
1124 case OMPD_teams_distribute:
1125 case OMPD_teams_distribute_simd:
1126 case OMPD_teams_distribute_parallel_for:
1127 case OMPD_teams_distribute_parallel_for_simd:
1128 case OMPD_target_update:
1129 case OMPD_declare_simd:
1130 case OMPD_declare_variant:
1131 case OMPD_declare_target:
1132 case OMPD_end_declare_target:
1133 case OMPD_declare_reduction:
1134 case OMPD_declare_mapper:
1135 case OMPD_taskloop:
1136 case OMPD_taskloop_simd:
1137 case OMPD_master_taskloop:
1138 case OMPD_master_taskloop_simd:
1139 case OMPD_parallel_master_taskloop:
1140 case OMPD_parallel_master_taskloop_simd:
1141 case OMPD_requires:
1142 case OMPD_unknown:
1143 break;
1144 }
1145 llvm_unreachable(::llvm::llvm_unreachable_internal("Unknown programming model for OpenMP directive on NVPTX target."
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 1146)
1146 "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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 1146)
;
1147}
1148
1149void CGOpenMPRuntimeNVPTX::emitNonSPMDKernel(const OMPExecutableDirective &D,
1150 StringRef ParentName,
1151 llvm::Function *&OutlinedFn,
1152 llvm::Constant *&OutlinedFnID,
1153 bool IsOffloadEntry,
1154 const RegionCodeGenTy &CodeGen) {
1155 ExecutionRuntimeModesRAII ModeRAII(CurrentExecutionMode);
1156 EntryFunctionState EST;
1157 WorkerFunctionState WST(CGM, D.getBeginLoc());
1158 Work.clear();
1159 WrapperFunctionsMap.clear();
1160
1161 // Emit target region as a standalone region.
1162 class NVPTXPrePostActionTy : public PrePostActionTy {
1163 CGOpenMPRuntimeNVPTX::EntryFunctionState &EST;
1164 CGOpenMPRuntimeNVPTX::WorkerFunctionState &WST;
1165
1166 public:
1167 NVPTXPrePostActionTy(CGOpenMPRuntimeNVPTX::EntryFunctionState &EST,
1168 CGOpenMPRuntimeNVPTX::WorkerFunctionState &WST)
1169 : EST(EST), WST(WST) {}
1170 void Enter(CodeGenFunction &CGF) override {
1171 auto &RT =
1172 static_cast<CGOpenMPRuntimeNVPTX &>(CGF.CGM.getOpenMPRuntime());
1173 RT.emitNonSPMDEntryHeader(CGF, EST, WST);
1174 // Skip target region initialization.
1175 RT.setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true);
1176 }
1177 void Exit(CodeGenFunction &CGF) override {
1178 auto &RT =
1179 static_cast<CGOpenMPRuntimeNVPTX &>(CGF.CGM.getOpenMPRuntime());
1180 RT.clearLocThreadIdInsertPt(CGF);
1181 RT.emitNonSPMDEntryFooter(CGF, EST);
1182 }
1183 } Action(EST, WST);
1184 CodeGen.setAction(Action);
1185 IsInTTDRegion = true;
1186 // Reserve place for the globalized memory.
1187 GlobalizedRecords.emplace_back();
1188 if (!KernelStaticGlobalized) {
1189 KernelStaticGlobalized = new llvm::GlobalVariable(
1190 CGM.getModule(), CGM.VoidPtrTy, /*isConstant=*/false,
1191 llvm::GlobalValue::InternalLinkage,
1192 llvm::ConstantPointerNull::get(CGM.VoidPtrTy),
1193 "_openmp_kernel_static_glob_rd$ptr", /*InsertBefore=*/nullptr,
1194 llvm::GlobalValue::NotThreadLocal,
1195 CGM.getContext().getTargetAddressSpace(LangAS::cuda_shared));
1196 }
1197 emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID,
1198 IsOffloadEntry, CodeGen);
1199 IsInTTDRegion = false;
1200
1201 // Now change the name of the worker function to correspond to this target
1202 // region's entry function.
1203 WST.WorkerFn->setName(Twine(OutlinedFn->getName(), "_worker"));
1204
1205 // Create the worker function
1206 emitWorkerFunction(WST);
1207}
1208
1209// Setup NVPTX threads for master-worker OpenMP scheme.
1210void CGOpenMPRuntimeNVPTX::emitNonSPMDEntryHeader(CodeGenFunction &CGF,
1211 EntryFunctionState &EST,
1212 WorkerFunctionState &WST) {
1213 CGBuilderTy &Bld = CGF.Builder;
1214
1215 llvm::BasicBlock *WorkerBB = CGF.createBasicBlock(".worker");
1216 llvm::BasicBlock *MasterCheckBB = CGF.createBasicBlock(".mastercheck");
1217 llvm::BasicBlock *MasterBB = CGF.createBasicBlock(".master");
1218 EST.ExitBB = CGF.createBasicBlock(".exit");
1219
1220 llvm::Value *IsWorker =
1221 Bld.CreateICmpULT(getNVPTXThreadID(CGF), getThreadLimit(CGF));
1222 Bld.CreateCondBr(IsWorker, WorkerBB, MasterCheckBB);
1223
1224 CGF.EmitBlock(WorkerBB);
1225 emitCall(CGF, WST.Loc, WST.WorkerFn);
1226 CGF.EmitBranch(EST.ExitBB);
1227
1228 CGF.EmitBlock(MasterCheckBB);
1229 llvm::Value *IsMaster =
1230 Bld.CreateICmpEQ(getNVPTXThreadID(CGF), getMasterThreadID(CGF));
1231 Bld.CreateCondBr(IsMaster, MasterBB, EST.ExitBB);
1232
1233 CGF.EmitBlock(MasterBB);
1234 IsInTargetMasterThreadRegion = true;
1235 // SEQUENTIAL (MASTER) REGION START
1236 // First action in sequential region:
1237 // Initialize the state of the OpenMP runtime library on the GPU.
1238 // TODO: Optimize runtime initialization and pass in correct value.
1239 llvm::Value *Args[] = {getThreadLimit(CGF),
1240 Bld.getInt16(/*RequiresOMPRuntime=*/1)};
1241 CGF.EmitRuntimeCall(
1242 createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_kernel_init), Args);
1243
1244 // For data sharing, we need to initialize the stack.
1245 CGF.EmitRuntimeCall(
1246 createNVPTXRuntimeFunction(
1247 OMPRTL_NVPTX__kmpc_data_sharing_init_stack));
1248
1249 emitGenericVarsProlog(CGF, WST.Loc);
1250}
1251
1252void CGOpenMPRuntimeNVPTX::emitNonSPMDEntryFooter(CodeGenFunction &CGF,
1253 EntryFunctionState &EST) {
1254 IsInTargetMasterThreadRegion = false;
1255 if (!CGF.HaveInsertPoint())
1256 return;
1257
1258 emitGenericVarsEpilog(CGF);
1259
1260 if (!EST.ExitBB)
1261 EST.ExitBB = CGF.createBasicBlock(".exit");
1262
1263 llvm::BasicBlock *TerminateBB = CGF.createBasicBlock(".termination.notifier");
1264 CGF.EmitBranch(TerminateBB);
1265
1266 CGF.EmitBlock(TerminateBB);
1267 // Signal termination condition.
1268 // TODO: Optimize runtime initialization and pass in correct value.
1269 llvm::Value *Args[] = {CGF.Builder.getInt16(/*IsOMPRuntimeInitialized=*/1)};
1270 CGF.EmitRuntimeCall(
1271 createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_kernel_deinit), Args);
1272 // Barrier to terminate worker threads.
1273 syncCTAThreads(CGF);
1274 // Master thread jumps to exit point.
1275 CGF.EmitBranch(EST.ExitBB);
1276
1277 CGF.EmitBlock(EST.ExitBB);
1278 EST.ExitBB = nullptr;
1279}
1280
1281void CGOpenMPRuntimeNVPTX::emitSPMDKernel(const OMPExecutableDirective &D,
1282 StringRef ParentName,
1283 llvm::Function *&OutlinedFn,
1284 llvm::Constant *&OutlinedFnID,
1285 bool IsOffloadEntry,
1286 const RegionCodeGenTy &CodeGen) {
1287 ExecutionRuntimeModesRAII ModeRAII(
1288 CurrentExecutionMode, RequiresFullRuntime,
1289 CGM.getLangOpts().OpenMPCUDAForceFullRuntime ||
1290 !supportsLightweightRuntime(CGM.getContext(), D));
1291 EntryFunctionState EST;
1292
1293 // Emit target region as a standalone region.
1294 class NVPTXPrePostActionTy : public PrePostActionTy {
1295 CGOpenMPRuntimeNVPTX &RT;
1296 CGOpenMPRuntimeNVPTX::EntryFunctionState &EST;
1297 const OMPExecutableDirective &D;
1298
1299 public:
1300 NVPTXPrePostActionTy(CGOpenMPRuntimeNVPTX &RT,
1301 CGOpenMPRuntimeNVPTX::EntryFunctionState &EST,
1302 const OMPExecutableDirective &D)
1303 : RT(RT), EST(EST), D(D) {}
1304 void Enter(CodeGenFunction &CGF) override {
1305 RT.emitSPMDEntryHeader(CGF, EST, D);
1306 // Skip target region initialization.
1307 RT.setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true);
1308 }
1309 void Exit(CodeGenFunction &CGF) override {
1310 RT.clearLocThreadIdInsertPt(CGF);
1311 RT.emitSPMDEntryFooter(CGF, EST);
1312 }
1313 } Action(*this, EST, D);
1314 CodeGen.setAction(Action);
1315 IsInTTDRegion = true;
1316 // Reserve place for the globalized memory.
1317 GlobalizedRecords.emplace_back();
1318 if (!KernelStaticGlobalized) {
1319 KernelStaticGlobalized = new llvm::GlobalVariable(
1320 CGM.getModule(), CGM.VoidPtrTy, /*isConstant=*/false,
1321 llvm::GlobalValue::InternalLinkage,
1322 llvm::ConstantPointerNull::get(CGM.VoidPtrTy),
1323 "_openmp_kernel_static_glob_rd$ptr", /*InsertBefore=*/nullptr,
1324 llvm::GlobalValue::NotThreadLocal,
1325 CGM.getContext().getTargetAddressSpace(LangAS::cuda_shared));
1326 }
1327 emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID,
1328 IsOffloadEntry, CodeGen);
1329 IsInTTDRegion = false;
1330}
1331
1332void CGOpenMPRuntimeNVPTX::emitSPMDEntryHeader(
1333 CodeGenFunction &CGF, EntryFunctionState &EST,
1334 const OMPExecutableDirective &D) {
1335 CGBuilderTy &Bld = CGF.Builder;
1336
1337 // Setup BBs in entry function.
1338 llvm::BasicBlock *ExecuteBB = CGF.createBasicBlock(".execute");
1339 EST.ExitBB = CGF.createBasicBlock(".exit");
1340
1341 llvm::Value *Args[] = {getThreadLimit(CGF, /*IsInSPMDExecutionMode=*/true),
1342 /*RequiresOMPRuntime=*/
1343 Bld.getInt16(RequiresFullRuntime ? 1 : 0),
1344 /*RequiresDataSharing=*/Bld.getInt16(0)};
1345 CGF.EmitRuntimeCall(
1346 createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_spmd_kernel_init), Args);
1347
1348 if (RequiresFullRuntime) {
1349 // For data sharing, we need to initialize the stack.
1350 CGF.EmitRuntimeCall(createNVPTXRuntimeFunction(
1351 OMPRTL_NVPTX__kmpc_data_sharing_init_stack_spmd));
1352 }
1353
1354 CGF.EmitBranch(ExecuteBB);
1355
1356 CGF.EmitBlock(ExecuteBB);
1357
1358 IsInTargetMasterThreadRegion = true;
1359}
1360
1361void CGOpenMPRuntimeNVPTX::emitSPMDEntryFooter(CodeGenFunction &CGF,
1362 EntryFunctionState &EST) {
1363 IsInTargetMasterThreadRegion = false;
1364 if (!CGF.HaveInsertPoint())
1365 return;
1366
1367 if (!EST.ExitBB)
1368 EST.ExitBB = CGF.createBasicBlock(".exit");
1369
1370 llvm::BasicBlock *OMPDeInitBB = CGF.createBasicBlock(".omp.deinit");
1371 CGF.EmitBranch(OMPDeInitBB);
1372
1373 CGF.EmitBlock(OMPDeInitBB);
1374 // DeInitialize the OMP state in the runtime; called by all active threads.
1375 llvm::Value *Args[] = {/*RequiresOMPRuntime=*/
1376 CGF.Builder.getInt16(RequiresFullRuntime ? 1 : 0)};
1377 CGF.EmitRuntimeCall(
1378 createNVPTXRuntimeFunction(
1379 OMPRTL_NVPTX__kmpc_spmd_kernel_deinit_v2), Args);
1380 CGF.EmitBranch(EST.ExitBB);
1381
1382 CGF.EmitBlock(EST.ExitBB);
1383 EST.ExitBB = nullptr;
1384}
1385
1386// Create a unique global variable to indicate the execution mode of this target
1387// region. The execution mode is either 'generic', or 'spmd' depending on the
1388// target directive. This variable is picked up by the offload library to setup
1389// the device appropriately before kernel launch. If the execution mode is
1390// 'generic', the runtime reserves one warp for the master, otherwise, all
1391// warps participate in parallel work.
1392static void setPropertyExecutionMode(CodeGenModule &CGM, StringRef Name,
1393 bool Mode) {
1394 auto *GVMode =
1395 new llvm::GlobalVariable(CGM.getModule(), CGM.Int8Ty, /*isConstant=*/true,
1396 llvm::GlobalValue::WeakAnyLinkage,
1397 llvm::ConstantInt::get(CGM.Int8Ty, Mode ? 0 : 1),
1398 Twine(Name, "_exec_mode"));
1399 CGM.addCompilerUsedGlobal(GVMode);
1400}
1401
1402void CGOpenMPRuntimeNVPTX::emitWorkerFunction(WorkerFunctionState &WST) {
1403 ASTContext &Ctx = CGM.getContext();
1404
1405 CodeGenFunction CGF(CGM, /*suppressNewContext=*/true);
1406 CGF.StartFunction(GlobalDecl(), Ctx.VoidTy, WST.WorkerFn, WST.CGFI, {},
1407 WST.Loc, WST.Loc);
1408 emitWorkerLoop(CGF, WST);
1409 CGF.FinishFunction();
1410}
1411
1412void CGOpenMPRuntimeNVPTX::emitWorkerLoop(CodeGenFunction &CGF,
1413 WorkerFunctionState &WST) {
1414 //
1415 // The workers enter this loop and wait for parallel work from the master.
1416 // When the master encounters a parallel region it sets up the work + variable
1417 // arguments, and wakes up the workers. The workers first check to see if
1418 // they are required for the parallel region, i.e., within the # of requested
1419 // parallel threads. The activated workers load the variable arguments and
1420 // execute the parallel work.
1421 //
1422
1423 CGBuilderTy &Bld = CGF.Builder;
1424
1425 llvm::BasicBlock *AwaitBB = CGF.createBasicBlock(".await.work");
1426 llvm::BasicBlock *SelectWorkersBB = CGF.createBasicBlock(".select.workers");
1427 llvm::BasicBlock *ExecuteBB = CGF.createBasicBlock(".execute.parallel");
1428 llvm::BasicBlock *TerminateBB = CGF.createBasicBlock(".terminate.parallel");
1429 llvm::BasicBlock *BarrierBB = CGF.createBasicBlock(".barrier.parallel");
1430 llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".exit");
1431
1432 CGF.EmitBranch(AwaitBB);
1433
1434 // Workers wait for work from master.
1435 CGF.EmitBlock(AwaitBB);
1436 // Wait for parallel work
1437 syncCTAThreads(CGF);
1438
1439 Address WorkFn =
1440 CGF.CreateDefaultAlignTempAlloca(CGF.Int8PtrTy, /*Name=*/"work_fn");
1441 Address ExecStatus =
1442 CGF.CreateDefaultAlignTempAlloca(CGF.Int8Ty, /*Name=*/"exec_status");
1443 CGF.InitTempAlloca(ExecStatus, Bld.getInt8(/*C=*/0));
1444 CGF.InitTempAlloca(WorkFn, llvm::Constant::getNullValue(CGF.Int8PtrTy));
1445
1446 // TODO: Optimize runtime initialization and pass in correct value.
1447 llvm::Value *Args[] = {WorkFn.getPointer(),
1448 /*RequiresOMPRuntime=*/Bld.getInt16(1)};
1449 llvm::Value *Ret = CGF.EmitRuntimeCall(
1450 createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_kernel_parallel), Args);
1451 Bld.CreateStore(Bld.CreateZExt(Ret, CGF.Int8Ty), ExecStatus);
1452
1453 // On termination condition (workid == 0), exit loop.
1454 llvm::Value *WorkID = Bld.CreateLoad(WorkFn);
1455 llvm::Value *ShouldTerminate = Bld.CreateIsNull(WorkID, "should_terminate");
1456 Bld.CreateCondBr(ShouldTerminate, ExitBB, SelectWorkersBB);
1457
1458 // Activate requested workers.
1459 CGF.EmitBlock(SelectWorkersBB);
1460 llvm::Value *IsActive =
1461 Bld.CreateIsNotNull(Bld.CreateLoad(ExecStatus), "is_active");
1462 Bld.CreateCondBr(IsActive, ExecuteBB, BarrierBB);
1463
1464 // Signal start of parallel region.
1465 CGF.EmitBlock(ExecuteBB);
1466 // Skip initialization.
1467 setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true);
1468
1469 // Process work items: outlined parallel functions.
1470 for (llvm::Function *W : Work) {
1471 // Try to match this outlined function.
1472 llvm::Value *ID = Bld.CreatePointerBitCastOrAddrSpaceCast(W, CGM.Int8PtrTy);
1473
1474 llvm::Value *WorkFnMatch =
1475 Bld.CreateICmpEQ(Bld.CreateLoad(WorkFn), ID, "work_match");
1476
1477 llvm::BasicBlock *ExecuteFNBB = CGF.createBasicBlock(".execute.fn");
1478 llvm::BasicBlock *CheckNextBB = CGF.createBasicBlock(".check.next");
1479 Bld.CreateCondBr(WorkFnMatch, ExecuteFNBB, CheckNextBB);
1480
1481 // Execute this outlined function.
1482 CGF.EmitBlock(ExecuteFNBB);
1483
1484 // Insert call to work function via shared wrapper. The shared
1485 // wrapper takes two arguments:
1486 // - the parallelism level;
1487 // - the thread ID;
1488 emitCall(CGF, WST.Loc, W,
1489 {Bld.getInt16(/*ParallelLevel=*/0), getThreadID(CGF, WST.Loc)});
1490
1491 // Go to end of parallel region.
1492 CGF.EmitBranch(TerminateBB);
1493
1494 CGF.EmitBlock(CheckNextBB);
1495 }
1496 // Default case: call to outlined function through pointer if the target
1497 // region makes a declare target call that may contain an orphaned parallel
1498 // directive.
1499 auto *ParallelFnTy =
1500 llvm::FunctionType::get(CGM.VoidTy, {CGM.Int16Ty, CGM.Int32Ty},
1501 /*isVarArg=*/false);
1502 llvm::Value *WorkFnCast =
1503 Bld.CreateBitCast(WorkID, ParallelFnTy->getPointerTo());
1504 // Insert call to work function via shared wrapper. The shared
1505 // wrapper takes two arguments:
1506 // - the parallelism level;
1507 // - the thread ID;
1508 emitCall(CGF, WST.Loc, {ParallelFnTy, WorkFnCast},
1509 {Bld.getInt16(/*ParallelLevel=*/0), getThreadID(CGF, WST.Loc)});
1510 // Go to end of parallel region.
1511 CGF.EmitBranch(TerminateBB);
1512
1513 // Signal end of parallel region.
1514 CGF.EmitBlock(TerminateBB);
1515 CGF.EmitRuntimeCall(
1516 createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_kernel_end_parallel),
1517 llvm::None);
1518 CGF.EmitBranch(BarrierBB);
1519
1520 // All active and inactive workers wait at a barrier after parallel region.
1521 CGF.EmitBlock(BarrierBB);
1522 // Barrier after parallel region.
1523 syncCTAThreads(CGF);
1524 CGF.EmitBranch(AwaitBB);
1525
1526 // Exit target region.
1527 CGF.EmitBlock(ExitBB);
1528 // Skip initialization.
1529 clearLocThreadIdInsertPt(CGF);
1530}
1531
1532/// Returns specified OpenMP runtime function for the current OpenMP
1533/// implementation. Specialized for the NVPTX device.
1534/// \param Function OpenMP runtime function.
1535/// \return Specified function.
1536llvm::FunctionCallee
1537CGOpenMPRuntimeNVPTX::createNVPTXRuntimeFunction(unsigned Function) {
1538 llvm::FunctionCallee RTLFn = nullptr;
1539 switch (static_cast<OpenMPRTLFunctionNVPTX>(Function)) {
1540 case OMPRTL_NVPTX__kmpc_kernel_init: {
1541 // Build void __kmpc_kernel_init(kmp_int32 thread_limit, int16_t
1542 // RequiresOMPRuntime);
1543 llvm::Type *TypeParams[] = {CGM.Int32Ty, CGM.Int16Ty};
1544 auto *FnTy =
1545 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
1546 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_kernel_init");
1547 break;
1548 }
1549 case OMPRTL_NVPTX__kmpc_kernel_deinit: {
1550 // Build void __kmpc_kernel_deinit(int16_t IsOMPRuntimeInitialized);
1551 llvm::Type *TypeParams[] = {CGM.Int16Ty};
1552 auto *FnTy =
1553 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
1554 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_kernel_deinit");
1555 break;
1556 }
1557 case OMPRTL_NVPTX__kmpc_spmd_kernel_init: {
1558 // Build void __kmpc_spmd_kernel_init(kmp_int32 thread_limit,
1559 // int16_t RequiresOMPRuntime, int16_t RequiresDataSharing);
1560 llvm::Type *TypeParams[] = {CGM.Int32Ty, CGM.Int16Ty, CGM.Int16Ty};
1561 auto *FnTy =
1562 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
1563 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_spmd_kernel_init");
1564 break;
1565 }
1566 case OMPRTL_NVPTX__kmpc_spmd_kernel_deinit_v2: {
1567 // Build void __kmpc_spmd_kernel_deinit_v2(int16_t RequiresOMPRuntime);
1568 llvm::Type *TypeParams[] = {CGM.Int16Ty};
1569 auto *FnTy =
1570 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
1571 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_spmd_kernel_deinit_v2");
1572 break;
1573 }
1574 case OMPRTL_NVPTX__kmpc_kernel_prepare_parallel: {
1575 /// Build void __kmpc_kernel_prepare_parallel(
1576 /// void *outlined_function, int16_t IsOMPRuntimeInitialized);
1577 llvm::Type *TypeParams[] = {CGM.Int8PtrTy, CGM.Int16Ty};
1578 auto *FnTy =
1579 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
1580 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_kernel_prepare_parallel");
1581 break;
1582 }
1583 case OMPRTL_NVPTX__kmpc_kernel_parallel: {
1584 /// Build bool __kmpc_kernel_parallel(void **outlined_function,
1585 /// int16_t IsOMPRuntimeInitialized);
1586 llvm::Type *TypeParams[] = {CGM.Int8PtrPtrTy, CGM.Int16Ty};
1587 llvm::Type *RetTy = CGM.getTypes().ConvertType(CGM.getContext().BoolTy);
1588 auto *FnTy =
1589 llvm::FunctionType::get(RetTy, TypeParams, /*isVarArg*/ false);
1590 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_kernel_parallel");
1591 break;
1592 }
1593 case OMPRTL_NVPTX__kmpc_kernel_end_parallel: {
1594 /// Build void __kmpc_kernel_end_parallel();
1595 auto *FnTy =
1596 llvm::FunctionType::get(CGM.VoidTy, llvm::None, /*isVarArg*/ false);
1597 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_kernel_end_parallel");
1598 break;
1599 }
1600 case OMPRTL_NVPTX__kmpc_serialized_parallel: {
1601 // Build void __kmpc_serialized_parallel(ident_t *loc, kmp_int32
1602 // global_tid);
1603 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty};
1604 auto *FnTy =
1605 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
1606 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_serialized_parallel");
1607 break;
1608 }
1609 case OMPRTL_NVPTX__kmpc_end_serialized_parallel: {
1610 // Build void __kmpc_end_serialized_parallel(ident_t *loc, kmp_int32
1611 // global_tid);
1612 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty};
1613 auto *FnTy =
1614 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
1615 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_end_serialized_parallel");
1616 break;
1617 }
1618 case OMPRTL_NVPTX__kmpc_shuffle_int32: {
1619 // Build int32_t __kmpc_shuffle_int32(int32_t element,
1620 // int16_t lane_offset, int16_t warp_size);
1621 llvm::Type *TypeParams[] = {CGM.Int32Ty, CGM.Int16Ty, CGM.Int16Ty};
1622 auto *FnTy =
1623 llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg*/ false);
1624 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_shuffle_int32");
1625 break;
1626 }
1627 case OMPRTL_NVPTX__kmpc_shuffle_int64: {
1628 // Build int64_t __kmpc_shuffle_int64(int64_t element,
1629 // int16_t lane_offset, int16_t warp_size);
1630 llvm::Type *TypeParams[] = {CGM.Int64Ty, CGM.Int16Ty, CGM.Int16Ty};
1631 auto *FnTy =
1632 llvm::FunctionType::get(CGM.Int64Ty, TypeParams, /*isVarArg*/ false);
1633 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_shuffle_int64");
1634 break;
1635 }
1636 case OMPRTL_NVPTX__kmpc_nvptx_parallel_reduce_nowait_v2: {
1637 // Build int32_t kmpc_nvptx_parallel_reduce_nowait_v2(ident_t *loc,
1638 // kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size, void*
1639 // reduce_data, void (*kmp_ShuffleReductFctPtr)(void *rhsData, int16_t
1640 // lane_id, int16_t lane_offset, int16_t Algorithm Version), void
1641 // (*kmp_InterWarpCopyFctPtr)(void* src, int warp_num));
1642 llvm::Type *ShuffleReduceTypeParams[] = {CGM.VoidPtrTy, CGM.Int16Ty,
1643 CGM.Int16Ty, CGM.Int16Ty};
1644 auto *ShuffleReduceFnTy =
1645 llvm::FunctionType::get(CGM.VoidTy, ShuffleReduceTypeParams,
1646 /*isVarArg=*/false);
1647 llvm::Type *InterWarpCopyTypeParams[] = {CGM.VoidPtrTy, CGM.Int32Ty};
1648 auto *InterWarpCopyFnTy =
1649 llvm::FunctionType::get(CGM.VoidTy, InterWarpCopyTypeParams,
1650 /*isVarArg=*/false);
1651 llvm::Type *TypeParams[] = {getIdentTyPointerTy(),
1652 CGM.Int32Ty,
1653 CGM.Int32Ty,
1654 CGM.SizeTy,
1655 CGM.VoidPtrTy,
1656 ShuffleReduceFnTy->getPointerTo(),
1657 InterWarpCopyFnTy->getPointerTo()};
1658 auto *FnTy =
1659 llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg=*/false);
1660 RTLFn = CGM.CreateRuntimeFunction(
1661 FnTy, /*Name=*/"__kmpc_nvptx_parallel_reduce_nowait_v2");
1662 break;
1663 }
1664 case OMPRTL_NVPTX__kmpc_end_reduce_nowait: {
1665 // Build __kmpc_end_reduce_nowait(kmp_int32 global_tid);
1666 llvm::Type *TypeParams[] = {CGM.Int32Ty};
1667 auto *FnTy =
1668 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false);
1669 RTLFn = CGM.CreateRuntimeFunction(
1670 FnTy, /*Name=*/"__kmpc_nvptx_end_reduce_nowait");
1671 break;
1672 }
1673 case OMPRTL_NVPTX__kmpc_nvptx_teams_reduce_nowait_v2: {
1674 // Build int32_t __kmpc_nvptx_teams_reduce_nowait_v2(ident_t *loc, kmp_int32
1675 // global_tid, void *global_buffer, int32_t num_of_records, void*
1676 // reduce_data,
1677 // void (*kmp_ShuffleReductFctPtr)(void *rhsData, int16_t lane_id, int16_t
1678 // lane_offset, int16_t shortCircuit),
1679 // void (*kmp_InterWarpCopyFctPtr)(void* src, int32_t warp_num), void
1680 // (*kmp_ListToGlobalCpyFctPtr)(void *buffer, int idx, void *reduce_data),
1681 // void (*kmp_GlobalToListCpyFctPtr)(void *buffer, int idx,
1682 // void *reduce_data), void (*kmp_GlobalToListCpyPtrsFctPtr)(void *buffer,
1683 // int idx, void *reduce_data), void (*kmp_GlobalToListRedFctPtr)(void
1684 // *buffer, int idx, void *reduce_data));
1685 llvm::Type *ShuffleReduceTypeParams[] = {CGM.VoidPtrTy, CGM.Int16Ty,
1686 CGM.Int16Ty, CGM.Int16Ty};
1687 auto *ShuffleReduceFnTy =
1688 llvm::FunctionType::get(CGM.VoidTy, ShuffleReduceTypeParams,
1689 /*isVarArg=*/false);
1690 llvm::Type *InterWarpCopyTypeParams[] = {CGM.VoidPtrTy, CGM.Int32Ty};
1691 auto *InterWarpCopyFnTy =
1692 llvm::FunctionType::get(CGM.VoidTy, InterWarpCopyTypeParams,
1693 /*isVarArg=*/false);
1694 llvm::Type *GlobalListTypeParams[] = {CGM.VoidPtrTy, CGM.IntTy,
1695 CGM.VoidPtrTy};
1696 auto *GlobalListFnTy =
1697 llvm::FunctionType::get(CGM.VoidTy, GlobalListTypeParams,
1698 /*isVarArg=*/false);
1699 llvm::Type *TypeParams[] = {getIdentTyPointerTy(),
1700 CGM.Int32Ty,
1701 CGM.VoidPtrTy,
1702 CGM.Int32Ty,
1703 CGM.VoidPtrTy,
1704 ShuffleReduceFnTy->getPointerTo(),
1705 InterWarpCopyFnTy->getPointerTo(),
1706 GlobalListFnTy->getPointerTo(),
1707 GlobalListFnTy->getPointerTo(),
1708 GlobalListFnTy->getPointerTo(),
1709 GlobalListFnTy->getPointerTo()};
1710 auto *FnTy =
1711 llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg=*/false);
1712 RTLFn = CGM.CreateRuntimeFunction(
1713 FnTy, /*Name=*/"__kmpc_nvptx_teams_reduce_nowait_v2");
1714 break;
1715 }
1716 case OMPRTL_NVPTX__kmpc_data_sharing_init_stack: {
1717 /// Build void __kmpc_data_sharing_init_stack();
1718 auto *FnTy =
1719 llvm::FunctionType::get(CGM.VoidTy, llvm::None, /*isVarArg*/ false);
1720 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_data_sharing_init_stack");
1721 break;
1722 }
1723 case OMPRTL_NVPTX__kmpc_data_sharing_init_stack_spmd: {
1724 /// Build void __kmpc_data_sharing_init_stack_spmd();
1725 auto *FnTy =
1726 llvm::FunctionType::get(CGM.VoidTy, llvm::None, /*isVarArg*/ false);
1727 RTLFn =
1728 CGM.CreateRuntimeFunction(FnTy, "__kmpc_data_sharing_init_stack_spmd");
1729 break;
1730 }
1731 case OMPRTL_NVPTX__kmpc_data_sharing_coalesced_push_stack: {
1732 // Build void *__kmpc_data_sharing_coalesced_push_stack(size_t size,
1733 // int16_t UseSharedMemory);
1734 llvm::Type *TypeParams[] = {CGM.SizeTy, CGM.Int16Ty};
1735 auto *FnTy =
1736 llvm::FunctionType::get(CGM.VoidPtrTy, TypeParams, /*isVarArg=*/false);
1737 RTLFn = CGM.CreateRuntimeFunction(
1738 FnTy, /*Name=*/"__kmpc_data_sharing_coalesced_push_stack");
1739 break;
1740 }
1741 case OMPRTL_NVPTX__kmpc_data_sharing_pop_stack: {
1742 // Build void __kmpc_data_sharing_pop_stack(void *a);
1743 llvm::Type *TypeParams[] = {CGM.VoidPtrTy};
1744 auto *FnTy =
1745 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false);
1746 RTLFn = CGM.CreateRuntimeFunction(FnTy,
1747 /*Name=*/"__kmpc_data_sharing_pop_stack");
1748 break;
1749 }
1750 case OMPRTL_NVPTX__kmpc_begin_sharing_variables: {
1751 /// Build void __kmpc_begin_sharing_variables(void ***args,
1752 /// size_t n_args);
1753 llvm::Type *TypeParams[] = {CGM.Int8PtrPtrTy->getPointerTo(), CGM.SizeTy};
1754 auto *FnTy =
1755 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
1756 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_begin_sharing_variables");
1757 break;
1758 }
1759 case OMPRTL_NVPTX__kmpc_end_sharing_variables: {
1760 /// Build void __kmpc_end_sharing_variables();
1761 auto *FnTy =
1762 llvm::FunctionType::get(CGM.VoidTy, llvm::None, /*isVarArg*/ false);
1763 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_end_sharing_variables");
1764 break;
1765 }
1766 case OMPRTL_NVPTX__kmpc_get_shared_variables: {
1767 /// Build void __kmpc_get_shared_variables(void ***GlobalArgs);
1768 llvm::Type *TypeParams[] = {CGM.Int8PtrPtrTy->getPointerTo()};
1769 auto *FnTy =
1770 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
1771 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_get_shared_variables");
1772 break;
1773 }
1774 case OMPRTL_NVPTX__kmpc_parallel_level: {
1775 // Build uint16_t __kmpc_parallel_level(ident_t *loc, kmp_int32 global_tid);
1776 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty};
1777 auto *FnTy =
1778 llvm::FunctionType::get(CGM.Int16Ty, TypeParams, /*isVarArg*/ false);
1779 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_parallel_level");
1780 break;
1781 }
1782 case OMPRTL_NVPTX__kmpc_is_spmd_exec_mode: {
1783 // Build int8_t __kmpc_is_spmd_exec_mode();
1784 auto *FnTy = llvm::FunctionType::get(CGM.Int8Ty, /*isVarArg=*/false);
1785 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_is_spmd_exec_mode");
1786 break;
1787 }
1788 case OMPRTL_NVPTX__kmpc_get_team_static_memory: {
1789 // Build void __kmpc_get_team_static_memory(int16_t isSPMDExecutionMode,
1790 // const void *buf, size_t size, int16_t is_shared, const void **res);
1791 llvm::Type *TypeParams[] = {CGM.Int16Ty, CGM.VoidPtrTy, CGM.SizeTy,
1792 CGM.Int16Ty, CGM.VoidPtrPtrTy};
1793 auto *FnTy =
1794 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
1795 RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_get_team_static_memory");
1796 break;
1797 }
1798 case OMPRTL_NVPTX__kmpc_restore_team_static_memory: {
1799 // Build void __kmpc_restore_team_static_memory(int16_t isSPMDExecutionMode,
1800 // int16_t is_shared);
1801 llvm::Type *TypeParams[] = {CGM.Int16Ty, CGM.Int16Ty};
1802 auto *FnTy =
1803 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false);
1804 RTLFn =
1805 CGM.CreateRuntimeFunction(FnTy, "__kmpc_restore_team_static_memory");
1806 break;
1807 }
1808 case OMPRTL__kmpc_barrier: {
1809 // Build void __kmpc_barrier(ident_t *loc, kmp_int32 global_tid);
1810 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty};
1811 auto *FnTy =
1812 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
1813 RTLFn =
1814 CGM.CreateConvergentRuntimeFunction(FnTy, /*Name*/ "__kmpc_barrier");
1815 break;
1816 }
1817 case OMPRTL__kmpc_barrier_simple_spmd: {
1818 // Build void __kmpc_barrier_simple_spmd(ident_t *loc, kmp_int32
1819 // global_tid);
1820 llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty};
1821 auto *FnTy =
1822 llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
1823 RTLFn = CGM.CreateConvergentRuntimeFunction(
1824 FnTy, /*Name*/ "__kmpc_barrier_simple_spmd");
1825 break;
1826 }
1827 case OMPRTL_NVPTX__kmpc_warp_active_thread_mask: {
1828 // Build int32_t __kmpc_warp_active_thread_mask(void);
1829 auto *FnTy =
1830 llvm::FunctionType::get(CGM.Int32Ty, llvm::None, /*isVarArg=*/false);
1831 RTLFn = CGM.CreateConvergentRuntimeFunction(FnTy, "__kmpc_warp_active_thread_mask");
1832 break;
1833 }
1834 case OMPRTL_NVPTX__kmpc_syncwarp: {
1835 // Build void __kmpc_syncwarp(kmp_int32 Mask);
1836 auto *FnTy =
1837 llvm::FunctionType::get(CGM.VoidTy, CGM.Int32Ty, /*isVarArg=*/false);
1838 RTLFn = CGM.CreateConvergentRuntimeFunction(FnTy, "__kmpc_syncwarp");
1839 break;
1840 }
1841 }
1842 return RTLFn;
1843}
1844
1845void CGOpenMPRuntimeNVPTX::createOffloadEntry(llvm::Constant *ID,
1846 llvm::Constant *Addr,
1847 uint64_t Size, int32_t,
1848 llvm::GlobalValue::LinkageTypes) {
1849 // TODO: Add support for global variables on the device after declare target
1850 // support.
1851 if (!isa<llvm::Function>(Addr))
1852 return;
1853 llvm::Module &M = CGM.getModule();
1854 llvm::LLVMContext &Ctx = CGM.getLLVMContext();
1855
1856 // Get "nvvm.annotations" metadata node
1857 llvm::NamedMDNode *MD = M.getOrInsertNamedMetadata("nvvm.annotations");
1858
1859 llvm::Metadata *MDVals[] = {
1860 llvm::ConstantAsMetadata::get(Addr), llvm::MDString::get(Ctx, "kernel"),
1861 llvm::ConstantAsMetadata::get(
1862 llvm::ConstantInt::get(llvm::Type::getInt32Ty(Ctx), 1))};
1863 // Append metadata to nvvm.annotations
1864 MD->addOperand(llvm::MDNode::get(Ctx, MDVals));
1865}
1866
1867void CGOpenMPRuntimeNVPTX::emitTargetOutlinedFunction(
1868 const OMPExecutableDirective &D, StringRef ParentName,
1869 llvm::Function *&OutlinedFn, llvm::Constant *&OutlinedFnID,
1870 bool IsOffloadEntry, const RegionCodeGenTy &CodeGen) {
1871 if (!IsOffloadEntry) // Nothing to do.
1872 return;
1873
1874 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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 1874, __PRETTY_FUNCTION__))
;
1875
1876 bool Mode = supportsSPMDExecutionMode(CGM.getContext(), D);
1877 if (Mode)
1878 emitSPMDKernel(D, ParentName, OutlinedFn, OutlinedFnID, IsOffloadEntry,
1879 CodeGen);
1880 else
1881 emitNonSPMDKernel(D, ParentName, OutlinedFn, OutlinedFnID, IsOffloadEntry,
1882 CodeGen);
1883
1884 setPropertyExecutionMode(CGM, OutlinedFn->getName(), Mode);
1885}
1886
1887namespace {
1888LLVM_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^=
;
1889/// Enum for accesseing the reserved_2 field of the ident_t struct.
1890enum ModeFlagsTy : unsigned {
1891 /// Bit set to 1 when in SPMD mode.
1892 KMP_IDENT_SPMD_MODE = 0x01,
1893 /// Bit set to 1 when a simplified runtime is used.
1894 KMP_IDENT_SIMPLE_RT_MODE = 0x02,
1895 LLVM_MARK_AS_BITMASK_ENUM(/*LargestValue=*/KMP_IDENT_SIMPLE_RT_MODE)LLVM_BITMASK_LARGEST_ENUMERATOR = KMP_IDENT_SIMPLE_RT_MODE
1896};
1897
1898/// Special mode Undefined. Is the combination of Non-SPMD mode + SimpleRuntime.
1899static const ModeFlagsTy UndefinedMode =
1900 (~KMP_IDENT_SPMD_MODE) & KMP_IDENT_SIMPLE_RT_MODE;
1901} // anonymous namespace
1902
1903unsigned CGOpenMPRuntimeNVPTX::getDefaultLocationReserved2Flags() const {
1904 switch (getExecutionMode()) {
1905 case EM_SPMD:
1906 if (requiresFullRuntime())
1907 return KMP_IDENT_SPMD_MODE & (~KMP_IDENT_SIMPLE_RT_MODE);
1908 return KMP_IDENT_SPMD_MODE | KMP_IDENT_SIMPLE_RT_MODE;
1909 case EM_NonSPMD:
1910 assert(requiresFullRuntime() && "Expected full runtime.")((requiresFullRuntime() && "Expected full runtime.") ?
static_cast<void> (0) : __assert_fail ("requiresFullRuntime() && \"Expected full runtime.\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 1910, __PRETTY_FUNCTION__))
;
1911 return (~KMP_IDENT_SPMD_MODE) & (~KMP_IDENT_SIMPLE_RT_MODE);
1912 case EM_Unknown:
1913 return UndefinedMode;
1914 }
1915 llvm_unreachable("Unknown flags are requested.")::llvm::llvm_unreachable_internal("Unknown flags are requested."
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 1915)
;
1916}
1917
1918bool CGOpenMPRuntimeNVPTX::tryEmitDeclareVariant(const GlobalDecl &NewGD,
1919 const GlobalDecl &OldGD,
1920 llvm::GlobalValue *OrigAddr,
1921 bool IsForDefinition) {
1922 // Emit the function in OldGD with the body from NewGD, if NewGD is defined.
1923 auto *NewFD = cast<FunctionDecl>(NewGD.getDecl());
1924 if (NewFD->isDefined()) {
1925 CGM.emitOpenMPDeviceFunctionRedefinition(OldGD, NewGD, OrigAddr);
1926 return true;
1927 }
1928 return false;
1929}
1930
1931CGOpenMPRuntimeNVPTX::CGOpenMPRuntimeNVPTX(CodeGenModule &CGM)
1932 : CGOpenMPRuntime(CGM, "_", "$") {
1933 if (!CGM.getLangOpts().OpenMPIsDevice)
1934 llvm_unreachable("OpenMP NVPTX can only handle device code.")::llvm::llvm_unreachable_internal("OpenMP NVPTX can only handle device code."
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 1934)
;
1935}
1936
1937void CGOpenMPRuntimeNVPTX::emitProcBindClause(CodeGenFunction &CGF,
1938 ProcBindKind ProcBind,
1939 SourceLocation Loc) {
1940 // Do nothing in case of SPMD mode and L0 parallel.
1941 if (getExecutionMode() == CGOpenMPRuntimeNVPTX::EM_SPMD)
1942 return;
1943
1944 CGOpenMPRuntime::emitProcBindClause(CGF, ProcBind, Loc);
1945}
1946
1947void CGOpenMPRuntimeNVPTX::emitNumThreadsClause(CodeGenFunction &CGF,
1948 llvm::Value *NumThreads,
1949 SourceLocation Loc) {
1950 // Do nothing in case of SPMD mode and L0 parallel.
1951 if (getExecutionMode() == CGOpenMPRuntimeNVPTX::EM_SPMD)
1952 return;
1953
1954 CGOpenMPRuntime::emitNumThreadsClause(CGF, NumThreads, Loc);
1955}
1956
1957void CGOpenMPRuntimeNVPTX::emitNumTeamsClause(CodeGenFunction &CGF,
1958 const Expr *NumTeams,
1959 const Expr *ThreadLimit,
1960 SourceLocation Loc) {}
1961
1962llvm::Function *CGOpenMPRuntimeNVPTX::emitParallelOutlinedFunction(
1963 const OMPExecutableDirective &D, const VarDecl *ThreadIDVar,
1964 OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) {
1965 // Emit target region as a standalone region.
1966 class NVPTXPrePostActionTy : public PrePostActionTy {
1967 bool &IsInParallelRegion;
1968 bool PrevIsInParallelRegion;
1969
1970 public:
1971 NVPTXPrePostActionTy(bool &IsInParallelRegion)
1972 : IsInParallelRegion(IsInParallelRegion) {}
1973 void Enter(CodeGenFunction &CGF) override {
1974 PrevIsInParallelRegion = IsInParallelRegion;
1975 IsInParallelRegion = true;
1976 }
1977 void Exit(CodeGenFunction &CGF) override {
1978 IsInParallelRegion = PrevIsInParallelRegion;
1979 }
1980 } Action(IsInParallelRegion);
1981 CodeGen.setAction(Action);
1982 bool PrevIsInTTDRegion = IsInTTDRegion;
1983 IsInTTDRegion = false;
1984 bool PrevIsInTargetMasterThreadRegion = IsInTargetMasterThreadRegion;
1985 IsInTargetMasterThreadRegion = false;
1986 auto *OutlinedFun =
1987 cast<llvm::Function>(CGOpenMPRuntime::emitParallelOutlinedFunction(
1988 D, ThreadIDVar, InnermostKind, CodeGen));
1989 if (CGM.getLangOpts().Optimize) {
1990 OutlinedFun->removeFnAttr(llvm::Attribute::NoInline);
1991 OutlinedFun->removeFnAttr(llvm::Attribute::OptimizeNone);
1992 OutlinedFun->addFnAttr(llvm::Attribute::AlwaysInline);
1993 }
1994 IsInTargetMasterThreadRegion = PrevIsInTargetMasterThreadRegion;
1995 IsInTTDRegion = PrevIsInTTDRegion;
1996 if (getExecutionMode() != CGOpenMPRuntimeNVPTX::EM_SPMD &&
1997 !IsInParallelRegion) {
1998 llvm::Function *WrapperFun =
1999 createParallelDataSharingWrapper(OutlinedFun, D);
2000 WrapperFunctionsMap[OutlinedFun] = WrapperFun;
2001 }
2002
2003 return OutlinedFun;
2004}
2005
2006/// Get list of lastprivate variables from the teams distribute ... or
2007/// teams {distribute ...} directives.
2008static void
2009getDistributeLastprivateVars(ASTContext &Ctx, const OMPExecutableDirective &D,
2010 llvm::SmallVectorImpl<const ValueDecl *> &Vars) {
2011 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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 2012, __PRETTY_FUNCTION__))
2012 "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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 2012, __PRETTY_FUNCTION__))
;
2013 const OMPExecutableDirective *Dir = &D;
2014 if (!isOpenMPDistributeDirective(D.getDirectiveKind())) {
2015 if (const Stmt *S = CGOpenMPRuntime::getSingleCompoundChild(
2016 Ctx,
2017 D.getInnermostCapturedStmt()->getCapturedStmt()->IgnoreContainers(
2018 /*IgnoreCaptured=*/true))) {
2019 Dir = dyn_cast_or_null<OMPExecutableDirective>(S);
2020 if (Dir && !isOpenMPDistributeDirective(Dir->getDirectiveKind()))
2021 Dir = nullptr;
2022 }
2023 }
2024 if (!Dir)
2025 return;
2026 for (const auto *C : Dir->getClausesOfKind<OMPLastprivateClause>()) {
2027 for (const Expr *E : C->getVarRefs())
2028 Vars.push_back(getPrivateItem(E));
2029 }
2030}
2031
2032/// Get list of reduction variables from the teams ... directives.
2033static void
2034getTeamsReductionVars(ASTContext &Ctx, const OMPExecutableDirective &D,
2035 llvm::SmallVectorImpl<const ValueDecl *> &Vars) {
2036 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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 2037, __PRETTY_FUNCTION__))
2037 "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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 2037, __PRETTY_FUNCTION__))
;
2038 for (const auto *C : D.getClausesOfKind<OMPReductionClause>()) {
2039 for (const Expr *E : C->privates())
2040 Vars.push_back(getPrivateItem(E));
2041 }
2042}
2043
2044llvm::Function *CGOpenMPRuntimeNVPTX::emitTeamsOutlinedFunction(
2045 const OMPExecutableDirective &D, const VarDecl *ThreadIDVar,
2046 OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) {
2047 SourceLocation Loc = D.getBeginLoc();
2048
2049 const RecordDecl *GlobalizedRD = nullptr;
2050 llvm::SmallVector<const ValueDecl *, 4> LastPrivatesReductions;
2051 llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> MappedDeclsFields;
2052 // Globalize team reductions variable unconditionally in all modes.
2053 if (getExecutionMode() != CGOpenMPRuntimeNVPTX::EM_SPMD)
2054 getTeamsReductionVars(CGM.getContext(), D, LastPrivatesReductions);
2055 if (getExecutionMode() == CGOpenMPRuntimeNVPTX::EM_SPMD) {
2056 getDistributeLastprivateVars(CGM.getContext(), D, LastPrivatesReductions);
2057 if (!LastPrivatesReductions.empty()) {
2058 GlobalizedRD = ::buildRecordForGlobalizedVars(
2059 CGM.getContext(), llvm::None, LastPrivatesReductions,
2060 MappedDeclsFields, WarpSize);
2061 }
2062 } else if (!LastPrivatesReductions.empty()) {
2063 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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 2064, __PRETTY_FUNCTION__))
2064 "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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 2064, __PRETTY_FUNCTION__))
;
2065 TeamAndReductions.first = D.getCapturedStmt(OMPD_teams)->getCapturedDecl();
2066 std::swap(TeamAndReductions.second, LastPrivatesReductions);
2067 }
2068
2069 // Emit target region as a standalone region.
2070 class NVPTXPrePostActionTy : public PrePostActionTy {
2071 SourceLocation &Loc;
2072 const RecordDecl *GlobalizedRD;
2073 llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
2074 &MappedDeclsFields;
2075
2076 public:
2077 NVPTXPrePostActionTy(
2078 SourceLocation &Loc, const RecordDecl *GlobalizedRD,
2079 llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
2080 &MappedDeclsFields)
2081 : Loc(Loc), GlobalizedRD(GlobalizedRD),
2082 MappedDeclsFields(MappedDeclsFields) {}
2083 void Enter(CodeGenFunction &CGF) override {
2084 auto &Rt =
2085 static_cast<CGOpenMPRuntimeNVPTX &>(CGF.CGM.getOpenMPRuntime());
2086 if (GlobalizedRD) {
2087 auto I = Rt.FunctionGlobalizedDecls.try_emplace(CGF.CurFn).first;
2088 I->getSecond().GlobalRecord = GlobalizedRD;
2089 I->getSecond().MappedParams =
2090 std::make_unique<CodeGenFunction::OMPMapVars>();
2091 DeclToAddrMapTy &Data = I->getSecond().LocalVarData;
2092 for (const auto &Pair : MappedDeclsFields) {
2093 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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 2094, __PRETTY_FUNCTION__))
2094 "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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 2094, __PRETTY_FUNCTION__))
;
2095 Data.insert(std::make_pair(Pair.getFirst(),
2096 MappedVarData(Pair.getSecond(),
2097 /*IsOnePerTeam=*/true)));
2098 }
2099 }
2100 Rt.emitGenericVarsProlog(CGF, Loc);
2101 }
2102 void Exit(CodeGenFunction &CGF) override {
2103 static_cast<CGOpenMPRuntimeNVPTX &>(CGF.CGM.getOpenMPRuntime())
2104 .emitGenericVarsEpilog(CGF);
2105 }
2106 } Action(Loc, GlobalizedRD, MappedDeclsFields);
2107 CodeGen.setAction(Action);
2108 llvm::Function *OutlinedFun = CGOpenMPRuntime::emitTeamsOutlinedFunction(
2109 D, ThreadIDVar, InnermostKind, CodeGen);
2110 if (CGM.getLangOpts().Optimize) {
2111 OutlinedFun->removeFnAttr(llvm::Attribute::NoInline);
2112 OutlinedFun->removeFnAttr(llvm::Attribute::OptimizeNone);
2113 OutlinedFun->addFnAttr(llvm::Attribute::AlwaysInline);
2114 }
2115
2116 return OutlinedFun;
2117}
2118
2119void CGOpenMPRuntimeNVPTX::emitGenericVarsProlog(CodeGenFunction &CGF,
2120 SourceLocation Loc,
2121 bool WithSPMDCheck) {
2122 if (getDataSharingMode(CGM) != CGOpenMPRuntimeNVPTX::Generic &&
2123 getExecutionMode() != CGOpenMPRuntimeNVPTX::EM_SPMD)
2124 return;
2125
2126 CGBuilderTy &Bld = CGF.Builder;
2127
2128 const auto I = FunctionGlobalizedDecls.find(CGF.CurFn);
2129 if (I == FunctionGlobalizedDecls.end())
2130 return;
2131 if (const RecordDecl *GlobalizedVarsRecord = I->getSecond().GlobalRecord) {
2132 QualType GlobalRecTy = CGM.getContext().getRecordType(GlobalizedVarsRecord);
2133 QualType SecGlobalRecTy;
2134
2135 // Recover pointer to this function's global record. The runtime will
2136 // handle the specifics of the allocation of the memory.
2137 // Use actual memory size of the record including the padding
2138 // for alignment purposes.
2139 unsigned Alignment =
2140 CGM.getContext().getTypeAlignInChars(GlobalRecTy).getQuantity();
2141 unsigned GlobalRecordSize =
2142 CGM.getContext().getTypeSizeInChars(GlobalRecTy).getQuantity();
2143 GlobalRecordSize = llvm::alignTo(GlobalRecordSize, Alignment);
2144
2145 llvm::PointerType *GlobalRecPtrTy =
2146 CGF.ConvertTypeForMem(GlobalRecTy)->getPointerTo();
2147 llvm::Value *GlobalRecCastAddr;
2148 llvm::Value *IsTTD = nullptr;
2149 if (!IsInTTDRegion &&
2150 (WithSPMDCheck ||
2151 getExecutionMode() == CGOpenMPRuntimeNVPTX::EM_Unknown)) {
2152 llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".exit");
2153 llvm::BasicBlock *SPMDBB = CGF.createBasicBlock(".spmd");
2154 llvm::BasicBlock *NonSPMDBB = CGF.createBasicBlock(".non-spmd");
2155 if (I->getSecond().SecondaryGlobalRecord.hasValue()) {
2156 llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc);
2157 llvm::Value *ThreadID = getThreadID(CGF, Loc);
2158 llvm::Value *PL = CGF.EmitRuntimeCall(
2159 createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_parallel_level),
2160 {RTLoc, ThreadID});
2161 IsTTD = Bld.CreateIsNull(PL);
2162 }
2163 llvm::Value *IsSPMD = Bld.CreateIsNotNull(CGF.EmitNounwindRuntimeCall(
2164 createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_is_spmd_exec_mode)));
2165 Bld.CreateCondBr(IsSPMD, SPMDBB, NonSPMDBB);
2166 // There is no need to emit line number for unconditional branch.
2167 (void)ApplyDebugLocation::CreateEmpty(CGF);
2168 CGF.EmitBlock(SPMDBB);
2169 Address RecPtr = Address(llvm::ConstantPointerNull::get(GlobalRecPtrTy),
2170 CharUnits::fromQuantity(Alignment));
2171 CGF.EmitBranch(ExitBB);
2172 // There is no need to emit line number for unconditional branch.
2173 (void)ApplyDebugLocation::CreateEmpty(CGF);
2174 CGF.EmitBlock(NonSPMDBB);
2175 llvm::Value *Size = llvm::ConstantInt::get(CGM.SizeTy, GlobalRecordSize);
2176 if (const RecordDecl *SecGlobalizedVarsRecord =
2177 I->getSecond().SecondaryGlobalRecord.getValueOr(nullptr)) {
2178 SecGlobalRecTy =
2179 CGM.getContext().getRecordType(SecGlobalizedVarsRecord);
2180
2181 // Recover pointer to this function's global record. The runtime will
2182 // handle the specifics of the allocation of the memory.
2183 // Use actual memory size of the record including the padding
2184 // for alignment purposes.
2185 unsigned Alignment =
2186 CGM.getContext().getTypeAlignInChars(SecGlobalRecTy).getQuantity();
2187 unsigned GlobalRecordSize =
2188 CGM.getContext().getTypeSizeInChars(SecGlobalRecTy).getQuantity();
2189 GlobalRecordSize = llvm::alignTo(GlobalRecordSize, Alignment);
2190 Size = Bld.CreateSelect(
2191 IsTTD, llvm::ConstantInt::get(CGM.SizeTy, GlobalRecordSize), Size);
2192 }
2193 // TODO: allow the usage of shared memory to be controlled by
2194 // the user, for now, default to global.
2195 llvm::Value *GlobalRecordSizeArg[] = {
2196 Size, CGF.Builder.getInt16(/*UseSharedMemory=*/0)};
2197 llvm::Value *GlobalRecValue = CGF.EmitRuntimeCall(
2198 createNVPTXRuntimeFunction(
2199 OMPRTL_NVPTX__kmpc_data_sharing_coalesced_push_stack),
2200 GlobalRecordSizeArg);
2201 GlobalRecCastAddr = Bld.CreatePointerBitCastOrAddrSpaceCast(
2202 GlobalRecValue, GlobalRecPtrTy);
2203 CGF.EmitBlock(ExitBB);
2204 auto *Phi = Bld.CreatePHI(GlobalRecPtrTy,
2205 /*NumReservedValues=*/2, "_select_stack");
2206 Phi->addIncoming(RecPtr.getPointer(), SPMDBB);
2207 Phi->addIncoming(GlobalRecCastAddr, NonSPMDBB);
2208 GlobalRecCastAddr = Phi;
2209 I->getSecond().GlobalRecordAddr = Phi;
2210 I->getSecond().IsInSPMDModeFlag = IsSPMD;
2211 } else if (IsInTTDRegion) {
2212 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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 2214, __PRETTY_FUNCTION__))
2213 "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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 2214, __PRETTY_FUNCTION__))
2214 "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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 2214, __PRETTY_FUNCTION__))
;
2215 unsigned Offset = 0;
2216 for (const RecordDecl *RD : GlobalizedRecords.back().Records) {
2217 QualType RDTy = CGM.getContext().getRecordType(RD);
2218 unsigned Alignment =
2219 CGM.getContext().getTypeAlignInChars(RDTy).getQuantity();
2220 unsigned Size = CGM.getContext().getTypeSizeInChars(RDTy).getQuantity();
2221 Offset =
2222 llvm::alignTo(llvm::alignTo(Offset, Alignment) + Size, Alignment);
2223 }
2224 unsigned Alignment =
2225 CGM.getContext().getTypeAlignInChars(GlobalRecTy).getQuantity();
2226 Offset = llvm::alignTo(Offset, Alignment);
2227 GlobalizedRecords.back().Records.push_back(GlobalizedVarsRecord);
2228 ++GlobalizedRecords.back().RegionCounter;
2229 if (GlobalizedRecords.back().Records.size() == 1) {
2230 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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 2231, __PRETTY_FUNCTION__))
2231 "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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 2231, __PRETTY_FUNCTION__))
;
2232 auto *UseSharedMemory = new llvm::GlobalVariable(
2233 CGM.getModule(), CGM.Int16Ty, /*isConstant=*/true,
2234 llvm::GlobalValue::InternalLinkage, nullptr,
2235 "_openmp_static_kernel$is_shared");
2236 UseSharedMemory->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
2237 QualType Int16Ty = CGM.getContext().getIntTypeForBitwidth(
2238 /*DestWidth=*/16, /*Signed=*/0);
2239 llvm::Value *IsInSharedMemory = CGF.EmitLoadOfScalar(
2240 Address(UseSharedMemory,
2241 CGM.getContext().getTypeAlignInChars(Int16Ty)),
2242 /*Volatile=*/false, Int16Ty, Loc);
2243 auto *StaticGlobalized = new llvm::GlobalVariable(
2244 CGM.getModule(), CGM.Int8Ty, /*isConstant=*/false,
2245 llvm::GlobalValue::CommonLinkage, nullptr);
2246 auto *RecSize = new llvm::GlobalVariable(
2247 CGM.getModule(), CGM.SizeTy, /*isConstant=*/true,
2248 llvm::GlobalValue::InternalLinkage, nullptr,
2249 "_openmp_static_kernel$size");
2250 RecSize->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
2251 llvm::Value *Ld = CGF.EmitLoadOfScalar(
2252 Address(RecSize, CGM.getSizeAlign()), /*Volatile=*/false,
2253 CGM.getContext().getSizeType(), Loc);
2254 llvm::Value *ResAddr = Bld.CreatePointerBitCastOrAddrSpaceCast(
2255 KernelStaticGlobalized, CGM.VoidPtrPtrTy);
2256 llvm::Value *GlobalRecordSizeArg[] = {
2257 llvm::ConstantInt::get(
2258 CGM.Int16Ty,
2259 getExecutionMode() == CGOpenMPRuntimeNVPTX::EM_SPMD ? 1 : 0),
2260 StaticGlobalized, Ld, IsInSharedMemory, ResAddr};
2261 CGF.EmitRuntimeCall(createNVPTXRuntimeFunction(
2262 OMPRTL_NVPTX__kmpc_get_team_static_memory),
2263 GlobalRecordSizeArg);
2264 GlobalizedRecords.back().Buffer = StaticGlobalized;
2265 GlobalizedRecords.back().RecSize = RecSize;
2266 GlobalizedRecords.back().UseSharedMemory = UseSharedMemory;
2267 GlobalizedRecords.back().Loc = Loc;
2268 }
2269 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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 2269, __PRETTY_FUNCTION__))
;
2270 Address FrameAddr = CGF.EmitLoadOfPointer(
2271 Address(KernelStaticGlobalized, CGM.getPointerAlign()),
2272 CGM.getContext()
2273 .getPointerType(CGM.getContext().VoidPtrTy)
2274 .castAs<PointerType>());
2275 llvm::Value *GlobalRecValue =
2276 Bld.CreateConstInBoundsGEP(FrameAddr, Offset).getPointer();
2277 I->getSecond().GlobalRecordAddr = GlobalRecValue;
2278 I->getSecond().IsInSPMDModeFlag = nullptr;
2279 GlobalRecCastAddr = Bld.CreatePointerBitCastOrAddrSpaceCast(
2280 GlobalRecValue, CGF.ConvertTypeForMem(GlobalRecTy)->getPointerTo());
2281 } else {
2282 // TODO: allow the usage of shared memory to be controlled by
2283 // the user, for now, default to global.
2284 llvm::Value *GlobalRecordSizeArg[] = {
2285 llvm::ConstantInt::get(CGM.SizeTy, GlobalRecordSize),
2286 CGF.Builder.getInt16(/*UseSharedMemory=*/0)};
2287 llvm::Value *GlobalRecValue = CGF.EmitRuntimeCall(
2288 createNVPTXRuntimeFunction(
2289 OMPRTL_NVPTX__kmpc_data_sharing_coalesced_push_stack),
2290 GlobalRecordSizeArg);
2291 GlobalRecCastAddr = Bld.CreatePointerBitCastOrAddrSpaceCast(
2292 GlobalRecValue, GlobalRecPtrTy);
2293 I->getSecond().GlobalRecordAddr = GlobalRecValue;
2294 I->getSecond().IsInSPMDModeFlag = nullptr;
2295 }
2296 LValue Base =
2297 CGF.MakeNaturalAlignPointeeAddrLValue(GlobalRecCastAddr, GlobalRecTy);
2298
2299 // Emit the "global alloca" which is a GEP from the global declaration
2300 // record using the pointer returned by the runtime.
2301 LValue SecBase;
2302 decltype(I->getSecond().LocalVarData)::const_iterator SecIt;
2303 if (IsTTD) {
2304 SecIt = I->getSecond().SecondaryLocalVarData->begin();
2305 llvm::PointerType *SecGlobalRecPtrTy =
2306 CGF.ConvertTypeForMem(SecGlobalRecTy)->getPointerTo();
2307 SecBase = CGF.MakeNaturalAlignPointeeAddrLValue(
2308 Bld.CreatePointerBitCastOrAddrSpaceCast(
2309 I->getSecond().GlobalRecordAddr, SecGlobalRecPtrTy),
2310 SecGlobalRecTy);
2311 }
2312 for (auto &Rec : I->getSecond().LocalVarData) {
2313 bool EscapedParam = I->getSecond().EscapedParameters.count(Rec.first);
2314 llvm::Value *ParValue;
2315 if (EscapedParam) {
2316 const auto *VD = cast<VarDecl>(Rec.first);
2317 LValue ParLVal =
2318 CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(VD), VD->getType());
2319 ParValue = CGF.EmitLoadOfScalar(ParLVal, Loc);
2320 }
2321 LValue VarAddr = CGF.EmitLValueForField(Base, Rec.second.FD);
2322 // Emit VarAddr basing on lane-id if required.
2323 QualType VarTy;
2324 if (Rec.second.IsOnePerTeam) {
2325 VarTy = Rec.second.FD->getType();
2326 } else {
2327 llvm::Value *Ptr = CGF.Builder.CreateInBoundsGEP(
2328 VarAddr.getAddress(CGF).getPointer(),
2329 {Bld.getInt32(0), getNVPTXLaneID(CGF)});
2330 VarTy =
2331 Rec.second.FD->getType()->castAsArrayTypeUnsafe()->getElementType();
2332 VarAddr = CGF.MakeAddrLValue(
2333 Address(Ptr, CGM.getContext().getDeclAlign(Rec.first)), VarTy,
2334 AlignmentSource::Decl);
2335 }
2336 Rec.second.PrivateAddr = VarAddr.getAddress(CGF);
2337 if (!IsInTTDRegion &&
2338 (WithSPMDCheck ||
2339 getExecutionMode() == CGOpenMPRuntimeNVPTX::EM_Unknown)) {
2340 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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 2341, __PRETTY_FUNCTION__))
2341 "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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 2341, __PRETTY_FUNCTION__))
;
2342 if (IsTTD) {
2343 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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 2344, __PRETTY_FUNCTION__))
2344 "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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 2344, __PRETTY_FUNCTION__))
;
2345 LValue SecVarAddr = CGF.EmitLValueForField(SecBase, SecIt->second.FD);
2346 VarAddr.setAddress(
2347 Address(Bld.CreateSelect(IsTTD, SecVarAddr.getPointer(CGF),
2348 VarAddr.getPointer(CGF)),
2349 VarAddr.getAlignment()));
2350 Rec.second.PrivateAddr = VarAddr.getAddress(CGF);
2351 }
2352 Address GlobalPtr = Rec.second.PrivateAddr;
2353 Address LocalAddr = CGF.CreateMemTemp(VarTy, Rec.second.FD->getName());
2354 Rec.second.PrivateAddr = Address(
2355 Bld.CreateSelect(I->getSecond().IsInSPMDModeFlag,
2356 LocalAddr.getPointer(), GlobalPtr.getPointer()),
2357 LocalAddr.getAlignment());
2358 }
2359 if (EscapedParam) {
2360 const auto *VD = cast<VarDecl>(Rec.first);
2361 CGF.EmitStoreOfScalar(ParValue, VarAddr);
2362 I->getSecond().MappedParams->setVarAddr(CGF, VD,
2363 VarAddr.getAddress(CGF));
2364 }
2365 if (IsTTD)
2366 ++SecIt;
2367 }
2368 }
2369 for (const ValueDecl *VD : I->getSecond().EscapedVariableLengthDecls) {
2370 // Recover pointer to this function's global record. The runtime will
2371 // handle the specifics of the allocation of the memory.
2372 // Use actual memory size of the record including the padding
2373 // for alignment purposes.
2374 CGBuilderTy &Bld = CGF.Builder;
2375 llvm::Value *Size = CGF.getTypeSize(VD->getType());
2376 CharUnits Align = CGM.getContext().getDeclAlign(VD);
2377 Size = Bld.CreateNUWAdd(
2378 Size, llvm::ConstantInt::get(CGF.SizeTy, Align.getQuantity() - 1));
2379 llvm::Value *AlignVal =
2380 llvm::ConstantInt::get(CGF.SizeTy, Align.getQuantity());
2381 Size = Bld.CreateUDiv(Size, AlignVal);
2382 Size = Bld.CreateNUWMul(Size, AlignVal);
2383 // TODO: allow the usage of shared memory to be controlled by
2384 // the user, for now, default to global.
2385 llvm::Value *GlobalRecordSizeArg[] = {
2386 Size, CGF.Builder.getInt16(/*UseSharedMemory=*/0)};
2387 llvm::Value *GlobalRecValue = CGF.EmitRuntimeCall(
2388 createNVPTXRuntimeFunction(
2389 OMPRTL_NVPTX__kmpc_data_sharing_coalesced_push_stack),
2390 GlobalRecordSizeArg);
2391 llvm::Value *GlobalRecCastAddr = Bld.CreatePointerBitCastOrAddrSpaceCast(
2392 GlobalRecValue, CGF.ConvertTypeForMem(VD->getType())->getPointerTo());
2393 LValue Base = CGF.MakeAddrLValue(GlobalRecCastAddr, VD->getType(),
2394 CGM.getContext().getDeclAlign(VD),
2395 AlignmentSource::Decl);
2396 I->getSecond().MappedParams->setVarAddr(CGF, cast<VarDecl>(VD),
2397 Base.getAddress(CGF));
2398 I->getSecond().EscapedVariableLengthDeclsAddrs.emplace_back(GlobalRecValue);
2399 }
2400 I->getSecond().MappedParams->apply(CGF);
2401}
2402
2403void CGOpenMPRuntimeNVPTX::emitGenericVarsEpilog(CodeGenFunction &CGF,
2404 bool WithSPMDCheck) {
2405 if (getDataSharingMode(CGM) != CGOpenMPRuntimeNVPTX::Generic &&
2406 getExecutionMode() != CGOpenMPRuntimeNVPTX::EM_SPMD)
2407 return;
2408
2409 const auto I = FunctionGlobalizedDecls.find(CGF.CurFn);
2410 if (I != FunctionGlobalizedDecls.end()) {
2411 I->getSecond().MappedParams->restore(CGF);
2412 if (!CGF.HaveInsertPoint())
2413 return;
2414 for (llvm::Value *Addr :
2415 llvm::reverse(I->getSecond().EscapedVariableLengthDeclsAddrs)) {
2416 CGF.EmitRuntimeCall(
2417 createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_data_sharing_pop_stack),
2418 Addr);
2419 }
2420 if (I->getSecond().GlobalRecordAddr) {
2421 if (!IsInTTDRegion &&
2422 (WithSPMDCheck ||
2423 getExecutionMode() == CGOpenMPRuntimeNVPTX::EM_Unknown)) {
2424 CGBuilderTy &Bld = CGF.Builder;
2425 llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".exit");
2426 llvm::BasicBlock *NonSPMDBB = CGF.createBasicBlock(".non-spmd");
2427 Bld.CreateCondBr(I->getSecond().IsInSPMDModeFlag, ExitBB, NonSPMDBB);
2428 // There is no need to emit line number for unconditional branch.
2429 (void)ApplyDebugLocation::CreateEmpty(CGF);
2430 CGF.EmitBlock(NonSPMDBB);
2431 CGF.EmitRuntimeCall(
2432 createNVPTXRuntimeFunction(
2433 OMPRTL_NVPTX__kmpc_data_sharing_pop_stack),
2434 CGF.EmitCastToVoidPtr(I->getSecond().GlobalRecordAddr));
2435 CGF.EmitBlock(ExitBB);
2436 } else if (IsInTTDRegion) {
2437 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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 2438, __PRETTY_FUNCTION__))
2438 "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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 2438, __PRETTY_FUNCTION__))
;
2439 --GlobalizedRecords.back().RegionCounter;
2440 // Emit the restore function only in the target region.
2441 if (GlobalizedRecords.back().RegionCounter == 0) {
2442 QualType Int16Ty = CGM.getContext().getIntTypeForBitwidth(
2443 /*DestWidth=*/16, /*Signed=*/0);
2444 llvm::Value *IsInSharedMemory = CGF.EmitLoadOfScalar(
2445 Address(GlobalizedRecords.back().UseSharedMemory,
2446 CGM.getContext().getTypeAlignInChars(Int16Ty)),
2447 /*Volatile=*/false, Int16Ty, GlobalizedRecords.back().Loc);
2448 llvm::Value *Args[] = {
2449 llvm::ConstantInt::get(
2450 CGM.Int16Ty,
2451 getExecutionMode() == CGOpenMPRuntimeNVPTX::EM_SPMD ? 1 : 0),
2452 IsInSharedMemory};
2453 CGF.EmitRuntimeCall(
2454 createNVPTXRuntimeFunction(
2455 OMPRTL_NVPTX__kmpc_restore_team_static_memory),
2456 Args);
2457 }
2458 } else {
2459 CGF.EmitRuntimeCall(createNVPTXRuntimeFunction(
2460 OMPRTL_NVPTX__kmpc_data_sharing_pop_stack),
2461 I->getSecond().GlobalRecordAddr);
2462 }
2463 }
2464 }
2465}
2466
2467void CGOpenMPRuntimeNVPTX::emitTeamsCall(CodeGenFunction &CGF,
2468 const OMPExecutableDirective &D,
2469 SourceLocation Loc,
2470 llvm::Function *OutlinedFn,
2471 ArrayRef<llvm::Value *> CapturedVars) {
2472 if (!CGF.HaveInsertPoint())
2473 return;
2474
2475 Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty,
2476 /*Name=*/".zero.addr");
2477 CGF.InitTempAlloca(ZeroAddr, CGF.Builder.getInt32(/*C*/ 0));
2478 llvm::SmallVector<llvm::Value *, 16> OutlinedFnArgs;
2479 OutlinedFnArgs.push_back(emitThreadIDAddress(CGF, Loc).getPointer());
2480 OutlinedFnArgs.push_back(ZeroAddr.getPointer());
2481 OutlinedFnArgs.append(CapturedVars.begin(), CapturedVars.end());
2482 emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, OutlinedFnArgs);
2483}
2484
2485void CGOpenMPRuntimeNVPTX::emitParallelCall(
2486 CodeGenFunction &CGF, SourceLocation Loc, llvm::Function *OutlinedFn,
2487 ArrayRef<llvm::Value *> CapturedVars, const Expr *IfCond) {
2488 if (!CGF.HaveInsertPoint())
2489 return;
2490
2491 if (getExecutionMode() == CGOpenMPRuntimeNVPTX::EM_SPMD)
2492 emitSPMDParallelCall(CGF, Loc, OutlinedFn, CapturedVars, IfCond);
2493 else
2494 emitNonSPMDParallelCall(CGF, Loc, OutlinedFn, CapturedVars, IfCond);
2495}
2496
2497void CGOpenMPRuntimeNVPTX::emitNonSPMDParallelCall(
2498 CodeGenFunction &CGF, SourceLocation Loc, llvm::Value *OutlinedFn,
2499 ArrayRef<llvm::Value *> CapturedVars, const Expr *IfCond) {
2500 llvm::Function *Fn = cast<llvm::Function>(OutlinedFn);
2501
2502 // Force inline this outlined function at its call site.
2503 Fn->setLinkage(llvm::GlobalValue::InternalLinkage);
2504
2505 Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty,
2506 /*Name=*/".zero.addr");
2507 CGF.InitTempAlloca(ZeroAddr, CGF.Builder.getInt32(/*C*/ 0));
2508 // ThreadId for serialized parallels is 0.
2509 Address ThreadIDAddr = ZeroAddr;
2510 auto &&CodeGen = [this, Fn, CapturedVars, Loc, &ThreadIDAddr](
2511 CodeGenFunction &CGF, PrePostActionTy &Action) {
2512 Action.Enter(CGF);
2513
2514 Address ZeroAddr =
2515 CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty,
2516 /*Name=*/".bound.zero.addr");
2517 CGF.InitTempAlloca(ZeroAddr, CGF.Builder.getInt32(/*C*/ 0));
2518 llvm::SmallVector<llvm::Value *, 16> OutlinedFnArgs;
2519 OutlinedFnArgs.push_back(ThreadIDAddr.getPointer());
2520 OutlinedFnArgs.push_back(ZeroAddr.getPointer());
2521 OutlinedFnArgs.append(CapturedVars.begin(), CapturedVars.end());
2522 emitOutlinedFunctionCall(CGF, Loc, Fn, OutlinedFnArgs);
2523 };
2524 auto &&SeqGen = [this, &CodeGen, Loc](CodeGenFunction &CGF,
2525 PrePostActionTy &) {
2526
2527 RegionCodeGenTy RCG(CodeGen);
2528 llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc);
2529 llvm::Value *ThreadID = getThreadID(CGF, Loc);
2530 llvm::Value *Args[] = {RTLoc, ThreadID};
2531
2532 NVPTXActionTy Action(
2533 createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_serialized_parallel),
2534 Args,
2535 createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_end_serialized_parallel),
2536 Args);
2537 RCG.setAction(Action);
2538 RCG(CGF);
2539 };
2540
2541 auto &&L0ParallelGen = [this, CapturedVars, Fn](CodeGenFunction &CGF,
2542 PrePostActionTy &Action) {
2543 CGBuilderTy &Bld = CGF.Builder;
2544 llvm::Function *WFn = WrapperFunctionsMap[Fn];
2545 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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 2545, __PRETTY_FUNCTION__))
;
2546 llvm::Value *ID = Bld.CreateBitOrPointerCast(WFn, CGM.Int8PtrTy);
2547
2548 // Prepare for parallel region. Indicate the outlined function.
2549 llvm::Value *Args[] = {ID, /*RequiresOMPRuntime=*/Bld.getInt16(1)};
2550 CGF.EmitRuntimeCall(
2551 createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_kernel_prepare_parallel),
2552 Args);
2553
2554 // Create a private scope that will globalize the arguments
2555 // passed from the outside of the target region.
2556 CodeGenFunction::OMPPrivateScope PrivateArgScope(CGF);
2557
2558 // There's something to share.
2559 if (!CapturedVars.empty()) {
2560 // Prepare for parallel region. Indicate the outlined function.
2561 Address SharedArgs =
2562 CGF.CreateDefaultAlignTempAlloca(CGF.VoidPtrPtrTy, "shared_arg_refs");
2563 llvm::Value *SharedArgsPtr = SharedArgs.getPointer();
2564
2565 llvm::Value *DataSharingArgs[] = {
2566 SharedArgsPtr,
2567 llvm::ConstantInt::get(CGM.SizeTy, CapturedVars.size())};
2568 CGF.EmitRuntimeCall(createNVPTXRuntimeFunction(
2569 OMPRTL_NVPTX__kmpc_begin_sharing_variables),
2570 DataSharingArgs);
2571
2572 // Store variable address in a list of references to pass to workers.
2573 unsigned Idx = 0;
2574 ASTContext &Ctx = CGF.getContext();
2575 Address SharedArgListAddress = CGF.EmitLoadOfPointer(
2576 SharedArgs, Ctx.getPointerType(Ctx.getPointerType(Ctx.VoidPtrTy))
2577 .castAs<PointerType>());
2578 for (llvm::Value *V : CapturedVars) {
2579 Address Dst = Bld.CreateConstInBoundsGEP(SharedArgListAddress, Idx);
2580 llvm::Value *PtrV;
2581 if (V->getType()->isIntegerTy())
2582 PtrV = Bld.CreateIntToPtr(V, CGF.VoidPtrTy);
2583 else
2584 PtrV = Bld.CreatePointerBitCastOrAddrSpaceCast(V, CGF.VoidPtrTy);
2585 CGF.EmitStoreOfScalar(PtrV, Dst, /*Volatile=*/false,
2586 Ctx.getPointerType(Ctx.VoidPtrTy));
2587 ++Idx;
2588 }
2589 }
2590
2591 // Activate workers. This barrier is used by the master to signal
2592 // work for the workers.
2593 syncCTAThreads(CGF);
2594
2595 // OpenMP [2.5, Parallel Construct, p.49]
2596 // There is an implied barrier at the end of a parallel region. After the
2597 // end of a parallel region, only the master thread of the team resumes
2598 // execution of the enclosing task region.
2599 //
2600 // The master waits at this barrier until all workers are done.
2601 syncCTAThreads(CGF);
2602
2603 if (!CapturedVars.empty())
2604 CGF.EmitRuntimeCall(
2605 createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_end_sharing_variables));
2606
2607 // Remember for post-processing in worker loop.
2608 Work.emplace_back(WFn);
2609 };
2610
2611 auto &&LNParallelGen = [this, Loc, &SeqGen, &L0ParallelGen](
2612 CodeGenFunction &CGF, PrePostActionTy &Action) {
2613 if (IsInParallelRegion) {
2614 SeqGen(CGF, Action);
2615 } else if (IsInTargetMasterThreadRegion) {
2616 L0ParallelGen(CGF, Action);
2617 } else {
2618 // Check for master and then parallelism:
2619 // if (__kmpc_is_spmd_exec_mode() || __kmpc_parallel_level(loc, gtid)) {
2620 // Serialized execution.
2621 // } else {
2622 // Worker call.
2623 // }
2624 CGBuilderTy &Bld = CGF.Builder;
2625 llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".exit");
2626 llvm::BasicBlock *SeqBB = CGF.createBasicBlock(".sequential");
2627 llvm::BasicBlock *ParallelCheckBB = CGF.createBasicBlock(".parcheck");
2628 llvm::BasicBlock *MasterBB = CGF.createBasicBlock(".master");
2629 llvm::Value *IsSPMD = Bld.CreateIsNotNull(CGF.EmitNounwindRuntimeCall(
2630 createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_is_spmd_exec_mode)));
2631 Bld.CreateCondBr(IsSPMD, SeqBB, ParallelCheckBB);
2632 // There is no need to emit line number for unconditional branch.
2633 (void)ApplyDebugLocation::CreateEmpty(CGF);
2634 CGF.EmitBlock(ParallelCheckBB);
2635 llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc);
2636 llvm::Value *ThreadID = getThreadID(CGF, Loc);
2637 llvm::Value *PL = CGF.EmitRuntimeCall(
2638 createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_parallel_level),
2639 {RTLoc, ThreadID});
2640 llvm::Value *Res = Bld.CreateIsNotNull(PL);
2641 Bld.CreateCondBr(Res, SeqBB, MasterBB);
2642 CGF.EmitBlock(SeqBB);
2643 SeqGen(CGF, Action);
2644 CGF.EmitBranch(ExitBB);
2645 // There is no need to emit line number for unconditional branch.
2646 (void)ApplyDebugLocation::CreateEmpty(CGF);
2647 CGF.EmitBlock(MasterBB);
2648 L0ParallelGen(CGF, Action);
2649 CGF.EmitBranch(ExitBB);
2650 // There is no need to emit line number for unconditional branch.
2651 (void)ApplyDebugLocation::CreateEmpty(CGF);
2652 // Emit the continuation block for code after the if.
2653 CGF.EmitBlock(ExitBB, /*IsFinished=*/true);
2654 }
2655 };
2656
2657 if (IfCond) {
2658 emitIfClause(CGF, IfCond, LNParallelGen, SeqGen);
2659 } else {
2660 CodeGenFunction::RunCleanupsScope Scope(CGF);
2661 RegionCodeGenTy ThenRCG(LNParallelGen);
2662 ThenRCG(CGF);
2663 }
2664}
2665
2666void CGOpenMPRuntimeNVPTX::emitSPMDParallelCall(
2667 CodeGenFunction &CGF, SourceLocation Loc, llvm::Function *OutlinedFn,
2668 ArrayRef<llvm::Value *> CapturedVars, const Expr *IfCond) {
2669 // Just call the outlined function to execute the parallel region.
2670 // OutlinedFn(&GTid, &zero, CapturedStruct);
2671 //
2672 llvm::SmallVector<llvm::Value *, 16> OutlinedFnArgs;
2673
2674 Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty,
2675 /*Name=*/".zero.addr");
2676 CGF.InitTempAlloca(ZeroAddr, CGF.Builder.getInt32(/*C*/ 0));
2677 // ThreadId for serialized parallels is 0.
2678 Address ThreadIDAddr = ZeroAddr;
2679 auto &&CodeGen = [this, OutlinedFn, CapturedVars, Loc, &ThreadIDAddr](
2680 CodeGenFunction &CGF, PrePostActionTy &Action) {
2681 Action.Enter(CGF);
2682
2683 Address ZeroAddr =
2684 CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty,
2685 /*Name=*/".bound.zero.addr");
2686 CGF.InitTempAlloca(ZeroAddr, CGF.Builder.getInt32(/*C*/ 0));
2687 llvm::SmallVector<llvm::Value *, 16> OutlinedFnArgs;
2688 OutlinedFnArgs.push_back(ThreadIDAddr.getPointer());
2689 OutlinedFnArgs.push_back(ZeroAddr.getPointer());
2690 OutlinedFnArgs.append(CapturedVars.begin(), CapturedVars.end());
2691 emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, OutlinedFnArgs);
2692 };
2693 auto &&SeqGen = [this, &CodeGen, Loc](CodeGenFunction &CGF,
2694 PrePostActionTy &) {
2695
2696 RegionCodeGenTy RCG(CodeGen);
2697 llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc);
2698 llvm::Value *ThreadID = getThreadID(CGF, Loc);
2699 llvm::Value *Args[] = {RTLoc, ThreadID};
2700
2701 NVPTXActionTy Action(
2702 createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_serialized_parallel),
2703 Args,
2704 createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_end_serialized_parallel),
2705 Args);
2706 RCG.setAction(Action);
2707 RCG(CGF);
2708 };
2709
2710 if (IsInTargetMasterThreadRegion) {
2711 // In the worker need to use the real thread id.
2712 ThreadIDAddr = emitThreadIDAddress(CGF, Loc);
2713 RegionCodeGenTy RCG(CodeGen);
2714 RCG(CGF);
2715 } else {
2716 // If we are not in the target region, it is definitely L2 parallelism or
2717 // more, because for SPMD mode we always has L1 parallel level, sowe don't
2718 // need to check for orphaned directives.
2719 RegionCodeGenTy RCG(SeqGen);
2720 RCG(CGF);
2721 }
2722}
2723
2724void CGOpenMPRuntimeNVPTX::syncCTAThreads(CodeGenFunction &CGF) {
2725 // Always emit simple barriers!
2726 if (!CGF.HaveInsertPoint())
2727 return;
2728 // Build call __kmpc_barrier_simple_spmd(nullptr, 0);
2729 // This function does not use parameters, so we can emit just default values.
2730 llvm::Value *Args[] = {
2731 llvm::ConstantPointerNull::get(
2732 cast<llvm::PointerType>(getIdentTyPointerTy())),
2733 llvm::ConstantInt::get(CGF.Int32Ty, /*V=*/0, /*isSigned=*/true)};
2734 llvm::CallInst *Call = CGF.EmitRuntimeCall(
2735 createNVPTXRuntimeFunction(OMPRTL__kmpc_barrier_simple_spmd), Args);
2736 Call->setConvergent();
2737}
2738
2739void CGOpenMPRuntimeNVPTX::emitBarrierCall(CodeGenFunction &CGF,
2740 SourceLocation Loc,
2741 OpenMPDirectiveKind Kind, bool,
2742 bool) {
2743 // Always emit simple barriers!
2744 if (!CGF.HaveInsertPoint())
2745 return;
2746 // Build call __kmpc_cancel_barrier(loc, thread_id);
2747 unsigned Flags = getDefaultFlagsForBarriers(Kind);
2748 llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc, Flags),
2749 getThreadID(CGF, Loc)};
2750 llvm::CallInst *Call = CGF.EmitRuntimeCall(
2751 createNVPTXRuntimeFunction(OMPRTL__kmpc_barrier), Args);
2752 Call->setConvergent();
2753}
2754
2755void CGOpenMPRuntimeNVPTX::emitCriticalRegion(
2756 CodeGenFunction &CGF, StringRef CriticalName,
2757 const RegionCodeGenTy &CriticalOpGen, SourceLocation Loc,
2758 const Expr *Hint) {
2759 llvm::BasicBlock *LoopBB = CGF.createBasicBlock("omp.critical.loop");
2760 llvm::BasicBlock *TestBB = CGF.createBasicBlock("omp.critical.test");
2761 llvm::BasicBlock *SyncBB = CGF.createBasicBlock("omp.critical.sync");
2762 llvm::BasicBlock *BodyBB = CGF.createBasicBlock("omp.critical.body");
2763 llvm::BasicBlock *ExitBB = CGF.createBasicBlock("omp.critical.exit");
2764
2765 // Get the mask of active threads in the warp.
2766 llvm::Value *Mask = CGF.EmitRuntimeCall(
2767 createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_warp_active_thread_mask));
2768 // Fetch team-local id of the thread.
2769 llvm::Value *ThreadID = getNVPTXThreadID(CGF);
2770
2771 // Get the width of the team.
2772 llvm::Value *TeamWidth = getNVPTXNumThreads(CGF);
2773
2774 // Initialize the counter variable for the loop.
2775 QualType Int32Ty =
2776 CGF.getContext().getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/0);
2777 Address Counter = CGF.CreateMemTemp(Int32Ty, "critical_counter");
2778 LValue CounterLVal = CGF.MakeAddrLValue(Counter, Int32Ty);
2779 CGF.EmitStoreOfScalar(llvm::Constant::getNullValue(CGM.Int32Ty), CounterLVal,
2780 /*isInit=*/true);
2781
2782 // Block checks if loop counter exceeds upper bound.
2783 CGF.EmitBlock(LoopBB);
2784 llvm::Value *CounterVal = CGF.EmitLoadOfScalar(CounterLVal, Loc);
2785 llvm::Value *CmpLoopBound = CGF.Builder.CreateICmpSLT(CounterVal, TeamWidth);
2786 CGF.Builder.CreateCondBr(CmpLoopBound, TestBB, ExitBB);
2787
2788 // Block tests which single thread should execute region, and which threads
2789 // should go straight to synchronisation point.
2790 CGF.EmitBlock(TestBB);
2791 CounterVal = CGF.EmitLoadOfScalar(CounterLVal, Loc);
2792 llvm::Value *CmpThreadToCounter =
2793 CGF.Builder.CreateICmpEQ(ThreadID, CounterVal);
2794 CGF.Builder.CreateCondBr(CmpThreadToCounter, BodyBB, SyncBB);
2795
2796 // Block emits the body of the critical region.
2797 CGF.EmitBlock(BodyBB);
2798
2799 // Output the critical statement.
2800 CGOpenMPRuntime::emitCriticalRegion(CGF, CriticalName, CriticalOpGen, Loc,
2801 Hint);
2802
2803 // After the body surrounded by the critical region, the single executing
2804 // thread will jump to the synchronisation point.
2805 // Block waits for all threads in current team to finish then increments the
2806 // counter variable and returns to the loop.
2807 CGF.EmitBlock(SyncBB);
2808 // Reconverge active threads in the warp.
2809 (void)CGF.EmitRuntimeCall(
2810 createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_syncwarp), Mask);
2811
2812 llvm::Value *IncCounterVal =
2813 CGF.Builder.CreateNSWAdd(CounterVal, CGF.Builder.getInt32(1));
2814 CGF.EmitStoreOfScalar(IncCounterVal, CounterLVal);
2815 CGF.EmitBranch(LoopBB);
2816
2817 // Block that is reached when all threads in the team complete the region.
2818 CGF.EmitBlock(ExitBB, /*IsFinished=*/true);
2819}
2820
2821/// Cast value to the specified type.
2822static llvm::Value *castValueToType(CodeGenFunction &CGF, llvm::Value *Val,
2823 QualType ValTy, QualType CastTy,
2824 SourceLocation Loc) {
2825 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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 2826, __PRETTY_FUNCTION__))
2826 "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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 2826, __PRETTY_FUNCTION__))
;
2827 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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 2828, __PRETTY_FUNCTION__))
2828 "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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 2828, __PRETTY_FUNCTION__))
;
2829 llvm::Type *LLVMCastTy = CGF.ConvertTypeForMem(CastTy);
2830 if (ValTy == CastTy)
2831 return Val;
2832 if (CGF.getContext().getTypeSizeInChars(ValTy) ==
2833 CGF.getContext().getTypeSizeInChars(CastTy))
2834 return CGF.Builder.CreateBitCast(Val, LLVMCastTy);
2835 if (CastTy->isIntegerType() && ValTy->isIntegerType())
2836 return CGF.Builder.CreateIntCast(Val, LLVMCastTy,
2837 CastTy->hasSignedIntegerRepresentation());
2838 Address CastItem = CGF.CreateMemTemp(CastTy);
2839 Address ValCastItem = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
2840 CastItem, Val->getType()->getPointerTo(CastItem.getAddressSpace()));
2841 CGF.EmitStoreOfScalar(Val, ValCastItem, /*Volatile=*/false, ValTy);
2842 return CGF.EmitLoadOfScalar(CastItem, /*Volatile=*/false, CastTy, Loc);
2843}
2844
2845/// This function creates calls to one of two shuffle functions to copy
2846/// variables between lanes in a warp.
2847static llvm::Value *createRuntimeShuffleFunction(CodeGenFunction &CGF,
2848 llvm::Value *Elem,
2849 QualType ElemType,
2850 llvm::Value *Offset,
2851 SourceLocation Loc) {
2852 CodeGenModule &CGM = CGF.CGM;
2853 CGBuilderTy &Bld = CGF.Builder;
2854 CGOpenMPRuntimeNVPTX &RT =
2855 *(static_cast<CGOpenMPRuntimeNVPTX *>(&CGM.getOpenMPRuntime()));
2856
2857 CharUnits Size = CGF.getContext().getTypeSizeInChars(ElemType);
2858 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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 2859, __PRETTY_FUNCTION__))
2859 "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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 2859, __PRETTY_FUNCTION__))
;
2860
2861 OpenMPRTLFunctionNVPTX ShuffleFn = Size.getQuantity() <= 4
2862 ? OMPRTL_NVPTX__kmpc_shuffle_int32
2863 : OMPRTL_NVPTX__kmpc_shuffle_int64;
2864
2865 // Cast all types to 32- or 64-bit values before calling shuffle routines.
2866 QualType CastTy = CGF.getContext().getIntTypeForBitwidth(
2867 Size.getQuantity() <= 4 ? 32 : 64, /*Signed=*/1);
2868 llvm::Value *ElemCast = castValueToType(CGF, Elem, ElemType, CastTy, Loc);
2869 llvm::Value *WarpSize =
2870 Bld.CreateIntCast(getNVPTXWarpSize(CGF), CGM.Int16Ty, /*isSigned=*/true);
2871
2872 llvm::Value *ShuffledVal = CGF.EmitRuntimeCall(
2873 RT.createNVPTXRuntimeFunction(ShuffleFn), {ElemCast, Offset, WarpSize});
2874
2875 return castValueToType(CGF, ShuffledVal, CastTy, ElemType, Loc);
2876}
2877
2878static void shuffleAndStore(CodeGenFunction &CGF, Address SrcAddr,
2879 Address DestAddr, QualType ElemType,
2880 llvm::Value *Offset, SourceLocation Loc) {
2881 CGBuilderTy &Bld = CGF.Builder;
2882
2883 CharUnits Size = CGF.getContext().getTypeSizeInChars(ElemType);
2884 // Create the loop over the big sized data.
2885 // ptr = (void*)Elem;
2886 // ptrEnd = (void*) Elem + 1;
2887 // Step = 8;
2888 // while (ptr + Step < ptrEnd)
2889 // shuffle((int64_t)*ptr);
2890 // Step = 4;
2891 // while (ptr + Step < ptrEnd)
2892 // shuffle((int32_t)*ptr);
2893 // ...
2894 Address ElemPtr = DestAddr;
2895 Address Ptr = SrcAddr;
2896 Address PtrEnd = Bld.CreatePointerBitCastOrAddrSpaceCast(
2897 Bld.CreateConstGEP(SrcAddr, 1), CGF.VoidPtrTy);
2898 for (int IntSize = 8; IntSize >= 1; IntSize /= 2) {
2899 if (Size < CharUnits::fromQuantity(IntSize))
2900 continue;
2901 QualType IntType = CGF.getContext().getIntTypeForBitwidth(
2902 CGF.getContext().toBits(CharUnits::fromQuantity(IntSize)),
2903 /*Signed=*/1);
2904 llvm::Type *IntTy = CGF.ConvertTypeForMem(IntType);
2905 Ptr = Bld.CreatePointerBitCastOrAddrSpaceCast(Ptr, IntTy->getPointerTo());
2906 ElemPtr =
2907 Bld.CreatePointerBitCastOrAddrSpaceCast(ElemPtr, IntTy->getPointerTo());
2908 if (Size.getQuantity() / IntSize > 1) {
2909 llvm::BasicBlock *PreCondBB = CGF.createBasicBlock(".shuffle.pre_cond");
2910 llvm::BasicBlock *ThenBB = CGF.createBasicBlock(".shuffle.then");
2911 llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".shuffle.exit");
2912 llvm::BasicBlock *CurrentBB = Bld.GetInsertBlock();
2913 CGF.EmitBlock(PreCondBB);
2914 llvm::PHINode *PhiSrc =
2915 Bld.CreatePHI(Ptr.getType(), /*NumReservedValues=*/2);
2916 PhiSrc->addIncoming(Ptr.getPointer(), CurrentBB);
2917 llvm::PHINode *PhiDest =
2918 Bld.CreatePHI(ElemPtr.getType(), /*NumReservedValues=*/2);
2919 PhiDest->addIncoming(ElemPtr.getPointer(), CurrentBB);
2920 Ptr = Address(PhiSrc, Ptr.getAlignment());
2921 ElemPtr = Address(PhiDest, ElemPtr.getAlignment());
2922 llvm::Value *PtrDiff = Bld.CreatePtrDiff(
2923 PtrEnd.getPointer(), Bld.CreatePointerBitCastOrAddrSpaceCast(
2924 Ptr.getPointer(), CGF.VoidPtrTy));
2925 Bld.CreateCondBr(Bld.CreateICmpSGT(PtrDiff, Bld.getInt64(IntSize - 1)),
2926 ThenBB, ExitBB);
2927 CGF.EmitBlock(ThenBB);
2928 llvm::Value *Res = createRuntimeShuffleFunction(
2929 CGF, CGF.EmitLoadOfScalar(Ptr, /*Volatile=*/false, IntType, Loc),
2930 IntType, Offset, Loc);
2931 CGF.EmitStoreOfScalar(Res, ElemPtr, /*Volatile=*/false, IntType);
2932 Address LocalPtr = Bld.CreateConstGEP(Ptr, 1);
2933 Address LocalElemPtr = Bld.CreateConstGEP(ElemPtr, 1);
2934 PhiSrc->addIncoming(LocalPtr.getPointer(), ThenBB);
2935 PhiDest->addIncoming(LocalElemPtr.getPointer(), ThenBB);
2936 CGF.EmitBranch(PreCondBB);
2937 CGF.EmitBlock(ExitBB);
2938 } else {
2939 llvm::Value *Res = createRuntimeShuffleFunction(
2940 CGF, CGF.EmitLoadOfScalar(Ptr, /*Volatile=*/false, IntType, Loc),
2941 IntType, Offset, Loc);
2942 CGF.EmitStoreOfScalar(Res, ElemPtr, /*Volatile=*/false, IntType);
2943 Ptr = Bld.CreateConstGEP(Ptr, 1);
2944 ElemPtr = Bld.CreateConstGEP(ElemPtr, 1);
2945 }
2946 Size = Size % IntSize;
2947 }
2948}
2949
2950namespace {
2951enum CopyAction : unsigned {
2952 // RemoteLaneToThread: Copy over a Reduce list from a remote lane in
2953 // the warp using shuffle instructions.
2954 RemoteLaneToThread,
2955 // ThreadCopy: Make a copy of a Reduce list on the thread's stack.
2956 ThreadCopy,
2957 // ThreadToScratchpad: Copy a team-reduced array to the scratchpad.
2958 ThreadToScratchpad,
2959 // ScratchpadToThread: Copy from a scratchpad array in global memory
2960 // containing team-reduced data to a thread's stack.
2961 ScratchpadToThread,
2962};
2963} // namespace
2964
2965struct CopyOptionsTy {
2966 llvm::Value *RemoteLaneOffset;
2967 llvm::Value *ScratchpadIndex;
2968 llvm::Value *ScratchpadWidth;
2969};
2970
2971/// Emit instructions to copy a Reduce list, which contains partially
2972/// aggregated values, in the specified direction.
2973static void emitReductionListCopy(
2974 CopyAction Action, CodeGenFunction &CGF, QualType ReductionArrayTy,
2975 ArrayRef<const Expr *> Privates, Address SrcBase, Address DestBase,
2976 CopyOptionsTy CopyOptions = {nullptr, nullptr, nullptr}) {
2977
2978 CodeGenModule &CGM = CGF.CGM;
2979 ASTContext &C = CGM.getContext();
2980 CGBuilderTy &Bld = CGF.Builder;
2981
2982 llvm::Value *RemoteLaneOffset = CopyOptions.RemoteLaneOffset;
2983 llvm::Value *ScratchpadIndex = CopyOptions.ScratchpadIndex;
2984 llvm::Value *ScratchpadWidth = CopyOptions.ScratchpadWidth;
2985
2986 // Iterates, element-by-element, through the source Reduce list and
2987 // make a copy.
2988 unsigned Idx = 0;
2989 unsigned Size = Privates.size();
2990 for (const Expr *Private : Privates) {
2991 Address SrcElementAddr = Address::invalid();
2992 Address DestElementAddr = Address::invalid();
2993 Address DestElementPtrAddr = Address::invalid();
2994 // Should we shuffle in an element from a remote lane?
2995 bool ShuffleInElement = false;
2996 // Set to true to update the pointer in the dest Reduce list to a
2997 // newly created element.
2998 bool UpdateDestListPtr = false;
2999 // Increment the src or dest pointer to the scratchpad, for each
3000 // new element.
3001 bool IncrScratchpadSrc = false;
3002 bool IncrScratchpadDest = false;
3003
3004 switch (Action) {
3005 case RemoteLaneToThread: {
3006 // Step 1.1: Get the address for the src element in the Reduce list.
3007 Address SrcElementPtrAddr = Bld.CreateConstArrayGEP(SrcBase, Idx);
3008 SrcElementAddr = CGF.EmitLoadOfPointer(
3009 SrcElementPtrAddr,
3010 C.getPointerType(Private->getType())->castAs<PointerType>());
3011
3012 // Step 1.2: Create a temporary to store the element in the destination
3013 // Reduce list.
3014 DestElementPtrAddr = Bld.CreateConstArrayGEP(DestBase, Idx);
3015 DestElementAddr =
3016 CGF.CreateMemTemp(Private->getType(), ".omp.reduction.element");
3017 ShuffleInElement = true;
3018 UpdateDestListPtr = true;
3019 break;
3020 }
3021 case ThreadCopy: {
3022 // Step 1.1: Get the address for the src element in the Reduce list.
3023 Address SrcElementPtrAddr = Bld.CreateConstArrayGEP(SrcBase, Idx);
3024 SrcElementAddr = CGF.EmitLoadOfPointer(
3025 SrcElementPtrAddr,
3026 C.getPointerType(Private->getType())->castAs<PointerType>());
3027
3028 // Step 1.2: Get the address for dest element. The destination
3029 // element has already been created on the thread's stack.
3030 DestElementPtrAddr = Bld.CreateConstArrayGEP(DestBase, Idx);
3031 DestElementAddr = CGF.EmitLoadOfPointer(
3032 DestElementPtrAddr,
3033 C.getPointerType(Private->getType())->castAs<PointerType>());
3034 break;
3035 }
3036 case ThreadToScratchpad: {
3037 // Step 1.1: Get the address for the src element in the Reduce list.
3038 Address SrcElementPtrAddr = Bld.CreateConstArrayGEP(SrcBase, Idx);
3039 SrcElementAddr = CGF.EmitLoadOfPointer(
3040 SrcElementPtrAddr,
3041 C.getPointerType(Private->getType())->castAs<PointerType>());
3042
3043 // Step 1.2: Get the address for dest element:
3044 // address = base + index * ElementSizeInChars.
3045 llvm::Value *ElementSizeInChars = CGF.getTypeSize(Private->getType());
3046 llvm::Value *CurrentOffset =
3047 Bld.CreateNUWMul(ElementSizeInChars, ScratchpadIndex);
3048 llvm::Value *ScratchPadElemAbsolutePtrVal =
3049 Bld.CreateNUWAdd(DestBase.getPointer(), CurrentOffset);
3050 ScratchPadElemAbsolutePtrVal =
3051 Bld.CreateIntToPtr(ScratchPadElemAbsolutePtrVal, CGF.VoidPtrTy);
3052 DestElementAddr = Address(ScratchPadElemAbsolutePtrVal,
3053 C.getTypeAlignInChars(Private->getType()));
3054 IncrScratchpadDest = true;
3055 break;
3056 }
3057 case ScratchpadToThread: {
3058 // Step 1.1: Get the address for the src element in the scratchpad.
3059 // address = base + index * ElementSizeInChars.
3060 llvm::Value *ElementSizeInChars = CGF.getTypeSize(Private->getType());
3061 llvm::Value *CurrentOffset =
3062 Bld.CreateNUWMul(ElementSizeInChars, ScratchpadIndex);
3063 llvm::Value *ScratchPadElemAbsolutePtrVal =
3064 Bld.CreateNUWAdd(SrcBase.getPointer(), CurrentOffset);
3065 ScratchPadElemAbsolutePtrVal =
3066 Bld.CreateIntToPtr(ScratchPadElemAbsolutePtrVal, CGF.VoidPtrTy);
3067 SrcElementAddr = Address(ScratchPadElemAbsolutePtrVal,
3068 C.getTypeAlignInChars(Private->getType()));
3069 IncrScratchpadSrc = true;
3070
3071 // Step 1.2: Create a temporary to store the element in the destination
3072 // Reduce list.
3073 DestElementPtrAddr = Bld.CreateConstArrayGEP(DestBase, Idx);
3074 DestElementAddr =
3075 CGF.CreateMemTemp(Private->getType(), ".omp.reduction.element");
3076 UpdateDestListPtr = true;
3077 break;
3078 }
3079 }
3080
3081 // Regardless of src and dest of copy, we emit the load of src
3082 // element as this is required in all directions
3083 SrcElementAddr = Bld.CreateElementBitCast(
3084 SrcElementAddr, CGF.ConvertTypeForMem(Private->getType()));
3085 DestElementAddr = Bld.CreateElementBitCast(DestElementAddr,
3086 SrcElementAddr.getElementType());
3087
3088 // Now that all active lanes have read the element in the
3089 // Reduce list, shuffle over the value from the remote lane.
3090 if (ShuffleInElement) {
3091 shuffleAndStore(CGF, SrcElementAddr, DestElementAddr, Private->getType(),
3092 RemoteLaneOffset, Private->getExprLoc());
3093 } else {
3094 switch (CGF.getEvaluationKind(Private->getType())) {
3095 case TEK_Scalar: {
3096 llvm::Value *Elem =
3097 CGF.EmitLoadOfScalar(SrcElementAddr, /*Volatile=*/false,
3098 Private->getType(), Private->getExprLoc());
3099 // Store the source element value to the dest element address.
3100 CGF.EmitStoreOfScalar(Elem, DestElementAddr, /*Volatile=*/false,
3101 Private->getType());
3102 break;
3103 }
3104 case TEK_Complex: {
3105 CodeGenFunction::ComplexPairTy Elem = CGF.EmitLoadOfComplex(
3106 CGF.MakeAddrLValue(SrcElementAddr, Private->getType()),
3107 Private->getExprLoc());
3108 CGF.EmitStoreOfComplex(
3109 Elem, CGF.MakeAddrLValue(DestElementAddr, Private->getType()),
3110 /*isInit=*/false);
3111 break;
3112 }
3113 case TEK_Aggregate:
3114 CGF.EmitAggregateCopy(
3115 CGF.MakeAddrLValue(DestElementAddr, Private->getType()),
3116 CGF.MakeAddrLValue(SrcElementAddr, Private->getType()),
3117 Private->getType(), AggValueSlot::DoesNotOverlap);
3118 break;
3119 }
3120 }
3121
3122 // Step 3.1: Modify reference in dest Reduce list as needed.
3123 // Modifying the reference in Reduce list to point to the newly
3124 // created element. The element is live in the current function
3125 // scope and that of functions it invokes (i.e., reduce_function).
3126 // RemoteReduceData[i] = (void*)&RemoteElem
3127 if (UpdateDestListPtr) {
3128 CGF.EmitStoreOfScalar(Bld.CreatePointerBitCastOrAddrSpaceCast(
3129 DestElementAddr.getPointer(), CGF.VoidPtrTy),
3130 DestElementPtrAddr, /*Volatile=*/false,
3131 C.VoidPtrTy);
3132 }
3133
3134 // Step 4.1: Increment SrcBase/DestBase so that it points to the starting
3135 // address of the next element in scratchpad memory, unless we're currently
3136 // processing the last one. Memory alignment is also taken care of here.
3137 if ((IncrScratchpadDest || IncrScratchpadSrc) && (Idx + 1 < Size)) {
3138 llvm::Value *ScratchpadBasePtr =
3139 IncrScratchpadDest ? DestBase.getPointer() : SrcBase.getPointer();
3140 llvm::Value *ElementSizeInChars = CGF.getTypeSize(Private->getType());
3141 ScratchpadBasePtr = Bld.CreateNUWAdd(
3142 ScratchpadBasePtr,
3143 Bld.CreateNUWMul(ScratchpadWidth, ElementSizeInChars));
3144
3145 // Take care of global memory alignment for performance
3146 ScratchpadBasePtr = Bld.CreateNUWSub(
3147 ScratchpadBasePtr, llvm::ConstantInt::get(CGM.SizeTy, 1));
3148 ScratchpadBasePtr = Bld.CreateUDiv(
3149 ScratchpadBasePtr,
3150 llvm::ConstantInt::get(CGM.SizeTy, GlobalMemoryAlignment));
3151 ScratchpadBasePtr = Bld.CreateNUWAdd(
3152 ScratchpadBasePtr, llvm::ConstantInt::get(CGM.SizeTy, 1));
3153 ScratchpadBasePtr = Bld.CreateNUWMul(
3154 ScratchpadBasePtr,
3155 llvm::ConstantInt::get(CGM.SizeTy, GlobalMemoryAlignment));
3156
3157 if (IncrScratchpadDest)
3158 DestBase = Address(ScratchpadBasePtr, CGF.getPointerAlign());
3159 else /* IncrScratchpadSrc = true */
3160 SrcBase = Address(ScratchpadBasePtr, CGF.getPointerAlign());
3161 }
3162
3163 ++Idx;
3164 }
3165}
3166
3167/// This function emits a helper that gathers Reduce lists from the first
3168/// lane of every active warp to lanes in the first warp.
3169///
3170/// void inter_warp_copy_func(void* reduce_data, num_warps)
3171/// shared smem[warp_size];
3172/// For all data entries D in reduce_data:
3173/// sync
3174/// If (I am the first lane in each warp)
3175/// Copy my local D to smem[warp_id]
3176/// sync
3177/// if (I am the first warp)
3178/// Copy smem[thread_id] to my local D
3179static llvm::Value *emitInterWarpCopyFunction(CodeGenModule &CGM,
3180 ArrayRef<const Expr *> Privates,
3181 QualType ReductionArrayTy,
3182 SourceLocation Loc) {
3183 ASTContext &C = CGM.getContext();
3184 llvm::Module &M = CGM.getModule();
3185
3186 // ReduceList: thread local Reduce list.
3187 // At the stage of the computation when this function is called, partially
3188 // aggregated values reside in the first lane of every active warp.
3189 ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
3190 C.VoidPtrTy, ImplicitParamDecl::Other);
3191 // NumWarps: number of warps active in the parallel region. This could
3192 // be smaller than 32 (max warps in a CTA) for partial block reduction.
3193 ImplicitParamDecl NumWarpsArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
3194 C.getIntTypeForBitwidth(32, /* Signed */ true),
3195 ImplicitParamDecl::Other);
3196 FunctionArgList Args;
3197 Args.push_back(&ReduceListArg);
3198 Args.push_back(&NumWarpsArg);
3199
3200 const CGFunctionInfo &CGFI =
3201 CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
3202 auto *Fn = llvm::Function::Create(CGM.getTypes().GetFunctionType(CGFI),
3203 llvm::GlobalValue::InternalLinkage,
3204 "_omp_reduction_inter_warp_copy_func", &M);
3205 CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
3206 Fn->setDoesNotRecurse();
3207 CodeGenFunction CGF(CGM);
3208 CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc);
3209
3210 CGBuilderTy &Bld = CGF.Builder;
3211
3212 // This array is used as a medium to transfer, one reduce element at a time,
3213 // the data from the first lane of every warp to lanes in the first warp
3214 // in order to perform the final step of a reduction in a parallel region
3215 // (reduction across warps). The array is placed in NVPTX __shared__ memory
3216 // for reduced latency, as well as to have a distinct copy for concurrently
3217 // executing target regions. The array is declared with common linkage so
3218 // as to be shared across compilation units.
3219 StringRef TransferMediumName =
3220 "__openmp_nvptx_data_transfer_temporary_storage";
3221 llvm::GlobalVariable *TransferMedium =
3222 M.getGlobalVariable(TransferMediumName);
3223 if (!TransferMedium) {
3224 auto *Ty = llvm::ArrayType::get(CGM.Int32Ty, WarpSize);
3225 unsigned SharedAddressSpace = C.getTargetAddressSpace(LangAS::cuda_shared);
3226 TransferMedium = new llvm::GlobalVariable(
3227 M, Ty, /*isConstant=*/false, llvm::GlobalVariable::CommonLinkage,
3228 llvm::Constant::getNullValue(Ty), TransferMediumName,
3229 /*InsertBefore=*/nullptr, llvm::GlobalVariable::NotThreadLocal,
3230 SharedAddressSpace);
3231 CGM.addCompilerUsedGlobal(TransferMedium);
3232 }
3233
3234 // Get the CUDA thread id of the current OpenMP thread on the GPU.
3235 llvm::Value *ThreadID = getNVPTXThreadID(CGF);
3236 // nvptx_lane_id = nvptx_id % warpsize
3237 llvm::Value *LaneID = getNVPTXLaneID(CGF);
3238 // nvptx_warp_id = nvptx_id / warpsize
3239 llvm::Value *WarpID = getNVPTXWarpID(CGF);
3240
3241 Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg);
3242 Address LocalReduceList(
3243 Bld.CreatePointerBitCastOrAddrSpaceCast(
3244 CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false,
3245 C.VoidPtrTy, Loc),
3246 CGF.ConvertTypeForMem(ReductionArrayTy)->getPointerTo()),
3247 CGF.getPointerAlign());
3248
3249 unsigned Idx = 0;
3250 for (const Expr *Private : Privates) {
3251 //
3252 // Warp master copies reduce element to transfer medium in __shared__
3253 // memory.
3254 //
3255 unsigned RealTySize =
3256 C.getTypeSizeInChars(Private->getType())
3257 .alignTo(C.getTypeAlignInChars(Private->getType()))
3258 .getQuantity();
3259 for (unsigned TySize = 4; TySize > 0 && RealTySize > 0; TySize /=2) {
3260 unsigned NumIters = RealTySize / TySize;
3261 if (NumIters == 0)
3262 continue;
3263 QualType CType = C.getIntTypeForBitwidth(
3264 C.toBits(CharUnits::fromQuantity(TySize)), /*Signed=*/1);
3265 llvm::Type *CopyType = CGF.ConvertTypeForMem(CType);
3266 CharUnits Align = CharUnits::fromQuantity(TySize);
3267 llvm::Value *Cnt = nullptr;
3268 Address CntAddr = Address::invalid();
3269 llvm::BasicBlock *PrecondBB = nullptr;
3270 llvm::BasicBlock *ExitBB = nullptr;
3271 if (NumIters > 1) {
3272 CntAddr = CGF.CreateMemTemp(C.IntTy, ".cnt.addr");
3273 CGF.EmitStoreOfScalar(llvm::Constant::getNullValue(CGM.IntTy), CntAddr,
3274 /*Volatile=*/false, C.IntTy);
3275 PrecondBB = CGF.createBasicBlock("precond");
3276 ExitBB = CGF.createBasicBlock("exit");
3277 llvm::BasicBlock *BodyBB = CGF.createBasicBlock("body");
3278 // There is no need to emit line number for unconditional branch.
3279 (void)ApplyDebugLocation::CreateEmpty(CGF);
3280 CGF.EmitBlock(PrecondBB);
3281 Cnt = CGF.EmitLoadOfScalar(CntAddr, /*Volatile=*/false, C.IntTy, Loc);
3282 llvm::Value *Cmp =
3283 Bld.CreateICmpULT(Cnt, llvm::ConstantInt::get(CGM.IntTy, NumIters));
3284 Bld.CreateCondBr(Cmp, BodyBB, ExitBB);
3285 CGF.EmitBlock(BodyBB);
3286 }
3287 // kmpc_barrier.
3288 CGM.getOpenMPRuntime().emitBarrierCall(CGF, Loc, OMPD_unknown,
3289 /*EmitChecks=*/false,
3290 /*ForceSimpleCall=*/true);
3291 llvm::BasicBlock *ThenBB = CGF.createBasicBlock("then");
3292 llvm::BasicBlock *ElseBB = CGF.createBasicBlock("else");
3293 llvm::BasicBlock *MergeBB = CGF.createBasicBlock("ifcont");
3294
3295 // if (lane_id == 0)
3296 llvm::Value *IsWarpMaster = Bld.CreateIsNull(LaneID, "warp_master");
3297 Bld.CreateCondBr(IsWarpMaster, ThenBB, ElseBB);
3298 CGF.EmitBlock(ThenBB);
3299
3300 // Reduce element = LocalReduceList[i]
3301 Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx);
3302 llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar(
3303 ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation());
3304 // elemptr = ((CopyType*)(elemptrptr)) + I
3305 Address ElemPtr = Address(ElemPtrPtr, Align);
3306 ElemPtr = Bld.CreateElementBitCast(ElemPtr, CopyType);
3307 if (NumIters > 1) {
3308 ElemPtr = Address(Bld.CreateGEP(ElemPtr.getPointer(), Cnt),
3309 ElemPtr.getAlignment());
3310 }
3311
3312 // Get pointer to location in transfer medium.
3313 // MediumPtr = &medium[warp_id]
3314 llvm::Value *MediumPtrVal = Bld.CreateInBoundsGEP(
3315 TransferMedium, {llvm::Constant::getNullValue(CGM.Int64Ty), WarpID});
3316 Address MediumPtr(MediumPtrVal, Align);
3317 // Casting to actual data type.
3318 // MediumPtr = (CopyType*)MediumPtrAddr;
3319 MediumPtr = Bld.CreateElementBitCast(MediumPtr, CopyType);
3320
3321 // elem = *elemptr
3322 //*MediumPtr = elem
3323 llvm::Value *Elem =
3324 CGF.EmitLoadOfScalar(ElemPtr, /*Volatile=*/false, CType, Loc);
3325 // Store the source element value to the dest element address.
3326 CGF.EmitStoreOfScalar(Elem, MediumPtr, /*Volatile=*/true, CType);
3327
3328 Bld.CreateBr(MergeBB);
3329
3330 CGF.EmitBlock(ElseBB);
3331 Bld.CreateBr(MergeBB);
3332
3333 CGF.EmitBlock(MergeBB);
3334
3335 // kmpc_barrier.
3336 CGM.getOpenMPRuntime().emitBarrierCall(CGF, Loc, OMPD_unknown,
3337 /*EmitChecks=*/false,
3338 /*ForceSimpleCall=*/true);
3339
3340 //
3341 // Warp 0 copies reduce element from transfer medium.
3342 //
3343 llvm::BasicBlock *W0ThenBB = CGF.createBasicBlock("then");
3344 llvm::BasicBlock *W0ElseBB = CGF.createBasicBlock("else");
3345 llvm::BasicBlock *W0MergeBB = CGF.createBasicBlock("ifcont");
3346
3347 Address AddrNumWarpsArg = CGF.GetAddrOfLocalVar(&NumWarpsArg);
3348 llvm::Value *NumWarpsVal = CGF.EmitLoadOfScalar(
3349 AddrNumWarpsArg, /*Volatile=*/false, C.IntTy, Loc);
3350
3351 // Up to 32 threads in warp 0 are active.
3352 llvm::Value *IsActiveThread =
3353 Bld.CreateICmpULT(ThreadID, NumWarpsVal, "is_active_thread");
3354 Bld.CreateCondBr(IsActiveThread, W0ThenBB, W0ElseBB);
3355
3356 CGF.EmitBlock(W0ThenBB);
3357
3358 // SrcMediumPtr = &medium[tid]
3359 llvm::Value *SrcMediumPtrVal = Bld.CreateInBoundsGEP(
3360 TransferMedium,
3361 {llvm::Constant::getNullValue(CGM.Int64Ty), ThreadID});
3362 Address SrcMediumPtr(SrcMediumPtrVal, Align);
3363 // SrcMediumVal = *SrcMediumPtr;
3364 SrcMediumPtr = Bld.CreateElementBitCast(SrcMediumPtr, CopyType);
3365
3366 // TargetElemPtr = (CopyType*)(SrcDataAddr[i]) + I
3367 Address TargetElemPtrPtr = Bld.CreateConstArrayGEP(LocalReduceList, Idx);
3368 llvm::Value *TargetElemPtrVal = CGF.EmitLoadOfScalar(
3369 TargetElemPtrPtr, /*Volatile=*/false, C.VoidPtrTy, Loc);
3370 Address TargetElemPtr = Address(TargetElemPtrVal, Align);
3371 TargetElemPtr = Bld.CreateElementBitCast(TargetElemPtr, CopyType);
3372 if (NumIters > 1) {
3373 TargetElemPtr = Address(Bld.CreateGEP(TargetElemPtr.getPointer(), Cnt),
3374 TargetElemPtr.getAlignment());
3375 }
3376
3377 // *TargetElemPtr = SrcMediumVal;
3378 llvm::Value *SrcMediumValue =
3379 CGF.EmitLoadOfScalar(SrcMediumPtr, /*Volatile=*/true, CType, Loc);
3380 CGF.EmitStoreOfScalar(SrcMediumValue, TargetElemPtr, /*Volatile=*/false,
3381 CType);
3382 Bld.CreateBr(W0MergeBB);
3383
3384 CGF.EmitBlock(W0ElseBB);
3385 Bld.CreateBr(W0MergeBB);
3386
3387 CGF.EmitBlock(W0MergeBB);
3388
3389 if (NumIters > 1) {
3390 Cnt = Bld.CreateNSWAdd(Cnt, llvm::ConstantInt::get(CGM.IntTy, /*V=*/1));
3391 CGF.EmitStoreOfScalar(Cnt, CntAddr, /*Volatile=*/false, C.IntTy);
3392 CGF.EmitBranch(PrecondBB);
3393 (void)ApplyDebugLocation::CreateEmpty(CGF);
3394 CGF.EmitBlock(ExitBB);
3395 }
3396 RealTySize %= TySize;
3397 }
3398 ++Idx;
3399 }
3400
3401 CGF.FinishFunction();
3402 return Fn;
3403}
3404
3405/// Emit a helper that reduces data across two OpenMP threads (lanes)
3406/// in the same warp. It uses shuffle instructions to copy over data from
3407/// a remote lane's stack. The reduction algorithm performed is specified
3408/// by the fourth parameter.
3409///
3410/// Algorithm Versions.
3411/// Full Warp Reduce (argument value 0):
3412/// This algorithm assumes that all 32 lanes are active and gathers
3413/// data from these 32 lanes, producing a single resultant value.
3414/// Contiguous Partial Warp Reduce (argument value 1):
3415/// This algorithm assumes that only a *contiguous* subset of lanes
3416/// are active. This happens for the last warp in a parallel region
3417/// when the user specified num_threads is not an integer multiple of
3418/// 32. This contiguous subset always starts with the zeroth lane.
3419/// Partial Warp Reduce (argument value 2):
3420/// This algorithm gathers data from any number of lanes at any position.
3421/// All reduced values are stored in the lowest possible lane. The set
3422/// of problems every algorithm addresses is a super set of those
3423/// addressable by algorithms with a lower version number. Overhead
3424/// increases as algorithm version increases.
3425///
3426/// Terminology
3427/// Reduce element:
3428/// Reduce element refers to the individual data field with primitive
3429/// data types to be combined and reduced across threads.
3430/// Reduce list:
3431/// Reduce list refers to a collection of local, thread-private
3432/// reduce elements.
3433/// Remote Reduce list:
3434/// Remote Reduce list refers to a collection of remote (relative to
3435/// the current thread) reduce elements.
3436///
3437/// We distinguish between three states of threads that are important to
3438/// the implementation of this function.
3439/// Alive threads:
3440/// Threads in a warp executing the SIMT instruction, as distinguished from
3441/// threads that are inactive due to divergent control flow.
3442/// Active threads:
3443/// The minimal set of threads that has to be alive upon entry to this
3444/// function. The computation is correct iff active threads are alive.
3445/// Some threads are alive but they are not active because they do not
3446/// contribute to the computation in any useful manner. Turning them off
3447/// may introduce control flow overheads without any tangible benefits.
3448/// Effective threads:
3449/// In order to comply with the argument requirements of the shuffle
3450/// function, we must keep all lanes holding data alive. But at most
3451/// half of them perform value aggregation; we refer to this half of
3452/// threads as effective. The other half is simply handing off their
3453/// data.
3454///
3455/// Procedure
3456/// Value shuffle:
3457/// In this step active threads transfer data from higher lane positions
3458/// in the warp to lower lane positions, creating Remote Reduce list.
3459/// Value aggregation:
3460/// In this step, effective threads combine their thread local Reduce list
3461/// with Remote Reduce list and store the result in the thread local
3462/// Reduce list.
3463/// Value copy:
3464/// In this step, we deal with the assumption made by algorithm 2
3465/// (i.e. contiguity assumption). When we have an odd number of lanes
3466/// active, say 2k+1, only k threads will be effective and therefore k
3467/// new values will be produced. However, the Reduce list owned by the
3468/// (2k+1)th thread is ignored in the value aggregation. Therefore
3469/// we copy the Reduce list from the (2k+1)th lane to (k+1)th lane so
3470/// that the contiguity assumption still holds.
3471static llvm::Function *emitShuffleAndReduceFunction(
3472 CodeGenModule &CGM, ArrayRef<const Expr *> Privates,
3473 QualType ReductionArrayTy, llvm::Function *ReduceFn, SourceLocation Loc) {
3474 ASTContext &C = CGM.getContext();
3475
3476 // Thread local Reduce list used to host the values of data to be reduced.
3477 ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
3478 C.VoidPtrTy, ImplicitParamDecl::Other);
3479 // Current lane id; could be logical.
3480 ImplicitParamDecl LaneIDArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.ShortTy,
3481 ImplicitParamDecl::Other);
3482 // Offset of the remote source lane relative to the current lane.
3483 ImplicitParamDecl RemoteLaneOffsetArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
3484 C.ShortTy, ImplicitParamDecl::Other);
3485 // Algorithm version. This is expected to be known at compile time.
3486 ImplicitParamDecl AlgoVerArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
3487 C.ShortTy, ImplicitParamDecl::Other);
3488 FunctionArgList Args;
3489 Args.push_back(&ReduceListArg);
3490 Args.push_back(&LaneIDArg);
3491 Args.push_back(&RemoteLaneOffsetArg);
3492 Args.push_back(&AlgoVerArg);
3493
3494 const CGFunctionInfo &CGFI =
3495 CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
3496 auto *Fn = llvm::Function::Create(
3497 CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
3498 "_omp_reduction_shuffle_and_reduce_func", &CGM.getModule());
3499 CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
3500 Fn->setDoesNotRecurse();
3501 if (CGM.getLangOpts().Optimize) {
3502 Fn->removeFnAttr(llvm::Attribute::NoInline);
3503 Fn->removeFnAttr(llvm::Attribute::OptimizeNone);
3504 Fn->addFnAttr(llvm::Attribute::AlwaysInline);
3505 }
3506
3507 CodeGenFunction CGF(CGM);
3508 CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc);
3509
3510 CGBuilderTy &Bld = CGF.Builder;
3511
3512 Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg);
3513 Address LocalReduceList(
3514 Bld.CreatePointerBitCastOrAddrSpaceCast(
3515 CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false,
3516 C.VoidPtrTy, SourceLocation()),
3517 CGF.ConvertTypeForMem(ReductionArrayTy)->getPointerTo()),
3518 CGF.getPointerAlign());
3519
3520 Address AddrLaneIDArg = CGF.GetAddrOfLocalVar(&LaneIDArg);
3521 llvm::Value *LaneIDArgVal = CGF.EmitLoadOfScalar(
3522 AddrLaneIDArg, /*Volatile=*/false, C.ShortTy, SourceLocation());
3523
3524 Address AddrRemoteLaneOffsetArg = CGF.GetAddrOfLocalVar(&RemoteLaneOffsetArg);
3525 llvm::Value *RemoteLaneOffsetArgVal = CGF.EmitLoadOfScalar(
3526 AddrRemoteLaneOffsetArg, /*Volatile=*/false, C.ShortTy, SourceLocation());
3527
3528 Address AddrAlgoVerArg = CGF.GetAddrOfLocalVar(&AlgoVerArg);
3529 llvm::Value *AlgoVerArgVal = CGF.EmitLoadOfScalar(
3530 AddrAlgoVerArg, /*Volatile=*/false, C.ShortTy, SourceLocation());
3531
3532 // Create a local thread-private variable to host the Reduce list
3533 // from a remote lane.
3534 Address RemoteReduceList =
3535 CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.remote_reduce_list");
3536
3537 // This loop iterates through the list of reduce elements and copies,
3538 // element by element, from a remote lane in the warp to RemoteReduceList,
3539 // hosted on the thread's stack.
3540 emitReductionListCopy(RemoteLaneToThread, CGF, ReductionArrayTy, Privates,
3541 LocalReduceList, RemoteReduceList,
3542 {/*RemoteLaneOffset=*/RemoteLaneOffsetArgVal,
3543 /*ScratchpadIndex=*/nullptr,
3544 /*ScratchpadWidth=*/nullptr});
3545
3546 // The actions to be performed on the Remote Reduce list is dependent
3547 // on the algorithm version.
3548 //
3549 // if (AlgoVer==0) || (AlgoVer==1 && (LaneId < Offset)) || (AlgoVer==2 &&
3550 // LaneId % 2 == 0 && Offset > 0):
3551 // do the reduction value aggregation
3552 //
3553 // The thread local variable Reduce list is mutated in place to host the
3554 // reduced data, which is the aggregated value produced from local and
3555 // remote lanes.
3556 //
3557 // Note that AlgoVer is expected to be a constant integer known at compile
3558 // time.
3559 // When AlgoVer==0, the first conjunction evaluates to true, making
3560 // the entire predicate true during compile time.
3561 // When AlgoVer==1, the second conjunction has only the second part to be
3562 // evaluated during runtime. Other conjunctions evaluates to false
3563 // during compile time.
3564 // When AlgoVer==2, the third conjunction has only the second part to be
3565 // evaluated during runtime. Other conjunctions evaluates to false
3566 // during compile time.
3567 llvm::Value *CondAlgo0 = Bld.CreateIsNull(AlgoVerArgVal);
3568
3569 llvm::Value *Algo1 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(1));
3570 llvm::Value *CondAlgo1 = Bld.CreateAnd(
3571 Algo1, Bld.CreateICmpULT(LaneIDArgVal, RemoteLaneOffsetArgVal));
3572
3573 llvm::Value *Algo2 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(2));
3574 llvm::Value *CondAlgo2 = Bld.CreateAnd(
3575 Algo2, Bld.CreateIsNull(Bld.CreateAnd(LaneIDArgVal, Bld.getInt16(1))));
3576 CondAlgo2 = Bld.CreateAnd(
3577 CondAlgo2, Bld.CreateICmpSGT(RemoteLaneOffsetArgVal, Bld.getInt16(0)));
3578
3579 llvm::Value *CondReduce = Bld.CreateOr(CondAlgo0, CondAlgo1);
3580 CondReduce = Bld.CreateOr(CondReduce, CondAlgo2);
3581
3582 llvm::BasicBlock *ThenBB = CGF.createBasicBlock("then");
3583 llvm::BasicBlock *ElseBB = CGF.createBasicBlock("else");
3584 llvm::BasicBlock *MergeBB = CGF.createBasicBlock("ifcont");
3585 Bld.CreateCondBr(CondReduce, ThenBB, ElseBB);
3586
3587 CGF.EmitBlock(ThenBB);
3588 // reduce_function(LocalReduceList, RemoteReduceList)
3589 llvm::Value *LocalReduceListPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
3590 LocalReduceList.getPointer(), CGF.VoidPtrTy);
3591 llvm::Value *RemoteReduceListPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
3592 RemoteReduceList.getPointer(), CGF.VoidPtrTy);
3593 CGM.getOpenMPRuntime().emitOutlinedFunctionCall(
3594 CGF, Loc, ReduceFn, {LocalReduceListPtr, RemoteReduceListPtr});
3595 Bld.CreateBr(MergeBB);
3596
3597 CGF.EmitBlock(ElseBB);
3598 Bld.CreateBr(MergeBB);
3599
3600 CGF.EmitBlock(MergeBB);
3601
3602 // if (AlgoVer==1 && (LaneId >= Offset)) copy Remote Reduce list to local
3603 // Reduce list.
3604 Algo1 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(1));
3605 llvm::Value *CondCopy = Bld.CreateAnd(
3606 Algo1, Bld.CreateICmpUGE(LaneIDArgVal, RemoteLaneOffsetArgVal));
3607
3608 llvm::BasicBlock *CpyThenBB = CGF.createBasicBlock("then");
3609 llvm::BasicBlock *CpyElseBB = CGF.createBasicBlock("else");
3610 llvm::BasicBlock *CpyMergeBB = CGF.createBasicBlock("ifcont");
3611 Bld.CreateCondBr(CondCopy, CpyThenBB, CpyElseBB);
3612
3613 CGF.EmitBlock(CpyThenBB);
3614 emitReductionListCopy(ThreadCopy, CGF, ReductionArrayTy, Privates,
3615 RemoteReduceList, LocalReduceList);
3616 Bld.CreateBr(CpyMergeBB);
3617
3618 CGF.EmitBlock(CpyElseBB);
3619 Bld.CreateBr(CpyMergeBB);
3620
3621 CGF.EmitBlock(CpyMergeBB);
3622
3623 CGF.FinishFunction();
3624 return Fn;
3625}
3626
3627/// This function emits a helper that copies all the reduction variables from
3628/// the team into the provided global buffer for the reduction variables.
3629///
3630/// void list_to_global_copy_func(void *buffer, int Idx, void *reduce_data)
3631/// For all data entries D in reduce_data:
3632/// Copy local D to buffer.D[Idx]
3633static llvm::Value *emitListToGlobalCopyFunction(
3634 CodeGenModule &CGM, ArrayRef<const Expr *> Privates,
3635 QualType ReductionArrayTy, SourceLocation Loc,
3636 const RecordDecl *TeamReductionRec,
3637 const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
3638 &VarFieldMap) {
3639 ASTContext &C = CGM.getContext();
3640
3641 // Buffer: global reduction buffer.
3642 ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
3643 C.VoidPtrTy, ImplicitParamDecl::Other);
3644 // Idx: index of the buffer.
3645 ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy,
3646 ImplicitParamDecl::Other);
3647 // ReduceList: thread local Reduce list.
3648 ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
3649 C.VoidPtrTy, ImplicitParamDecl::Other);
3650 FunctionArgList Args;
3651 Args.push_back(&BufferArg);
3652 Args.push_back(&IdxArg);
3653 Args.push_back(&ReduceListArg);
3654
3655 const CGFunctionInfo &CGFI =
3656 CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
3657 auto *Fn = llvm::Function::Create(
3658 CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
3659 "_omp_reduction_list_to_global_copy_func", &CGM.getModule());
3660 CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
3661 Fn->setDoesNotRecurse();
3662 CodeGenFunction CGF(CGM);
3663 CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc);
3664
3665 CGBuilderTy &Bld = CGF.Builder;
3666
3667 Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg);
3668 Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg);
3669 Address LocalReduceList(
3670 Bld.CreatePointerBitCastOrAddrSpaceCast(
3671 CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false,
3672 C.VoidPtrTy, Loc),
3673 CGF.ConvertTypeForMem(ReductionArrayTy)->getPointerTo()),
3674 CGF.getPointerAlign());
3675 QualType StaticTy = C.getRecordType(TeamReductionRec);
3676 llvm::Type *LLVMReductionsBufferTy =
3677 CGM.getTypes().ConvertTypeForMem(StaticTy);
3678 llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
3679 CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc),
3680 LLVMReductionsBufferTy->getPointerTo());
3681 llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty),
3682 CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg),
3683 /*Volatile=*/false, C.IntTy,
3684 Loc)};
3685 unsigned Idx = 0;
3686 for (const Expr *Private : Privates) {
3687 // Reduce element = LocalReduceList[i]
3688 Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx);
3689 llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar(
3690 ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation());
3691 // elemptr = ((CopyType*)(elemptrptr)) + I
3692 ElemPtrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
3693 ElemPtrPtr, CGF.ConvertTypeForMem(Private->getType())->getPointerTo());
3694 Address ElemPtr =
3695 Address(ElemPtrPtr, C.getTypeAlignInChars(Private->getType()));
3696 const ValueDecl *VD = cast<DeclRefExpr>(Private)->getDecl();
3697 // Global = Buffer.VD[Idx];
3698 const FieldDecl *FD = VarFieldMap.lookup(VD);
3699 LValue GlobLVal = CGF.EmitLValueForField(
3700 CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD);
3701 llvm::Value *BufferPtr =
3702 Bld.CreateInBoundsGEP(GlobLVal.getPointer(CGF), Idxs);
3703 GlobLVal.setAddress(Address(BufferPtr, GlobLVal.getAlignment()));
3704 switch (CGF.getEvaluationKind(Private->getType())) {
3705 case TEK_Scalar: {
3706 llvm::Value *V = CGF.EmitLoadOfScalar(ElemPtr, /*Volatile=*/false,
3707 Private->getType(), Loc);
3708 CGF.EmitStoreOfScalar(V, GlobLVal);
3709 break;
3710 }
3711 case TEK_Complex: {
3712 CodeGenFunction::ComplexPairTy V = CGF.EmitLoadOfComplex(
3713 CGF.MakeAddrLValue(ElemPtr, Private->getType()), Loc);
3714 CGF.EmitStoreOfComplex(V, GlobLVal, /*isInit=*/false);
3715 break;
3716 }
3717 case TEK_Aggregate:
3718 CGF.EmitAggregateCopy(GlobLVal,
3719 CGF.MakeAddrLValue(ElemPtr, Private->getType()),
3720 Private->getType(), AggValueSlot::DoesNotOverlap);
3721 break;
3722 }
3723 ++Idx;
3724 }
3725
3726 CGF.FinishFunction();
3727 return Fn;
3728}
3729
3730/// This function emits a helper that reduces all the reduction variables from
3731/// the team into the provided global buffer for the reduction variables.
3732///
3733/// void list_to_global_reduce_func(void *buffer, int Idx, void *reduce_data)
3734/// void *GlobPtrs[];
3735/// GlobPtrs[0] = (void*)&buffer.D0[Idx];
3736/// ...
3737/// GlobPtrs[N] = (void*)&buffer.DN[Idx];
3738/// reduce_function(GlobPtrs, reduce_data);
3739static llvm::Value *emitListToGlobalReduceFunction(
3740 CodeGenModule &CGM, ArrayRef<const Expr *> Privates,
3741 QualType ReductionArrayTy, SourceLocation Loc,
3742 const RecordDecl *TeamReductionRec,
3743 const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
3744 &VarFieldMap,
3745 llvm::Function *ReduceFn) {
3746 ASTContext &C = CGM.getContext();
3747
3748 // Buffer: global reduction buffer.
3749 ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
3750 C.VoidPtrTy, ImplicitParamDecl::Other);
3751 // Idx: index of the buffer.
3752 ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy,
3753 ImplicitParamDecl::Other);
3754 // ReduceList: thread local Reduce list.
3755 ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
3756 C.VoidPtrTy, ImplicitParamDecl::Other);
3757 FunctionArgList Args;
3758 Args.push_back(&BufferArg);
3759 Args.push_back(&IdxArg);
3760 Args.push_back(&ReduceListArg);
3761
3762 const CGFunctionInfo &CGFI =
3763 CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
3764 auto *Fn = llvm::Function::Create(
3765 CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
3766 "_omp_reduction_list_to_global_reduce_func", &CGM.getModule());
3767 CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
3768 Fn->setDoesNotRecurse();
3769 CodeGenFunction CGF(CGM);
3770 CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc);
3771
3772 CGBuilderTy &Bld = CGF.Builder;
3773
3774 Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg);
3775 QualType StaticTy = C.getRecordType(TeamReductionRec);
3776 llvm::Type *LLVMReductionsBufferTy =
3777 CGM.getTypes().ConvertTypeForMem(StaticTy);
3778 llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
3779 CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc),
3780 LLVMReductionsBufferTy->getPointerTo());
3781
3782 // 1. Build a list of reduction variables.
3783 // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]};
3784 Address ReductionList =
3785 CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list");
3786 auto IPriv = Privates.begin();
3787 llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty),
3788 CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg),
3789 /*Volatile=*/false, C.IntTy,
3790 Loc)};
3791 unsigned Idx = 0;
3792 for (unsigned I = 0, E = Privates.size(); I < E; ++I, ++IPriv, ++Idx) {
3793 Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx);
3794 // Global = Buffer.VD[Idx];
3795 const ValueDecl *VD = cast<DeclRefExpr>(*IPriv)->getDecl();
3796 const FieldDecl *FD = VarFieldMap.lookup(VD);
3797 LValue GlobLVal = CGF.EmitLValueForField(
3798 CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD);
3799 llvm::Value *BufferPtr =
3800 Bld.CreateInBoundsGEP(GlobLVal.getPointer(CGF), Idxs);
3801 llvm::Value *Ptr = CGF.EmitCastToVoidPtr(BufferPtr);
3802 CGF.EmitStoreOfScalar(Ptr, Elem, /*Volatile=*/false, C.VoidPtrTy);
3803 if ((*IPriv)->getType()->isVariablyModifiedType()) {
3804 // Store array size.
3805 ++Idx;
3806 Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx);
3807 llvm::Value *Size = CGF.Builder.CreateIntCast(
3808 CGF.getVLASize(
3809 CGF.getContext().getAsVariableArrayType((*IPriv)->getType()))
3810 .NumElts,
3811 CGF.SizeTy, /*isSigned=*/false);
3812 CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy),
3813 Elem);
3814 }
3815 }
3816
3817 // Call reduce_function(GlobalReduceList, ReduceList)
3818 llvm::Value *GlobalReduceList =
3819 CGF.EmitCastToVoidPtr(ReductionList.getPointer());
3820 Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg);
3821 llvm::Value *ReducedPtr = CGF.EmitLoadOfScalar(
3822 AddrReduceListArg, /*Volatile=*/false, C.VoidPtrTy, Loc);
3823 CGM.getOpenMPRuntime().emitOutlinedFunctionCall(
3824 CGF, Loc, ReduceFn, {GlobalReduceList, ReducedPtr});
3825 CGF.FinishFunction();
3826 return Fn;
3827}
3828
3829/// This function emits a helper that copies all the reduction variables from
3830/// the team into the provided global buffer for the reduction variables.
3831///
3832/// void list_to_global_copy_func(void *buffer, int Idx, void *reduce_data)
3833/// For all data entries D in reduce_data:
3834/// Copy buffer.D[Idx] to local D;
3835static llvm::Value *emitGlobalToListCopyFunction(
3836 CodeGenModule &CGM, ArrayRef<const Expr *> Privates,
3837 QualType ReductionArrayTy, SourceLocation Loc,
3838 const RecordDecl *TeamReductionRec,
3839 const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
3840 &VarFieldMap) {
3841 ASTContext &C = CGM.getContext();
3842
3843 // Buffer: global reduction buffer.
3844 ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
3845 C.VoidPtrTy, ImplicitParamDecl::Other);
3846 // Idx: index of the buffer.
3847 ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy,
3848 ImplicitParamDecl::Other);
3849 // ReduceList: thread local Reduce list.
3850 ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
3851 C.VoidPtrTy, ImplicitParamDecl::Other);
3852 FunctionArgList Args;
3853 Args.push_back(&BufferArg);
3854 Args.push_back(&IdxArg);
3855 Args.push_back(&ReduceListArg);
3856
3857 const CGFunctionInfo &CGFI =
3858 CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
3859 auto *Fn = llvm::Function::Create(
3860 CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
3861 "_omp_reduction_global_to_list_copy_func", &CGM.getModule());
3862 CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
3863 Fn->setDoesNotRecurse();
3864 CodeGenFunction CGF(CGM);
3865 CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc);
3866
3867 CGBuilderTy &Bld = CGF.Builder;
3868
3869 Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg);
3870 Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg);
3871 Address LocalReduceList(
3872 Bld.CreatePointerBitCastOrAddrSpaceCast(
3873 CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false,
3874 C.VoidPtrTy, Loc),
3875 CGF.ConvertTypeForMem(ReductionArrayTy)->getPointerTo()),
3876 CGF.getPointerAlign());
3877 QualType StaticTy = C.getRecordType(TeamReductionRec);
3878 llvm::Type *LLVMReductionsBufferTy =
3879 CGM.getTypes().ConvertTypeForMem(StaticTy);
3880 llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
3881 CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc),
3882 LLVMReductionsBufferTy->getPointerTo());
3883
3884 llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty),
3885 CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg),
3886 /*Volatile=*/false, C.IntTy,
3887 Loc)};
3888 unsigned Idx = 0;
3889 for (const Expr *Private : Privates) {
3890 // Reduce element = LocalReduceList[i]
3891 Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx);
3892 llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar(
3893 ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation());
3894 // elemptr = ((CopyType*)(elemptrptr)) + I
3895 ElemPtrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
3896 ElemPtrPtr, CGF.ConvertTypeForMem(Private->getType())->getPointerTo());
3897 Address ElemPtr =
3898 Address(ElemPtrPtr, C.getTypeAlignInChars(Private->getType()));
3899 const ValueDecl *VD = cast<DeclRefExpr>(Private)->getDecl();
3900 // Global = Buffer.VD[Idx];
3901 const FieldDecl *FD = VarFieldMap.lookup(VD);
3902 LValue GlobLVal = CGF.EmitLValueForField(
3903 CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD);
3904 llvm::Value *BufferPtr =
3905 Bld.CreateInBoundsGEP(GlobLVal.getPointer(CGF), Idxs);
3906 GlobLVal.setAddress(Address(BufferPtr, GlobLVal.getAlignment()));
3907 switch (CGF.getEvaluationKind(Private->getType())) {
3908 case TEK_Scalar: {
3909 llvm::Value *V = CGF.EmitLoadOfScalar(GlobLVal, Loc);
3910 CGF.EmitStoreOfScalar(V, ElemPtr, /*Volatile=*/false, Private->getType());
3911 break;
3912 }
3913 case TEK_Complex: {
3914 CodeGenFunction::ComplexPairTy V = CGF.EmitLoadOfComplex(GlobLVal, Loc);
3915 CGF.EmitStoreOfComplex(V, CGF.MakeAddrLValue(ElemPtr, Private->getType()),
3916 /*isInit=*/false);
3917 break;
3918 }
3919 case TEK_Aggregate:
3920 CGF.EmitAggregateCopy(CGF.MakeAddrLValue(ElemPtr, Private->getType()),
3921 GlobLVal, Private->getType(),
3922 AggValueSlot::DoesNotOverlap);
3923 break;
3924 }
3925 ++Idx;
3926 }
3927
3928 CGF.FinishFunction();
3929 return Fn;
3930}
3931
3932/// This function emits a helper that reduces all the reduction variables from
3933/// the team into the provided global buffer for the reduction variables.
3934///
3935/// void global_to_list_reduce_func(void *buffer, int Idx, void *reduce_data)
3936/// void *GlobPtrs[];
3937/// GlobPtrs[0] = (void*)&buffer.D0[Idx];
3938/// ...
3939/// GlobPtrs[N] = (void*)&buffer.DN[Idx];
3940/// reduce_function(reduce_data, GlobPtrs);
3941static llvm::Value *emitGlobalToListReduceFunction(
3942 CodeGenModule &CGM, ArrayRef<const Expr *> Privates,
3943 QualType ReductionArrayTy, SourceLocation Loc,
3944 const RecordDecl *TeamReductionRec,
3945 const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
3946 &VarFieldMap,
3947 llvm::Function *ReduceFn) {
3948 ASTContext &C = CGM.getContext();
3949
3950 // Buffer: global reduction buffer.
3951 ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
3952 C.VoidPtrTy, ImplicitParamDecl::Other);
3953 // Idx: index of the buffer.
3954 ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy,
3955 ImplicitParamDecl::Other);
3956 // ReduceList: thread local Reduce list.
3957 ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
3958 C.VoidPtrTy, ImplicitParamDecl::Other);
3959 FunctionArgList Args;
3960 Args.push_back(&BufferArg);
3961 Args.push_back(&IdxArg);
3962 Args.push_back(&ReduceListArg);
3963
3964 const CGFunctionInfo &CGFI =
3965 CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
3966 auto *Fn = llvm::Function::Create(
3967 CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
3968 "_omp_reduction_global_to_list_reduce_func", &CGM.getModule());
3969 CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
3970 Fn->setDoesNotRecurse();
3971 CodeGenFunction CGF(CGM);
3972 CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc);
3973
3974 CGBuilderTy &Bld = CGF.Builder;
3975
3976 Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg);
3977 QualType StaticTy = C.getRecordType(TeamReductionRec);
3978 llvm::Type *LLVMReductionsBufferTy =
3979 CGM.getTypes().ConvertTypeForMem(StaticTy);
3980 llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
3981 CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc),
3982 LLVMReductionsBufferTy->getPointerTo());
3983
3984 // 1. Build a list of reduction variables.
3985 // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]};
3986 Address ReductionList =
3987 CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list");
3988 auto IPriv = Privates.begin();
3989 llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty),
3990 CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg),
3991 /*Volatile=*/false, C.IntTy,
3992 Loc)};
3993 unsigned Idx = 0;
3994 for (unsigned I = 0, E = Privates.size(); I < E; ++I, ++IPriv, ++Idx) {
3995 Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx);
3996 // Global = Buffer.VD[Idx];
3997 const ValueDecl *VD = cast<DeclRefExpr>(*IPriv)->getDecl();
3998 const FieldDecl *FD = VarFieldMap.lookup(VD);
3999 LValue GlobLVal = CGF.EmitLValueForField(
4000 CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD);
4001 llvm::Value *BufferPtr =
4002 Bld.CreateInBoundsGEP(GlobLVal.getPointer(CGF), Idxs);
4003 llvm::Value *Ptr = CGF.EmitCastToVoidPtr(BufferPtr);
4004 CGF.EmitStoreOfScalar(Ptr, Elem, /*Volatile=*/false, C.VoidPtrTy);
4005 if ((*IPriv)->getType()->isVariablyModifiedType()) {
4006 // Store array size.
4007 ++Idx;
4008 Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx);
4009 llvm::Value *Size = CGF.Builder.CreateIntCast(
4010 CGF.getVLASize(
4011 CGF.getContext().getAsVariableArrayType((*IPriv)->getType()))
4012 .NumElts,
4013 CGF.SizeTy, /*isSigned=*/false);
4014 CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy),
4015 Elem);
4016 }
4017 }
4018
4019 // Call reduce_function(ReduceList, GlobalReduceList)
4020 llvm::Value *GlobalReduceList =
4021 CGF.EmitCastToVoidPtr(ReductionList.getPointer());
4022 Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg);
4023 llvm::Value *ReducedPtr = CGF.EmitLoadOfScalar(
4024 AddrReduceListArg, /*Volatile=*/false, C.VoidPtrTy, Loc);
4025 CGM.getOpenMPRuntime().emitOutlinedFunctionCall(
4026 CGF, Loc, ReduceFn, {ReducedPtr, GlobalReduceList});
4027 CGF.FinishFunction();
4028 return Fn;
4029}
4030
4031///
4032/// Design of OpenMP reductions on the GPU
4033///
4034/// Consider a typical OpenMP program with one or more reduction
4035/// clauses:
4036///
4037/// float foo;
4038/// double bar;
4039/// #pragma omp target teams distribute parallel for \
4040/// reduction(+:foo) reduction(*:bar)
4041/// for (int i = 0; i < N; i++) {
4042/// foo += A[i]; bar *= B[i];
4043/// }
4044///
4045/// where 'foo' and 'bar' are reduced across all OpenMP threads in
4046/// all teams. In our OpenMP implementation on the NVPTX device an
4047/// OpenMP team is mapped to a CUDA threadblock and OpenMP threads
4048/// within a team are mapped to CUDA threads within a threadblock.
4049/// Our goal is to efficiently aggregate values across all OpenMP
4050/// threads such that:
4051///
4052/// - the compiler and runtime are logically concise, and
4053/// - the reduction is performed efficiently in a hierarchical
4054/// manner as follows: within OpenMP threads in the same warp,
4055/// across warps in a threadblock, and finally across teams on
4056/// the NVPTX device.
4057///
4058/// Introduction to Decoupling
4059///
4060/// We would like to decouple the compiler and the runtime so that the
4061/// latter is ignorant of the reduction variables (number, data types)
4062/// and the reduction operators. This allows a simpler interface
4063/// and implementation while still attaining good performance.
4064///
4065/// Pseudocode for the aforementioned OpenMP program generated by the
4066/// compiler is as follows:
4067///
4068/// 1. Create private copies of reduction variables on each OpenMP
4069/// thread: 'foo_private', 'bar_private'
4070/// 2. Each OpenMP thread reduces the chunk of 'A' and 'B' assigned
4071/// to it and writes the result in 'foo_private' and 'bar_private'
4072/// respectively.
4073/// 3. Call the OpenMP runtime on the GPU to reduce within a team
4074/// and store the result on the team master:
4075///
4076/// __kmpc_nvptx_parallel_reduce_nowait_v2(...,
4077/// reduceData, shuffleReduceFn, interWarpCpyFn)
4078///
4079/// where:
4080/// struct ReduceData {
4081/// double *foo;
4082/// double *bar;
4083/// } reduceData
4084/// reduceData.foo = &foo_private
4085/// reduceData.bar = &bar_private
4086///
4087/// 'shuffleReduceFn' and 'interWarpCpyFn' are pointers to two
4088/// auxiliary functions generated by the compiler that operate on
4089/// variables of type 'ReduceData'. They aid the runtime perform
4090/// algorithmic steps in a data agnostic manner.
4091///
4092/// 'shuffleReduceFn' is a pointer to a function that reduces data
4093/// of type 'ReduceData' across two OpenMP threads (lanes) in the
4094/// same warp. It takes the following arguments as input:
4095///
4096/// a. variable of type 'ReduceData' on the calling lane,
4097/// b. its lane_id,
4098/// c. an offset relative to the current lane_id to generate a
4099/// remote_lane_id. The remote lane contains the second
4100/// variable of type 'ReduceData' that is to be reduced.
4101/// d. an algorithm version parameter determining which reduction
4102/// algorithm to use.
4103///
4104/// 'shuffleReduceFn' retrieves data from the remote lane using
4105/// efficient GPU shuffle intrinsics and reduces, using the
4106/// algorithm specified by the 4th parameter, the two operands
4107/// element-wise. The result is written to the first operand.
4108///
4109/// Different reduction algorithms are implemented in different
4110/// runtime functions, all calling 'shuffleReduceFn' to perform
4111/// the essential reduction step. Therefore, based on the 4th
4112/// parameter, this function behaves slightly differently to
4113/// cooperate with the runtime to ensure correctness under
4114/// different circumstances.
4115///
4116/// 'InterWarpCpyFn' is a pointer to a function that transfers
4117/// reduced variables across warps. It tunnels, through CUDA
4118/// shared memory, the thread-private data of type 'ReduceData'
4119/// from lane 0 of each warp to a lane in the first warp.
4120/// 4. Call the OpenMP runtime on the GPU to reduce across teams.
4121/// The last team writes the global reduced value to memory.
4122///
4123/// ret = __kmpc_nvptx_teams_reduce_nowait(...,
4124/// reduceData, shuffleReduceFn, interWarpCpyFn,
4125/// scratchpadCopyFn, loadAndReduceFn)
4126///
4127/// 'scratchpadCopyFn' is a helper that stores reduced
4128/// data from the team master to a scratchpad array in
4129/// global memory.
4130///
4131/// 'loadAndReduceFn' is a helper that loads data from
4132/// the scratchpad array and reduces it with the input
4133/// operand.
4134///
4135/// These compiler generated functions hide address
4136/// calculation and alignment information from the runtime.
4137/// 5. if ret == 1:
4138/// The team master of the last team stores the reduced
4139/// result to the globals in memory.
4140/// foo += reduceData.foo; bar *= reduceData.bar
4141///
4142///
4143/// Warp Reduction Algorithms
4144///
4145/// On the warp level, we have three algorithms implemented in the
4146/// OpenMP runtime depending on the number of active lanes:
4147///
4148/// Full Warp Reduction
4149///
4150/// The reduce algorithm within a warp where all lanes are active
4151/// is implemented in the runtime as follows:
4152///
4153/// full_warp_reduce(void *reduce_data,
4154/// kmp_ShuffleReductFctPtr ShuffleReduceFn) {
4155/// for (int offset = WARPSIZE/2; offset > 0; offset /= 2)
4156/// ShuffleReduceFn(reduce_data, 0, offset, 0);
4157/// }
4158///
4159/// The algorithm completes in log(2, WARPSIZE) steps.
4160///
4161/// 'ShuffleReduceFn' is used here with lane_id set to 0 because it is
4162/// not used therefore we save instructions by not retrieving lane_id
4163/// from the corresponding special registers. The 4th parameter, which
4164/// represents the version of the algorithm being used, is set to 0 to
4165/// signify full warp reduction.
4166///
4167/// In this version, 'ShuffleReduceFn' behaves, per element, as follows:
4168///
4169/// #reduce_elem refers to an element in the local lane's data structure
4170/// #remote_elem is retrieved from a remote lane
4171/// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE);
4172/// reduce_elem = reduce_elem REDUCE_OP remote_elem;
4173///
4174/// Contiguous Partial Warp Reduction
4175///
4176/// This reduce algorithm is used within a warp where only the first
4177/// 'n' (n <= WARPSIZE) lanes are active. It is typically used when the
4178/// number of OpenMP threads in a parallel region is not a multiple of
4179/// WARPSIZE. The algorithm is implemented in the runtime as follows:
4180///
4181/// void
4182/// contiguous_partial_reduce(void *reduce_data,
4183/// kmp_ShuffleReductFctPtr ShuffleReduceFn,
4184/// int size, int lane_id) {
4185/// int curr_size;
4186/// int offset;
4187/// curr_size = size;
4188/// mask = curr_size/2;
4189/// while (offset>0) {
4190/// ShuffleReduceFn(reduce_data, lane_id, offset, 1);
4191/// curr_size = (curr_size+1)/2;
4192/// offset = curr_size/2;
4193/// }
4194/// }
4195///
4196/// In this version, 'ShuffleReduceFn' behaves, per element, as follows:
4197///
4198/// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE);
4199/// if (lane_id < offset)
4200/// reduce_elem = reduce_elem REDUCE_OP remote_elem
4201/// else
4202/// reduce_elem = remote_elem
4203///
4204/// This algorithm assumes that the data to be reduced are located in a
4205/// contiguous subset of lanes starting from the first. When there is
4206/// an odd number of active lanes, the data in the last lane is not
4207/// aggregated with any other lane's dat but is instead copied over.
4208///
4209/// Dispersed Partial Warp Reduction
4210///
4211/// This algorithm is used within a warp when any discontiguous subset of
4212/// lanes are active. It is used to implement the reduction operation
4213/// across lanes in an OpenMP simd region or in a nested parallel region.
4214///
4215/// void
4216/// dispersed_partial_reduce(void *reduce_data,
4217/// kmp_ShuffleReductFctPtr ShuffleReduceFn) {
4218/// int size, remote_id;
4219/// int logical_lane_id = number_of_active_lanes_before_me() * 2;
4220/// do {
4221/// remote_id = next_active_lane_id_right_after_me();
4222/// # the above function returns 0 of no active lane
4223/// # is present right after the current lane.
4224/// size = number_of_active_lanes_in_this_warp();
4225/// logical_lane_id /= 2;
4226/// ShuffleReduceFn(reduce_data, logical_lane_id,
4227/// remote_id-1-threadIdx.x, 2);
4228/// } while (logical_lane_id % 2 == 0 && size > 1);
4229/// }
4230///
4231/// There is no assumption made about the initial state of the reduction.
4232/// Any number of lanes (>=1) could be active at any position. The reduction
4233/// result is returned in the first active lane.
4234///
4235/// In this version, 'ShuffleReduceFn' behaves, per element, as follows:
4236///
4237/// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE);
4238/// if (lane_id % 2 == 0 && offset > 0)
4239/// reduce_elem = reduce_elem REDUCE_OP remote_elem
4240/// else
4241/// reduce_elem = remote_elem
4242///
4243///
4244/// Intra-Team Reduction
4245///
4246/// This function, as implemented in the runtime call
4247/// '__kmpc_nvptx_parallel_reduce_nowait_v2', aggregates data across OpenMP
4248/// threads in a team. It first reduces within a warp using the
4249/// aforementioned algorithms. We then proceed to gather all such
4250/// reduced values at the first warp.
4251///
4252/// The runtime makes use of the function 'InterWarpCpyFn', which copies
4253/// data from each of the "warp master" (zeroth lane of each warp, where
4254/// warp-reduced data is held) to the zeroth warp. This step reduces (in
4255/// a mathematical sense) the problem of reduction across warp masters in
4256/// a block to the problem of warp reduction.
4257///
4258///
4259/// Inter-Team Reduction
4260///
4261/// Once a team has reduced its data to a single value, it is stored in
4262/// a global scratchpad array. Since each team has a distinct slot, this
4263/// can be done without locking.
4264///
4265/// The last team to write to the scratchpad array proceeds to reduce the
4266/// scratchpad array. One or more workers in the last team use the helper
4267/// 'loadAndReduceDataFn' to load and reduce values from the array, i.e.,
4268/// the k'th worker reduces every k'th element.
4269///
4270/// Finally, a call is made to '__kmpc_nvptx_parallel_reduce_nowait_v2' to
4271/// reduce across workers and compute a globally reduced value.
4272///
4273void CGOpenMPRuntimeNVPTX::emitReduction(
4274 CodeGenFunction &CGF, SourceLocation Loc, ArrayRef<const Expr *> Privates,
4275 ArrayRef<const Expr *> LHSExprs, ArrayRef<const Expr *> RHSExprs,
4276 ArrayRef<const Expr *> ReductionOps, ReductionOptionsTy Options) {
4277 if (!CGF.HaveInsertPoint())
4278 return;
4279
4280 bool ParallelReduction = isOpenMPParallelDirective(Options.ReductionKind);
4281#ifndef NDEBUG
4282 bool TeamsReduction = isOpenMPTeamsDirective(Options.ReductionKind);
4283#endif
4284
4285 if (Options.SimpleReduction) {
4286 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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 4287, __PRETTY_FUNCTION__))
4287 "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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 4287, __PRETTY_FUNCTION__))
;
4288 CGOpenMPRuntime::emitReduction(CGF, Loc, Privates, LHSExprs, RHSExprs,
4289 ReductionOps, Options);
4290 return;
4291 }
4292
4293 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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 4294, __PRETTY_FUNCTION__))
4294 "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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 4294, __PRETTY_FUNCTION__))
;
4295
4296 // Build res = __kmpc_reduce{_nowait}(<gtid>, <n>, sizeof(RedList),
4297 // RedList, shuffle_reduce_func, interwarp_copy_func);
4298 // or
4299 // Build res = __kmpc_reduce_teams_nowait_simple(<loc>, <gtid>, <lck>);
4300 llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc);
4301 llvm::Value *ThreadId = getThreadID(CGF, Loc);
4302
4303 llvm::Value *Res;
4304 ASTContext &C = CGM.getContext();
4305 // 1. Build a list of reduction variables.
4306 // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]};
4307 auto Size = RHSExprs.size();
4308 for (const Expr *E : Privates) {
4309 if (E->getType()->isVariablyModifiedType())
4310 // Reserve place for array size.
4311 ++Size;
4312 }
4313 llvm::APInt ArraySize(/*unsigned int numBits=*/32, Size);
4314 QualType ReductionArrayTy =
4315 C.getConstantArrayType(C.VoidPtrTy, ArraySize, nullptr, ArrayType::Normal,
4316 /*IndexTypeQuals=*/0);
4317 Address ReductionList =
4318 CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list");
4319 auto IPriv = Privates.begin();
4320 unsigned Idx = 0;
4321 for (unsigned I = 0, E = RHSExprs.size(); I < E; ++I, ++IPriv, ++Idx) {
4322 Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx);
4323 CGF.Builder.CreateStore(
4324 CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
4325 CGF.EmitLValue(RHSExprs[I]).getPointer(CGF), CGF.VoidPtrTy),
4326 Elem);
4327 if ((*IPriv)->getType()->isVariablyModifiedType()) {
4328 // Store array size.
4329 ++Idx;
4330 Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx);
4331 llvm::Value *Size = CGF.Builder.CreateIntCast(
4332 CGF.getVLASize(
4333 CGF.getContext().getAsVariableArrayType((*IPriv)->getType()))
4334 .NumElts,
4335 CGF.SizeTy, /*isSigned=*/false);
4336 CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy),
4337 Elem);
4338 }
4339 }
4340
4341 llvm::Value *RL = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
4342 ReductionList.getPointer(), CGF.VoidPtrTy);
4343 llvm::Function *ReductionFn = emitReductionFunction(
4344 Loc, CGF.ConvertTypeForMem(ReductionArrayTy)->getPointerTo(), Privates,
4345 LHSExprs, RHSExprs, ReductionOps);
4346 llvm::Value *ReductionArrayTySize = CGF.getTypeSize(ReductionArrayTy);
4347 llvm::Function *ShuffleAndReduceFn = emitShuffleAndReduceFunction(
4348 CGM, Privates, ReductionArrayTy, ReductionFn, Loc);
4349 llvm::Value *InterWarpCopyFn =
4350 emitInterWarpCopyFunction(CGM, Privates, ReductionArrayTy, Loc);
4351
4352 if (ParallelReduction) {
4353 llvm::Value *Args[] = {RTLoc,
4354 ThreadId,
4355 CGF.Builder.getInt32(RHSExprs.size()),
4356 ReductionArrayTySize,
4357 RL,
4358 ShuffleAndReduceFn,
4359 InterWarpCopyFn};
4360
4361 Res = CGF.EmitRuntimeCall(
4362 createNVPTXRuntimeFunction(
4363 OMPRTL_NVPTX__kmpc_nvptx_parallel_reduce_nowait_v2),
4364 Args);
4365 } else {
4366 assert(TeamsReduction && "expected teams reduction.")((TeamsReduction && "expected teams reduction.") ? static_cast
<void> (0) : __assert_fail ("TeamsReduction && \"expected teams reduction.\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 4366, __PRETTY_FUNCTION__))
;
4367 llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> VarFieldMap;
4368 llvm::SmallVector<const ValueDecl *, 4> PrivatesReductions(Privates.size());
4369 int Cnt = 0;
4370 for (const Expr *DRE : Privates) {
4371 PrivatesReductions[Cnt] = cast<DeclRefExpr>(DRE)->getDecl();
4372 ++Cnt;
4373 }
4374 const RecordDecl *TeamReductionRec = ::buildRecordForGlobalizedVars(
4375 CGM.getContext(), PrivatesReductions, llvm::None, VarFieldMap,
4376 C.getLangOpts().OpenMPCUDAReductionBufNum);
4377 TeamsReductions.push_back(TeamReductionRec);
4378 if (!KernelTeamsReductionPtr) {
4379 KernelTeamsReductionPtr = new llvm::GlobalVariable(
4380 CGM.getModule(), CGM.VoidPtrTy, /*isConstant=*/true,
4381 llvm::GlobalValue::InternalLinkage, nullptr,
4382 "_openmp_teams_reductions_buffer_$_$ptr");
4383 }
4384 llvm::Value *GlobalBufferPtr = CGF.EmitLoadOfScalar(
4385 Address(KernelTeamsReductionPtr, CGM.getPointerAlign()),
4386 /*Volatile=*/false, C.getPointerType(C.VoidPtrTy), Loc);
4387 llvm::Value *GlobalToBufferCpyFn = ::emitListToGlobalCopyFunction(
4388 CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap);
4389 llvm::Value *GlobalToBufferRedFn = ::emitListToGlobalReduceFunction(
4390 CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap,
4391 ReductionFn);
4392 llvm::Value *BufferToGlobalCpyFn = ::emitGlobalToListCopyFunction(
4393 CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap);
4394 llvm::Value *BufferToGlobalRedFn = ::emitGlobalToListReduceFunction(
4395 CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap,
4396 ReductionFn);
4397
4398 llvm::Value *Args[] = {
4399 RTLoc,
4400 ThreadId,
4401 GlobalBufferPtr,
4402 CGF.Builder.getInt32(C.getLangOpts().OpenMPCUDAReductionBufNum),
4403 RL,
4404 ShuffleAndReduceFn,
4405 InterWarpCopyFn,
4406 GlobalToBufferCpyFn,
4407 GlobalToBufferRedFn,
4408 BufferToGlobalCpyFn,
4409 BufferToGlobalRedFn};
4410
4411 Res = CGF.EmitRuntimeCall(
4412 createNVPTXRuntimeFunction(
4413 OMPRTL_NVPTX__kmpc_nvptx_teams_reduce_nowait_v2),
4414 Args);
4415 }
4416
4417 // 5. Build if (res == 1)
4418 llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".omp.reduction.done");
4419 llvm::BasicBlock *ThenBB = CGF.createBasicBlock(".omp.reduction.then");
4420 llvm::Value *Cond = CGF.Builder.CreateICmpEQ(
4421 Res, llvm::ConstantInt::get(CGM.Int32Ty, /*V=*/1));
4422 CGF.Builder.CreateCondBr(Cond, ThenBB, ExitBB);
4423
4424 // 6. Build then branch: where we have reduced values in the master
4425 // thread in each team.
4426 // __kmpc_end_reduce{_nowait}(<gtid>);
4427 // break;
4428 CGF.EmitBlock(ThenBB);
4429
4430 // Add emission of __kmpc_end_reduce{_nowait}(<gtid>);
4431 auto &&CodeGen = [Privates, LHSExprs, RHSExprs, ReductionOps,
4432 this](CodeGenFunction &CGF, PrePostActionTy &Action) {
4433 auto IPriv = Privates.begin();
4434 auto ILHS = LHSExprs.begin();
4435 auto IRHS = RHSExprs.begin();
4436 for (const Expr *E : ReductionOps) {
4437 emitSingleReductionCombiner(CGF, E, *IPriv, cast<DeclRefExpr>(*ILHS),
4438 cast<DeclRefExpr>(*IRHS));
4439 ++IPriv;
4440 ++ILHS;
4441 ++IRHS;
4442 }
4443 };
4444 llvm::Value *EndArgs[] = {ThreadId};
4445 RegionCodeGenTy RCG(CodeGen);
4446 NVPTXActionTy Action(
4447 nullptr, llvm::None,
4448 createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_end_reduce_nowait),
4449 EndArgs);
4450 RCG.setAction(Action);
4451 RCG(CGF);
4452 // There is no need to emit line number for unconditional branch.
4453 (void)ApplyDebugLocation::CreateEmpty(CGF);
4454 CGF.EmitBlock(ExitBB, /*IsFinished=*/true);
4455}
4456
4457const VarDecl *
4458CGOpenMPRuntimeNVPTX::translateParameter(const FieldDecl *FD,
4459 const VarDecl *NativeParam) const {
4460 if (!NativeParam->getType()->isReferenceType())
4461 return NativeParam;
4462 QualType ArgType = NativeParam->getType();
4463 QualifierCollector QC;
4464 const Type *NonQualTy = QC.strip(ArgType);
4465 QualType PointeeTy = cast<ReferenceType>(NonQualTy)->getPointeeType();
4466 if (const auto *Attr = FD->getAttr<OMPCaptureKindAttr>()) {
4467 if (Attr->getCaptureKind() == OMPC_map) {
4468 PointeeTy = CGM.getContext().getAddrSpaceQualType(PointeeTy,
4469 LangAS::opencl_global);
4470 } else if (Attr->getCaptureKind() == OMPC_firstprivate &&
4471 PointeeTy.isConstant(CGM.getContext())) {
4472 PointeeTy = CGM.getContext().getAddrSpaceQualType(PointeeTy,
4473 LangAS::opencl_generic);
4474 }
4475 }
4476 ArgType = CGM.getContext().getPointerType(PointeeTy);
4477 QC.addRestrict();
4478 enum { NVPTX_local_addr = 5 };
4479 QC.addAddressSpace(getLangASFromTargetAS(NVPTX_local_addr));
4480 ArgType = QC.apply(CGM.getContext(), ArgType);
4481 if (isa<ImplicitParamDecl>(NativeParam))
4482 return ImplicitParamDecl::Create(
4483 CGM.getContext(), /*DC=*/nullptr, NativeParam->getLocation(),
4484 NativeParam->getIdentifier(), ArgType, ImplicitParamDecl::Other);
4485 return ParmVarDecl::Create(
4486 CGM.getContext(),
4487 const_cast<DeclContext *>(NativeParam->getDeclContext()),
4488 NativeParam->getBeginLoc(), NativeParam->getLocation(),
4489 NativeParam->getIdentifier(), ArgType,
4490 /*TInfo=*/nullptr, SC_None, /*DefArg=*/nullptr);
4491}
4492
4493Address
4494CGOpenMPRuntimeNVPTX::getParameterAddress(CodeGenFunction &CGF,
4495 const VarDecl *NativeParam,
4496 const VarDecl *TargetParam) const {
4497 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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 4499, __PRETTY_FUNCTION__))
4498 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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 4499, __PRETTY_FUNCTION__))
4499 "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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 4499, __PRETTY_FUNCTION__))
;
4500 Address LocalAddr = CGF.GetAddrOfLocalVar(TargetParam);
4501 QualType NativeParamType = NativeParam->getType();
4502 QualifierCollector QC;
4503 const Type *NonQualTy = QC.strip(NativeParamType);
4504 QualType NativePointeeTy = cast<ReferenceType>(NonQualTy)->getPointeeType();
4505 unsigned NativePointeeAddrSpace =
4506 CGF.getContext().getTargetAddressSpace(NativePointeeTy);
4507 QualType TargetTy = TargetParam->getType();
4508 llvm::Value *TargetAddr = CGF.EmitLoadOfScalar(
4509 LocalAddr, /*Volatile=*/false, TargetTy, SourceLocation());
4510 // First cast to generic.
4511 TargetAddr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
4512 TargetAddr, TargetAddr->getType()->getPointerElementType()->getPointerTo(
4513 /*AddrSpace=*/0));
4514 // Cast from generic to native address space.
4515 TargetAddr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
4516 TargetAddr, TargetAddr->getType()->getPointerElementType()->getPointerTo(
4517 NativePointeeAddrSpace));
4518 Address NativeParamAddr = CGF.CreateMemTemp(NativeParamType);
4519 CGF.EmitStoreOfScalar(TargetAddr, NativeParamAddr, /*Volatile=*/false,
4520 NativeParamType);
4521 return NativeParamAddr;
4522}
4523
4524void CGOpenMPRuntimeNVPTX::emitOutlinedFunctionCall(
4525 CodeGenFunction &CGF, SourceLocation Loc, llvm::FunctionCallee OutlinedFn,
4526 ArrayRef<llvm::Value *> Args) const {
4527 SmallVector<llvm::Value *, 4> TargetArgs;
4528 TargetArgs.reserve(Args.size());
4529 auto *FnType = OutlinedFn.getFunctionType();
4530 for (unsigned I = 0, E = Args.size(); I < E; ++I) {
4531 if (FnType->isVarArg() && FnType->getNumParams() <= I) {
4532 TargetArgs.append(std::next(Args.begin(), I), Args.end());
4533 break;
4534 }
4535 llvm::Type *TargetType = FnType->getParamType(I);
4536 llvm::Value *NativeArg = Args[I];
4537 if (!TargetType->isPointerTy()) {
4538 TargetArgs.emplace_back(NativeArg);
4539 continue;
4540 }
4541 llvm::Value *TargetArg = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
4542 NativeArg,
4543 NativeArg->getType()->getPointerElementType()->getPointerTo());
4544 TargetArgs.emplace_back(
4545 CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(TargetArg, TargetType));
4546 }
4547 CGOpenMPRuntime::emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, TargetArgs);
4548}
4549
4550/// Emit function which wraps the outline parallel region
4551/// and controls the arguments which are passed to this function.
4552/// The wrapper ensures that the outlined function is called
4553/// with the correct arguments when data is shared.
4554llvm::Function *CGOpenMPRuntimeNVPTX::createParallelDataSharingWrapper(
4555 llvm::Function *OutlinedParallelFn, const OMPExecutableDirective &D) {
4556 ASTContext &Ctx = CGM.getContext();
4557 const auto &CS = *D.getCapturedStmt(OMPD_parallel);
4558
4559 // Create a function that takes as argument the source thread.
4560 FunctionArgList WrapperArgs;
4561 QualType Int16QTy =
4562 Ctx.getIntTypeForBitwidth(/*DestWidth=*/16, /*Signed=*/false);
4563 QualType Int32QTy =
4564 Ctx.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/false);
4565 ImplicitParamDecl ParallelLevelArg(Ctx, /*DC=*/nullptr, D.getBeginLoc(),
4566 /*Id=*/nullptr, Int16QTy,
4567 ImplicitParamDecl::Other);
4568 ImplicitParamDecl WrapperArg(Ctx, /*DC=*/nullptr, D.getBeginLoc(),
4569 /*Id=*/nullptr, Int32QTy,
4570 ImplicitParamDecl::Other);
4571 WrapperArgs.emplace_back(&ParallelLevelArg);
4572 WrapperArgs.emplace_back(&WrapperArg);
4573
4574 const CGFunctionInfo &CGFI =
4575 CGM.getTypes().arrangeBuiltinFunctionDeclaration(Ctx.VoidTy, WrapperArgs);
4576
4577 auto *Fn = llvm::Function::Create(
4578 CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
4579 Twine(OutlinedParallelFn->getName(), "_wrapper"), &CGM.getModule());
4580 CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
4581 Fn->setLinkage(llvm::GlobalValue::InternalLinkage);
4582 Fn->setDoesNotRecurse();
4583
4584 CodeGenFunction CGF(CGM, /*suppressNewContext=*/true);
4585 CGF.StartFunction(GlobalDecl(), Ctx.VoidTy, Fn, CGFI, WrapperArgs,
4586 D.getBeginLoc(), D.getBeginLoc());
4587
4588 const auto *RD = CS.getCapturedRecordDecl();
4589 auto CurField = RD->field_begin();
4590
4591 Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty,
4592 /*Name=*/".zero.addr");
4593 CGF.InitTempAlloca(ZeroAddr, CGF.Builder.getInt32(/*C*/ 0));
4594 // Get the array of arguments.
4595 SmallVector<llvm::Value *, 8> Args;
4596
4597 Args.emplace_back(CGF.GetAddrOfLocalVar(&WrapperArg).getPointer());
4598 Args.emplace_back(ZeroAddr.getPointer());
4599
4600 CGBuilderTy &Bld = CGF.Builder;
4601 auto CI = CS.capture_begin();
4602
4603 // Use global memory for data sharing.
4604 // Handle passing of global args to workers.
4605 Address GlobalArgs =
4606 CGF.CreateDefaultAlignTempAlloca(CGF.VoidPtrPtrTy, "global_args");
4607 llvm::Value *GlobalArgsPtr = GlobalArgs.getPointer();
4608 llvm::Value *DataSharingArgs[] = {GlobalArgsPtr};
4609 CGF.EmitRuntimeCall(
4610 createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_get_shared_variables),
4611 DataSharingArgs);
4612
4613 // Retrieve the shared variables from the list of references returned
4614 // by the runtime. Pass the variables to the outlined function.
4615 Address SharedArgListAddress = Address::invalid();
4616 if (CS.capture_size() > 0 ||
4617 isOpenMPLoopBoundSharingDirective(D.getDirectiveKind())) {
4618 SharedArgListAddress = CGF.EmitLoadOfPointer(
4619 GlobalArgs, CGF.getContext()
4620 .getPointerType(CGF.getContext().getPointerType(
4621 CGF.getContext().VoidPtrTy))
4622 .castAs<PointerType>());
4623 }
4624 unsigned Idx = 0;
4625 if (isOpenMPLoopBoundSharingDirective(D.getDirectiveKind())) {
4626 Address Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, Idx);
4627 Address TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast(
4628 Src, CGF.SizeTy->getPointerTo());
4629 llvm::Value *LB = CGF.EmitLoadOfScalar(
4630 TypedAddress,
4631 /*Volatile=*/false,
4632 CGF.getContext().getPointerType(CGF.getContext().getSizeType()),
4633 cast<OMPLoopDirective>(D).getLowerBoundVariable()->getExprLoc());
4634 Args.emplace_back(LB);
4635 ++Idx;
4636 Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, Idx);
4637 TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast(
4638 Src, CGF.SizeTy->getPointerTo());
4639 llvm::Value *UB = CGF.EmitLoadOfScalar(
4640 TypedAddress,
4641 /*Volatile=*/false,
4642 CGF.getContext().getPointerType(CGF.getContext().getSizeType()),
4643 cast<OMPLoopDirective>(D).getUpperBoundVariable()->getExprLoc());
4644 Args.emplace_back(UB);
4645 ++Idx;
4646 }
4647 if (CS.capture_size() > 0) {
4648 ASTContext &CGFContext = CGF.getContext();
4649 for (unsigned I = 0, E = CS.capture_size(); I < E; ++I, ++CI, ++CurField) {
4650 QualType ElemTy = CurField->getType();
4651 Address Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, I + Idx);
4652 Address TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast(
4653 Src, CGF.ConvertTypeForMem(CGFContext.getPointerType(ElemTy)));
4654 llvm::Value *Arg = CGF.EmitLoadOfScalar(TypedAddress,
4655 /*Volatile=*/false,
4656 CGFContext.getPointerType(ElemTy),
4657 CI->getLocation());
4658 if (CI->capturesVariableByCopy() &&
4659 !CI->getCapturedVar()->getType()->isAnyPointerType()) {
4660 Arg = castValueToType(CGF, Arg, ElemTy, CGFContext.getUIntPtrType(),
4661 CI->getLocation());
4662 }
4663 Args.emplace_back(Arg);
4664 }
4665 }
4666
4667 emitOutlinedFunctionCall(CGF, D.getBeginLoc(), OutlinedParallelFn, Args);
4668 CGF.FinishFunction();
4669 return Fn;
4670}
4671
4672void CGOpenMPRuntimeNVPTX::emitFunctionProlog(CodeGenFunction &CGF,
4673 const Decl *D) {
4674 if (getDataSharingMode(CGM) != CGOpenMPRuntimeNVPTX::Generic)
4675 return;
4676
4677 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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 4677, __PRETTY_FUNCTION__))
;
4678 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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 4679, __PRETTY_FUNCTION__))
4679 "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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 4679, __PRETTY_FUNCTION__))
;
4680 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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 4681, __PRETTY_FUNCTION__))
4681 "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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 4681, __PRETTY_FUNCTION__))
;
4682 const Stmt *Body = nullptr;
4683 bool NeedToDelayGlobalization = false;
4684 if (const auto *FD = dyn_cast<FunctionDecl>(D)) {
4685 Body = FD->getBody();
4686 } else if (const auto *BD = dyn_cast<BlockDecl>(D)) {
4687 Body = BD->getBody();
4688 } else if (const auto *CD = dyn_cast<CapturedDecl>(D)) {
4689 Body = CD->getBody();
4690 NeedToDelayGlobalization = CGF.CapturedStmtInfo->getKind() == CR_OpenMP;
4691 if (NeedToDelayGlobalization &&
4692 getExecutionMode() == CGOpenMPRuntimeNVPTX::EM_SPMD)
4693 return;
4694 }
4695 if (!Body)
4696 return;
4697 CheckVarsEscapingDeclContext VarChecker(CGF, TeamAndReductions.second);
4698 VarChecker.Visit(Body);
4699 const RecordDecl *GlobalizedVarsRecord =
4700 VarChecker.getGlobalizedRecord(IsInTTDRegion);
4701 TeamAndReductions.first = nullptr;
4702 TeamAndReductions.second.clear();
4703 ArrayRef<const ValueDecl *> EscapedVariableLengthDecls =
4704 VarChecker.getEscapedVariableLengthDecls();
4705 if (!GlobalizedVarsRecord && EscapedVariableLengthDecls.empty())
4706 return;
4707 auto I = FunctionGlobalizedDecls.try_emplace(CGF.CurFn).first;
4708 I->getSecond().MappedParams =
4709 std::make_unique<CodeGenFunction::OMPMapVars>();
4710 I->getSecond().GlobalRecord = GlobalizedVarsRecord;
4711 I->getSecond().EscapedParameters.insert(
4712 VarChecker.getEscapedParameters().begin(),
4713 VarChecker.getEscapedParameters().end());
4714 I->getSecond().EscapedVariableLengthDecls.append(
4715 EscapedVariableLengthDecls.begin(), EscapedVariableLengthDecls.end());
4716 DeclToAddrMapTy &Data = I->getSecond().LocalVarData;
4717 for (const ValueDecl *VD : VarChecker.getEscapedDecls()) {
4718 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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 4718, __PRETTY_FUNCTION__))
;
4719 const FieldDecl *FD = VarChecker.getFieldForGlobalizedVar(VD);
4720 Data.insert(std::make_pair(VD, MappedVarData(FD, IsInTTDRegion)));
4721 }
4722 if (!IsInTTDRegion && !NeedToDelayGlobalization && !IsInParallelRegion) {
4723 CheckVarsEscapingDeclContext VarChecker(CGF, llvm::None);
4724 VarChecker.Visit(Body);
4725 I->getSecond().SecondaryGlobalRecord =
4726 VarChecker.getGlobalizedRecord(/*IsInTTDRegion=*/true);
4727 I->getSecond().SecondaryLocalVarData.emplace();
4728 DeclToAddrMapTy &Data = I->getSecond().SecondaryLocalVarData.getValue();
4729 for (const ValueDecl *VD : VarChecker.getEscapedDecls()) {
4730 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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 4730, __PRETTY_FUNCTION__))
;
4731 const FieldDecl *FD = VarChecker.getFieldForGlobalizedVar(VD);
4732 Data.insert(
4733 std::make_pair(VD, MappedVarData(FD, /*IsInTTDRegion=*/true)));
4734 }
4735 }
4736 if (!NeedToDelayGlobalization) {
4737 emitGenericVarsProlog(CGF, D->getBeginLoc(), /*WithSPMDCheck=*/true);
4738 struct GlobalizationScope final : EHScopeStack::Cleanup {
4739 GlobalizationScope() = default;
4740
4741 void Emit(CodeGenFunction &CGF, Flags flags) override {
4742 static_cast<CGOpenMPRuntimeNVPTX &>(CGF.CGM.getOpenMPRuntime())
4743 .emitGenericVarsEpilog(CGF, /*WithSPMDCheck=*/true);
4744 }
4745 };
4746 CGF.EHStack.pushCleanup<GlobalizationScope>(NormalAndEHCleanup);
4747 }
4748}
4749
4750Address CGOpenMPRuntimeNVPTX::getAddressOfLocalVariable(CodeGenFunction &CGF,
4751 const VarDecl *VD) {
4752 if (VD && VD->hasAttr<OMPAllocateDeclAttr>()) {
1
Assuming 'VD' is null
4753 const auto *A = VD->getAttr<OMPAllocateDeclAttr>();
4754 switch (A->getAllocatorType()) {
4755 // Use the default allocator here as by default local vars are
4756 // threadlocal.
4757 case OMPAllocateDeclAttr::OMPDefaultMemAlloc:
4758 case OMPAllocateDeclAttr::OMPThreadMemAlloc:
4759 case OMPAllocateDeclAttr::OMPHighBWMemAlloc:
4760 case OMPAllocateDeclAttr::OMPLowLatMemAlloc:
4761 // Follow the user decision - use default allocation.
4762 return Address::invalid();
4763 case OMPAllocateDeclAttr::OMPUserDefinedMemAlloc:
4764 // TODO: implement aupport for user-defined allocators.
4765 return Address::invalid();
4766 case OMPAllocateDeclAttr::OMPConstMemAlloc: {
4767 llvm::Type *VarTy = CGF.ConvertTypeForMem(VD->getType());
4768 auto *GV = new llvm::GlobalVariable(
4769 CGM.getModule(), VarTy, /*isConstant=*/false,
4770 llvm::GlobalValue::InternalLinkage,
4771 llvm::Constant::getNullValue(VarTy), VD->getName(),
4772 /*InsertBefore=*/nullptr, llvm::GlobalValue::NotThreadLocal,
4773 CGM.getContext().getTargetAddressSpace(LangAS::cuda_constant));
4774 CharUnits Align = CGM.getContext().getDeclAlign(VD);
4775 GV->setAlignment(Align.getAsAlign());
4776 return Address(GV, Align);
4777 }
4778 case OMPAllocateDeclAttr::OMPPTeamMemAlloc: {
4779 llvm::Type *VarTy = CGF.ConvertTypeForMem(VD->getType());
4780 auto *GV = new llvm::GlobalVariable(
4781 CGM.getModule(), VarTy, /*isConstant=*/false,
4782 llvm::GlobalValue::InternalLinkage,
4783 llvm::Constant::getNullValue(VarTy), VD->getName(),
4784 /*InsertBefore=*/nullptr, llvm::GlobalValue::NotThreadLocal,
4785 CGM.getContext().getTargetAddressSpace(LangAS::cuda_shared));
4786 CharUnits Align = CGM.getContext().getDeclAlign(VD);
4787 GV->setAlignment(Align.getAsAlign());
4788 return Address(GV, Align);
4789 }
4790 case OMPAllocateDeclAttr::OMPLargeCapMemAlloc:
4791 case OMPAllocateDeclAttr::OMPCGroupMemAlloc: {
4792 llvm::Type *VarTy = CGF.ConvertTypeForMem(VD->getType());
4793 auto *GV = new llvm::GlobalVariable(
4794 CGM.getModule(), VarTy, /*isConstant=*/false,
4795 llvm::GlobalValue::InternalLinkage,
4796 llvm::Constant::getNullValue(VarTy), VD->getName());
4797 CharUnits Align = CGM.getContext().getDeclAlign(VD);
4798 GV->setAlignment(Align.getAsAlign());
4799 return Address(GV, Align);
4800 }
4801 }
4802 }
4803
4804 if (getDataSharingMode(CGM) != CGOpenMPRuntimeNVPTX::Generic)
2
Taking false branch
4805 return Address::invalid();
4806
4807 VD = VD->getCanonicalDecl();
3
Called C++ object pointer is null
4808 auto I = FunctionGlobalizedDecls.find(CGF.CurFn);
4809 if (I == FunctionGlobalizedDecls.end())
4810 return Address::invalid();
4811 auto VDI = I->getSecond().LocalVarData.find(VD);
4812 if (VDI != I->getSecond().LocalVarData.end())
4813 return VDI->second.PrivateAddr;
4814 if (VD->hasAttrs()) {
4815 for (specific_attr_iterator<OMPReferencedVarAttr> IT(VD->attr_begin()),
4816 E(VD->attr_end());
4817 IT != E; ++IT) {
4818 auto VDI = I->getSecond().LocalVarData.find(
4819 cast<VarDecl>(cast<DeclRefExpr>(IT->getRef())->getDecl())
4820 ->getCanonicalDecl());
4821 if (VDI != I->getSecond().LocalVarData.end())
4822 return VDI->second.PrivateAddr;
4823 }
4824 }
4825
4826 return Address::invalid();
4827}
4828
4829void CGOpenMPRuntimeNVPTX::functionFinished(CodeGenFunction &CGF) {
4830 FunctionGlobalizedDecls.erase(CGF.CurFn);
4831 CGOpenMPRuntime::functionFinished(CGF);
4832}
4833
4834void CGOpenMPRuntimeNVPTX::getDefaultDistScheduleAndChunk(
4835 CodeGenFunction &CGF, const OMPLoopDirective &S,
4836 OpenMPDistScheduleClauseKind &ScheduleKind,
4837 llvm::Value *&Chunk) const {
4838 if (getExecutionMode() == CGOpenMPRuntimeNVPTX::EM_SPMD) {
4839 ScheduleKind = OMPC_DIST_SCHEDULE_static;
4840 Chunk = CGF.EmitScalarConversion(getNVPTXNumThreads(CGF),
4841 CGF.getContext().getIntTypeForBitwidth(32, /*Signed=*/0),
4842 S.getIterationVariable()->getType(), S.getBeginLoc());
4843 return;
4844 }
4845 CGOpenMPRuntime::getDefaultDistScheduleAndChunk(
4846 CGF, S, ScheduleKind, Chunk);
4847}
4848
4849void CGOpenMPRuntimeNVPTX::getDefaultScheduleAndChunk(
4850 CodeGenFunction &CGF, const OMPLoopDirective &S,
4851 OpenMPScheduleClauseKind &ScheduleKind,
4852 const Expr *&ChunkExpr) const {
4853 ScheduleKind = OMPC_SCHEDULE_static;
4854 // Chunk size is 1 in this case.
4855 llvm::APInt ChunkSize(32, 1);
4856 ChunkExpr = IntegerLiteral::Create(CGF.getContext(), ChunkSize,
4857 CGF.getContext().getIntTypeForBitwidth(32, /*Signed=*/0),
4858 SourceLocation());
4859}
4860
4861void CGOpenMPRuntimeNVPTX::adjustTargetSpecificDataForLambdas(
4862 CodeGenFunction &CGF, const OMPExecutableDirective &D) const {
4863 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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 4864, __PRETTY_FUNCTION__))
4864 " 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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 4864, __PRETTY_FUNCTION__))
;
4865 const CapturedStmt *CS = D.getCapturedStmt(OMPD_target);
4866 for (const CapturedStmt::Capture &C : CS->captures()) {
4867 // Capture variables captured by reference in lambdas for target-based
4868 // directives.
4869 if (!C.capturesVariable())
4870 continue;
4871 const VarDecl *VD = C.getCapturedVar();
4872 const auto *RD = VD->getType()
4873 .getCanonicalType()
4874 .getNonReferenceType()
4875 ->getAsCXXRecordDecl();
4876 if (!RD || !RD->isLambda())
4877 continue;
4878 Address VDAddr = CGF.GetAddrOfLocalVar(VD);
4879 LValue VDLVal;
4880 if (VD->getType().getCanonicalType()->isReferenceType())
4881 VDLVal = CGF.EmitLoadOfReferenceLValue(VDAddr, VD->getType());
4882 else
4883 VDLVal = CGF.MakeAddrLValue(
4884 VDAddr, VD->getType().getCanonicalType().getNonReferenceType());
4885 llvm::DenseMap<const VarDecl *, FieldDecl *> Captures;
4886 FieldDecl *ThisCapture = nullptr;
4887 RD->getCaptureFields(Captures, ThisCapture);
4888 if (ThisCapture && CGF.CapturedStmtInfo->isCXXThisExprCaptured()) {
4889 LValue ThisLVal =
4890 CGF.EmitLValueForFieldInitialization(VDLVal, ThisCapture);
4891 llvm::Value *CXXThis = CGF.LoadCXXThis();
4892 CGF.EmitStoreOfScalar(CXXThis, ThisLVal);
4893 }
4894 for (const LambdaCapture &LC : RD->captures()) {
4895 if (LC.getCaptureKind() != LCK_ByRef)
4896 continue;
4897 const VarDecl *VD = LC.getCapturedVar();
4898 if (!CS->capturesVariable(VD))
4899 continue;
4900 auto It = Captures.find(VD);
4901 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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 4901, __PRETTY_FUNCTION__))
;
4902 LValue VarLVal = CGF.EmitLValueForFieldInitialization(VDLVal, It->second);
4903 Address VDAddr = CGF.GetAddrOfLocalVar(VD);
4904 if (VD->getType().getCanonicalType()->isReferenceType())
4905 VDAddr = CGF.EmitLoadOfReferenceLValue(VDAddr,
4906 VD->getType().getCanonicalType())
4907 .getAddress(CGF);
4908 CGF.EmitStoreOfScalar(VDAddr.getPointer(), VarLVal);
4909 }
4910 }
4911}
4912
4913unsigned CGOpenMPRuntimeNVPTX::getDefaultFirstprivateAddressSpace() const {
4914 return CGM.getContext().getTargetAddressSpace(LangAS::cuda_constant);
4915}
4916
4917bool CGOpenMPRuntimeNVPTX::hasAllocateAttributeForGlobalVar(const VarDecl *VD,
4918 LangAS &AS) {
4919 if (!VD || !VD->hasAttr<OMPAllocateDeclAttr>())
4920 return false;
4921 const auto *A = VD->getAttr<OMPAllocateDeclAttr>();
4922 switch(A->getAllocatorType()) {
4923 case OMPAllocateDeclAttr::OMPDefaultMemAlloc:
4924 // Not supported, fallback to the default mem space.
4925 case OMPAllocateDeclAttr::OMPThreadMemAlloc:
4926 case OMPAllocateDeclAttr::OMPLargeCapMemAlloc:
4927 case OMPAllocateDeclAttr::OMPCGroupMemAlloc:
4928 case OMPAllocateDeclAttr::OMPHighBWMemAlloc:
4929 case OMPAllocateDeclAttr::OMPLowLatMemAlloc:
4930 AS = LangAS::Default;
4931 return true;
4932 case OMPAllocateDeclAttr::OMPConstMemAlloc:
4933 AS = LangAS::cuda_constant;
4934 return true;
4935 case OMPAllocateDeclAttr::OMPPTeamMemAlloc:
4936 AS = LangAS::cuda_shared;
4937 return true;
4938 case OMPAllocateDeclAttr::OMPUserDefinedMemAlloc:
4939 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-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 4940)
4940 "static storage.")::llvm::llvm_unreachable_internal("Expected predefined allocator for the variables with the "
"static storage.", "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 4940)
;
4941 }
4942 return false;
4943}
4944
4945// Get current CudaArch and ignore any unknown values
4946static CudaArch getCudaArch(CodeGenModule &CGM) {
4947 if (!CGM.getTarget().hasFeature("ptx"))
4948 return CudaArch::UNKNOWN;
4949 llvm::StringMap<bool> Features;
4950 CGM.getTarget().initFeatureMap(Features, CGM.getDiags(),
4951 CGM.getTarget().getTargetOpts().CPU,
4952 CGM.getTarget().getTargetOpts().Features);
4953 for (const auto &Feature : Features) {
4954 if (Feature.getValue()) {
4955 CudaArch Arch = StringToCudaArch(Feature.getKey());
4956 if (Arch != CudaArch::UNKNOWN)
4957 return Arch;
4958 }
4959 }
4960 return CudaArch::UNKNOWN;
4961}
4962
4963/// Check to see if target architecture supports unified addressing which is
4964/// a restriction for OpenMP requires clause "unified_shared_memory".
4965void CGOpenMPRuntimeNVPTX::checkArchForUnifiedAddressing(
4966 const OMPRequiresDecl *D) {
4967 for (const OMPClause *Clause : D->clauselists()) {
4968 if (Clause->getClauseKind() == OMPC_unified_shared_memory) {
4969 CudaArch Arch = getCudaArch(CGM);
4970 switch (Arch) {
4971 case CudaArch::SM_20:
4972 case CudaArch::SM_21:
4973 case CudaArch::SM_30:
4974 case CudaArch::SM_32:
4975 case CudaArch::SM_35:
4976 case CudaArch::SM_37:
4977 case CudaArch::SM_50:
4978 case CudaArch::SM_52:
4979 case CudaArch::SM_53:
4980 case CudaArch::SM_60:
4981 case CudaArch::SM_61:
4982 case CudaArch::SM_62: {
4983 SmallString<256> Buffer;
4984 llvm::raw_svector_ostream Out(Buffer);
4985 Out << "Target architecture " << CudaArchToString(Arch)
4986 << " does not support unified addressing";
4987 CGM.Error(Clause->getBeginLoc(), Out.str());
4988 return;
4989 }
4990 case CudaArch::SM_70:
4991 case CudaArch::SM_72:
4992 case CudaArch::SM_75:
4993 case CudaArch::GFX600:
4994 case CudaArch::GFX601:
4995 case CudaArch::GFX700:
4996 case CudaArch::GFX701:
4997 case CudaArch::GFX702:
4998 case CudaArch::GFX703:
4999 case CudaArch::GFX704:
5000 case CudaArch::GFX801:
5001 case CudaArch::GFX802:
5002 case CudaArch::GFX803:
5003 case CudaArch::GFX810:
5004 case CudaArch::GFX900:
5005 case CudaArch::GFX902:
5006 case CudaArch::GFX904:
5007 case CudaArch::GFX906:
5008 case CudaArch::GFX908:
5009 case CudaArch::GFX909:
5010 case CudaArch::GFX1010:
5011 case CudaArch::GFX1011:
5012 case CudaArch::GFX1012:
5013 case CudaArch::UNKNOWN:
5014 break;
5015 case CudaArch::LAST:
5016 llvm_unreachable("Unexpected Cuda arch.")::llvm::llvm_unreachable_internal("Unexpected Cuda arch.", "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 5016)
;
5017 }
5018 }
5019 }
5020 CGOpenMPRuntime::checkArchForUnifiedAddressing(D);
5021}
5022
5023/// Get number of SMs and number of blocks per SM.
5024static std::pair<unsigned, unsigned> getSMsBlocksPerSM(CodeGenModule &CGM) {
5025 std::pair<unsigned, unsigned> Data;
5026 if (CGM.getLangOpts().OpenMPCUDANumSMs)
5027 Data.first = CGM.getLangOpts().OpenMPCUDANumSMs;
5028 if (CGM.getLangOpts().OpenMPCUDABlocksPerSM)
5029 Data.second = CGM.getLangOpts().OpenMPCUDABlocksPerSM;
5030 if (Data.first && Data.second)
5031 return Data;
5032 switch (getCudaArch(CGM)) {
5033 case CudaArch::SM_20:
5034 case CudaArch::SM_21:
5035 case CudaArch::SM_30:
5036 case CudaArch::SM_32:
5037 case CudaArch::SM_35:
5038 case CudaArch::SM_37:
5039 case CudaArch::SM_50:
5040 case CudaArch::SM_52:
5041 case CudaArch::SM_53:
5042 return {16, 16};
5043 case CudaArch::SM_60:
5044 case CudaArch::SM_61:
5045 case CudaArch::SM_62:
5046 return {56, 32};
5047 case CudaArch::SM_70:
5048 case CudaArch::SM_72:
5049 case CudaArch::SM_75:
5050 return {84, 32};
5051 case CudaArch::GFX600:
5052 case CudaArch::GFX601:
5053 case CudaArch::GFX700:
5054 case CudaArch::GFX701:
5055 case CudaArch::GFX702:
5056 case CudaArch::GFX703:
5057 case CudaArch::GFX704:
5058 case CudaArch::GFX801:
5059 case CudaArch::GFX802:
5060 case CudaArch::GFX803:
5061 case CudaArch::GFX810:
5062 case CudaArch::GFX900:
5063 case CudaArch::GFX902:
5064 case CudaArch::GFX904:
5065 case CudaArch::GFX906:
5066 case CudaArch::GFX908:
5067 case CudaArch::GFX909:
5068 case CudaArch::GFX1010:
5069 case CudaArch::GFX1011:
5070 case CudaArch::GFX1012:
5071 case CudaArch::UNKNOWN:
5072 break;
5073 case CudaArch::LAST:
5074 llvm_unreachable("Unexpected Cuda arch.")::llvm::llvm_unreachable_internal("Unexpected Cuda arch.", "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 5074)
;
5075 }
5076 llvm_unreachable("Unexpected NVPTX target without ptx feature.")::llvm::llvm_unreachable_internal("Unexpected NVPTX target without ptx feature."
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp"
, 5076)
;
5077}
5078
5079void CGOpenMPRuntimeNVPTX::clear() {
5080 if (!GlobalizedRecords.empty()) {
5081 ASTContext &C = CGM.getContext();
5082 llvm::SmallVector<const GlobalPtrSizeRecsTy *, 4> GlobalRecs;
5083 llvm::SmallVector<const GlobalPtrSizeRecsTy *, 4> SharedRecs;
5084 RecordDecl *StaticRD = C.buildImplicitRecord(
5085 "_openmp_static_memory_type_$_", RecordDecl::TagKind::TTK_Union);
5086 StaticRD->startDefinition();
5087 RecordDecl *SharedStaticRD = C.buildImplicitRecord(
5088 "_shared_openmp_static_memory_type_$_", RecordDecl::TagKind::TTK_Union);
5089 SharedStaticRD->startDefinition();
5090 for (const GlobalPtrSizeRecsTy &Records : GlobalizedRecords) {
5091 if (Records.Records.empty())
5092 continue;
5093 unsigned Size = 0;
5094 unsigned RecAlignment = 0;
5095 for (const RecordDecl *RD : Records.Records) {
5096 QualType RDTy = C.getRecordType(RD);
5097 unsigned Alignment = C.getTypeAlignInChars(RDTy).getQuantity();
5098 RecAlignment = std::max(RecAlignment, Alignment);
5099 unsigned RecSize = C.getTypeSizeInChars(RDTy).getQuantity();
5100 Size =
5101 llvm::alignTo(llvm::alignTo(Size, Alignment) + RecSize, Alignment);
5102 }
5103 Size = llvm::alignTo(Size, RecAlignment);
5104 llvm::APInt ArySize(/*numBits=*/64, Size);
5105 QualType SubTy = C.getConstantArrayType(
5106 C.CharTy, ArySize, nullptr, ArrayType::Normal, /*IndexTypeQuals=*/0);
5107 const bool UseSharedMemory = Size <= SharedMemorySize;
5108 auto *Field =
5109 FieldDecl::Create(C, UseSharedMemory ? SharedStaticRD : StaticRD,
5110 SourceLocation(), SourceLocation(), nullptr, SubTy,
5111 C.getTrivialTypeSourceInfo(SubTy, SourceLocation()),
5112 /*BW=*/nullptr, /*Mutable=*/false,
5113 /*InitStyle=*/ICIS_NoInit);
5114 Field->setAccess(AS_public);
5115 if (UseSharedMemory) {
5116 SharedStaticRD->addDecl(Field);
5117 SharedRecs.push_back(&Records);
5118 } else {
5119 StaticRD->addDecl(Field);
5120 GlobalRecs.push_back(&Records);
5121 }
5122 Records.RecSize->setInitializer(llvm::ConstantInt::get(CGM.SizeTy, Size));
5123 Records.UseSharedMemory->setInitializer(
5124 llvm::ConstantInt::get(CGM.Int16Ty, UseSharedMemory ? 1 : 0));
5125 }
5126 // Allocate SharedMemorySize buffer for the shared memory.
5127 // FIXME: nvlink does not handle weak linkage correctly (object with the
5128 // different size are reported as erroneous).
5129 // Restore this code as sson as nvlink is fixed.
5130 if (!SharedStaticRD->field_empty()) {
5131 llvm::APInt ArySize(/*numBits=*/64, SharedMemorySize);
5132 QualType SubTy = C.getConstantArrayType(
5133 C.CharTy, ArySize, nullptr, ArrayType::Normal, /*IndexTypeQuals=*/0);
5134 auto *Field = FieldDecl::Create(
5135 C, SharedStaticRD, SourceLocation(), SourceLocation(), nullptr, SubTy,
5136 C.getTrivialTypeSourceInfo(SubTy, SourceLocation()),
5137 /*BW=*/nullptr, /*Mutable=*/false,
5138 /*InitStyle=*/ICIS_NoInit);
5139 Field->setAccess(AS_public);
5140 SharedStaticRD->addDecl(Field);
5141 }
5142 SharedStaticRD->completeDefinition();
5143 if (!SharedStaticRD->field_empty()) {
5144 QualType StaticTy = C.getRecordType(SharedStaticRD);
5145 llvm::Type *LLVMStaticTy = CGM.getTypes().ConvertTypeForMem(StaticTy);
5146 auto *GV = new llvm::GlobalVariable(
5147 CGM.getModule(), LLVMStaticTy,
5148 /*isConstant=*/false, llvm::GlobalValue::CommonLinkage,
5149 llvm::Constant::getNullValue(LLVMStaticTy),
5150 "_openmp_shared_static_glob_rd_$_", /*InsertBefore=*/nullptr,
5151 llvm::GlobalValue::NotThreadLocal,
5152 C.getTargetAddressSpace(LangAS::cuda_shared));
5153 auto *Replacement = llvm::ConstantExpr::getPointerBitCastOrAddrSpaceCast(
5154 GV, CGM.VoidPtrTy);
5155 for (const GlobalPtrSizeRecsTy *Rec : SharedRecs) {
5156 Rec->Buffer->replaceAllUsesWith(Replacement);
5157 Rec->Buffer->eraseFromParent();
5158 }
5159 }
5160 StaticRD->completeDefinition();
5161 if (!StaticRD->field_empty()) {
5162 QualType StaticTy = C.getRecordType(StaticRD);
5163 std::pair<unsigned, unsigned> SMsBlockPerSM = getSMsBlocksPerSM(CGM);
5164 llvm::APInt Size1(32, SMsBlockPerSM.second);
5165 QualType Arr1Ty =
5166 C.getConstantArrayType(StaticTy, Size1, nullptr, ArrayType::Normal,
5167 /*IndexTypeQuals=*/0);
5168 llvm::APInt Size2(32, SMsBlockPerSM.first);
5169 QualType Arr2Ty =
5170 C.getConstantArrayType(Arr1Ty, Size2, nullptr, ArrayType::Normal,
5171 /*IndexTypeQuals=*/0);
5172 llvm::Type *LLVMArr2Ty = CGM.getTypes().ConvertTypeForMem(Arr2Ty);
5173 // FIXME: nvlink does not handle weak linkage correctly (object with the
5174 // different size are reported as erroneous).
5175 // Restore CommonLinkage as soon as nvlink is fixed.
5176 auto *GV = new llvm::GlobalVariable(
5177 CGM.getModule(), LLVMArr2Ty,
5178 /*isConstant=*/false, llvm::GlobalValue::InternalLinkage,
5179 llvm::Constant::getNullValue(LLVMArr2Ty),
5180 "_openmp_static_glob_rd_$_");
5181 auto *Replacement = llvm::ConstantExpr::getPointerBitCastOrAddrSpaceCast(
5182 GV, CGM.VoidPtrTy);
5183 for (const GlobalPtrSizeRecsTy *Rec : GlobalRecs) {
5184 Rec->Buffer->replaceAllUsesWith(Replacement);
5185 Rec->Buffer->eraseFromParent();
5186 }
5187 }
5188 }
5189 if (!TeamsReductions.empty()) {
5190 ASTContext &C = CGM.getContext();
5191 RecordDecl *StaticRD = C.buildImplicitRecord(
5192 "_openmp_teams_reduction_type_$_", RecordDecl::TagKind::TTK_Union);
5193 StaticRD->startDefinition();
5194 for (const RecordDecl *TeamReductionRec : TeamsReductions) {
5195 QualType RecTy = C.getRecordType(TeamReductionRec);
5196 auto *Field = FieldDecl::Create(
5197 C, StaticRD, SourceLocation(), SourceLocation(), nullptr, RecTy,
5198 C.getTrivialTypeSourceInfo(RecTy, SourceLocation()),
5199 /*BW=*/nullptr, /*Mutable=*/false,
5200 /*InitStyle=*/ICIS_NoInit);
5201 Field->setAccess(AS_public);
5202 StaticRD->addDecl(Field);
5203 }
5204 StaticRD->completeDefinition();
5205 QualType StaticTy = C.getRecordType(StaticRD);
5206 llvm::Type *LLVMReductionsBufferTy =
5207 CGM.getTypes().ConvertTypeForMem(StaticTy);
5208 // FIXME: nvlink does not handle weak linkage correctly (object with the
5209 // different size are reported as erroneous).
5210 // Restore CommonLinkage as soon as nvlink is fixed.
5211 auto *GV = new llvm::GlobalVariable(
5212 CGM.getModule(), LLVMReductionsBufferTy,
5213 /*isConstant=*/false, llvm::GlobalValue::InternalLinkage,
5214 llvm::Constant::getNullValue(LLVMReductionsBufferTy),
5215 "_openmp_teams_reductions_buffer_$_");
5216 KernelTeamsReductionPtr->setInitializer(
5217 llvm::ConstantExpr::getPointerBitCastOrAddrSpaceCast(GV,
5218 CGM.VoidPtrTy));
5219 }
5220 CGOpenMPRuntime::clear();
5221}